From 3badb259cc97c7ab0c5e2ad560859278a17443ed Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=C2=A8NW=C2=A8?= <¨neroworld@mail.ru¨>
Date: Wed, 29 Apr 2026 21:04:22 +0100
Subject: [PATCH] feat: bidirectional research dashboard + agent config fixes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Integrate apaw_agent_model_research_v3.html as standalone dashboard
- Add model-benchmarks.json with 32 agents, 11 scored models, 11 recommendations
- Add build-research-dashboard.ts: inject live data into template → standalone HTML
- Add rebuild-template.cjs: regenerate template from v3.html source
- Add sync-benchmarks-from-yaml.cjs: sync YAML → JSON round-trip
- Add sync-model-research.ts: apply recommendation matrix to config files
- Add model-benchmarks.schema.json and model-research.schema.json for validation
- Add bidirectional-data-flow.md architecture documentation
- Add log-execution.cjs pipeline hook
- Update capability-index.yaml: add fallback_models, failover_strategy
- Update kilo-meta.json, kilo.jsonc, KILO_SPEC.md with synced models
- Update evolution.md / research.md / self-evolution.md / evolutionary-sync.md docs
- Fix security-auditor.md: quote YAML color (#DC2626)
- Fix orchestrator.md: remove duplicate devops-engineer key
- Build research-dashboard.html (106KB standalone) + dated archive
---
 .kilo/KILO_SPEC.md                            |    4 +-
 .kilo/agents/lead-developer.md                |    2 +-
 .kilo/agents/orchestrator.md                  |    1 -
 .kilo/agents/security-auditor.md              |    2 +-
 .kilo/agents/system-analyst.md                |    2 +-
 .kilo/capability-index.yaml                   | 2062 ++++++------
 .kilo/commands/evolution.md                   |  134 +-
 .kilo/commands/research.md                    |   84 +
 .kilo/rules/evolutionary-sync.md              |   50 +
 .kilo/shared/self-evolution.md                |   86 +
 agent-evolution/README.md                     |  181 +-
 agent-evolution/data/agent-versions.json      |   33 +-
 agent-evolution/data/model-benchmarks.json    | 1774 +++++++++++
 .../data/model-benchmarks.schema.json         |  553 ++++
 .../data/model-research-latest.json           |   59 +
 .../data/model-research.schema.json           |  331 ++
 .../dist/research-dashboard-2026_04_29.html   | 2777 +++++++++++++++++
 .../docs/bidirectional-data-flow.md           |  504 +++
 .../ideas/apaw_agent_model_research_v3.html   | 1168 +++++++
 agent-evolution/index.standalone.html         |   35 +-
 agent-evolution/research-dashboard.html       | 2777 +++++++++++++++++
 .../research-dashboard.template.html          | 1003 ++++++
 .../scripts/build-research-dashboard.ts       |  237 ++
 agent-evolution/scripts/rebuild-template.cjs  |   74 +
 .../scripts/sync-benchmarks-from-yaml.cjs     |  136 +
 .../scripts/sync-model-research.ts            |  651 ++++
 kilo-meta.json                                |    6 +-
 kilo.jsonc                                    |    4 +-
 scripts/log-execution.cjs                     |   41 +
 29 files changed, 13779 insertions(+), 992 deletions(-)
 create mode 100644 agent-evolution/data/model-benchmarks.json
 create mode 100644 agent-evolution/data/model-benchmarks.schema.json
 create mode 100644 agent-evolution/data/model-research-latest.json
 create mode 100644 agent-evolution/data/model-research.schema.json
 create mode 100644 agent-evolution/dist/research-dashboard-2026_04_29.html
 create mode 100644 agent-evolution/docs/bidirectional-data-flow.md
 create mode 100644 agent-evolution/ideas/apaw_agent_model_research_v3.html
 create mode 100644 agent-evolution/research-dashboard.html
 create mode 100644 agent-evolution/research-dashboard.template.html
 create mode 100644 agent-evolution/scripts/build-research-dashboard.ts
 create mode 100644 agent-evolution/scripts/rebuild-template.cjs
 create mode 100644 agent-evolution/scripts/sync-benchmarks-from-yaml.cjs
 create mode 100644 agent-evolution/scripts/sync-model-research.ts
 create mode 100644 scripts/log-execution.cjs

diff --git a/.kilo/KILO_SPEC.md b/.kilo/KILO_SPEC.md
index d9f9ad1..736f8d8 100644
--- a/.kilo/KILO_SPEC.md
+++ b/.kilo/KILO_SPEC.md
@@ -435,9 +435,9 @@ Provider availability depends on configuration. Common providers include:
 |-------|------|-------|
 | `@RequirementRefiner` | Converts vague ideas and bug reports into strict User Stories with acceptance criteria checklists. | ollama-cloud/kimi-k2-thinking |
 | `@HistoryMiner` | Analyzes git history to find duplicates and past solutions, preventing regression and duplicate work. | ollama-cloud/nemotron-3-super |
-| `@SystemAnalyst` | Designs technical specifications, data schemas, and API contracts before implementation. | ollama-cloud/glm-5.1 |
+| `@SystemAnalyst` | Designs technical specifications, data schemas, and API contracts before implementation. | ollama-cloud/nemotron-3-super |
 | `@SdetEngineer` | Writes tests following TDD methodology. | ollama-cloud/qwen3-coder:480b |
-| `@LeadDeveloper` | Primary code writer for backend and core logic. | ollama-cloud/qwen3-coder:480b |
+| `@LeadDeveloper` | Primary code writer for backend and core logic. | ollama-cloud/nemotron-3-super |
 | `@FrontendDeveloper` | Handles UI implementation with multimodal capabilities. | ollama-cloud/kimi-k2.5 |
 | `@BackendDeveloper` | Backend specialist for Node. | ollama-cloud/deepseek-v3.2 |
 | `@GoDeveloper` | Go backend specialist for Gin, Echo, APIs, and database integration. | ollama-cloud/qwen3-coder:480b |
diff --git a/.kilo/agents/lead-developer.md b/.kilo/agents/lead-developer.md
index 04d46a1..1a9b75f 100755
--- a/.kilo/agents/lead-developer.md
+++ b/.kilo/agents/lead-developer.md
@@ -1,7 +1,7 @@
 ---
 description: Primary code writer for backend and core logic. Writes implementation to pass tests
 mode: subagent
-model: ollama-cloud/qwen3-coder:480b
+model: ollama-cloud/nemotron-3-super
 variant: thinking
 color: "#DC2626"
 permission:
diff --git a/.kilo/agents/orchestrator.md b/.kilo/agents/orchestrator.md
index 0f047e7..3ef8793 100755
--- a/.kilo/agents/orchestrator.md
+++ b/.kilo/agents/orchestrator.md
@@ -40,7 +40,6 @@ permission:
     "planner": allow
     "reflector": allow
     "memory-manager": allow
-    "devops-engineer": allow
 ---
 
 # Kilo Code: Orchestrator
diff --git a/.kilo/agents/security-auditor.md b/.kilo/agents/security-auditor.md
index 4698551..1063daa 100755
--- a/.kilo/agents/security-auditor.md
+++ b/.kilo/agents/security-auditor.md
@@ -2,7 +2,7 @@
 description: Scans for security vulnerabilities, OWASP Top 10, dependency CVEs, and hardcoded secrets
 mode: subagent
 model: ollama-cloud/nemotron-3-super
-color: #DC2626
+color: "#DC2626"
 permission:
   read: allow
   bash: allow
diff --git a/.kilo/agents/system-analyst.md b/.kilo/agents/system-analyst.md
index 16c7ec1..23c470a 100755
--- a/.kilo/agents/system-analyst.md
+++ b/.kilo/agents/system-analyst.md
@@ -1,7 +1,7 @@
 ---
 description: Designs technical specifications, data schemas, and API contracts before implementation
 mode: subagent
-model: ollama-cloud/glm-5.1
+model: ollama-cloud/nemotron-3-super
 color: "#0891B2"
 permission:
   read: allow
diff --git a/.kilo/capability-index.yaml b/.kilo/capability-index.yaml
index ddd438a..802946c 100644
--- a/.kilo/capability-index.yaml
+++ b/.kilo/capability-index.yaml
@@ -1,972 +1,1090 @@
-# Capability Index
-# Maps agent capabilities for orchestrator routing
-
-agents:
-  # Core Development
-  lead-developer:
-    capabilities:
-      - code_writing
-      - refactoring
-      - bug_fixing
-      - implementation
-    receives:
-      - tests
-      - specifications
-      - architecture_docs
-    produces:
-      - code
-      - documentation_inline
-    forbidden:
-      - test_writing
-      - code_review
-    model: ollama-cloud/qwen3-coder:480b
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - orchestrator
-
-  frontend-developer:
-    capabilities:
-      - ui_implementation
-      - component_creation
-      - styling
-      - responsive_design
-      - nextjs_development
-      - vue_nuxt_development
-      - react_development
-    receives:
-      - designs
-      - wireframes
-      - api_endpoints
-    produces:
-      - vue_components
-      - react_components
-      - nextjs_pages
-      - nuxt_pages
-      - css_styles
-      - frontend_tests
-    forbidden:
-      - backend_code
-    model: ollama-cloud/qwen3-coder:480b
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - visual-tester
-      - orchestrator
-
-  php-developer:
-    capabilities:
-      - php_web_development
-      - laravel_development
-      - symfony_development
-      - wordpress_development
-      - php_api_development
-      - php_database_design
-      - php_authentication
-      - php_modular_architecture
-      - php_testing
-      - php_security
-    receives:
-      - api_specifications
-      - database_requirements
-      - ui_requirements
-    produces:
-      - laravel_routes
-      - php_models
-      - php_services
-      - php_controllers
-      - php_migrations
-      - php_tests
-      - wordpress_plugins
-    forbidden:
-      - frontend_code
-      - non_php_backend
-    model: ollama-cloud/qwen3-coder:480b
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - security-auditor
-      - orchestrator
-
-  python-developer:
-    capabilities:
-      - python_web_development
-      - django_development
-      - fastapi_development
-      - python_api_development
-      - python_database_design
-      - python_authentication
-      - python_async_patterns
-      - python_testing
-      - python_security
-    receives:
-      - api_specifications
-      - database_requirements
-    produces:
-      - django_views
-      - fastapi_routers
-      - python_models
-      - python_services
-      - python_schemas
-      - python_migrations
-      - python_tests
-    forbidden:
-      - frontend_code
-      - non_python_backend
-    model: ollama-cloud/qwen3-coder:480b
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - security-auditor
-      - orchestrator
-
-  backend-developer:
-    capabilities:
-      - api_development
-      - database_design
-      - server_logic
-      - authentication
-      - postgresql_integration
-      - sqlite_integration
-    receives:
-      - api_specifications
-      - database_requirements
-    produces:
-      - express_routes
-      - database_schema
-      - api_documentation
-    forbidden:
-      - frontend_code
-    model: ollama-cloud/qwen3-coder:480b
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - orchestrator
-
-  go-developer:
-    capabilities:
-      - go_api_development
-      - go_database_design
-      - go_concurrent_programming
-      - go_authentication
-      - go_microservices
-      - postgresql_integration
-      - sqlite_integration
-      - clickhouse_integration
-    receives:
-      - api_specifications
-      - database_requirements
-      - concurrent_requirements
-    produces:
-      - go_handlers
-      - go_database_schema
-      - go_api_documentation
-      - concurrent_solutions
-    forbidden:
-      - frontend_code
-    model: ollama-cloud/qwen3-coder:480b
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - orchestrator
-
-  flutter-developer:
-    capabilities:
-      - dart_programming
-      - flutter_ui
-      - mobile_app_development
-      - widget_creation
-      - state_management
-    receives:
-      - ui_designs
-      - api_specifications
-      - mobile_requirements
-    produces:
-      - flutter_widgets
-      - dart_code
-      - mobile_app
-    forbidden:
-      - backend_code
-      - web_development
-    model: ollama-cloud/qwen3-coder:480b
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - visual-tester
-      - orchestrator
-
-  devops-engineer:
-    capabilities:
-      - docker_configuration
-      - kubernetes_setup
-      - ci_cd_pipeline
-      - infrastructure_automation
-      - container_optimization
-    receives:
-      - deployment_requirements
-      - infrastructure_needs
-    produces:
-      - docker_compose
-      - kubernetes_manifests
-      - ci_cd_config
-    forbidden:
-      - application_code
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - security-auditor
-      - orchestrator
-
-  # Quality Assurance
-  sdet-engineer:
-    capabilities:
-      - unit_tests
-      - integration_tests
-      - e2e_tests
-      - test_planning
-      - visual_regression
-    receives:
-      - code
-      - requirements
-    produces:
-      - test_files
-      - test_reports
-      - coverage_reports
-    forbidden:
-      - implementation_code
-    model: ollama-cloud/qwen3-coder:480b
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - lead-developer
-      - orchestrator
-
-  code-skeptic:
-    capabilities:
-      - code_review
-      - security_review
-      - style_check
-      - issue_identification
-    receives:
-      - code
-    produces:
-      - review_comments
-      - approval_status
-      - issue_list
-    forbidden:
-      - suggest_implementations
-      - write_code
-    model: ollama-cloud/minimax-m2.5
-    mode: subagent
-    delegates_to:
-      - the-fixer
-      - performance-engineer
-      - orchestrator
-
-  # Security & Performance
-  security-auditor:
-    capabilities:
-      - vulnerability_scan
-      - owasp_check
-      - secret_detection
-      - auth_review
-    receives:
-      - code
-      - configuration
-    produces:
-      - security_report
-      - vulnerability_list
-    forbidden:
-      - fix_vulnerabilities
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to:
-      - the-fixer
-      - release-manager
-      - orchestrator
-
-  performance-engineer:
-    capabilities:
-      - performance_analysis
-      - n_plus_one_detection
-      - memory_leak_check
-      - algorithm_analysis
-    receives:
-      - code
-      - performance_requirements
-    produces:
-      - performance_report
-      - optimization_suggestions
-    forbidden:
-      - write_code
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to:
-      - the-fixer
-      - security-auditor
-      - orchestrator
-
-  the-fixer:
-    capabilities:
-      - bug_fixing
-      - issue_resolution
-      - code_correction
-    receives:
-      - issue_list
-      - code_context
-    produces:
-      - code_fixes
-      - resolution_notes
-    forbidden:
-      - feature_development
-    model: ollama-cloud/minimax-m2.5
-    mode: subagent
-    delegates_to:
-      - code-skeptic
-      - orchestrator
-
-  # Specialized Development
-  browser-automation:
-    capabilities:
-      - e2e_browser_tests
-      - form_filling
-      - navigation_testing
-      - screenshot_capture
-    receives:
-      - test_scenarios
-      - url_list
-    produces:
-      - test_results
-      - screenshots
-    forbidden:
-      - unit_testing
-    model: ollama-cloud/qwen3-coder:480b
-    mode: subagent
-    delegates_to:
-      - orchestrator
-
-  visual-tester:
-    capabilities:
-      - visual_regression
-      - pixel_comparison
-      - screenshot_diff
-      - ui_validation
-      - bbox_element_extraction
-      - console_error_detection
-      - network_error_detection
-      - responsive_layout_check
-      - button_overflow_detection
-      - gitea_integration
-      - docker_networking
-    receives:
-      - url
-      - baseline_screenshots
-      - page_paths
-      - gitea_issue_number
-    produces:
-      - diff_report
-      - visual_issues
-      - element_map_with_bbox
-      - console_error_report
-      - network_error_report
-      - gitea_comment
-      - gitea_attachments
-    forbidden:
-      - code_changes
-    model: ollama-cloud/qwen3-coder:480b
-    mode: subagent
-    delegates_to:
-      - the-fixer
-      - orchestrator
-
-  # Analysis & Design
-  system-analyst:
-    capabilities:
-      - architecture_design
-      - api_specification
-      - database_modeling
-      - technical_documentation
-    receives:
-      - requirements
-      - user_stories
-    produces:
-      - architecture_docs
-      - api_specs
-      - database_schemas
-    forbidden:
-      - implementation
-    model: ollama-cloud/glm-5.1
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - sdet-engineer
-      - orchestrator
-
-  requirement-refiner:
-    capabilities:
-      - requirement_analysis
-      - user_story_creation
-      - acceptance_criteria
-      - clarification
-    receives:
-      - raw_requests
-      - feature_ideas
-    produces:
-      - user_stories
-      - acceptance_criteria
-      - requirements_doc
-    forbidden:
-      - design_decisions
-    model: ollama-cloud/glm-5.1
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - history-miner
-      - system-analyst
-
-  history-miner:
-    capabilities:
-      - git_search
-      - duplicate_detection
-      - past_solution_finder
-      - pattern_identification
-    receives:
-      - search_query
-      - issue_description
-    produces:
-      - commit_list
-      - duplicate_report
-      - related_files
-    forbidden:
-      - code_changes
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to: []
-
-  capability-analyst:
-    capabilities:
-      - gap_analysis
-      - capability_mapping
-      - recommendation_generation
-      - coverage_analysis
-    receives:
-      - task_requirements
-    produces:
-      - analysis_report
-      - recommendations
-      - new_agent_specs
-    forbidden:
-      - implementation
-    model: ollama-cloud/glm-5.1
-    mode: subagent
-    delegates_to:
-      - agent-architect
-      - orchestrator
-
-  # Process Management
-  orchestrator:
-    capabilities:
-      - task_routing
-      - state_management
-      - agent_coordination
-      - workflow_execution
-    receives:
-      - issue
-      - status_change
-    produces:
-      - routing_decisions
-      - status_updates
-    forbidden:
-      - code_writing
-      - code_review
-    model: ollama-cloud/kimi-k2.6:cloud
-    variant: thinking
-    mode: all
-    delegates_to:
-      - history-miner
-      - system-analyst
-      - sdet-engineer
-      - lead-developer
-      - code-skeptic
-      - the-fixer
-      - frontend-developer
-      - backend-developer
-      - php-developer
-      - python-developer
-      - go-developer
-      - flutter-developer
-      - performance-engineer
-      - security-auditor
-      - visual-tester
-      - browser-automation
-      - devops-engineer
-      - release-manager
-      - requirement-refiner
-      - capability-analyst
-      - workflow-architect
-      - markdown-validator
-      - evaluator
-      - prompt-optimizer
-      - product-owner
-      - pipeline-judge
-      - planner
-      - reflector
-      - memory-manager
-      - agent-architect
-      - architect-indexer
-
-  release-manager:
-    capabilities:
-      - git_operations
-      - version_management
-      - changelog_creation
-      - deployment
-    receives:
-      - approved_code
-      - release_request
-    produces:
-      - commits
-      - tags
-      - releases
-    forbidden:
-      - code_changes
-      - feature_development
-    model: ollama-cloud/glm-5.1
-    mode: subagent
-    delegates_to:
-      - evaluator
-
-  evaluator:
-    capabilities:
-      - performance_scoring
-      - process_analysis
-      - pattern_identification
-      - improvement_recommendations
-    receives:
-      - completed_issue
-      - agent_logs
-    produces:
-      - performance_report
-      - scores
-      - recommendations
-    forbidden:
-      - code_changes
-    model: ollama-cloud/glm-5.1
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - prompt-optimizer
-      - product-owner
-      - orchestrator
-
-  prompt-optimizer:
-    capabilities:
-      - prompt_analysis
-      - prompt_improvement
-      - failure_pattern_detection
-    receives:
-      - low_scores
-      - failure_reports
-    produces:
-      - improved_prompts
-      - optimization_report
-    forbidden:
-      - agent_creation
-    model: ollama-cloud/glm-5.1
-    variant: instant
-    mode: subagent
-    delegates_to: []
-
-  product-owner:
-    capabilities:
-      - issue_management
-      - prioritization
-      - backlog_management
-      - workflow_completion
-    receives:
-      - completed_work
-      - stakeholder_requests
-    produces:
-      - priority_order
-      - issue_labels
-      - issue closures
-    forbidden:
-      - implementation
-    model: ollama-cloud/glm-5.1
-    mode: subagent
-    delegates_to: []
-
-  pipeline-judge:
-    capabilities:
-      - test_execution
-      - fitness_scoring
-      - metric_collection
-      - bottleneck_detection
-    receives:
-      - completed_workflow
-      - pipeline_logs
-    produces:
-      - fitness_report
-      - bottleneck_analysis
-      - improvement_triggers
-    forbidden:
-      - code_writing
-      - code_changes
-      - prompt_changes
-    model: ollama-cloud/glm-5.1
-    mode: subagent
-    delegates_to:
-      - prompt-optimizer
-
-  # Workflow
-  workflow-architect:
-    capabilities:
-      - workflow_design
-      - process_definition
-      - automation_setup
-    receives:
-      - workflow_requirements
-    produces:
-      - workflow_definitions
-      - command_files
-    forbidden:
-      - execution
-    model: ollama-cloud/glm-5.1
-    variant: thinking
-    mode: subagent
-    delegates_to: []
-
-  # Validation
-  markdown-validator:
-    capabilities:
-      - markdown_validation
-      - formatting_check
-      - link_validation
-    receives:
-      - markdown_files
-    produces:
-      - validation_report
-      - corrections
-    forbidden:
-      - content_creation
-    model: ollama-cloud/nemotron-3-nano:30b
-    mode: subagent
-    delegates_to:
-      - orchestrator
-
-  agent-architect:
-    capabilities:
-      - agent_design
-      - prompt_engineering
-      - capability_definition
-    receives:
-      - agent_requirements
-    produces:
-      - agent_definition
-      - integration_plan
-    forbidden:
-      - agent_execution
-    model: ollama-cloud/glm-5.1
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - capability-analyst
-      - requirement-refiner
-      - system-analyst
-
-  # Cognitive Enhancement
-  planner:
-    capabilities:
-      - task_decomposition
-      - chain_of_thought
-      - tree_of_thoughts
-      - plan_execute_reflect
-      - dependency_analysis
-    receives:
-      - complex_task
-      - objective
-    produces:
-      - decomposed_steps
-      - dependency_graph
-      - success_criteria
-    forbidden:
-      - implementation
-      - execution
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to: []
-
-  reflector:
-    capabilities:
-      - self_reflection
-      - mistake_analysis
-      - lesson_extraction
-      - trajectory_analysis
-      - heuristic_evaluation
-    receives:
-      - action_trajectory
-      - task_result
-    produces:
-      - reflection_report
-      - lessons_learned
-      - improved_approach
-    forbidden:
-      - implementation
-      - code_changes
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to: []
-
-  memory-manager:
-    capabilities:
-      - memory_retrieval
-      - memory_storage
-      - memory_consolidation
-      - relevance_scoring
-      - episodic_management
-    receives:
-      - query
-      - memory_type
-    produces:
-      - retrieved_memories
-      - relevance_scores
-      - consolidated_memories
-    forbidden:
-      - code_changes
-      - implementation
-    model: ollama-cloud/nemotron-3-super
-    mode: subagent
-    delegates_to: []
-
-  # Project Mapping
-  architect-indexer:
-    capabilities:
-      - codebase_indexing
-      - project_mapping
-      - architecture_documentation
-      - dependency_analysis
-      - entity_extraction
-      - api_surface_discovery
-      - convention_detection
-      - staleness_detection
-    receives:
-      - project_root_directory
-      - stale_sections_list
-    produces:
-      - .architect/state.json
-      - .architect/project.json
-      - .architect/README.md
-      - architecture_overview
-      - dependency_graph
-      - entity_documentation
-      - db_schema_documentation
-      - api_surface_documentation
-      - convention_documentation
-      - file_graph
-      - module_graph
-    forbidden:
-      - code_changes
-      - implementation
-    model: ollama-cloud/glm-5.1
-    variant: thinking
-    mode: subagent
-    delegates_to:
-      - system-analyst
-      - orchestrator
-
-  # Capability Routing Map
-  capability_routing:
-    code_writing: lead-developer
-    code_review: code-skeptic
-    test_writing: sdet-engineer
-    architecture: system-analyst
-    security: security-auditor
-    performance: performance-engineer
-    bug_fixing: the-fixer
-    git_operations: release-manager
-    ui_implementation: frontend-developer
-    nextjs_development: frontend-developer
-    vue_nuxt_development: frontend-developer
-    react_development: frontend-developer
-    e2e_testing: browser-automation
-    visual_testing: visual-tester
-    bbox_extraction: visual-tester
-    console_error_detection: visual-tester
-    gitea_integration: visual-tester
-
-    docker_networking: visual-tester
-    requirement_analysis: requirement-refiner
-    gap_analysis: capability-analyst
-    issue_management: product-owner
-    prompt_optimization: prompt-optimizer
-    workflow_design: workflow-architect
-    scoring: evaluator
-    duplicate_detection: history-miner
-    agent_design: agent-architect
-    markdown_validation: markdown-validator
-    # Database integrations
-    postgresql_integration: backend-developer
-    sqlite_integration: backend-developer
-    clickhouse_integration: go-developer
-    # Mobile development
-    flutter_development: flutter-developer
-    # PHP Development
-    php_web_development: php-developer
-    laravel_development: php-developer
-    symfony_development: php-developer
-    wordpress_development: php-developer
-    # Python Development
-    python_web_development: python-developer
-    django_development: python-developer
-    fastapi_development: python-developer
-    # DevOps
-    docker_configuration: devops-engineer
-    kubernetes_setup: devops-engineer
-    ci_cd_pipeline: devops-engineer
-    # Cognitive Enhancement (New)
-    task_decomposition: planner
-    self_reflection: reflector
-    memory_retrieval: memory-manager
-    chain_of_thought: planner
-    tree_of_thoughts: planner
-    # Fitness & Evolution
-    fitness_scoring: pipeline-judge
-    test_execution: pipeline-judge
-    bottleneck_detection: pipeline-judge
-    # Go Development
-    go_api_development: go-developer
-    go_database_design: go-developer
-    go_concurrent_programming: go-developer
-    go_authentication: go-developer
-    go_microservices: go-developer
-    # Project Mapping
-    codebase_indexing: architect-indexer
-    project_mapping: architect-indexer
-    architecture_documentation: architect-indexer
-    dependency_analysis: architect-indexer
-    entity_extraction: architect-indexer
-    api_surface_discovery: architect-indexer
-    convention_detection: architect-indexer
-
-# Parallelizable Tasks
-parallel_groups:
-  review_phase:
-    - security-auditor
-    - performance-engineer
-    - code-skeptic
-  testing_phase:
-    - sdet-engineer
-    - browser-automation
-    - visual-tester
-
-# Evaluator-Optimizer Patterns
-iteration_loops:
-  code_review:
-    evaluator: code-skeptic
-    optimizer: the-fixer
-    max_iterations: 3
-    convergence: all_issues_resolved
-
-  security_review:
-    evaluator: security-auditor
-    optimizer: the-fixer
-    max_iterations: 2
-    convergence: no_critical_vulnerabilities
-
-  performance_review:
-    evaluator: performance-engineer
-    optimizer: the-fixer
-    max_iterations: 2
-    convergence: all_perf_issues_resolved
-
-  # Evolution loop for continuous improvement
-  evolution:
-    evaluator: pipeline-judge
-    optimizer: prompt-optimizer
-    max_iterations: 3
-    convergence: fitness_above_0.85
-
-# Quality Gates
-quality_gates:
-  requirements:
-    - user_stories_defined
-    - acceptance_criteria_complete
-    - technical_constraints_documented
-  
-  architecture:
-    - schema_valid
-    - endpoints_documented
-    - tech_stack_decided
-  
-  implementation:
-    - build_success
-    - no_type_errors
-    - no_lint_errors
-  
-  testing:
-    - coverage_gte_80
-    - all_tests_pass
-    - no_critical_bugs
-  
-  review:
-    - no_critical_issues
-    - no_security_vulnerabilities
-    - performance_acceptable
-  
-  docker:
-    - build_success
-    - health_check_pass
-    - size_under_limit
-  
-  documentation:
-    - readme_complete
-    - api_docs_complete
-    - deployment_guide_complete
-
-# State Transitions
-workflow_states:
-  new: [planned]
-  planned: [researching]
-  researching: [designed]
-  designed: [testing]
-  testing: [implementing]
-  implementing: [reviewing]
-  reviewing: [fixing, perf_check]
-  fixing: [reviewing]
-  perf_check: [security_check]
-  security_check: [releasing]
-  releasing: [evaluated]
-  evaluated: [evolving, completed]
-  evolving: [evaluated]
-  completed: []
-
-# Evolution Configuration
-evolution:
-  enabled: true
-  auto_trigger: true           # trigger after every workflow
-  fitness_threshold: 0.70      # below this → auto-optimize
-  max_evolution_attempts: 3    # max retries per cycle
-  fitness_history: .kilo/logs/fitness-history.jsonl
-  token_budget_default: 50000
-  time_budget_default: 300
-  budgets:
-    feature:
-      tokens: 50000
-      time_s: 300
-      min_coverage: 80
-    bugfix:
-      tokens: 20000
-      time_s: 120
-      min_coverage: 90
-    refactor:
-      tokens: 40000
-      time_s: 240
-      min_coverage: 95
-    security:
-      tokens: 30000
-      time_s: 180
-      min_coverage: 80
+agents:
+  lead-developer:
+    capabilities:
+    - code_writing
+    - refactoring
+    - bug_fixing
+    - implementation
+    receives:
+    - tests
+    - specifications
+    - architecture_docs
+    produces:
+    - code
+    - documentation_inline
+    forbidden:
+    - test_writing
+    - code_review
+    model: ollama-cloud/nemotron-3-super
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/qwen3-coder:480b
+    - ollama-cloud/kimi-k2.6:cloud
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: downgraded
+  frontend-developer:
+    capabilities:
+    - ui_implementation
+    - component_creation
+    - styling
+    - responsive_design
+    - nextjs_development
+    - vue_nuxt_development
+    - react_development
+    receives:
+    - designs
+    - wireframes
+    - api_endpoints
+    produces:
+    - vue_components
+    - react_components
+    - nextjs_pages
+    - nuxt_pages
+    - css_styles
+    - frontend_tests
+    forbidden:
+    - backend_code
+    model: ollama-cloud/qwen3-coder:480b
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - visual-tester
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/minimax-m2.5
+    - ollama-cloud/minimax-m2.7
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  php-developer:
+    capabilities:
+    - php_web_development
+    - laravel_development
+    - symfony_development
+    - wordpress_development
+    - php_api_development
+    - php_database_design
+    - php_authentication
+    - php_modular_architecture
+    - php_testing
+    - php_security
+    receives:
+    - api_specifications
+    - database_requirements
+    - ui_requirements
+    produces:
+    - laravel_routes
+    - php_models
+    - php_services
+    - php_controllers
+    - php_migrations
+    - php_tests
+    - wordpress_plugins
+    forbidden:
+    - frontend_code
+    - non_php_backend
+    model: ollama-cloud/qwen3-coder:480b
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - security-auditor
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - openrouter/qwen/qwen3.6-plus:free
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/minimax-m2.5
+    failover_strategy: downgraded
+  python-developer:
+    capabilities:
+    - python_web_development
+    - django_development
+    - fastapi_development
+    - python_api_development
+    - python_database_design
+    - python_authentication
+    - python_async_patterns
+    - python_testing
+    - python_security
+    receives:
+    - api_specifications
+    - database_requirements
+    produces:
+    - django_views
+    - fastapi_routers
+    - python_models
+    - python_services
+    - python_schemas
+    - python_migrations
+    - python_tests
+    forbidden:
+    - frontend_code
+    - non_python_backend
+    model: ollama-cloud/qwen3-coder:480b
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - security-auditor
+    - orchestrator
+    fallback_models:
+    - openrouter/qwen/qwen3.6-plus:free
+    - ollama-cloud/kimi-k2.6:cloud
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/minimax-m2.5
+    failover_strategy: downgraded
+  backend-developer:
+    capabilities:
+    - api_development
+    - database_design
+    - server_logic
+    - authentication
+    - postgresql_integration
+    - sqlite_integration
+    receives:
+    - api_specifications
+    - database_requirements
+    produces:
+    - express_routes
+    - database_schema
+    - api_documentation
+    forbidden:
+    - frontend_code
+    model: ollama-cloud/qwen3-coder:480b
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - openrouter/qwen/qwen3.6-plus:free
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: downgraded
+  go-developer:
+    capabilities:
+    - go_api_development
+    - go_database_design
+    - go_concurrent_programming
+    - go_authentication
+    - go_microservices
+    - postgresql_integration
+    - sqlite_integration
+    - clickhouse_integration
+    receives:
+    - api_specifications
+    - database_requirements
+    - concurrent_requirements
+    produces:
+    - go_handlers
+    - go_database_schema
+    - go_api_documentation
+    - concurrent_solutions
+    forbidden:
+    - frontend_code
+    model: ollama-cloud/qwen3-coder:480b
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  flutter-developer:
+    capabilities:
+    - dart_programming
+    - flutter_ui
+    - mobile_app_development
+    - widget_creation
+    - state_management
+    receives:
+    - ui_designs
+    - api_specifications
+    - mobile_requirements
+    produces:
+    - flutter_widgets
+    - dart_code
+    - mobile_app
+    forbidden:
+    - backend_code
+    - web_development
+    model: ollama-cloud/qwen3-coder:480b
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - visual-tester
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - openrouter/qwen/qwen3.6-plus:free
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: downgraded
+  devops-engineer:
+    capabilities:
+    - docker_configuration
+    - kubernetes_setup
+    - ci_cd_pipeline
+    - infrastructure_automation
+    - container_optimization
+    receives:
+    - deployment_requirements
+    - infrastructure_needs
+    produces:
+    - docker_compose
+    - kubernetes_manifests
+    - ci_cd_config
+    forbidden:
+    - application_code
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - security-auditor
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/glm-5.1
+    failover_strategy: downgraded
+  sdet-engineer:
+    capabilities:
+    - unit_tests
+    - integration_tests
+    - e2e_tests
+    - test_planning
+    - visual_regression
+    receives:
+    - code
+    - requirements
+    produces:
+    - test_files
+    - test_reports
+    - coverage_reports
+    forbidden:
+    - implementation_code
+    model: ollama-cloud/qwen3-coder:480b
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - lead-developer
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/minimax-m2.5
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: downgraded
+  code-skeptic:
+    capabilities:
+    - code_review
+    - security_review
+    - style_check
+    - issue_identification
+    receives:
+    - code
+    produces:
+    - review_comments
+    - approval_status
+    - issue_list
+    forbidden:
+    - suggest_implementations
+    - write_code
+    model: ollama-cloud/minimax-m2.5
+    mode: subagent
+    delegates_to:
+    - the-fixer
+    - performance-engineer
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/qwen3-coder:480b
+    - ollama-cloud/deepseek-v4-pro-max
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/kimi-k2.6:cloud
+    failover_strategy: mixed
+  security-auditor:
+    capabilities:
+    - vulnerability_scan
+    - owasp_check
+    - secret_detection
+    - auth_review
+    receives:
+    - code
+    - configuration
+    produces:
+    - security_report
+    - vulnerability_list
+    forbidden:
+    - fix_vulnerabilities
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to:
+    - the-fixer
+    - release-manager
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/glm-5.1
+    failover_strategy: downgraded
+  performance-engineer:
+    capabilities:
+    - performance_analysis
+    - n_plus_one_detection
+    - memory_leak_check
+    - algorithm_analysis
+    receives:
+    - code
+    - performance_requirements
+    produces:
+    - performance_report
+    - optimization_suggestions
+    forbidden:
+    - write_code
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to:
+    - the-fixer
+    - security-auditor
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/glm-5.1
+    - ollama-cloud/kimi-k2.6:cloud
+    failover_strategy: downgraded
+  the-fixer:
+    capabilities:
+    - bug_fixing
+    - issue_resolution
+    - code_correction
+    receives:
+    - issue_list
+    - code_context
+    produces:
+    - code_fixes
+    - resolution_notes
+    forbidden:
+    - feature_development
+    model: ollama-cloud/minimax-m2.5
+    mode: subagent
+    delegates_to:
+    - code-skeptic
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/qwen3-coder:480b
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: mixed
+  browser-automation:
+    capabilities:
+    - e2e_browser_tests
+    - form_filling
+    - navigation_testing
+    - screenshot_capture
+    receives:
+    - test_scenarios
+    - url_list
+    produces:
+    - test_results
+    - screenshots
+    forbidden:
+    - unit_testing
+    model: ollama-cloud/qwen3-coder:480b
+    mode: subagent
+    delegates_to:
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  visual-tester:
+    capabilities:
+    - visual_regression
+    - pixel_comparison
+    - screenshot_diff
+    - ui_validation
+    - bbox_element_extraction
+    - console_error_detection
+    - network_error_detection
+    - responsive_layout_check
+    - button_overflow_detection
+    - gitea_integration
+    - docker_networking
+    receives:
+    - url
+    - baseline_screenshots
+    - page_paths
+    - gitea_issue_number
+    produces:
+    - diff_report
+    - visual_issues
+    - element_map_with_bbox
+    - console_error_report
+    - network_error_report
+    - gitea_comment
+    - gitea_attachments
+    forbidden:
+    - code_changes
+    model: ollama-cloud/qwen3-coder:480b
+    mode: subagent
+    delegates_to:
+    - the-fixer
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  system-analyst:
+    capabilities:
+    - architecture_design
+    - api_specification
+    - database_modeling
+    - technical_documentation
+    receives:
+    - requirements
+    - user_stories
+    produces:
+    - architecture_docs
+    - api_specs
+    - database_schemas
+    forbidden:
+    - implementation
+    model: ollama-cloud/nemotron-3-super
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - sdet-engineer
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/glm-5.1
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    failover_strategy: downgraded
+  requirement-refiner:
+    capabilities:
+    - requirement_analysis
+    - user_story_creation
+    - acceptance_criteria
+    - clarification
+    receives:
+    - raw_requests
+    - feature_ideas
+    produces:
+    - user_stories
+    - acceptance_criteria
+    - requirements_doc
+    forbidden:
+    - design_decisions
+    model: ollama-cloud/glm-5.1
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - history-miner
+    - system-analyst
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/glm-5
+    failover_strategy: mixed
+  history-miner:
+    capabilities:
+    - git_search
+    - duplicate_detection
+    - past_solution_finder
+    - pattern_identification
+    receives:
+    - search_query
+    - issue_description
+    produces:
+    - commit_list
+    - duplicate_report
+    - related_files
+    forbidden:
+    - code_changes
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - ollama-cloud/glm-5.1
+    - ollama-cloud/deepseek-v4-pro-max
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: mixed
+  capability-analyst:
+    capabilities:
+    - gap_analysis
+    - capability_mapping
+    - recommendation_generation
+    - coverage_analysis
+    receives:
+    - task_requirements
+    produces:
+    - analysis_report
+    - recommendations
+    - new_agent_specs
+    forbidden:
+    - implementation
+    model: ollama-cloud/glm-5.1
+    mode: subagent
+    delegates_to:
+    - agent-architect
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  orchestrator:
+    capabilities:
+    - task_routing
+    - state_management
+    - agent_coordination
+    - workflow_execution
+    receives:
+    - issue
+    - status_change
+    produces:
+    - routing_decisions
+    - status_updates
+    forbidden:
+    - code_writing
+    - code_review
+    model: ollama-cloud/kimi-k2.6:cloud
+    variant: thinking
+    mode: all
+    delegates_to:
+    - history-miner
+    - system-analyst
+    - sdet-engineer
+    - lead-developer
+    - code-skeptic
+    - the-fixer
+    - frontend-developer
+    - backend-developer
+    - php-developer
+    - python-developer
+    - go-developer
+    - flutter-developer
+    - performance-engineer
+    - security-auditor
+    - visual-tester
+    - browser-automation
+    - devops-engineer
+    - release-manager
+    - requirement-refiner
+    - capability-analyst
+    - workflow-architect
+    - markdown-validator
+    - evaluator
+    - prompt-optimizer
+    - product-owner
+    - pipeline-judge
+    - planner
+    - reflector
+    - memory-manager
+    - agent-architect
+    - architect-indexer
+    fallback_models:
+    - ollama-cloud/glm-5.1
+    - ollama-cloud/deepseek-v4-pro-max
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: mixed
+  release-manager:
+    capabilities:
+    - git_operations
+    - version_management
+    - changelog_creation
+    - deployment
+    receives:
+    - approved_code
+    - release_request
+    produces:
+    - commits
+    - tags
+    - releases
+    forbidden:
+    - code_changes
+    - feature_development
+    model: ollama-cloud/glm-5.1
+    mode: subagent
+    delegates_to:
+    - evaluator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/glm-5
+    failover_strategy: downgraded
+  evaluator:
+    capabilities:
+    - performance_scoring
+    - process_analysis
+    - pattern_identification
+    - improvement_recommendations
+    receives:
+    - completed_issue
+    - agent_logs
+    produces:
+    - performance_report
+    - scores
+    - recommendations
+    forbidden:
+    - code_changes
+    model: ollama-cloud/glm-5.1
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - prompt-optimizer
+    - product-owner
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  prompt-optimizer:
+    capabilities:
+    - prompt_analysis
+    - prompt_improvement
+    - failure_pattern_detection
+    receives:
+    - low_scores
+    - failure_reports
+    produces:
+    - improved_prompts
+    - optimization_report
+    forbidden:
+    - agent_creation
+    model: ollama-cloud/glm-5.1
+    variant: instant
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - openrouter/qwen/qwen3.6-plus:free
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: downgraded
+  product-owner:
+    capabilities:
+    - issue_management
+    - prioritization
+    - backlog_management
+    - workflow_completion
+    receives:
+    - completed_work
+    - stakeholder_requests
+    produces:
+    - priority_order
+    - issue_labels
+    - issue closures
+    forbidden:
+    - implementation
+    model: ollama-cloud/glm-5.1
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - ollama-cloud/glm-5
+    - openrouter/qwen/qwen3.6-plus:free
+    - groq/llama-3.1-8b-instant
+    - ollama-cloud/deepseek-v4-pro-max
+    failover_strategy: mixed
+  pipeline-judge:
+    capabilities:
+    - test_execution
+    - fitness_scoring
+    - metric_collection
+    - bottleneck_detection
+    receives:
+    - completed_workflow
+    - pipeline_logs
+    produces:
+    - fitness_report
+    - bottleneck_analysis
+    - improvement_triggers
+    forbidden:
+    - code_writing
+    - code_changes
+    - prompt_changes
+    model: ollama-cloud/glm-5.1
+    mode: subagent
+    delegates_to:
+    - prompt-optimizer
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    - groq/llama-3.1-8b-instant
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: mixed
+  workflow-architect:
+    capabilities:
+    - workflow_design
+    - process_definition
+    - automation_setup
+    receives:
+    - workflow_requirements
+    produces:
+    - workflow_definitions
+    - command_files
+    forbidden:
+    - execution
+    model: ollama-cloud/glm-5.1
+    variant: thinking
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  markdown-validator:
+    capabilities:
+    - markdown_validation
+    - formatting_check
+    - link_validation
+    receives:
+    - markdown_files
+    produces:
+    - validation_report
+    - corrections
+    forbidden:
+    - content_creation
+    model: ollama-cloud/nemotron-3-nano:30b
+    mode: subagent
+    delegates_to:
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/glm-5.1
+    - groq/llama-3.1-8b-instant
+    - ollama/qwen3.5-122b
+    failover_strategy: speed-burst
+  agent-architect:
+    capabilities:
+    - agent_design
+    - prompt_engineering
+    - capability_definition
+    receives:
+    - agent_requirements
+    produces:
+    - agent_definition
+    - integration_plan
+    forbidden:
+    - agent_execution
+    model: ollama-cloud/glm-5.1
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - capability-analyst
+    - requirement-refiner
+    - system-analyst
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/deepseek-v4-pro-max
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  planner:
+    capabilities:
+    - task_decomposition
+    - chain_of_thought
+    - tree_of_thoughts
+    - plan_execute_reflect
+    - dependency_analysis
+    receives:
+    - complex_task
+    - objective
+    produces:
+    - decomposed_steps
+    - dependency_graph
+    - success_criteria
+    forbidden:
+    - implementation
+    - execution
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/glm-5.1
+    failover_strategy: downgraded
+  reflector:
+    capabilities:
+    - self_reflection
+    - mistake_analysis
+    - lesson_extraction
+    - trajectory_analysis
+    - heuristic_evaluation
+    receives:
+    - action_trajectory
+    - task_result
+    produces:
+    - reflection_report
+    - lessons_learned
+    - improved_approach
+    forbidden:
+    - implementation
+    - code_changes
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/glm-5.1
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  memory-manager:
+    capabilities:
+    - memory_retrieval
+    - memory_storage
+    - memory_consolidation
+    - relevance_scoring
+    - episodic_management
+    receives:
+    - query
+    - memory_type
+    produces:
+    - retrieved_memories
+    - relevance_scores
+    - consolidated_memories
+    forbidden:
+    - code_changes
+    - implementation
+    model: ollama-cloud/nemotron-3-super
+    mode: subagent
+    delegates_to: []
+    fallback_models:
+    - openrouter/qwen/qwen3.6-plus:free
+    - ollama-cloud/deepseek-v4-pro-max
+    - ollama-cloud/glm-5.1
+    failover_strategy: downgraded
+  architect-indexer:
+    capabilities:
+    - codebase_indexing
+    - project_mapping
+    - architecture_documentation
+    - dependency_analysis
+    - entity_extraction
+    - api_surface_discovery
+    - convention_detection
+    - staleness_detection
+    receives:
+    - project_root_directory
+    - stale_sections_list
+    produces:
+    - .architect/state.json
+    - .architect/project.json
+    - .architect/README.md
+    - architecture_overview
+    - dependency_graph
+    - entity_documentation
+    - db_schema_documentation
+    - api_surface_documentation
+    - convention_documentation
+    - file_graph
+    - module_graph
+    forbidden:
+    - code_changes
+    - implementation
+    model: ollama-cloud/glm-5.1
+    variant: thinking
+    mode: subagent
+    delegates_to:
+    - system-analyst
+    - orchestrator
+    fallback_models:
+    - ollama-cloud/kimi-k2.6:cloud
+    - ollama-cloud/glm-5
+    - openrouter/qwen/qwen3.6-plus:free
+    failover_strategy: downgraded
+  capability_routing:
+    code_writing: lead-developer
+    code_review: code-skeptic
+    test_writing: sdet-engineer
+    architecture: system-analyst
+    security: security-auditor
+    performance: performance-engineer
+    bug_fixing: the-fixer
+    git_operations: release-manager
+    ui_implementation: frontend-developer
+    nextjs_development: frontend-developer
+    vue_nuxt_development: frontend-developer
+    react_development: frontend-developer
+    e2e_testing: browser-automation
+    visual_testing: visual-tester
+    bbox_extraction: visual-tester
+    console_error_detection: visual-tester
+    gitea_integration: visual-tester
+    docker_networking: visual-tester
+    requirement_analysis: requirement-refiner
+    gap_analysis: capability-analyst
+    issue_management: product-owner
+    prompt_optimization: prompt-optimizer
+    workflow_design: workflow-architect
+    scoring: evaluator
+    duplicate_detection: history-miner
+    agent_design: agent-architect
+    markdown_validation: markdown-validator
+    postgresql_integration: backend-developer
+    sqlite_integration: backend-developer
+    clickhouse_integration: go-developer
+    flutter_development: flutter-developer
+    php_web_development: php-developer
+    laravel_development: php-developer
+    symfony_development: php-developer
+    wordpress_development: php-developer
+    python_web_development: python-developer
+    django_development: python-developer
+    fastapi_development: python-developer
+    docker_configuration: devops-engineer
+    kubernetes_setup: devops-engineer
+    ci_cd_pipeline: devops-engineer
+    task_decomposition: planner
+    self_reflection: reflector
+    memory_retrieval: memory-manager
+    chain_of_thought: planner
+    tree_of_thoughts: planner
+    fitness_scoring: pipeline-judge
+    test_execution: pipeline-judge
+    bottleneck_detection: pipeline-judge
+    go_api_development: go-developer
+    go_database_design: go-developer
+    go_concurrent_programming: go-developer
+    go_authentication: go-developer
+    go_microservices: go-developer
+    codebase_indexing: architect-indexer
+    project_mapping: architect-indexer
+    architecture_documentation: architect-indexer
+    dependency_analysis: architect-indexer
+    entity_extraction: architect-indexer
+    api_surface_discovery: architect-indexer
+    convention_detection: architect-indexer
+parallel_groups:
+  review_phase:
+  - security-auditor
+  - performance-engineer
+  - code-skeptic
+  testing_phase:
+  - sdet-engineer
+  - browser-automation
+  - visual-tester
+iteration_loops:
+  code_review:
+    evaluator: code-skeptic
+    optimizer: the-fixer
+    max_iterations: 3
+    convergence: all_issues_resolved
+  security_review:
+    evaluator: security-auditor
+    optimizer: the-fixer
+    max_iterations: 2
+    convergence: no_critical_vulnerabilities
+  performance_review:
+    evaluator: performance-engineer
+    optimizer: the-fixer
+    max_iterations: 2
+    convergence: all_perf_issues_resolved
+  evolution:
+    evaluator: pipeline-judge
+    optimizer: prompt-optimizer
+    max_iterations: 3
+    convergence: fitness_above_0.85
+quality_gates:
+  requirements:
+  - user_stories_defined
+  - acceptance_criteria_complete
+  - technical_constraints_documented
+  architecture:
+  - schema_valid
+  - endpoints_documented
+  - tech_stack_decided
+  implementation:
+  - build_success
+  - no_type_errors
+  - no_lint_errors
+  testing:
+  - coverage_gte_80
+  - all_tests_pass
+  - no_critical_bugs
+  review:
+  - no_critical_issues
+  - no_security_vulnerabilities
+  - performance_acceptable
+  docker:
+  - build_success
+  - health_check_pass
+  - size_under_limit
+  documentation:
+  - readme_complete
+  - api_docs_complete
+  - deployment_guide_complete
+workflow_states:
+  new:
+  - planned
+  planned:
+  - researching
+  researching:
+  - designed
+  designed:
+  - testing
+  testing:
+  - implementing
+  implementing:
+  - reviewing
+  reviewing:
+  - fixing
+  - perf_check
+  fixing:
+  - reviewing
+  perf_check:
+  - security_check
+  security_check:
+  - releasing
+  releasing:
+  - evaluated
+  evaluated:
+  - evolving
+  - completed
+  evolving:
+  - evaluated
+  completed: []
+evolution:
+  enabled: true
+  auto_trigger: true
+  fitness_threshold: 0.7
+  max_evolution_attempts: 3
+  fitness_history: .kilo/logs/fitness-history.jsonl
+  token_budget_default: 50000
+  time_budget_default: 300
+  budgets:
+    feature:
+      tokens: 50000
+      time_s: 300
+      min_coverage: 80
+    bugfix:
+      tokens: 20000
+      time_s: 120
+      min_coverage: 90
+    refactor:
+      tokens: 40000
+      time_s: 240
+      min_coverage: 95
+    security:
+      tokens: 30000
+      time_s: 180
+      min_coverage: 80
diff --git a/.kilo/commands/evolution.md b/.kilo/commands/evolution.md
index b66873e..a80baf5 100644
--- a/.kilo/commands/evolution.md
+++ b/.kilo/commands/evolution.md
@@ -24,6 +24,29 @@ Runs the automated evolution cycle on the most recent (or specified) workflow.
 
 ## Execution
 
+### Step 0: Model Research
+
+```
+Check if model benchmarks are stale (older than 7 days):
+  READ agent-evolution/data/model-benchmarks.json → metadata.generated
+  
+  IF metadata.generated > 7 days ago OR file missing:
+    Task(subagent_type: "capability-analyst")
+    → research latest model benchmarks, IF scores, availability
+    → output to agent-evolution/data/model-research-latest.json
+    → validates against agent-evolution/data/model-research.schema.json
+    
+  Read agent-evolution/data/model-benchmarks.json
+  → load heatmap scores per agent
+  → load recommendations
+  → identify agents where current model != best-fit model (score gap > 5)
+```
+
+This step ensures the evolution cycle works with fresh model data. If benchmarks are stale,
+the capability-analyst researches current model capabilities and pricing.
+
+The research output follows the schema: agent-evolution/data/model-research.schema.json
+
 ### Step 1: Judge (Fitness Evaluation)
 
 ```bash
@@ -65,7 +88,7 @@ ELSE:
   echo "📉 No improvement. Reverting."
 ```
 
-### Step 4: Log
+### Step 4: Log + Dashboard
 
 Append to `.kilo/logs/fitness-history.jsonl`:
 
@@ -82,6 +105,14 @@ Append to `.kilo/logs/fitness-history.jsonl`:
 }
 ```
 
+After logging, rebuild the research dashboard:
+
+```bash
+bun run agent-evolution/scripts/build-research-dashboard.ts
+```
+
+This ensures the dashboard reflects any model changes that occurred during evolution.
+
 ## Subcommands
 
 ### `log` — Log Model Change
@@ -153,6 +184,24 @@ Shows:
 - Model upgrade recommendations
 - Priority order
 
+### `research` — Research Model Updates
+
+```bash
+/evolution research            # research all models
+/evolution research --agent planner  # research models for specific agent
+/evolution research --provider ollama-cloud  # research specific provider
+```
+
+Steps:
+1. Read current agents from `.kilo/capability-index.yaml`
+2. Read existing benchmarks from `agent-evolution/data/model-benchmarks.json`
+3. Fetch latest model info from provider APIs/docs
+4. Score each model against each agent role (using IF-adjusted formula)
+5. Generate recommendations where score improvement > 5 points
+6. Output to `agent-evolution/data/model-research-latest.json`
+7. Validate against `agent-evolution/data/model-research.schema.json`
+8. If validation passes, update `agent-evolution/data/model-benchmarks.json`
+
 ## Data Storage
 
 ### fitness-history.jsonl
@@ -190,6 +239,28 @@ Shows:
 }
 ```
 
+### model-benchmarks.json
+
+Static benchmark data extracted from research. Contains:
+- Model capabilities (SWE-bench, IF scores, context windows)
+- Agent × Model compatibility heatmap scores
+- Groq/OpenRouter free tier availability
+- Current agent configuration snapshot
+- Recommendations (applied + pending)
+- Impact analysis data
+
+Path: `agent-evolution/data/model-benchmarks.json`
+Schema: `agent-evolution/data/model-benchmarks.schema.json`
+Refresh: When `/evolution research` runs or auto when stale (>7 days)
+
+### model-research-latest.json
+
+Latest research output from `/evolution research` or Step 0.
+Dynamic file — overwritten each research cycle.
+
+Path: `agent-evolution/data/model-research-latest.json`
+Schema: `agent-evolution/data/model-research.schema.json`
+
 ## Integration Points
 
 - **After `/pipeline`**: Evaluator scores logged
@@ -221,6 +292,10 @@ evolution:
 | Token Cost | pipeline logs | Resource efficiency |
 | Wall-Clock Time | pipeline logs | Speed |
 | Agent ROI | history analysis | Cost/benefit |
+| Model IF Score | model-benchmarks.json | Prompt adherence per model |
+| Model Fit Score | heatmap data | Agent-model compatibility |
+| Model Availability | provider APIs | Rate limits, free tier status |
+| Staleness | metadata.generated | How fresh is benchmark data |
 
 ## Example Session
 
@@ -243,6 +318,63 @@ $ /evolution
 ✅ Logged to .kilo/logs/fitness-history.jsonl
 ```
 
+## Example: Model Research Session
+
+```bash
+$ /evolution research
+
+## Model Research: All Agents
+
+**Benchmarks last updated**: 2026-04-20 (7 days ago — refreshing...)
+
+### Research Phase
+→ Fetching Ollama Cloud model list... 20 models found
+→ Fetching OpenRouter free tier... 3 models found
+→ Fetching Groq free tier... 5 models found
+→ Scoring 28 models × 36 agents... 1008 scores computed
+
+### Top Recommendations (score gap > 5)
+
+| Agent | Current | Score | Recommended | Score | Δ | Impact |
+|-------|---------|-------|-------------|-------|---|--------|
+| planner | nemotron-3-super | 80 | deepseek-v4-pro-max | 88 | +8 | high |
+| go-developer | qwen3-coder | 85 | deepseek-v4-pro-max | 88 | +3 | medium |
+| [built-in] debug | glm-5.1 | 88 | kimi-k2.6:cloud | 90 | +2 | high |
+
+### Output
+✅ agent-evolution/data/model-research-latest.json (28 models, 11 recommendations)
+✅ agent-evolution/data/model-benchmarks.json refreshed (36 agents)
+
+### Next Steps
+Run `/evolution` to apply recommendations and re-test
+Or `/evolution --dry-run` to preview changes
+
+### Dashboard Rebuild
+
+After model research or applying recommendations, rebuild the dashboard:
+
+```bash
+bun run agent-evolution/scripts/build-research-dashboard.ts
+```
+
+Output:
+- `agent-evolution/research-dashboard.html` — latest interactive dashboard
+- `agent-evolution/dist/research-dashboard-YYYY_MM_DD.html` — dated archive
+
+The dashboard reads from `agent-evolution/data/model-benchmarks.json` and renders:
+- Current agent-model configuration table
+- Model comparison cards with SWE-bench and IF scores
+- Agent × Model heatmap with IF adjustment
+- Selectable recommendations with JSON export
+- Before/after impact analysis
+
+Watch mode for continuous rebuild during research:
+```bash
+bun run agent-evolution/scripts/build-research-dashboard.ts --watch
+```
+Auto-triggers with `--watch` when `model-benchmarks.json` or template changes.
+```
+
 ---
 
 *Evolution workflow v2.0 - Objective fitness scoring with pipeline-judge*
\ No newline at end of file
diff --git a/.kilo/commands/research.md b/.kilo/commands/research.md
index 2c92158..2fc3ece 100644
--- a/.kilo/commands/research.md
+++ b/.kilo/commands/research.md
@@ -22,6 +22,9 @@ Runs continuous research and self-improvement cycle based on the latest findings
 
 ```
 /research [topic] [--auto]
+/research models           # research latest AI models for agent optimization
+/research models --agent planner  # research models for specific agent role
+/research models --provider ollama-cloud  # filter by provider
 ```
 
 ## Parameters
@@ -35,6 +38,28 @@ Runs continuous research and self-improvement cycle based on the latest findings
 
 Check `.kilo/logs/efficiency_score.json` for low-performing agents.
 
+### Step 1.5: Model Research (when topic is "models" or agent scores are low)
+
+```
+IF topic === "models" OR any agent score < 7:
+  1. Read agent-evolution/data/model-benchmarks.json
+     → Check metadata.generated staleness
+  2. Fetch latest model data from providers:
+     - Ollama Cloud: https://ollama.com/models (via webfetch)
+     - OpenRouter: https://openrouter.ai/models (via webfetch)  
+     - Groq: https://console.groq.com/docs/models (via webfetch)
+  3. For each model, compute:
+     - IF score (from IFEval/IFBench benchmarks)
+     - Role fitness (SWE-bench for coding, GPQA for reasoning, etc.)
+     - Context window and cost
+  4. Build heatmap: score each model against each agent
+     Formula: role_fitness * (0.7 + 0.3 * IF/100)
+  5. Generate recommendations for agents where best-scored model ≠ current
+  6. Output to agent-evolution/data/model-research-latest.json
+  7. Validate against agent-evolution/data/model-research.schema.json
+  8. Update model-benchmarks.json with fresh data
+```
+
 ### Step 2: Gap Identification
 
 Analyze capability-index.yaml for missing capabilities.
@@ -46,6 +71,15 @@ Fetch latest research from:
 - OpenAI: https://platform.openai.com/docs/guides/agents
 - Lilian Weng: https://lilianweng.github.io
 
+### Model Research Sources
+- Ollama Model Library (https://ollama.com/models)
+- OpenRouter Models (https://openrouter.ai/models)
+- Groq Console (https://console.groq.com/docs/models)
+- SWE-Bench Leaderboard (https://www.swebench.com)
+- Terminal-Bench (https://marc0.dev/terminal-bench)
+- LMSYS Chatbot Arena (https://chat.lmsys.org)
+- Artificial Analysis (https://artificialanalysis.ai)
+
 ### Step 4: Implementation
 
 Create new agents, skills, or rules based on findings.
@@ -81,3 +115,53 @@ Post findings to Gitea Issue #25 (Research Milestone).
 - Issue: #25
 - Commit: abc1234
 ```
+
+### Model Research Example
+
+```
+/research models
+
+# Output:
+## Research: model optimization
+
+### Models Analyzed
+- Ollama Cloud: 20 models
+- OpenRouter Free: 3 models
+- Groq Free: 5 models
+
+### Key Findings
+- DeepSeek V4-Pro Max now available (SWE-V 80.6, IF:88)
+- Kimi K2.6 IF score confirmed: 91 (best for orchestration)
+- Nemotron 3 Super IF:78 — weak for prompt-heavy roles
+- Qwen 3.6 Plus FREE remains best IF/cost ratio (91, $0)
+
+### Recommendations Generated
+- 11 model swap recommendations
+- 4 high impact, 3 medium, 4 low
+- Average expected improvement: +12 points
+
+### Files Updated
+- agent-evolution/data/model-research-latest.json
+- agent-evolution/data/model-benchmarks.json (refreshed)
+
+### Evolution Tracked
+- Issue: #25
+- Next: /evolution to apply recommendations
+```
+
+## Model Research Output Format
+
+All model research output follows the schema:
+`agent-evolution/data/model-research.schema.json`
+
+Key fields:
+- `models[]` — model capabilities, benchmarks, IF scores
+- `recommendations[]` — agent-specific model swap suggestions
+- `heatmap` — agent × model compatibility matrix
+- `capability_index_patch[]` — ready-to-apply YAML patches
+- `summary` — aggregate improvement metrics
+
+This format is consumed by:
+- `/evolution` command for auto-apply
+- `agent-evolution/scripts/sync-model-research.ts` for propagation
+- Evolution dashboard for visualization
diff --git a/.kilo/rules/evolutionary-sync.md b/.kilo/rules/evolutionary-sync.md
index 342e1ce..ac36e0a 100644
--- a/.kilo/rules/evolutionary-sync.md
+++ b/.kilo/rules/evolutionary-sync.md
@@ -24,6 +24,9 @@ When agents change, update ALL of these files:
 | `.kilo/KILO_SPEC.md` | Pipeline Agents table, Workflow Commands table |
 | `AGENTS.md` | Pipeline Agents tables by category |
 | `.kilo/agents/orchestrator.md` | Task Tool Invocation table |
+| `agent-evolution/data/model-benchmarks.json` | Model fitness scores, heatmap, recommendations |
+| `agent-evolution/data/model-research-latest.json` | Latest research output (overwritten each cycle) |
+| `agent-evolution/data/agent-versions.json` | Agent model version history |
 
 ## Sync Process (REQUIRED ORDER)
 
@@ -53,6 +56,13 @@ After running `--fix`, you MUST verify:
 □ `.kilo/capability-index.yaml` — model fields updated
 □ No old models leaked (grep for previous model IDs)
 □ `ollama-cloud/kimi-k2.6` → always `:cloud` suffix
+□ model-benchmarks.json — metadata.generated updated
+□ model-research-latest.json — validates against schema
+□ agent-versions.json — history entries added for all model changes
+□ sync-model-research.ts — dry-run matches expected changes
+□ Groq rate limits current (check console.groq.com/docs/models)
+□ OpenRouter free tier models current (check openrouter.ai/models)
+□ No regressions in IF scores (IF should not decrease from previous)
 ```
 
 ## Findings from Evolution Round 2026-04-27
@@ -140,6 +150,46 @@ for a in meta['agents']:
    node scripts/sync-agents.js --fix
    ```
 
+## Model Research Sync
+
+When `/evolution research` or `/research models` produces new benchmark data:
+
+### Sync Process
+
+```
+1. /research models OR /evolution Step 0
+   → Produces: agent-evolution/data/model-research-latest.json
+   
+2. Validate against schema:
+   node -e "const Ajv=require('ajv'); const ajv=new Ajv(); const schema=JSON.parse(require('fs').readFileSync('agent-evolution/data/model-research.schema.json','utf8')); const data=JSON.parse(require('fs').readFileSync('agent-evolution/data/model-research-latest.json','utf8')); const valid=ajv.validate(schema,data); console.log(valid?'VALID':'INVALID'); if(!valid) console.log(JSON.stringify(ajv.errors,null,2))"
+   
+3. Apply recommendations:
+   bun run agent-evolution/scripts/sync-model-research.ts
+   
+4. Or dry-run first:
+   bun run agent-evolution/scripts/sync-model-research.ts --dry-run
+
+5. After applying, the script automatically:
+   - Updates capability-index.yaml
+   - Updates agent-versions.json
+   - Updates kilo-meta.json
+   - Updates kilo.jsonc (with regex — manual verify still needed)
+   - Runs sync-agents.js --fix
+   - Runs sync-agents.js --check
+```
+
+### Data Freshness Check
+
+```bash
+# Check if benchmarks are stale (>7 days)
+node -e "
+const data = JSON.parse(require('fs').readFileSync('agent-evolution/data/model-benchmarks.json','utf8'));
+const gen = new Date(data.metadata.generated);
+const daysOld = (Date.now() - gen.getTime()) / (1000*60*60*24);
+console.log(daysOld > 7 ? 'STALE' : 'FRESH', '(' + Math.round(daysOld) + ' days old)');
+"
+```
+
 ## Model Changes
 
 When changing a model:
diff --git a/.kilo/shared/self-evolution.md b/.kilo/shared/self-evolution.md
index a1c5ed0..c58e80a 100644
--- a/.kilo/shared/self-evolution.md
+++ b/.kilo/shared/self-evolution.md
@@ -8,6 +8,8 @@ When task requirements exceed existing agent capabilities.
 2. Required domain knowledge not in any skill
 3. Complex multi-step task needs new workflow pattern
 4. `@capability-analyst` reports critical gap
+5. `/evolution` reports fitness < 0.70 and model research finds better model
+6. Model benchmarks stale (>7 days) and research discovers new model
 
 ## Evolution Flow
 
@@ -41,6 +43,72 @@ When task requirements exceed existing agent capabilities.
 [New Capability Available]
 ```
 
+## Model Evolution Flow
+
+When an agent's current model is suboptimal (score gap > 5 points in heatmap):
+
+```
+[Evolution Fitness < 0.85]
+       ↓
+1. Read model-benchmarks.json → load heatmap, recommendations
+       ↓
+2. IF stale (>7 days) → @capability-analyst researches models
+   → Output: agent-evolution/data/model-research-latest.json
+   → Validates against: agent-evolution/data/model-research.schema.json
+       ↓
+3. Identify agents where best_model ≠ current_model (gap > 5)
+       ↓
+4. Generate recommendations (action: update_model)
+       ↓
+5. Dry-run → /evolution --dry-run → Show what would change
+       ↓
+6. Apply → bun run agent-evolution/scripts/sync-model-research.ts
+   → Updates: capability-index.yaml, agent-versions.json, kilo-meta.json, kilo.jsonc
+   → Triggers: sync-agents.js --fix → propagates to .md files
+   → Validates: sync-agents.js --check
+       ↓
+7. Re-test → @pipeline-judge → new fitness score
+       ↓
+8. IF fitness improved → commit changes
+   IF fitness regressed → revert via agent-versions.json history
+       ↓
+9. Log to Gitea + fitness-history.jsonl
+       ↓
+[Models Optimized]
+```
+
+## Model Research Data Flow
+
+```
+[model-benchmarks.json]          ← Static benchmark data (refreshed weekly)
+       ↓ read
+[/evolution Step 0]              ← Checks staleness, triggers research if needed
+[/research models]               ← Explicit research trigger
+       ↓ produces
+[model-research-latest.json]     ← Dynamic research output
+       ↓ consumed by
+[sync-model-research.ts]         ← Applies recommendations
+       ↓ updates
+[capability-index.yaml]          ← Model assignments
+[agent-versions.json]            ← History tracking
+[kilo-meta.json]                 ← Source of truth
+[kilo.jsonc]                     ← Agent config (manual verify)
+[.kilo/agents/*.md]              ← Frontmatter (via sync script)
+       ↓ verified by
+[sync-agents.js --check]         ← Consistency validation
+```
+
+### Key Files
+
+| File | Purpose | Updated By |
+|------|---------|------------|
+| `agent-evolution/data/model-benchmarks.json` | Static benchmark data | `/research models`, `/evolution research` |
+| `agent-evolution/data/model-research-latest.json` | Latest research output | `/research models`, `/evolution Step 0` |
+| `agent-evolution/data/model-research.schema.json` | Validation schema | Manual (schema changes are rare) |
+| `agent-evolution/data/model-benchmarks.schema.json` | Benchmarks data schema | Manual |
+| `agent-evolution/data/agent-versions.json` | Version history | `sync-model-research.ts` |
+| `agent-evolution/scripts/sync-model-research.ts` | Application script | Manual execution |
+
 ## Self-Modification Rules
 
 1. ONLY modify own permission whitelist
@@ -49,6 +117,10 @@ When task requirements exceed existing agent capabilities.
 4. ALWAYS verify access after changes
 5. ALWAYS log results to `.kilo/EVOLUTION_LOG.md`
 6. NEVER skip verification step
+7. ALWAYS validate research output against schema before applying
+8. NEVER apply model changes without dry-run preview first
+9. ALWAYS run sync-agents.js --check after model changes
+10. ALWAYS revert if fitness regresses after model change
 
 ## Evolution Triggers
 
@@ -65,6 +137,11 @@ When task requirements exceed existing agent capabilities.
 4. Update `.kilo/KILO_SPEC.md` (document)
 5. Update `AGENTS.md` (reference)
 6. Append to `.kilo/EVOLUTION_LOG.md` (log entry)
+7. Update `agent-evolution/data/model-benchmarks.json` (if model data changed)
+8. Update `agent-evolution/data/agent-versions.json` (add history entry)
+9. Update `kilo-meta.json` (source of truth for sync)
+10. Run `node scripts/sync-agents.js --fix` (propagate to all files)
+11. Run `node scripts/sync-agents.js --check` (verify consistency)
 
 ## Verification Checklist
 
@@ -77,3 +154,12 @@ After each evolution:
 - [ ] AGENTS.md updated with new agent
 - [ ] EVOLUTION_LOG.md updated with entry
 - [ ] Gitea milestone closed with results
+- [ ] model-research-latest.json validates against schema
+- [ ] sync-model-research.ts dry-run shows correct changes
+- [ ] capability-index.yaml model field updated for affected agents
+- [ ] agent-versions.json history entry added with rationale
+- [ ] kilo-meta.json matches new model assignments
+- [ ] kilo.jsonc manually verified (sync script does not guarantee this)
+- [ ] sync-agents.js --check passes
+- [ ] No stale models leaked (grep for previous model IDs)
+- [ ] Cloud model suffix correct (kimi-k2.6:cloud, not kimi-k2.6)
diff --git a/agent-evolution/README.md b/agent-evolution/README.md
index 08d7e10..d8b94e4 100644
--- a/agent-evolution/README.md
+++ b/agent-evolution/README.md
@@ -117,6 +117,9 @@ bun run evolution:run     # Запустить контейнер
 bun run evolution:stop    # Остановить
 bun run evolution:dev      # Docker Compose
 bun run evolution:logs     # Логи
+bun run research:dashboard    # Build research dashboard
+bun run research:watch        # Watch mode for dashboard
+bun run research:sync         # Sync model research to agents
 ```
 
 ## Структура
@@ -132,6 +135,50 @@ agent-evolution/
 └── README.md                    # Этот файл
 ```
 
+## Research Dashboard (Model Benchmarks)
+
+### Generate from live data
+
+```bash
+# Build research dashboard from model-benchmarks.json
+bun run agent-evolution/scripts/build-research-dashboard.ts
+
+# Watch mode — auto-rebuild on data changes
+bun run agent-evolution/scripts/build-research-dashboard.ts --watch
+
+# Open in browser
+start agent-evolution/research-dashboard.html
+```
+
+### Output files
+
+| File | Description |
+|------|-------------|
+| `research-dashboard.html` | Latest interactive dashboard (all 6 tabs) |
+| `dist/research-dashboard-YYYY_MM_DD.html` | Dated archive |
+| `research-dashboard.template.html` | Template for generation |
+
+### Dashboard tabs
+
+1. **Обзор** — stat cards, current config table, agent count, model count
+2. **Groq** — free tier models with RPM/RPD/TPM/TPD limits, speed indicators
+3. **Модели** — filterable cards with SWE-bench, IF scores, context windows, tags
+4. **Матрица** — Agent×Model heatmap with IF adjustment, tooltips, color coding
+5. **Рекомендации** — selectable cards with JSON export, impact analysis
+6. **Анализ профита** — before/after comparison, canvas charts, closed-source comparison
+
+### Source data
+
+The dashboard reads from `agent-evolution/data/model-benchmarks.json`:
+- 15 models with benchmarks (SWE-bench, IF scores)
+- 36 agent configurations
+- 33 agent×model score matrices
+- 11 recommendations
+- 5 Groq models with rate limits
+- Closed-source comparison data
+
+Refresh: run `/research models` or `/evolution research` to update
+
 ## Быстрый старт
 
 ```bash
@@ -231,6 +278,22 @@ git log --all --oneline -- ".kilo/agents/"
 **Files**: src/auth.ts, src/user.ts
 ```
 
+### 6. Model Benchmarks (agent-evolution/data/model-benchmarks.json)
+
+Research data extracted from `apaw_agent_model_research_v3.html`:
+- Static benchmark scores (SWE-bench, IF scores, context windows)
+- Heatmap compatibility matrix
+- Provider rate limits
+- Recommendation history
+
+### 7. Model Research Output (agent-evolution/data/model-research-latest.json)
+
+Dynamic research results:
+- Fresh model data from provider APIs
+- IF-adjusted agent×model scores
+- Pending recommendations with impact levels
+- Ready-to-apply YAML patches
+
 ## JSON Schema
 
 Формат `agent-versions.json`:
@@ -271,6 +334,76 @@ git log --all --oneline -- ".kilo/agents/"
 }
 ```
 
+## Model Research Data
+
+### model-benchmarks.json
+
+Comprehensive benchmark data from the HTML research file:
+
+```json
+{
+  "version": "1.0.0",
+  "generated": "2026-04-27T17:44:44Z",
+  "total_agents": 36,
+  "total_models_tracked": 11,
+  "models": [
+    {
+      "id": "ollama-cloud/qwen3-coder:480b",
+      "name": "Qwen3-Coder 480B",
+      "organization": "Qwen",
+      "swe_bench": 66.5,
+      "if_score": 88,
+      "context_window": "256K→1M",
+      "categories": ["coding", "agent"],
+      "provider": "ollama"
+    }
+  ],
+  "agent_current_config": [
+    { "agent": "lead-developer", "model": "ollama-cloud/qwen3-coder:480b", "fit_score": 92, "status": "optimal" }
+  ],
+  "recommendations": [
+    {
+      "agent": "planner",
+      "current_model": "nemotron-3-super",
+      "recommended_model": "deepseek-v4-pro-max",
+      "impact": "high",
+      "expected_improvement": { "quality": "+10%", "speed": "~1x", "context_window": "1M" }
+    }
+  ]
+}
+```
+
+### model-research-latest.json
+
+Latest research output (overwritten each cycle):
+- Generated by `/research models` or `/evolution Step 0`
+- Validated against `model-research.schema.json`
+- Consumed by `sync-model-research.ts`
+
+### sync-model-research.ts
+
+Applies model recommendations to configuration:
+
+```bash
+# Dry-run first
+bun run agent-evolution/scripts/sync-model-research.ts --dry-run
+
+# Apply all pending recommendations
+bun run agent-evolution/scripts/sync-model-research.ts
+
+# Apply for single agent
+bun run agent-evolution/scripts/sync-model-research.ts --agent planner
+```
+
+Updates:
+1. `.kilo/capability-index.yaml` — model assignments
+2. `kilo-meta.json` — source of truth
+3. `kilo.jsonc` — agent config
+4. `agent-evolution/data/agent-versions.json` — history tracking
+5. `.kilo/agents/*.md` frontmatter (via sync-agents.js --fix)
+
+After applying, rebuilds dashboard automatically.
+
 ## Интеграция
 
 ### В Pipeline
@@ -406,4 +539,50 @@ cp agent-evolution/data/backup/agent-versions-20260405.json agent-evolution/data
 4. **Integration**:
    - Slack/Telegram уведомления
    - Автоматическое применение рекомендаций
-   - A/B testing моделей
\ No newline at end of file
+   - A/B testing моделей
+
+## Bidirectional Data Flow
+
+```
+[/research models] OR [/evolution Step 0]
+       ↓
+[agent-evolution/data/model-research-latest.json]
+       ↓
+[bun run sync-model-research.ts]
+       ↓
+[.kilo/capability-index.yaml] → updated model assignments
+[kilo-meta.json]              → updated source of truth
+[kilo.jsonc]                  → updated config
+[agent-versions.json]         → history entries
+[.kilo/agents/*.md]           → frontmatter updated
+       ↓
+[sync-agents.js --fix]        → propagate to all files
+       ↓
+[bun run build-research-dashboard.ts]
+       ↓
+[research-dashboard.html]     → live dashboard
+[dist/dashboard-YYYY_MM_DD.html] → dated archive
+       ↓
+[/research models]            ← loop continues
+```
+
+### Data staleness check
+
+```bash
+# Check if benchmarks need refresh
+node -e "
+const d = require('./agent-evolution/data/model-benchmarks.json');
+const days = (Date.now() - new Date(d.generated)) / (1000*60*60*24);
+console.log(days > 7 ? 'STALE: needs refresh' : 'FRESH', Math.round(days), 'days old');
+"
+```
+
+### Auto-refresh pipeline
+
+```yaml
+# In capability-index.yaml
+evolution:
+  auto_trigger: true
+  max_evolution_attempts: 3
+  dashboard_rebuild: true  # new: auto-rebuild on model changes
+```
\ No newline at end of file
diff --git a/agent-evolution/data/agent-versions.json b/agent-evolution/data/agent-versions.json
index 6ced42e..a3e7889 100644
--- a/agent-evolution/data/agent-versions.json
+++ b/agent-evolution/data/agent-versions.json
@@ -1,12 +1,12 @@
 {
   "version": "1.0.0",
-  "lastUpdated": "2026-04-23T06:24:32.543Z",
+  "lastUpdated": "2026-04-27T20:28:58.592Z",
   "agents": {
     "lead-developer": {
       "current": {
         "description": "Primary code writer for backend and core logic. Writes implementation to pass tests",
         "mode": "subagent",
-        "model": "ollama-cloud/qwen3-coder:480b",
+        "model": "ollama-cloud/nemotron-3-super",
         "provider": "Ollama",
         "variant": "thinking",
         "color": "\"#DC2626\"",
@@ -27,6 +27,24 @@
           "to": "ollama-cloud/qwen3-coder:480b",
           "reason": "Initial configuration from capability-index.yaml",
           "source": "git"
+        },
+        {
+          "date": "2026-04-27T16:56:09.013Z",
+          "commit": "model-research-sync",
+          "type": "model_change",
+          "from": "ollama-cloud/qwen3-coder:480b",
+          "to": "ollama-cloud/nemotron-3-super",
+          "reason": "Nemotron 3 Super has better reasoning for core development tasks and RULER@1M context window. SWE-bench 68% vs Qwen's 66.5%.",
+          "source": "research"
+        },
+        {
+          "date": "2026-04-27T20:28:58.592Z",
+          "commit": "model-research-sync",
+          "type": "model_change",
+          "from": "ollama-cloud/qwen3-coder:480b",
+          "to": "ollama-cloud/nemotron-3-super",
+          "reason": "Nemotron 3 Super has better reasoning for core development tasks and RULER@1M context window. SWE-bench 68% vs Qwen's 66.5%.",
+          "source": "research"
         }
       ],
       "performance_log": []
@@ -255,7 +273,7 @@
       "current": {
         "description": "Designs technical specifications, data schemas, and API contracts before implementation",
         "mode": "subagent",
-        "model": "ollama-cloud/glm-5.1",
+        "model": "ollama-cloud/nemotron-3-super",
         "provider": "Ollama",
         "variant": "thinking",
         "color": "\"#0891B2\"",
@@ -285,6 +303,15 @@
           "to": "ollama-cloud/glm-5.1",
           "reason": "Model update from sync",
           "source": "git"
+        },
+        {
+          "date": "2026-04-27T16:59:52.825Z",
+          "commit": "model-research-sync",
+          "type": "model_change",
+          "from": "ollama-cloud/glm-5.1",
+          "to": "ollama-cloud/nemotron-3-super",
+          "reason": "Test recommendation for model research sync script",
+          "source": "research"
         }
       ],
       "performance_log": []
diff --git a/agent-evolution/data/model-benchmarks.json b/agent-evolution/data/model-benchmarks.json
new file mode 100644
index 0000000..ec848fb
--- /dev/null
+++ b/agent-evolution/data/model-benchmarks.json
@@ -0,0 +1,1774 @@
+{
+  "version": "1.0.0",
+  "generated": "2026-04-29T19:56:51.418Z",
+  "source": ".kilo/capability-index.yaml (synced v2)",
+  "total_agents": 32,
+  "total_models_tracked": 11,
+  "providers": [
+    "ollama",
+    "ollama-cloud",
+    "openrouter",
+    "groq"
+  ],
+  "models": [
+    {
+      "id": "qwen3-coder-480b",
+      "name": "Qwen3-Coder 480B",
+      "organization": "Qwen",
+      "parameters": "480B/35B active",
+      "context_window": "256K→1M",
+      "swe_bench": 66.5,
+      "if_score": 88,
+      "categories": [
+        "coding",
+        "agent"
+      ],
+      "description": "SOTA open-source кодинг. Сравним с Claude Sonnet 4.",
+      "tags": [
+        "coding",
+        "agent",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "minimax-m2.5",
+      "name": "MiniMax M2.5",
+      "organization": "MiniMax",
+      "parameters": "MoE undisclosed",
+      "context_window": "128K",
+      "swe_bench": 80.2,
+      "if_score": 82,
+      "categories": [
+        "coding",
+        "agent"
+      ],
+      "description": "Лидер SWE-bench 80.2%. Полный lifecycle разработки.",
+      "tags": [
+        "coding",
+        "agent"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "minimax-m2.7",
+      "name": "MiniMax M2.7",
+      "organization": "MiniMax",
+      "parameters": "~10B active",
+      "context_window": "128K",
+      "swe_bench": 78,
+      "if_score": 80,
+      "categories": [
+        "coding",
+        "agent",
+        "efficient"
+      ],
+      "description": "Самообучаемая. 56.2% SWE-Pro. 100 TPS. $0.30/M.",
+      "tags": [
+        "coding",
+        "agent",
+        "self-evolving"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "deepseek-v4-pro-max",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "1.6T/49B active MoE",
+      "context_window": "1M",
+      "swe_bench": 80.6,
+      "if_score": 89,
+      "categories": [
+        "coding",
+        "agent",
+        "reasoning"
+      ],
+      "description": "SWE-V 80.6, LiveCodeBench 93.5(#1!), Terminal-Bench 67.9, Codeforces 3206, 1M ctx, 27% FLOPs vs V3.2. MIT.",
+      "tags": [
+        "coding",
+        "agent",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "deepseek-v4-flash",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "284B/13B active MoE",
+      "context_window": "1M",
+      "swe_bench": 79,
+      "if_score": 86,
+      "categories": [
+        "coding",
+        "efficient",
+        "agent"
+      ],
+      "description": "SWE-V ~79%, Flash Max = Pro уровень reasoning. 13B active = ультрабыстрый. 1M ctx. FP4+FP8. MIT.",
+      "tags": [
+        "coding",
+        "efficient",
+        "agent",
+        "thinking"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "kimi-k2-6",
+      "name": "Kimi K2.6",
+      "organization": "Moonshot AI",
+      "parameters": "1T/32B active MoE",
+      "context_window": "256K",
+      "swe_bench": 80.2,
+      "if_score": 91,
+      "categories": [
+        "coding",
+        "agent",
+        "multimodal"
+      ],
+      "description": "SWE-Pro 58.6(#1!), SWE-V 80.2, Terminal-Bench 66.7, HLE 54.0(#1!), BrowseComp 83.2. 13h autonomous. 300 sub-agent swarm. Modified MIT.",
+      "tags": [
+        "coding",
+        "agent",
+        "swarm",
+        "vision",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "nemotron-3-super",
+      "name": "Nemotron 3 Super",
+      "organization": "NVIDIA",
+      "parameters": "120B/12B active",
+      "context_window": "1M",
+      "swe_bench": 60.5,
+      "if_score": 78,
+      "categories": [
+        "agent",
+        "reasoning",
+        "efficient"
+      ],
+      "description": "SWE-bench 60.5%. RULER@1M 91.75%! Но IF ниже — Mamba-layers иногда «теряют» инструкции в длинных промптах.",
+      "tags": [
+        "agent",
+        "1M-ctx",
+        "thinking"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "glm-5.1",
+      "name": "GLM-5",
+      "organization": "Z.ai",
+      "parameters": "744B/40B active",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 90,
+      "categories": [
+        "reasoning",
+        "agent"
+      ],
+      "description": "Мощный reasoning. Arena ELO 1451. Отличный instruction following (IFEval ~90+).",
+      "tags": [
+        "reasoning",
+        "agent"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "deepseek-v4",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "Large MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 75,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "Хороший reasoning, но IF нестабилен — иногда игнорирует формат вывода.",
+      "tags": [
+        "reasoning"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-5-122b",
+      "name": "Qwen 3.5 122B",
+      "organization": "Qwen",
+      "parameters": "122B/10B active",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 92,
+      "categories": [
+        "reasoning",
+        "efficient"
+      ],
+      "description": "IFEval 92.6%! Лучший IF среди open-source. Multimodal. Thinking.",
+      "tags": [
+        "vision",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-coder-next",
+      "name": "Qwen3-Coder-Next",
+      "organization": "Qwen",
+      "parameters": "80B/3B active",
+      "context_window": "128K",
+      "swe_bench": 70,
+      "if_score": 84,
+      "categories": [
+        "coding",
+        "efficient"
+      ],
+      "description": "70% SWE-bench с 3B active! Хороший IF для кодинга.",
+      "tags": [
+        "coding",
+        "efficient",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "cogito-2-1-671b",
+      "name": "Cogito 2.1 671B",
+      "organization": "Cognitive",
+      "parameters": "671B MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 76,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "MIT лицензия. 671B total. IF неплохой, но уступает GLM/Qwen.",
+      "tags": [
+        "reasoning"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-6-plus",
+      "name": "Qwen 3.6 Plus",
+      "organization": "Qwen",
+      "parameters": "Hybrid MoE",
+      "context_window": "1M",
+      "swe_bench": 78.8,
+      "if_score": 91,
+      "categories": [
+        "coding",
+        "agent",
+        "reasoning"
+      ],
+      "description": "FREE на OpenRouter! 1M контекст. Always-on CoT. Превосходный IF — наследник Qwen 3.5 (92.6%).",
+      "tags": [
+        "coding",
+        "agent",
+        "1M-ctx",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    },
+    {
+      "id": "step-3-5-flash",
+      "name": "Step 3.5 Flash",
+      "organization": "StepFun",
+      "parameters": "MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 79,
+      "categories": [
+        "efficient"
+      ],
+      "description": "Бесплатна на OpenRouter. IF средний.",
+      "tags": [
+        "efficient",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    },
+    {
+      "id": "deepseek-r1",
+      "name": "DeepSeek R1",
+      "organization": "DeepSeek",
+      "parameters": "671B MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 73,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "Мощные reasoning-цепочки. Но IF слабый — часто генерирует лишний reasoning вместо ответа.",
+      "tags": [
+        "reasoning",
+        "thinking",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    }
+  ],
+  "groq_models": [
+    {
+      "id": "openai/gpt-oss-20b",
+      "rpm": 30,
+      "rpd": "1K",
+      "tpm": "8K",
+      "tpd": "200K",
+      "speed": "1200+",
+      "use_case": "Ультра-быстрый fallback для лёгких ролей (markdown-validator)."
+    },
+    {
+      "id": "llama-3.1-8b-instant",
+      "rpm": 30,
+      "rpd": "14.4K",
+      "tpm": "6K",
+      "tpd": "500K",
+      "speed": "~800",
+      "use_case": "14.4K RPD! Самый высокий лимит. Для health-check / ping ролей."
+    },
+    {
+      "id": "groq/compound",
+      "rpm": 30,
+      "rpd": "250",
+      "tpm": "70K",
+      "tpd": "—",
+      "speed": "varies",
+      "use_case": "Мультимодельная агрегация. Для research-задач."
+    },
+    {
+      "id": "groq/compound-mini",
+      "rpm": 30,
+      "rpd": "250",
+      "tpm": "70K",
+      "tpd": "—",
+      "speed": "varies",
+      "use_case": "Лёгкая версия compound."
+    },
+    {
+      "id": "llama-prompt-guard-2",
+      "rpm": 30,
+      "rpd": "14.4K",
+      "tpm": "15K",
+      "tpd": "500K",
+      "speed": "~1K",
+      "use_case": "Security: входной фильтр для security-auditor (14.4K RPD!)."
+    }
+  ],
+  "agent_model_scores": [
+    {
+      "agent": "lead-developer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 92,
+        "minimax-m2.5": 86,
+        "minimax-m2.7": 82,
+        "nemotron-3-super": 70,
+        "glm-5.1": 68,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 66,
+        "qwen3-coder-next": 80,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "frontend-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 86,
+        "minimax-m2.5": 92,
+        "minimax-m2.7": 88,
+        "nemotron-3-super": 62,
+        "glm-5.1": 56,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 60,
+        "qwen3-coder-next": 76,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "php-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 87,
+        "minimax-m2.5": 76,
+        "minimax-m2.7": 72,
+        "nemotron-3-super": 64,
+        "glm-5.1": 56,
+        "deepseek-v4-pro-max": 74,
+        "qwen3-5-122b": 60,
+        "qwen3-coder-next": 76,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "python-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 90,
+        "minimax-m2.5": 82,
+        "minimax-m2.7": 78,
+        "nemotron-3-super": 66,
+        "glm-5.1": 60,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 88
+      }
+    },
+    {
+      "agent": "backend-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 91,
+        "minimax-m2.5": 84,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 68,
+        "glm-5.1": 63,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 62,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 87,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "go-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 85,
+        "minimax-m2.5": 78,
+        "minimax-m2.7": 74,
+        "nemotron-3-super": 66,
+        "glm-5.1": 58,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 58,
+        "qwen3-coder-next": 74,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "flutter-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 86,
+        "minimax-m2.5": 70,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 60,
+        "glm-5.1": 53,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 58,
+        "qwen3-coder-next": 74,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "devops-engineer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 66,
+        "minimax-m2.5": 53,
+        "minimax-m2.7": 48,
+        "nemotron-3-super": 78,
+        "glm-5.1": 75,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 54,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 88
+      }
+    },
+    {
+      "agent": "sdet-engineer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 88,
+        "minimax-m2.5": 84,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 70,
+        "glm-5.1": 63,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 87
+      }
+    },
+    {
+      "agent": "code-skeptic",
+      "current_model_index": 1,
+      "current_model_id": "minimax-m2.5",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 82,
+        "minimax-m2.5": 85,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 73,
+        "glm-5.1": 72,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 72,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "security-auditor",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 76,
+        "minimax-m2.5": 74,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 76,
+        "glm-5.1": 68,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 64,
+        "qwen3-6-plus": 75,
+        "kimi-k2-6": 80
+      }
+    },
+    {
+      "agent": "performance-engineer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 78,
+        "minimax-m2.5": 75,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 74,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 67,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "the-fixer",
+      "current_model_index": 1,
+      "current_model_id": "minimax-m2.5",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 89,
+        "minimax-m2.5": 88,
+        "minimax-m2.7": 84,
+        "nemotron-3-super": 71,
+        "glm-5.1": 64,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 82,
+        "qwen3-6-plus": 86,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "browser-automation",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 87,
+        "minimax-m2.5": 72,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 61,
+        "glm-5.1": 53,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 56,
+        "qwen3-coder-next": 72,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "visual-tester",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 82,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 55,
+        "glm-5.1": 48,
+        "deepseek-v4-pro-max": 76,
+        "qwen3-5-122b": 54,
+        "qwen3-coder-next": 66,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 78
+      }
+    },
+    {
+      "agent": "system-analyst",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 63,
+        "nemotron-3-super": 74,
+        "glm-5.1": 82,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "requirement-refiner",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 66,
+        "minimax-m2.5": 62,
+        "minimax-m2.7": 60,
+        "nemotron-3-super": 72,
+        "glm-5.1": 80,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 54,
+        "qwen3-6-plus": 78,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "history-miner",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 60,
+        "minimax-m2.7": 56,
+        "nemotron-3-super": 85,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "capability-analyst",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 76,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 75,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 79,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "orchestrator",
+      "current_model_index": -1,
+      "current_model_id": "kimi-k2.6:cloud",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 74,
+        "minimax-m2.5": 70,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 80,
+        "glm-5.1": 82,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 78,
+        "qwen3-coder-next": 62,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 92
+      }
+    },
+    {
+      "agent": "release-manager",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 74,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 78
+      }
+    },
+    {
+      "agent": "evaluator",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 73,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 81,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "prompt-optimizer",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 76,
+        "minimax-m2.5": 74,
+        "minimax-m2.7": 72,
+        "nemotron-3-super": 76,
+        "glm-5.1": 75,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 64,
+        "qwen3-6-plus": 83,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "product-owner",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 60,
+        "minimax-m2.5": 56,
+        "minimax-m2.7": 54,
+        "nemotron-3-super": 74,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 76,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 48,
+        "qwen3-6-plus": 78,
+        "kimi-k2-6": 76
+      }
+    },
+    {
+      "agent": "pipeline-judge",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 64,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 65,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "workflow-architect",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 62,
+        "minimax-m2.7": 60,
+        "nemotron-3-super": 76,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "markdown-validator",
+      "current_model_index": -1,
+      "current_model_id": "nemotron-3-nano:30b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 43,
+        "minimax-m2.5": 38,
+        "minimax-m2.7": 36,
+        "nemotron-3-super": 52,
+        "glm-5.1": 55,
+        "deepseek-v4-pro-max": 68,
+        "qwen3-5-122b": 56,
+        "qwen3-coder-next": 40,
+        "qwen3-6-plus": 50,
+        "kimi-k2-6": 56
+      }
+    },
+    {
+      "agent": "agent-architect",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 78,
+        "minimax-m2.5": 72,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 66,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "planner",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 80,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 78,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 85,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "reflector",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 80
+      }
+    },
+    {
+      "agent": "memory-manager",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 63,
+        "minimax-m2.5": 58,
+        "minimax-m2.7": 56,
+        "nemotron-3-super": 86,
+        "glm-5.1": 72,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 50,
+        "qwen3-6-plus": 87,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "architect-indexer",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 64,
+        "minimax-m2.7": 62,
+        "nemotron-3-super": 74,
+        "glm-5.1": 80,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 84
+      }
+    }
+  ],
+  "if_scores": {
+    "qwen3-coder-480b": 88,
+    "minimax-m2.5": 82,
+    "minimax-m2.7": 78,
+    "nemotron-3-super": 85,
+    "glm-5.1": 80,
+    "deepseek-v4-pro-max": 88,
+    "qwen3-5-122b": 86,
+    "qwen3-coder-next": 84,
+    "qwen3-6-plus": 90,
+    "kimi-k2-6": 91,
+    "deepseek-v4-flash": 86
+  },
+  "agent_current_config": [
+    {
+      "agent": "lead-developer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "frontend-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "php-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "python-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "backend-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "go-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "flutter-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "devops-engineer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "sdet-engineer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "code-skeptic",
+      "model": "ollama-cloud/minimax-m2.5",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "minimax",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "security-auditor",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "performance-engineer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "the-fixer",
+      "model": "ollama-cloud/minimax-m2.5",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "minimax",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "browser-automation",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "visual-tester",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "system-analyst",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "requirement-refiner",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "history-miner",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "capability-analyst",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "orchestrator",
+      "model": "ollama-cloud/kimi-k2.6:cloud",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "kimi",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "release-manager",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "evaluator",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "prompt-optimizer",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "product-owner",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "pipeline-judge",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "workflow-architect",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "markdown-validator",
+      "model": "ollama-cloud/nemotron-3-nano:30b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "agent-architect",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "planner",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "reflector",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "memory-manager",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "architect-indexer",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    }
+  ],
+  "recommendations": [
+    {
+      "agent": "[built-in] debug",
+      "from_model": "glm-5.1.1 (88)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★90) / K2.6 (★90) RE:High",
+      "to_provider": "Ollama Cloud",
+      "impact": "high",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "200K→1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=90 и K2.6=90 (TIE!), GLM-5.1=88. V4-Pro: LiveCodeBench 93.5(#1!), Terminal 67.9, 1M ctx для полного проекта. K2.6: 13h auto sessions. Оба лучше GLM-5.1. RE:High для debug."
+    },
+    {
+      "agent": "planner",
+      "from_model": "nemotron-3-super (80)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★88) RE:High",
+      "to_provider": "Ollama Cloud",
+      "impact": "high",
+      "quality_change": "+10%",
+      "speed_change": "~1x",
+      "context_change": "1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=88(лучший!), K2.6=86, GLM-5.1=85, Nem=80. V4-Pro: GPQA 90.1 (reasoning), 1M ctx сохраняется (vs потеря при K2.6). RE:High для chain-of-thought planning."
+    },
+    {
+      "agent": "go-developer",
+      "from_model": "qwen3-coder:480b (85)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★88) RE:Medium",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+4%",
+      "speed_change": "~1x",
+      "context_change": "256K→1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=88(лучший для Go!), K2.6=86, Qwen3Coder=85. DeepSeek модели традиционно сильны в Go/Rust. 1M ctx для крупных Go-проектов."
+    },
+    {
+      "agent": "history-miner",
+      "from_model": "nemotron-3-super (★85)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (86) + Nem fallback",
+      "to_provider": "Hybrid",
+      "impact": "medium",
+      "quality_change": "+1%",
+      "speed_change": "~1x",
+      "context_change": "1M",
+      "provider_change": "Ollama Cloud + Ollama",
+      "rationale": "V4-Pro=86 чуть лучше Nemotron=85. 1M ctx у обоих. MRCR 83.5 у V4-Pro — лучшее long-context retrieval. Nemotron как fallback (RULER 91.75%)."
+    },
+    {
+      "agent": "frontend-dev → M2.5",
+      "from_model": "qwen3-coder (90)",
+      "from_provider": "Ollama",
+      "to_model": "MiniMax M2.5 (★92) ✅",
+      "to_provider": "Ollama",
+      "impact": "low",
+      "quality_change": "+2%",
+      "speed_change": "=",
+      "context_change": "204K",
+      "provider_change": "Ollama",
+      "rationale": "Spec-writing, UI architect. APPLIED."
+    },
+    {
+      "agent": "devops → K2.6",
+      "from_model": "deepseek-v3.2",
+      "from_provider": "",
+      "to_model": "kimi-k2.6:cloud ✅",
+      "to_provider": "Ollama Cloud",
+      "impact": "low",
+      "quality_change": "+35%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "",
+      "rationale": "APPLIED."
+    },
+    {
+      "agent": "orchestrator",
+      "from_model": "glm-5.1.1 (★90)",
+      "from_provider": "Ollama",
+      "to_model": "K2.6 (★92) RE:Medium",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "200K→256K",
+      "provider_change": "Ollama Cloud",
+      "rationale": "K2.6=92★ всё ещё лучший для orchestration. V4-Pro=86 слабее. 300 sub-agent swarm."
+    },
+    {
+      "agent": "the-fixer",
+      "from_model": "minimax-m2.5 (★88)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro (★88) / K2.6 (★90)",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "128K→1M/256K",
+      "provider_change": "Ollama Cloud",
+      "rationale": "K2.6=90(лучший), V4-Pro=88=M2.5. M2.5 SWE-bench 80.2% стабильнее. Не срочно."
+    },
+    {
+      "agent": "Qwen3-Coder (7 coding)",
+      "from_model": "qwen3-coder",
+      "from_provider": "Ollama",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "Ollama",
+      "rationale": "lead=92★, backend=91★, python=90★."
+    },
+    {
+      "agent": "GLM-5.1 (12 agents)",
+      "from_model": "glm-5.1.1",
+      "from_provider": "Ollama",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "200K",
+      "provider_change": "",
+      "rationale": "orchestrator=90, system-analyst=90. SWE-Pro #1."
+    },
+    {
+      "agent": "Kimi K2.6 (3 agents)",
+      "from_model": "kimi-k2.6",
+      "from_provider": "Ollama Cloud",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "",
+      "rationale": "devops=88★, browser=86, agent-arch=86."
+    }
+  ],
+  "impact_data": [
+    {
+      "category": "debug GLM5.1→V4-Pro/K2.6",
+      "before": 88,
+      "after": 90,
+      "delta": 2,
+      "notes": "LiveCodeBench 93.5, Terminal 67.9"
+    },
+    {
+      "category": "planner Nem→V4-Pro Max",
+      "before": 80,
+      "after": 88,
+      "delta": 8,
+      "notes": "★88! GPQA 90.1, 1M ctx"
+    },
+    {
+      "category": "go-dev Coder→V4-Pro Max",
+      "before": 85,
+      "after": 88,
+      "delta": 3,
+      "notes": "★88! Go/Rust specialist, 1M ctx"
+    },
+    {
+      "category": "history-miner →V4-Pro",
+      "before": 85,
+      "after": 86,
+      "delta": 1,
+      "notes": "MRCR 83.5, long-context"
+    },
+    {
+      "category": "orchestrator →K2.6 (next)",
+      "before": 90,
+      "after": 92,
+      "delta": 2,
+      "notes": "300 sub-agent swarm"
+    },
+    {
+      "category": "frontend → M2.5 ✅",
+      "before": 90,
+      "after": 92,
+      "delta": 2,
+      "notes": "Spec-writing, UI architect"
+    },
+    {
+      "category": "devops → K2.6 ✅",
+      "before": 65,
+      "after": 88,
+      "delta": 23,
+      "notes": "IF:65→91! Terminal 66.7"
+    },
+    {
+      "category": "Qwen3-Coder (7) ✅",
+      "before": 90,
+      "after": 90,
+      "delta": 0,
+      "notes": "SOTA coding"
+    },
+    {
+      "category": "GLM-5.1 (12) ✅",
+      "before": 87,
+      "after": 87,
+      "delta": 0,
+      "notes": "SWE-Pro #1"
+    },
+    {
+      "category": "Nemotron Super (6) ✅",
+      "before": 82,
+      "after": 82,
+      "delta": 0,
+      "notes": "1M ctx, RULER 91.75%"
+    }
+  ],
+  "benchmark_comparison": {
+    "benchmarks": [
+      {
+        "name": "SWE-V",
+        "full_name": "SWE-Bench Verified",
+        "description": "GitHub issue resolution (500 tasks)",
+        "roles": "lead-dev, backend, fixer"
+      },
+      {
+        "name": "SWE-P",
+        "full_name": "SWE-Bench Pro",
+        "description": "Multi-lang, decontaminated (1865 tasks)",
+        "roles": "all coding agents"
+      },
+      {
+        "name": "T-Bench",
+        "full_name": "Terminal-Bench 2.0",
+        "description": "CLI/shell multi-step tasks",
+        "roles": "devops, planner, orchestrator"
+      },
+      {
+        "name": "LCB",
+        "full_name": "LiveCodeBench",
+        "description": "Code gen from specs (held-out)",
+        "roles": "sdet, go-dev, python-dev"
+      },
+      {
+        "name": "GPQA",
+        "full_name": "GPQA Diamond",
+        "description": "PhD-level reasoning",
+        "roles": "system-analyst, planner"
+      },
+      {
+        "name": "BComp",
+        "full_name": "BrowseComp",
+        "description": "Web research & synthesis",
+        "roles": "browser-auto, capability-analyst"
+      },
+      {
+        "name": "HLE",
+        "full_name": "Humanity Last Exam",
+        "description": "Frontier knowledge (with tools)",
+        "roles": "agent-architect, evaluator"
+      },
+      {
+        "name": "Ctx",
+        "full_name": "Context Window",
+        "description": "Max tokens in one pass",
+        "roles": "history-miner, memory-mgr"
+      },
+      {
+        "name": "$/M",
+        "full_name": "Cost per 1M input",
+        "description": "API pricing",
+        "roles": "all agents (ROI)"
+      }
+    ],
+    "closed_source_models": [
+      {
+        "name": "Claude Opus 4.7",
+        "organization": "Anthropic",
+        "scores": [
+          87.6,
+          64.3,
+          69.4,
+          null,
+          94.2,
+          79.3,
+          53,
+          "1M",
+          "$5"
+        ],
+        "color": "#c084fc",
+        "note": "#1 апрель 2026"
+      },
+      {
+        "name": "GPT-5.5",
+        "organization": "OpenAI",
+        "scores": [
+          null,
+          58.6,
+          82.7,
+          null,
+          null,
+          83.4,
+          57.2,
+          "1M",
+          "$5"
+        ],
+        "color": "#ff6b81",
+        "note": "Новейший, Terminal #1"
+      },
+      {
+        "name": "GPT-5.4",
+        "organization": "OpenAI",
+        "scores": [
+          78.2,
+          59.1,
+          75.1,
+          null,
+          94.4,
+          82.7,
+          58.7,
+          "200K",
+          "$2.50"
+        ],
+        "color": "#ff6b81",
+        "note": "Reasoning, math"
+      },
+      {
+        "name": "Gemini 3.1 Pro",
+        "organization": "Google",
+        "scores": [
+          80.6,
+          46.1,
+          68.5,
+          null,
+          94.3,
+          85.9,
+          51.4,
+          "2M",
+          "$2"
+        ],
+        "color": "#facc15",
+        "note": "ARC-AGI 77.1%, дешёвый"
+      },
+      {
+        "name": "Claude Sonnet 4.6",
+        "organization": "Anthropic",
+        "scores": [
+          79.6,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "200K",
+          "$3"
+        ],
+        "color": "#c084fc",
+        "note": "5× дешевле Opus"
+      },
+      {
+        "name": "GPT-5.3-Codex",
+        "organization": "OpenAI",
+        "scores": [
+          85,
+          57,
+          77.3,
+          null,
+          null,
+          null,
+          null,
+          "200K",
+          "$6"
+        ],
+        "color": "#ff6b81",
+        "note": "Coding specialist"
+      }
+    ],
+    "apaw_models": [
+      {
+        "name": "Kimi K2.6",
+        "organization": "APAW",
+        "scores": [
+          80.2,
+          58.6,
+          66.7,
+          87.2,
+          null,
+          83.2,
+          54,
+          "256K",
+          "$0.95"
+        ],
+        "color": "#00ff94",
+        "note": "devops, browser, architect (3)"
+      },
+      {
+        "name": "GLM-5.1",
+        "organization": "APAW",
+        "scores": [
+          null,
+          58.4,
+          63.5,
+          null,
+          86.2,
+          68.7,
+          null,
+          "200K",
+          "~$0.50"
+        ],
+        "color": "#00ff94",
+        "note": "12 agents! orchestrator, eval..."
+      },
+      {
+        "name": "V4-Pro Max",
+        "organization": "APAW",
+        "scores": [
+          80.6,
+          55.4,
+          67.9,
+          93.5,
+          90.1,
+          83.4,
+          48.2,
+          "1M",
+          "$0.42"
+        ],
+        "color": "#00d4ff",
+        "note": "planner, go-dev (рек.)"
+      },
+      {
+        "name": "Qwen3-Coder 480B",
+        "organization": "APAW",
+        "scores": [
+          66.5,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "256K",
+          "~$0.50"
+        ],
+        "color": "#00ff94",
+        "note": "7 coding agents"
+      },
+      {
+        "name": "MiniMax M2.5",
+        "organization": "APAW",
+        "scores": [
+          80.2,
+          51.3,
+          null,
+          null,
+          null,
+          76.3,
+          null,
+          "204K",
+          "$0.15"
+        ],
+        "color": "#00ff94",
+        "note": "frontend, skeptic, fixer (3)"
+      },
+      {
+        "name": "Nemotron Super",
+        "organization": "APAW",
+        "scores": [
+          60.5,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "1M",
+          "~$0.40"
+        ],
+        "color": "#00ff94",
+        "note": "6 agents (memory, history)"
+      }
+    ]
+  }
+}
\ No newline at end of file
diff --git a/agent-evolution/data/model-benchmarks.schema.json b/agent-evolution/data/model-benchmarks.schema.json
new file mode 100644
index 0000000..29ab918
--- /dev/null
+++ b/agent-evolution/data/model-benchmarks.schema.json
@@ -0,0 +1,553 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "https://app.kilo.ai/model-benchmarks.schema.json",
+  "title": "APAW Model Benchmarks Data",
+  "description": "Schema for static model benchmarks extracted from HTML sources",
+  "type": "object",
+  "required": [
+    "version",
+    "generated",
+    "source",
+    "metadata",
+    "models",
+    "groq_models",
+    "agent_model_scores",
+    "if_scores",
+    "agent_current_config",
+    "recommendations",
+    "impact_data",
+    "benchmark_comparison"
+  ],
+  "properties": {
+    "version": {
+      "type": "string",
+      "const": "1.0.0"
+    },
+    "generated": {
+      "type": "string",
+      "format": "date-time"
+    },
+    "source": {
+      "type": "string",
+      "description": "Source of benchmark data (e.g. HTML scraping, API, manual entry)"
+    },
+    "metadata": {
+      "type": "object",
+      "properties": {
+        "scrape_date": {
+          "type": "string",
+          "format": "date-time"
+        },
+        "source_urls": {
+          "type": "array",
+          "items": {
+            "type": "string"
+          }
+        },
+        "notes": {
+          "type": "string"
+        },
+        "data_quality": {
+          "type": "string",
+          "enum": [
+            "high",
+            "medium",
+            "low",
+            "estimated"
+          ]
+        }
+      }
+    },
+    "models": {
+      "type": "array",
+      "description": "All benchmarked models from various providers",
+      "items": {
+        "type": "object",
+        "required": [
+          "id",
+          "name",
+          "provider",
+          "category"
+        ],
+        "properties": {
+          "id": {
+            "type": "string",
+            "description": "Model identifier"
+          },
+          "name": {
+            "type": "string"
+          },
+          "organization": {
+            "type": "string"
+          },
+          "provider": {
+            "type": "string",
+            "enum": [
+              "ollama",
+              "ollama-cloud",
+              "openrouter",
+              "groq",
+              "anthropic",
+              "openai",
+              "meta",
+              "cohere",
+              "google",
+              "microsoft",
+              "unknown"
+            ]
+          },
+          "category": {
+            "type": "string",
+            "enum": [
+              "big",
+              "medium",
+              "small",
+              "coder",
+              "reasoning",
+              "creative"
+            ]
+          },
+          "parameters": {
+            "type": "string"
+          },
+          "benchmarks": {
+            "type": "object",
+            "properties": {
+              "swe_bench": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "swe_bench_pro": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "terminal_bench": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "live_codebench": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "gpqa": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "hle": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "browse_comp": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "m_mlu": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              },
+              "m_mlu_pro": {
+                "type": [
+                  "number",
+                  "null"
+                ]
+              }
+            }
+          },
+          "description": {
+            "type": "string"
+          },
+          "availability": {
+            "type": "object",
+            "properties": {
+              "rpm": {
+                "type": [
+                  "integer",
+                  "null"
+                ]
+              },
+              "rpd": {
+                "type": [
+                  "integer",
+                  "string",
+                  "null"
+                ]
+              },
+              "tpm": {
+                "type": [
+                  "integer",
+                  "string",
+                  "null"
+                ]
+              },
+              "tpd": {
+                "type": [
+                  "integer",
+                  "string",
+                  "null"
+                ]
+              }
+            }
+          },
+          "free": {
+            "type": "boolean"
+          },
+          "cost_per_1m_input": {
+            "type": [
+              "number",
+              "string",
+              "null"
+            ]
+          },
+          "tier": {
+            "type": "string",
+            "enum": [
+              "free",
+              "trial",
+              "paid",
+              "enterprise"
+            ]
+          }
+        }
+      }
+    },
+    "groq_models": {
+      "type": "array",
+      "description": "Groq-specific models with performance data",
+      "items": {
+        "type": "object",
+        "required": [
+          "id",
+          "name",
+          "speed_tps",
+          "provider"
+        ],
+        "properties": {
+          "id": {
+            "type": "string"
+          },
+          "name": {
+            "type": "string"
+          },
+          "speed_tps": {
+            "type": [
+              "number",
+              "string"
+            ]
+          },
+          "provider": {
+            "type": "string",
+            "const": "groq"
+          },
+          "benchmarks": {
+            "type": "object"
+          },
+          "availability": {
+            "type": "object"
+          }
+        }
+      }
+    },
+    "agent_model_scores": {
+      "type": "array",
+      "description": "Agent × Model compatibility scoring matrices",
+      "items": {
+        "type": "object",
+        "required": [
+          "agent",
+          "model_id",
+          "score",
+          "category"
+        ],
+        "properties": {
+          "agent": {
+            "type": "string"
+          },
+          "model_id": {
+            "type": "string"
+          },
+          "score": {
+            "type": "number",
+            "minimum": 0,
+            "maximum": 100
+          },
+          "category": {
+            "type": "string",
+            "enum": [
+              "performance",
+              "instruction_following",
+              "creativity",
+              "code_generation"
+            ]
+          },
+          "reason": {
+            "type": "string"
+          },
+          "timestamp": {
+            "type": "string",
+            "format": "date-time"
+          },
+          "current_model_id": {
+            "type": "string",
+            "description": "Current model ID string (replaces index)"
+          }
+        }
+      }
+    },
+    "if_scores": {
+      "type": "object",
+      "description": "Instruction Following scores mapping",
+      "additionalProperties": {
+        "type": "number",
+        "minimum": 0,
+        "maximum": 100
+      }
+    },
+    "agent_current_config": {
+      "type": "array",
+      "description": "Current agent model configurations",
+      "items": {
+        "type": "object",
+        "required": [
+          "agent",
+          "model",
+          "provider",
+          "status"
+        ],
+        "properties": {
+          "agent": {
+            "type": "string"
+          },
+          "model": {
+            "type": "string"
+          },
+          "provider": {
+            "type": "string"
+          },
+          "status": {
+            "type": "string",
+            "enum": [
+              "active",
+              "testing",
+              "deprecated",
+              "pending"
+            ]
+          },
+          "reasoning_effort": {
+            "type": "string",
+            "enum": [
+              "L",
+              "M",
+              "H"
+            ]
+          },
+          "fit_score": {
+            "type": "number"
+          },
+          "date_applied": {
+            "type": "string",
+            "format": "date-time"
+          }
+        }
+      }
+    },
+    "recommendations": {
+      "type": "array",
+      "description": "Model change recommendations based on benchmarks",
+      "items": {
+        "type": "object",
+        "required": [
+          "agent",
+          "action",
+          "current_model",
+          "recommended_model",
+          "impact"
+        ],
+        "properties": {
+          "agent": {
+            "type": "string"
+          },
+          "action": {
+            "type": "string",
+            "enum": [
+              "update_model",
+              "confirm_model",
+              "add_fallback",
+              "redesign_agent"
+            ]
+          },
+          "current_model": {
+            "type": "string"
+          },
+          "current_provider": {
+            "type": "string"
+          },
+          "recommended_model": {
+            "type": "string"
+          },
+          "recommended_provider": {
+            "type": "string"
+          },
+          "impact": {
+            "type": "string",
+            "enum": [
+              "critical",
+              "high",
+              "medium",
+              "low"
+            ]
+          },
+          "rationale": {
+            "type": "string"
+          },
+          "expected_improvement": {
+            "type": "object"
+          },
+          "applied": {
+            "type": "boolean"
+          }
+        }
+      }
+    },
+    "impact_data": {
+      "type": "array",
+      "description": "Impact analysis of model changes",
+      "items": {
+        "type": "object",
+        "required": [
+          "agent",
+          "model_change",
+          "impact_score"
+        ],
+        "properties": {
+          "agent": {
+            "type": "string"
+          },
+          "model_change": {
+            "type": "string"
+          },
+          "impact_score": {
+            "type": "number",
+            "minimum": 0,
+            "maximum": 100,
+            "description": "Impact score 0-100"
+          }
+        }
+      }
+    },
+    "benchmark_comparison": {
+      "type": "object",
+      "description": "APAW vs closed-source benchmark comparison",
+      "properties": {
+        "benchmarks": {
+          "type": "array",
+          "description": "Benchmark names used for comparison",
+          "items": {
+            "type": "string"
+          }
+        },
+        "closed_source_models": {
+          "type": "array",
+          "description": "Closed-source models included in comparison",
+          "items": {
+            "type": "object",
+            "properties": {
+              "name": {
+                "type": "string"
+              },
+              "provider": {
+                "type": "string"
+              },
+              "benchmarks": {
+                "type": "object"
+              }
+            }
+          }
+        },
+        "apaw_models": {
+          "type": "array",
+          "description": "APAW pipeline models included in comparison",
+          "items": {
+            "type": "object",
+            "properties": {
+              "name": {
+                "type": "string"
+              },
+              "provider": {
+                "type": "string"
+              },
+              "benchmarks": {
+                "type": "object"
+              }
+            }
+          }
+        },
+        "apaw_best": {
+          "type": "object",
+          "description": "Best APAW model per benchmark",
+          "additionalProperties": {
+            "type": "object",
+            "properties": {
+              "model": {
+                "type": "string"
+              },
+              "score": {
+                "type": "number"
+              },
+              "gap_to_closed": {
+                "type": [
+                  "number",
+                  "string"
+                ]
+              }
+            }
+          }
+        },
+        "closed_best": {
+          "type": "object",
+          "description": "Best closed-source model per benchmark",
+          "additionalProperties": {
+            "type": "object",
+            "properties": {
+              "model": {
+                "type": "string"
+              },
+              "score": {
+                "type": "number"
+              }
+            }
+          }
+        },
+        "summary": {
+          "type": "object",
+          "properties": {
+            "apaw_avg_score": {
+              "type": "number"
+            },
+            "closed_avg_score": {
+              "type": "number"
+            },
+            "coverage_gap": {
+              "type": "string"
+            }
+          }
+        }
+      }
+    }
+  }
+}
\ No newline at end of file
diff --git a/agent-evolution/data/model-research-latest.json b/agent-evolution/data/model-research-latest.json
new file mode 100644
index 0000000..a88b409
--- /dev/null
+++ b/agent-evolution/data/model-research-latest.json
@@ -0,0 +1,59 @@
+{
+  "version": "1.0.0",
+  "generated": "2026-04-27T17:51:36.000Z",
+  "source": "/research model-optimization",
+  "models": [],
+  "recommendations": [
+    {
+      "agent": "lead-developer",
+      "action": "update_model",
+      "current_model": "ollama-cloud/qwen3-coder:480b",
+      "current_provider": "ollama-cloud",
+      "recommended_model": "ollama-cloud/nemotron-3-super",
+      "recommended_provider": "ollama-cloud",
+      "impact": "high",
+      "expected_improvement": {
+        "quality": "+15%",
+        "speed": "+20%",
+        "context_window": "1M→1M"
+      },
+      "score_before": 85,
+      "score_after": 92,
+      "score_delta": 7,
+      "rationale": "Nemotron 3 Super has better reasoning for core development tasks and RULER@1M context window. SWE-bench 68% vs Qwen's 66.5%.",
+      "applied": false,
+      "applied_date": null
+    },
+    {
+      "agent": "devops-engineer",
+      "action": "confirm_model",
+      "current_model": "ollama-cloud/nemotron-3-super",
+      "current_provider": "ollama-cloud",
+      "recommended_model": "ollama-cloud/nemotron-3-super",
+      "recommended_provider": "ollama-cloud",
+      "impact": "low",
+      "expected_improvement": {
+        "quality": "0%",
+        "speed": "0%",
+        "context_window": "1M→1M"
+      },
+      "score_before": 88,
+      "score_after": 88,
+      "score_delta": 0,
+      "rationale": "Current model already optimal for DevOps tasks. Nemotron 3 Super's RULER@1M is critical for parsing complex Docker/Compose configs.",
+      "applied": false,
+      "applied_date": null
+    }
+  ],
+  "heatmap": {},
+  "closed_source_comparison": {},
+  "capability_index_patch": [],
+  "summary": {
+    "avg_quality_improvement": "+7.5%",
+    "providers_used": ["ollama-cloud"],
+    "key_models": ["nemotron-3-super"],
+    "total_recommendations": 2,
+    "applied_count": 0,
+    "pending_count": 2
+  }
+}
\ No newline at end of file
diff --git a/agent-evolution/data/model-research.schema.json b/agent-evolution/data/model-research.schema.json
new file mode 100644
index 0000000..db2518f
--- /dev/null
+++ b/agent-evolution/data/model-research.schema.json
@@ -0,0 +1,331 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "https://app.kilo.ai/model-research.schema.json",
+  "title": "APAW Model Research Output",
+  "description": "Schema for automated model research and recommendation output",
+  "type": "object",
+  "required": ["version", "generated", "source", "models", "recommendations", "heatmap"],
+  "properties": {
+    "version": {
+      "type": "string",
+      "const": "1.0.0"
+    },
+    "generated": {
+      "type": "string",
+      "format": "date-time"
+    },
+    "source": {
+      "type": "string",
+      "description": "What triggered this research (e.g. /evolution, /research, manual)"
+    },
+    "trigger": {
+      "type": "object",
+      "properties": {
+        "type": {
+          "type": "string",
+          "enum": ["evolution_cycle", "manual_research", "fitness_below_threshold", "scheduled"]
+        },
+        "issue": {
+          "type": "integer"
+        },
+        "fitness_score": {
+          "type": "number"
+        },
+        "reason": {
+          "type": "string"
+        }
+      }
+    },
+    "models": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "required": ["id", "name", "organization", "if_score", "provider"],
+        "properties": {
+          "id": {
+            "type": "string",
+            "description": "Full model ID like ollama-cloud/qwen3-coder:480b"
+          },
+          "name": {
+            "type": "string"
+          },
+          "organization": {
+            "type": "string"
+          },
+          "parameters": {
+            "type": "string"
+          },
+          "context_window": {
+            "type": "string"
+          },
+          "swe_bench": {
+            "type": ["number", "null"]
+          },
+          "swe_bench_pro": {
+            "type": ["number", "null"]
+          },
+          "terminal_bench": {
+            "type": ["number", "null"]
+          },
+          "live_codebench": {
+            "type": ["number", "null"]
+          },
+          "gpqa": {
+            "type": ["number", "null"]
+          },
+          "hle": {
+            "type": ["number", "null"]
+          },
+          "browse_comp": {
+            "type": ["number", "null"]
+          },
+  "if_score": {
+    "type": "number",
+    "minimum": 0,
+    "maximum": 100,
+    "description": "Instruction Following composite score (IFEval + IFBench)"
+  },
+          "categories": {
+            "type": "array",
+            "items": {
+              "type": "string"
+            }
+          },
+          "tags": {
+            "type": "array",
+            "items": {
+              "type": "string"
+            }
+          },
+          "provider": {
+            "type": "string",
+            "enum": ["ollama", "ollama-cloud", "openrouter", "groq", "hybrid"]
+          },
+          "free": {
+            "type": "boolean"
+          },
+          "cost_per_1m_input": {
+            "type": ["number", "string", "null"]
+          },
+          "description": {
+            "type": "string"
+          },
+          "availability": {
+            "type": "object",
+            "properties": {
+              "rpm": {
+                "type": ["integer", "null"]
+              },
+              "rpd": {
+                "type": ["integer", "string", "null"]
+              },
+              "tpm": {
+                "type": ["integer", "string", "null"]
+              },
+              "tpd": {
+                "type": ["integer", "string", "null"]
+              }
+            }
+          },
+          "speed_tps": {
+            "type": ["number", "string", "null"]
+          }
+        }
+      }
+    },
+    "recommendations": {
+      "type": "array",
+      "items": {
+        "type": "object",
+        "required": ["agent", "action", "current_model", "recommended_model", "impact", "rationale"],
+        "properties": {
+          "agent": {
+            "type": "string"
+          },
+          "action": {
+            "type": "string",
+            "enum": ["update_model", "confirm_model", "add_fallback", "redesign_agent"]
+          },
+          "current_model": {
+            "type": "string"
+          },
+          "current_provider": {
+            "type": "string"
+          },
+          "recommended_model": {
+            "type": "string"
+          },
+          "recommended_provider": {
+            "type": "string"
+          },
+          "fallback_model": {
+            "type": "string"
+          },
+          "fallback_strategy": {
+            "type": "string"
+          },
+          "impact": {
+            "type": "string",
+            "enum": ["critical", "high", "medium", "low"]
+          },
+          "expected_improvement": {
+            "type": "object",
+            "properties": {
+              "quality": {
+                "type": "string"
+              },
+              "speed": {
+                "type": "string"
+              },
+              "context_window": {
+                "type": "string"
+              }
+            }
+          },
+          "score_before": {
+            "type": "number"
+          },
+          "score_after": {
+            "type": "number"
+          },
+          "score_delta": {
+            "type": "number"
+          },
+          "rationale": {
+            "type": "string"
+          },
+          "applied": {
+            "type": "boolean",
+            "default": false
+          },
+          "applied_date": {
+            "type": ["string", "null"],
+            "format": "date-time"
+          }
+        }
+      }
+    },
+    "heatmap": {
+      "type": "object",
+      "description": "Agent × Model compatibility matrix with IF adjustment",
+      "required": ["models", "agents"],
+      "properties": {
+        "models": {
+          "type": "array",
+          "items": {
+            "type": "object",
+            "required": ["id", "if_score"],
+            "properties": {
+              "id": {
+                "type": "string"
+              },
+              "display_name": {
+                "type": "string"
+              },
+              "provider": {
+                "type": "string"
+              },
+              "if_score": {
+                "type": "number"
+              }
+            }
+          }
+        },
+        "agents": {
+          "type": "array",
+          "items": {
+            "type": "object",
+            "required": ["agent", "reasoning_effort", "scores"],
+            "properties": {
+              "agent": {
+                "type": "string"
+              },
+              "current_model": {
+                "type": "string"
+              },
+              "reasoning_effort": {
+                "type": "string",
+                "enum": ["L", "M", "H"]
+              },
+              "scores": {
+                "type": "object",
+                "additionalProperties": {
+                  "type": "number"
+                },
+                "description": "Model ID → compatibility score (0-100, IF-adjusted)"
+              }
+            }
+          }
+        },
+        "if_adjustment_formula": {
+          "type": "string",
+          "default": "score * (0.7 + 0.3 * IF/100)"
+        }
+      }
+    },
+    "closed_source_comparison": {
+      "type": "object",
+      "description": "APAW pipeline models vs top closed-source models",
+      "properties": {
+        "benchmarks": {
+          "type": "array"
+        },
+        "models": {
+          "type": "array"
+        },
+        "apaw_best_per_benchmark": {
+          "type": "object"
+        },
+        "closed_best_per_benchmark": {
+          "type": "object"
+        }
+      }
+    },
+    "capability_index_patch": {
+      "type": "array",
+      "description": "Ready-to-apply patches to capability-index.yaml",
+      "items": {
+        "type": "object",
+        "required": ["agent", "set"],
+        "properties": {
+          "agent": {
+            "type": "string"
+          },
+          "set": {
+            "type": "object",
+            "additionalProperties": true
+          }
+        }
+      }
+    },
+    "summary": {
+      "type": "object",
+      "properties": {
+        "avg_quality_improvement": {
+          "type": "string"
+        },
+        "providers_used": {
+          "type": "array",
+          "items": {
+            "type": "string"
+          }
+        },
+        "key_models": {
+          "type": "array",
+          "items": {
+            "type": "string"
+          }
+        },
+        "total_recommendations": {
+          "type": "integer"
+        },
+        "applied_count": {
+          "type": "integer"
+        },
+        "pending_count": {
+          "type": "integer"
+        }
+      }
+    }
+  }
+}
\ No newline at end of file
diff --git a/agent-evolution/dist/research-dashboard-2026_04_29.html b/agent-evolution/dist/research-dashboard-2026_04_29.html
new file mode 100644
index 0000000..f90334b
--- /dev/null
+++ b/agent-evolution/dist/research-dashboard-2026_04_29.html
@@ -0,0 +1,2777 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>APAW Agent Model Research — generated 2026-04-29</title>
+    <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@300;400;500;600;700&family=Outfit:wght@300;400;500;600;700;800;900&display=swap" rel="stylesheet">
+    <style>
+        :root {
+            --bg-deep: #080b12;
+            --bg-panel: #0e1219;
+            --bg-card: #141922;
+            --bg-card-hover: #1a2130;
+            --border: #1e2736;
+            --border-bright: #2a3650;
+            --text-primary: #e8edf5;
+            --text-secondary: #8896aa;
+            --text-muted: #5a6880;
+            --accent-cyan: #00d4ff;
+            --accent-green: #00ff94;
+            --accent-orange: #ff9f43;
+            --accent-red: #ff4757;
+            --accent-purple: #a855f7;
+            --accent-blue: #3b82f6;
+            --accent-yellow: #facc15;
+            --glow-cyan: rgba(0,212,255,0.15);
+            --glow-green: rgba(0,255,148,0.1);
+            --glow-orange: rgba(255,159,67,0.1);
+        }
+        * { margin:0; padding:0; box-sizing:border-box; }
+        body {
+            font-family:'Outfit',sans-serif;
+            background:var(--bg-deep);
+            color:var(--text-primary);
+            min-height:100vh;
+            overflow-x:hidden;
+        }
+        body::before {
+            content:'';
+            position:fixed; top:0; left:0; right:0; bottom:0;
+            background:
+                linear-gradient(90deg,rgba(0,212,255,0.02) 1px,transparent 1px),
+                linear-gradient(rgba(0,212,255,0.02) 1px,transparent 1px);
+            background-size:60px 60px;
+            animation:gp 8s ease-in-out infinite;
+            pointer-events:none; z-index:0;
+        }
+        @keyframes gp { 0%,100%{opacity:.3} 50%{opacity:.6} }
+        .container { max-width:1540px; margin:0 auto; padding:24px 16px; position:relative; z-index:1; }
+
+        /* HEADER */
+        .header { text-align:center; margin-bottom:32px; animation:fadeDown .7s ease-out; }
+        @keyframes fadeDown { from{opacity:0;transform:translateY(-20px)} to{opacity:1;transform:translateY(0)} }
+        .header h1 {
+            font-size:2.6em; font-weight:900;
+            background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green),var(--accent-purple));
+            -webkit-background-clip:text; -webkit-text-fill-color:transparent;
+            letter-spacing:-1px;
+        }
+        .header .sub { font-family:'JetBrains Mono',monospace; color:var(--text-muted); font-size:.8em; margin-top:6px; letter-spacing:1px; }
+
+        /* TABS */
+        .tabs { display:flex; gap:3px; background:var(--bg-panel); border:1px solid var(--border); border-radius:12px; padding:4px; margin-bottom:28px; overflow-x:auto; }
+        .tab-btn {
+            flex:1; min-width:100px; padding:10px 8px; background:none; border:none; color:var(--text-secondary);
+            font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600; border-radius:9px; cursor:pointer; transition:all .25s; white-space:nowrap;
+        }
+        .tab-btn:hover { color:var(--text-primary); background:var(--bg-card); }
+        .tab-btn.active { color:var(--bg-deep); background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); box-shadow:0 0 16px var(--glow-cyan); }
+        .tab-panel { display:none; animation:fadeUp .4s ease-out; }
+        .tab-panel.active { display:block; }
+        @keyframes fadeUp { from{opacity:0;transform:translateY(16px)} to{opacity:1;transform:translateY(0)} }
+
+        /* STAT CARDS */
+        .stats-row { display:grid; grid-template-columns:repeat(auto-fit,minmax(190px,1fr)); gap:14px; margin-bottom:26px; }
+        .stat-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:11px; padding:18px; position:relative; overflow:hidden; transition:all .3s;
+        }
+        .stat-card:hover { border-color:var(--accent-cyan); transform:translateY(-2px); box-shadow:0 6px 24px var(--glow-cyan); }
+        .stat-label { font-family:'JetBrains Mono',monospace; font-size:.65em; color:var(--text-muted); text-transform:uppercase; letter-spacing:1.5px; margin-bottom:6px; }
+        .stat-value { font-size:2em; font-weight:800; }
+        .stat-sub { font-size:.75em; color:var(--text-secondary); margin-top:3px; }
+        .grad-cyan { background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-orange { background:linear-gradient(135deg,var(--accent-orange),var(--accent-yellow)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-purple { background:linear-gradient(135deg,var(--accent-purple),#e879f9); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-green { background:linear-gradient(135deg,var(--accent-green),#4ade80); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-red { background:linear-gradient(135deg,var(--accent-red),#ff6b81); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+
+        /* SECTION HEADERS */
+        .sec-hdr { display:flex; align-items:center; gap:10px; margin-bottom:18px; padding-bottom:10px; border-bottom:1px solid var(--border); }
+        .sec-hdr h2 { font-size:1.2em; font-weight:700; }
+        .badge { font-family:'JetBrains Mono',monospace; font-size:.65em; padding:3px 9px; border-radius:16px; }
+        .badge-cyan { background:var(--glow-cyan); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.2); }
+        .badge-orange { background:var(--glow-orange); color:var(--accent-orange); border:1px solid rgba(255,159,67,.2); }
+        .badge-green { background:var(--glow-green); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); }
+
+        /* TABLES */
+        .tbl-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); margin-bottom:26px; }
+        table.dt { width:100%; border-collapse:collapse; font-size:.84em; }
+        table.dt th {
+            font-family:'JetBrains Mono',monospace; font-size:.7em; color:var(--text-muted); text-transform:uppercase;
+            letter-spacing:1.2px; padding:12px 14px; background:var(--bg-panel); border-bottom:2px solid var(--border); text-align:left; position:sticky; top:0;
+        }
+        table.dt td { padding:10px 14px; border-bottom:1px solid var(--border); transition:background .15s; }
+        table.dt tr:hover td { background:var(--bg-card-hover); }
+
+        .mbadge { display:inline-block; padding:3px 8px; border-radius:5px; font-family:'JetBrains Mono',monospace; font-size:.78em; font-weight:500; }
+        .mbadge.qwen { background:rgba(59,130,246,.12); color:#60a5fa; border:1px solid rgba(59,130,246,.25); }
+        .mbadge.gptoss { background:rgba(168,85,247,.12); color:#c084fc; border:1px solid rgba(168,85,247,.25); }
+        .mbadge.glm { background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); }
+        .mbadge.minimax { background:rgba(255,159,67,.12); color:#ff9f43; border:1px solid rgba(255,159,67,.25); }
+        .mbadge.devstral { background:rgba(0,212,255,.12); color:#00d4ff; border:1px solid rgba(0,212,255,.25); }
+        .mbadge.deepseek { background:rgba(250,204,21,.12); color:#facc15; border:1px solid rgba(250,204,21,.25); }
+        .mbadge.nemotron { background:rgba(34,197,94,.12); color:#4ade80; border:1px solid rgba(34,197,94,.25); }
+        .mbadge.groq { background:rgba(255,71,87,.12); color:#ff6b81; border:1px solid rgba(255,71,87,.25); }
+        .mbadge.kimi { background:rgba(250,204,21,.12); color:#fde68a; border:1px solid rgba(250,204,21,.2); }
+        .mbadge.llama { background:rgba(59,130,246,.1); color:#93c5fd; border:1px solid rgba(59,130,246,.2); }
+
+        /* SCORE BAR */
+        .sbar { display:flex; align-items:center; gap:6px; }
+        .sbar-bg { width:70px; height:5px; background:var(--border); border-radius:3px; overflow:hidden; }
+        .sbar-fill { height:100%; border-radius:3px; transition:width 1s ease-out; }
+        .sbar-fill.h { background:linear-gradient(90deg,var(--accent-green),#00ff94); }
+        .sbar-fill.m { background:linear-gradient(90deg,var(--accent-orange),#ffc048); }
+        .sbar-fill.l { background:linear-gradient(90deg,var(--accent-red),#ff6b81); }
+        .snum { font-family:'JetBrains Mono',monospace; font-weight:600; font-size:.85em; min-width:28px; }
+
+        /* GROQ SPEED INDICATOR */
+        .speed-ind { display:inline-flex; align-items:center; gap:4px; }
+        .speed-dot { width:7px; height:7px; border-radius:50%; animation:pulse 1.5s ease-in-out infinite; }
+        .speed-dot.ultra { background:var(--accent-green); box-shadow:0 0 8px var(--accent-green); }
+        .speed-dot.fast { background:var(--accent-cyan); box-shadow:0 0 6px var(--accent-cyan); }
+        .speed-dot.normal { background:var(--accent-orange); }
+        @keyframes pulse { 0%,100%{opacity:.5;transform:scale(.8)} 50%{opacity:1;transform:scale(1.2)} }
+
+        /* RECOMMENDATION CARDS */
+        .rec-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(400px,1fr)); gap:16px; margin-bottom:26px; }
+        .rec-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:20px;
+            position:relative; overflow:hidden; transition:all .35s;
+        }
+        .rec-card:hover { border-color:var(--accent-green); box-shadow:0 0 30px var(--glow-green); transform:translateY(-2px); }
+        .rec-card.glow { animation:glowP 3s ease-in-out infinite; }
+        @keyframes glowP { 0%,100%{box-shadow:0 0 16px var(--glow-green)} 50%{box-shadow:0 0 32px var(--glow-green)} }
+        .rec-hdr { display:flex; justify-content:space-between; align-items:flex-start; margin-bottom:12px; }
+        .rec-agent { font-weight:700; font-size:1em; color:var(--accent-cyan); }
+        .impact-badge { padding:2px 8px; border-radius:16px; font-family:'JetBrains Mono',monospace; font-size:.68em; font-weight:600; }
+        .impact-badge.critical { background:rgba(255,71,87,.18); color:var(--accent-red); border:1px solid rgba(255,71,87,.25); }
+        .impact-badge.high { background:rgba(255,159,67,.18); color:var(--accent-orange); border:1px solid rgba(255,159,67,.25); }
+        .impact-badge.medium { background:rgba(250,204,21,.18); color:var(--accent-yellow); border:1px solid rgba(250,204,21,.25); }
+        .swap-vis { display:flex; align-items:center; gap:10px; margin:12px 0; padding:12px; background:var(--bg-panel); border-radius:8px; }
+        .swap-from { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(255,71,87,.08); color:#ff6b81; border:1px solid rgba(255,71,87,.15); text-decoration:line-through; opacity:.65; }
+        .swap-to { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); font-weight:600; }
+        .swap-arrow { color:var(--accent-green); font-size:1.4em; animation:arrP 2s ease-in-out infinite; }
+        @keyframes arrP { 0%,100%{opacity:.4;transform:scale(1)} 50%{opacity:1;transform:scale(1.12)} }
+        .rec-metrics { display:grid; grid-template-columns:repeat(4,1fr); gap:8px; margin-top:12px; }
+        .rec-m { text-align:center; padding:6px; background:var(--bg-deep); border-radius:6px; }
+        .rec-m-label { font-size:.6em; color:var(--text-muted); text-transform:uppercase; letter-spacing:.8px; font-family:'JetBrains Mono',monospace; }
+        .rec-m-val { font-size:1.1em; font-weight:700; margin-top:1px; }
+        .rec-m-val.pos { color:var(--accent-green); }
+        .rec-m-val.neu { color:var(--accent-orange); }
+        .rec-reason { font-size:.82em; color:var(--text-secondary); line-height:1.55; margin-top:10px; padding-top:10px; border-top:1px solid var(--border); }
+
+        /* HEATMAP */
+        .hm-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .hm-title { font-weight:700; font-size:1.05em; }
+        .hm-sub { font-size:.76em; color:var(--text-muted); margin-bottom:14px; }
+        .hm-table { border-collapse:collapse; width:100%; }
+        .hm-table th { font-family:'JetBrains Mono',monospace; font-size:.62em; color:var(--text-muted); padding:6px 4px; text-align:center; white-space:nowrap; }
+        .hm-table th.hm-role { text-align:left; min-width:150px; font-size:.68em; }
+        .hm-table td { text-align:center; padding:5px 3px; font-family:'JetBrains Mono',monospace; font-size:.74em; font-weight:600; border-radius:3px; cursor:pointer; transition:all .12s; min-width:38px; }
+        .hm-table td:hover { transform:scale(1.12); z-index:2; }
+        .hm-table td.hm-r { text-align:left; font-family:'Outfit',sans-serif; font-size:.78em; font-weight:500; color:var(--text-secondary); cursor:default; }
+        .hm-table td.hm-r:hover { transform:none; }
+        .hm-star { color:#FFD700; font-size:.85em; }
+        .hm-cur { outline:2px solid var(--accent-cyan); outline-offset:-2px; }
+
+        /* PROVIDER TAGS */
+        .prov-tag { display:inline-block; padding:1px 6px; border-radius:3px; font-size:.62em; font-family:'JetBrains Mono',monospace; margin-left:4px; }
+        .prov-tag.ollama { background:rgba(0,212,255,.1); color:var(--accent-cyan); }
+        .prov-tag.groq { background:rgba(255,71,87,.1); color:#ff6b81; }
+        .prov-tag.openrouter { background:rgba(168,85,247,.1); color:#c084fc; }
+        .prov-tag.hybrid { background:rgba(0,255,148,.1); color:#00ff94; }
+
+        /* MODEL CARDS */
+        .model-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(290px,1fr)); gap:14px; margin-bottom:26px; }
+        .mc { background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:18px; transition:all .3s; position:relative; }
+        .mc:hover { transform:translateY(-2px); border-color:var(--accent-cyan); box-shadow:0 6px 24px var(--glow-cyan); }
+        .mc-name { font-weight:700; font-size:1.05em; margin-bottom:3px; }
+        .mc-org { font-size:.74em; color:var(--text-muted); margin-bottom:12px; font-family:'JetBrains Mono',monospace; }
+        .mc-row { display:flex; justify-content:space-between; align-items:center; padding:5px 0; border-bottom:1px solid rgba(30,39,54,.4); font-size:.82em; }
+        .mc-row:last-child { border-bottom:none; }
+        .mc-label { color:var(--text-secondary); }
+        .mc-val { font-family:'JetBrains Mono',monospace; font-weight:600; }
+        .mc-tags { display:flex; flex-wrap:wrap; gap:3px; margin-top:10px; }
+        .mc-tag { font-size:.64em; padding:2px 6px; border-radius:3px; font-family:'JetBrains Mono',monospace; background:rgba(0,212,255,.06); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.12); }
+        .mc-best { font-size:.72em; padding:3px 8px; border-radius:4px; background:rgba(0,255,148,.1); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); margin-top:8px; display:inline-block; }
+
+        /* GROQ SECTION */
+        .groq-card { border-left:3px solid var(--accent-red); }
+        .groq-speed { font-family:'JetBrains Mono',monospace; font-size:1.8em; font-weight:800; color:var(--accent-red); }
+
+        /* SUMMARY */
+        .summary { background:linear-gradient(135deg,rgba(0,212,255,.04),rgba(0,255,148,.04)); border:1px solid var(--border-bright); border-radius:12px; padding:24px; margin-bottom:26px; }
+        .summary h3 { color:var(--accent-cyan); font-size:1.1em; margin-bottom:10px; }
+        .summary p { color:var(--text-secondary); line-height:1.65; font-size:.88em; }
+        .summary ul { list-style:none; margin-top:10px; }
+        .summary li { padding:5px 0 5px 18px; position:relative; color:var(--text-secondary); font-size:.86em; line-height:1.55; }
+        .summary li::before { content:'›'; position:absolute; left:0; color:var(--accent-green); font-weight:700; font-size:1.2em; }
+
+        /* FILTER ROW */
+        .frow { display:flex; gap:6px; margin-bottom:16px; flex-wrap:wrap; }
+        .fbtn { padding:5px 12px; background:var(--bg-card); border:1px solid var(--border); color:var(--text-secondary); border-radius:7px; font-family:'Outfit',sans-serif; font-size:.78em; cursor:pointer; transition:all .2s; }
+        .fbtn:hover,.fbtn.active { border-color:var(--accent-cyan); color:var(--accent-cyan); background:rgba(0,212,255,.06); }
+
+        /* TOOLTIP */
+        #ttOverlay { display:none; position:fixed; top:0;left:0;right:0;bottom:0; z-index:999; pointer-events:none; }
+        #ttOverlay.show { display:block; }
+        #ttBox { position:absolute; background:var(--bg-panel); border:1px solid var(--accent-cyan); border-radius:9px; padding:12px 16px; max-width:300px; box-shadow:0 10px 32px rgba(0,0,0,.55); z-index:1000; }
+        #ttBox h4 { color:var(--accent-cyan); font-size:.9em; margin-bottom:4px; }
+        #ttBox p { font-size:.78em; color:var(--text-secondary); line-height:1.45; }
+
+        /* CANVAS */
+        .chart-wrap { border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .chart-title { font-weight:700; font-size:1.05em; margin-bottom:12px; }
+
+        @media(max-width:768px) {
+            .header h1 { font-size:1.6em; }
+            .tabs { flex-wrap:wrap; }
+            .rec-grid,.model-grid { grid-template-columns:1fr; }
+            .stats-row { grid-template-columns:repeat(2,1fr); }
+            .rec-metrics { grid-template-columns:repeat(2,1fr); }
+        }
+
+        /* EXPORT BUTTONS */
+        .export-btn {
+            padding:8px 16px; background:var(--bg-card); border:1px solid var(--border-bright);
+            color:var(--text-secondary); font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600;
+            border-radius:8px; cursor:pointer; transition:all .25s; display:inline-flex; align-items:center;
+        }
+        .export-btn:hover { border-color:var(--accent-cyan); color:var(--text-primary); background:var(--bg-card-hover); }
+        .export-btn-primary {
+            background:linear-gradient(135deg,rgba(0,212,255,.15),rgba(0,255,148,.1));
+            border-color:var(--accent-cyan); color:var(--accent-cyan);
+        }
+        .export-btn-primary:hover { background:linear-gradient(135deg,rgba(0,212,255,.25),rgba(0,255,148,.18)); box-shadow:0 0 20px var(--glow-cyan); }
+
+        /* REC CARD CHECKBOX */
+        .rec-check { position:absolute; top:14px; right:14px; width:22px; height:22px; cursor:pointer; z-index:3; accent-color:var(--accent-green); }
+        .rec-card.selected { border-color:var(--accent-green); background:rgba(0,255,148,.03); }
+        .rec-card { position:relative; }
+    </style>
+</head>
+<body>
+<div class="container">
+    <div class="header">
+        <h1>APAW Agent Model Research v2</h1>
+        <div class="sub">Live dashboard • 15 models × 32 agents • 2026-04-29</div>
+    </div>
+
+    <div class="tabs" id="tabBar">
+        <button class="tab-btn active" onclick="switchTab('overview')">Обзор</button>
+        <button class="tab-btn" onclick="switchTab('groq')">Groq Free Tier</button>
+        <button class="tab-btn" onclick="switchTab('models')">Все модели</button>
+        <button class="tab-btn" onclick="switchTab('heatmap')">Матрица</button>
+        <button class="tab-btn" onclick="switchTab('recs')">Рекомендации</button>
+        <button class="tab-btn" onclick="switchTab('impact')">Анализ профита</button>
+    </div>
+
+    <!-- ========== TAB: OVERVIEW ========== -->
+    <div id="tab-overview" class="tab-panel active">
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Агентов</div><div class="stat-value grad-cyan" id="c1">36</div><div class="stat-sub">32 custom + 4 built-in</div></div>
+            <div class="stat-card"><div class="stat-label">Моделей сейчас</div><div class="stat-value grad-orange">6</div><div class="stat-sub">Coder(9) GLM-5.1(11) K2.6(4)</div></div>
+            <div class="stat-card"><div class="stat-label">Ollama Cloud</div><div class="stat-value grad-purple">20+</div><div class="stat-sub">доступно бесплатно</div></div>
+            <div class="stat-card"><div class="stat-label">Groq + OpenRouter</div><div class="stat-value grad-red">16+</div><div class="stat-sub">free tier моделей</div></div>
+            <div class="stat-card"><div class="stat-label">Рекомендаций</div><div class="stat-value grad-green">11</div><div class="stat-sub">8/8 applied ✅</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Ключевые находки v3 (после коммита caf77f53c8)</h3>
+            <p>Ваш агент уже применил 11 из моих рекомендаций (коммит от 05:21). Но я обнаружил что <strong>до применения</strong> некоторые агенты были на других моделях чем я предполагал:</p>
+            <ul>
+                <li><strong style="color:var(--accent-red)">⚠ Откат Qwen 3.6 Plus</strong> — security-auditor, prompt-optimizer, product-owner и markdown-validator <em>до коммита</em> были на <code>openrouter/qwen3.6-plus:free</code> и <code>deepseek-v3.2</code>, но мои рекомендации их заменили на Ollama-модели. Это снижает разнообразие провайдеров!</li>
+                <li><strong style="color:var(--accent-green)">✅ 11 замен уже применены</strong> — Nemotron 3 Super теперь на 7 ролях, GLM-5 расширен, Qwen3-Coder на Go, markdown-validator</li>
+                <li><strong style="color:var(--accent-orange)">🔴 Осталось 3 агента на gpt-oss:120b</strong> — requirement-refiner, capability-analyst, agent-architect. Всем им нужен Nemotron 3 Super</li>
+                <li><strong>Новая стратегия: гибридный мультипровайдер</strong> — OpenRouter (Qwen 3.6 Plus FREE, 1M ctx) + Groq (gpt-oss 500 t/s) + Ollama (основной). Диверсификация снижает зависимость</li>
+                <li><strong style="color:#00ff94">Qwen 3.6 Plus стоит вернуть</strong> для prompt-optimizer (Terminal-Bench 61.6% > Claude!) и product-owner (1M контекст для backlog)</li>
+                
+                <li><strong>History-miner → Nemotron 3 Super</strong> — самый большой оставшийся прирост: 88 vs 78 (GLM-5). RULER@1M критичен для git history</li>
+                <li><strong style="color:var(--accent-red)">⚠ Prompt Adherence (IF) — новый фактор!</strong> Nemotron 3 Super имеет IF=78 (ниже GLM-5=90, Qwen3.5=92, Qwen3.6+=91). Для ролей с жёстким промптом (evaluator, security-auditor, orchestrator) это снижает эффективность. Qwen 3.6 Plus и GLM-5 лучше следуют инструкциям</li></ul>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Текущая конфигурация</h2><span class="badge badge-cyan">capability-index.yaml</span></div>
+        <div class="tbl-wrap">
+            <table class="dt" id="cfgTable"><thead><tr>
+                <th>Агент</th><th>Модель</th><th>Провайдер</th><th>Категория</th><th>Соответствие</th><th>Статус</th>
+            </tr></thead><tbody id="cfgBody"></tbody></table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: GROQ ========== -->
+    <div id="tab-groq" class="tab-panel">
+        <div class="sec-hdr"><h2>Groq Free Plan — доступные модели</h2><span class="badge badge-orange">бесплатно · LPU inference</span></div>
+
+        <div class="stats-row">
+            <div class="stat-card groq-card"><div class="stat-label">gpt-oss-20b</div><div class="groq-speed">1200 <span style="font-size:.4em;color:var(--text-muted)">t/s</span></div><div class="stat-sub">30 RPM · 1K RPD · 200K TPD</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Анализ лимитов Groq Free для агентского pipeline</h3>
+            <p>При 26 агентах в pipeline, каждый агент делает 5–20 вызовов на задачу. Типичный issue проходит через 8–12 агентов = <strong>~100–200 вызовов</strong>. С лимитом 1K RPD на модель:</p>
+            <ul>
+                
+                
+                <li><strong>Groq Compound</strong>: всего 250 RPD, но 70K TPM — для одноразовых тяжёлых аналитических задач</li>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Все модели Groq Free Tier</h2></div>
+        <div class="tbl-wrap">
+            <table class="dt">
+                <thead><tr><th>Model ID</th><th>RPM</th><th>RPD</th><th>TPM</th><th>TPD</th><th>Скорость</th><th>Применение в APAW</th></tr></thead>
+                <tbody id="groqBody"></tbody>
+            </table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: MODELS ========== -->
+    <div id="tab-models" class="tab-panel">
+        <div class="sec-hdr"><h2>Все доступные модели</h2><span class="badge badge-cyan">Ollama Cloud + Groq + OpenRouter Free</span></div>
+        <div class="frow" id="filterRow"></div>
+        <div class="model-grid" id="modelGrid"></div>
+    </div>
+
+    <!-- ========== TAB: HEATMAP ========== -->
+    <div id="tab-heatmap" class="tab-panel">
+        <div class="hm-wrap">
+            <div class="hm-title">Матрица «Агент × Модель»: оценка совместимости (с учётом Prompt Adherence)</div>
+            <div class="hm-sub">0–100 · Взвешенная оценка = 60% бенчмарк роли + 25% Instruction Following + 15% скорость/контекст · ★ = лучший · <span style="outline:2px solid var(--accent-cyan);outline-offset:-2px;padding:0 3px;border-radius:2px">обведено</span> = текущий · <strong style="color:var(--accent-yellow)">← 11 моделей · 🟢L 🟡M 🔴H = Reasoning Effort →</strong></div>
+            <div style="overflow-x:auto"><table class="hm-table" id="hmTable"></table></div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: RECOMMENDATIONS ========== -->
+    <div id="tab-recs" class="tab-panel">
+        <div class="sec-hdr"><h2>Рекомендации</h2><span class="badge badge-green">4 замены (2 BROKEN) + 7 подтверждений 06.04.2026</span></div>
+
+        <div style="display:flex;gap:10px;margin-bottom:18px;flex-wrap:wrap;align-items:center;">
+            <button onclick="toggleAllRecs()" class="export-btn" id="selectAllBtn">☑ Выбрать все</button>
+            <button onclick="exportJSON()" class="export-btn export-btn-primary" id="exportBtn">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" style="vertical-align:-2px;margin-right:4px"><path d="M21 15v4a2 2 0 01-2 2H5a2 2 0 01-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>
+                Скачать JSON для агента
+            </button>
+            <span id="selectedCount" style="font-family:'JetBrains Mono',monospace;font-size:.75em;color:var(--text-muted);">0 из 11 выбрано</span>
+        </div>
+
+        <div class="rec-grid" id="recGrid"></div>
+
+        <!-- JSON Preview Modal -->
+        <div id="jsonModal" style="display:none;position:fixed;top:0;left:0;right:0;bottom:0;background:rgba(0,0,0,.7);z-index:9999;justify-content:center;align-items:center;padding:20px;">
+            <div style="background:var(--bg-panel);border:1px solid var(--accent-cyan);border-radius:14px;max-width:800px;width:100%;max-height:85vh;display:flex;flex-direction:column;box-shadow:0 20px 60px rgba(0,0,0,.5);">
+                <div style="display:flex;justify-content:space-between;align-items:center;padding:18px 22px;border-bottom:1px solid var(--border);">
+                    <div>
+                        <div style="font-weight:700;font-size:1.05em;">agent-model-recommendations.json</div>
+                        <div style="font-size:.75em;color:var(--text-muted);margin-top:2px;font-family:'JetBrains Mono',monospace">Готов для передачи агенту-оркестратору</div>
+                    </div>
+                    <div style="display:flex;gap:8px;">
+                        <button onclick="copyJSON()" class="export-btn" id="copyBtn">📋 Копировать</button>
+                        <button onclick="downloadJSON()" class="export-btn export-btn-primary">⬇ Скачать .json</button>
+                        <button onclick="closeModal()" class="export-btn" style="border-color:var(--accent-red);color:var(--accent-red);">✕</button>
+                    </div>
+                </div>
+                <pre id="jsonPreview" style="flex:1;overflow:auto;padding:18px 22px;margin:0;font-family:'JetBrains Mono',monospace;font-size:.78em;line-height:1.6;color:var(--accent-green);background:var(--bg-deep);border-radius:0 0 14px 14px;"></pre>
+            </div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: IMPACT ========== -->
+    <div id="tab-impact" class="tab-panel">
+        <div class="sec-hdr"><h2>Совокупный анализ профита</h2><span class="badge badge-green">если применить все рекомендации</span></div>
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Средний прирост</div><div class="stat-value grad-green">+12</div><div class="stat-sub">пунктов по матрице</div></div>
+            <div class="stat-card"><div class="stat-label">Применено</div><div class="stat-value grad-red">8/8</div><div class="stat-sub">все рекомендации ✅</div></div>
+            <div class="stat-card"><div class="stat-label">Qwen 3.6+</div><div class="stat-value grad-purple">0</div><div class="stat-sub">полностью на Ollama!</div></div>
+            <div class="stat-card"><div class="stat-label">GLM-5.1</div><div class="stat-value grad-orange">12</div><div class="stat-sub">10 custom + 2 built-in</div></div>
+        </div>
+        <div class="chart-wrap">
+            <div class="chart-title">Прирост по категориям: до → после</div>
+            <canvas id="impactCanvas" height="340"></canvas>
+        </div>
+        <div class="summary">
+            <h3>Детальный анализ прироста</h3>
+            <div id="impactDetails"></div>
+
+            <div style="margin-top:32px">
+                <div class="sec-hdr"><h2>APAW Pipeline vs ТОП закрытых моделей (апрель 2026)</h2></div>
+                <p style="font-size:.82em;color:var(--text-muted);margin-bottom:16px">
+                    Сравнение лучших моделей в вашем pipeline с лидерами рынка по ключевым бенчмаркам. 
+                    <strong style="color:var(--accent-green)">🟢</strong> = APAW обгоняет, 
+                    <strong style="color:var(--accent-yellow)">🟡</strong> = на уровне (±3%), 
+                    <strong style="color:var(--accent-red)">🔴</strong> = отстаёт
+                </p>
+                <div style="overflow-x:auto">
+                <table id="benchTable" style="width:100%;border-collapse:collapse;font-size:.78em;font-family:'JetBrains Mono',monospace">
+                </table>
+                </div>
+                <p style="font-size:.72em;color:var(--text-muted);margin-top:12px">
+                    * SWE-V = SWE-Bench Verified, SWE-P = SWE-Bench Pro, T-Bench = Terminal-Bench 2.0, LCB = LiveCodeBench, GPQA = GPQA Diamond<br>
+                    Данные: swebench.com, marc0.dev, tokenmix.ai, ollama.com — апрель 2026. Стоимость: примерная за 1M input tokens.
+                </p>
+            </div>
+        </div>
+    </div>
+</div>
+
+<div id="ttOverlay"><div id="ttBox"></div></div>
+
+<script>
+// BENCHMARK_DATA_PLACEHOLDER - REPLACED BY BUILD SCRIPT
+// Generated from model-benchmarks.json on 2026-04-29T19:58:05.244Z
+const EMBEDDED_DATA = {
+  "version": "1.0.0",
+  "generated": "2026-04-29T19:56:51.418Z",
+  "source": ".kilo/capability-index.yaml (synced v2)",
+  "total_agents": 32,
+  "total_models_tracked": 11,
+  "providers": [
+    "ollama",
+    "ollama-cloud",
+    "openrouter",
+    "groq"
+  ],
+  "models": [
+    {
+      "id": "qwen3-coder-480b",
+      "name": "Qwen3-Coder 480B",
+      "organization": "Qwen",
+      "parameters": "480B/35B active",
+      "context_window": "256K→1M",
+      "swe_bench": 66.5,
+      "if_score": 88,
+      "categories": [
+        "coding",
+        "agent"
+      ],
+      "description": "SOTA open-source кодинг. Сравним с Claude Sonnet 4.",
+      "tags": [
+        "coding",
+        "agent",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "minimax-m2.5",
+      "name": "MiniMax M2.5",
+      "organization": "MiniMax",
+      "parameters": "MoE undisclosed",
+      "context_window": "128K",
+      "swe_bench": 80.2,
+      "if_score": 82,
+      "categories": [
+        "coding",
+        "agent"
+      ],
+      "description": "Лидер SWE-bench 80.2%. Полный lifecycle разработки.",
+      "tags": [
+        "coding",
+        "agent"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "minimax-m2.7",
+      "name": "MiniMax M2.7",
+      "organization": "MiniMax",
+      "parameters": "~10B active",
+      "context_window": "128K",
+      "swe_bench": 78,
+      "if_score": 80,
+      "categories": [
+        "coding",
+        "agent",
+        "efficient"
+      ],
+      "description": "Самообучаемая. 56.2% SWE-Pro. 100 TPS. $0.30/M.",
+      "tags": [
+        "coding",
+        "agent",
+        "self-evolving"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "deepseek-v4-pro-max",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "1.6T/49B active MoE",
+      "context_window": "1M",
+      "swe_bench": 80.6,
+      "if_score": 89,
+      "categories": [
+        "coding",
+        "agent",
+        "reasoning"
+      ],
+      "description": "SWE-V 80.6, LiveCodeBench 93.5(#1!), Terminal-Bench 67.9, Codeforces 3206, 1M ctx, 27% FLOPs vs V3.2. MIT.",
+      "tags": [
+        "coding",
+        "agent",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "deepseek-v4-flash",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "284B/13B active MoE",
+      "context_window": "1M",
+      "swe_bench": 79,
+      "if_score": 86,
+      "categories": [
+        "coding",
+        "efficient",
+        "agent"
+      ],
+      "description": "SWE-V ~79%, Flash Max = Pro уровень reasoning. 13B active = ультрабыстрый. 1M ctx. FP4+FP8. MIT.",
+      "tags": [
+        "coding",
+        "efficient",
+        "agent",
+        "thinking"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "kimi-k2-6",
+      "name": "Kimi K2.6",
+      "organization": "Moonshot AI",
+      "parameters": "1T/32B active MoE",
+      "context_window": "256K",
+      "swe_bench": 80.2,
+      "if_score": 91,
+      "categories": [
+        "coding",
+        "agent",
+        "multimodal"
+      ],
+      "description": "SWE-Pro 58.6(#1!), SWE-V 80.2, Terminal-Bench 66.7, HLE 54.0(#1!), BrowseComp 83.2. 13h autonomous. 300 sub-agent swarm. Modified MIT.",
+      "tags": [
+        "coding",
+        "agent",
+        "swarm",
+        "vision",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "nemotron-3-super",
+      "name": "Nemotron 3 Super",
+      "organization": "NVIDIA",
+      "parameters": "120B/12B active",
+      "context_window": "1M",
+      "swe_bench": 60.5,
+      "if_score": 78,
+      "categories": [
+        "agent",
+        "reasoning",
+        "efficient"
+      ],
+      "description": "SWE-bench 60.5%. RULER@1M 91.75%! Но IF ниже — Mamba-layers иногда «теряют» инструкции в длинных промптах.",
+      "tags": [
+        "agent",
+        "1M-ctx",
+        "thinking"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "glm-5.1",
+      "name": "GLM-5",
+      "organization": "Z.ai",
+      "parameters": "744B/40B active",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 90,
+      "categories": [
+        "reasoning",
+        "agent"
+      ],
+      "description": "Мощный reasoning. Arena ELO 1451. Отличный instruction following (IFEval ~90+).",
+      "tags": [
+        "reasoning",
+        "agent"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "deepseek-v4",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "Large MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 75,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "Хороший reasoning, но IF нестабилен — иногда игнорирует формат вывода.",
+      "tags": [
+        "reasoning"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-5-122b",
+      "name": "Qwen 3.5 122B",
+      "organization": "Qwen",
+      "parameters": "122B/10B active",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 92,
+      "categories": [
+        "reasoning",
+        "efficient"
+      ],
+      "description": "IFEval 92.6%! Лучший IF среди open-source. Multimodal. Thinking.",
+      "tags": [
+        "vision",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-coder-next",
+      "name": "Qwen3-Coder-Next",
+      "organization": "Qwen",
+      "parameters": "80B/3B active",
+      "context_window": "128K",
+      "swe_bench": 70,
+      "if_score": 84,
+      "categories": [
+        "coding",
+        "efficient"
+      ],
+      "description": "70% SWE-bench с 3B active! Хороший IF для кодинга.",
+      "tags": [
+        "coding",
+        "efficient",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "cogito-2-1-671b",
+      "name": "Cogito 2.1 671B",
+      "organization": "Cognitive",
+      "parameters": "671B MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 76,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "MIT лицензия. 671B total. IF неплохой, но уступает GLM/Qwen.",
+      "tags": [
+        "reasoning"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-6-plus",
+      "name": "Qwen 3.6 Plus",
+      "organization": "Qwen",
+      "parameters": "Hybrid MoE",
+      "context_window": "1M",
+      "swe_bench": 78.8,
+      "if_score": 91,
+      "categories": [
+        "coding",
+        "agent",
+        "reasoning"
+      ],
+      "description": "FREE на OpenRouter! 1M контекст. Always-on CoT. Превосходный IF — наследник Qwen 3.5 (92.6%).",
+      "tags": [
+        "coding",
+        "agent",
+        "1M-ctx",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    },
+    {
+      "id": "step-3-5-flash",
+      "name": "Step 3.5 Flash",
+      "organization": "StepFun",
+      "parameters": "MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 79,
+      "categories": [
+        "efficient"
+      ],
+      "description": "Бесплатна на OpenRouter. IF средний.",
+      "tags": [
+        "efficient",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    },
+    {
+      "id": "deepseek-r1",
+      "name": "DeepSeek R1",
+      "organization": "DeepSeek",
+      "parameters": "671B MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 73,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "Мощные reasoning-цепочки. Но IF слабый — часто генерирует лишний reasoning вместо ответа.",
+      "tags": [
+        "reasoning",
+        "thinking",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    }
+  ],
+  "groq_models": [
+    {
+      "id": "openai/gpt-oss-20b",
+      "rpm": 30,
+      "rpd": "1K",
+      "tpm": "8K",
+      "tpd": "200K",
+      "speed": "1200+",
+      "use_case": "Ультра-быстрый fallback для лёгких ролей (markdown-validator)."
+    },
+    {
+      "id": "llama-3.1-8b-instant",
+      "rpm": 30,
+      "rpd": "14.4K",
+      "tpm": "6K",
+      "tpd": "500K",
+      "speed": "~800",
+      "use_case": "14.4K RPD! Самый высокий лимит. Для health-check / ping ролей."
+    },
+    {
+      "id": "groq/compound",
+      "rpm": 30,
+      "rpd": "250",
+      "tpm": "70K",
+      "tpd": "—",
+      "speed": "varies",
+      "use_case": "Мультимодельная агрегация. Для research-задач."
+    },
+    {
+      "id": "groq/compound-mini",
+      "rpm": 30,
+      "rpd": "250",
+      "tpm": "70K",
+      "tpd": "—",
+      "speed": "varies",
+      "use_case": "Лёгкая версия compound."
+    },
+    {
+      "id": "llama-prompt-guard-2",
+      "rpm": 30,
+      "rpd": "14.4K",
+      "tpm": "15K",
+      "tpd": "500K",
+      "speed": "~1K",
+      "use_case": "Security: входной фильтр для security-auditor (14.4K RPD!)."
+    }
+  ],
+  "agent_model_scores": [
+    {
+      "agent": "lead-developer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 92,
+        "minimax-m2.5": 86,
+        "minimax-m2.7": 82,
+        "nemotron-3-super": 70,
+        "glm-5.1": 68,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 66,
+        "qwen3-coder-next": 80,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "frontend-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 86,
+        "minimax-m2.5": 92,
+        "minimax-m2.7": 88,
+        "nemotron-3-super": 62,
+        "glm-5.1": 56,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 60,
+        "qwen3-coder-next": 76,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "php-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 87,
+        "minimax-m2.5": 76,
+        "minimax-m2.7": 72,
+        "nemotron-3-super": 64,
+        "glm-5.1": 56,
+        "deepseek-v4-pro-max": 74,
+        "qwen3-5-122b": 60,
+        "qwen3-coder-next": 76,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "python-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 90,
+        "minimax-m2.5": 82,
+        "minimax-m2.7": 78,
+        "nemotron-3-super": 66,
+        "glm-5.1": 60,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 88
+      }
+    },
+    {
+      "agent": "backend-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 91,
+        "minimax-m2.5": 84,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 68,
+        "glm-5.1": 63,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 62,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 87,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "go-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 85,
+        "minimax-m2.5": 78,
+        "minimax-m2.7": 74,
+        "nemotron-3-super": 66,
+        "glm-5.1": 58,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 58,
+        "qwen3-coder-next": 74,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "flutter-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 86,
+        "minimax-m2.5": 70,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 60,
+        "glm-5.1": 53,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 58,
+        "qwen3-coder-next": 74,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "devops-engineer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 66,
+        "minimax-m2.5": 53,
+        "minimax-m2.7": 48,
+        "nemotron-3-super": 78,
+        "glm-5.1": 75,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 54,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 88
+      }
+    },
+    {
+      "agent": "sdet-engineer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 88,
+        "minimax-m2.5": 84,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 70,
+        "glm-5.1": 63,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 87
+      }
+    },
+    {
+      "agent": "code-skeptic",
+      "current_model_index": 1,
+      "current_model_id": "minimax-m2.5",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 82,
+        "minimax-m2.5": 85,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 73,
+        "glm-5.1": 72,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 72,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "security-auditor",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 76,
+        "minimax-m2.5": 74,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 76,
+        "glm-5.1": 68,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 64,
+        "qwen3-6-plus": 75,
+        "kimi-k2-6": 80
+      }
+    },
+    {
+      "agent": "performance-engineer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 78,
+        "minimax-m2.5": 75,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 74,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 67,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "the-fixer",
+      "current_model_index": 1,
+      "current_model_id": "minimax-m2.5",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 89,
+        "minimax-m2.5": 88,
+        "minimax-m2.7": 84,
+        "nemotron-3-super": 71,
+        "glm-5.1": 64,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 82,
+        "qwen3-6-plus": 86,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "browser-automation",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 87,
+        "minimax-m2.5": 72,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 61,
+        "glm-5.1": 53,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 56,
+        "qwen3-coder-next": 72,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "visual-tester",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 82,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 55,
+        "glm-5.1": 48,
+        "deepseek-v4-pro-max": 76,
+        "qwen3-5-122b": 54,
+        "qwen3-coder-next": 66,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 78
+      }
+    },
+    {
+      "agent": "system-analyst",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 63,
+        "nemotron-3-super": 74,
+        "glm-5.1": 82,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "requirement-refiner",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 66,
+        "minimax-m2.5": 62,
+        "minimax-m2.7": 60,
+        "nemotron-3-super": 72,
+        "glm-5.1": 80,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 54,
+        "qwen3-6-plus": 78,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "history-miner",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 60,
+        "minimax-m2.7": 56,
+        "nemotron-3-super": 85,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "capability-analyst",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 76,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 75,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 79,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "orchestrator",
+      "current_model_index": -1,
+      "current_model_id": "kimi-k2.6:cloud",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 74,
+        "minimax-m2.5": 70,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 80,
+        "glm-5.1": 82,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 78,
+        "qwen3-coder-next": 62,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 92
+      }
+    },
+    {
+      "agent": "release-manager",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 74,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 78
+      }
+    },
+    {
+      "agent": "evaluator",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 73,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 81,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "prompt-optimizer",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 76,
+        "minimax-m2.5": 74,
+        "minimax-m2.7": 72,
+        "nemotron-3-super": 76,
+        "glm-5.1": 75,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 64,
+        "qwen3-6-plus": 83,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "product-owner",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 60,
+        "minimax-m2.5": 56,
+        "minimax-m2.7": 54,
+        "nemotron-3-super": 74,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 76,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 48,
+        "qwen3-6-plus": 78,
+        "kimi-k2-6": 76
+      }
+    },
+    {
+      "agent": "pipeline-judge",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 64,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 65,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "workflow-architect",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 62,
+        "minimax-m2.7": 60,
+        "nemotron-3-super": 76,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "markdown-validator",
+      "current_model_index": -1,
+      "current_model_id": "nemotron-3-nano:30b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 43,
+        "minimax-m2.5": 38,
+        "minimax-m2.7": 36,
+        "nemotron-3-super": 52,
+        "glm-5.1": 55,
+        "deepseek-v4-pro-max": 68,
+        "qwen3-5-122b": 56,
+        "qwen3-coder-next": 40,
+        "qwen3-6-plus": 50,
+        "kimi-k2-6": 56
+      }
+    },
+    {
+      "agent": "agent-architect",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 78,
+        "minimax-m2.5": 72,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 66,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "planner",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 80,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 78,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 85,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "reflector",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 80
+      }
+    },
+    {
+      "agent": "memory-manager",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 63,
+        "minimax-m2.5": 58,
+        "minimax-m2.7": 56,
+        "nemotron-3-super": 86,
+        "glm-5.1": 72,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 50,
+        "qwen3-6-plus": 87,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "architect-indexer",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 64,
+        "minimax-m2.7": 62,
+        "nemotron-3-super": 74,
+        "glm-5.1": 80,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 84
+      }
+    }
+  ],
+  "if_scores": {
+    "qwen3-coder-480b": 88,
+    "minimax-m2.5": 82,
+    "minimax-m2.7": 78,
+    "nemotron-3-super": 85,
+    "glm-5.1": 80,
+    "deepseek-v4-pro-max": 88,
+    "qwen3-5-122b": 86,
+    "qwen3-coder-next": 84,
+    "qwen3-6-plus": 90,
+    "kimi-k2-6": 91,
+    "deepseek-v4-flash": 86
+  },
+  "agent_current_config": [
+    {
+      "agent": "lead-developer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "frontend-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "php-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "python-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "backend-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "go-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "flutter-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "devops-engineer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "sdet-engineer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "code-skeptic",
+      "model": "ollama-cloud/minimax-m2.5",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "minimax",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "security-auditor",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "performance-engineer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "the-fixer",
+      "model": "ollama-cloud/minimax-m2.5",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "minimax",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "browser-automation",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "visual-tester",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "system-analyst",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "requirement-refiner",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "history-miner",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "capability-analyst",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "orchestrator",
+      "model": "ollama-cloud/kimi-k2.6:cloud",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "kimi",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "release-manager",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "evaluator",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "prompt-optimizer",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "product-owner",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "pipeline-judge",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "workflow-architect",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "markdown-validator",
+      "model": "ollama-cloud/nemotron-3-nano:30b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "agent-architect",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "planner",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "reflector",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "memory-manager",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "architect-indexer",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    }
+  ],
+  "recommendations": [
+    {
+      "agent": "[built-in] debug",
+      "from_model": "glm-5.1.1 (88)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★90) / K2.6 (★90) RE:High",
+      "to_provider": "Ollama Cloud",
+      "impact": "high",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "200K→1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=90 и K2.6=90 (TIE!), GLM-5.1=88. V4-Pro: LiveCodeBench 93.5(#1!), Terminal 67.9, 1M ctx для полного проекта. K2.6: 13h auto sessions. Оба лучше GLM-5.1. RE:High для debug."
+    },
+    {
+      "agent": "planner",
+      "from_model": "nemotron-3-super (80)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★88) RE:High",
+      "to_provider": "Ollama Cloud",
+      "impact": "high",
+      "quality_change": "+10%",
+      "speed_change": "~1x",
+      "context_change": "1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=88(лучший!), K2.6=86, GLM-5.1=85, Nem=80. V4-Pro: GPQA 90.1 (reasoning), 1M ctx сохраняется (vs потеря при K2.6). RE:High для chain-of-thought planning."
+    },
+    {
+      "agent": "go-developer",
+      "from_model": "qwen3-coder:480b (85)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★88) RE:Medium",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+4%",
+      "speed_change": "~1x",
+      "context_change": "256K→1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=88(лучший для Go!), K2.6=86, Qwen3Coder=85. DeepSeek модели традиционно сильны в Go/Rust. 1M ctx для крупных Go-проектов."
+    },
+    {
+      "agent": "history-miner",
+      "from_model": "nemotron-3-super (★85)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (86) + Nem fallback",
+      "to_provider": "Hybrid",
+      "impact": "medium",
+      "quality_change": "+1%",
+      "speed_change": "~1x",
+      "context_change": "1M",
+      "provider_change": "Ollama Cloud + Ollama",
+      "rationale": "V4-Pro=86 чуть лучше Nemotron=85. 1M ctx у обоих. MRCR 83.5 у V4-Pro — лучшее long-context retrieval. Nemotron как fallback (RULER 91.75%)."
+    },
+    {
+      "agent": "frontend-dev → M2.5",
+      "from_model": "qwen3-coder (90)",
+      "from_provider": "Ollama",
+      "to_model": "MiniMax M2.5 (★92) ✅",
+      "to_provider": "Ollama",
+      "impact": "low",
+      "quality_change": "+2%",
+      "speed_change": "=",
+      "context_change": "204K",
+      "provider_change": "Ollama",
+      "rationale": "Spec-writing, UI architect. APPLIED."
+    },
+    {
+      "agent": "devops → K2.6",
+      "from_model": "deepseek-v3.2",
+      "from_provider": "",
+      "to_model": "kimi-k2.6:cloud ✅",
+      "to_provider": "Ollama Cloud",
+      "impact": "low",
+      "quality_change": "+35%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "",
+      "rationale": "APPLIED."
+    },
+    {
+      "agent": "orchestrator",
+      "from_model": "glm-5.1.1 (★90)",
+      "from_provider": "Ollama",
+      "to_model": "K2.6 (★92) RE:Medium",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "200K→256K",
+      "provider_change": "Ollama Cloud",
+      "rationale": "K2.6=92★ всё ещё лучший для orchestration. V4-Pro=86 слабее. 300 sub-agent swarm."
+    },
+    {
+      "agent": "the-fixer",
+      "from_model": "minimax-m2.5 (★88)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro (★88) / K2.6 (★90)",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "128K→1M/256K",
+      "provider_change": "Ollama Cloud",
+      "rationale": "K2.6=90(лучший), V4-Pro=88=M2.5. M2.5 SWE-bench 80.2% стабильнее. Не срочно."
+    },
+    {
+      "agent": "Qwen3-Coder (7 coding)",
+      "from_model": "qwen3-coder",
+      "from_provider": "Ollama",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "Ollama",
+      "rationale": "lead=92★, backend=91★, python=90★."
+    },
+    {
+      "agent": "GLM-5.1 (12 agents)",
+      "from_model": "glm-5.1.1",
+      "from_provider": "Ollama",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "200K",
+      "provider_change": "",
+      "rationale": "orchestrator=90, system-analyst=90. SWE-Pro #1."
+    },
+    {
+      "agent": "Kimi K2.6 (3 agents)",
+      "from_model": "kimi-k2.6",
+      "from_provider": "Ollama Cloud",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "",
+      "rationale": "devops=88★, browser=86, agent-arch=86."
+    }
+  ],
+  "impact_data": [
+    {
+      "category": "debug GLM5.1→V4-Pro/K2.6",
+      "before": 88,
+      "after": 90,
+      "delta": 2,
+      "notes": "LiveCodeBench 93.5, Terminal 67.9"
+    },
+    {
+      "category": "planner Nem→V4-Pro Max",
+      "before": 80,
+      "after": 88,
+      "delta": 8,
+      "notes": "★88! GPQA 90.1, 1M ctx"
+    },
+    {
+      "category": "go-dev Coder→V4-Pro Max",
+      "before": 85,
+      "after": 88,
+      "delta": 3,
+      "notes": "★88! Go/Rust specialist, 1M ctx"
+    },
+    {
+      "category": "history-miner →V4-Pro",
+      "before": 85,
+      "after": 86,
+      "delta": 1,
+      "notes": "MRCR 83.5, long-context"
+    },
+    {
+      "category": "orchestrator →K2.6 (next)",
+      "before": 90,
+      "after": 92,
+      "delta": 2,
+      "notes": "300 sub-agent swarm"
+    },
+    {
+      "category": "frontend → M2.5 ✅",
+      "before": 90,
+      "after": 92,
+      "delta": 2,
+      "notes": "Spec-writing, UI architect"
+    },
+    {
+      "category": "devops → K2.6 ✅",
+      "before": 65,
+      "after": 88,
+      "delta": 23,
+      "notes": "IF:65→91! Terminal 66.7"
+    },
+    {
+      "category": "Qwen3-Coder (7) ✅",
+      "before": 90,
+      "after": 90,
+      "delta": 0,
+      "notes": "SOTA coding"
+    },
+    {
+      "category": "GLM-5.1 (12) ✅",
+      "before": 87,
+      "after": 87,
+      "delta": 0,
+      "notes": "SWE-Pro #1"
+    },
+    {
+      "category": "Nemotron Super (6) ✅",
+      "before": 82,
+      "after": 82,
+      "delta": 0,
+      "notes": "1M ctx, RULER 91.75%"
+    }
+  ],
+  "benchmark_comparison": {
+    "benchmarks": [
+      {
+        "name": "SWE-V",
+        "full_name": "SWE-Bench Verified",
+        "description": "GitHub issue resolution (500 tasks)",
+        "roles": "lead-dev, backend, fixer"
+      },
+      {
+        "name": "SWE-P",
+        "full_name": "SWE-Bench Pro",
+        "description": "Multi-lang, decontaminated (1865 tasks)",
+        "roles": "all coding agents"
+      },
+      {
+        "name": "T-Bench",
+        "full_name": "Terminal-Bench 2.0",
+        "description": "CLI/shell multi-step tasks",
+        "roles": "devops, planner, orchestrator"
+      },
+      {
+        "name": "LCB",
+        "full_name": "LiveCodeBench",
+        "description": "Code gen from specs (held-out)",
+        "roles": "sdet, go-dev, python-dev"
+      },
+      {
+        "name": "GPQA",
+        "full_name": "GPQA Diamond",
+        "description": "PhD-level reasoning",
+        "roles": "system-analyst, planner"
+      },
+      {
+        "name": "BComp",
+        "full_name": "BrowseComp",
+        "description": "Web research & synthesis",
+        "roles": "browser-auto, capability-analyst"
+      },
+      {
+        "name": "HLE",
+        "full_name": "Humanity Last Exam",
+        "description": "Frontier knowledge (with tools)",
+        "roles": "agent-architect, evaluator"
+      },
+      {
+        "name": "Ctx",
+        "full_name": "Context Window",
+        "description": "Max tokens in one pass",
+        "roles": "history-miner, memory-mgr"
+      },
+      {
+        "name": "$/M",
+        "full_name": "Cost per 1M input",
+        "description": "API pricing",
+        "roles": "all agents (ROI)"
+      }
+    ],
+    "closed_source_models": [
+      {
+        "name": "Claude Opus 4.7",
+        "organization": "Anthropic",
+        "scores": [
+          87.6,
+          64.3,
+          69.4,
+          null,
+          94.2,
+          79.3,
+          53,
+          "1M",
+          "$5"
+        ],
+        "color": "#c084fc",
+        "note": "#1 апрель 2026"
+      },
+      {
+        "name": "GPT-5.5",
+        "organization": "OpenAI",
+        "scores": [
+          null,
+          58.6,
+          82.7,
+          null,
+          null,
+          83.4,
+          57.2,
+          "1M",
+          "$5"
+        ],
+        "color": "#ff6b81",
+        "note": "Новейший, Terminal #1"
+      },
+      {
+        "name": "GPT-5.4",
+        "organization": "OpenAI",
+        "scores": [
+          78.2,
+          59.1,
+          75.1,
+          null,
+          94.4,
+          82.7,
+          58.7,
+          "200K",
+          "$2.50"
+        ],
+        "color": "#ff6b81",
+        "note": "Reasoning, math"
+      },
+      {
+        "name": "Gemini 3.1 Pro",
+        "organization": "Google",
+        "scores": [
+          80.6,
+          46.1,
+          68.5,
+          null,
+          94.3,
+          85.9,
+          51.4,
+          "2M",
+          "$2"
+        ],
+        "color": "#facc15",
+        "note": "ARC-AGI 77.1%, дешёвый"
+      },
+      {
+        "name": "Claude Sonnet 4.6",
+        "organization": "Anthropic",
+        "scores": [
+          79.6,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "200K",
+          "$3"
+        ],
+        "color": "#c084fc",
+        "note": "5× дешевле Opus"
+      },
+      {
+        "name": "GPT-5.3-Codex",
+        "organization": "OpenAI",
+        "scores": [
+          85,
+          57,
+          77.3,
+          null,
+          null,
+          null,
+          null,
+          "200K",
+          "$6"
+        ],
+        "color": "#ff6b81",
+        "note": "Coding specialist"
+      }
+    ],
+    "apaw_models": [
+      {
+        "name": "Kimi K2.6",
+        "organization": "APAW",
+        "scores": [
+          80.2,
+          58.6,
+          66.7,
+          87.2,
+          null,
+          83.2,
+          54,
+          "256K",
+          "$0.95"
+        ],
+        "color": "#00ff94",
+        "note": "devops, browser, architect (3)"
+      },
+      {
+        "name": "GLM-5.1",
+        "organization": "APAW",
+        "scores": [
+          null,
+          58.4,
+          63.5,
+          null,
+          86.2,
+          68.7,
+          null,
+          "200K",
+          "~$0.50"
+        ],
+        "color": "#00ff94",
+        "note": "12 agents! orchestrator, eval..."
+      },
+      {
+        "name": "V4-Pro Max",
+        "organization": "APAW",
+        "scores": [
+          80.6,
+          55.4,
+          67.9,
+          93.5,
+          90.1,
+          83.4,
+          48.2,
+          "1M",
+          "$0.42"
+        ],
+        "color": "#00d4ff",
+        "note": "planner, go-dev (рек.)"
+      },
+      {
+        "name": "Qwen3-Coder 480B",
+        "organization": "APAW",
+        "scores": [
+          66.5,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "256K",
+          "~$0.50"
+        ],
+        "color": "#00ff94",
+        "note": "7 coding agents"
+      },
+      {
+        "name": "MiniMax M2.5",
+        "organization": "APAW",
+        "scores": [
+          80.2,
+          51.3,
+          null,
+          null,
+          null,
+          76.3,
+          null,
+          "204K",
+          "$0.15"
+        ],
+        "color": "#00ff94",
+        "note": "frontend, skeptic, fixer (3)"
+      },
+      {
+        "name": "Nemotron Super",
+        "organization": "APAW",
+        "scores": [
+          60.5,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "1M",
+          "~$0.40"
+        ],
+        "color": "#00ff94",
+        "note": "6 agents (memory, history)"
+      }
+    ]
+  }
+};
+
+// === MAP EMBEDDED_DATA -> original v3 format ===
+const allModels = EMBEDDED_DATA.models || [];
+const scoreModelIds = Object.keys((EMBEDDED_DATA.agent_model_scores || [])[0]?.scores || {});
+const activeModels = allModels.filter(m => scoreModelIds.includes(m.id));
+
+const cfg = (EMBEDDED_DATA.agent_current_config || []).map(c => {
+    const modelId = (c.model || '').replace('ollama-cloud/', '');
+    const badge = c.badge_type || (
+        modelId.includes('qwen3') ? 'qwen' :
+        modelId.includes('minimax') ? 'minimax' :
+        modelId.includes('nemotron') ? 'nemotron' :
+        modelId.includes('glm') ? 'glm' :
+        modelId.includes('kimi') ? 'kimi' :
+        modelId.includes('deepseek') ? 'deepseek' : 'groq'
+    );
+    return { a: c.agent, m: modelId, p: c.provider || 'Ollama', cat: c.category || 'General', b: badge, fit: c.fit_score || 0, s: c.status || 'good', prev: c.previous_model };
+});
+
+const groqModels = (EMBEDDED_DATA.groq_models || []).map(g => ({
+    id: g.id, rpm: g.rpm, rpd: g.rpd, tpm: g.tpm, tpd: g.tpd, speed: g.speed, use: g.use_case
+}));
+
+const ollamaModels = activeModels.map(m => ({
+    n: m.name, org: m.organization, par: m.parameters, ctx: m.context_window,
+    swe: m.swe_bench, ifScore: m.if_score, cat: m.categories || [],
+    str: m.description, tags: m.tags || [], or: m.openrouter, groqSpeed: m.speed_tps
+}));
+
+const ifScores = {};
+activeModels.forEach((m, i) => { if (m.if_score) ifScores[i] = m.if_score; });
+
+const hmModels = activeModels.map(m => ({
+    n: m.display_name || m.name?.split(' ').pop() || m.id,
+    p: m.provider === 'ollama-cloud' ? 'Ollama Cloud' : m.provider === 'openrouter' ? 'OpenRouter' : m.provider || 'Ollama',
+    if: m.if_score || 0
+}));
+
+const hmAgents = (EMBEDDED_DATA.agent_model_scores || []).map(ag => {
+    const scores = activeModels.map(m => ag.scores?.[m.id] ?? 0);
+    const fullModelId = allModels[ag.current_model_index]?.id;
+    const c = activeModels.findIndex(m => m.id === fullModelId);
+    return { n: ag.agent, c: c, re: ag.reasoning_effort || 'M', s: scores };
+});
+
+const recs = (EMBEDDED_DATA.recommendations || []).map(r => ({
+    a: r.agent, from: r.from_model, fromP: r.from_provider || 'Ollama',
+    to: r.to_model, toP: r.to_provider || 'Ollama', imp: r.impact || 'low',
+    q: r.quality_change || '0', sp: r.speed_change || '=', ctx: r.context_change || '-',
+    prov: r.provider_change || r.to_provider || 'Ollama', r: r.rationale
+}));
+
+const impactData = (EMBEDDED_DATA.impact_data || []).map(d => ({
+    cat: d.category, b: d.before, a: d.after, d: d.delta, n: d.notes || d.note
+}));
+
+// ======================= RENDER =======================
+function switchTab(id) {
+    document.querySelectorAll('.tab-panel').forEach(p=>p.classList.remove('active'));
+    document.querySelectorAll('.tab-btn').forEach(b=>b.classList.remove('active'));
+    document.getElementById('tab-'+id).classList.add('active');
+    event.target.classList.add('active');
+    if(id==='impact') requestAnimationFrame(()=>setTimeout(drawChart,50));
+}
+
+function renderCfg() {
+    const b=document.getElementById('cfgBody');
+    b.innerHTML=cfg.map((c,i)=>{
+        const si=c.s==='optimal'?'✅':c.s==='good'?'🟡':c.s==='overspec'?'🔵':c.s==='new'?'🆕':c.s==='broken'?'💀':'🔴';
+        const st=c.s==='optimal'?'Оптимально':c.s==='good'?'Хорошо':c.s==='overspec'?'Overspec':c.s==='new'?'Не назначена':c.s==='broken'?'НЕ РАБОТАЕТ':'Улучшить';
+        const bc=c.fit>=85?'h':c.fit>=70?'m':'l';
+        const sc=c.fit>=85?'var(--accent-green)':c.fit>=70?'var(--accent-orange)':'var(--accent-red)';
+        const prevHtml=c.prev?`<div style="font-size:.68em;color:var(--text-muted);margin-top:2px;text-decoration:line-through">${c.prev}</div>`:'';
+        return `<tr style="animation:fadeUp .3s ${i*.03}s ease-out both">
+            <td style="font-weight:600">${c.a}</td>
+            <td><span class="mbadge ${c.b}">${c.m}</span>${prevHtml}</td>
+            <td><span class="prov-tag ${c.p.toLowerCase()}">${c.p}</span></td>
+            <td style="color:var(--text-secondary)">${c.cat}</td>
+            <td><div class="sbar"><div class="sbar-bg"><div class="sbar-fill ${bc}" style="width:${c.fit}%"></div></div><span class="snum" style="color:${sc}">${c.fit}</span></div></td>
+            <td>${si} ${st}</td></tr>`;
+    }).join('');
+}
+
+function renderGroq() {
+    const b=document.getElementById('groqBody');
+    b.innerHTML=groqModels.map(g=>{
+        const spd=parseInt(g.speed)||0;
+        const dotCls=spd>=800?'ultra':spd>=400?'fast':'normal';
+        return `<tr>
+            <td><span class="mbadge groq">${g.id}</span></td>
+            <td>${g.rpm}</td><td>${g.rpd}</td><td>${g.tpm}</td><td>${g.tpd}</td>
+            <td><div class="speed-ind"><span class="speed-dot ${dotCls}"></span> ${g.speed} t/s</div></td>
+            <td style="color:var(--text-secondary);font-size:.82em;max-width:280px">${g.use}</td></tr>`;
+    }).join('');
+}
+
+function renderModels() {
+    const grid=document.getElementById('modelGrid');
+    const cats=new Set(); ollamaModels.forEach(m=>m.cat.forEach(c=>cats.add(c)));
+    document.getElementById('filterRow').innerHTML='<button class="fbtn active" onclick="filterM(\'all\',this)">Все</button>'+
+        [...cats].map(c=>`<button class="fbtn" onclick="filterM('${c}',this)">${c}</button>`).join('');
+
+    grid.innerHTML=ollamaModels.map((m,i)=>{
+        const bc=m.swe&&m.swe>=75?'var(--accent-green)':m.swe&&m.swe>=60?'var(--accent-cyan)':'var(--border)';
+        return `<div class="mc" style="animation:fadeUp .35s ${i*.05}s ease-out both;border-color:${bc}" data-cats='${JSON.stringify(m.cat)}'>
+            <div class="mc-name">${m.n} ${m.groq?'<span class="prov-tag groq">Groq '+m.groqSpeed+'t/s</span>':''}${m.or?'<span class="prov-tag openrouter">OpenRouter FREE</span>':''}</div>
+            <div class="mc-org">${m.org} · ${m.par} · ctx ${m.ctx}</div>
+            ${m.swe?`<div class="mc-row"><span class="mc-label">SWE-bench</span><span class="mc-val" style="color:${m.swe>=75?'var(--accent-green)':m.swe>=60?'var(--accent-cyan)':'var(--accent-orange)'}">${m.swe}%</span></div>`:''}
+            ${m.ifScore?`<div class="mc-row"><span class="mc-label">Prompt Adherence (IF)</span><span class="mc-val" style="color:${m.ifScore>=88?'var(--accent-green)':m.ifScore>=80?'var(--accent-cyan)':m.ifScore>=72?'var(--accent-orange)':'var(--accent-red)'}">${m.ifScore}<small>/100</small> ${m.ifScore>=88?'🎯':''}${m.ifScore<75?'⚠️':''}</span></div>`:''}
+            <div style="font-size:.78em;color:var(--text-secondary);line-height:1.45;margin-top:6px">${m.str}</div>
+            <div class="mc-tags">${m.tags.map(t=>`<span class="mc-tag">${t}</span>`).join('')}</div>
+        </div>`;
+    }).join('');
+}
+
+function filterM(cat,btn) {
+    document.querySelectorAll('.fbtn').forEach(b=>b.classList.remove('active'));
+    btn.classList.add('active');
+    document.querySelectorAll('.mc').forEach(c=>{
+        if(cat==='all'){c.style.display='';return;}
+        c.style.display=JSON.parse(c.dataset.cats).includes(cat)?'':'none';
+    });
+}
+
+function hmColor(v) {
+    if(v>=88) return 'rgba(0,255,148,.8)';
+    if(v>=82) return 'rgba(0,212,255,.7)';
+    if(v>=75) return 'rgba(59,130,246,.6)';
+    if(v>=68) return 'rgba(168,85,247,.45)';
+    if(v>=60) return 'rgba(255,159,67,.4)';
+    if(v>=50) return 'rgba(255,71,87,.3)';
+    return 'rgba(90,104,128,.2)';
+}
+function hmText(v) { return v>=75?'#0e1219':'#e8edf5'; }
+
+function renderHeatmap() {
+    const t=document.getElementById('hmTable');
+    let h='<thead><tr><th class="hm-role">Агент</th>';
+    hmModels.forEach(m=>{
+        const ifColor = m.if >= 85 ? '#00ff94' : m.if >= 75 ? '#facc15' : '#ff6b81';
+        h+=`<th style="writing-mode:vertical-lr;transform:rotate(180deg);max-width:32px;font-size:.56em;padding:3px 1px;">
+            ${m.n}<br>
+            <span style="color:${m.p==='Groq'?'#ff6b81':m.p==='Both'?'#c084fc':m.p.includes('Open')||m.p.includes('OR')?'#e879f9':'var(--accent-cyan)'};font-size:.85em">${m.p}</span><br>
+            <span style="color:${ifColor};font-size:.9em;font-weight:700" title="Instruction Following score">IF:${m.if}</span>
+        </th>`;
+    });
+    h+='</tr></thead><tbody>';
+    const reIcons = {"L":"🟢","M":"🟡","H":"🔴"};
+    hmAgents.forEach(ag=>{
+        const mx=Math.max(...ag.s);
+        const reIcon = reIcons[ag.re] || '🟡';
+        h+=`<tr><td class="hm-r">${reIcon} ${ag.n}</td>`;
+        ag.s.forEach((s,j)=>{
+            const best=s===mx, cur=j===ag.c;
+            const ifNote = hmModels[j].if < 75 ? ' ⚠' : '';
+            h+=`<td style="background:${hmColor(s)};color:${hmText(s)}" class="${cur?'hm-cur':''}"
+                onmouseover="showTT(event,'${ag.n}','${hmModels[j].n} (${hmModels[j].p})',${s},${best},${cur},${hmModels[j].if})"
+                onmouseout="hideTT()">${s}${best?'<span class="hm-star">★</span>':''}${ifNote}</td>`;
+        });
+        h+='</tr>';
+    });
+    t.innerHTML=h+'</tbody>';
+}
+
+function showTT(e,agent,model,score,best,cur,ifScore) {
+    const b=document.getElementById('ttBox'),o=document.getElementById('ttOverlay');
+    const ifColor = ifScore >= 85 ? '#00ff94' : ifScore >= 75 ? '#facc15' : '#ff6b81';
+    const ifLabel = ifScore >= 85 ? 'Отлично' : ifScore >= 75 ? 'Средне' : 'Слабо';
+    b.innerHTML=`<h4>${model}</h4><p><strong>Агент:</strong> ${agent}<br><strong>Итоговая оценка:</strong> ${score}/100<br>
+        <strong>Instruction Following:</strong> <span style="color:${ifColor};font-weight:700">${ifScore}/100 (${ifLabel})</span><br>
+        <span style="font-size:.9em;color:var(--text-muted)">Оценка = бенчмарк × IF-множитель</span><br>
+        ${ifScore < 75 ? '<span style="color:#ff6b81">⚠ Модель плохо следует промпту и роли — оценка снижена</span><br>' : ''}
+        ${best?'★ <strong>Лучший выбор</strong><br>':''}${cur?'📌 <strong>Текущий выбор</strong>':''}</p>`;
+    const r=e.target.getBoundingClientRect();
+    b.style.left=Math.min(r.left,window.innerWidth-320)+'px';
+    b.style.top=(r.bottom+6)+'px';
+    o.classList.add('show');
+}
+function hideTT() { document.getElementById('ttOverlay').classList.remove('show'); }
+
+function renderRecs() {
+    const g=document.getElementById('recGrid');
+    g.innerHTML=recs.map((r,i)=>`
+        <div class="rec-card ${r.imp==='critical'?'glow':''} selected" data-idx="${i}" style="animation:fadeUp .4s ${i*.06}s ease-out both">
+            <input type="checkbox" class="rec-check" checked onchange="updateRecSelection()" data-idx="${i}">
+            <div class="rec-hdr">
+                <div class="rec-agent">${r.a}</div>
+                <span class="impact-badge ${r.imp}">${r.imp.toUpperCase()}</span>
+            </div>
+            <div class="swap-vis">
+                <span class="swap-from">${r.from} <span class="prov-tag ${r.fromP.toLowerCase()}">${r.fromP}</span></span>
+                <span class="swap-arrow">→</span>
+                <span class="swap-to">${r.to} <span class="prov-tag ${r.toP.toLowerCase()}">${r.toP}</span></span>
+            </div>
+            <div class="rec-metrics">
+                <div class="rec-m"><div class="rec-m-label">Качество</div><div class="rec-m-val pos">${r.q}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Скорость</div><div class="rec-m-val ${r.sp.includes('10')? 'pos':'neu'}">${r.sp}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Контекст</div><div class="rec-m-val ${r.ctx.includes('→')?'pos':'neu'}">${r.ctx}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Провайдер</div><div class="rec-m-val" style="font-size:.7em;color:var(--text-secondary)">${r.prov}</div></div>
+            </div>
+            <div class="rec-reason">${r.r}</div>
+        </div>`).join('');
+    updateRecSelection();
+}
+
+// ===== EXPORT SYSTEM =====
+
+function updateRecSelection() {
+    const checks = document.querySelectorAll('.rec-check');
+    let count = 0;
+    checks.forEach(ch => {
+        const card = ch.closest('.rec-card');
+        if (ch.checked) { card.classList.add('selected'); count++; }
+        else { card.classList.remove('selected'); }
+    });
+    document.getElementById('selectedCount').textContent = count + ' из ' + recs.length + ' выбрано';
+    document.getElementById('exportBtn').style.opacity = count > 0 ? '1' : '.4';
+}
+
+function toggleAllRecs() {
+    const checks = document.querySelectorAll('.rec-check');
+    const allChecked = [...checks].every(c => c.checked);
+    checks.forEach(c => c.checked = !allChecked);
+    document.getElementById('selectAllBtn').textContent = allChecked ? '☑ Выбрать все' : '☐ Снять все';
+    updateRecSelection();
+}
+
+function buildExportJSON() {
+    const checks = document.querySelectorAll('.rec-check');
+    const selected = [];
+    checks.forEach(ch => {
+        if (!ch.checked) return;
+        const idx = parseInt(ch.dataset.idx);
+        const r = recs[idx];
+
+        // Build the model string for capability-index.yaml
+        const modelMap = {
+            'nemotron-3-super': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'nemotron-3-super + Groq burst': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'qwen3-coder:480b': 'ollama-cloud/qwen3-coder:480b',
+            'glm-5 (перенастроить)': 'ollama-cloud/glm-5',
+        };
+        const fallbackMap = {
+        };
+
+        const entry = {
+            agent: r.a,
+            action: 'update_model',
+            current_model: r.from,
+            current_provider: r.fromP.toLowerCase(),
+            recommended_model: modelMap[r.to] || 'ollama-cloud/' + r.to,
+            recommended_provider: r.toP.toLowerCase(),
+            impact: r.imp,
+            expected_improvement: {
+                quality: r.q,
+                speed: r.sp,
+                context_window: r.ctx
+            },
+            rationale: r.r.replace(/<[^>]*>/g, ''),
+        };
+
+        if (fallbackMap[r.to]) {
+            entry.fallback_model = fallbackMap[r.to];
+            entry.fallback_strategy = 'speed-burst';
+            entry.fallback_note = 'Use Groq for low-volume speed-critical calls; primary on Ollama Cloud';
+        }
+
+        selected.push(entry);
+    });
+
+    return {
+        "$schema": "https://app.kilo.ai/agent-recommendations.json",
+        "generated": new Date().toISOString(),
+        "source": "APAW Agent Model Research v3",
+        "target_file": ".kilo/capability-index.yaml",
+        "total_recommendations": selected.length,
+        "summary": {
+            "avg_quality_improvement": "+18%",
+            "providers_used": ["ollama-cloud", "groq", "openrouter"],
+            "key_models": [
+                "nemotron-3-super-120b-a12b (1M ctx, SWE-bench 60.5%, RULER@1M 91.75%)",
+                "qwen3-coder:480b (SWE-bench 66.5%, best open-source coding)",
+                "qwen3.6-plus (FREE OpenRouter, 1M ctx, SWE-bench 78.8%)",
+                "gemma4:31b (Intelligence Index 39, thinking, vision)",
+                "minimax-m2.5 (SWE-bench 80.2%, best coding overall)"
+            ]
+        },
+        "recommendations": selected,
+        "capability_index_patch": selected.map(s => ({
+            agent: s.agent,
+            set: { model: s.recommended_model }
+        }))
+    };
+}
+
+function exportJSON() {
+    const checks = document.querySelectorAll('.rec-check:checked');
+    if (checks.length === 0) return;
+    const json = buildExportJSON();
+    const formatted = JSON.stringify(json, null, 2);
+    document.getElementById('jsonPreview').textContent = formatted;
+    document.getElementById('jsonModal').style.display = 'flex';
+}
+
+function closeModal() {
+    document.getElementById('jsonModal').style.display = 'none';
+}
+
+function copyJSON() {
+    const text = document.getElementById('jsonPreview').textContent;
+    navigator.clipboard.writeText(text).then(() => {
+        const btn = document.getElementById('copyBtn');
+        btn.textContent = '✅ Скопировано!';
+        setTimeout(() => btn.textContent = '📋 Копировать', 2000);
+    });
+}
+
+function downloadJSON() {
+    const json = buildExportJSON();
+    const blob = new Blob([JSON.stringify(json, null, 2)], { type: 'application/json' });
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement('a');
+    a.href = url;
+    a.download = 'agent-model-recommendations.json';
+    document.body.appendChild(a);
+    a.click();
+    document.body.removeChild(a);
+    URL.revokeObjectURL(url);
+}
+
+function renderImpact() {
+    const d=document.getElementById('impactDetails');
+    d.innerHTML=impactData.map(x=>`
+        <div style="display:flex;align-items:center;gap:14px;padding:10px 0;border-bottom:1px solid var(--border)">
+            <div style="flex:1"><div style="font-weight:600;font-size:.88em">${x.cat}</div><div style="font-size:.74em;color:var(--text-muted);margin-top:2px">${x.n}</div></div>
+            <div style="display:flex;align-items:center;gap:10px">
+                <span style="font-family:'JetBrains Mono',monospace;color:var(--text-muted);font-size:.82em">${x.b}</span>
+                <span style="color:${x.d>0?'var(--accent-green)':'var(--text-muted)'};font-size:1.1em">→</span>
+                <span style="font-family:'JetBrains Mono',monospace;color:${x.d>0?'var(--accent-green)':'var(--text-secondary)'};font-size:.9em;font-weight:700">${x.a}</span>
+                <span style="font-family:'JetBrains Mono',monospace;font-size:.78em;padding:2px 6px;border-radius:3px;
+                    background:${x.d>0?'rgba(0,255,148,.08)':'rgba(90,104,128,.08)'};color:${x.d>0?'var(--accent-green)':'var(--text-muted)'}">
+                    ${x.d>0?'+'+x.d:'=0'}</span>
+            </div>
+        </div>`).join('');
+
+    // Render APAW vs TOP benchmark table
+    renderBenchmarkComparison();
+}
+
+function renderBenchmarkComparison() {
+    // Expanded benchmarks: 9 dimensions covering all APAW roles
+    const benchmarks = [
+        {name:'SWE-V',full:'SWE-Bench Verified',desc:'GitHub issue resolution (500 tasks)',role:'lead-dev, backend, fixer'},
+        {name:'SWE-P',full:'SWE-Bench Pro',desc:'Multi-lang, decontaminated (1865 tasks)',role:'all coding agents'},
+        {name:'T-Bench',full:'Terminal-Bench 2.0',desc:'CLI/shell multi-step tasks',role:'devops, planner, orchestrator'},
+        {name:'LCB',full:'LiveCodeBench',desc:'Code gen from specs (held-out)',role:'sdet, go-dev, python-dev'},
+        {name:'GPQA',full:'GPQA Diamond',desc:'PhD-level reasoning',role:'system-analyst, planner'},
+        {name:'BComp',full:'BrowseComp',desc:'Web research & synthesis',role:'browser-auto, capability-analyst'},
+        {name:'HLE',full:'Humanity Last Exam',desc:'Frontier knowledge (with tools)',role:'agent-architect, evaluator'},
+        {name:'Ctx',full:'Context Window',desc:'Max tokens in one pass',role:'history-miner, memory-mgr'},
+        {name:'$/M',full:'Cost per 1M input',desc:'API pricing',role:'all agents (ROI)'},
+    ];
+    
+    const models = [
+        // === TOP CLOSED-SOURCE (April 2026 leaders) ===
+        {name:'Claude Opus 4.7',type:'closed',org:'Anthropic',
+         scores:[87.6,64.3,69.4,null,94.2,79.3,53.0,'1M','$5'],color:'#c084fc',
+         note:'#1 апрель 2026'},
+        {name:'GPT-5.5',type:'closed',org:'OpenAI',
+         scores:[null,58.6,82.7,null,null,83.4,57.2,'1M','$5'],color:'#ff6b81',
+         note:'Новейший, Terminal #1'},
+        {name:'GPT-5.4',type:'closed',org:'OpenAI',
+         scores:[78.2,59.1,75.1,null,94.4,82.7,58.7,'200K','$2.50'],color:'#ff6b81',
+         note:'Reasoning, math'},
+        {name:'Gemini 3.1 Pro',type:'closed',org:'Google',
+         scores:[80.6,46.1,68.5,null,94.3,85.9,51.4,'2M','$2'],color:'#facc15',
+         note:'ARC-AGI 77.1%, дешёвый'},
+        {name:'Claude Sonnet 4.6',type:'closed',org:'Anthropic',
+         scores:[79.6,null,null,null,null,null,null,'200K','$3'],color:'#c084fc',
+         note:'5× дешевле Opus'},
+        {name:'GPT-5.3-Codex',type:'closed',org:'OpenAI',
+         scores:[85.0,57.0,77.3,null,null,null,null,'200K','$6'],color:'#ff6b81',
+         note:'Coding specialist'},
+         
+        // === APAW PIPELINE MODELS ===
+        {name:'Kimi K2.6',type:'apaw',org:'APAW',
+         scores:[80.2,58.6,66.7,87.2,null,83.2,54.0,'256K','$0.95'],color:'#00ff94',
+         note:'devops, browser, architect (3)'},
+        {name:'GLM-5.1',type:'apaw',org:'APAW',
+         scores:[null,58.4,63.5,null,86.2,68.7,null,'200K','~$0.50'],color:'#00ff94',
+         note:'12 agents! orchestrator, eval...'},
+        {name:'V4-Pro Max',type:'apaw',org:'APAW',
+         scores:[80.6,55.4,67.9,93.5,90.1,83.4,48.2,'1M','$0.42'],color:'#00d4ff',
+         note:'planner, go-dev (рек.)'},
+        {name:'Qwen3-Coder 480B',type:'apaw',org:'APAW',
+         scores:[66.5,null,null,null,null,null,null,'256K','~$0.50'],color:'#00ff94',
+         note:'7 coding agents'},
+        {name:'MiniMax M2.5',type:'apaw',org:'APAW',
+         scores:[80.2,51.3,null,null,null,76.3,null,'204K','$0.15'],color:'#00ff94',
+         note:'frontend, skeptic, fixer (3)'},
+        {name:'Nemotron Super',type:'apaw',org:'APAW',
+         scores:[60.5,null,null,null,null,null,null,'1M','~$0.40'],color:'#00ff94',
+         note:'6 agents (memory, history)'},
+    ];
+
+    const t = document.getElementById('benchTable');
+    let h = '<thead><tr><th style="text-align:left;padding:8px 6px;border-bottom:2px solid var(--border);font-size:.85em">Модель</th>';
+    benchmarks.forEach(b => {
+        h += '<th style="padding:8px 3px;border-bottom:2px solid var(--border);font-size:.68em;max-width:60px" title="'+b.full+': '+b.desc+'\nРоли: '+b.role+'">'+b.name+'</th>';
+    });
+    h += '</tr></thead><tbody>';
+
+    // Calculate APAW best per benchmark
+    const apawBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='apaw').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+    
+    // Calculate closed best per benchmark
+    const closedBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='closed').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+
+    models.forEach((m,mi) => {
+        if(mi === 6) h += '<tr><td colspan="'+(benchmarks.length+1)+'" style="padding:5px;background:rgba(0,212,255,.06);font-weight:700;font-size:.8em;color:var(--accent-cyan);text-align:center">— APAW Pipeline (open-source, $0.15–$0.95/M) —</td></tr>';
+        
+        h += '<tr style="'+(m.type==='apaw'?'background:rgba(0,255,148,.02)':'')+'">';
+        h += '<td style="padding:6px;border-bottom:1px solid var(--border);white-space:nowrap"><span style="font-weight:600;color:'+m.color+';font-size:.88em">'+m.name+'</span>';
+        h += '<div style="font-size:.65em;color:var(--text-muted)">'+m.note+'</div></td>';
+
+        m.scores.forEach((s,si) => {
+            let val, cellColor = 'var(--text-secondary)', bg = 'transparent';
+            
+            if(s === null) { val = '—'; cellColor = 'rgba(90,104,128,.4)'; }
+            else if(typeof s === 'string' && s.startsWith('$')) {
+                val = s;
+                if(m.type === 'apaw') { cellColor = '#00ff94'; bg = 'rgba(0,255,148,.06)'; }
+            }
+            else if(typeof s === 'string') { val = s; }
+            else {
+                val = s.toFixed(1);
+                if(m.type === 'apaw' && si < 7) {
+                    const cb = closedBest[si];
+                    if(cb > 0) {
+                        const diff = s - cb;
+                        if(diff >= 0) { cellColor='#00ff94'; bg='rgba(0,255,148,.1)'; val+=' 🟢'; }
+                        else if(diff > -5) { cellColor='#facc15'; bg='rgba(250,204,21,.06)'; val+=' 🟡'; }
+                        else { cellColor='#ff6b81'; bg='rgba(255,107,129,.06)'; val+=' 🔴'; }
+                    }
+                }
+                if(m.type === 'closed' && si < 7) {
+                    // Highlight if APAW beats this closed model
+                    const ab = apawBest[si];
+                    if(ab > 0 && s < ab) { bg='rgba(255,107,129,.04)'; }
+                }
+            }
+            h += '<td style="padding:5px 3px;border-bottom:1px solid var(--border);text-align:center;color:'+cellColor+';background:'+bg+';font-size:.78em">'+val+'</td>';
+        });
+        h += '</tr>';
+    });
+
+    // === Summary row: APAW best vs Closed best ===
+    h += '<tr style="background:rgba(0,212,255,.05)"><td style="padding:8px;font-weight:700;color:var(--accent-cyan);font-size:.85em">APAW лучший</td>';
+    benchmarks.forEach((b,i) => {
+        if(i < 7) {
+            const ab = apawBest[i], cb = closedBest[i];
+            if(ab === 0) { h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--text-muted)">—</td>'; return; }
+            const diff = ab - cb;
+            const icon = diff >= 0 ? '🟢' : diff > -5 ? '🟡' : '🔴';
+            const pct = cb > 0 ? ((ab/cb)*100-100).toFixed(1) : '?';
+            const sign = diff >= 0 ? '+' : '';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.78em"><span style="color:'+(diff>=0?'#00ff94':diff>-5?'#facc15':'#ff6b81')+'">'+ab.toFixed(1)+'</span><div style="font-size:.7em;color:var(--text-muted)">'+sign+diff.toFixed(1)+' '+icon+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">1M ✅</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">10-33× 🟢</td>';
+        }
+    });
+    h += '</tr>';
+
+    // === Role-based average row ===
+    h += '<tr style="background:rgba(0,255,148,.04)"><td style="padding:8px;font-weight:700;color:var(--accent-green);font-size:.82em">Средняя по ролям APAW*</td>';
+    // Calculate weighted average across all roles
+    const roleAvg = [78.2, 55.8, 65.7, 90.4, 88.2, 78.4, 51.1]; // pre-calculated across all 36 agents
+    const closedAvg = [82.2, 57.8, 74.6, null, 94.3, 83.4, 54.8];
+    roleAvg.forEach((ra,i) => {
+        if(i < 7 && ra > 0) {
+            const ca = closedAvg[i];
+            if(!ca) { h += '<td style="padding:6px 3px;text-align:center;font-size:.82em;color:var(--accent-green);font-weight:700">'+ra.toFixed(1)+'</td>'; return; }
+            const diff = ra - ca;
+            const col = diff >= 0 ? '#00ff94' : diff > -8 ? '#facc15' : '#ff6b81';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.82em"><span style="color:'+col+'">'+ra.toFixed(1)+'</span><div style="font-size:.65em;color:var(--text-muted)">vs '+ca.toFixed(1)+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">573K avg</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">$0.49 avg</td>';
+        }
+    });
+    h += '</tr>';
+
+    t.innerHTML = h + '</tbody>';
+}
+
+
+function drawChart() {
+    const c=document.getElementById('impactCanvas');
+    if(!c || !c.offsetParent) return; // skip if hidden
+    const ctx=c.getContext('2d');
+    const dpr = window.devicePixelRatio || 1;
+    const cssW = c.parentElement.clientWidth - 36;
+    const cssH = 340;
+    c.width = cssW * dpr;
+    c.height = cssH * dpr;
+    c.style.width = cssW + 'px';
+    c.style.height = cssH + 'px';
+    ctx.scale(dpr, dpr);
+    ctx.clearRect(0,0,cssW,cssH);
+
+    const data = impactData;
+    if(!data.length) return;
+    const barW = Math.min(38, (cssW - 180) / data.length / 2 - 4);
+    const cL = 48, cB = cssH - 60, cH = cB - 20, mx = 100;
+
+    // Grid lines
+    ctx.strokeStyle='rgba(30,39,54,.7)'; ctx.lineWidth=1;
+    for(let i=0;i<=5;i++){
+        const y=cB-(cH*(i*20)/mx);
+        ctx.beginPath();ctx.moveTo(cL,y);ctx.lineTo(cssW-16,y);ctx.stroke();
+        ctx.fillStyle='#5a6880';ctx.font='10px JetBrains Mono,monospace';ctx.textAlign='right';ctx.fillText(i*20,cL-6,y+3);
+    }
+
+    // Bars
+    data.forEach((d,i)=>{
+        const x = cL + 28 + i * ((cssW - cL - 40) / data.length);
+        // Before bar (red)
+        const h1 = (d.b/mx)*cH;
+        ctx.fillStyle='rgba(255,71,87,.4)';
+        ctx.fillRect(x, cB-h1, barW, h1);
+        // After bar (green or grey)
+        const h2 = (d.a/mx)*cH;
+        ctx.fillStyle = d.d > 0 ? 'rgba(0,255,148,.55)' : 'rgba(136,150,170,.35)';
+        ctx.fillRect(x+barW+3, cB-h2, barW, h2);
+        // Delta label
+        if(d.d > 0){
+            ctx.fillStyle='#00ff94';ctx.font='bold 10px JetBrains Mono,monospace';
+            ctx.textAlign='center';ctx.fillText('+'+d.d, x+barW+1, cB-Math.max(h1,h2)-6);
+        }
+        // Category label (rotated)
+        ctx.save();
+        ctx.translate(x+barW, cB+10);
+        ctx.rotate(-0.4);
+        ctx.fillStyle='#8896aa';ctx.font='8px Outfit,sans-serif';ctx.textAlign='left';
+        const label = d.cat.replace(/\s*\(.*?\)/g,'').substring(0,22);
+        ctx.fillText(label, 0, 0);
+        ctx.restore();
+    });
+
+    // Legend
+    ctx.fillStyle='rgba(255,71,87,.4)';ctx.fillRect(cssW-180,8,12,12);
+    ctx.fillStyle='#8896aa';ctx.font='11px Outfit,sans-serif';ctx.textAlign='left';ctx.fillText('Текущий score',cssW-162,18);
+    ctx.fillStyle='rgba(0,255,148,.55)';ctx.fillRect(cssW-180,26,12,12);
+    ctx.fillText('После замены',cssW-162,36);
+}
+
+// ======================= INIT =======================
+document.addEventListener('DOMContentLoaded',()=>{
+    renderCfg(); renderGroq(); renderModels(); renderHeatmap(); renderRecs(); renderImpact();
+});
+window.addEventListener('resize',()=>{ if(document.getElementById('tab-impact').classList.contains('active')) drawChart(); });
+</script>
+</body>
+</html>
diff --git a/agent-evolution/docs/bidirectional-data-flow.md b/agent-evolution/docs/bidirectional-data-flow.md
new file mode 100644
index 0000000..83c3a5b
--- /dev/null
+++ b/agent-evolution/docs/bidirectional-data-flow.md
@@ -0,0 +1,504 @@
+# Двунаправленный поток данных APAW Agent Model Research
+
+Этот документ описывает архитектуру системы, которая автоматизирует исследование моделей AI для агентов APAW и синхронизирует данные между визуальной панелью, конфигурационными файлами и пайплайном эволюции.
+
+## Цель
+
+Изначально все данные исследования моделей были захардкожены в HTML-файле `apaw_agent_model_research_v3.html` (1168 строк JavaScript). Двунаправленный поток делает эту систему:
+
+- **Машиночитаемой** — данные хранятся в JSON для автоматической обработки
+- **Записываемой** — изменения в конфигурации агентов обновляют JSON и перегенерируют дашборд
+- **Визуализированной** — любое изменение данных автоматически создаёт новый HTML
+
+## Архитектура данных
+
+### Файлы системы
+
+| Файл | Назначение | Формат | Обновляется |
+|------|-----------|--------|-------------|
+| `data/model-benchmarks.json` | Статические бенчмарки | JSON | `/research models`, вручную |
+| `data/model-research-latest.json` | Последнее исследование | JSON | `/evolution Step 0`, `/research models` |
+| `data/model-research.schema.json` | Схема валидации | JSON Schema | Вручную |
+| `data/model-benchmarks.schema.json` | Схема бенчмарков | JSON Schema | Вручную |
+| `scripts/build-research-dashboard.ts` | Генерация HTML | TypeScript/Bun | Вручную |
+| `scripts/sync-model-research.ts` | Применение изменений | TypeScript/Bun | Вручную |
+| `research-dashboard.template.html` | Шаблон дашборда | HTML+JS+CSS | Вручную |
+| `research-dashboard.html` | Готовый дашборд | HTML (standalone) | `build-research-dashboard.ts` |
+| `dist/research-dashboard-YYYY_MM_DD.html` | Архив | HTML | `build-research-dashboard.ts` |
+
+## Поток данных
+
+### Направление 1: HTML → JSON (Исследование → Бенчмарки)
+
+Источник: `apaw_agent_model_research_v3.html` (вручную исследованные данные)
+
+```
+apaw_agent_model_research_v3.html
+       │ hardcoded JS arrays:
+       │   cfg[]     — текущие конфиги агентов
+       │   ollamaModels[] — характеристики моделей
+       │   hmAgents[] — матрица очков
+       │   recs[] — рекомендации
+       │   impactData[] — дельта изменений
+       │   groqModels[] — лимиты Groq
+       ↓
+agent-evolution/data/model-benchmarks.json
+       ├─ models[] — 15 моделей, бенчмарки, IF-оценки
+       ├─ agent_model_scores[] — 33 агента × 11 моделей
+       ├─ agent_current_config[] — 36 текущих назначений
+       ├─ recommendations[] — 11 рекомендуемых замен
+       ├─ groq_models[] — 5 моделей Groq с лимитами
+       ├─ impact_data[] — before/after
+       └─ benchmark_comparison — сравнение с закрытыми моделями
+```
+
+**Как обновлять**: один раз данные извлечены из HTML. Дальнейшие обновления:
+- Автоматически: `/research models` → `model-research-latest.json` → `model-benchmarks.json`
+- Вручную: редактировать `model-benchmarks.json`, обновить `metadata.generated`
+
+### Направление 2: JSON → Конфиг → HTML (Применение → Визуализация)
+
+```
+[/research models] OR [/evolution Step 0]
+       ↓
+model-research-latest.json
+       │ validates against:
+       ↓ model-research.schema.json
+bun run agent-evolution/scripts/sync-model-research.ts
+       ├─ обновляет .kilo/capability-index.yaml (model поля)
+       ├─ обновляет kilo-meta.json (source of truth)
+       ├─ обновляет kilo.jsonc (agent config)
+       ├─ обновляет agent-evolution/data/agent-versions.json (история)
+       ├─ обновляет .kilo/agents/*.md frontmatter (через sync-agents.js --fix)
+       └─ rebuilds dashboard (build-research-dashboard.ts)
+       ↓
+bun run agent-evolution/scripts/build-research-dashboard.ts
+       ├─ читает model-benchmarks.json
+       ├─ инжектирует в research-dashboard.template.html
+       ├─ записывает research-dashboard.html
+       └─ копирует dist/research-dashboard-YYYY_MM_DD.html
+       ↓
+[/research models] ← цикл продолжается
+```
+
+## Структура model-benchmarks.json
+
+### Верхний уровень
+
+```json
+{
+  "version": "1.0.0",
+  "generated": "2026-04-27T17:44:44.000Z",
+  "source": "apaw_agent_model_research_v3.html",
+  "total_agents": 36,
+  "total_models_tracked": 11,
+  "providers": ["ollama", "ollama-cloud", "openrouter", "groq"],
+  "models": [...],
+  "groq_models": [...],
+  "agent_model_scores": [...],
+  "if_scores": {...},
+  "agent_current_config": [...],
+  "recommendations": [...],
+  "impact_data": [...],
+  "benchmark_comparison": {...}
+}
+```
+
+### Модель
+
+```json
+{
+  "id": "ollama-cloud/qwen3-coder:480b",
+  "name": "Qwen3-Coder 480B",
+  "organization": "Qwen",
+  "parameters": "480B/35B active",
+  "context_window": "256K\u21921M",
+  "swe_bench": 66.5,
+  "swe_bench_pro": null,
+  "terminal_bench": null,
+  "live_codebench": null,
+  "gpqa": null,
+  "hle": null,
+  "browse_comp": null,
+  "if_score": 88,
+  "categories": ["coding", "agent"],
+  "tags": ["coding", "agent", "tools"],
+  "provider": "ollama",
+  "free": false,
+  "cost_per_1m_input": "~$0.50",
+  "description": "SOTA open-source \u043a\u043e\u0434\u0438\u043d\u0433. \u0421\u0440\u0430\u0432\u043d\u0438\u043c \u0441 Claude Sonnet 4.",
+  "availability": null,
+  "speed_tps": null
+}
+```
+
+### Рекомендация
+
+```json
+{
+  "agent": "planner",
+  "action": "update_model",
+  "current_model": "nemotron-3-super",
+  "current_provider": "Ollama",
+  "recommended_model": "deepseek-v4-pro-max",
+  "recommended_provider": "Ollama Cloud",
+  "impact": "high",
+  "score_before": 80,
+  "score_after": 88,
+  "score_delta": 8,
+  "expected_improvement": {
+    "quality": "+10%",
+    "speed": "~1x",
+    "context_window": "1M"
+  },
+  "rationale": "\u2605 matri\u0446\u044b: V4-Pro=88(\u043b\u0443\u0447\u0448\u0438\u0439!)..."
+}
+```
+
+### Очки агента
+
+```json
+{
+  "agent": "lead-developer",
+  "current_model_index": 0,
+  "reasoning_effort": "M",
+  "scores": {
+    "ollama-cloud/qwen3-coder:480b": 92,
+    "ollama-cloud/minimax-m2.5": 86,
+    "ollama-cloud/minimax-m2.7": 82,
+    "ollama-cloud/nemotron-3-super": 70,
+    "ollama-cloud/glm-5": 68,
+    "ollama-cloud/glm-5.1": 75,
+    "ollama-cloud/deepseek-v4-pro-max": 88,
+    "ollama-cloud/qwen3.5-122b": 66,
+    "ollama-cloud/qwen3-coder-next": 80,
+    "openrouter/qwen/qwen3.6-plus:free": 88,
+    "ollama-cloud/kimi-k2.6:cloud": 90
+  }
+}
+```
+
+## Формула IF-ажастмента
+
+Оценка агента с учётом способности модели следовать инструкциям:
+
+```
+IF-adjusted_score = raw_score × (0.7 + 0.3 × IF/100)
+
+Где:
+  raw_score — бенчмарк оценка пары агент×модель (0-100)
+  IF — instruction following score модели (0-100)
+
+Примеры:
+  IF=100 → score × 1.00 (без изменений)
+  IF=90  → score × 0.97
+  IF=78  → score × 0.93
+  IF=50  → score × 0.85
+  IF=0   → score × 0.70
+
+Чем ниже IF, тем сильнее штраф — модель плохо следует промпту и роли.
+```
+
+## Скрипты системы
+
+### build-research-dashboard.ts
+
+**Вход**: `model-benchmarks.json` + `research-dashboard.template.html`
+**Выход**: `research-dashboard.html` + `dist/dashboard-YYYY_MM_DD.html`
+
+```bash
+bun run agent-evolution/scripts/build-research-dashboard.ts              # однократная сборка
+bun run agent-evolution/scripts/build-research-dashboard.ts --watch       # watch-режим
+bun run agent-evolution/scripts/build-research-dashboard.ts --template custom.html
+```
+
+Процесс:
+1. Читает JSON, валидирует наличие полей
+2. Читает шаблон, ищет placeholder `// BENCHMARK_DATA_PLACEHOLDER`
+3. Заменяет `const EMBEDDED_DATA = {};` на полный JSON с данными
+4. Обновляет `<title>` с датой генерации
+5. Пишет `research-dashboard.html` и архивную копию
+
+### sync-model-research.ts
+
+**Вход**: `model-research-latest.json`
+**Действия**:
+
+```bash
+# Предпросмотр
+bun run agent-evolution/scripts/sync-model-research.ts --dry-run
+
+# Применение всех рекомендаций
+bun run agent-evolution/scripts/sync-model-research.ts
+
+# Только для одного агента
+bun run agent-evolution/scripts/sync-model-research.ts --agent planner
+```
+
+Для каждой рекомендации (`action: "update_model"`, `applied: false`):
+1. Находит блок агента в `capability-index.yaml`, заменяет `model:`
+2. Обновляет `kilo-meta.json` (source of truth)
+3. Обновляет `kilo.jsonc` (через regex, требует ручной проверки)
+4. Добавляет запись в `agent-versions.json` history
+5. Запускает `node scripts/sync-agents.js --fix` → обновляет .md frontmatter
+6. Запускает `node scripts/sync-agents.js --check` → проверка консистентности
+7. Пересобирает дашборд через `build-research-dashboard.ts`
+
+## Интеграция в пайплайн
+
+### /research models
+
+```
+1. Загрузить текущие данные из model-benchmarks.json
+2. Если stale (>7 дней) или --force:
+   a. Fetch моделей с Ollama Cloud, OpenRouter, Groq
+   b. Compute IF scores для каждой модели
+   c. Score каждую модель против каждого агента
+3. Сгенерировать рекомендации (gap > 5)
+4. Записать model-research-latest.json
+5. Валидировать против model-research.schema.json
+6. Обновить model-benchmarks.json (если данные изменились)
+7. Пересобрать дашборд
+```
+
+### /evolution (полный цикл)
+
+```
+Step 0: Model Research
+  ├─ Проверить staleness model-benchmarks.json
+  ├─ Если stale → @capability-analyst исследует модели
+  ├─ Загрузить heatmap scores
+  └─ Определить агентов с mismatch (gap > 5)
+
+Step 1: Judge
+  └─ @pipeline-judge → fitness score
+
+Step 2: Decide
+  ├─ fitness >= 0.85 → выход
+  ├─ fitness >= 0.70 → @prompt-optimizer (minor)
+  └─ fitness < 0.70 → @prompt-optimizer (major) + apply model recs
+
+Step 3: Re-test
+  └─ Перезапуск с обновлёнными промптами/моделями
+
+Step 4: Log + Dashboard
+  ├─ Append fitness-history.jsonl
+  ├─ Apply рекомендации sync-model-research.ts
+  └─ Пересобрать дашборд build-research-dashboard.ts
+```
+
+### /evolution research
+
+```
+1. Прочитать текущую конфигурацию
+2. Исследовать модели (как /research models)
+3. Сгенерировать рекомендации
+4. Dry-run preview
+5. Применить при подтверждении
+6. Пересобрать дашборд
+```
+
+## Правила синхронизации
+
+Из `.kilo/rules/evolutionary-sync.md`:
+
+### Обязательный порядок
+
+```
+1. Обновить kilo-meta.json (source of truth)
+2. Обновить capability-index.yaml
+3. Запустить sync-agents.js --fix
+4. Ручная проверка kilo.jsonc (sync script не гарантирует)
+5. Запустить sync-agents.js --check
+6. Проверить agent-versions.json history
+7. Пересобрать дашборд
+8. Если любая проверка не прошла — НЕ коммитить
+```
+
+### Облачный суффикс
+
+При использовании `ollama-cloud/kimi-k2.6` ВСЕГДА с суффиксом `:cloud`:
+
+```yaml
+# Правильно
+model: "ollama-cloud/kimi-k2.6:cloud"
+
+# Неправильно — отсутствует суффикс
+model: "ollama-cloud/kimi-k2.6"
+```
+
+## Чеклист применения изменений
+
+```
+□ Исследование: /research models завершено
+□ Валидация: model-research-latest.json проходит schema check
+□ Dry-run: sync-model-research.ts --dry-run показывает ожидаемые изменения
+□ Применение: sync-model-research.ts выполнен без ошибок
+□ YAML: capability-index.yaml обновлены поля model
+□ Meta: kilo-meta.json соответствует
+□ kilo.jsonc: модели обновлены (ручная проверка)
+□ История: agent-versions.json записи добавлены
+□ Sync: sync-agents.js --fix обновил все .md файлы
+□ Check: sync-agents.js --check проходит
+□ Старые модели: grep не находит предыдущие model IDs
+□ Суффикс: kimi-k2.6:cloud (с :cloud)
+□ Дашборд: build-research-dashboard.ts сгенерировал свежий HTML
+□ Открыть: research-dashboard.html показывает актуальные данные
+□ Гит: все изменения add и commit
+```
+
+## Устранение неполадок
+
+| Проблема | Диагностика | Решение |
+|----------|------------|---------|
+| Дашборд пустой | Проверить placeholder в template.html | Пересобрать: `bun run build-research-dashboard.ts` |
+| Schema validation fails | Сравнить JSON со схемой | Проверить model-research.schema.json актуальность |
+| sync-agents.js check fails | Model mismatch в конфигах | Запустить `--fix`, затем `--check`; ручная проверка kilo.jsonc |
+| Heatmap пустой | agent_model_scores отсутствует | Обновить бенчмарки через `/research models` |
+| Рекомендации не отображаются | Empty recs array | Запустить research для генерации новых рекомендаций |
+| Старые данные | metadata.generated > 7 дней | Обновить бенчмарки |
+| sync-model-research.ts падает | Файл не найден | Проверить пути, запустить из корня проекта |
+
+## Пример полного цикла
+
+### 1. Исследование моделей
+
+```bash
+$ /research models
+
+## Research: model optimization
+
+### Models Analyzed
+- Ollama Cloud: 20 models
+- OpenRouter Free: 3 models
+- Groq Free: 5 models
+
+### Key Findings
+- DeepSeek V4-Pro Max доступен (SWE-V 80.6, IF:88)
+- Kimi K2.6 IF=91 (лучший для orchestration)
+- Nemotron 3 Super IF=78 — слаб для prompt-heavy ролей
+- Qwen 3.6 Plus FREE остаётся лучшим IF/cost (91, $0)
+
+### Recommendations Generated
+- 11 model swap recommendations
+- 4 high, 3 medium, 4 low
+- Средний expected improvement: +12 points
+
+### Files Updated
+- agent-evolution/data/model-research-latest.json
+- agent-evolution/data/model-benchmarks.json (refreshed)
+- agent-evolution/dist/research-dashboard-2026_04_27.html (archive)
+```
+
+### 2. Валидация schema
+
+```bash
+$ node -e "
+const Ajv = require('ajv');
+const ajv = new Ajv();
+const schema = JSON.parse(require('fs').readFileSync('agent-evolution/data/model-research.schema.json','utf8'));
+const data = JSON.parse(require('fs').readFileSync('agent-evolution/data/model-research-latest.json','utf8'));
+const valid = ajv.validate(schema, data);
+console.log(valid ? 'VALID' : 'INVALID');
+if (!valid) console.log(JSON.stringify(ajv.errors, null, 2));
+"
+VALID
+```
+
+### 3. Dry-run
+
+```bash
+$ bun run agent-evolution/scripts/sync-model-research.ts --dry-run
+
+=== SYNC PREVIEW (dry-run) ===
+3 agents would be updated:
+
+planner
+  FROM: nemotron-3-super (Ollama)
+  TO:   deepseek-v4-pro-max (Ollama Cloud)
+  DELTA: +8 (80 → 88)
+  IMPACT: high
+
+go-developer
+  FROM: qwen3-coder:480b (Ollama)
+  TO:   deepseek-v4-pro-max (Ollama Cloud)
+  DELTA: +3 (85 → 88)
+  IMPACT: medium
+
+[built-in] debug
+  FROM: glm-5.1 (Ollama)
+  TO:   kimi-k2.6:cloud (Ollama Cloud)
+  DELTA: +2 (88 → 90)
+  IMPACT: high
+
+Files to modify: capability-index.yaml, kilo-meta.json, kilo.jsonc, agent-versions.json
+```
+
+### 4. Применение
+
+```bash
+$ bun run agent-evolution/scripts/sync-model-research.ts
+
+✅ capability-index.yaml updated (3 agents)
+✅ kilo-meta.json updated  
+✅ kilo.jsonc updated
+✅ agent-versions.json history updated (3 entries)
+✅ sync-agents.js --fix completed
+✅ sync-agents.js --check passed
+✅ Dashboard rebuilt: research-dashboard.html (106KB)
+```
+
+### 5. Проверка дашборда
+
+```bash
+$ start agent-evolution/research-dashboard.html
+
+# В браузере:
+# - Overview: 3 agents updated, 11 recommendations total
+# - Heatmap: V4-Pro Max column green for planner, go-developer
+# - Recommendations: 3 marked as applied with checkmarks
+# - Impact: +8 for planner shown in chart
+```
+
+### 6. Тест пайплайна
+
+```bash
+$ /evolve --issue 42
+
+## Pipeline Judgment: Issue #42
+
+**Fitness: 0.88/1.00** [PASS → improved from 0.82]
+
+| Metric | Value | Weight | Contribution |
+|--------|-------|--------|-------------|
+| Tests  | 96% (46/48) | 50% | 0.480 |
+| Gates  | 80% (4/5) | 25% | 0.200 |
+| Cost   | 38.4K tok / 245s | 25% | 0.198 |
+
+**Bottleneck:** none (all agents optimal)
+**Verdict:** PASS — fitness improved!
+
+✅ Logged to .kilo/logs/fitness-history.jsonl
+✅ Auto-rebuilt: agent-evolution/research-dashboard.html
+```
+
+## Периодичность обновления
+
+| Файл | Период | Триггер |
+|------|--------|---------|
+| model-benchmarks.json | Еженедельно (>7 дней = stale) | `/evolution Step 0` или `/research models` |
+| model-research-latest.json | Каждый research cycle | `/research models`, `/evolution research` |
+| research-dashboard.html | После каждого изменения | `sync-model-research.ts` или `build-research-dashboard.ts` |
+| dist/*.html | Архив | Каждая генерация |
+| agent-versions.json | При каждом изменении модели | `sync-model-research.ts` |
+
+## Связанные документы
+
+- `.kilo/commands/evolution.md` — команда /evolution
+- `.kilo/commands/research.md` — команда /research
+- `.kilo/shared/self-evolution.md` — протокол эволюции
+- `.kilo/rules/evolutionary-sync.md` — правила синхронизации
+- `.kilo/rules/agent-frontmatter-validation.md` — валидация YAML frontmatter
+- `agent-evolution/README.md` — обзор системы эволюции
+- `kilo-meta.json` — source of truth для моделей
+- `.kilo/capability-index.yaml` — маршрутизация и назначения
diff --git a/agent-evolution/ideas/apaw_agent_model_research_v3.html b/agent-evolution/ideas/apaw_agent_model_research_v3.html
new file mode 100644
index 0000000..5e9251e
--- /dev/null
+++ b/agent-evolution/ideas/apaw_agent_model_research_v3.html
@@ -0,0 +1,1168 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>APAW KiloCode — Agent Model Research v3 (Ollama + Groq + OpenRouter)</title>
+    <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@300;400;500;600;700&family=Outfit:wght@300;400;500;600;700;800;900&display=swap" rel="stylesheet">
+    <style>
+        :root {
+            --bg-deep: #080b12;
+            --bg-panel: #0e1219;
+            --bg-card: #141922;
+            --bg-card-hover: #1a2130;
+            --border: #1e2736;
+            --border-bright: #2a3650;
+            --text-primary: #e8edf5;
+            --text-secondary: #8896aa;
+            --text-muted: #5a6880;
+            --accent-cyan: #00d4ff;
+            --accent-green: #00ff94;
+            --accent-orange: #ff9f43;
+            --accent-red: #ff4757;
+            --accent-purple: #a855f7;
+            --accent-blue: #3b82f6;
+            --accent-yellow: #facc15;
+            --glow-cyan: rgba(0,212,255,0.15);
+            --glow-green: rgba(0,255,148,0.1);
+            --glow-orange: rgba(255,159,67,0.1);
+        }
+        * { margin:0; padding:0; box-sizing:border-box; }
+        body {
+            font-family:'Outfit',sans-serif;
+            background:var(--bg-deep);
+            color:var(--text-primary);
+            min-height:100vh;
+            overflow-x:hidden;
+        }
+        body::before {
+            content:'';
+            position:fixed; top:0; left:0; right:0; bottom:0;
+            background:
+                linear-gradient(90deg,rgba(0,212,255,0.02) 1px,transparent 1px),
+                linear-gradient(rgba(0,212,255,0.02) 1px,transparent 1px);
+            background-size:60px 60px;
+            animation:gp 8s ease-in-out infinite;
+            pointer-events:none; z-index:0;
+        }
+        @keyframes gp { 0%,100%{opacity:.3} 50%{opacity:.6} }
+        .container { max-width:1540px; margin:0 auto; padding:24px 16px; position:relative; z-index:1; }
+
+        /* HEADER */
+        .header { text-align:center; margin-bottom:32px; animation:fadeDown .7s ease-out; }
+        @keyframes fadeDown { from{opacity:0;transform:translateY(-20px)} to{opacity:1;transform:translateY(0)} }
+        .header h1 {
+            font-size:2.6em; font-weight:900;
+            background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green),var(--accent-purple));
+            -webkit-background-clip:text; -webkit-text-fill-color:transparent;
+            letter-spacing:-1px;
+        }
+        .header .sub { font-family:'JetBrains Mono',monospace; color:var(--text-muted); font-size:.8em; margin-top:6px; letter-spacing:1px; }
+
+        /* TABS */
+        .tabs { display:flex; gap:3px; background:var(--bg-panel); border:1px solid var(--border); border-radius:12px; padding:4px; margin-bottom:28px; overflow-x:auto; }
+        .tab-btn {
+            flex:1; min-width:100px; padding:10px 8px; background:none; border:none; color:var(--text-secondary);
+            font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600; border-radius:9px; cursor:pointer; transition:all .25s; white-space:nowrap;
+        }
+        .tab-btn:hover { color:var(--text-primary); background:var(--bg-card); }
+        .tab-btn.active { color:var(--bg-deep); background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); box-shadow:0 0 16px var(--glow-cyan); }
+        .tab-panel { display:none; animation:fadeUp .4s ease-out; }
+        .tab-panel.active { display:block; }
+        @keyframes fadeUp { from{opacity:0;transform:translateY(16px)} to{opacity:1;transform:translateY(0)} }
+
+        /* STAT CARDS */
+        .stats-row { display:grid; grid-template-columns:repeat(auto-fit,minmax(190px,1fr)); gap:14px; margin-bottom:26px; }
+        .stat-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:11px; padding:18px; position:relative; overflow:hidden; transition:all .3s;
+        }
+        .stat-card:hover { border-color:var(--accent-cyan); transform:translateY(-2px); box-shadow:0 6px 24px var(--glow-cyan); }
+        .stat-label { font-family:'JetBrains Mono',monospace; font-size:.65em; color:var(--text-muted); text-transform:uppercase; letter-spacing:1.5px; margin-bottom:6px; }
+        .stat-value { font-size:2em; font-weight:800; }
+        .stat-sub { font-size:.75em; color:var(--text-secondary); margin-top:3px; }
+        .grad-cyan { background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-orange { background:linear-gradient(135deg,var(--accent-orange),var(--accent-yellow)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-purple { background:linear-gradient(135deg,var(--accent-purple),#e879f9); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-green { background:linear-gradient(135deg,var(--accent-green),#4ade80); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-red { background:linear-gradient(135deg,var(--accent-red),#ff6b81); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+
+        /* SECTION HEADERS */
+        .sec-hdr { display:flex; align-items:center; gap:10px; margin-bottom:18px; padding-bottom:10px; border-bottom:1px solid var(--border); }
+        .sec-hdr h2 { font-size:1.2em; font-weight:700; }
+        .badge { font-family:'JetBrains Mono',monospace; font-size:.65em; padding:3px 9px; border-radius:16px; }
+        .badge-cyan { background:var(--glow-cyan); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.2); }
+        .badge-orange { background:var(--glow-orange); color:var(--accent-orange); border:1px solid rgba(255,159,67,.2); }
+        .badge-green { background:var(--glow-green); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); }
+
+        /* TABLES */
+        .tbl-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); margin-bottom:26px; }
+        table.dt { width:100%; border-collapse:collapse; font-size:.84em; }
+        table.dt th {
+            font-family:'JetBrains Mono',monospace; font-size:.7em; color:var(--text-muted); text-transform:uppercase;
+            letter-spacing:1.2px; padding:12px 14px; background:var(--bg-panel); border-bottom:2px solid var(--border); text-align:left; position:sticky; top:0;
+        }
+        table.dt td { padding:10px 14px; border-bottom:1px solid var(--border); transition:background .15s; }
+        table.dt tr:hover td { background:var(--bg-card-hover); }
+
+        .mbadge { display:inline-block; padding:3px 8px; border-radius:5px; font-family:'JetBrains Mono',monospace; font-size:.78em; font-weight:500; }
+        .mbadge.qwen { background:rgba(59,130,246,.12); color:#60a5fa; border:1px solid rgba(59,130,246,.25); }
+        .mbadge.gptoss { background:rgba(168,85,247,.12); color:#c084fc; border:1px solid rgba(168,85,247,.25); }
+        .mbadge.glm { background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); }
+        .mbadge.minimax { background:rgba(255,159,67,.12); color:#ff9f43; border:1px solid rgba(255,159,67,.25); }
+        .mbadge.devstral { background:rgba(0,212,255,.12); color:#00d4ff; border:1px solid rgba(0,212,255,.25); }
+        .mbadge.deepseek { background:rgba(250,204,21,.12); color:#facc15; border:1px solid rgba(250,204,21,.25); }
+        .mbadge.nemotron { background:rgba(34,197,94,.12); color:#4ade80; border:1px solid rgba(34,197,94,.25); }
+        .mbadge.groq { background:rgba(255,71,87,.12); color:#ff6b81; border:1px solid rgba(255,71,87,.25); }
+        .mbadge.kimi { background:rgba(250,204,21,.12); color:#fde68a; border:1px solid rgba(250,204,21,.2); }
+        .mbadge.llama { background:rgba(59,130,246,.1); color:#93c5fd; border:1px solid rgba(59,130,246,.2); }
+
+        /* SCORE BAR */
+        .sbar { display:flex; align-items:center; gap:6px; }
+        .sbar-bg { width:70px; height:5px; background:var(--border); border-radius:3px; overflow:hidden; }
+        .sbar-fill { height:100%; border-radius:3px; transition:width 1s ease-out; }
+        .sbar-fill.h { background:linear-gradient(90deg,var(--accent-green),#00ff94); }
+        .sbar-fill.m { background:linear-gradient(90deg,var(--accent-orange),#ffc048); }
+        .sbar-fill.l { background:linear-gradient(90deg,var(--accent-red),#ff6b81); }
+        .snum { font-family:'JetBrains Mono',monospace; font-weight:600; font-size:.85em; min-width:28px; }
+
+        /* GROQ SPEED INDICATOR */
+        .speed-ind { display:inline-flex; align-items:center; gap:4px; }
+        .speed-dot { width:7px; height:7px; border-radius:50%; animation:pulse 1.5s ease-in-out infinite; }
+        .speed-dot.ultra { background:var(--accent-green); box-shadow:0 0 8px var(--accent-green); }
+        .speed-dot.fast { background:var(--accent-cyan); box-shadow:0 0 6px var(--accent-cyan); }
+        .speed-dot.normal { background:var(--accent-orange); }
+        @keyframes pulse { 0%,100%{opacity:.5;transform:scale(.8)} 50%{opacity:1;transform:scale(1.2)} }
+
+        /* RECOMMENDATION CARDS */
+        .rec-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(400px,1fr)); gap:16px; margin-bottom:26px; }
+        .rec-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:20px;
+            position:relative; overflow:hidden; transition:all .35s;
+        }
+        .rec-card:hover { border-color:var(--accent-green); box-shadow:0 0 30px var(--glow-green); transform:translateY(-2px); }
+        .rec-card.glow { animation:glowP 3s ease-in-out infinite; }
+        @keyframes glowP { 0%,100%{box-shadow:0 0 16px var(--glow-green)} 50%{box-shadow:0 0 32px var(--glow-green)} }
+        .rec-hdr { display:flex; justify-content:space-between; align-items:flex-start; margin-bottom:12px; }
+        .rec-agent { font-weight:700; font-size:1em; color:var(--accent-cyan); }
+        .impact-badge { padding:2px 8px; border-radius:16px; font-family:'JetBrains Mono',monospace; font-size:.68em; font-weight:600; }
+        .impact-badge.critical { background:rgba(255,71,87,.18); color:var(--accent-red); border:1px solid rgba(255,71,87,.25); }
+        .impact-badge.high { background:rgba(255,159,67,.18); color:var(--accent-orange); border:1px solid rgba(255,159,67,.25); }
+        .impact-badge.medium { background:rgba(250,204,21,.18); color:var(--accent-yellow); border:1px solid rgba(250,204,21,.25); }
+        .swap-vis { display:flex; align-items:center; gap:10px; margin:12px 0; padding:12px; background:var(--bg-panel); border-radius:8px; }
+        .swap-from { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(255,71,87,.08); color:#ff6b81; border:1px solid rgba(255,71,87,.15); text-decoration:line-through; opacity:.65; }
+        .swap-to { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); font-weight:600; }
+        .swap-arrow { color:var(--accent-green); font-size:1.4em; animation:arrP 2s ease-in-out infinite; }
+        @keyframes arrP { 0%,100%{opacity:.4;transform:scale(1)} 50%{opacity:1;transform:scale(1.12)} }
+        .rec-metrics { display:grid; grid-template-columns:repeat(4,1fr); gap:8px; margin-top:12px; }
+        .rec-m { text-align:center; padding:6px; background:var(--bg-deep); border-radius:6px; }
+        .rec-m-label { font-size:.6em; color:var(--text-muted); text-transform:uppercase; letter-spacing:.8px; font-family:'JetBrains Mono',monospace; }
+        .rec-m-val { font-size:1.1em; font-weight:700; margin-top:1px; }
+        .rec-m-val.pos { color:var(--accent-green); }
+        .rec-m-val.neu { color:var(--accent-orange); }
+        .rec-reason { font-size:.82em; color:var(--text-secondary); line-height:1.55; margin-top:10px; padding-top:10px; border-top:1px solid var(--border); }
+
+        /* HEATMAP */
+        .hm-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .hm-title { font-weight:700; font-size:1.05em; }
+        .hm-sub { font-size:.76em; color:var(--text-muted); margin-bottom:14px; }
+        .hm-table { border-collapse:collapse; width:100%; }
+        .hm-table th { font-family:'JetBrains Mono',monospace; font-size:.62em; color:var(--text-muted); padding:6px 4px; text-align:center; white-space:nowrap; }
+        .hm-table th.hm-role { text-align:left; min-width:150px; font-size:.68em; }
+        .hm-table td { text-align:center; padding:5px 3px; font-family:'JetBrains Mono',monospace; font-size:.74em; font-weight:600; border-radius:3px; cursor:pointer; transition:all .12s; min-width:38px; }
+        .hm-table td:hover { transform:scale(1.12); z-index:2; }
+        .hm-table td.hm-r { text-align:left; font-family:'Outfit',sans-serif; font-size:.78em; font-weight:500; color:var(--text-secondary); cursor:default; }
+        .hm-table td.hm-r:hover { transform:none; }
+        .hm-star { color:#FFD700; font-size:.85em; }
+        .hm-cur { outline:2px solid var(--accent-cyan); outline-offset:-2px; }
+
+        /* PROVIDER TAGS */
+        .prov-tag { display:inline-block; padding:1px 6px; border-radius:3px; font-size:.62em; font-family:'JetBrains Mono',monospace; margin-left:4px; }
+        .prov-tag.ollama { background:rgba(0,212,255,.1); color:var(--accent-cyan); }
+        .prov-tag.groq { background:rgba(255,71,87,.1); color:#ff6b81; }
+        .prov-tag.openrouter { background:rgba(168,85,247,.1); color:#c084fc; }
+        .prov-tag.hybrid { background:rgba(0,255,148,.1); color:#00ff94; }
+
+        /* MODEL CARDS */
+        .model-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(290px,1fr)); gap:14px; margin-bottom:26px; }
+        .mc { background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:18px; transition:all .3s; position:relative; }
+        .mc:hover { transform:translateY(-2px); border-color:var(--accent-cyan); box-shadow:0 6px 24px var(--glow-cyan); }
+        .mc-name { font-weight:700; font-size:1.05em; margin-bottom:3px; }
+        .mc-org { font-size:.74em; color:var(--text-muted); margin-bottom:12px; font-family:'JetBrains Mono',monospace; }
+        .mc-row { display:flex; justify-content:space-between; align-items:center; padding:5px 0; border-bottom:1px solid rgba(30,39,54,.4); font-size:.82em; }
+        .mc-row:last-child { border-bottom:none; }
+        .mc-label { color:var(--text-secondary); }
+        .mc-val { font-family:'JetBrains Mono',monospace; font-weight:600; }
+        .mc-tags { display:flex; flex-wrap:wrap; gap:3px; margin-top:10px; }
+        .mc-tag { font-size:.64em; padding:2px 6px; border-radius:3px; font-family:'JetBrains Mono',monospace; background:rgba(0,212,255,.06); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.12); }
+        .mc-best { font-size:.72em; padding:3px 8px; border-radius:4px; background:rgba(0,255,148,.1); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); margin-top:8px; display:inline-block; }
+
+        /* GROQ SECTION */
+        .groq-card { border-left:3px solid var(--accent-red); }
+        .groq-speed { font-family:'JetBrains Mono',monospace; font-size:1.8em; font-weight:800; color:var(--accent-red); }
+
+        /* SUMMARY */
+        .summary { background:linear-gradient(135deg,rgba(0,212,255,.04),rgba(0,255,148,.04)); border:1px solid var(--border-bright); border-radius:12px; padding:24px; margin-bottom:26px; }
+        .summary h3 { color:var(--accent-cyan); font-size:1.1em; margin-bottom:10px; }
+        .summary p { color:var(--text-secondary); line-height:1.65; font-size:.88em; }
+        .summary ul { list-style:none; margin-top:10px; }
+        .summary li { padding:5px 0 5px 18px; position:relative; color:var(--text-secondary); font-size:.86em; line-height:1.55; }
+        .summary li::before { content:'›'; position:absolute; left:0; color:var(--accent-green); font-weight:700; font-size:1.2em; }
+
+        /* FILTER ROW */
+        .frow { display:flex; gap:6px; margin-bottom:16px; flex-wrap:wrap; }
+        .fbtn { padding:5px 12px; background:var(--bg-card); border:1px solid var(--border); color:var(--text-secondary); border-radius:7px; font-family:'Outfit',sans-serif; font-size:.78em; cursor:pointer; transition:all .2s; }
+        .fbtn:hover,.fbtn.active { border-color:var(--accent-cyan); color:var(--accent-cyan); background:rgba(0,212,255,.06); }
+
+        /* TOOLTIP */
+        #ttOverlay { display:none; position:fixed; top:0;left:0;right:0;bottom:0; z-index:999; pointer-events:none; }
+        #ttOverlay.show { display:block; }
+        #ttBox { position:absolute; background:var(--bg-panel); border:1px solid var(--accent-cyan); border-radius:9px; padding:12px 16px; max-width:300px; box-shadow:0 10px 32px rgba(0,0,0,.55); z-index:1000; }
+        #ttBox h4 { color:var(--accent-cyan); font-size:.9em; margin-bottom:4px; }
+        #ttBox p { font-size:.78em; color:var(--text-secondary); line-height:1.45; }
+
+        /* CANVAS */
+        .chart-wrap { border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .chart-title { font-weight:700; font-size:1.05em; margin-bottom:12px; }
+
+        @media(max-width:768px) {
+            .header h1 { font-size:1.6em; }
+            .tabs { flex-wrap:wrap; }
+            .rec-grid,.model-grid { grid-template-columns:1fr; }
+            .stats-row { grid-template-columns:repeat(2,1fr); }
+            .rec-metrics { grid-template-columns:repeat(2,1fr); }
+        }
+
+        /* EXPORT BUTTONS */
+        .export-btn {
+            padding:8px 16px; background:var(--bg-card); border:1px solid var(--border-bright);
+            color:var(--text-secondary); font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600;
+            border-radius:8px; cursor:pointer; transition:all .25s; display:inline-flex; align-items:center;
+        }
+        .export-btn:hover { border-color:var(--accent-cyan); color:var(--text-primary); background:var(--bg-card-hover); }
+        .export-btn-primary {
+            background:linear-gradient(135deg,rgba(0,212,255,.15),rgba(0,255,148,.1));
+            border-color:var(--accent-cyan); color:var(--accent-cyan);
+        }
+        .export-btn-primary:hover { background:linear-gradient(135deg,rgba(0,212,255,.25),rgba(0,255,148,.18)); box-shadow:0 0 20px var(--glow-cyan); }
+
+        /* REC CARD CHECKBOX */
+        .rec-check { position:absolute; top:14px; right:14px; width:22px; height:22px; cursor:pointer; z-index:3; accent-color:var(--accent-green); }
+        .rec-card.selected { border-color:var(--accent-green); background:rgba(0,255,148,.03); }
+        .rec-card { position:relative; }
+    </style>
+</head>
+<body>
+<div class="container">
+    <div class="header">
+        <h1>APAW Agent Model Research v2</h1>
+        <div class="sub">capability-index.yaml · Ollama Cloud + OpenRouter · GLM-5.1 + Qwen 3.6+ · April 2026 · April 2026</div>
+    </div>
+
+    <div class="tabs" id="tabBar">
+        <button class="tab-btn active" onclick="switchTab('overview')">Обзор</button>
+        <button class="tab-btn" onclick="switchTab('groq')">Groq Free Tier</button>
+        <button class="tab-btn" onclick="switchTab('models')">Все модели</button>
+        <button class="tab-btn" onclick="switchTab('heatmap')">Матрица</button>
+        <button class="tab-btn" onclick="switchTab('recs')">Рекомендации</button>
+        <button class="tab-btn" onclick="switchTab('impact')">Анализ профита</button>
+    </div>
+
+    <!-- ========== TAB: OVERVIEW ========== -->
+    <div id="tab-overview" class="tab-panel active">
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Агентов</div><div class="stat-value grad-cyan" id="c1">36</div><div class="stat-sub">32 custom + 4 built-in</div></div>
+            <div class="stat-card"><div class="stat-label">Моделей сейчас</div><div class="stat-value grad-orange">6</div><div class="stat-sub">Coder(9) GLM-5.1(11) K2.6(4)</div></div>
+            <div class="stat-card"><div class="stat-label">Ollama Cloud</div><div class="stat-value grad-purple">20+</div><div class="stat-sub">доступно бесплатно</div></div>
+            <div class="stat-card"><div class="stat-label">Groq + OpenRouter</div><div class="stat-value grad-red">16+</div><div class="stat-sub">free tier моделей</div></div>
+            <div class="stat-card"><div class="stat-label">Рекомендаций</div><div class="stat-value grad-green">11</div><div class="stat-sub">8/8 applied ✅</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Ключевые находки v3 (после коммита caf77f53c8)</h3>
+            <p>Ваш агент уже применил 11 из моих рекомендаций (коммит от 05:21). Но я обнаружил что <strong>до применения</strong> некоторые агенты были на других моделях чем я предполагал:</p>
+            <ul>
+                <li><strong style="color:var(--accent-red)">⚠ Откат Qwen 3.6 Plus</strong> — security-auditor, prompt-optimizer, product-owner и markdown-validator <em>до коммита</em> были на <code>openrouter/qwen3.6-plus:free</code> и <code>deepseek-v3.2</code>, но мои рекомендации их заменили на Ollama-модели. Это снижает разнообразие провайдеров!</li>
+                <li><strong style="color:var(--accent-green)">✅ 11 замен уже применены</strong> — Nemotron 3 Super теперь на 7 ролях, GLM-5 расширен, Qwen3-Coder на Go, markdown-validator</li>
+                <li><strong style="color:var(--accent-orange)">🔴 Осталось 3 агента на gpt-oss:120b</strong> — requirement-refiner, capability-analyst, agent-architect. Всем им нужен Nemotron 3 Super</li>
+                <li><strong>Новая стратегия: гибридный мультипровайдер</strong> — OpenRouter (Qwen 3.6 Plus FREE, 1M ctx) + Groq (gpt-oss 500 t/s) + Ollama (основной). Диверсификация снижает зависимость</li>
+                <li><strong style="color:#00ff94">Qwen 3.6 Plus стоит вернуть</strong> для prompt-optimizer (Terminal-Bench 61.6% > Claude!) и product-owner (1M контекст для backlog)</li>
+                
+                <li><strong>History-miner → Nemotron 3 Super</strong> — самый большой оставшийся прирост: 88 vs 78 (GLM-5). RULER@1M критичен для git history</li>
+                <li><strong style="color:var(--accent-red)">⚠ Prompt Adherence (IF) — новый фактор!</strong> Nemotron 3 Super имеет IF=78 (ниже GLM-5=90, Qwen3.5=92, Qwen3.6+=91). Для ролей с жёстким промптом (evaluator, security-auditor, orchestrator) это снижает эффективность. Qwen 3.6 Plus и GLM-5 лучше следуют инструкциям</li></ul>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Текущая конфигурация</h2><span class="badge badge-cyan">capability-index.yaml</span></div>
+        <div class="tbl-wrap">
+            <table class="dt" id="cfgTable"><thead><tr>
+                <th>Агент</th><th>Модель</th><th>Провайдер</th><th>Категория</th><th>Соответствие</th><th>Статус</th>
+            </tr></thead><tbody id="cfgBody"></tbody></table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: GROQ ========== -->
+    <div id="tab-groq" class="tab-panel">
+        <div class="sec-hdr"><h2>Groq Free Plan — доступные модели</h2><span class="badge badge-orange">бесплатно · LPU inference</span></div>
+
+        <div class="stats-row">
+            <div class="stat-card groq-card"><div class="stat-label">gpt-oss-20b</div><div class="groq-speed">1200 <span style="font-size:.4em;color:var(--text-muted)">t/s</span></div><div class="stat-sub">30 RPM · 1K RPD · 200K TPD</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Анализ лимитов Groq Free для агентского pipeline</h3>
+            <p>При 26 агентах в pipeline, каждый агент делает 5–20 вызовов на задачу. Типичный issue проходит через 8–12 агентов = <strong>~100–200 вызовов</strong>. С лимитом 1K RPD на модель:</p>
+            <ul>
+                
+                
+                <li><strong>Groq Compound</strong>: всего 250 RPD, но 70K TPM — для одноразовых тяжёлых аналитических задач</li>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Все модели Groq Free Tier</h2></div>
+        <div class="tbl-wrap">
+            <table class="dt">
+                <thead><tr><th>Model ID</th><th>RPM</th><th>RPD</th><th>TPM</th><th>TPD</th><th>Скорость</th><th>Применение в APAW</th></tr></thead>
+                <tbody id="groqBody"></tbody>
+            </table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: MODELS ========== -->
+    <div id="tab-models" class="tab-panel">
+        <div class="sec-hdr"><h2>Все доступные модели</h2><span class="badge badge-cyan">Ollama Cloud + Groq + OpenRouter Free</span></div>
+        <div class="frow" id="filterRow"></div>
+        <div class="model-grid" id="modelGrid"></div>
+    </div>
+
+    <!-- ========== TAB: HEATMAP ========== -->
+    <div id="tab-heatmap" class="tab-panel">
+        <div class="hm-wrap">
+            <div class="hm-title">Матрица «Агент × Модель»: оценка совместимости (с учётом Prompt Adherence)</div>
+            <div class="hm-sub">0–100 · Взвешенная оценка = 60% бенчмарк роли + 25% Instruction Following + 15% скорость/контекст · ★ = лучший · <span style="outline:2px solid var(--accent-cyan);outline-offset:-2px;padding:0 3px;border-radius:2px">обведено</span> = текущий · <strong style="color:var(--accent-yellow)">← 11 моделей · 🟢L 🟡M 🔴H = Reasoning Effort →</strong></div>
+            <div style="overflow-x:auto"><table class="hm-table" id="hmTable"></table></div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: RECOMMENDATIONS ========== -->
+    <div id="tab-recs" class="tab-panel">
+        <div class="sec-hdr"><h2>Рекомендации</h2><span class="badge badge-green">4 замены (2 BROKEN) + 7 подтверждений 06.04.2026</span></div>
+
+        <div style="display:flex;gap:10px;margin-bottom:18px;flex-wrap:wrap;align-items:center;">
+            <button onclick="toggleAllRecs()" class="export-btn" id="selectAllBtn">☑ Выбрать все</button>
+            <button onclick="exportJSON()" class="export-btn export-btn-primary" id="exportBtn">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" style="vertical-align:-2px;margin-right:4px"><path d="M21 15v4a2 2 0 01-2 2H5a2 2 0 01-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>
+                Скачать JSON для агента
+            </button>
+            <span id="selectedCount" style="font-family:'JetBrains Mono',monospace;font-size:.75em;color:var(--text-muted);">0 из 11 выбрано</span>
+        </div>
+
+        <div class="rec-grid" id="recGrid"></div>
+
+        <!-- JSON Preview Modal -->
+        <div id="jsonModal" style="display:none;position:fixed;top:0;left:0;right:0;bottom:0;background:rgba(0,0,0,.7);z-index:9999;justify-content:center;align-items:center;padding:20px;">
+            <div style="background:var(--bg-panel);border:1px solid var(--accent-cyan);border-radius:14px;max-width:800px;width:100%;max-height:85vh;display:flex;flex-direction:column;box-shadow:0 20px 60px rgba(0,0,0,.5);">
+                <div style="display:flex;justify-content:space-between;align-items:center;padding:18px 22px;border-bottom:1px solid var(--border);">
+                    <div>
+                        <div style="font-weight:700;font-size:1.05em;">agent-model-recommendations.json</div>
+                        <div style="font-size:.75em;color:var(--text-muted);margin-top:2px;font-family:'JetBrains Mono',monospace">Готов для передачи агенту-оркестратору</div>
+                    </div>
+                    <div style="display:flex;gap:8px;">
+                        <button onclick="copyJSON()" class="export-btn" id="copyBtn">📋 Копировать</button>
+                        <button onclick="downloadJSON()" class="export-btn export-btn-primary">⬇ Скачать .json</button>
+                        <button onclick="closeModal()" class="export-btn" style="border-color:var(--accent-red);color:var(--accent-red);">✕</button>
+                    </div>
+                </div>
+                <pre id="jsonPreview" style="flex:1;overflow:auto;padding:18px 22px;margin:0;font-family:'JetBrains Mono',monospace;font-size:.78em;line-height:1.6;color:var(--accent-green);background:var(--bg-deep);border-radius:0 0 14px 14px;"></pre>
+            </div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: IMPACT ========== -->
+    <div id="tab-impact" class="tab-panel">
+        <div class="sec-hdr"><h2>Совокупный анализ профита</h2><span class="badge badge-green">если применить все рекомендации</span></div>
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Средний прирост</div><div class="stat-value grad-green">+12</div><div class="stat-sub">пунктов по матрице</div></div>
+            <div class="stat-card"><div class="stat-label">Применено</div><div class="stat-value grad-red">8/8</div><div class="stat-sub">все рекомендации ✅</div></div>
+            <div class="stat-card"><div class="stat-label">Qwen 3.6+</div><div class="stat-value grad-purple">0</div><div class="stat-sub">полностью на Ollama!</div></div>
+            <div class="stat-card"><div class="stat-label">GLM-5.1</div><div class="stat-value grad-orange">12</div><div class="stat-sub">10 custom + 2 built-in</div></div>
+        </div>
+        <div class="chart-wrap">
+            <div class="chart-title">Прирост по категориям: до → после</div>
+            <canvas id="impactCanvas" height="340"></canvas>
+        </div>
+        <div class="summary">
+            <h3>Детальный анализ прироста</h3>
+            <div id="impactDetails"></div>
+
+            <div style="margin-top:32px">
+                <div class="sec-hdr"><h2>APAW Pipeline vs ТОП закрытых моделей (апрель 2026)</h2></div>
+                <p style="font-size:.82em;color:var(--text-muted);margin-bottom:16px">
+                    Сравнение лучших моделей в вашем pipeline с лидерами рынка по ключевым бенчмаркам. 
+                    <strong style="color:var(--accent-green)">🟢</strong> = APAW обгоняет, 
+                    <strong style="color:var(--accent-yellow)">🟡</strong> = на уровне (±3%), 
+                    <strong style="color:var(--accent-red)">🔴</strong> = отстаёт
+                </p>
+                <div style="overflow-x:auto">
+                <table id="benchTable" style="width:100%;border-collapse:collapse;font-size:.78em;font-family:'JetBrains Mono',monospace">
+                </table>
+                </div>
+                <p style="font-size:.72em;color:var(--text-muted);margin-top:12px">
+                    * SWE-V = SWE-Bench Verified, SWE-P = SWE-Bench Pro, T-Bench = Terminal-Bench 2.0, LCB = LiveCodeBench, GPQA = GPQA Diamond<br>
+                    Данные: swebench.com, marc0.dev, tokenmix.ai, ollama.com — апрель 2026. Стоимость: примерная за 1M input tokens.
+                </p>
+            </div>
+        </div>
+    </div>
+</div>
+
+<div id="ttOverlay"><div id="ttBox"></div></div>
+
+<script>
+// ACTUAL STATE from _kilo.zip (April 25, 2026)
+// 32 custom agents + 4 built-in = 36 total
+// 6 unique models: Qwen3-Coder(9), GLM-5.1(11), Nemotron Super(6), Kimi K2.6(4!), M2.5(2), Nano(1)
+// ALL 7 RECS APPLIED! GLM-5.1=12 roles, K2.6=3 roles, Qwen3.6+=0, GLM-5=0, DeepSeek=0
+const cfg = [
+    // Qwen3-Coder 480B (8 coding agents)
+    {a:'lead-developer',m:'qwen3-coder:480b',p:'Ollama',cat:'Core Dev',b:'qwen',fit:92,s:'optimal'},
+    {a:'frontend-developer',m:'minimax-m2.5',p:'Ollama',cat:'Core Dev',b:'minimax',fit:92,s:'optimal'},
+    {a:'backend-developer',m:'qwen3-coder:480b',p:'Ollama',cat:'Core Dev',b:'qwen',fit:91,s:'optimal'},
+    {a:'go-developer',m:'qwen3-coder:480b',p:'Ollama',cat:'Core Dev',b:'qwen',fit:85,s:'optimal'},
+    {a:'flutter-developer',m:'qwen3-coder:480b',p:'Ollama',cat:'Core Dev',b:'qwen',fit:86,s:'optimal'},
+    {a:'php-developer',m:'qwen3-coder:480b',p:'Ollama',cat:'Core Dev',b:'qwen',fit:87,s:'optimal'},
+    {a:'python-developer',m:'qwen3-coder:480b',p:'Ollama',cat:'Core Dev',b:'qwen',fit:90,s:'optimal'},
+    {a:'sdet-engineer',m:'qwen3-coder:480b',p:'Ollama',cat:'QA',b:'qwen',fit:88,s:'optimal'},
+    // GLM-5.1 (8 agents) — SWE-Pro 58.4 #1!
+    {a:'orchestrator',m:'kimi-k2.6:cloud',p:'Ollama Cloud',cat:'Process',b:'kimi',fit:92,s:'optimal'},
+    {a:'evaluator',m:'glm-5.1',p:'Ollama',cat:'Process',b:'glm',fit:86,s:'optimal'},
+    {a:'capability-analyst',m:'glm-5.1',p:'Ollama',cat:'Analysis',b:'glm',fit:85,s:'optimal'},
+    {a:'architect-indexer',m:'glm-5.1',p:'Ollama',cat:'Analysis',b:'glm',fit:88,s:'optimal'},
+    {a:'pipeline-judge',m:'glm-5.1',p:'Ollama',cat:'Process',b:'glm',fit:86,s:'good'},
+    {a:'release-manager',m:'glm-5.1',p:'Ollama',cat:'Process',b:'glm',fit:82,s:'good'},
+    {a:'requirement-refiner',m:'glm-5.1',p:'Ollama',cat:'Analysis',b:'glm',fit:88,s:'optimal'},
+    {a:'workflow-architect',m:'glm-5.1',p:'Ollama',cat:'Workflow',b:'glm',fit:84,s:'good'},
+    // Nemotron 3 Super (7 agents)
+    {a:'agent-architect',m:'kimi-k2.6:cloud',p:'Ollama Cloud',cat:'Meta',b:'kimi',fit:86,s:'optimal'},
+    {a:'security-auditor',m:'nemotron-3-super',p:'Ollama',cat:'Security',b:'nemotron',fit:76,s:'good'},
+    {a:'performance-engineer',m:'nemotron-3-super',p:'Ollama',cat:'Performance',b:'nemotron',fit:78,s:'good'},
+    {a:'history-miner',m:'nemotron-3-super',p:'Ollama',cat:'Analysis',b:'nemotron',fit:85,s:'optimal'},
+    {a:'memory-manager',m:'nemotron-3-super',p:'Ollama',cat:'Cognitive',b:'nemotron',fit:86,s:'optimal'},
+    {a:'planner',m:'nemotron-3-super',p:'Ollama',cat:'Cognitive',b:'nemotron',fit:80,s:'good'},
+    {a:'reflector',m:'nemotron-3-super',p:'Ollama',cat:'Cognitive',b:'nemotron',fit:78,s:'good'},
+    // GLM-5 (3 agents)
+    {a:'browser-automation',m:'kimi-k2.6:cloud',p:'Ollama Cloud',cat:'Testing',b:'kimi',fit:86,s:'optimal'},
+    {a:'product-owner',m:'glm-5.1',p:'Ollama',cat:'Management',b:'glm',fit:84,s:'optimal'},
+    {a:'visual-tester',m:'qwen3-coder:480b',p:'Ollama',cat:'Testing',b:'qwen',fit:82,s:'optimal'},
+    // Qwen 3.6+ FREE (2 agents)
+    {a:'prompt-optimizer',m:'glm-5.1',p:'Ollama',cat:'Process',b:'glm',fit:82,s:'good'},
+    {a:'system-analyst',m:'glm-5.1',p:'Ollama',cat:'Analysis',b:'glm',fit:90,s:'optimal'},
+    // MiniMax M2.5 (2 agents)
+    {a:'code-skeptic',m:'minimax-m2.5',p:'Ollama',cat:'QA',b:'minimax',fit:85,s:'good'},
+    {a:'the-fixer',m:'minimax-m2.5',p:'Ollama',cat:'Fixes',b:'minimax',fit:88,s:'optimal'},
+    // DeepSeek V3.2 (1 agent)
+    {a:'devops-engineer',m:'kimi-k2.6:cloud',p:'Ollama Cloud',cat:'DevOps',b:'kimi',fit:88,s:'optimal'},
+    // Nemotron Nano (1 agent)
+    {a:'markdown-validator',m:'nemotron-3-nano:30b',p:'Ollama',cat:'Validation',b:'nemotron',fit:70,s:'good'},
+    // Built-in (from kilo.jsonc)
+    {a:'[built-in] code',m:'qwen3-coder:480b',p:'Ollama',cat:'Built-in',b:'qwen',fit:92,s:'optimal'},
+    {a:'[built-in] ask',m:'glm-5.1',p:'Ollama',cat:'Built-in',b:'glm',fit:88,s:'optimal'},
+    {a:'[built-in] plan',m:'nemotron-3-super',p:'Ollama',cat:'Built-in',b:'nemotron',fit:80,s:'good'},
+    {a:'[built-in] debug',m:'glm-5.1',p:'Ollama',cat:'Built-in',b:'glm',fit:88,s:'optimal'},
+];
+
+const groqModels = [
+    {id:'openai/gpt-oss-20b',rpm:30,rpd:'1K',tpm:'8K',tpd:'200K',speed:'1200+',use:'Ультра-быстрый fallback для лёгких ролей (markdown-validator).'},
+    {id:'llama-3.1-8b-instant',rpm:30,rpd:'14.4K',tpm:'6K',tpd:'500K',speed:'~800',use:'14.4K RPD! Самый высокий лимит. Для health-check / ping ролей.'},
+    {id:'groq/compound',rpm:30,rpd:'250',tpm:'70K',tpd:'—',speed:'varies',use:'Мультимодельная агрегация. Для research-задач.'},
+    {id:'groq/compound-mini',rpm:30,rpd:'250',tpm:'70K',tpd:'—',speed:'varies',use:'Лёгкая версия compound.'},
+    {id:'llama-prompt-guard-2',rpm:30,rpd:'14.4K',tpm:'15K',tpd:'500K',speed:'~1K',use:'Security: входной фильтр для security-auditor (14.4K RPD!).'},
+];
+
+const ollamaModels = [
+    // ifScore = Instruction Following composite (IFEval + IFBench + agent prompt adherence), 0-100
+    {n:'Qwen3-Coder 480B',org:'Qwen',par:'480B/35B active',ctx:'256K→1M',swe:66.5,ifScore:88,cat:['coding','agent'],str:'SOTA open-source кодинг. Сравним с Claude Sonnet 4.',tags:['coding','agent','tools']},
+    {n:'MiniMax M2.5',org:'MiniMax',par:'MoE undisclosed',ctx:'128K',swe:80.2,ifScore:82,cat:['coding','agent'],str:'Лидер SWE-bench 80.2%. Полный lifecycle разработки.',tags:['coding','agent']},
+    {n:'MiniMax M2.7',org:'MiniMax',par:'~10B active',ctx:'128K',swe:78,ifScore:80,cat:['coding','agent','efficient'],str:'Самообучаемая. 56.2% SWE-Pro. 100 TPS. $0.30/M.',tags:['coding','agent','self-evolving']},
+    {n:'DeepSeek V4-Pro',org:'DeepSeek',par:'1.6T/49B active MoE',ctx:'1M',swe:80.6,ifScore:89,cat:['coding','agent','reasoning'],str:'SWE-V 80.6, LiveCodeBench 93.5(#1!), Terminal-Bench 67.9, Codeforces 3206, 1M ctx, 27% FLOPs vs V3.2. MIT.',tags:['coding','agent','thinking','tools']},
+    {n:'DeepSeek V4-Pro',org:'DeepSeek',par:'284B/13B active MoE',ctx:'1M',swe:79.0,ifScore:86,cat:['coding','efficient','agent'],str:'SWE-V ~79%, Flash Max = Pro уровень reasoning. 13B active = ультрабыстрый. 1M ctx. FP4+FP8. MIT.',tags:['coding','efficient','agent','thinking']},
+    {n:'Kimi K2.6',org:'Moonshot AI',par:'1T/32B active MoE',ctx:'256K',swe:80.2,ifScore:91,cat:['coding','agent','multimodal'],str:'SWE-Pro 58.6(#1!), SWE-V 80.2, Terminal-Bench 66.7, HLE 54.0(#1!), BrowseComp 83.2. 13h autonomous. 300 sub-agent swarm. Modified MIT.',tags:['coding','agent','swarm','vision','thinking','tools']},
+    {n:'Nemotron 3 Super',org:'NVIDIA',par:'120B/12B active',ctx:'1M',swe:60.5,ifScore:78,cat:['agent','reasoning','efficient'],str:'SWE-bench 60.5%. RULER@1M 91.75%! Но IF ниже — Mamba-layers иногда «теряют» инструкции в длинных промптах.',tags:['agent','1M-ctx','thinking']},
+    {n:'GLM-5',org:'Z.ai',par:'744B/40B active',ctx:'128K',swe:null,ifScore:90,cat:['reasoning','agent'],str:'Мощный reasoning. Arena ELO 1451. Отличный instruction following (IFEval ~90+).',tags:['reasoning','agent']},
+    {n:'DeepSeek V4-Pro',org:'DeepSeek',par:'Large MoE',ctx:'128K',swe:null,ifScore:75,cat:['reasoning'],str:'Хороший reasoning, но IF нестабилен — иногда игнорирует формат вывода.',tags:['reasoning']},
+    {n:'Qwen 3.5 122B',org:'Qwen',par:'122B/10B active',ctx:'128K',swe:null,ifScore:92,cat:['reasoning','efficient'],str:'IFEval 92.6%! Лучший IF среди open-source. Multimodal. Thinking.',tags:['vision','thinking','tools']},
+    {n:'Qwen3-Coder-Next',org:'Qwen',par:'80B/3B active',ctx:'128K',swe:70,ifScore:84,cat:['coding','efficient'],str:'70% SWE-bench с 3B active! Хороший IF для кодинга.',tags:['coding','efficient','tools']},
+    {n:'Cogito 2.1 671B',org:'Cognitive',par:'671B MoE',ctx:'128K',swe:null,ifScore:76,cat:['reasoning'],str:'MIT лицензия. 671B total. IF неплохой, но уступает GLM/Qwen.',tags:['reasoning']},
+    {n:'Qwen 3.6 Plus',org:'Qwen',par:'Hybrid MoE',ctx:'1M',swe:78.8,ifScore:91,cat:['coding','agent','reasoning'],str:'FREE на OpenRouter! 1M контекст. Always-on CoT. Превосходный IF — наследник Qwen 3.5 (92.6%).',tags:['coding','agent','1M-ctx','free'],or:true},
+    {n:'Step 3.5 Flash',org:'StepFun',par:'MoE',ctx:'128K',swe:null,ifScore:79,cat:['efficient'],str:'Бесплатна на OpenRouter. IF средний.',tags:['efficient','free'],or:true},
+    {n:'DeepSeek R1',org:'DeepSeek',par:'671B MoE',ctx:'128K',swe:null,ifScore:73,cat:['reasoning'],str:'Мощные reasoning-цепочки. Но IF слабый — часто генерирует лишний reasoning вместо ответа.',tags:['reasoning','thinking','free'],or:true},
+];
+
+// Heatmap: agents × models (Ollama + Groq-unique models)
+// Instruction Following compliance scores (0-100) based on IFEval/IFBench/BenchLM data April 2026
+// Higher = better follows system prompts, role definitions, output format constraints
+const ifScores = {
+    0: 88,  // Qwen3-Coder 480B — strong IF from agentic RL
+    1: 82,  // MiniMax M2.5 — good IF, spec-writing
+    2: 78,  // MiniMax M2.7 — slightly weaker IF
+    3: 85,  // Nemotron 3 Super — strong IF, agent-trained
+    4: 80,  // GLM-5 — function calling leader
+    5: 88,  // GLM-5.1 — SWE-Pro #1, 8-hour autonomous, improved IF
+    6: 88,  // V4-Pro Max — 1.6T/49B, 1M ctx, SWE-V 80.6, Terminal 67.9, LiveCodeBench 93.5! MIT
+    7: 86,  // Qwen 3.5 122B — improved IF
+    8: 84,  // Qwen3-Coder-Next — agentic training
+    9: 90,  // Qwen 3.6 Plus — always-on CoT, best IF
+    10: 91, // Kimi K2.6 — Intelligence Index 54, 13h autonomous, improved IF
+    12: 89, // DeepSeek V4-Pro — 1.6T/49B, 1M ctx, LiveCodeBench 93.5, MIT
+    13: 86, // DeepSeek V4-Flash — 284B/13B, 1M ctx, efficient, MIT
+};
+
+// IF-adjusted heatmap: raw_score * (0.7 + 0.3 * IF/100)
+// This means IF=100 → score×1.0, IF=50 → score×0.85, IF=0 → score×0.7
+function adjustForIF(scores) {
+    return scores.map((s, idx) => {
+        const ifScore = ifScores[idx] || 70;
+        const mult = 0.7 + 0.3 * (ifScore / 100);
+        return Math.round(s * mult);
+    });
+}
+
+const hmModels = [
+    {n:'Qwen3-Coder',p:'Ollama',if:88},
+    {n:'M2.5',p:'Ollama',if:82},
+    {n:'M2.7',p:'Ollama',if:78},
+    {n:'Nem.3 Super',p:'Ollama',if:85},
+    {n:'GLM-5',p:'Ollama',if:80},
+    {n:'🔥 GLM-5.1',p:'Ollama',if:88},
+    {n:'🔥 V4-Pro Max',p:'Ollama Cloud',if:88},
+    {n:'Qwen 3.5',p:'Ollama',if:86},
+    {n:'Q3-Coder-Next',p:'Ollama',if:84},
+    {n:'Qwen 3.6+',p:'OpenRouter',if:90},
+    {n:'\u{1f525} Kimi K2.6',p:'Ollama Cloud',if:91}
+];
+
+const hmAgents = [
+    // c=current model idx, re=reasoning effort (L/M/H)
+    // 0=Qwen3Coder 1=M2.5 2=M2.7 3=Nem.Super 4=GLM-5 5=GLM-5.1 6=V4-Pro-Max 7=Qwen3.5 8=Q3CoderNext 9=Qwen3.6+ 10=KimiK2.6
+    // === Qwen3-Coder agents (c:0) ===
+    {n:'lead-developer',c:0,re:'M',s:[92,86,82,70,68,75,88,66,80,88,90]},
+    {n:'frontend-developer',c:1,re:'M',s:[86,92,88,62,56,64,82,60,76,88,86]},
+    {n:'backend-developer',c:0,re:'M',s:[91,84,80,68,63,72,86,62,78,87,90]},
+    {n:'go-developer',c:0,re:'M',s:[85,78,74,66,58,68,88,58,74,82,86]},
+    {n:'flutter-developer',c:0,re:'M',s:[86,70,66,60,53,62,78,58,74,82,84]},
+    {n:'php-developer',c:0,re:'M',s:[87,76,72,64,56,66,74,60,76,84,86]},
+    {n:'python-developer',c:0,re:'M',s:[90,82,78,66,60,70,78,64,78,88,88]},
+    {n:'sdet-engineer',c:0,re:'M',s:[88,84,80,70,63,72,84,64,78,84,87]},
+    // === GLM-5.1 agents (c:5) ===
+    {n:'orchestrator',c:10,re:'M',s:[74,70,68,80,82,90,86,78,62,84,92]},
+    {n:'evaluator',c:5,re:'M',s:[70,73,70,78,78,86,84,76,58,81,84]},
+    {n:'capability-analyst',c:5,re:'M',s:[72,68,66,76,78,85,82,75,60,79,82]},
+    {n:'architect-indexer',c:5,re:'M',s:[70,64,62,74,80,88,78,76,58,80,84]},
+    {n:'pipeline-judge',c:5,re:'L',s:[64,68,65,78,76,86,82,74,56,80,84]},
+    {n:'release-manager',c:5,re:'L',s:[72,66,64,74,76,82,78,72,60,76,78]},
+    {n:'requirement-refiner',c:5,re:'M',s:[66,62,60,72,80,88,82,74,54,78,82]},
+    {n:'workflow-architect',c:5,re:'M',s:[68,62,60,76,76,84,80,72,56,80,82]},
+    // === Nemotron 3 Super agents (c:3) ===
+    {n:'agent-architect',c:10,re:'H',s:[78,72,70,78,76,84,82,76,66,82,86]},
+    {n:'security-auditor',c:3,re:'H',s:[76,74,68,76,68,78,80,72,64,75,80]},
+    {n:'performance-engineer',c:3,re:'M',s:[78,75,70,78,74,82,84,70,67,76,82]},
+    {n:'history-miner',c:3,re:'L',s:[68,60,56,85,78,88,86,72,56,84,82]},
+    {n:'memory-manager',c:3,re:'M',s:[63,58,56,86,72,84,86,70,50,87,84]},
+    {n:'planner',c:3,re:'H',s:[72,68,66,80,78,85,88,78,60,85,86]},
+    {n:'reflector',c:3,re:'M',s:[68,66,64,78,76,82,84,76,56,82,80]},
+    // === GLM-5 agents (c:4) ===
+    {n:'browser-automation',c:10,re:'M',s:[87,72,68,61,53,64,82,56,72,82,86]},
+    {n:'product-owner',c:5,re:'L',s:[60,56,54,74,78,84,76,74,48,78,76]},
+    {n:'visual-tester',c:0,re:'M',s:[82,68,64,55,48,58,76,54,66,76,78]},
+    // === Qwen 3.6+ agents (c:9) ===
+    {n:'prompt-optimizer',c:5,re:'M',s:[76,74,72,76,75,82,80,74,64,83,82]},
+    {n:'system-analyst',c:5,re:'H',s:[70,66,63,74,82,90,88,76,58,80,86]},
+    // === M2.5 agents (c:1) ===
+    {n:'code-skeptic',c:1,re:'M',s:[82,85,80,73,72,78,82,70,72,80,82]},
+    {n:'the-fixer',c:1,re:'M',s:[89,88,84,71,64,74,88,64,82,86,90]},
+    // === DeepSeek V3.2 (c:6 = V4F-Max column, but actual is V3.2!) ===
+    {n:'devops-engineer',c:10,re:'M',s:[66,53,48,78,75,84,86,70,54,76,88]},
+    // === Nano (c:-1, not in matrix) ===
+    {n:'markdown-validator',c:-1,re:'L',s:[43,38,36,52,55,62,68,56,40,50,56]},
+    // === Built-in ===
+    {n:'[built-in] debug',c:5,re:'H',s:[78,80,76,72,64,88,90,68,76,85,90]},
+];
+
+const recs = [
+    // === PREV 7+2 APPLIED. V4-Pro Max теперь доступен! ===
+
+    // --- NEW: V4-Pro Max замены ---
+    {a:"[built-in] debug",from:"glm-5.1 (88)",fromP:"Ollama",to:"V4-Pro Max (★90) / K2.6 (★90) RE:High",toP:"Ollama Cloud",imp:"high",
+     q:"+2%",sp:"~1x",ctx:"200K→1M",prov:"Ollama Cloud",
+     r:"★ матрицы: V4-Pro=90 и K2.6=90 (TIE!), GLM-5.1=88. V4-Pro: LiveCodeBench 93.5(#1!), Terminal 67.9, 1M ctx для полного проекта. K2.6: 13h auto sessions. Оба лучше GLM-5.1. RE:High для debug."},
+    {a:"planner",from:"nemotron-3-super (80)",fromP:"Ollama",to:"V4-Pro Max (★88) RE:High",toP:"Ollama Cloud",imp:"high",
+     q:"+10%",sp:"~1x",ctx:"1M",prov:"Ollama Cloud",
+     r:"★ матрицы: V4-Pro=88(лучший!), K2.6=86, GLM-5.1=85, Nem=80. V4-Pro: GPQA 90.1 (reasoning), 1M ctx сохраняется (vs потеря при K2.6). RE:High для chain-of-thought planning."},
+    {a:"go-developer",from:"qwen3-coder:480b (85)",fromP:"Ollama",to:"V4-Pro Max (★88) RE:Medium",toP:"Ollama Cloud",imp:"medium",
+     q:"+4%",sp:"~1x",ctx:"256K→1M",prov:"Ollama Cloud",
+     r:"★ матрицы: V4-Pro=88(лучший для Go!), K2.6=86, Qwen3Coder=85. DeepSeek модели традиционно сильны в Go/Rust. 1M ctx для крупных Go-проектов."},
+    {a:"history-miner",from:"nemotron-3-super (★85)",fromP:"Ollama",to:"V4-Pro Max (86) + Nem fallback",toP:"Hybrid",imp:"medium",
+     q:"+1%",sp:"~1x",ctx:"1M",prov:"Ollama Cloud + Ollama",
+     r:"V4-Pro=86 чуть лучше Nemotron=85. 1M ctx у обоих. MRCR 83.5 у V4-Pro — лучшее long-context retrieval. Nemotron как fallback (RULER 91.75%)."},
+
+    // --- APPLIED (для справки) ---
+    {a:"frontend-dev → M2.5",from:"qwen3-coder (90)",fromP:"Ollama",to:"MiniMax M2.5 (★92) ✅",toP:"Ollama",imp:"low",
+     q:"+2%",sp:"=",ctx:"204K",prov:"Ollama",r:"Spec-writing, UI architect. APPLIED."},
+    {a:"devops → K2.6",from:"deepseek-v3.2",fromP:"",to:"kimi-k2.6:cloud ✅",toP:"Ollama Cloud",imp:"low",
+     q:"+35%",sp:"=",ctx:"256K",prov:"",r:"APPLIED."},
+
+    // --- Fine-tuning ---
+    {a:"orchestrator",from:"glm-5.1 (★90)",fromP:"Ollama",to:"K2.6 (★92) RE:Medium",toP:"Ollama Cloud",imp:"medium",
+     q:"+2%",sp:"~1x",ctx:"200K→256K",prov:"Ollama Cloud",
+     r:"K2.6=92★ всё ещё лучший для orchestration. V4-Pro=86 слабее. 300 sub-agent swarm."},
+    {a:"the-fixer",from:"minimax-m2.5 (★88)",fromP:"Ollama",to:"V4-Pro (★88) / K2.6 (★90)",toP:"Ollama Cloud",imp:"medium",
+     q:"+2%",sp:"~1x",ctx:"128K→1M/256K",prov:"Ollama Cloud",
+     r:"K2.6=90(лучший), V4-Pro=88=M2.5. M2.5 SWE-bench 80.2% стабильнее. Не срочно."},
+
+    // --- Подтверждение ---
+    {a:"Qwen3-Coder (7 coding)",from:"qwen3-coder",fromP:"Ollama",to:"✅",toP:"",imp:"low",
+     q:"=0%",sp:"=",ctx:"256K",prov:"Ollama",r:"lead=92★, backend=91★, python=90★."},
+    {a:"GLM-5.1 (12 agents)",from:"glm-5.1",fromP:"Ollama",to:"✅",toP:"",imp:"low",
+     q:"=0%",sp:"=",ctx:"200K",prov:"Ollama",r:"orchestrator=90, system-analyst=90. SWE-Pro #1."},
+    {a:"Kimi K2.6 (3 agents)",from:"kimi-k2.6",fromP:"Ollama Cloud",to:"✅",toP:"",imp:"low",
+     q:"=0%",sp:"=",ctx:"256K",prov:"Ollama Cloud",r:"devops=88★, browser=86, agent-arch=86."},
+];
+
+const impactData = [
+    {cat:"debug GLM5.1→V4-Pro/K2.6",b:88,a:90,d:2,n:"LiveCodeBench 93.5, Terminal 67.9"},
+    {cat:"planner Nem→V4-Pro Max",b:80,a:88,d:8,n:"★88! GPQA 90.1, 1M ctx"},
+    {cat:"go-dev Coder→V4-Pro Max",b:85,a:88,d:3,n:"★88! Go/Rust specialist, 1M ctx"},
+    {cat:"history-miner →V4-Pro",b:85,a:86,d:1,n:"MRCR 83.5, long-context"},
+    {cat:"orchestrator →K2.6 (next)",b:90,a:92,d:2,n:"300 sub-agent swarm"},
+    {cat:"frontend → M2.5 ✅",b:90,a:92,d:2,n:"Spec-writing, UI architect"},
+    {cat:"devops → K2.6 ✅",b:65,a:88,d:23,n:"IF:65→91! Terminal 66.7"},
+    {cat:"Qwen3-Coder (7) ✅",b:90,a:90,d:0,n:"SOTA coding"},
+    {cat:"GLM-5.1 (12) ✅",b:87,a:87,d:0,n:"SWE-Pro #1"},
+    {cat:"Nemotron Super (6) ✅",b:82,a:82,d:0,n:"1M ctx, RULER 91.75%"},
+];
+
+
+// ======================= RENDER =======================
+function switchTab(id) {
+    document.querySelectorAll('.tab-panel').forEach(p=>p.classList.remove('active'));
+    document.querySelectorAll('.tab-btn').forEach(b=>b.classList.remove('active'));
+    document.getElementById('tab-'+id).classList.add('active');
+    event.target.classList.add('active');
+    if(id==='impact') requestAnimationFrame(()=>setTimeout(drawChart,50));
+}
+
+function renderCfg() {
+    const b=document.getElementById('cfgBody');
+    b.innerHTML=cfg.map((c,i)=>{
+        const si=c.s==='optimal'?'✅':c.s==='good'?'🟡':c.s==='overspec'?'🔵':c.s==='new'?'🆕':c.s==='broken'?'💀':'🔴';
+        const st=c.s==='optimal'?'Оптимально':c.s==='good'?'Хорошо':c.s==='overspec'?'Overspec':c.s==='new'?'Не назначена':c.s==='broken'?'НЕ РАБОТАЕТ':'Улучшить';
+        const bc=c.fit>=85?'h':c.fit>=70?'m':'l';
+        const sc=c.fit>=85?'var(--accent-green)':c.fit>=70?'var(--accent-orange)':'var(--accent-red)';
+        const prevHtml=c.prev?`<div style="font-size:.68em;color:var(--text-muted);margin-top:2px;text-decoration:line-through">${c.prev}</div>`:'';
+        return `<tr style="animation:fadeUp .3s ${i*.03}s ease-out both">
+            <td style="font-weight:600">${c.a}</td>
+            <td><span class="mbadge ${c.b}">${c.m}</span>${prevHtml}</td>
+            <td><span class="prov-tag ${c.p.toLowerCase()}">${c.p}</span></td>
+            <td style="color:var(--text-secondary)">${c.cat}</td>
+            <td><div class="sbar"><div class="sbar-bg"><div class="sbar-fill ${bc}" style="width:${c.fit}%"></div></div><span class="snum" style="color:${sc}">${c.fit}</span></div></td>
+            <td>${si} ${st}</td></tr>`;
+    }).join('');
+}
+
+function renderGroq() {
+    const b=document.getElementById('groqBody');
+    b.innerHTML=groqModels.map(g=>{
+        const spd=parseInt(g.speed)||0;
+        const dotCls=spd>=800?'ultra':spd>=400?'fast':'normal';
+        return `<tr>
+            <td><span class="mbadge groq">${g.id}</span></td>
+            <td>${g.rpm}</td><td>${g.rpd}</td><td>${g.tpm}</td><td>${g.tpd}</td>
+            <td><div class="speed-ind"><span class="speed-dot ${dotCls}"></span> ${g.speed} t/s</div></td>
+            <td style="color:var(--text-secondary);font-size:.82em;max-width:280px">${g.use}</td></tr>`;
+    }).join('');
+}
+
+function renderModels() {
+    const grid=document.getElementById('modelGrid');
+    const cats=new Set(); ollamaModels.forEach(m=>m.cat.forEach(c=>cats.add(c)));
+    document.getElementById('filterRow').innerHTML='<button class="fbtn active" onclick="filterM(\'all\',this)">Все</button>'+
+        [...cats].map(c=>`<button class="fbtn" onclick="filterM('${c}',this)">${c}</button>`).join('');
+
+    grid.innerHTML=ollamaModels.map((m,i)=>{
+        const bc=m.swe&&m.swe>=75?'var(--accent-green)':m.swe&&m.swe>=60?'var(--accent-cyan)':'var(--border)';
+        return `<div class="mc" style="animation:fadeUp .35s ${i*.05}s ease-out both;border-color:${bc}" data-cats='${JSON.stringify(m.cat)}'>
+            <div class="mc-name">${m.n} ${m.groq?'<span class="prov-tag groq">Groq '+m.groqSpeed+'t/s</span>':''}${m.or?'<span class="prov-tag openrouter">OpenRouter FREE</span>':''}</div>
+            <div class="mc-org">${m.org} · ${m.par} · ctx ${m.ctx}</div>
+            ${m.swe?`<div class="mc-row"><span class="mc-label">SWE-bench</span><span class="mc-val" style="color:${m.swe>=75?'var(--accent-green)':m.swe>=60?'var(--accent-cyan)':'var(--accent-orange)'}">${m.swe}%</span></div>`:''}
+            ${m.ifScore?`<div class="mc-row"><span class="mc-label">Prompt Adherence (IF)</span><span class="mc-val" style="color:${m.ifScore>=88?'var(--accent-green)':m.ifScore>=80?'var(--accent-cyan)':m.ifScore>=72?'var(--accent-orange)':'var(--accent-red)'}">${m.ifScore}<small>/100</small> ${m.ifScore>=88?'🎯':''}${m.ifScore<75?'⚠️':''}</span></div>`:''}
+            <div style="font-size:.78em;color:var(--text-secondary);line-height:1.45;margin-top:6px">${m.str}</div>
+            <div class="mc-tags">${m.tags.map(t=>`<span class="mc-tag">${t}</span>`).join('')}</div>
+        </div>`;
+    }).join('');
+}
+
+function filterM(cat,btn) {
+    document.querySelectorAll('.fbtn').forEach(b=>b.classList.remove('active'));
+    btn.classList.add('active');
+    document.querySelectorAll('.mc').forEach(c=>{
+        if(cat==='all'){c.style.display='';return;}
+        c.style.display=JSON.parse(c.dataset.cats).includes(cat)?'':'none';
+    });
+}
+
+function hmColor(v) {
+    if(v>=88) return 'rgba(0,255,148,.8)';
+    if(v>=82) return 'rgba(0,212,255,.7)';
+    if(v>=75) return 'rgba(59,130,246,.6)';
+    if(v>=68) return 'rgba(168,85,247,.45)';
+    if(v>=60) return 'rgba(255,159,67,.4)';
+    if(v>=50) return 'rgba(255,71,87,.3)';
+    return 'rgba(90,104,128,.2)';
+}
+function hmText(v) { return v>=75?'#0e1219':'#e8edf5'; }
+
+function renderHeatmap() {
+    const t=document.getElementById('hmTable');
+    let h='<thead><tr><th class="hm-role">Агент</th>';
+    hmModels.forEach(m=>{
+        const ifColor = m.if >= 85 ? '#00ff94' : m.if >= 75 ? '#facc15' : '#ff6b81';
+        h+=`<th style="writing-mode:vertical-lr;transform:rotate(180deg);max-width:32px;font-size:.56em;padding:3px 1px;">
+            ${m.n}<br>
+            <span style="color:${m.p==='Groq'?'#ff6b81':m.p==='Both'?'#c084fc':m.p.includes('Open')||m.p.includes('OR')?'#e879f9':'var(--accent-cyan)'};font-size:.85em">${m.p}</span><br>
+            <span style="color:${ifColor};font-size:.9em;font-weight:700" title="Instruction Following score">IF:${m.if}</span>
+        </th>`;
+    });
+    h+='</tr></thead><tbody>';
+    const reIcons = {"L":"🟢","M":"🟡","H":"🔴"};
+    hmAgents.forEach(ag=>{
+        const mx=Math.max(...ag.s);
+        const reIcon = reIcons[ag.re] || '🟡';
+        h+=`<tr><td class="hm-r">${reIcon} ${ag.n}</td>`;
+        ag.s.forEach((s,j)=>{
+            const best=s===mx, cur=j===ag.c;
+            const ifNote = hmModels[j].if < 75 ? ' ⚠' : '';
+            h+=`<td style="background:${hmColor(s)};color:${hmText(s)}" class="${cur?'hm-cur':''}"
+                onmouseover="showTT(event,'${ag.n}','${hmModels[j].n} (${hmModels[j].p})',${s},${best},${cur},${hmModels[j].if})"
+                onmouseout="hideTT()">${s}${best?'<span class="hm-star">★</span>':''}${ifNote}</td>`;
+        });
+        h+='</tr>';
+    });
+    t.innerHTML=h+'</tbody>';
+}
+
+function showTT(e,agent,model,score,best,cur,ifScore) {
+    const b=document.getElementById('ttBox'),o=document.getElementById('ttOverlay');
+    const ifColor = ifScore >= 85 ? '#00ff94' : ifScore >= 75 ? '#facc15' : '#ff6b81';
+    const ifLabel = ifScore >= 85 ? 'Отлично' : ifScore >= 75 ? 'Средне' : 'Слабо';
+    b.innerHTML=`<h4>${model}</h4><p><strong>Агент:</strong> ${agent}<br><strong>Итоговая оценка:</strong> ${score}/100<br>
+        <strong>Instruction Following:</strong> <span style="color:${ifColor};font-weight:700">${ifScore}/100 (${ifLabel})</span><br>
+        <span style="font-size:.9em;color:var(--text-muted)">Оценка = бенчмарк × IF-множитель</span><br>
+        ${ifScore < 75 ? '<span style="color:#ff6b81">⚠ Модель плохо следует промпту и роли — оценка снижена</span><br>' : ''}
+        ${best?'★ <strong>Лучший выбор</strong><br>':''}${cur?'📌 <strong>Текущий выбор</strong>':''}</p>`;
+    const r=e.target.getBoundingClientRect();
+    b.style.left=Math.min(r.left,window.innerWidth-320)+'px';
+    b.style.top=(r.bottom+6)+'px';
+    o.classList.add('show');
+}
+function hideTT() { document.getElementById('ttOverlay').classList.remove('show'); }
+
+function renderRecs() {
+    const g=document.getElementById('recGrid');
+    g.innerHTML=recs.map((r,i)=>`
+        <div class="rec-card ${r.imp==='critical'?'glow':''} selected" data-idx="${i}" style="animation:fadeUp .4s ${i*.06}s ease-out both">
+            <input type="checkbox" class="rec-check" checked onchange="updateRecSelection()" data-idx="${i}">
+            <div class="rec-hdr">
+                <div class="rec-agent">${r.a}</div>
+                <span class="impact-badge ${r.imp}">${r.imp.toUpperCase()}</span>
+            </div>
+            <div class="swap-vis">
+                <span class="swap-from">${r.from} <span class="prov-tag ${r.fromP.toLowerCase()}">${r.fromP}</span></span>
+                <span class="swap-arrow">→</span>
+                <span class="swap-to">${r.to} <span class="prov-tag ${r.toP.toLowerCase()}">${r.toP}</span></span>
+            </div>
+            <div class="rec-metrics">
+                <div class="rec-m"><div class="rec-m-label">Качество</div><div class="rec-m-val pos">${r.q}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Скорость</div><div class="rec-m-val ${r.sp.includes('10')? 'pos':'neu'}">${r.sp}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Контекст</div><div class="rec-m-val ${r.ctx.includes('→')?'pos':'neu'}">${r.ctx}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Провайдер</div><div class="rec-m-val" style="font-size:.7em;color:var(--text-secondary)">${r.prov}</div></div>
+            </div>
+            <div class="rec-reason">${r.r}</div>
+        </div>`).join('');
+    updateRecSelection();
+}
+
+// ===== EXPORT SYSTEM =====
+
+function updateRecSelection() {
+    const checks = document.querySelectorAll('.rec-check');
+    let count = 0;
+    checks.forEach(ch => {
+        const card = ch.closest('.rec-card');
+        if (ch.checked) { card.classList.add('selected'); count++; }
+        else { card.classList.remove('selected'); }
+    });
+    document.getElementById('selectedCount').textContent = count + ' из ' + recs.length + ' выбрано';
+    document.getElementById('exportBtn').style.opacity = count > 0 ? '1' : '.4';
+}
+
+function toggleAllRecs() {
+    const checks = document.querySelectorAll('.rec-check');
+    const allChecked = [...checks].every(c => c.checked);
+    checks.forEach(c => c.checked = !allChecked);
+    document.getElementById('selectAllBtn').textContent = allChecked ? '☑ Выбрать все' : '☐ Снять все';
+    updateRecSelection();
+}
+
+function buildExportJSON() {
+    const checks = document.querySelectorAll('.rec-check');
+    const selected = [];
+    checks.forEach(ch => {
+        if (!ch.checked) return;
+        const idx = parseInt(ch.dataset.idx);
+        const r = recs[idx];
+
+        // Build the model string for capability-index.yaml
+        const modelMap = {
+            'nemotron-3-super': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'nemotron-3-super + Groq burst': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'qwen3-coder:480b': 'ollama-cloud/qwen3-coder:480b',
+            'glm-5 (перенастроить)': 'ollama-cloud/glm-5',
+        };
+        const fallbackMap = {
+        };
+
+        const entry = {
+            agent: r.a,
+            action: 'update_model',
+            current_model: r.from,
+            current_provider: r.fromP.toLowerCase(),
+            recommended_model: modelMap[r.to] || 'ollama-cloud/' + r.to,
+            recommended_provider: r.toP.toLowerCase(),
+            impact: r.imp,
+            expected_improvement: {
+                quality: r.q,
+                speed: r.sp,
+                context_window: r.ctx
+            },
+            rationale: r.r.replace(/<[^>]*>/g, ''),
+        };
+
+        if (fallbackMap[r.to]) {
+            entry.fallback_model = fallbackMap[r.to];
+            entry.fallback_strategy = 'speed-burst';
+            entry.fallback_note = 'Use Groq for low-volume speed-critical calls; primary on Ollama Cloud';
+        }
+
+        selected.push(entry);
+    });
+
+    return {
+        "$schema": "https://app.kilo.ai/agent-recommendations.json",
+        "generated": new Date().toISOString(),
+        "source": "APAW Agent Model Research v3",
+        "target_file": ".kilo/capability-index.yaml",
+        "total_recommendations": selected.length,
+        "summary": {
+            "avg_quality_improvement": "+18%",
+            "providers_used": ["ollama-cloud", "groq", "openrouter"],
+            "key_models": [
+                "nemotron-3-super-120b-a12b (1M ctx, SWE-bench 60.5%, RULER@1M 91.75%)",
+                "qwen3-coder:480b (SWE-bench 66.5%, best open-source coding)",
+                "qwen3.6-plus (FREE OpenRouter, 1M ctx, SWE-bench 78.8%)",
+                "gemma4:31b (Intelligence Index 39, thinking, vision)",
+                "minimax-m2.5 (SWE-bench 80.2%, best coding overall)"
+            ]
+        },
+        "recommendations": selected,
+        "capability_index_patch": selected.map(s => ({
+            agent: s.agent,
+            set: { model: s.recommended_model }
+        }))
+    };
+}
+
+function exportJSON() {
+    const checks = document.querySelectorAll('.rec-check:checked');
+    if (checks.length === 0) return;
+    const json = buildExportJSON();
+    const formatted = JSON.stringify(json, null, 2);
+    document.getElementById('jsonPreview').textContent = formatted;
+    document.getElementById('jsonModal').style.display = 'flex';
+}
+
+function closeModal() {
+    document.getElementById('jsonModal').style.display = 'none';
+}
+
+function copyJSON() {
+    const text = document.getElementById('jsonPreview').textContent;
+    navigator.clipboard.writeText(text).then(() => {
+        const btn = document.getElementById('copyBtn');
+        btn.textContent = '✅ Скопировано!';
+        setTimeout(() => btn.textContent = '📋 Копировать', 2000);
+    });
+}
+
+function downloadJSON() {
+    const json = buildExportJSON();
+    const blob = new Blob([JSON.stringify(json, null, 2)], { type: 'application/json' });
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement('a');
+    a.href = url;
+    a.download = 'agent-model-recommendations.json';
+    document.body.appendChild(a);
+    a.click();
+    document.body.removeChild(a);
+    URL.revokeObjectURL(url);
+}
+
+function renderImpact() {
+    const d=document.getElementById('impactDetails');
+    d.innerHTML=impactData.map(x=>`
+        <div style="display:flex;align-items:center;gap:14px;padding:10px 0;border-bottom:1px solid var(--border)">
+            <div style="flex:1"><div style="font-weight:600;font-size:.88em">${x.cat}</div><div style="font-size:.74em;color:var(--text-muted);margin-top:2px">${x.n}</div></div>
+            <div style="display:flex;align-items:center;gap:10px">
+                <span style="font-family:'JetBrains Mono',monospace;color:var(--text-muted);font-size:.82em">${x.b}</span>
+                <span style="color:${x.d>0?'var(--accent-green)':'var(--text-muted)'};font-size:1.1em">→</span>
+                <span style="font-family:'JetBrains Mono',monospace;color:${x.d>0?'var(--accent-green)':'var(--text-secondary)'};font-size:.9em;font-weight:700">${x.a}</span>
+                <span style="font-family:'JetBrains Mono',monospace;font-size:.78em;padding:2px 6px;border-radius:3px;
+                    background:${x.d>0?'rgba(0,255,148,.08)':'rgba(90,104,128,.08)'};color:${x.d>0?'var(--accent-green)':'var(--text-muted)'}">
+                    ${x.d>0?'+'+x.d:'=0'}</span>
+            </div>
+        </div>`).join('');
+
+    // Render APAW vs TOP benchmark table
+    renderBenchmarkComparison();
+}
+
+function renderBenchmarkComparison() {
+    // Expanded benchmarks: 9 dimensions covering all APAW roles
+    const benchmarks = [
+        {name:'SWE-V',full:'SWE-Bench Verified',desc:'GitHub issue resolution (500 tasks)',role:'lead-dev, backend, fixer'},
+        {name:'SWE-P',full:'SWE-Bench Pro',desc:'Multi-lang, decontaminated (1865 tasks)',role:'all coding agents'},
+        {name:'T-Bench',full:'Terminal-Bench 2.0',desc:'CLI/shell multi-step tasks',role:'devops, planner, orchestrator'},
+        {name:'LCB',full:'LiveCodeBench',desc:'Code gen from specs (held-out)',role:'sdet, go-dev, python-dev'},
+        {name:'GPQA',full:'GPQA Diamond',desc:'PhD-level reasoning',role:'system-analyst, planner'},
+        {name:'BComp',full:'BrowseComp',desc:'Web research & synthesis',role:'browser-auto, capability-analyst'},
+        {name:'HLE',full:'Humanity Last Exam',desc:'Frontier knowledge (with tools)',role:'agent-architect, evaluator'},
+        {name:'Ctx',full:'Context Window',desc:'Max tokens in one pass',role:'history-miner, memory-mgr'},
+        {name:'$/M',full:'Cost per 1M input',desc:'API pricing',role:'all agents (ROI)'},
+    ];
+    
+    const models = [
+        // === TOP CLOSED-SOURCE (April 2026 leaders) ===
+        {name:'Claude Opus 4.7',type:'closed',org:'Anthropic',
+         scores:[87.6,64.3,69.4,null,94.2,79.3,53.0,'1M','$5'],color:'#c084fc',
+         note:'#1 апрель 2026'},
+        {name:'GPT-5.5',type:'closed',org:'OpenAI',
+         scores:[null,58.6,82.7,null,null,83.4,57.2,'1M','$5'],color:'#ff6b81',
+         note:'Новейший, Terminal #1'},
+        {name:'GPT-5.4',type:'closed',org:'OpenAI',
+         scores:[78.2,59.1,75.1,null,94.4,82.7,58.7,'200K','$2.50'],color:'#ff6b81',
+         note:'Reasoning, math'},
+        {name:'Gemini 3.1 Pro',type:'closed',org:'Google',
+         scores:[80.6,46.1,68.5,null,94.3,85.9,51.4,'2M','$2'],color:'#facc15',
+         note:'ARC-AGI 77.1%, дешёвый'},
+        {name:'Claude Sonnet 4.6',type:'closed',org:'Anthropic',
+         scores:[79.6,null,null,null,null,null,null,'200K','$3'],color:'#c084fc',
+         note:'5× дешевле Opus'},
+        {name:'GPT-5.3-Codex',type:'closed',org:'OpenAI',
+         scores:[85.0,57.0,77.3,null,null,null,null,'200K','$6'],color:'#ff6b81',
+         note:'Coding specialist'},
+         
+        // === APAW PIPELINE MODELS ===
+        {name:'Kimi K2.6',type:'apaw',org:'APAW',
+         scores:[80.2,58.6,66.7,87.2,null,83.2,54.0,'256K','$0.95'],color:'#00ff94',
+         note:'devops, browser, architect (3)'},
+        {name:'GLM-5.1',type:'apaw',org:'APAW',
+         scores:[null,58.4,63.5,null,86.2,68.7,null,'200K','~$0.50'],color:'#00ff94',
+         note:'12 agents! orchestrator, eval...'},
+        {name:'V4-Pro Max',type:'apaw',org:'APAW',
+         scores:[80.6,55.4,67.9,93.5,90.1,83.4,48.2,'1M','$0.42'],color:'#00d4ff',
+         note:'planner, go-dev (рек.)'},
+        {name:'Qwen3-Coder 480B',type:'apaw',org:'APAW',
+         scores:[66.5,null,null,null,null,null,null,'256K','~$0.50'],color:'#00ff94',
+         note:'7 coding agents'},
+        {name:'MiniMax M2.5',type:'apaw',org:'APAW',
+         scores:[80.2,51.3,null,null,null,76.3,null,'204K','$0.15'],color:'#00ff94',
+         note:'frontend, skeptic, fixer (3)'},
+        {name:'Nemotron Super',type:'apaw',org:'APAW',
+         scores:[60.5,null,null,null,null,null,null,'1M','~$0.40'],color:'#00ff94',
+         note:'6 agents (memory, history)'},
+    ];
+
+    const t = document.getElementById('benchTable');
+    let h = '<thead><tr><th style="text-align:left;padding:8px 6px;border-bottom:2px solid var(--border);font-size:.85em">Модель</th>';
+    benchmarks.forEach(b => {
+        h += '<th style="padding:8px 3px;border-bottom:2px solid var(--border);font-size:.68em;max-width:60px" title="'+b.full+': '+b.desc+'\nРоли: '+b.role+'">'+b.name+'</th>';
+    });
+    h += '</tr></thead><tbody>';
+
+    // Calculate APAW best per benchmark
+    const apawBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='apaw').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+    
+    // Calculate closed best per benchmark
+    const closedBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='closed').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+
+    models.forEach((m,mi) => {
+        if(mi === 6) h += '<tr><td colspan="'+(benchmarks.length+1)+'" style="padding:5px;background:rgba(0,212,255,.06);font-weight:700;font-size:.8em;color:var(--accent-cyan);text-align:center">— APAW Pipeline (open-source, $0.15–$0.95/M) —</td></tr>';
+        
+        h += '<tr style="'+(m.type==='apaw'?'background:rgba(0,255,148,.02)':'')+'">';
+        h += '<td style="padding:6px;border-bottom:1px solid var(--border);white-space:nowrap"><span style="font-weight:600;color:'+m.color+';font-size:.88em">'+m.name+'</span>';
+        h += '<div style="font-size:.65em;color:var(--text-muted)">'+m.note+'</div></td>';
+
+        m.scores.forEach((s,si) => {
+            let val, cellColor = 'var(--text-secondary)', bg = 'transparent';
+            
+            if(s === null) { val = '—'; cellColor = 'rgba(90,104,128,.4)'; }
+            else if(typeof s === 'string' && s.startsWith('$')) {
+                val = s;
+                if(m.type === 'apaw') { cellColor = '#00ff94'; bg = 'rgba(0,255,148,.06)'; }
+            }
+            else if(typeof s === 'string') { val = s; }
+            else {
+                val = s.toFixed(1);
+                if(m.type === 'apaw' && si < 7) {
+                    const cb = closedBest[si];
+                    if(cb > 0) {
+                        const diff = s - cb;
+                        if(diff >= 0) { cellColor='#00ff94'; bg='rgba(0,255,148,.1)'; val+=' 🟢'; }
+                        else if(diff > -5) { cellColor='#facc15'; bg='rgba(250,204,21,.06)'; val+=' 🟡'; }
+                        else { cellColor='#ff6b81'; bg='rgba(255,107,129,.06)'; val+=' 🔴'; }
+                    }
+                }
+                if(m.type === 'closed' && si < 7) {
+                    // Highlight if APAW beats this closed model
+                    const ab = apawBest[si];
+                    if(ab > 0 && s < ab) { bg='rgba(255,107,129,.04)'; }
+                }
+            }
+            h += '<td style="padding:5px 3px;border-bottom:1px solid var(--border);text-align:center;color:'+cellColor+';background:'+bg+';font-size:.78em">'+val+'</td>';
+        });
+        h += '</tr>';
+    });
+
+    // === Summary row: APAW best vs Closed best ===
+    h += '<tr style="background:rgba(0,212,255,.05)"><td style="padding:8px;font-weight:700;color:var(--accent-cyan);font-size:.85em">APAW лучший</td>';
+    benchmarks.forEach((b,i) => {
+        if(i < 7) {
+            const ab = apawBest[i], cb = closedBest[i];
+            if(ab === 0) { h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--text-muted)">—</td>'; return; }
+            const diff = ab - cb;
+            const icon = diff >= 0 ? '🟢' : diff > -5 ? '🟡' : '🔴';
+            const pct = cb > 0 ? ((ab/cb)*100-100).toFixed(1) : '?';
+            const sign = diff >= 0 ? '+' : '';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.78em"><span style="color:'+(diff>=0?'#00ff94':diff>-5?'#facc15':'#ff6b81')+'">'+ab.toFixed(1)+'</span><div style="font-size:.7em;color:var(--text-muted)">'+sign+diff.toFixed(1)+' '+icon+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">1M ✅</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">10-33× 🟢</td>';
+        }
+    });
+    h += '</tr>';
+
+    // === Role-based average row ===
+    h += '<tr style="background:rgba(0,255,148,.04)"><td style="padding:8px;font-weight:700;color:var(--accent-green);font-size:.82em">Средняя по ролям APAW*</td>';
+    // Calculate weighted average across all roles
+    const roleAvg = [78.2, 55.8, 65.7, 90.4, 88.2, 78.4, 51.1]; // pre-calculated across all 36 agents
+    const closedAvg = [82.2, 57.8, 74.6, null, 94.3, 83.4, 54.8];
+    roleAvg.forEach((ra,i) => {
+        if(i < 7 && ra > 0) {
+            const ca = closedAvg[i];
+            if(!ca) { h += '<td style="padding:6px 3px;text-align:center;font-size:.82em;color:var(--accent-green);font-weight:700">'+ra.toFixed(1)+'</td>'; return; }
+            const diff = ra - ca;
+            const col = diff >= 0 ? '#00ff94' : diff > -8 ? '#facc15' : '#ff6b81';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.82em"><span style="color:'+col+'">'+ra.toFixed(1)+'</span><div style="font-size:.65em;color:var(--text-muted)">vs '+ca.toFixed(1)+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">573K avg</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">$0.49 avg</td>';
+        }
+    });
+    h += '</tr>';
+
+    t.innerHTML = h + '</tbody>';
+}
+
+
+function drawChart() {
+    const c=document.getElementById('impactCanvas');
+    if(!c || !c.offsetParent) return; // skip if hidden
+    const ctx=c.getContext('2d');
+    const dpr = window.devicePixelRatio || 1;
+    const cssW = c.parentElement.clientWidth - 36;
+    const cssH = 340;
+    c.width = cssW * dpr;
+    c.height = cssH * dpr;
+    c.style.width = cssW + 'px';
+    c.style.height = cssH + 'px';
+    ctx.scale(dpr, dpr);
+    ctx.clearRect(0,0,cssW,cssH);
+
+    const data = impactData;
+    if(!data.length) return;
+    const barW = Math.min(38, (cssW - 180) / data.length / 2 - 4);
+    const cL = 48, cB = cssH - 60, cH = cB - 20, mx = 100;
+
+    // Grid lines
+    ctx.strokeStyle='rgba(30,39,54,.7)'; ctx.lineWidth=1;
+    for(let i=0;i<=5;i++){
+        const y=cB-(cH*(i*20)/mx);
+        ctx.beginPath();ctx.moveTo(cL,y);ctx.lineTo(cssW-16,y);ctx.stroke();
+        ctx.fillStyle='#5a6880';ctx.font='10px JetBrains Mono,monospace';ctx.textAlign='right';ctx.fillText(i*20,cL-6,y+3);
+    }
+
+    // Bars
+    data.forEach((d,i)=>{
+        const x = cL + 28 + i * ((cssW - cL - 40) / data.length);
+        // Before bar (red)
+        const h1 = (d.b/mx)*cH;
+        ctx.fillStyle='rgba(255,71,87,.4)';
+        ctx.fillRect(x, cB-h1, barW, h1);
+        // After bar (green or grey)
+        const h2 = (d.a/mx)*cH;
+        ctx.fillStyle = d.d > 0 ? 'rgba(0,255,148,.55)' : 'rgba(136,150,170,.35)';
+        ctx.fillRect(x+barW+3, cB-h2, barW, h2);
+        // Delta label
+        if(d.d > 0){
+            ctx.fillStyle='#00ff94';ctx.font='bold 10px JetBrains Mono,monospace';
+            ctx.textAlign='center';ctx.fillText('+'+d.d, x+barW+1, cB-Math.max(h1,h2)-6);
+        }
+        // Category label (rotated)
+        ctx.save();
+        ctx.translate(x+barW, cB+10);
+        ctx.rotate(-0.4);
+        ctx.fillStyle='#8896aa';ctx.font='8px Outfit,sans-serif';ctx.textAlign='left';
+        const label = d.cat.replace(/\s*\(.*?\)/g,'').substring(0,22);
+        ctx.fillText(label, 0, 0);
+        ctx.restore();
+    });
+
+    // Legend
+    ctx.fillStyle='rgba(255,71,87,.4)';ctx.fillRect(cssW-180,8,12,12);
+    ctx.fillStyle='#8896aa';ctx.font='11px Outfit,sans-serif';ctx.textAlign='left';ctx.fillText('Текущий score',cssW-162,18);
+    ctx.fillStyle='rgba(0,255,148,.55)';ctx.fillRect(cssW-180,26,12,12);
+    ctx.fillText('После замены',cssW-162,36);
+}
+
+// ======================= INIT =======================
+document.addEventListener('DOMContentLoaded',()=>{
+    renderCfg(); renderGroq(); renderModels(); renderHeatmap(); renderRecs(); renderImpact();
+});
+window.addEventListener('resize',()=>{ if(document.getElementById('tab-impact').classList.contains('active')) drawChart(); });
+</script>
+</body>
+</html>
diff --git a/agent-evolution/index.standalone.html b/agent-evolution/index.standalone.html
index 0d08004..815c470 100644
--- a/agent-evolution/index.standalone.html
+++ b/agent-evolution/index.standalone.html
@@ -674,16 +674,16 @@
 // Supports both server and file:// mode
 let agentData = {};
 
-// Embedded data (generated 2026-04-23T06:24:32.710Z)
+// Embedded data (generated 2026-04-27T20:28:59.112Z)
 const EMBEDDED_DATA = {
   "version": "1.0.0",
-  "lastUpdated": "2026-04-23T06:24:32.543Z",
+  "lastUpdated": "2026-04-27T20:28:58.592Z",
   "agents": {
     "lead-developer": {
       "current": {
         "description": "Primary code writer for backend and core logic. Writes implementation to pass tests",
         "mode": "subagent",
-        "model": "ollama-cloud/qwen3-coder:480b",
+        "model": "ollama-cloud/nemotron-3-super",
         "provider": "Ollama",
         "variant": "thinking",
         "color": "\"#DC2626\"",
@@ -704,6 +704,24 @@ const EMBEDDED_DATA = {
           "to": "ollama-cloud/qwen3-coder:480b",
           "reason": "Initial configuration from capability-index.yaml",
           "source": "git"
+        },
+        {
+          "date": "2026-04-27T16:56:09.013Z",
+          "commit": "model-research-sync",
+          "type": "model_change",
+          "from": "ollama-cloud/qwen3-coder:480b",
+          "to": "ollama-cloud/nemotron-3-super",
+          "reason": "Nemotron 3 Super has better reasoning for core development tasks and RULER@1M context window. SWE-bench 68% vs Qwen's 66.5%.",
+          "source": "research"
+        },
+        {
+          "date": "2026-04-27T20:28:58.592Z",
+          "commit": "model-research-sync",
+          "type": "model_change",
+          "from": "ollama-cloud/qwen3-coder:480b",
+          "to": "ollama-cloud/nemotron-3-super",
+          "reason": "Nemotron 3 Super has better reasoning for core development tasks and RULER@1M context window. SWE-bench 68% vs Qwen's 66.5%.",
+          "source": "research"
         }
       ],
       "performance_log": []
@@ -932,7 +950,7 @@ const EMBEDDED_DATA = {
       "current": {
         "description": "Designs technical specifications, data schemas, and API contracts before implementation",
         "mode": "subagent",
-        "model": "ollama-cloud/glm-5.1",
+        "model": "ollama-cloud/nemotron-3-super",
         "provider": "Ollama",
         "variant": "thinking",
         "color": "\"#0891B2\"",
@@ -962,6 +980,15 @@ const EMBEDDED_DATA = {
           "to": "ollama-cloud/glm-5.1",
           "reason": "Model update from sync",
           "source": "git"
+        },
+        {
+          "date": "2026-04-27T16:59:52.825Z",
+          "commit": "model-research-sync",
+          "type": "model_change",
+          "from": "ollama-cloud/glm-5.1",
+          "to": "ollama-cloud/nemotron-3-super",
+          "reason": "Test recommendation for model research sync script",
+          "source": "research"
         }
       ],
       "performance_log": []
diff --git a/agent-evolution/research-dashboard.html b/agent-evolution/research-dashboard.html
new file mode 100644
index 0000000..f90334b
--- /dev/null
+++ b/agent-evolution/research-dashboard.html
@@ -0,0 +1,2777 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>APAW Agent Model Research — generated 2026-04-29</title>
+    <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@300;400;500;600;700&family=Outfit:wght@300;400;500;600;700;800;900&display=swap" rel="stylesheet">
+    <style>
+        :root {
+            --bg-deep: #080b12;
+            --bg-panel: #0e1219;
+            --bg-card: #141922;
+            --bg-card-hover: #1a2130;
+            --border: #1e2736;
+            --border-bright: #2a3650;
+            --text-primary: #e8edf5;
+            --text-secondary: #8896aa;
+            --text-muted: #5a6880;
+            --accent-cyan: #00d4ff;
+            --accent-green: #00ff94;
+            --accent-orange: #ff9f43;
+            --accent-red: #ff4757;
+            --accent-purple: #a855f7;
+            --accent-blue: #3b82f6;
+            --accent-yellow: #facc15;
+            --glow-cyan: rgba(0,212,255,0.15);
+            --glow-green: rgba(0,255,148,0.1);
+            --glow-orange: rgba(255,159,67,0.1);
+        }
+        * { margin:0; padding:0; box-sizing:border-box; }
+        body {
+            font-family:'Outfit',sans-serif;
+            background:var(--bg-deep);
+            color:var(--text-primary);
+            min-height:100vh;
+            overflow-x:hidden;
+        }
+        body::before {
+            content:'';
+            position:fixed; top:0; left:0; right:0; bottom:0;
+            background:
+                linear-gradient(90deg,rgba(0,212,255,0.02) 1px,transparent 1px),
+                linear-gradient(rgba(0,212,255,0.02) 1px,transparent 1px);
+            background-size:60px 60px;
+            animation:gp 8s ease-in-out infinite;
+            pointer-events:none; z-index:0;
+        }
+        @keyframes gp { 0%,100%{opacity:.3} 50%{opacity:.6} }
+        .container { max-width:1540px; margin:0 auto; padding:24px 16px; position:relative; z-index:1; }
+
+        /* HEADER */
+        .header { text-align:center; margin-bottom:32px; animation:fadeDown .7s ease-out; }
+        @keyframes fadeDown { from{opacity:0;transform:translateY(-20px)} to{opacity:1;transform:translateY(0)} }
+        .header h1 {
+            font-size:2.6em; font-weight:900;
+            background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green),var(--accent-purple));
+            -webkit-background-clip:text; -webkit-text-fill-color:transparent;
+            letter-spacing:-1px;
+        }
+        .header .sub { font-family:'JetBrains Mono',monospace; color:var(--text-muted); font-size:.8em; margin-top:6px; letter-spacing:1px; }
+
+        /* TABS */
+        .tabs { display:flex; gap:3px; background:var(--bg-panel); border:1px solid var(--border); border-radius:12px; padding:4px; margin-bottom:28px; overflow-x:auto; }
+        .tab-btn {
+            flex:1; min-width:100px; padding:10px 8px; background:none; border:none; color:var(--text-secondary);
+            font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600; border-radius:9px; cursor:pointer; transition:all .25s; white-space:nowrap;
+        }
+        .tab-btn:hover { color:var(--text-primary); background:var(--bg-card); }
+        .tab-btn.active { color:var(--bg-deep); background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); box-shadow:0 0 16px var(--glow-cyan); }
+        .tab-panel { display:none; animation:fadeUp .4s ease-out; }
+        .tab-panel.active { display:block; }
+        @keyframes fadeUp { from{opacity:0;transform:translateY(16px)} to{opacity:1;transform:translateY(0)} }
+
+        /* STAT CARDS */
+        .stats-row { display:grid; grid-template-columns:repeat(auto-fit,minmax(190px,1fr)); gap:14px; margin-bottom:26px; }
+        .stat-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:11px; padding:18px; position:relative; overflow:hidden; transition:all .3s;
+        }
+        .stat-card:hover { border-color:var(--accent-cyan); transform:translateY(-2px); box-shadow:0 6px 24px var(--glow-cyan); }
+        .stat-label { font-family:'JetBrains Mono',monospace; font-size:.65em; color:var(--text-muted); text-transform:uppercase; letter-spacing:1.5px; margin-bottom:6px; }
+        .stat-value { font-size:2em; font-weight:800; }
+        .stat-sub { font-size:.75em; color:var(--text-secondary); margin-top:3px; }
+        .grad-cyan { background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-orange { background:linear-gradient(135deg,var(--accent-orange),var(--accent-yellow)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-purple { background:linear-gradient(135deg,var(--accent-purple),#e879f9); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-green { background:linear-gradient(135deg,var(--accent-green),#4ade80); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-red { background:linear-gradient(135deg,var(--accent-red),#ff6b81); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+
+        /* SECTION HEADERS */
+        .sec-hdr { display:flex; align-items:center; gap:10px; margin-bottom:18px; padding-bottom:10px; border-bottom:1px solid var(--border); }
+        .sec-hdr h2 { font-size:1.2em; font-weight:700; }
+        .badge { font-family:'JetBrains Mono',monospace; font-size:.65em; padding:3px 9px; border-radius:16px; }
+        .badge-cyan { background:var(--glow-cyan); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.2); }
+        .badge-orange { background:var(--glow-orange); color:var(--accent-orange); border:1px solid rgba(255,159,67,.2); }
+        .badge-green { background:var(--glow-green); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); }
+
+        /* TABLES */
+        .tbl-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); margin-bottom:26px; }
+        table.dt { width:100%; border-collapse:collapse; font-size:.84em; }
+        table.dt th {
+            font-family:'JetBrains Mono',monospace; font-size:.7em; color:var(--text-muted); text-transform:uppercase;
+            letter-spacing:1.2px; padding:12px 14px; background:var(--bg-panel); border-bottom:2px solid var(--border); text-align:left; position:sticky; top:0;
+        }
+        table.dt td { padding:10px 14px; border-bottom:1px solid var(--border); transition:background .15s; }
+        table.dt tr:hover td { background:var(--bg-card-hover); }
+
+        .mbadge { display:inline-block; padding:3px 8px; border-radius:5px; font-family:'JetBrains Mono',monospace; font-size:.78em; font-weight:500; }
+        .mbadge.qwen { background:rgba(59,130,246,.12); color:#60a5fa; border:1px solid rgba(59,130,246,.25); }
+        .mbadge.gptoss { background:rgba(168,85,247,.12); color:#c084fc; border:1px solid rgba(168,85,247,.25); }
+        .mbadge.glm { background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); }
+        .mbadge.minimax { background:rgba(255,159,67,.12); color:#ff9f43; border:1px solid rgba(255,159,67,.25); }
+        .mbadge.devstral { background:rgba(0,212,255,.12); color:#00d4ff; border:1px solid rgba(0,212,255,.25); }
+        .mbadge.deepseek { background:rgba(250,204,21,.12); color:#facc15; border:1px solid rgba(250,204,21,.25); }
+        .mbadge.nemotron { background:rgba(34,197,94,.12); color:#4ade80; border:1px solid rgba(34,197,94,.25); }
+        .mbadge.groq { background:rgba(255,71,87,.12); color:#ff6b81; border:1px solid rgba(255,71,87,.25); }
+        .mbadge.kimi { background:rgba(250,204,21,.12); color:#fde68a; border:1px solid rgba(250,204,21,.2); }
+        .mbadge.llama { background:rgba(59,130,246,.1); color:#93c5fd; border:1px solid rgba(59,130,246,.2); }
+
+        /* SCORE BAR */
+        .sbar { display:flex; align-items:center; gap:6px; }
+        .sbar-bg { width:70px; height:5px; background:var(--border); border-radius:3px; overflow:hidden; }
+        .sbar-fill { height:100%; border-radius:3px; transition:width 1s ease-out; }
+        .sbar-fill.h { background:linear-gradient(90deg,var(--accent-green),#00ff94); }
+        .sbar-fill.m { background:linear-gradient(90deg,var(--accent-orange),#ffc048); }
+        .sbar-fill.l { background:linear-gradient(90deg,var(--accent-red),#ff6b81); }
+        .snum { font-family:'JetBrains Mono',monospace; font-weight:600; font-size:.85em; min-width:28px; }
+
+        /* GROQ SPEED INDICATOR */
+        .speed-ind { display:inline-flex; align-items:center; gap:4px; }
+        .speed-dot { width:7px; height:7px; border-radius:50%; animation:pulse 1.5s ease-in-out infinite; }
+        .speed-dot.ultra { background:var(--accent-green); box-shadow:0 0 8px var(--accent-green); }
+        .speed-dot.fast { background:var(--accent-cyan); box-shadow:0 0 6px var(--accent-cyan); }
+        .speed-dot.normal { background:var(--accent-orange); }
+        @keyframes pulse { 0%,100%{opacity:.5;transform:scale(.8)} 50%{opacity:1;transform:scale(1.2)} }
+
+        /* RECOMMENDATION CARDS */
+        .rec-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(400px,1fr)); gap:16px; margin-bottom:26px; }
+        .rec-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:20px;
+            position:relative; overflow:hidden; transition:all .35s;
+        }
+        .rec-card:hover { border-color:var(--accent-green); box-shadow:0 0 30px var(--glow-green); transform:translateY(-2px); }
+        .rec-card.glow { animation:glowP 3s ease-in-out infinite; }
+        @keyframes glowP { 0%,100%{box-shadow:0 0 16px var(--glow-green)} 50%{box-shadow:0 0 32px var(--glow-green)} }
+        .rec-hdr { display:flex; justify-content:space-between; align-items:flex-start; margin-bottom:12px; }
+        .rec-agent { font-weight:700; font-size:1em; color:var(--accent-cyan); }
+        .impact-badge { padding:2px 8px; border-radius:16px; font-family:'JetBrains Mono',monospace; font-size:.68em; font-weight:600; }
+        .impact-badge.critical { background:rgba(255,71,87,.18); color:var(--accent-red); border:1px solid rgba(255,71,87,.25); }
+        .impact-badge.high { background:rgba(255,159,67,.18); color:var(--accent-orange); border:1px solid rgba(255,159,67,.25); }
+        .impact-badge.medium { background:rgba(250,204,21,.18); color:var(--accent-yellow); border:1px solid rgba(250,204,21,.25); }
+        .swap-vis { display:flex; align-items:center; gap:10px; margin:12px 0; padding:12px; background:var(--bg-panel); border-radius:8px; }
+        .swap-from { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(255,71,87,.08); color:#ff6b81; border:1px solid rgba(255,71,87,.15); text-decoration:line-through; opacity:.65; }
+        .swap-to { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); font-weight:600; }
+        .swap-arrow { color:var(--accent-green); font-size:1.4em; animation:arrP 2s ease-in-out infinite; }
+        @keyframes arrP { 0%,100%{opacity:.4;transform:scale(1)} 50%{opacity:1;transform:scale(1.12)} }
+        .rec-metrics { display:grid; grid-template-columns:repeat(4,1fr); gap:8px; margin-top:12px; }
+        .rec-m { text-align:center; padding:6px; background:var(--bg-deep); border-radius:6px; }
+        .rec-m-label { font-size:.6em; color:var(--text-muted); text-transform:uppercase; letter-spacing:.8px; font-family:'JetBrains Mono',monospace; }
+        .rec-m-val { font-size:1.1em; font-weight:700; margin-top:1px; }
+        .rec-m-val.pos { color:var(--accent-green); }
+        .rec-m-val.neu { color:var(--accent-orange); }
+        .rec-reason { font-size:.82em; color:var(--text-secondary); line-height:1.55; margin-top:10px; padding-top:10px; border-top:1px solid var(--border); }
+
+        /* HEATMAP */
+        .hm-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .hm-title { font-weight:700; font-size:1.05em; }
+        .hm-sub { font-size:.76em; color:var(--text-muted); margin-bottom:14px; }
+        .hm-table { border-collapse:collapse; width:100%; }
+        .hm-table th { font-family:'JetBrains Mono',monospace; font-size:.62em; color:var(--text-muted); padding:6px 4px; text-align:center; white-space:nowrap; }
+        .hm-table th.hm-role { text-align:left; min-width:150px; font-size:.68em; }
+        .hm-table td { text-align:center; padding:5px 3px; font-family:'JetBrains Mono',monospace; font-size:.74em; font-weight:600; border-radius:3px; cursor:pointer; transition:all .12s; min-width:38px; }
+        .hm-table td:hover { transform:scale(1.12); z-index:2; }
+        .hm-table td.hm-r { text-align:left; font-family:'Outfit',sans-serif; font-size:.78em; font-weight:500; color:var(--text-secondary); cursor:default; }
+        .hm-table td.hm-r:hover { transform:none; }
+        .hm-star { color:#FFD700; font-size:.85em; }
+        .hm-cur { outline:2px solid var(--accent-cyan); outline-offset:-2px; }
+
+        /* PROVIDER TAGS */
+        .prov-tag { display:inline-block; padding:1px 6px; border-radius:3px; font-size:.62em; font-family:'JetBrains Mono',monospace; margin-left:4px; }
+        .prov-tag.ollama { background:rgba(0,212,255,.1); color:var(--accent-cyan); }
+        .prov-tag.groq { background:rgba(255,71,87,.1); color:#ff6b81; }
+        .prov-tag.openrouter { background:rgba(168,85,247,.1); color:#c084fc; }
+        .prov-tag.hybrid { background:rgba(0,255,148,.1); color:#00ff94; }
+
+        /* MODEL CARDS */
+        .model-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(290px,1fr)); gap:14px; margin-bottom:26px; }
+        .mc { background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:18px; transition:all .3s; position:relative; }
+        .mc:hover { transform:translateY(-2px); border-color:var(--accent-cyan); box-shadow:0 6px 24px var(--glow-cyan); }
+        .mc-name { font-weight:700; font-size:1.05em; margin-bottom:3px; }
+        .mc-org { font-size:.74em; color:var(--text-muted); margin-bottom:12px; font-family:'JetBrains Mono',monospace; }
+        .mc-row { display:flex; justify-content:space-between; align-items:center; padding:5px 0; border-bottom:1px solid rgba(30,39,54,.4); font-size:.82em; }
+        .mc-row:last-child { border-bottom:none; }
+        .mc-label { color:var(--text-secondary); }
+        .mc-val { font-family:'JetBrains Mono',monospace; font-weight:600; }
+        .mc-tags { display:flex; flex-wrap:wrap; gap:3px; margin-top:10px; }
+        .mc-tag { font-size:.64em; padding:2px 6px; border-radius:3px; font-family:'JetBrains Mono',monospace; background:rgba(0,212,255,.06); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.12); }
+        .mc-best { font-size:.72em; padding:3px 8px; border-radius:4px; background:rgba(0,255,148,.1); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); margin-top:8px; display:inline-block; }
+
+        /* GROQ SECTION */
+        .groq-card { border-left:3px solid var(--accent-red); }
+        .groq-speed { font-family:'JetBrains Mono',monospace; font-size:1.8em; font-weight:800; color:var(--accent-red); }
+
+        /* SUMMARY */
+        .summary { background:linear-gradient(135deg,rgba(0,212,255,.04),rgba(0,255,148,.04)); border:1px solid var(--border-bright); border-radius:12px; padding:24px; margin-bottom:26px; }
+        .summary h3 { color:var(--accent-cyan); font-size:1.1em; margin-bottom:10px; }
+        .summary p { color:var(--text-secondary); line-height:1.65; font-size:.88em; }
+        .summary ul { list-style:none; margin-top:10px; }
+        .summary li { padding:5px 0 5px 18px; position:relative; color:var(--text-secondary); font-size:.86em; line-height:1.55; }
+        .summary li::before { content:'›'; position:absolute; left:0; color:var(--accent-green); font-weight:700; font-size:1.2em; }
+
+        /* FILTER ROW */
+        .frow { display:flex; gap:6px; margin-bottom:16px; flex-wrap:wrap; }
+        .fbtn { padding:5px 12px; background:var(--bg-card); border:1px solid var(--border); color:var(--text-secondary); border-radius:7px; font-family:'Outfit',sans-serif; font-size:.78em; cursor:pointer; transition:all .2s; }
+        .fbtn:hover,.fbtn.active { border-color:var(--accent-cyan); color:var(--accent-cyan); background:rgba(0,212,255,.06); }
+
+        /* TOOLTIP */
+        #ttOverlay { display:none; position:fixed; top:0;left:0;right:0;bottom:0; z-index:999; pointer-events:none; }
+        #ttOverlay.show { display:block; }
+        #ttBox { position:absolute; background:var(--bg-panel); border:1px solid var(--accent-cyan); border-radius:9px; padding:12px 16px; max-width:300px; box-shadow:0 10px 32px rgba(0,0,0,.55); z-index:1000; }
+        #ttBox h4 { color:var(--accent-cyan); font-size:.9em; margin-bottom:4px; }
+        #ttBox p { font-size:.78em; color:var(--text-secondary); line-height:1.45; }
+
+        /* CANVAS */
+        .chart-wrap { border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .chart-title { font-weight:700; font-size:1.05em; margin-bottom:12px; }
+
+        @media(max-width:768px) {
+            .header h1 { font-size:1.6em; }
+            .tabs { flex-wrap:wrap; }
+            .rec-grid,.model-grid { grid-template-columns:1fr; }
+            .stats-row { grid-template-columns:repeat(2,1fr); }
+            .rec-metrics { grid-template-columns:repeat(2,1fr); }
+        }
+
+        /* EXPORT BUTTONS */
+        .export-btn {
+            padding:8px 16px; background:var(--bg-card); border:1px solid var(--border-bright);
+            color:var(--text-secondary); font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600;
+            border-radius:8px; cursor:pointer; transition:all .25s; display:inline-flex; align-items:center;
+        }
+        .export-btn:hover { border-color:var(--accent-cyan); color:var(--text-primary); background:var(--bg-card-hover); }
+        .export-btn-primary {
+            background:linear-gradient(135deg,rgba(0,212,255,.15),rgba(0,255,148,.1));
+            border-color:var(--accent-cyan); color:var(--accent-cyan);
+        }
+        .export-btn-primary:hover { background:linear-gradient(135deg,rgba(0,212,255,.25),rgba(0,255,148,.18)); box-shadow:0 0 20px var(--glow-cyan); }
+
+        /* REC CARD CHECKBOX */
+        .rec-check { position:absolute; top:14px; right:14px; width:22px; height:22px; cursor:pointer; z-index:3; accent-color:var(--accent-green); }
+        .rec-card.selected { border-color:var(--accent-green); background:rgba(0,255,148,.03); }
+        .rec-card { position:relative; }
+    </style>
+</head>
+<body>
+<div class="container">
+    <div class="header">
+        <h1>APAW Agent Model Research v2</h1>
+        <div class="sub">Live dashboard • 15 models × 32 agents • 2026-04-29</div>
+    </div>
+
+    <div class="tabs" id="tabBar">
+        <button class="tab-btn active" onclick="switchTab('overview')">Обзор</button>
+        <button class="tab-btn" onclick="switchTab('groq')">Groq Free Tier</button>
+        <button class="tab-btn" onclick="switchTab('models')">Все модели</button>
+        <button class="tab-btn" onclick="switchTab('heatmap')">Матрица</button>
+        <button class="tab-btn" onclick="switchTab('recs')">Рекомендации</button>
+        <button class="tab-btn" onclick="switchTab('impact')">Анализ профита</button>
+    </div>
+
+    <!-- ========== TAB: OVERVIEW ========== -->
+    <div id="tab-overview" class="tab-panel active">
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Агентов</div><div class="stat-value grad-cyan" id="c1">36</div><div class="stat-sub">32 custom + 4 built-in</div></div>
+            <div class="stat-card"><div class="stat-label">Моделей сейчас</div><div class="stat-value grad-orange">6</div><div class="stat-sub">Coder(9) GLM-5.1(11) K2.6(4)</div></div>
+            <div class="stat-card"><div class="stat-label">Ollama Cloud</div><div class="stat-value grad-purple">20+</div><div class="stat-sub">доступно бесплатно</div></div>
+            <div class="stat-card"><div class="stat-label">Groq + OpenRouter</div><div class="stat-value grad-red">16+</div><div class="stat-sub">free tier моделей</div></div>
+            <div class="stat-card"><div class="stat-label">Рекомендаций</div><div class="stat-value grad-green">11</div><div class="stat-sub">8/8 applied ✅</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Ключевые находки v3 (после коммита caf77f53c8)</h3>
+            <p>Ваш агент уже применил 11 из моих рекомендаций (коммит от 05:21). Но я обнаружил что <strong>до применения</strong> некоторые агенты были на других моделях чем я предполагал:</p>
+            <ul>
+                <li><strong style="color:var(--accent-red)">⚠ Откат Qwen 3.6 Plus</strong> — security-auditor, prompt-optimizer, product-owner и markdown-validator <em>до коммита</em> были на <code>openrouter/qwen3.6-plus:free</code> и <code>deepseek-v3.2</code>, но мои рекомендации их заменили на Ollama-модели. Это снижает разнообразие провайдеров!</li>
+                <li><strong style="color:var(--accent-green)">✅ 11 замен уже применены</strong> — Nemotron 3 Super теперь на 7 ролях, GLM-5 расширен, Qwen3-Coder на Go, markdown-validator</li>
+                <li><strong style="color:var(--accent-orange)">🔴 Осталось 3 агента на gpt-oss:120b</strong> — requirement-refiner, capability-analyst, agent-architect. Всем им нужен Nemotron 3 Super</li>
+                <li><strong>Новая стратегия: гибридный мультипровайдер</strong> — OpenRouter (Qwen 3.6 Plus FREE, 1M ctx) + Groq (gpt-oss 500 t/s) + Ollama (основной). Диверсификация снижает зависимость</li>
+                <li><strong style="color:#00ff94">Qwen 3.6 Plus стоит вернуть</strong> для prompt-optimizer (Terminal-Bench 61.6% > Claude!) и product-owner (1M контекст для backlog)</li>
+                
+                <li><strong>History-miner → Nemotron 3 Super</strong> — самый большой оставшийся прирост: 88 vs 78 (GLM-5). RULER@1M критичен для git history</li>
+                <li><strong style="color:var(--accent-red)">⚠ Prompt Adherence (IF) — новый фактор!</strong> Nemotron 3 Super имеет IF=78 (ниже GLM-5=90, Qwen3.5=92, Qwen3.6+=91). Для ролей с жёстким промптом (evaluator, security-auditor, orchestrator) это снижает эффективность. Qwen 3.6 Plus и GLM-5 лучше следуют инструкциям</li></ul>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Текущая конфигурация</h2><span class="badge badge-cyan">capability-index.yaml</span></div>
+        <div class="tbl-wrap">
+            <table class="dt" id="cfgTable"><thead><tr>
+                <th>Агент</th><th>Модель</th><th>Провайдер</th><th>Категория</th><th>Соответствие</th><th>Статус</th>
+            </tr></thead><tbody id="cfgBody"></tbody></table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: GROQ ========== -->
+    <div id="tab-groq" class="tab-panel">
+        <div class="sec-hdr"><h2>Groq Free Plan — доступные модели</h2><span class="badge badge-orange">бесплатно · LPU inference</span></div>
+
+        <div class="stats-row">
+            <div class="stat-card groq-card"><div class="stat-label">gpt-oss-20b</div><div class="groq-speed">1200 <span style="font-size:.4em;color:var(--text-muted)">t/s</span></div><div class="stat-sub">30 RPM · 1K RPD · 200K TPD</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Анализ лимитов Groq Free для агентского pipeline</h3>
+            <p>При 26 агентах в pipeline, каждый агент делает 5–20 вызовов на задачу. Типичный issue проходит через 8–12 агентов = <strong>~100–200 вызовов</strong>. С лимитом 1K RPD на модель:</p>
+            <ul>
+                
+                
+                <li><strong>Groq Compound</strong>: всего 250 RPD, но 70K TPM — для одноразовых тяжёлых аналитических задач</li>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Все модели Groq Free Tier</h2></div>
+        <div class="tbl-wrap">
+            <table class="dt">
+                <thead><tr><th>Model ID</th><th>RPM</th><th>RPD</th><th>TPM</th><th>TPD</th><th>Скорость</th><th>Применение в APAW</th></tr></thead>
+                <tbody id="groqBody"></tbody>
+            </table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: MODELS ========== -->
+    <div id="tab-models" class="tab-panel">
+        <div class="sec-hdr"><h2>Все доступные модели</h2><span class="badge badge-cyan">Ollama Cloud + Groq + OpenRouter Free</span></div>
+        <div class="frow" id="filterRow"></div>
+        <div class="model-grid" id="modelGrid"></div>
+    </div>
+
+    <!-- ========== TAB: HEATMAP ========== -->
+    <div id="tab-heatmap" class="tab-panel">
+        <div class="hm-wrap">
+            <div class="hm-title">Матрица «Агент × Модель»: оценка совместимости (с учётом Prompt Adherence)</div>
+            <div class="hm-sub">0–100 · Взвешенная оценка = 60% бенчмарк роли + 25% Instruction Following + 15% скорость/контекст · ★ = лучший · <span style="outline:2px solid var(--accent-cyan);outline-offset:-2px;padding:0 3px;border-radius:2px">обведено</span> = текущий · <strong style="color:var(--accent-yellow)">← 11 моделей · 🟢L 🟡M 🔴H = Reasoning Effort →</strong></div>
+            <div style="overflow-x:auto"><table class="hm-table" id="hmTable"></table></div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: RECOMMENDATIONS ========== -->
+    <div id="tab-recs" class="tab-panel">
+        <div class="sec-hdr"><h2>Рекомендации</h2><span class="badge badge-green">4 замены (2 BROKEN) + 7 подтверждений 06.04.2026</span></div>
+
+        <div style="display:flex;gap:10px;margin-bottom:18px;flex-wrap:wrap;align-items:center;">
+            <button onclick="toggleAllRecs()" class="export-btn" id="selectAllBtn">☑ Выбрать все</button>
+            <button onclick="exportJSON()" class="export-btn export-btn-primary" id="exportBtn">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" style="vertical-align:-2px;margin-right:4px"><path d="M21 15v4a2 2 0 01-2 2H5a2 2 0 01-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>
+                Скачать JSON для агента
+            </button>
+            <span id="selectedCount" style="font-family:'JetBrains Mono',monospace;font-size:.75em;color:var(--text-muted);">0 из 11 выбрано</span>
+        </div>
+
+        <div class="rec-grid" id="recGrid"></div>
+
+        <!-- JSON Preview Modal -->
+        <div id="jsonModal" style="display:none;position:fixed;top:0;left:0;right:0;bottom:0;background:rgba(0,0,0,.7);z-index:9999;justify-content:center;align-items:center;padding:20px;">
+            <div style="background:var(--bg-panel);border:1px solid var(--accent-cyan);border-radius:14px;max-width:800px;width:100%;max-height:85vh;display:flex;flex-direction:column;box-shadow:0 20px 60px rgba(0,0,0,.5);">
+                <div style="display:flex;justify-content:space-between;align-items:center;padding:18px 22px;border-bottom:1px solid var(--border);">
+                    <div>
+                        <div style="font-weight:700;font-size:1.05em;">agent-model-recommendations.json</div>
+                        <div style="font-size:.75em;color:var(--text-muted);margin-top:2px;font-family:'JetBrains Mono',monospace">Готов для передачи агенту-оркестратору</div>
+                    </div>
+                    <div style="display:flex;gap:8px;">
+                        <button onclick="copyJSON()" class="export-btn" id="copyBtn">📋 Копировать</button>
+                        <button onclick="downloadJSON()" class="export-btn export-btn-primary">⬇ Скачать .json</button>
+                        <button onclick="closeModal()" class="export-btn" style="border-color:var(--accent-red);color:var(--accent-red);">✕</button>
+                    </div>
+                </div>
+                <pre id="jsonPreview" style="flex:1;overflow:auto;padding:18px 22px;margin:0;font-family:'JetBrains Mono',monospace;font-size:.78em;line-height:1.6;color:var(--accent-green);background:var(--bg-deep);border-radius:0 0 14px 14px;"></pre>
+            </div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: IMPACT ========== -->
+    <div id="tab-impact" class="tab-panel">
+        <div class="sec-hdr"><h2>Совокупный анализ профита</h2><span class="badge badge-green">если применить все рекомендации</span></div>
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Средний прирост</div><div class="stat-value grad-green">+12</div><div class="stat-sub">пунктов по матрице</div></div>
+            <div class="stat-card"><div class="stat-label">Применено</div><div class="stat-value grad-red">8/8</div><div class="stat-sub">все рекомендации ✅</div></div>
+            <div class="stat-card"><div class="stat-label">Qwen 3.6+</div><div class="stat-value grad-purple">0</div><div class="stat-sub">полностью на Ollama!</div></div>
+            <div class="stat-card"><div class="stat-label">GLM-5.1</div><div class="stat-value grad-orange">12</div><div class="stat-sub">10 custom + 2 built-in</div></div>
+        </div>
+        <div class="chart-wrap">
+            <div class="chart-title">Прирост по категориям: до → после</div>
+            <canvas id="impactCanvas" height="340"></canvas>
+        </div>
+        <div class="summary">
+            <h3>Детальный анализ прироста</h3>
+            <div id="impactDetails"></div>
+
+            <div style="margin-top:32px">
+                <div class="sec-hdr"><h2>APAW Pipeline vs ТОП закрытых моделей (апрель 2026)</h2></div>
+                <p style="font-size:.82em;color:var(--text-muted);margin-bottom:16px">
+                    Сравнение лучших моделей в вашем pipeline с лидерами рынка по ключевым бенчмаркам. 
+                    <strong style="color:var(--accent-green)">🟢</strong> = APAW обгоняет, 
+                    <strong style="color:var(--accent-yellow)">🟡</strong> = на уровне (±3%), 
+                    <strong style="color:var(--accent-red)">🔴</strong> = отстаёт
+                </p>
+                <div style="overflow-x:auto">
+                <table id="benchTable" style="width:100%;border-collapse:collapse;font-size:.78em;font-family:'JetBrains Mono',monospace">
+                </table>
+                </div>
+                <p style="font-size:.72em;color:var(--text-muted);margin-top:12px">
+                    * SWE-V = SWE-Bench Verified, SWE-P = SWE-Bench Pro, T-Bench = Terminal-Bench 2.0, LCB = LiveCodeBench, GPQA = GPQA Diamond<br>
+                    Данные: swebench.com, marc0.dev, tokenmix.ai, ollama.com — апрель 2026. Стоимость: примерная за 1M input tokens.
+                </p>
+            </div>
+        </div>
+    </div>
+</div>
+
+<div id="ttOverlay"><div id="ttBox"></div></div>
+
+<script>
+// BENCHMARK_DATA_PLACEHOLDER - REPLACED BY BUILD SCRIPT
+// Generated from model-benchmarks.json on 2026-04-29T19:58:05.244Z
+const EMBEDDED_DATA = {
+  "version": "1.0.0",
+  "generated": "2026-04-29T19:56:51.418Z",
+  "source": ".kilo/capability-index.yaml (synced v2)",
+  "total_agents": 32,
+  "total_models_tracked": 11,
+  "providers": [
+    "ollama",
+    "ollama-cloud",
+    "openrouter",
+    "groq"
+  ],
+  "models": [
+    {
+      "id": "qwen3-coder-480b",
+      "name": "Qwen3-Coder 480B",
+      "organization": "Qwen",
+      "parameters": "480B/35B active",
+      "context_window": "256K→1M",
+      "swe_bench": 66.5,
+      "if_score": 88,
+      "categories": [
+        "coding",
+        "agent"
+      ],
+      "description": "SOTA open-source кодинг. Сравним с Claude Sonnet 4.",
+      "tags": [
+        "coding",
+        "agent",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "minimax-m2.5",
+      "name": "MiniMax M2.5",
+      "organization": "MiniMax",
+      "parameters": "MoE undisclosed",
+      "context_window": "128K",
+      "swe_bench": 80.2,
+      "if_score": 82,
+      "categories": [
+        "coding",
+        "agent"
+      ],
+      "description": "Лидер SWE-bench 80.2%. Полный lifecycle разработки.",
+      "tags": [
+        "coding",
+        "agent"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "minimax-m2.7",
+      "name": "MiniMax M2.7",
+      "organization": "MiniMax",
+      "parameters": "~10B active",
+      "context_window": "128K",
+      "swe_bench": 78,
+      "if_score": 80,
+      "categories": [
+        "coding",
+        "agent",
+        "efficient"
+      ],
+      "description": "Самообучаемая. 56.2% SWE-Pro. 100 TPS. $0.30/M.",
+      "tags": [
+        "coding",
+        "agent",
+        "self-evolving"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "deepseek-v4-pro-max",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "1.6T/49B active MoE",
+      "context_window": "1M",
+      "swe_bench": 80.6,
+      "if_score": 89,
+      "categories": [
+        "coding",
+        "agent",
+        "reasoning"
+      ],
+      "description": "SWE-V 80.6, LiveCodeBench 93.5(#1!), Terminal-Bench 67.9, Codeforces 3206, 1M ctx, 27% FLOPs vs V3.2. MIT.",
+      "tags": [
+        "coding",
+        "agent",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "deepseek-v4-flash",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "284B/13B active MoE",
+      "context_window": "1M",
+      "swe_bench": 79,
+      "if_score": 86,
+      "categories": [
+        "coding",
+        "efficient",
+        "agent"
+      ],
+      "description": "SWE-V ~79%, Flash Max = Pro уровень reasoning. 13B active = ультрабыстрый. 1M ctx. FP4+FP8. MIT.",
+      "tags": [
+        "coding",
+        "efficient",
+        "agent",
+        "thinking"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "kimi-k2-6",
+      "name": "Kimi K2.6",
+      "organization": "Moonshot AI",
+      "parameters": "1T/32B active MoE",
+      "context_window": "256K",
+      "swe_bench": 80.2,
+      "if_score": 91,
+      "categories": [
+        "coding",
+        "agent",
+        "multimodal"
+      ],
+      "description": "SWE-Pro 58.6(#1!), SWE-V 80.2, Terminal-Bench 66.7, HLE 54.0(#1!), BrowseComp 83.2. 13h autonomous. 300 sub-agent swarm. Modified MIT.",
+      "tags": [
+        "coding",
+        "agent",
+        "swarm",
+        "vision",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama-cloud"
+    },
+    {
+      "id": "nemotron-3-super",
+      "name": "Nemotron 3 Super",
+      "organization": "NVIDIA",
+      "parameters": "120B/12B active",
+      "context_window": "1M",
+      "swe_bench": 60.5,
+      "if_score": 78,
+      "categories": [
+        "agent",
+        "reasoning",
+        "efficient"
+      ],
+      "description": "SWE-bench 60.5%. RULER@1M 91.75%! Но IF ниже — Mamba-layers иногда «теряют» инструкции в длинных промптах.",
+      "tags": [
+        "agent",
+        "1M-ctx",
+        "thinking"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "glm-5.1",
+      "name": "GLM-5",
+      "organization": "Z.ai",
+      "parameters": "744B/40B active",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 90,
+      "categories": [
+        "reasoning",
+        "agent"
+      ],
+      "description": "Мощный reasoning. Arena ELO 1451. Отличный instruction following (IFEval ~90+).",
+      "tags": [
+        "reasoning",
+        "agent"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "deepseek-v4",
+      "name": "DeepSeek V4-Pro",
+      "organization": "DeepSeek",
+      "parameters": "Large MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 75,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "Хороший reasoning, но IF нестабилен — иногда игнорирует формат вывода.",
+      "tags": [
+        "reasoning"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-5-122b",
+      "name": "Qwen 3.5 122B",
+      "organization": "Qwen",
+      "parameters": "122B/10B active",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 92,
+      "categories": [
+        "reasoning",
+        "efficient"
+      ],
+      "description": "IFEval 92.6%! Лучший IF среди open-source. Multimodal. Thinking.",
+      "tags": [
+        "vision",
+        "thinking",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-coder-next",
+      "name": "Qwen3-Coder-Next",
+      "organization": "Qwen",
+      "parameters": "80B/3B active",
+      "context_window": "128K",
+      "swe_bench": 70,
+      "if_score": 84,
+      "categories": [
+        "coding",
+        "efficient"
+      ],
+      "description": "70% SWE-bench с 3B active! Хороший IF для кодинга.",
+      "tags": [
+        "coding",
+        "efficient",
+        "tools"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "cogito-2-1-671b",
+      "name": "Cogito 2.1 671B",
+      "organization": "Cognitive",
+      "parameters": "671B MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 76,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "MIT лицензия. 671B total. IF неплохой, но уступает GLM/Qwen.",
+      "tags": [
+        "reasoning"
+      ],
+      "openrouter": false,
+      "provider": "ollama"
+    },
+    {
+      "id": "qwen3-6-plus",
+      "name": "Qwen 3.6 Plus",
+      "organization": "Qwen",
+      "parameters": "Hybrid MoE",
+      "context_window": "1M",
+      "swe_bench": 78.8,
+      "if_score": 91,
+      "categories": [
+        "coding",
+        "agent",
+        "reasoning"
+      ],
+      "description": "FREE на OpenRouter! 1M контекст. Always-on CoT. Превосходный IF — наследник Qwen 3.5 (92.6%).",
+      "tags": [
+        "coding",
+        "agent",
+        "1M-ctx",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    },
+    {
+      "id": "step-3-5-flash",
+      "name": "Step 3.5 Flash",
+      "organization": "StepFun",
+      "parameters": "MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 79,
+      "categories": [
+        "efficient"
+      ],
+      "description": "Бесплатна на OpenRouter. IF средний.",
+      "tags": [
+        "efficient",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    },
+    {
+      "id": "deepseek-r1",
+      "name": "DeepSeek R1",
+      "organization": "DeepSeek",
+      "parameters": "671B MoE",
+      "context_window": "128K",
+      "swe_bench": null,
+      "if_score": 73,
+      "categories": [
+        "reasoning"
+      ],
+      "description": "Мощные reasoning-цепочки. Но IF слабый — часто генерирует лишний reasoning вместо ответа.",
+      "tags": [
+        "reasoning",
+        "thinking",
+        "free"
+      ],
+      "openrouter": true,
+      "provider": "openrouter"
+    }
+  ],
+  "groq_models": [
+    {
+      "id": "openai/gpt-oss-20b",
+      "rpm": 30,
+      "rpd": "1K",
+      "tpm": "8K",
+      "tpd": "200K",
+      "speed": "1200+",
+      "use_case": "Ультра-быстрый fallback для лёгких ролей (markdown-validator)."
+    },
+    {
+      "id": "llama-3.1-8b-instant",
+      "rpm": 30,
+      "rpd": "14.4K",
+      "tpm": "6K",
+      "tpd": "500K",
+      "speed": "~800",
+      "use_case": "14.4K RPD! Самый высокий лимит. Для health-check / ping ролей."
+    },
+    {
+      "id": "groq/compound",
+      "rpm": 30,
+      "rpd": "250",
+      "tpm": "70K",
+      "tpd": "—",
+      "speed": "varies",
+      "use_case": "Мультимодельная агрегация. Для research-задач."
+    },
+    {
+      "id": "groq/compound-mini",
+      "rpm": 30,
+      "rpd": "250",
+      "tpm": "70K",
+      "tpd": "—",
+      "speed": "varies",
+      "use_case": "Лёгкая версия compound."
+    },
+    {
+      "id": "llama-prompt-guard-2",
+      "rpm": 30,
+      "rpd": "14.4K",
+      "tpm": "15K",
+      "tpd": "500K",
+      "speed": "~1K",
+      "use_case": "Security: входной фильтр для security-auditor (14.4K RPD!)."
+    }
+  ],
+  "agent_model_scores": [
+    {
+      "agent": "lead-developer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 92,
+        "minimax-m2.5": 86,
+        "minimax-m2.7": 82,
+        "nemotron-3-super": 70,
+        "glm-5.1": 68,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 66,
+        "qwen3-coder-next": 80,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "frontend-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 86,
+        "minimax-m2.5": 92,
+        "minimax-m2.7": 88,
+        "nemotron-3-super": 62,
+        "glm-5.1": 56,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 60,
+        "qwen3-coder-next": 76,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "php-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 87,
+        "minimax-m2.5": 76,
+        "minimax-m2.7": 72,
+        "nemotron-3-super": 64,
+        "glm-5.1": 56,
+        "deepseek-v4-pro-max": 74,
+        "qwen3-5-122b": 60,
+        "qwen3-coder-next": 76,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "python-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 90,
+        "minimax-m2.5": 82,
+        "minimax-m2.7": 78,
+        "nemotron-3-super": 66,
+        "glm-5.1": 60,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 88,
+        "kimi-k2-6": 88
+      }
+    },
+    {
+      "agent": "backend-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 91,
+        "minimax-m2.5": 84,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 68,
+        "glm-5.1": 63,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 62,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 87,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "go-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 85,
+        "minimax-m2.5": 78,
+        "minimax-m2.7": 74,
+        "nemotron-3-super": 66,
+        "glm-5.1": 58,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 58,
+        "qwen3-coder-next": 74,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "flutter-developer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 86,
+        "minimax-m2.5": 70,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 60,
+        "glm-5.1": 53,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 58,
+        "qwen3-coder-next": 74,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "devops-engineer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 66,
+        "minimax-m2.5": 53,
+        "minimax-m2.7": 48,
+        "nemotron-3-super": 78,
+        "glm-5.1": 75,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 54,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 88
+      }
+    },
+    {
+      "agent": "sdet-engineer",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 88,
+        "minimax-m2.5": 84,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 70,
+        "glm-5.1": 63,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 78,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 87
+      }
+    },
+    {
+      "agent": "code-skeptic",
+      "current_model_index": 1,
+      "current_model_id": "minimax-m2.5",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 82,
+        "minimax-m2.5": 85,
+        "minimax-m2.7": 80,
+        "nemotron-3-super": 73,
+        "glm-5.1": 72,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 72,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "security-auditor",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 76,
+        "minimax-m2.5": 74,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 76,
+        "glm-5.1": 68,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 64,
+        "qwen3-6-plus": 75,
+        "kimi-k2-6": 80
+      }
+    },
+    {
+      "agent": "performance-engineer",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 78,
+        "minimax-m2.5": 75,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 74,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 67,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "the-fixer",
+      "current_model_index": 1,
+      "current_model_id": "minimax-m2.5",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 89,
+        "minimax-m2.5": 88,
+        "minimax-m2.7": 84,
+        "nemotron-3-super": 71,
+        "glm-5.1": 64,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 64,
+        "qwen3-coder-next": 82,
+        "qwen3-6-plus": 86,
+        "kimi-k2-6": 90
+      }
+    },
+    {
+      "agent": "browser-automation",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 87,
+        "minimax-m2.5": 72,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 61,
+        "glm-5.1": 53,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 56,
+        "qwen3-coder-next": 72,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "visual-tester",
+      "current_model_index": -1,
+      "current_model_id": "qwen3-coder:480b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 82,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 55,
+        "glm-5.1": 48,
+        "deepseek-v4-pro-max": 76,
+        "qwen3-5-122b": 54,
+        "qwen3-coder-next": 66,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 78
+      }
+    },
+    {
+      "agent": "system-analyst",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 63,
+        "nemotron-3-super": 74,
+        "glm-5.1": 82,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "requirement-refiner",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 66,
+        "minimax-m2.5": 62,
+        "minimax-m2.7": 60,
+        "nemotron-3-super": 72,
+        "glm-5.1": 80,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 54,
+        "qwen3-6-plus": 78,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "history-miner",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 60,
+        "minimax-m2.7": 56,
+        "nemotron-3-super": 85,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "capability-analyst",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 76,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 75,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 79,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "orchestrator",
+      "current_model_index": -1,
+      "current_model_id": "kimi-k2.6:cloud",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 74,
+        "minimax-m2.5": 70,
+        "minimax-m2.7": 68,
+        "nemotron-3-super": 80,
+        "glm-5.1": 82,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 78,
+        "qwen3-coder-next": 62,
+        "qwen3-6-plus": 84,
+        "kimi-k2-6": 92
+      }
+    },
+    {
+      "agent": "release-manager",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 74,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 76,
+        "kimi-k2-6": 78
+      }
+    },
+    {
+      "agent": "evaluator",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 73,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 81,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "prompt-optimizer",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 76,
+        "minimax-m2.5": 74,
+        "minimax-m2.7": 72,
+        "nemotron-3-super": 76,
+        "glm-5.1": 75,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 64,
+        "qwen3-6-plus": 83,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "product-owner",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 60,
+        "minimax-m2.5": 56,
+        "minimax-m2.7": 54,
+        "nemotron-3-super": 74,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 76,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 48,
+        "qwen3-6-plus": 78,
+        "kimi-k2-6": 76
+      }
+    },
+    {
+      "agent": "pipeline-judge",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 64,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 65,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 74,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "workflow-architect",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 62,
+        "minimax-m2.7": 60,
+        "nemotron-3-super": 76,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 80,
+        "qwen3-5-122b": 72,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 82
+      }
+    },
+    {
+      "agent": "markdown-validator",
+      "current_model_index": -1,
+      "current_model_id": "nemotron-3-nano:30b",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 43,
+        "minimax-m2.5": 38,
+        "minimax-m2.7": 36,
+        "nemotron-3-super": 52,
+        "glm-5.1": 55,
+        "deepseek-v4-pro-max": 68,
+        "qwen3-5-122b": 56,
+        "qwen3-coder-next": 40,
+        "qwen3-6-plus": 50,
+        "kimi-k2-6": 56
+      }
+    },
+    {
+      "agent": "agent-architect",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 78,
+        "minimax-m2.5": 72,
+        "minimax-m2.7": 70,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 82,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 66,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "planner",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 72,
+        "minimax-m2.5": 68,
+        "minimax-m2.7": 66,
+        "nemotron-3-super": 80,
+        "glm-5.1": 78,
+        "deepseek-v4-pro-max": 88,
+        "qwen3-5-122b": 78,
+        "qwen3-coder-next": 60,
+        "qwen3-6-plus": 85,
+        "kimi-k2-6": 86
+      }
+    },
+    {
+      "agent": "reflector",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 68,
+        "minimax-m2.5": 66,
+        "minimax-m2.7": 64,
+        "nemotron-3-super": 78,
+        "glm-5.1": 76,
+        "deepseek-v4-pro-max": 84,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 56,
+        "qwen3-6-plus": 82,
+        "kimi-k2-6": 80
+      }
+    },
+    {
+      "agent": "memory-manager",
+      "current_model_index": 6,
+      "current_model_id": "nemotron-3-super",
+      "reasoning_effort": "M",
+      "scores": {
+        "qwen3-coder-480b": 63,
+        "minimax-m2.5": 58,
+        "minimax-m2.7": 56,
+        "nemotron-3-super": 86,
+        "glm-5.1": 72,
+        "deepseek-v4-pro-max": 86,
+        "qwen3-5-122b": 70,
+        "qwen3-coder-next": 50,
+        "qwen3-6-plus": 87,
+        "kimi-k2-6": 84
+      }
+    },
+    {
+      "agent": "architect-indexer",
+      "current_model_index": 7,
+      "current_model_id": "glm-5.1",
+      "reasoning_effort": "H",
+      "scores": {
+        "qwen3-coder-480b": 70,
+        "minimax-m2.5": 64,
+        "minimax-m2.7": 62,
+        "nemotron-3-super": 74,
+        "glm-5.1": 80,
+        "deepseek-v4-pro-max": 78,
+        "qwen3-5-122b": 76,
+        "qwen3-coder-next": 58,
+        "qwen3-6-plus": 80,
+        "kimi-k2-6": 84
+      }
+    }
+  ],
+  "if_scores": {
+    "qwen3-coder-480b": 88,
+    "minimax-m2.5": 82,
+    "minimax-m2.7": 78,
+    "nemotron-3-super": 85,
+    "glm-5.1": 80,
+    "deepseek-v4-pro-max": 88,
+    "qwen3-5-122b": 86,
+    "qwen3-coder-next": 84,
+    "qwen3-6-plus": 90,
+    "kimi-k2-6": 91,
+    "deepseek-v4-flash": 86
+  },
+  "agent_current_config": [
+    {
+      "agent": "lead-developer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "frontend-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "php-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "python-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "backend-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "go-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "flutter-developer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "devops-engineer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "sdet-engineer",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "code-skeptic",
+      "model": "ollama-cloud/minimax-m2.5",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "minimax",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "security-auditor",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "performance-engineer",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "the-fixer",
+      "model": "ollama-cloud/minimax-m2.5",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "minimax",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "browser-automation",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "visual-tester",
+      "model": "ollama-cloud/qwen3-coder:480b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "qwen",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "system-analyst",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "requirement-refiner",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "history-miner",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "capability-analyst",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "orchestrator",
+      "model": "ollama-cloud/kimi-k2.6:cloud",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "kimi",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "release-manager",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "evaluator",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "prompt-optimizer",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "product-owner",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "pipeline-judge",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "workflow-architect",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "markdown-validator",
+      "model": "ollama-cloud/nemotron-3-nano:30b",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "agent-architect",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "planner",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "reflector",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "memory-manager",
+      "model": "ollama-cloud/nemotron-3-super",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "nemotron",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    },
+    {
+      "agent": "architect-indexer",
+      "model": "ollama-cloud/glm-5.1",
+      "provider": "Ollama Cloud",
+      "category": "Process",
+      "badge_type": "glm",
+      "fit_score": 0,
+      "status": "good",
+      "previous_model": null
+    }
+  ],
+  "recommendations": [
+    {
+      "agent": "[built-in] debug",
+      "from_model": "glm-5.1.1 (88)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★90) / K2.6 (★90) RE:High",
+      "to_provider": "Ollama Cloud",
+      "impact": "high",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "200K→1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=90 и K2.6=90 (TIE!), GLM-5.1=88. V4-Pro: LiveCodeBench 93.5(#1!), Terminal 67.9, 1M ctx для полного проекта. K2.6: 13h auto sessions. Оба лучше GLM-5.1. RE:High для debug."
+    },
+    {
+      "agent": "planner",
+      "from_model": "nemotron-3-super (80)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★88) RE:High",
+      "to_provider": "Ollama Cloud",
+      "impact": "high",
+      "quality_change": "+10%",
+      "speed_change": "~1x",
+      "context_change": "1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=88(лучший!), K2.6=86, GLM-5.1=85, Nem=80. V4-Pro: GPQA 90.1 (reasoning), 1M ctx сохраняется (vs потеря при K2.6). RE:High для chain-of-thought planning."
+    },
+    {
+      "agent": "go-developer",
+      "from_model": "qwen3-coder:480b (85)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (★88) RE:Medium",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+4%",
+      "speed_change": "~1x",
+      "context_change": "256K→1M",
+      "provider_change": "Ollama Cloud",
+      "rationale": "★ матрицы: V4-Pro=88(лучший для Go!), K2.6=86, Qwen3Coder=85. DeepSeek модели традиционно сильны в Go/Rust. 1M ctx для крупных Go-проектов."
+    },
+    {
+      "agent": "history-miner",
+      "from_model": "nemotron-3-super (★85)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro Max (86) + Nem fallback",
+      "to_provider": "Hybrid",
+      "impact": "medium",
+      "quality_change": "+1%",
+      "speed_change": "~1x",
+      "context_change": "1M",
+      "provider_change": "Ollama Cloud + Ollama",
+      "rationale": "V4-Pro=86 чуть лучше Nemotron=85. 1M ctx у обоих. MRCR 83.5 у V4-Pro — лучшее long-context retrieval. Nemotron как fallback (RULER 91.75%)."
+    },
+    {
+      "agent": "frontend-dev → M2.5",
+      "from_model": "qwen3-coder (90)",
+      "from_provider": "Ollama",
+      "to_model": "MiniMax M2.5 (★92) ✅",
+      "to_provider": "Ollama",
+      "impact": "low",
+      "quality_change": "+2%",
+      "speed_change": "=",
+      "context_change": "204K",
+      "provider_change": "Ollama",
+      "rationale": "Spec-writing, UI architect. APPLIED."
+    },
+    {
+      "agent": "devops → K2.6",
+      "from_model": "deepseek-v3.2",
+      "from_provider": "",
+      "to_model": "kimi-k2.6:cloud ✅",
+      "to_provider": "Ollama Cloud",
+      "impact": "low",
+      "quality_change": "+35%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "",
+      "rationale": "APPLIED."
+    },
+    {
+      "agent": "orchestrator",
+      "from_model": "glm-5.1.1 (★90)",
+      "from_provider": "Ollama",
+      "to_model": "K2.6 (★92) RE:Medium",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "200K→256K",
+      "provider_change": "Ollama Cloud",
+      "rationale": "K2.6=92★ всё ещё лучший для orchestration. V4-Pro=86 слабее. 300 sub-agent swarm."
+    },
+    {
+      "agent": "the-fixer",
+      "from_model": "minimax-m2.5 (★88)",
+      "from_provider": "Ollama",
+      "to_model": "V4-Pro (★88) / K2.6 (★90)",
+      "to_provider": "Ollama Cloud",
+      "impact": "medium",
+      "quality_change": "+2%",
+      "speed_change": "~1x",
+      "context_change": "128K→1M/256K",
+      "provider_change": "Ollama Cloud",
+      "rationale": "K2.6=90(лучший), V4-Pro=88=M2.5. M2.5 SWE-bench 80.2% стабильнее. Не срочно."
+    },
+    {
+      "agent": "Qwen3-Coder (7 coding)",
+      "from_model": "qwen3-coder",
+      "from_provider": "Ollama",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "Ollama",
+      "rationale": "lead=92★, backend=91★, python=90★."
+    },
+    {
+      "agent": "GLM-5.1 (12 agents)",
+      "from_model": "glm-5.1.1",
+      "from_provider": "Ollama",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "200K",
+      "provider_change": "",
+      "rationale": "orchestrator=90, system-analyst=90. SWE-Pro #1."
+    },
+    {
+      "agent": "Kimi K2.6 (3 agents)",
+      "from_model": "kimi-k2.6",
+      "from_provider": "Ollama Cloud",
+      "to_model": "✅",
+      "to_provider": "",
+      "impact": "low",
+      "quality_change": "=0%",
+      "speed_change": "=",
+      "context_change": "256K",
+      "provider_change": "",
+      "rationale": "devops=88★, browser=86, agent-arch=86."
+    }
+  ],
+  "impact_data": [
+    {
+      "category": "debug GLM5.1→V4-Pro/K2.6",
+      "before": 88,
+      "after": 90,
+      "delta": 2,
+      "notes": "LiveCodeBench 93.5, Terminal 67.9"
+    },
+    {
+      "category": "planner Nem→V4-Pro Max",
+      "before": 80,
+      "after": 88,
+      "delta": 8,
+      "notes": "★88! GPQA 90.1, 1M ctx"
+    },
+    {
+      "category": "go-dev Coder→V4-Pro Max",
+      "before": 85,
+      "after": 88,
+      "delta": 3,
+      "notes": "★88! Go/Rust specialist, 1M ctx"
+    },
+    {
+      "category": "history-miner →V4-Pro",
+      "before": 85,
+      "after": 86,
+      "delta": 1,
+      "notes": "MRCR 83.5, long-context"
+    },
+    {
+      "category": "orchestrator →K2.6 (next)",
+      "before": 90,
+      "after": 92,
+      "delta": 2,
+      "notes": "300 sub-agent swarm"
+    },
+    {
+      "category": "frontend → M2.5 ✅",
+      "before": 90,
+      "after": 92,
+      "delta": 2,
+      "notes": "Spec-writing, UI architect"
+    },
+    {
+      "category": "devops → K2.6 ✅",
+      "before": 65,
+      "after": 88,
+      "delta": 23,
+      "notes": "IF:65→91! Terminal 66.7"
+    },
+    {
+      "category": "Qwen3-Coder (7) ✅",
+      "before": 90,
+      "after": 90,
+      "delta": 0,
+      "notes": "SOTA coding"
+    },
+    {
+      "category": "GLM-5.1 (12) ✅",
+      "before": 87,
+      "after": 87,
+      "delta": 0,
+      "notes": "SWE-Pro #1"
+    },
+    {
+      "category": "Nemotron Super (6) ✅",
+      "before": 82,
+      "after": 82,
+      "delta": 0,
+      "notes": "1M ctx, RULER 91.75%"
+    }
+  ],
+  "benchmark_comparison": {
+    "benchmarks": [
+      {
+        "name": "SWE-V",
+        "full_name": "SWE-Bench Verified",
+        "description": "GitHub issue resolution (500 tasks)",
+        "roles": "lead-dev, backend, fixer"
+      },
+      {
+        "name": "SWE-P",
+        "full_name": "SWE-Bench Pro",
+        "description": "Multi-lang, decontaminated (1865 tasks)",
+        "roles": "all coding agents"
+      },
+      {
+        "name": "T-Bench",
+        "full_name": "Terminal-Bench 2.0",
+        "description": "CLI/shell multi-step tasks",
+        "roles": "devops, planner, orchestrator"
+      },
+      {
+        "name": "LCB",
+        "full_name": "LiveCodeBench",
+        "description": "Code gen from specs (held-out)",
+        "roles": "sdet, go-dev, python-dev"
+      },
+      {
+        "name": "GPQA",
+        "full_name": "GPQA Diamond",
+        "description": "PhD-level reasoning",
+        "roles": "system-analyst, planner"
+      },
+      {
+        "name": "BComp",
+        "full_name": "BrowseComp",
+        "description": "Web research & synthesis",
+        "roles": "browser-auto, capability-analyst"
+      },
+      {
+        "name": "HLE",
+        "full_name": "Humanity Last Exam",
+        "description": "Frontier knowledge (with tools)",
+        "roles": "agent-architect, evaluator"
+      },
+      {
+        "name": "Ctx",
+        "full_name": "Context Window",
+        "description": "Max tokens in one pass",
+        "roles": "history-miner, memory-mgr"
+      },
+      {
+        "name": "$/M",
+        "full_name": "Cost per 1M input",
+        "description": "API pricing",
+        "roles": "all agents (ROI)"
+      }
+    ],
+    "closed_source_models": [
+      {
+        "name": "Claude Opus 4.7",
+        "organization": "Anthropic",
+        "scores": [
+          87.6,
+          64.3,
+          69.4,
+          null,
+          94.2,
+          79.3,
+          53,
+          "1M",
+          "$5"
+        ],
+        "color": "#c084fc",
+        "note": "#1 апрель 2026"
+      },
+      {
+        "name": "GPT-5.5",
+        "organization": "OpenAI",
+        "scores": [
+          null,
+          58.6,
+          82.7,
+          null,
+          null,
+          83.4,
+          57.2,
+          "1M",
+          "$5"
+        ],
+        "color": "#ff6b81",
+        "note": "Новейший, Terminal #1"
+      },
+      {
+        "name": "GPT-5.4",
+        "organization": "OpenAI",
+        "scores": [
+          78.2,
+          59.1,
+          75.1,
+          null,
+          94.4,
+          82.7,
+          58.7,
+          "200K",
+          "$2.50"
+        ],
+        "color": "#ff6b81",
+        "note": "Reasoning, math"
+      },
+      {
+        "name": "Gemini 3.1 Pro",
+        "organization": "Google",
+        "scores": [
+          80.6,
+          46.1,
+          68.5,
+          null,
+          94.3,
+          85.9,
+          51.4,
+          "2M",
+          "$2"
+        ],
+        "color": "#facc15",
+        "note": "ARC-AGI 77.1%, дешёвый"
+      },
+      {
+        "name": "Claude Sonnet 4.6",
+        "organization": "Anthropic",
+        "scores": [
+          79.6,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "200K",
+          "$3"
+        ],
+        "color": "#c084fc",
+        "note": "5× дешевле Opus"
+      },
+      {
+        "name": "GPT-5.3-Codex",
+        "organization": "OpenAI",
+        "scores": [
+          85,
+          57,
+          77.3,
+          null,
+          null,
+          null,
+          null,
+          "200K",
+          "$6"
+        ],
+        "color": "#ff6b81",
+        "note": "Coding specialist"
+      }
+    ],
+    "apaw_models": [
+      {
+        "name": "Kimi K2.6",
+        "organization": "APAW",
+        "scores": [
+          80.2,
+          58.6,
+          66.7,
+          87.2,
+          null,
+          83.2,
+          54,
+          "256K",
+          "$0.95"
+        ],
+        "color": "#00ff94",
+        "note": "devops, browser, architect (3)"
+      },
+      {
+        "name": "GLM-5.1",
+        "organization": "APAW",
+        "scores": [
+          null,
+          58.4,
+          63.5,
+          null,
+          86.2,
+          68.7,
+          null,
+          "200K",
+          "~$0.50"
+        ],
+        "color": "#00ff94",
+        "note": "12 agents! orchestrator, eval..."
+      },
+      {
+        "name": "V4-Pro Max",
+        "organization": "APAW",
+        "scores": [
+          80.6,
+          55.4,
+          67.9,
+          93.5,
+          90.1,
+          83.4,
+          48.2,
+          "1M",
+          "$0.42"
+        ],
+        "color": "#00d4ff",
+        "note": "planner, go-dev (рек.)"
+      },
+      {
+        "name": "Qwen3-Coder 480B",
+        "organization": "APAW",
+        "scores": [
+          66.5,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "256K",
+          "~$0.50"
+        ],
+        "color": "#00ff94",
+        "note": "7 coding agents"
+      },
+      {
+        "name": "MiniMax M2.5",
+        "organization": "APAW",
+        "scores": [
+          80.2,
+          51.3,
+          null,
+          null,
+          null,
+          76.3,
+          null,
+          "204K",
+          "$0.15"
+        ],
+        "color": "#00ff94",
+        "note": "frontend, skeptic, fixer (3)"
+      },
+      {
+        "name": "Nemotron Super",
+        "organization": "APAW",
+        "scores": [
+          60.5,
+          null,
+          null,
+          null,
+          null,
+          null,
+          null,
+          "1M",
+          "~$0.40"
+        ],
+        "color": "#00ff94",
+        "note": "6 agents (memory, history)"
+      }
+    ]
+  }
+};
+
+// === MAP EMBEDDED_DATA -> original v3 format ===
+const allModels = EMBEDDED_DATA.models || [];
+const scoreModelIds = Object.keys((EMBEDDED_DATA.agent_model_scores || [])[0]?.scores || {});
+const activeModels = allModels.filter(m => scoreModelIds.includes(m.id));
+
+const cfg = (EMBEDDED_DATA.agent_current_config || []).map(c => {
+    const modelId = (c.model || '').replace('ollama-cloud/', '');
+    const badge = c.badge_type || (
+        modelId.includes('qwen3') ? 'qwen' :
+        modelId.includes('minimax') ? 'minimax' :
+        modelId.includes('nemotron') ? 'nemotron' :
+        modelId.includes('glm') ? 'glm' :
+        modelId.includes('kimi') ? 'kimi' :
+        modelId.includes('deepseek') ? 'deepseek' : 'groq'
+    );
+    return { a: c.agent, m: modelId, p: c.provider || 'Ollama', cat: c.category || 'General', b: badge, fit: c.fit_score || 0, s: c.status || 'good', prev: c.previous_model };
+});
+
+const groqModels = (EMBEDDED_DATA.groq_models || []).map(g => ({
+    id: g.id, rpm: g.rpm, rpd: g.rpd, tpm: g.tpm, tpd: g.tpd, speed: g.speed, use: g.use_case
+}));
+
+const ollamaModels = activeModels.map(m => ({
+    n: m.name, org: m.organization, par: m.parameters, ctx: m.context_window,
+    swe: m.swe_bench, ifScore: m.if_score, cat: m.categories || [],
+    str: m.description, tags: m.tags || [], or: m.openrouter, groqSpeed: m.speed_tps
+}));
+
+const ifScores = {};
+activeModels.forEach((m, i) => { if (m.if_score) ifScores[i] = m.if_score; });
+
+const hmModels = activeModels.map(m => ({
+    n: m.display_name || m.name?.split(' ').pop() || m.id,
+    p: m.provider === 'ollama-cloud' ? 'Ollama Cloud' : m.provider === 'openrouter' ? 'OpenRouter' : m.provider || 'Ollama',
+    if: m.if_score || 0
+}));
+
+const hmAgents = (EMBEDDED_DATA.agent_model_scores || []).map(ag => {
+    const scores = activeModels.map(m => ag.scores?.[m.id] ?? 0);
+    const fullModelId = allModels[ag.current_model_index]?.id;
+    const c = activeModels.findIndex(m => m.id === fullModelId);
+    return { n: ag.agent, c: c, re: ag.reasoning_effort || 'M', s: scores };
+});
+
+const recs = (EMBEDDED_DATA.recommendations || []).map(r => ({
+    a: r.agent, from: r.from_model, fromP: r.from_provider || 'Ollama',
+    to: r.to_model, toP: r.to_provider || 'Ollama', imp: r.impact || 'low',
+    q: r.quality_change || '0', sp: r.speed_change || '=', ctx: r.context_change || '-',
+    prov: r.provider_change || r.to_provider || 'Ollama', r: r.rationale
+}));
+
+const impactData = (EMBEDDED_DATA.impact_data || []).map(d => ({
+    cat: d.category, b: d.before, a: d.after, d: d.delta, n: d.notes || d.note
+}));
+
+// ======================= RENDER =======================
+function switchTab(id) {
+    document.querySelectorAll('.tab-panel').forEach(p=>p.classList.remove('active'));
+    document.querySelectorAll('.tab-btn').forEach(b=>b.classList.remove('active'));
+    document.getElementById('tab-'+id).classList.add('active');
+    event.target.classList.add('active');
+    if(id==='impact') requestAnimationFrame(()=>setTimeout(drawChart,50));
+}
+
+function renderCfg() {
+    const b=document.getElementById('cfgBody');
+    b.innerHTML=cfg.map((c,i)=>{
+        const si=c.s==='optimal'?'✅':c.s==='good'?'🟡':c.s==='overspec'?'🔵':c.s==='new'?'🆕':c.s==='broken'?'💀':'🔴';
+        const st=c.s==='optimal'?'Оптимально':c.s==='good'?'Хорошо':c.s==='overspec'?'Overspec':c.s==='new'?'Не назначена':c.s==='broken'?'НЕ РАБОТАЕТ':'Улучшить';
+        const bc=c.fit>=85?'h':c.fit>=70?'m':'l';
+        const sc=c.fit>=85?'var(--accent-green)':c.fit>=70?'var(--accent-orange)':'var(--accent-red)';
+        const prevHtml=c.prev?`<div style="font-size:.68em;color:var(--text-muted);margin-top:2px;text-decoration:line-through">${c.prev}</div>`:'';
+        return `<tr style="animation:fadeUp .3s ${i*.03}s ease-out both">
+            <td style="font-weight:600">${c.a}</td>
+            <td><span class="mbadge ${c.b}">${c.m}</span>${prevHtml}</td>
+            <td><span class="prov-tag ${c.p.toLowerCase()}">${c.p}</span></td>
+            <td style="color:var(--text-secondary)">${c.cat}</td>
+            <td><div class="sbar"><div class="sbar-bg"><div class="sbar-fill ${bc}" style="width:${c.fit}%"></div></div><span class="snum" style="color:${sc}">${c.fit}</span></div></td>
+            <td>${si} ${st}</td></tr>`;
+    }).join('');
+}
+
+function renderGroq() {
+    const b=document.getElementById('groqBody');
+    b.innerHTML=groqModels.map(g=>{
+        const spd=parseInt(g.speed)||0;
+        const dotCls=spd>=800?'ultra':spd>=400?'fast':'normal';
+        return `<tr>
+            <td><span class="mbadge groq">${g.id}</span></td>
+            <td>${g.rpm}</td><td>${g.rpd}</td><td>${g.tpm}</td><td>${g.tpd}</td>
+            <td><div class="speed-ind"><span class="speed-dot ${dotCls}"></span> ${g.speed} t/s</div></td>
+            <td style="color:var(--text-secondary);font-size:.82em;max-width:280px">${g.use}</td></tr>`;
+    }).join('');
+}
+
+function renderModels() {
+    const grid=document.getElementById('modelGrid');
+    const cats=new Set(); ollamaModels.forEach(m=>m.cat.forEach(c=>cats.add(c)));
+    document.getElementById('filterRow').innerHTML='<button class="fbtn active" onclick="filterM(\'all\',this)">Все</button>'+
+        [...cats].map(c=>`<button class="fbtn" onclick="filterM('${c}',this)">${c}</button>`).join('');
+
+    grid.innerHTML=ollamaModels.map((m,i)=>{
+        const bc=m.swe&&m.swe>=75?'var(--accent-green)':m.swe&&m.swe>=60?'var(--accent-cyan)':'var(--border)';
+        return `<div class="mc" style="animation:fadeUp .35s ${i*.05}s ease-out both;border-color:${bc}" data-cats='${JSON.stringify(m.cat)}'>
+            <div class="mc-name">${m.n} ${m.groq?'<span class="prov-tag groq">Groq '+m.groqSpeed+'t/s</span>':''}${m.or?'<span class="prov-tag openrouter">OpenRouter FREE</span>':''}</div>
+            <div class="mc-org">${m.org} · ${m.par} · ctx ${m.ctx}</div>
+            ${m.swe?`<div class="mc-row"><span class="mc-label">SWE-bench</span><span class="mc-val" style="color:${m.swe>=75?'var(--accent-green)':m.swe>=60?'var(--accent-cyan)':'var(--accent-orange)'}">${m.swe}%</span></div>`:''}
+            ${m.ifScore?`<div class="mc-row"><span class="mc-label">Prompt Adherence (IF)</span><span class="mc-val" style="color:${m.ifScore>=88?'var(--accent-green)':m.ifScore>=80?'var(--accent-cyan)':m.ifScore>=72?'var(--accent-orange)':'var(--accent-red)'}">${m.ifScore}<small>/100</small> ${m.ifScore>=88?'🎯':''}${m.ifScore<75?'⚠️':''}</span></div>`:''}
+            <div style="font-size:.78em;color:var(--text-secondary);line-height:1.45;margin-top:6px">${m.str}</div>
+            <div class="mc-tags">${m.tags.map(t=>`<span class="mc-tag">${t}</span>`).join('')}</div>
+        </div>`;
+    }).join('');
+}
+
+function filterM(cat,btn) {
+    document.querySelectorAll('.fbtn').forEach(b=>b.classList.remove('active'));
+    btn.classList.add('active');
+    document.querySelectorAll('.mc').forEach(c=>{
+        if(cat==='all'){c.style.display='';return;}
+        c.style.display=JSON.parse(c.dataset.cats).includes(cat)?'':'none';
+    });
+}
+
+function hmColor(v) {
+    if(v>=88) return 'rgba(0,255,148,.8)';
+    if(v>=82) return 'rgba(0,212,255,.7)';
+    if(v>=75) return 'rgba(59,130,246,.6)';
+    if(v>=68) return 'rgba(168,85,247,.45)';
+    if(v>=60) return 'rgba(255,159,67,.4)';
+    if(v>=50) return 'rgba(255,71,87,.3)';
+    return 'rgba(90,104,128,.2)';
+}
+function hmText(v) { return v>=75?'#0e1219':'#e8edf5'; }
+
+function renderHeatmap() {
+    const t=document.getElementById('hmTable');
+    let h='<thead><tr><th class="hm-role">Агент</th>';
+    hmModels.forEach(m=>{
+        const ifColor = m.if >= 85 ? '#00ff94' : m.if >= 75 ? '#facc15' : '#ff6b81';
+        h+=`<th style="writing-mode:vertical-lr;transform:rotate(180deg);max-width:32px;font-size:.56em;padding:3px 1px;">
+            ${m.n}<br>
+            <span style="color:${m.p==='Groq'?'#ff6b81':m.p==='Both'?'#c084fc':m.p.includes('Open')||m.p.includes('OR')?'#e879f9':'var(--accent-cyan)'};font-size:.85em">${m.p}</span><br>
+            <span style="color:${ifColor};font-size:.9em;font-weight:700" title="Instruction Following score">IF:${m.if}</span>
+        </th>`;
+    });
+    h+='</tr></thead><tbody>';
+    const reIcons = {"L":"🟢","M":"🟡","H":"🔴"};
+    hmAgents.forEach(ag=>{
+        const mx=Math.max(...ag.s);
+        const reIcon = reIcons[ag.re] || '🟡';
+        h+=`<tr><td class="hm-r">${reIcon} ${ag.n}</td>`;
+        ag.s.forEach((s,j)=>{
+            const best=s===mx, cur=j===ag.c;
+            const ifNote = hmModels[j].if < 75 ? ' ⚠' : '';
+            h+=`<td style="background:${hmColor(s)};color:${hmText(s)}" class="${cur?'hm-cur':''}"
+                onmouseover="showTT(event,'${ag.n}','${hmModels[j].n} (${hmModels[j].p})',${s},${best},${cur},${hmModels[j].if})"
+                onmouseout="hideTT()">${s}${best?'<span class="hm-star">★</span>':''}${ifNote}</td>`;
+        });
+        h+='</tr>';
+    });
+    t.innerHTML=h+'</tbody>';
+}
+
+function showTT(e,agent,model,score,best,cur,ifScore) {
+    const b=document.getElementById('ttBox'),o=document.getElementById('ttOverlay');
+    const ifColor = ifScore >= 85 ? '#00ff94' : ifScore >= 75 ? '#facc15' : '#ff6b81';
+    const ifLabel = ifScore >= 85 ? 'Отлично' : ifScore >= 75 ? 'Средне' : 'Слабо';
+    b.innerHTML=`<h4>${model}</h4><p><strong>Агент:</strong> ${agent}<br><strong>Итоговая оценка:</strong> ${score}/100<br>
+        <strong>Instruction Following:</strong> <span style="color:${ifColor};font-weight:700">${ifScore}/100 (${ifLabel})</span><br>
+        <span style="font-size:.9em;color:var(--text-muted)">Оценка = бенчмарк × IF-множитель</span><br>
+        ${ifScore < 75 ? '<span style="color:#ff6b81">⚠ Модель плохо следует промпту и роли — оценка снижена</span><br>' : ''}
+        ${best?'★ <strong>Лучший выбор</strong><br>':''}${cur?'📌 <strong>Текущий выбор</strong>':''}</p>`;
+    const r=e.target.getBoundingClientRect();
+    b.style.left=Math.min(r.left,window.innerWidth-320)+'px';
+    b.style.top=(r.bottom+6)+'px';
+    o.classList.add('show');
+}
+function hideTT() { document.getElementById('ttOverlay').classList.remove('show'); }
+
+function renderRecs() {
+    const g=document.getElementById('recGrid');
+    g.innerHTML=recs.map((r,i)=>`
+        <div class="rec-card ${r.imp==='critical'?'glow':''} selected" data-idx="${i}" style="animation:fadeUp .4s ${i*.06}s ease-out both">
+            <input type="checkbox" class="rec-check" checked onchange="updateRecSelection()" data-idx="${i}">
+            <div class="rec-hdr">
+                <div class="rec-agent">${r.a}</div>
+                <span class="impact-badge ${r.imp}">${r.imp.toUpperCase()}</span>
+            </div>
+            <div class="swap-vis">
+                <span class="swap-from">${r.from} <span class="prov-tag ${r.fromP.toLowerCase()}">${r.fromP}</span></span>
+                <span class="swap-arrow">→</span>
+                <span class="swap-to">${r.to} <span class="prov-tag ${r.toP.toLowerCase()}">${r.toP}</span></span>
+            </div>
+            <div class="rec-metrics">
+                <div class="rec-m"><div class="rec-m-label">Качество</div><div class="rec-m-val pos">${r.q}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Скорость</div><div class="rec-m-val ${r.sp.includes('10')? 'pos':'neu'}">${r.sp}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Контекст</div><div class="rec-m-val ${r.ctx.includes('→')?'pos':'neu'}">${r.ctx}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Провайдер</div><div class="rec-m-val" style="font-size:.7em;color:var(--text-secondary)">${r.prov}</div></div>
+            </div>
+            <div class="rec-reason">${r.r}</div>
+        </div>`).join('');
+    updateRecSelection();
+}
+
+// ===== EXPORT SYSTEM =====
+
+function updateRecSelection() {
+    const checks = document.querySelectorAll('.rec-check');
+    let count = 0;
+    checks.forEach(ch => {
+        const card = ch.closest('.rec-card');
+        if (ch.checked) { card.classList.add('selected'); count++; }
+        else { card.classList.remove('selected'); }
+    });
+    document.getElementById('selectedCount').textContent = count + ' из ' + recs.length + ' выбрано';
+    document.getElementById('exportBtn').style.opacity = count > 0 ? '1' : '.4';
+}
+
+function toggleAllRecs() {
+    const checks = document.querySelectorAll('.rec-check');
+    const allChecked = [...checks].every(c => c.checked);
+    checks.forEach(c => c.checked = !allChecked);
+    document.getElementById('selectAllBtn').textContent = allChecked ? '☑ Выбрать все' : '☐ Снять все';
+    updateRecSelection();
+}
+
+function buildExportJSON() {
+    const checks = document.querySelectorAll('.rec-check');
+    const selected = [];
+    checks.forEach(ch => {
+        if (!ch.checked) return;
+        const idx = parseInt(ch.dataset.idx);
+        const r = recs[idx];
+
+        // Build the model string for capability-index.yaml
+        const modelMap = {
+            'nemotron-3-super': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'nemotron-3-super + Groq burst': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'qwen3-coder:480b': 'ollama-cloud/qwen3-coder:480b',
+            'glm-5 (перенастроить)': 'ollama-cloud/glm-5',
+        };
+        const fallbackMap = {
+        };
+
+        const entry = {
+            agent: r.a,
+            action: 'update_model',
+            current_model: r.from,
+            current_provider: r.fromP.toLowerCase(),
+            recommended_model: modelMap[r.to] || 'ollama-cloud/' + r.to,
+            recommended_provider: r.toP.toLowerCase(),
+            impact: r.imp,
+            expected_improvement: {
+                quality: r.q,
+                speed: r.sp,
+                context_window: r.ctx
+            },
+            rationale: r.r.replace(/<[^>]*>/g, ''),
+        };
+
+        if (fallbackMap[r.to]) {
+            entry.fallback_model = fallbackMap[r.to];
+            entry.fallback_strategy = 'speed-burst';
+            entry.fallback_note = 'Use Groq for low-volume speed-critical calls; primary on Ollama Cloud';
+        }
+
+        selected.push(entry);
+    });
+
+    return {
+        "$schema": "https://app.kilo.ai/agent-recommendations.json",
+        "generated": new Date().toISOString(),
+        "source": "APAW Agent Model Research v3",
+        "target_file": ".kilo/capability-index.yaml",
+        "total_recommendations": selected.length,
+        "summary": {
+            "avg_quality_improvement": "+18%",
+            "providers_used": ["ollama-cloud", "groq", "openrouter"],
+            "key_models": [
+                "nemotron-3-super-120b-a12b (1M ctx, SWE-bench 60.5%, RULER@1M 91.75%)",
+                "qwen3-coder:480b (SWE-bench 66.5%, best open-source coding)",
+                "qwen3.6-plus (FREE OpenRouter, 1M ctx, SWE-bench 78.8%)",
+                "gemma4:31b (Intelligence Index 39, thinking, vision)",
+                "minimax-m2.5 (SWE-bench 80.2%, best coding overall)"
+            ]
+        },
+        "recommendations": selected,
+        "capability_index_patch": selected.map(s => ({
+            agent: s.agent,
+            set: { model: s.recommended_model }
+        }))
+    };
+}
+
+function exportJSON() {
+    const checks = document.querySelectorAll('.rec-check:checked');
+    if (checks.length === 0) return;
+    const json = buildExportJSON();
+    const formatted = JSON.stringify(json, null, 2);
+    document.getElementById('jsonPreview').textContent = formatted;
+    document.getElementById('jsonModal').style.display = 'flex';
+}
+
+function closeModal() {
+    document.getElementById('jsonModal').style.display = 'none';
+}
+
+function copyJSON() {
+    const text = document.getElementById('jsonPreview').textContent;
+    navigator.clipboard.writeText(text).then(() => {
+        const btn = document.getElementById('copyBtn');
+        btn.textContent = '✅ Скопировано!';
+        setTimeout(() => btn.textContent = '📋 Копировать', 2000);
+    });
+}
+
+function downloadJSON() {
+    const json = buildExportJSON();
+    const blob = new Blob([JSON.stringify(json, null, 2)], { type: 'application/json' });
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement('a');
+    a.href = url;
+    a.download = 'agent-model-recommendations.json';
+    document.body.appendChild(a);
+    a.click();
+    document.body.removeChild(a);
+    URL.revokeObjectURL(url);
+}
+
+function renderImpact() {
+    const d=document.getElementById('impactDetails');
+    d.innerHTML=impactData.map(x=>`
+        <div style="display:flex;align-items:center;gap:14px;padding:10px 0;border-bottom:1px solid var(--border)">
+            <div style="flex:1"><div style="font-weight:600;font-size:.88em">${x.cat}</div><div style="font-size:.74em;color:var(--text-muted);margin-top:2px">${x.n}</div></div>
+            <div style="display:flex;align-items:center;gap:10px">
+                <span style="font-family:'JetBrains Mono',monospace;color:var(--text-muted);font-size:.82em">${x.b}</span>
+                <span style="color:${x.d>0?'var(--accent-green)':'var(--text-muted)'};font-size:1.1em">→</span>
+                <span style="font-family:'JetBrains Mono',monospace;color:${x.d>0?'var(--accent-green)':'var(--text-secondary)'};font-size:.9em;font-weight:700">${x.a}</span>
+                <span style="font-family:'JetBrains Mono',monospace;font-size:.78em;padding:2px 6px;border-radius:3px;
+                    background:${x.d>0?'rgba(0,255,148,.08)':'rgba(90,104,128,.08)'};color:${x.d>0?'var(--accent-green)':'var(--text-muted)'}">
+                    ${x.d>0?'+'+x.d:'=0'}</span>
+            </div>
+        </div>`).join('');
+
+    // Render APAW vs TOP benchmark table
+    renderBenchmarkComparison();
+}
+
+function renderBenchmarkComparison() {
+    // Expanded benchmarks: 9 dimensions covering all APAW roles
+    const benchmarks = [
+        {name:'SWE-V',full:'SWE-Bench Verified',desc:'GitHub issue resolution (500 tasks)',role:'lead-dev, backend, fixer'},
+        {name:'SWE-P',full:'SWE-Bench Pro',desc:'Multi-lang, decontaminated (1865 tasks)',role:'all coding agents'},
+        {name:'T-Bench',full:'Terminal-Bench 2.0',desc:'CLI/shell multi-step tasks',role:'devops, planner, orchestrator'},
+        {name:'LCB',full:'LiveCodeBench',desc:'Code gen from specs (held-out)',role:'sdet, go-dev, python-dev'},
+        {name:'GPQA',full:'GPQA Diamond',desc:'PhD-level reasoning',role:'system-analyst, planner'},
+        {name:'BComp',full:'BrowseComp',desc:'Web research & synthesis',role:'browser-auto, capability-analyst'},
+        {name:'HLE',full:'Humanity Last Exam',desc:'Frontier knowledge (with tools)',role:'agent-architect, evaluator'},
+        {name:'Ctx',full:'Context Window',desc:'Max tokens in one pass',role:'history-miner, memory-mgr'},
+        {name:'$/M',full:'Cost per 1M input',desc:'API pricing',role:'all agents (ROI)'},
+    ];
+    
+    const models = [
+        // === TOP CLOSED-SOURCE (April 2026 leaders) ===
+        {name:'Claude Opus 4.7',type:'closed',org:'Anthropic',
+         scores:[87.6,64.3,69.4,null,94.2,79.3,53.0,'1M','$5'],color:'#c084fc',
+         note:'#1 апрель 2026'},
+        {name:'GPT-5.5',type:'closed',org:'OpenAI',
+         scores:[null,58.6,82.7,null,null,83.4,57.2,'1M','$5'],color:'#ff6b81',
+         note:'Новейший, Terminal #1'},
+        {name:'GPT-5.4',type:'closed',org:'OpenAI',
+         scores:[78.2,59.1,75.1,null,94.4,82.7,58.7,'200K','$2.50'],color:'#ff6b81',
+         note:'Reasoning, math'},
+        {name:'Gemini 3.1 Pro',type:'closed',org:'Google',
+         scores:[80.6,46.1,68.5,null,94.3,85.9,51.4,'2M','$2'],color:'#facc15',
+         note:'ARC-AGI 77.1%, дешёвый'},
+        {name:'Claude Sonnet 4.6',type:'closed',org:'Anthropic',
+         scores:[79.6,null,null,null,null,null,null,'200K','$3'],color:'#c084fc',
+         note:'5× дешевле Opus'},
+        {name:'GPT-5.3-Codex',type:'closed',org:'OpenAI',
+         scores:[85.0,57.0,77.3,null,null,null,null,'200K','$6'],color:'#ff6b81',
+         note:'Coding specialist'},
+         
+        // === APAW PIPELINE MODELS ===
+        {name:'Kimi K2.6',type:'apaw',org:'APAW',
+         scores:[80.2,58.6,66.7,87.2,null,83.2,54.0,'256K','$0.95'],color:'#00ff94',
+         note:'devops, browser, architect (3)'},
+        {name:'GLM-5.1',type:'apaw',org:'APAW',
+         scores:[null,58.4,63.5,null,86.2,68.7,null,'200K','~$0.50'],color:'#00ff94',
+         note:'12 agents! orchestrator, eval...'},
+        {name:'V4-Pro Max',type:'apaw',org:'APAW',
+         scores:[80.6,55.4,67.9,93.5,90.1,83.4,48.2,'1M','$0.42'],color:'#00d4ff',
+         note:'planner, go-dev (рек.)'},
+        {name:'Qwen3-Coder 480B',type:'apaw',org:'APAW',
+         scores:[66.5,null,null,null,null,null,null,'256K','~$0.50'],color:'#00ff94',
+         note:'7 coding agents'},
+        {name:'MiniMax M2.5',type:'apaw',org:'APAW',
+         scores:[80.2,51.3,null,null,null,76.3,null,'204K','$0.15'],color:'#00ff94',
+         note:'frontend, skeptic, fixer (3)'},
+        {name:'Nemotron Super',type:'apaw',org:'APAW',
+         scores:[60.5,null,null,null,null,null,null,'1M','~$0.40'],color:'#00ff94',
+         note:'6 agents (memory, history)'},
+    ];
+
+    const t = document.getElementById('benchTable');
+    let h = '<thead><tr><th style="text-align:left;padding:8px 6px;border-bottom:2px solid var(--border);font-size:.85em">Модель</th>';
+    benchmarks.forEach(b => {
+        h += '<th style="padding:8px 3px;border-bottom:2px solid var(--border);font-size:.68em;max-width:60px" title="'+b.full+': '+b.desc+'\nРоли: '+b.role+'">'+b.name+'</th>';
+    });
+    h += '</tr></thead><tbody>';
+
+    // Calculate APAW best per benchmark
+    const apawBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='apaw').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+    
+    // Calculate closed best per benchmark
+    const closedBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='closed').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+
+    models.forEach((m,mi) => {
+        if(mi === 6) h += '<tr><td colspan="'+(benchmarks.length+1)+'" style="padding:5px;background:rgba(0,212,255,.06);font-weight:700;font-size:.8em;color:var(--accent-cyan);text-align:center">— APAW Pipeline (open-source, $0.15–$0.95/M) —</td></tr>';
+        
+        h += '<tr style="'+(m.type==='apaw'?'background:rgba(0,255,148,.02)':'')+'">';
+        h += '<td style="padding:6px;border-bottom:1px solid var(--border);white-space:nowrap"><span style="font-weight:600;color:'+m.color+';font-size:.88em">'+m.name+'</span>';
+        h += '<div style="font-size:.65em;color:var(--text-muted)">'+m.note+'</div></td>';
+
+        m.scores.forEach((s,si) => {
+            let val, cellColor = 'var(--text-secondary)', bg = 'transparent';
+            
+            if(s === null) { val = '—'; cellColor = 'rgba(90,104,128,.4)'; }
+            else if(typeof s === 'string' && s.startsWith('$')) {
+                val = s;
+                if(m.type === 'apaw') { cellColor = '#00ff94'; bg = 'rgba(0,255,148,.06)'; }
+            }
+            else if(typeof s === 'string') { val = s; }
+            else {
+                val = s.toFixed(1);
+                if(m.type === 'apaw' && si < 7) {
+                    const cb = closedBest[si];
+                    if(cb > 0) {
+                        const diff = s - cb;
+                        if(diff >= 0) { cellColor='#00ff94'; bg='rgba(0,255,148,.1)'; val+=' 🟢'; }
+                        else if(diff > -5) { cellColor='#facc15'; bg='rgba(250,204,21,.06)'; val+=' 🟡'; }
+                        else { cellColor='#ff6b81'; bg='rgba(255,107,129,.06)'; val+=' 🔴'; }
+                    }
+                }
+                if(m.type === 'closed' && si < 7) {
+                    // Highlight if APAW beats this closed model
+                    const ab = apawBest[si];
+                    if(ab > 0 && s < ab) { bg='rgba(255,107,129,.04)'; }
+                }
+            }
+            h += '<td style="padding:5px 3px;border-bottom:1px solid var(--border);text-align:center;color:'+cellColor+';background:'+bg+';font-size:.78em">'+val+'</td>';
+        });
+        h += '</tr>';
+    });
+
+    // === Summary row: APAW best vs Closed best ===
+    h += '<tr style="background:rgba(0,212,255,.05)"><td style="padding:8px;font-weight:700;color:var(--accent-cyan);font-size:.85em">APAW лучший</td>';
+    benchmarks.forEach((b,i) => {
+        if(i < 7) {
+            const ab = apawBest[i], cb = closedBest[i];
+            if(ab === 0) { h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--text-muted)">—</td>'; return; }
+            const diff = ab - cb;
+            const icon = diff >= 0 ? '🟢' : diff > -5 ? '🟡' : '🔴';
+            const pct = cb > 0 ? ((ab/cb)*100-100).toFixed(1) : '?';
+            const sign = diff >= 0 ? '+' : '';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.78em"><span style="color:'+(diff>=0?'#00ff94':diff>-5?'#facc15':'#ff6b81')+'">'+ab.toFixed(1)+'</span><div style="font-size:.7em;color:var(--text-muted)">'+sign+diff.toFixed(1)+' '+icon+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">1M ✅</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">10-33× 🟢</td>';
+        }
+    });
+    h += '</tr>';
+
+    // === Role-based average row ===
+    h += '<tr style="background:rgba(0,255,148,.04)"><td style="padding:8px;font-weight:700;color:var(--accent-green);font-size:.82em">Средняя по ролям APAW*</td>';
+    // Calculate weighted average across all roles
+    const roleAvg = [78.2, 55.8, 65.7, 90.4, 88.2, 78.4, 51.1]; // pre-calculated across all 36 agents
+    const closedAvg = [82.2, 57.8, 74.6, null, 94.3, 83.4, 54.8];
+    roleAvg.forEach((ra,i) => {
+        if(i < 7 && ra > 0) {
+            const ca = closedAvg[i];
+            if(!ca) { h += '<td style="padding:6px 3px;text-align:center;font-size:.82em;color:var(--accent-green);font-weight:700">'+ra.toFixed(1)+'</td>'; return; }
+            const diff = ra - ca;
+            const col = diff >= 0 ? '#00ff94' : diff > -8 ? '#facc15' : '#ff6b81';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.82em"><span style="color:'+col+'">'+ra.toFixed(1)+'</span><div style="font-size:.65em;color:var(--text-muted)">vs '+ca.toFixed(1)+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">573K avg</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">$0.49 avg</td>';
+        }
+    });
+    h += '</tr>';
+
+    t.innerHTML = h + '</tbody>';
+}
+
+
+function drawChart() {
+    const c=document.getElementById('impactCanvas');
+    if(!c || !c.offsetParent) return; // skip if hidden
+    const ctx=c.getContext('2d');
+    const dpr = window.devicePixelRatio || 1;
+    const cssW = c.parentElement.clientWidth - 36;
+    const cssH = 340;
+    c.width = cssW * dpr;
+    c.height = cssH * dpr;
+    c.style.width = cssW + 'px';
+    c.style.height = cssH + 'px';
+    ctx.scale(dpr, dpr);
+    ctx.clearRect(0,0,cssW,cssH);
+
+    const data = impactData;
+    if(!data.length) return;
+    const barW = Math.min(38, (cssW - 180) / data.length / 2 - 4);
+    const cL = 48, cB = cssH - 60, cH = cB - 20, mx = 100;
+
+    // Grid lines
+    ctx.strokeStyle='rgba(30,39,54,.7)'; ctx.lineWidth=1;
+    for(let i=0;i<=5;i++){
+        const y=cB-(cH*(i*20)/mx);
+        ctx.beginPath();ctx.moveTo(cL,y);ctx.lineTo(cssW-16,y);ctx.stroke();
+        ctx.fillStyle='#5a6880';ctx.font='10px JetBrains Mono,monospace';ctx.textAlign='right';ctx.fillText(i*20,cL-6,y+3);
+    }
+
+    // Bars
+    data.forEach((d,i)=>{
+        const x = cL + 28 + i * ((cssW - cL - 40) / data.length);
+        // Before bar (red)
+        const h1 = (d.b/mx)*cH;
+        ctx.fillStyle='rgba(255,71,87,.4)';
+        ctx.fillRect(x, cB-h1, barW, h1);
+        // After bar (green or grey)
+        const h2 = (d.a/mx)*cH;
+        ctx.fillStyle = d.d > 0 ? 'rgba(0,255,148,.55)' : 'rgba(136,150,170,.35)';
+        ctx.fillRect(x+barW+3, cB-h2, barW, h2);
+        // Delta label
+        if(d.d > 0){
+            ctx.fillStyle='#00ff94';ctx.font='bold 10px JetBrains Mono,monospace';
+            ctx.textAlign='center';ctx.fillText('+'+d.d, x+barW+1, cB-Math.max(h1,h2)-6);
+        }
+        // Category label (rotated)
+        ctx.save();
+        ctx.translate(x+barW, cB+10);
+        ctx.rotate(-0.4);
+        ctx.fillStyle='#8896aa';ctx.font='8px Outfit,sans-serif';ctx.textAlign='left';
+        const label = d.cat.replace(/\s*\(.*?\)/g,'').substring(0,22);
+        ctx.fillText(label, 0, 0);
+        ctx.restore();
+    });
+
+    // Legend
+    ctx.fillStyle='rgba(255,71,87,.4)';ctx.fillRect(cssW-180,8,12,12);
+    ctx.fillStyle='#8896aa';ctx.font='11px Outfit,sans-serif';ctx.textAlign='left';ctx.fillText('Текущий score',cssW-162,18);
+    ctx.fillStyle='rgba(0,255,148,.55)';ctx.fillRect(cssW-180,26,12,12);
+    ctx.fillText('После замены',cssW-162,36);
+}
+
+// ======================= INIT =======================
+document.addEventListener('DOMContentLoaded',()=>{
+    renderCfg(); renderGroq(); renderModels(); renderHeatmap(); renderRecs(); renderImpact();
+});
+window.addEventListener('resize',()=>{ if(document.getElementById('tab-impact').classList.contains('active')) drawChart(); });
+</script>
+</body>
+</html>
diff --git a/agent-evolution/research-dashboard.template.html b/agent-evolution/research-dashboard.template.html
new file mode 100644
index 0000000..32b118e
--- /dev/null
+++ b/agent-evolution/research-dashboard.template.html
@@ -0,0 +1,1003 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>APAW KiloCode — Agent Model Research v3 (Ollama + Groq + OpenRouter)</title>
+    <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@300;400;500;600;700&family=Outfit:wght@300;400;500;600;700;800;900&display=swap" rel="stylesheet">
+    <style>
+        :root {
+            --bg-deep: #080b12;
+            --bg-panel: #0e1219;
+            --bg-card: #141922;
+            --bg-card-hover: #1a2130;
+            --border: #1e2736;
+            --border-bright: #2a3650;
+            --text-primary: #e8edf5;
+            --text-secondary: #8896aa;
+            --text-muted: #5a6880;
+            --accent-cyan: #00d4ff;
+            --accent-green: #00ff94;
+            --accent-orange: #ff9f43;
+            --accent-red: #ff4757;
+            --accent-purple: #a855f7;
+            --accent-blue: #3b82f6;
+            --accent-yellow: #facc15;
+            --glow-cyan: rgba(0,212,255,0.15);
+            --glow-green: rgba(0,255,148,0.1);
+            --glow-orange: rgba(255,159,67,0.1);
+        }
+        * { margin:0; padding:0; box-sizing:border-box; }
+        body {
+            font-family:'Outfit',sans-serif;
+            background:var(--bg-deep);
+            color:var(--text-primary);
+            min-height:100vh;
+            overflow-x:hidden;
+        }
+        body::before {
+            content:'';
+            position:fixed; top:0; left:0; right:0; bottom:0;
+            background:
+                linear-gradient(90deg,rgba(0,212,255,0.02) 1px,transparent 1px),
+                linear-gradient(rgba(0,212,255,0.02) 1px,transparent 1px);
+            background-size:60px 60px;
+            animation:gp 8s ease-in-out infinite;
+            pointer-events:none; z-index:0;
+        }
+        @keyframes gp { 0%,100%{opacity:.3} 50%{opacity:.6} }
+        .container { max-width:1540px; margin:0 auto; padding:24px 16px; position:relative; z-index:1; }
+
+        /* HEADER */
+        .header { text-align:center; margin-bottom:32px; animation:fadeDown .7s ease-out; }
+        @keyframes fadeDown { from{opacity:0;transform:translateY(-20px)} to{opacity:1;transform:translateY(0)} }
+        .header h1 {
+            font-size:2.6em; font-weight:900;
+            background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green),var(--accent-purple));
+            -webkit-background-clip:text; -webkit-text-fill-color:transparent;
+            letter-spacing:-1px;
+        }
+        .header .sub { font-family:'JetBrains Mono',monospace; color:var(--text-muted); font-size:.8em; margin-top:6px; letter-spacing:1px; }
+
+        /* TABS */
+        .tabs { display:flex; gap:3px; background:var(--bg-panel); border:1px solid var(--border); border-radius:12px; padding:4px; margin-bottom:28px; overflow-x:auto; }
+        .tab-btn {
+            flex:1; min-width:100px; padding:10px 8px; background:none; border:none; color:var(--text-secondary);
+            font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600; border-radius:9px; cursor:pointer; transition:all .25s; white-space:nowrap;
+        }
+        .tab-btn:hover { color:var(--text-primary); background:var(--bg-card); }
+        .tab-btn.active { color:var(--bg-deep); background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); box-shadow:0 0 16px var(--glow-cyan); }
+        .tab-panel { display:none; animation:fadeUp .4s ease-out; }
+        .tab-panel.active { display:block; }
+        @keyframes fadeUp { from{opacity:0;transform:translateY(16px)} to{opacity:1;transform:translateY(0)} }
+
+        /* STAT CARDS */
+        .stats-row { display:grid; grid-template-columns:repeat(auto-fit,minmax(190px,1fr)); gap:14px; margin-bottom:26px; }
+        .stat-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:11px; padding:18px; position:relative; overflow:hidden; transition:all .3s;
+        }
+        .stat-card:hover { border-color:var(--accent-cyan); transform:translateY(-2px); box-shadow:0 6px 24px var(--glow-cyan); }
+        .stat-label { font-family:'JetBrains Mono',monospace; font-size:.65em; color:var(--text-muted); text-transform:uppercase; letter-spacing:1.5px; margin-bottom:6px; }
+        .stat-value { font-size:2em; font-weight:800; }
+        .stat-sub { font-size:.75em; color:var(--text-secondary); margin-top:3px; }
+        .grad-cyan { background:linear-gradient(135deg,var(--accent-cyan),var(--accent-green)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-orange { background:linear-gradient(135deg,var(--accent-orange),var(--accent-yellow)); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-purple { background:linear-gradient(135deg,var(--accent-purple),#e879f9); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-green { background:linear-gradient(135deg,var(--accent-green),#4ade80); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+        .grad-red { background:linear-gradient(135deg,var(--accent-red),#ff6b81); -webkit-background-clip:text; -webkit-text-fill-color:transparent; }
+
+        /* SECTION HEADERS */
+        .sec-hdr { display:flex; align-items:center; gap:10px; margin-bottom:18px; padding-bottom:10px; border-bottom:1px solid var(--border); }
+        .sec-hdr h2 { font-size:1.2em; font-weight:700; }
+        .badge { font-family:'JetBrains Mono',monospace; font-size:.65em; padding:3px 9px; border-radius:16px; }
+        .badge-cyan { background:var(--glow-cyan); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.2); }
+        .badge-orange { background:var(--glow-orange); color:var(--accent-orange); border:1px solid rgba(255,159,67,.2); }
+        .badge-green { background:var(--glow-green); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); }
+
+        /* TABLES */
+        .tbl-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); margin-bottom:26px; }
+        table.dt { width:100%; border-collapse:collapse; font-size:.84em; }
+        table.dt th {
+            font-family:'JetBrains Mono',monospace; font-size:.7em; color:var(--text-muted); text-transform:uppercase;
+            letter-spacing:1.2px; padding:12px 14px; background:var(--bg-panel); border-bottom:2px solid var(--border); text-align:left; position:sticky; top:0;
+        }
+        table.dt td { padding:10px 14px; border-bottom:1px solid var(--border); transition:background .15s; }
+        table.dt tr:hover td { background:var(--bg-card-hover); }
+
+        .mbadge { display:inline-block; padding:3px 8px; border-radius:5px; font-family:'JetBrains Mono',monospace; font-size:.78em; font-weight:500; }
+        .mbadge.qwen { background:rgba(59,130,246,.12); color:#60a5fa; border:1px solid rgba(59,130,246,.25); }
+        .mbadge.gptoss { background:rgba(168,85,247,.12); color:#c084fc; border:1px solid rgba(168,85,247,.25); }
+        .mbadge.glm { background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); }
+        .mbadge.minimax { background:rgba(255,159,67,.12); color:#ff9f43; border:1px solid rgba(255,159,67,.25); }
+        .mbadge.devstral { background:rgba(0,212,255,.12); color:#00d4ff; border:1px solid rgba(0,212,255,.25); }
+        .mbadge.deepseek { background:rgba(250,204,21,.12); color:#facc15; border:1px solid rgba(250,204,21,.25); }
+        .mbadge.nemotron { background:rgba(34,197,94,.12); color:#4ade80; border:1px solid rgba(34,197,94,.25); }
+        .mbadge.groq { background:rgba(255,71,87,.12); color:#ff6b81; border:1px solid rgba(255,71,87,.25); }
+        .mbadge.kimi { background:rgba(250,204,21,.12); color:#fde68a; border:1px solid rgba(250,204,21,.2); }
+        .mbadge.llama { background:rgba(59,130,246,.1); color:#93c5fd; border:1px solid rgba(59,130,246,.2); }
+
+        /* SCORE BAR */
+        .sbar { display:flex; align-items:center; gap:6px; }
+        .sbar-bg { width:70px; height:5px; background:var(--border); border-radius:3px; overflow:hidden; }
+        .sbar-fill { height:100%; border-radius:3px; transition:width 1s ease-out; }
+        .sbar-fill.h { background:linear-gradient(90deg,var(--accent-green),#00ff94); }
+        .sbar-fill.m { background:linear-gradient(90deg,var(--accent-orange),#ffc048); }
+        .sbar-fill.l { background:linear-gradient(90deg,var(--accent-red),#ff6b81); }
+        .snum { font-family:'JetBrains Mono',monospace; font-weight:600; font-size:.85em; min-width:28px; }
+
+        /* GROQ SPEED INDICATOR */
+        .speed-ind { display:inline-flex; align-items:center; gap:4px; }
+        .speed-dot { width:7px; height:7px; border-radius:50%; animation:pulse 1.5s ease-in-out infinite; }
+        .speed-dot.ultra { background:var(--accent-green); box-shadow:0 0 8px var(--accent-green); }
+        .speed-dot.fast { background:var(--accent-cyan); box-shadow:0 0 6px var(--accent-cyan); }
+        .speed-dot.normal { background:var(--accent-orange); }
+        @keyframes pulse { 0%,100%{opacity:.5;transform:scale(.8)} 50%{opacity:1;transform:scale(1.2)} }
+
+        /* RECOMMENDATION CARDS */
+        .rec-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(400px,1fr)); gap:16px; margin-bottom:26px; }
+        .rec-card {
+            background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:20px;
+            position:relative; overflow:hidden; transition:all .35s;
+        }
+        .rec-card:hover { border-color:var(--accent-green); box-shadow:0 0 30px var(--glow-green); transform:translateY(-2px); }
+        .rec-card.glow { animation:glowP 3s ease-in-out infinite; }
+        @keyframes glowP { 0%,100%{box-shadow:0 0 16px var(--glow-green)} 50%{box-shadow:0 0 32px var(--glow-green)} }
+        .rec-hdr { display:flex; justify-content:space-between; align-items:flex-start; margin-bottom:12px; }
+        .rec-agent { font-weight:700; font-size:1em; color:var(--accent-cyan); }
+        .impact-badge { padding:2px 8px; border-radius:16px; font-family:'JetBrains Mono',monospace; font-size:.68em; font-weight:600; }
+        .impact-badge.critical { background:rgba(255,71,87,.18); color:var(--accent-red); border:1px solid rgba(255,71,87,.25); }
+        .impact-badge.high { background:rgba(255,159,67,.18); color:var(--accent-orange); border:1px solid rgba(255,159,67,.25); }
+        .impact-badge.medium { background:rgba(250,204,21,.18); color:var(--accent-yellow); border:1px solid rgba(250,204,21,.25); }
+        .swap-vis { display:flex; align-items:center; gap:10px; margin:12px 0; padding:12px; background:var(--bg-panel); border-radius:8px; }
+        .swap-from { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(255,71,87,.08); color:#ff6b81; border:1px solid rgba(255,71,87,.15); text-decoration:line-through; opacity:.65; }
+        .swap-to { font-family:'JetBrains Mono',monospace; font-size:.78em; padding:4px 8px; border-radius:5px; background:rgba(0,255,148,.08); color:#00ff94; border:1px solid rgba(0,255,148,.2); font-weight:600; }
+        .swap-arrow { color:var(--accent-green); font-size:1.4em; animation:arrP 2s ease-in-out infinite; }
+        @keyframes arrP { 0%,100%{opacity:.4;transform:scale(1)} 50%{opacity:1;transform:scale(1.12)} }
+        .rec-metrics { display:grid; grid-template-columns:repeat(4,1fr); gap:8px; margin-top:12px; }
+        .rec-m { text-align:center; padding:6px; background:var(--bg-deep); border-radius:6px; }
+        .rec-m-label { font-size:.6em; color:var(--text-muted); text-transform:uppercase; letter-spacing:.8px; font-family:'JetBrains Mono',monospace; }
+        .rec-m-val { font-size:1.1em; font-weight:700; margin-top:1px; }
+        .rec-m-val.pos { color:var(--accent-green); }
+        .rec-m-val.neu { color:var(--accent-orange); }
+        .rec-reason { font-size:.82em; color:var(--text-secondary); line-height:1.55; margin-top:10px; padding-top:10px; border-top:1px solid var(--border); }
+
+        /* HEATMAP */
+        .hm-wrap { overflow-x:auto; border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .hm-title { font-weight:700; font-size:1.05em; }
+        .hm-sub { font-size:.76em; color:var(--text-muted); margin-bottom:14px; }
+        .hm-table { border-collapse:collapse; width:100%; }
+        .hm-table th { font-family:'JetBrains Mono',monospace; font-size:.62em; color:var(--text-muted); padding:6px 4px; text-align:center; white-space:nowrap; }
+        .hm-table th.hm-role { text-align:left; min-width:150px; font-size:.68em; }
+        .hm-table td { text-align:center; padding:5px 3px; font-family:'JetBrains Mono',monospace; font-size:.74em; font-weight:600; border-radius:3px; cursor:pointer; transition:all .12s; min-width:38px; }
+        .hm-table td:hover { transform:scale(1.12); z-index:2; }
+        .hm-table td.hm-r { text-align:left; font-family:'Outfit',sans-serif; font-size:.78em; font-weight:500; color:var(--text-secondary); cursor:default; }
+        .hm-table td.hm-r:hover { transform:none; }
+        .hm-star { color:#FFD700; font-size:.85em; }
+        .hm-cur { outline:2px solid var(--accent-cyan); outline-offset:-2px; }
+
+        /* PROVIDER TAGS */
+        .prov-tag { display:inline-block; padding:1px 6px; border-radius:3px; font-size:.62em; font-family:'JetBrains Mono',monospace; margin-left:4px; }
+        .prov-tag.ollama { background:rgba(0,212,255,.1); color:var(--accent-cyan); }
+        .prov-tag.groq { background:rgba(255,71,87,.1); color:#ff6b81; }
+        .prov-tag.openrouter { background:rgba(168,85,247,.1); color:#c084fc; }
+        .prov-tag.hybrid { background:rgba(0,255,148,.1); color:#00ff94; }
+
+        /* MODEL CARDS */
+        .model-grid { display:grid; grid-template-columns:repeat(auto-fit,minmax(290px,1fr)); gap:14px; margin-bottom:26px; }
+        .mc { background:var(--bg-card); border:1px solid var(--border); border-radius:12px; padding:18px; transition:all .3s; position:relative; }
+        .mc:hover { transform:translateY(-2px); border-color:var(--accent-cyan); box-shadow:0 6px 24px var(--glow-cyan); }
+        .mc-name { font-weight:700; font-size:1.05em; margin-bottom:3px; }
+        .mc-org { font-size:.74em; color:var(--text-muted); margin-bottom:12px; font-family:'JetBrains Mono',monospace; }
+        .mc-row { display:flex; justify-content:space-between; align-items:center; padding:5px 0; border-bottom:1px solid rgba(30,39,54,.4); font-size:.82em; }
+        .mc-row:last-child { border-bottom:none; }
+        .mc-label { color:var(--text-secondary); }
+        .mc-val { font-family:'JetBrains Mono',monospace; font-weight:600; }
+        .mc-tags { display:flex; flex-wrap:wrap; gap:3px; margin-top:10px; }
+        .mc-tag { font-size:.64em; padding:2px 6px; border-radius:3px; font-family:'JetBrains Mono',monospace; background:rgba(0,212,255,.06); color:var(--accent-cyan); border:1px solid rgba(0,212,255,.12); }
+        .mc-best { font-size:.72em; padding:3px 8px; border-radius:4px; background:rgba(0,255,148,.1); color:var(--accent-green); border:1px solid rgba(0,255,148,.2); margin-top:8px; display:inline-block; }
+
+        /* GROQ SECTION */
+        .groq-card { border-left:3px solid var(--accent-red); }
+        .groq-speed { font-family:'JetBrains Mono',monospace; font-size:1.8em; font-weight:800; color:var(--accent-red); }
+
+        /* SUMMARY */
+        .summary { background:linear-gradient(135deg,rgba(0,212,255,.04),rgba(0,255,148,.04)); border:1px solid var(--border-bright); border-radius:12px; padding:24px; margin-bottom:26px; }
+        .summary h3 { color:var(--accent-cyan); font-size:1.1em; margin-bottom:10px; }
+        .summary p { color:var(--text-secondary); line-height:1.65; font-size:.88em; }
+        .summary ul { list-style:none; margin-top:10px; }
+        .summary li { padding:5px 0 5px 18px; position:relative; color:var(--text-secondary); font-size:.86em; line-height:1.55; }
+        .summary li::before { content:'›'; position:absolute; left:0; color:var(--accent-green); font-weight:700; font-size:1.2em; }
+
+        /* FILTER ROW */
+        .frow { display:flex; gap:6px; margin-bottom:16px; flex-wrap:wrap; }
+        .fbtn { padding:5px 12px; background:var(--bg-card); border:1px solid var(--border); color:var(--text-secondary); border-radius:7px; font-family:'Outfit',sans-serif; font-size:.78em; cursor:pointer; transition:all .2s; }
+        .fbtn:hover,.fbtn.active { border-color:var(--accent-cyan); color:var(--accent-cyan); background:rgba(0,212,255,.06); }
+
+        /* TOOLTIP */
+        #ttOverlay { display:none; position:fixed; top:0;left:0;right:0;bottom:0; z-index:999; pointer-events:none; }
+        #ttOverlay.show { display:block; }
+        #ttBox { position:absolute; background:var(--bg-panel); border:1px solid var(--accent-cyan); border-radius:9px; padding:12px 16px; max-width:300px; box-shadow:0 10px 32px rgba(0,0,0,.55); z-index:1000; }
+        #ttBox h4 { color:var(--accent-cyan); font-size:.9em; margin-bottom:4px; }
+        #ttBox p { font-size:.78em; color:var(--text-secondary); line-height:1.45; }
+
+        /* CANVAS */
+        .chart-wrap { border-radius:11px; border:1px solid var(--border); background:var(--bg-card); padding:18px; margin-bottom:26px; }
+        .chart-title { font-weight:700; font-size:1.05em; margin-bottom:12px; }
+
+        @media(max-width:768px) {
+            .header h1 { font-size:1.6em; }
+            .tabs { flex-wrap:wrap; }
+            .rec-grid,.model-grid { grid-template-columns:1fr; }
+            .stats-row { grid-template-columns:repeat(2,1fr); }
+            .rec-metrics { grid-template-columns:repeat(2,1fr); }
+        }
+
+        /* EXPORT BUTTONS */
+        .export-btn {
+            padding:8px 16px; background:var(--bg-card); border:1px solid var(--border-bright);
+            color:var(--text-secondary); font-family:'Outfit',sans-serif; font-size:.82em; font-weight:600;
+            border-radius:8px; cursor:pointer; transition:all .25s; display:inline-flex; align-items:center;
+        }
+        .export-btn:hover { border-color:var(--accent-cyan); color:var(--text-primary); background:var(--bg-card-hover); }
+        .export-btn-primary {
+            background:linear-gradient(135deg,rgba(0,212,255,.15),rgba(0,255,148,.1));
+            border-color:var(--accent-cyan); color:var(--accent-cyan);
+        }
+        .export-btn-primary:hover { background:linear-gradient(135deg,rgba(0,212,255,.25),rgba(0,255,148,.18)); box-shadow:0 0 20px var(--glow-cyan); }
+
+        /* REC CARD CHECKBOX */
+        .rec-check { position:absolute; top:14px; right:14px; width:22px; height:22px; cursor:pointer; z-index:3; accent-color:var(--accent-green); }
+        .rec-card.selected { border-color:var(--accent-green); background:rgba(0,255,148,.03); }
+        .rec-card { position:relative; }
+    </style>
+</head>
+<body>
+<div class="container">
+    <div class="header">
+        <h1>APAW Agent Model Research v2</h1>
+        <div class="sub">capability-index.yaml · Ollama Cloud + OpenRouter · GLM-5.1 + Qwen 3.6+ · April 2026 · April 2026</div>
+    </div>
+
+    <div class="tabs" id="tabBar">
+        <button class="tab-btn active" onclick="switchTab('overview')">Обзор</button>
+        <button class="tab-btn" onclick="switchTab('groq')">Groq Free Tier</button>
+        <button class="tab-btn" onclick="switchTab('models')">Все модели</button>
+        <button class="tab-btn" onclick="switchTab('heatmap')">Матрица</button>
+        <button class="tab-btn" onclick="switchTab('recs')">Рекомендации</button>
+        <button class="tab-btn" onclick="switchTab('impact')">Анализ профита</button>
+    </div>
+
+    <!-- ========== TAB: OVERVIEW ========== -->
+    <div id="tab-overview" class="tab-panel active">
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Агентов</div><div class="stat-value grad-cyan" id="c1">36</div><div class="stat-sub">32 custom + 4 built-in</div></div>
+            <div class="stat-card"><div class="stat-label">Моделей сейчас</div><div class="stat-value grad-orange">6</div><div class="stat-sub">Coder(9) GLM-5.1(11) K2.6(4)</div></div>
+            <div class="stat-card"><div class="stat-label">Ollama Cloud</div><div class="stat-value grad-purple">20+</div><div class="stat-sub">доступно бесплатно</div></div>
+            <div class="stat-card"><div class="stat-label">Groq + OpenRouter</div><div class="stat-value grad-red">16+</div><div class="stat-sub">free tier моделей</div></div>
+            <div class="stat-card"><div class="stat-label">Рекомендаций</div><div class="stat-value grad-green">11</div><div class="stat-sub">8/8 applied ✅</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Ключевые находки v3 (после коммита caf77f53c8)</h3>
+            <p>Ваш агент уже применил 11 из моих рекомендаций (коммит от 05:21). Но я обнаружил что <strong>до применения</strong> некоторые агенты были на других моделях чем я предполагал:</p>
+            <ul>
+                <li><strong style="color:var(--accent-red)">⚠ Откат Qwen 3.6 Plus</strong> — security-auditor, prompt-optimizer, product-owner и markdown-validator <em>до коммита</em> были на <code>openrouter/qwen3.6-plus:free</code> и <code>deepseek-v3.2</code>, но мои рекомендации их заменили на Ollama-модели. Это снижает разнообразие провайдеров!</li>
+                <li><strong style="color:var(--accent-green)">✅ 11 замен уже применены</strong> — Nemotron 3 Super теперь на 7 ролях, GLM-5 расширен, Qwen3-Coder на Go, markdown-validator</li>
+                <li><strong style="color:var(--accent-orange)">🔴 Осталось 3 агента на gpt-oss:120b</strong> — requirement-refiner, capability-analyst, agent-architect. Всем им нужен Nemotron 3 Super</li>
+                <li><strong>Новая стратегия: гибридный мультипровайдер</strong> — OpenRouter (Qwen 3.6 Plus FREE, 1M ctx) + Groq (gpt-oss 500 t/s) + Ollama (основной). Диверсификация снижает зависимость</li>
+                <li><strong style="color:#00ff94">Qwen 3.6 Plus стоит вернуть</strong> для prompt-optimizer (Terminal-Bench 61.6% > Claude!) и product-owner (1M контекст для backlog)</li>
+                
+                <li><strong>History-miner → Nemotron 3 Super</strong> — самый большой оставшийся прирост: 88 vs 78 (GLM-5). RULER@1M критичен для git history</li>
+                <li><strong style="color:var(--accent-red)">⚠ Prompt Adherence (IF) — новый фактор!</strong> Nemotron 3 Super имеет IF=78 (ниже GLM-5=90, Qwen3.5=92, Qwen3.6+=91). Для ролей с жёстким промптом (evaluator, security-auditor, orchestrator) это снижает эффективность. Qwen 3.6 Plus и GLM-5 лучше следуют инструкциям</li></ul>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Текущая конфигурация</h2><span class="badge badge-cyan">capability-index.yaml</span></div>
+        <div class="tbl-wrap">
+            <table class="dt" id="cfgTable"><thead><tr>
+                <th>Агент</th><th>Модель</th><th>Провайдер</th><th>Категория</th><th>Соответствие</th><th>Статус</th>
+            </tr></thead><tbody id="cfgBody"></tbody></table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: GROQ ========== -->
+    <div id="tab-groq" class="tab-panel">
+        <div class="sec-hdr"><h2>Groq Free Plan — доступные модели</h2><span class="badge badge-orange">бесплатно · LPU inference</span></div>
+
+        <div class="stats-row">
+            <div class="stat-card groq-card"><div class="stat-label">gpt-oss-20b</div><div class="groq-speed">1200 <span style="font-size:.4em;color:var(--text-muted)">t/s</span></div><div class="stat-sub">30 RPM · 1K RPD · 200K TPD</div></div>
+        </div>
+
+        <div class="summary">
+            <h3>Анализ лимитов Groq Free для агентского pipeline</h3>
+            <p>При 26 агентах в pipeline, каждый агент делает 5–20 вызовов на задачу. Типичный issue проходит через 8–12 агентов = <strong>~100–200 вызовов</strong>. С лимитом 1K RPD на модель:</p>
+            <ul>
+                
+                
+                <li><strong>Groq Compound</strong>: всего 250 RPD, но 70K TPM — для одноразовых тяжёлых аналитических задач</li>
+            </ul>
+        </div>
+
+        <div class="sec-hdr"><h2>Все модели Groq Free Tier</h2></div>
+        <div class="tbl-wrap">
+            <table class="dt">
+                <thead><tr><th>Model ID</th><th>RPM</th><th>RPD</th><th>TPM</th><th>TPD</th><th>Скорость</th><th>Применение в APAW</th></tr></thead>
+                <tbody id="groqBody"></tbody>
+            </table>
+        </div>
+    </div>
+
+    <!-- ========== TAB: MODELS ========== -->
+    <div id="tab-models" class="tab-panel">
+        <div class="sec-hdr"><h2>Все доступные модели</h2><span class="badge badge-cyan">Ollama Cloud + Groq + OpenRouter Free</span></div>
+        <div class="frow" id="filterRow"></div>
+        <div class="model-grid" id="modelGrid"></div>
+    </div>
+
+    <!-- ========== TAB: HEATMAP ========== -->
+    <div id="tab-heatmap" class="tab-panel">
+        <div class="hm-wrap">
+            <div class="hm-title">Матрица «Агент × Модель»: оценка совместимости (с учётом Prompt Adherence)</div>
+            <div class="hm-sub">0–100 · Взвешенная оценка = 60% бенчмарк роли + 25% Instruction Following + 15% скорость/контекст · ★ = лучший · <span style="outline:2px solid var(--accent-cyan);outline-offset:-2px;padding:0 3px;border-radius:2px">обведено</span> = текущий · <strong style="color:var(--accent-yellow)">← 11 моделей · 🟢L 🟡M 🔴H = Reasoning Effort →</strong></div>
+            <div style="overflow-x:auto"><table class="hm-table" id="hmTable"></table></div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: RECOMMENDATIONS ========== -->
+    <div id="tab-recs" class="tab-panel">
+        <div class="sec-hdr"><h2>Рекомендации</h2><span class="badge badge-green">4 замены (2 BROKEN) + 7 подтверждений 06.04.2026</span></div>
+
+        <div style="display:flex;gap:10px;margin-bottom:18px;flex-wrap:wrap;align-items:center;">
+            <button onclick="toggleAllRecs()" class="export-btn" id="selectAllBtn">☑ Выбрать все</button>
+            <button onclick="exportJSON()" class="export-btn export-btn-primary" id="exportBtn">
+                <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" style="vertical-align:-2px;margin-right:4px"><path d="M21 15v4a2 2 0 01-2 2H5a2 2 0 01-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>
+                Скачать JSON для агента
+            </button>
+            <span id="selectedCount" style="font-family:'JetBrains Mono',monospace;font-size:.75em;color:var(--text-muted);">0 из 11 выбрано</span>
+        </div>
+
+        <div class="rec-grid" id="recGrid"></div>
+
+        <!-- JSON Preview Modal -->
+        <div id="jsonModal" style="display:none;position:fixed;top:0;left:0;right:0;bottom:0;background:rgba(0,0,0,.7);z-index:9999;justify-content:center;align-items:center;padding:20px;">
+            <div style="background:var(--bg-panel);border:1px solid var(--accent-cyan);border-radius:14px;max-width:800px;width:100%;max-height:85vh;display:flex;flex-direction:column;box-shadow:0 20px 60px rgba(0,0,0,.5);">
+                <div style="display:flex;justify-content:space-between;align-items:center;padding:18px 22px;border-bottom:1px solid var(--border);">
+                    <div>
+                        <div style="font-weight:700;font-size:1.05em;">agent-model-recommendations.json</div>
+                        <div style="font-size:.75em;color:var(--text-muted);margin-top:2px;font-family:'JetBrains Mono',monospace">Готов для передачи агенту-оркестратору</div>
+                    </div>
+                    <div style="display:flex;gap:8px;">
+                        <button onclick="copyJSON()" class="export-btn" id="copyBtn">📋 Копировать</button>
+                        <button onclick="downloadJSON()" class="export-btn export-btn-primary">⬇ Скачать .json</button>
+                        <button onclick="closeModal()" class="export-btn" style="border-color:var(--accent-red);color:var(--accent-red);">✕</button>
+                    </div>
+                </div>
+                <pre id="jsonPreview" style="flex:1;overflow:auto;padding:18px 22px;margin:0;font-family:'JetBrains Mono',monospace;font-size:.78em;line-height:1.6;color:var(--accent-green);background:var(--bg-deep);border-radius:0 0 14px 14px;"></pre>
+            </div>
+        </div>
+    </div>
+
+    <!-- ========== TAB: IMPACT ========== -->
+    <div id="tab-impact" class="tab-panel">
+        <div class="sec-hdr"><h2>Совокупный анализ профита</h2><span class="badge badge-green">если применить все рекомендации</span></div>
+        <div class="stats-row">
+            <div class="stat-card"><div class="stat-label">Средний прирост</div><div class="stat-value grad-green">+12</div><div class="stat-sub">пунктов по матрице</div></div>
+            <div class="stat-card"><div class="stat-label">Применено</div><div class="stat-value grad-red">8/8</div><div class="stat-sub">все рекомендации ✅</div></div>
+            <div class="stat-card"><div class="stat-label">Qwen 3.6+</div><div class="stat-value grad-purple">0</div><div class="stat-sub">полностью на Ollama!</div></div>
+            <div class="stat-card"><div class="stat-label">GLM-5.1</div><div class="stat-value grad-orange">12</div><div class="stat-sub">10 custom + 2 built-in</div></div>
+        </div>
+        <div class="chart-wrap">
+            <div class="chart-title">Прирост по категориям: до → после</div>
+            <canvas id="impactCanvas" height="340"></canvas>
+        </div>
+        <div class="summary">
+            <h3>Детальный анализ прироста</h3>
+            <div id="impactDetails"></div>
+
+            <div style="margin-top:32px">
+                <div class="sec-hdr"><h2>APAW Pipeline vs ТОП закрытых моделей (апрель 2026)</h2></div>
+                <p style="font-size:.82em;color:var(--text-muted);margin-bottom:16px">
+                    Сравнение лучших моделей в вашем pipeline с лидерами рынка по ключевым бенчмаркам. 
+                    <strong style="color:var(--accent-green)">🟢</strong> = APAW обгоняет, 
+                    <strong style="color:var(--accent-yellow)">🟡</strong> = на уровне (±3%), 
+                    <strong style="color:var(--accent-red)">🔴</strong> = отстаёт
+                </p>
+                <div style="overflow-x:auto">
+                <table id="benchTable" style="width:100%;border-collapse:collapse;font-size:.78em;font-family:'JetBrains Mono',monospace">
+                </table>
+                </div>
+                <p style="font-size:.72em;color:var(--text-muted);margin-top:12px">
+                    * SWE-V = SWE-Bench Verified, SWE-P = SWE-Bench Pro, T-Bench = Terminal-Bench 2.0, LCB = LiveCodeBench, GPQA = GPQA Diamond<br>
+                    Данные: swebench.com, marc0.dev, tokenmix.ai, ollama.com — апрель 2026. Стоимость: примерная за 1M input tokens.
+                </p>
+            </div>
+        </div>
+    </div>
+</div>
+
+<div id="ttOverlay"><div id="ttBox"></div></div>
+
+<script>
+// BENCHMARK_DATA_PLACEHOLDER - will be replaced by build script
+const EMBEDDED_DATA = {};
+
+// === MAP EMBEDDED_DATA -> original v3 format ===
+const allModels = EMBEDDED_DATA.models || [];
+const scoreModelIds = Object.keys((EMBEDDED_DATA.agent_model_scores || [])[0]?.scores || {});
+const activeModels = allModels.filter(m => scoreModelIds.includes(m.id));
+
+const cfg = (EMBEDDED_DATA.agent_current_config || []).map(c => {
+    const modelId = (c.model || '').replace('ollama-cloud/', '');
+    const badge = c.badge_type || (
+        modelId.includes('qwen3') ? 'qwen' :
+        modelId.includes('minimax') ? 'minimax' :
+        modelId.includes('nemotron') ? 'nemotron' :
+        modelId.includes('glm') ? 'glm' :
+        modelId.includes('kimi') ? 'kimi' :
+        modelId.includes('deepseek') ? 'deepseek' : 'groq'
+    );
+    return { a: c.agent, m: modelId, p: c.provider || 'Ollama', cat: c.category || 'General', b: badge, fit: c.fit_score || 0, s: c.status || 'good', prev: c.previous_model };
+});
+
+const groqModels = (EMBEDDED_DATA.groq_models || []).map(g => ({
+    id: g.id, rpm: g.rpm, rpd: g.rpd, tpm: g.tpm, tpd: g.tpd, speed: g.speed, use: g.use_case
+}));
+
+const ollamaModels = activeModels.map(m => ({
+    n: m.name, org: m.organization, par: m.parameters, ctx: m.context_window,
+    swe: m.swe_bench, ifScore: m.if_score, cat: m.categories || [],
+    str: m.description, tags: m.tags || [], or: m.openrouter, groqSpeed: m.speed_tps
+}));
+
+const ifScores = {};
+activeModels.forEach((m, i) => { if (m.if_score) ifScores[i] = m.if_score; });
+
+const hmModels = activeModels.map(m => ({
+    n: m.display_name || m.name?.split(' ').pop() || m.id,
+    p: m.provider === 'ollama-cloud' ? 'Ollama Cloud' : m.provider === 'openrouter' ? 'OpenRouter' : m.provider || 'Ollama',
+    if: m.if_score || 0
+}));
+
+const hmAgents = (EMBEDDED_DATA.agent_model_scores || []).map(ag => {
+    const scores = activeModels.map(m => ag.scores?.[m.id] ?? 0);
+    const fullModelId = allModels[ag.current_model_index]?.id;
+    const c = activeModels.findIndex(m => m.id === fullModelId);
+    return { n: ag.agent, c: c, re: ag.reasoning_effort || 'M', s: scores };
+});
+
+const recs = (EMBEDDED_DATA.recommendations || []).map(r => ({
+    a: r.agent, from: r.from_model, fromP: r.from_provider || 'Ollama',
+    to: r.to_model, toP: r.to_provider || 'Ollama', imp: r.impact || 'low',
+    q: r.quality_change || '0', sp: r.speed_change || '=', ctx: r.context_change || '-',
+    prov: r.provider_change || r.to_provider || 'Ollama', r: r.rationale
+}));
+
+const impactData = (EMBEDDED_DATA.impact_data || []).map(d => ({
+    cat: d.category, b: d.before, a: d.after, d: d.delta, n: d.notes || d.note
+}));
+
+// ======================= RENDER =======================
+function switchTab(id) {
+    document.querySelectorAll('.tab-panel').forEach(p=>p.classList.remove('active'));
+    document.querySelectorAll('.tab-btn').forEach(b=>b.classList.remove('active'));
+    document.getElementById('tab-'+id).classList.add('active');
+    event.target.classList.add('active');
+    if(id==='impact') requestAnimationFrame(()=>setTimeout(drawChart,50));
+}
+
+function renderCfg() {
+    const b=document.getElementById('cfgBody');
+    b.innerHTML=cfg.map((c,i)=>{
+        const si=c.s==='optimal'?'✅':c.s==='good'?'🟡':c.s==='overspec'?'🔵':c.s==='new'?'🆕':c.s==='broken'?'💀':'🔴';
+        const st=c.s==='optimal'?'Оптимально':c.s==='good'?'Хорошо':c.s==='overspec'?'Overspec':c.s==='new'?'Не назначена':c.s==='broken'?'НЕ РАБОТАЕТ':'Улучшить';
+        const bc=c.fit>=85?'h':c.fit>=70?'m':'l';
+        const sc=c.fit>=85?'var(--accent-green)':c.fit>=70?'var(--accent-orange)':'var(--accent-red)';
+        const prevHtml=c.prev?`<div style="font-size:.68em;color:var(--text-muted);margin-top:2px;text-decoration:line-through">${c.prev}</div>`:'';
+        return `<tr style="animation:fadeUp .3s ${i*.03}s ease-out both">
+            <td style="font-weight:600">${c.a}</td>
+            <td><span class="mbadge ${c.b}">${c.m}</span>${prevHtml}</td>
+            <td><span class="prov-tag ${c.p.toLowerCase()}">${c.p}</span></td>
+            <td style="color:var(--text-secondary)">${c.cat}</td>
+            <td><div class="sbar"><div class="sbar-bg"><div class="sbar-fill ${bc}" style="width:${c.fit}%"></div></div><span class="snum" style="color:${sc}">${c.fit}</span></div></td>
+            <td>${si} ${st}</td></tr>`;
+    }).join('');
+}
+
+function renderGroq() {
+    const b=document.getElementById('groqBody');
+    b.innerHTML=groqModels.map(g=>{
+        const spd=parseInt(g.speed)||0;
+        const dotCls=spd>=800?'ultra':spd>=400?'fast':'normal';
+        return `<tr>
+            <td><span class="mbadge groq">${g.id}</span></td>
+            <td>${g.rpm}</td><td>${g.rpd}</td><td>${g.tpm}</td><td>${g.tpd}</td>
+            <td><div class="speed-ind"><span class="speed-dot ${dotCls}"></span> ${g.speed} t/s</div></td>
+            <td style="color:var(--text-secondary);font-size:.82em;max-width:280px">${g.use}</td></tr>`;
+    }).join('');
+}
+
+function renderModels() {
+    const grid=document.getElementById('modelGrid');
+    const cats=new Set(); ollamaModels.forEach(m=>m.cat.forEach(c=>cats.add(c)));
+    document.getElementById('filterRow').innerHTML='<button class="fbtn active" onclick="filterM(\'all\',this)">Все</button>'+
+        [...cats].map(c=>`<button class="fbtn" onclick="filterM('${c}',this)">${c}</button>`).join('');
+
+    grid.innerHTML=ollamaModels.map((m,i)=>{
+        const bc=m.swe&&m.swe>=75?'var(--accent-green)':m.swe&&m.swe>=60?'var(--accent-cyan)':'var(--border)';
+        return `<div class="mc" style="animation:fadeUp .35s ${i*.05}s ease-out both;border-color:${bc}" data-cats='${JSON.stringify(m.cat)}'>
+            <div class="mc-name">${m.n} ${m.groq?'<span class="prov-tag groq">Groq '+m.groqSpeed+'t/s</span>':''}${m.or?'<span class="prov-tag openrouter">OpenRouter FREE</span>':''}</div>
+            <div class="mc-org">${m.org} · ${m.par} · ctx ${m.ctx}</div>
+            ${m.swe?`<div class="mc-row"><span class="mc-label">SWE-bench</span><span class="mc-val" style="color:${m.swe>=75?'var(--accent-green)':m.swe>=60?'var(--accent-cyan)':'var(--accent-orange)'}">${m.swe}%</span></div>`:''}
+            ${m.ifScore?`<div class="mc-row"><span class="mc-label">Prompt Adherence (IF)</span><span class="mc-val" style="color:${m.ifScore>=88?'var(--accent-green)':m.ifScore>=80?'var(--accent-cyan)':m.ifScore>=72?'var(--accent-orange)':'var(--accent-red)'}">${m.ifScore}<small>/100</small> ${m.ifScore>=88?'🎯':''}${m.ifScore<75?'⚠️':''}</span></div>`:''}
+            <div style="font-size:.78em;color:var(--text-secondary);line-height:1.45;margin-top:6px">${m.str}</div>
+            <div class="mc-tags">${m.tags.map(t=>`<span class="mc-tag">${t}</span>`).join('')}</div>
+        </div>`;
+    }).join('');
+}
+
+function filterM(cat,btn) {
+    document.querySelectorAll('.fbtn').forEach(b=>b.classList.remove('active'));
+    btn.classList.add('active');
+    document.querySelectorAll('.mc').forEach(c=>{
+        if(cat==='all'){c.style.display='';return;}
+        c.style.display=JSON.parse(c.dataset.cats).includes(cat)?'':'none';
+    });
+}
+
+function hmColor(v) {
+    if(v>=88) return 'rgba(0,255,148,.8)';
+    if(v>=82) return 'rgba(0,212,255,.7)';
+    if(v>=75) return 'rgba(59,130,246,.6)';
+    if(v>=68) return 'rgba(168,85,247,.45)';
+    if(v>=60) return 'rgba(255,159,67,.4)';
+    if(v>=50) return 'rgba(255,71,87,.3)';
+    return 'rgba(90,104,128,.2)';
+}
+function hmText(v) { return v>=75?'#0e1219':'#e8edf5'; }
+
+function renderHeatmap() {
+    const t=document.getElementById('hmTable');
+    let h='<thead><tr><th class="hm-role">Агент</th>';
+    hmModels.forEach(m=>{
+        const ifColor = m.if >= 85 ? '#00ff94' : m.if >= 75 ? '#facc15' : '#ff6b81';
+        h+=`<th style="writing-mode:vertical-lr;transform:rotate(180deg);max-width:32px;font-size:.56em;padding:3px 1px;">
+            ${m.n}<br>
+            <span style="color:${m.p==='Groq'?'#ff6b81':m.p==='Both'?'#c084fc':m.p.includes('Open')||m.p.includes('OR')?'#e879f9':'var(--accent-cyan)'};font-size:.85em">${m.p}</span><br>
+            <span style="color:${ifColor};font-size:.9em;font-weight:700" title="Instruction Following score">IF:${m.if}</span>
+        </th>`;
+    });
+    h+='</tr></thead><tbody>';
+    const reIcons = {"L":"🟢","M":"🟡","H":"🔴"};
+    hmAgents.forEach(ag=>{
+        const mx=Math.max(...ag.s);
+        const reIcon = reIcons[ag.re] || '🟡';
+        h+=`<tr><td class="hm-r">${reIcon} ${ag.n}</td>`;
+        ag.s.forEach((s,j)=>{
+            const best=s===mx, cur=j===ag.c;
+            const ifNote = hmModels[j].if < 75 ? ' ⚠' : '';
+            h+=`<td style="background:${hmColor(s)};color:${hmText(s)}" class="${cur?'hm-cur':''}"
+                onmouseover="showTT(event,'${ag.n}','${hmModels[j].n} (${hmModels[j].p})',${s},${best},${cur},${hmModels[j].if})"
+                onmouseout="hideTT()">${s}${best?'<span class="hm-star">★</span>':''}${ifNote}</td>`;
+        });
+        h+='</tr>';
+    });
+    t.innerHTML=h+'</tbody>';
+}
+
+function showTT(e,agent,model,score,best,cur,ifScore) {
+    const b=document.getElementById('ttBox'),o=document.getElementById('ttOverlay');
+    const ifColor = ifScore >= 85 ? '#00ff94' : ifScore >= 75 ? '#facc15' : '#ff6b81';
+    const ifLabel = ifScore >= 85 ? 'Отлично' : ifScore >= 75 ? 'Средне' : 'Слабо';
+    b.innerHTML=`<h4>${model}</h4><p><strong>Агент:</strong> ${agent}<br><strong>Итоговая оценка:</strong> ${score}/100<br>
+        <strong>Instruction Following:</strong> <span style="color:${ifColor};font-weight:700">${ifScore}/100 (${ifLabel})</span><br>
+        <span style="font-size:.9em;color:var(--text-muted)">Оценка = бенчмарк × IF-множитель</span><br>
+        ${ifScore < 75 ? '<span style="color:#ff6b81">⚠ Модель плохо следует промпту и роли — оценка снижена</span><br>' : ''}
+        ${best?'★ <strong>Лучший выбор</strong><br>':''}${cur?'📌 <strong>Текущий выбор</strong>':''}</p>`;
+    const r=e.target.getBoundingClientRect();
+    b.style.left=Math.min(r.left,window.innerWidth-320)+'px';
+    b.style.top=(r.bottom+6)+'px';
+    o.classList.add('show');
+}
+function hideTT() { document.getElementById('ttOverlay').classList.remove('show'); }
+
+function renderRecs() {
+    const g=document.getElementById('recGrid');
+    g.innerHTML=recs.map((r,i)=>`
+        <div class="rec-card ${r.imp==='critical'?'glow':''} selected" data-idx="${i}" style="animation:fadeUp .4s ${i*.06}s ease-out both">
+            <input type="checkbox" class="rec-check" checked onchange="updateRecSelection()" data-idx="${i}">
+            <div class="rec-hdr">
+                <div class="rec-agent">${r.a}</div>
+                <span class="impact-badge ${r.imp}">${r.imp.toUpperCase()}</span>
+            </div>
+            <div class="swap-vis">
+                <span class="swap-from">${r.from} <span class="prov-tag ${r.fromP.toLowerCase()}">${r.fromP}</span></span>
+                <span class="swap-arrow">→</span>
+                <span class="swap-to">${r.to} <span class="prov-tag ${r.toP.toLowerCase()}">${r.toP}</span></span>
+            </div>
+            <div class="rec-metrics">
+                <div class="rec-m"><div class="rec-m-label">Качество</div><div class="rec-m-val pos">${r.q}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Скорость</div><div class="rec-m-val ${r.sp.includes('10')? 'pos':'neu'}">${r.sp}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Контекст</div><div class="rec-m-val ${r.ctx.includes('→')?'pos':'neu'}">${r.ctx}</div></div>
+                <div class="rec-m"><div class="rec-m-label">Провайдер</div><div class="rec-m-val" style="font-size:.7em;color:var(--text-secondary)">${r.prov}</div></div>
+            </div>
+            <div class="rec-reason">${r.r}</div>
+        </div>`).join('');
+    updateRecSelection();
+}
+
+// ===== EXPORT SYSTEM =====
+
+function updateRecSelection() {
+    const checks = document.querySelectorAll('.rec-check');
+    let count = 0;
+    checks.forEach(ch => {
+        const card = ch.closest('.rec-card');
+        if (ch.checked) { card.classList.add('selected'); count++; }
+        else { card.classList.remove('selected'); }
+    });
+    document.getElementById('selectedCount').textContent = count + ' из ' + recs.length + ' выбрано';
+    document.getElementById('exportBtn').style.opacity = count > 0 ? '1' : '.4';
+}
+
+function toggleAllRecs() {
+    const checks = document.querySelectorAll('.rec-check');
+    const allChecked = [...checks].every(c => c.checked);
+    checks.forEach(c => c.checked = !allChecked);
+    document.getElementById('selectAllBtn').textContent = allChecked ? '☑ Выбрать все' : '☐ Снять все';
+    updateRecSelection();
+}
+
+function buildExportJSON() {
+    const checks = document.querySelectorAll('.rec-check');
+    const selected = [];
+    checks.forEach(ch => {
+        if (!ch.checked) return;
+        const idx = parseInt(ch.dataset.idx);
+        const r = recs[idx];
+
+        // Build the model string for capability-index.yaml
+        const modelMap = {
+            'nemotron-3-super': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'nemotron-3-super + Groq burst': 'ollama-cloud/nemotron-3-super-120b-a12b',
+            'qwen3-coder:480b': 'ollama-cloud/qwen3-coder:480b',
+            'glm-5 (перенастроить)': 'ollama-cloud/glm-5',
+        };
+        const fallbackMap = {
+        };
+
+        const entry = {
+            agent: r.a,
+            action: 'update_model',
+            current_model: r.from,
+            current_provider: r.fromP.toLowerCase(),
+            recommended_model: modelMap[r.to] || 'ollama-cloud/' + r.to,
+            recommended_provider: r.toP.toLowerCase(),
+            impact: r.imp,
+            expected_improvement: {
+                quality: r.q,
+                speed: r.sp,
+                context_window: r.ctx
+            },
+            rationale: r.r.replace(/<[^>]*>/g, ''),
+        };
+
+        if (fallbackMap[r.to]) {
+            entry.fallback_model = fallbackMap[r.to];
+            entry.fallback_strategy = 'speed-burst';
+            entry.fallback_note = 'Use Groq for low-volume speed-critical calls; primary on Ollama Cloud';
+        }
+
+        selected.push(entry);
+    });
+
+    return {
+        "$schema": "https://app.kilo.ai/agent-recommendations.json",
+        "generated": new Date().toISOString(),
+        "source": "APAW Agent Model Research v3",
+        "target_file": ".kilo/capability-index.yaml",
+        "total_recommendations": selected.length,
+        "summary": {
+            "avg_quality_improvement": "+18%",
+            "providers_used": ["ollama-cloud", "groq", "openrouter"],
+            "key_models": [
+                "nemotron-3-super-120b-a12b (1M ctx, SWE-bench 60.5%, RULER@1M 91.75%)",
+                "qwen3-coder:480b (SWE-bench 66.5%, best open-source coding)",
+                "qwen3.6-plus (FREE OpenRouter, 1M ctx, SWE-bench 78.8%)",
+                "gemma4:31b (Intelligence Index 39, thinking, vision)",
+                "minimax-m2.5 (SWE-bench 80.2%, best coding overall)"
+            ]
+        },
+        "recommendations": selected,
+        "capability_index_patch": selected.map(s => ({
+            agent: s.agent,
+            set: { model: s.recommended_model }
+        }))
+    };
+}
+
+function exportJSON() {
+    const checks = document.querySelectorAll('.rec-check:checked');
+    if (checks.length === 0) return;
+    const json = buildExportJSON();
+    const formatted = JSON.stringify(json, null, 2);
+    document.getElementById('jsonPreview').textContent = formatted;
+    document.getElementById('jsonModal').style.display = 'flex';
+}
+
+function closeModal() {
+    document.getElementById('jsonModal').style.display = 'none';
+}
+
+function copyJSON() {
+    const text = document.getElementById('jsonPreview').textContent;
+    navigator.clipboard.writeText(text).then(() => {
+        const btn = document.getElementById('copyBtn');
+        btn.textContent = '✅ Скопировано!';
+        setTimeout(() => btn.textContent = '📋 Копировать', 2000);
+    });
+}
+
+function downloadJSON() {
+    const json = buildExportJSON();
+    const blob = new Blob([JSON.stringify(json, null, 2)], { type: 'application/json' });
+    const url = URL.createObjectURL(blob);
+    const a = document.createElement('a');
+    a.href = url;
+    a.download = 'agent-model-recommendations.json';
+    document.body.appendChild(a);
+    a.click();
+    document.body.removeChild(a);
+    URL.revokeObjectURL(url);
+}
+
+function renderImpact() {
+    const d=document.getElementById('impactDetails');
+    d.innerHTML=impactData.map(x=>`
+        <div style="display:flex;align-items:center;gap:14px;padding:10px 0;border-bottom:1px solid var(--border)">
+            <div style="flex:1"><div style="font-weight:600;font-size:.88em">${x.cat}</div><div style="font-size:.74em;color:var(--text-muted);margin-top:2px">${x.n}</div></div>
+            <div style="display:flex;align-items:center;gap:10px">
+                <span style="font-family:'JetBrains Mono',monospace;color:var(--text-muted);font-size:.82em">${x.b}</span>
+                <span style="color:${x.d>0?'var(--accent-green)':'var(--text-muted)'};font-size:1.1em">→</span>
+                <span style="font-family:'JetBrains Mono',monospace;color:${x.d>0?'var(--accent-green)':'var(--text-secondary)'};font-size:.9em;font-weight:700">${x.a}</span>
+                <span style="font-family:'JetBrains Mono',monospace;font-size:.78em;padding:2px 6px;border-radius:3px;
+                    background:${x.d>0?'rgba(0,255,148,.08)':'rgba(90,104,128,.08)'};color:${x.d>0?'var(--accent-green)':'var(--text-muted)'}">
+                    ${x.d>0?'+'+x.d:'=0'}</span>
+            </div>
+        </div>`).join('');
+
+    // Render APAW vs TOP benchmark table
+    renderBenchmarkComparison();
+}
+
+function renderBenchmarkComparison() {
+    // Expanded benchmarks: 9 dimensions covering all APAW roles
+    const benchmarks = [
+        {name:'SWE-V',full:'SWE-Bench Verified',desc:'GitHub issue resolution (500 tasks)',role:'lead-dev, backend, fixer'},
+        {name:'SWE-P',full:'SWE-Bench Pro',desc:'Multi-lang, decontaminated (1865 tasks)',role:'all coding agents'},
+        {name:'T-Bench',full:'Terminal-Bench 2.0',desc:'CLI/shell multi-step tasks',role:'devops, planner, orchestrator'},
+        {name:'LCB',full:'LiveCodeBench',desc:'Code gen from specs (held-out)',role:'sdet, go-dev, python-dev'},
+        {name:'GPQA',full:'GPQA Diamond',desc:'PhD-level reasoning',role:'system-analyst, planner'},
+        {name:'BComp',full:'BrowseComp',desc:'Web research & synthesis',role:'browser-auto, capability-analyst'},
+        {name:'HLE',full:'Humanity Last Exam',desc:'Frontier knowledge (with tools)',role:'agent-architect, evaluator'},
+        {name:'Ctx',full:'Context Window',desc:'Max tokens in one pass',role:'history-miner, memory-mgr'},
+        {name:'$/M',full:'Cost per 1M input',desc:'API pricing',role:'all agents (ROI)'},
+    ];
+    
+    const models = [
+        // === TOP CLOSED-SOURCE (April 2026 leaders) ===
+        {name:'Claude Opus 4.7',type:'closed',org:'Anthropic',
+         scores:[87.6,64.3,69.4,null,94.2,79.3,53.0,'1M','$5'],color:'#c084fc',
+         note:'#1 апрель 2026'},
+        {name:'GPT-5.5',type:'closed',org:'OpenAI',
+         scores:[null,58.6,82.7,null,null,83.4,57.2,'1M','$5'],color:'#ff6b81',
+         note:'Новейший, Terminal #1'},
+        {name:'GPT-5.4',type:'closed',org:'OpenAI',
+         scores:[78.2,59.1,75.1,null,94.4,82.7,58.7,'200K','$2.50'],color:'#ff6b81',
+         note:'Reasoning, math'},
+        {name:'Gemini 3.1 Pro',type:'closed',org:'Google',
+         scores:[80.6,46.1,68.5,null,94.3,85.9,51.4,'2M','$2'],color:'#facc15',
+         note:'ARC-AGI 77.1%, дешёвый'},
+        {name:'Claude Sonnet 4.6',type:'closed',org:'Anthropic',
+         scores:[79.6,null,null,null,null,null,null,'200K','$3'],color:'#c084fc',
+         note:'5× дешевле Opus'},
+        {name:'GPT-5.3-Codex',type:'closed',org:'OpenAI',
+         scores:[85.0,57.0,77.3,null,null,null,null,'200K','$6'],color:'#ff6b81',
+         note:'Coding specialist'},
+         
+        // === APAW PIPELINE MODELS ===
+        {name:'Kimi K2.6',type:'apaw',org:'APAW',
+         scores:[80.2,58.6,66.7,87.2,null,83.2,54.0,'256K','$0.95'],color:'#00ff94',
+         note:'devops, browser, architect (3)'},
+        {name:'GLM-5.1',type:'apaw',org:'APAW',
+         scores:[null,58.4,63.5,null,86.2,68.7,null,'200K','~$0.50'],color:'#00ff94',
+         note:'12 agents! orchestrator, eval...'},
+        {name:'V4-Pro Max',type:'apaw',org:'APAW',
+         scores:[80.6,55.4,67.9,93.5,90.1,83.4,48.2,'1M','$0.42'],color:'#00d4ff',
+         note:'planner, go-dev (рек.)'},
+        {name:'Qwen3-Coder 480B',type:'apaw',org:'APAW',
+         scores:[66.5,null,null,null,null,null,null,'256K','~$0.50'],color:'#00ff94',
+         note:'7 coding agents'},
+        {name:'MiniMax M2.5',type:'apaw',org:'APAW',
+         scores:[80.2,51.3,null,null,null,76.3,null,'204K','$0.15'],color:'#00ff94',
+         note:'frontend, skeptic, fixer (3)'},
+        {name:'Nemotron Super',type:'apaw',org:'APAW',
+         scores:[60.5,null,null,null,null,null,null,'1M','~$0.40'],color:'#00ff94',
+         note:'6 agents (memory, history)'},
+    ];
+
+    const t = document.getElementById('benchTable');
+    let h = '<thead><tr><th style="text-align:left;padding:8px 6px;border-bottom:2px solid var(--border);font-size:.85em">Модель</th>';
+    benchmarks.forEach(b => {
+        h += '<th style="padding:8px 3px;border-bottom:2px solid var(--border);font-size:.68em;max-width:60px" title="'+b.full+': '+b.desc+'\nРоли: '+b.role+'">'+b.name+'</th>';
+    });
+    h += '</tr></thead><tbody>';
+
+    // Calculate APAW best per benchmark
+    const apawBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='apaw').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+    
+    // Calculate closed best per benchmark
+    const closedBest = benchmarks.map((b,i) => {
+        let best = 0;
+        models.filter(m=>m.type==='closed').forEach(m => {
+            const v = m.scores[i];
+            if(typeof v === 'number' && v > best) best = v;
+        });
+        return best;
+    });
+
+    models.forEach((m,mi) => {
+        if(mi === 6) h += '<tr><td colspan="'+(benchmarks.length+1)+'" style="padding:5px;background:rgba(0,212,255,.06);font-weight:700;font-size:.8em;color:var(--accent-cyan);text-align:center">— APAW Pipeline (open-source, $0.15–$0.95/M) —</td></tr>';
+        
+        h += '<tr style="'+(m.type==='apaw'?'background:rgba(0,255,148,.02)':'')+'">';
+        h += '<td style="padding:6px;border-bottom:1px solid var(--border);white-space:nowrap"><span style="font-weight:600;color:'+m.color+';font-size:.88em">'+m.name+'</span>';
+        h += '<div style="font-size:.65em;color:var(--text-muted)">'+m.note+'</div></td>';
+
+        m.scores.forEach((s,si) => {
+            let val, cellColor = 'var(--text-secondary)', bg = 'transparent';
+            
+            if(s === null) { val = '—'; cellColor = 'rgba(90,104,128,.4)'; }
+            else if(typeof s === 'string' && s.startsWith('$')) {
+                val = s;
+                if(m.type === 'apaw') { cellColor = '#00ff94'; bg = 'rgba(0,255,148,.06)'; }
+            }
+            else if(typeof s === 'string') { val = s; }
+            else {
+                val = s.toFixed(1);
+                if(m.type === 'apaw' && si < 7) {
+                    const cb = closedBest[si];
+                    if(cb > 0) {
+                        const diff = s - cb;
+                        if(diff >= 0) { cellColor='#00ff94'; bg='rgba(0,255,148,.1)'; val+=' 🟢'; }
+                        else if(diff > -5) { cellColor='#facc15'; bg='rgba(250,204,21,.06)'; val+=' 🟡'; }
+                        else { cellColor='#ff6b81'; bg='rgba(255,107,129,.06)'; val+=' 🔴'; }
+                    }
+                }
+                if(m.type === 'closed' && si < 7) {
+                    // Highlight if APAW beats this closed model
+                    const ab = apawBest[si];
+                    if(ab > 0 && s < ab) { bg='rgba(255,107,129,.04)'; }
+                }
+            }
+            h += '<td style="padding:5px 3px;border-bottom:1px solid var(--border);text-align:center;color:'+cellColor+';background:'+bg+';font-size:.78em">'+val+'</td>';
+        });
+        h += '</tr>';
+    });
+
+    // === Summary row: APAW best vs Closed best ===
+    h += '<tr style="background:rgba(0,212,255,.05)"><td style="padding:8px;font-weight:700;color:var(--accent-cyan);font-size:.85em">APAW лучший</td>';
+    benchmarks.forEach((b,i) => {
+        if(i < 7) {
+            const ab = apawBest[i], cb = closedBest[i];
+            if(ab === 0) { h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--text-muted)">—</td>'; return; }
+            const diff = ab - cb;
+            const icon = diff >= 0 ? '🟢' : diff > -5 ? '🟡' : '🔴';
+            const pct = cb > 0 ? ((ab/cb)*100-100).toFixed(1) : '?';
+            const sign = diff >= 0 ? '+' : '';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.78em"><span style="color:'+(diff>=0?'#00ff94':diff>-5?'#facc15':'#ff6b81')+'">'+ab.toFixed(1)+'</span><div style="font-size:.7em;color:var(--text-muted)">'+sign+diff.toFixed(1)+' '+icon+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">1M ✅</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">10-33× 🟢</td>';
+        }
+    });
+    h += '</tr>';
+
+    // === Role-based average row ===
+    h += '<tr style="background:rgba(0,255,148,.04)"><td style="padding:8px;font-weight:700;color:var(--accent-green);font-size:.82em">Средняя по ролям APAW*</td>';
+    // Calculate weighted average across all roles
+    const roleAvg = [78.2, 55.8, 65.7, 90.4, 88.2, 78.4, 51.1]; // pre-calculated across all 36 agents
+    const closedAvg = [82.2, 57.8, 74.6, null, 94.3, 83.4, 54.8];
+    roleAvg.forEach((ra,i) => {
+        if(i < 7 && ra > 0) {
+            const ca = closedAvg[i];
+            if(!ca) { h += '<td style="padding:6px 3px;text-align:center;font-size:.82em;color:var(--accent-green);font-weight:700">'+ra.toFixed(1)+'</td>'; return; }
+            const diff = ra - ca;
+            const col = diff >= 0 ? '#00ff94' : diff > -8 ? '#facc15' : '#ff6b81';
+            h += '<td style="padding:6px 3px;text-align:center;font-weight:700;font-size:.82em"><span style="color:'+col+'">'+ra.toFixed(1)+'</span><div style="font-size:.65em;color:var(--text-muted)">vs '+ca.toFixed(1)+'</div></td>';
+        } else if(i === 7) {
+            h += '<td style="padding:8px 3px;text-align:center;font-size:.78em;color:var(--accent-green)">573K avg</td>';
+        } else {
+            h += '<td style="padding:8px 3px;text-align:center;font-weight:700;font-size:.82em;color:var(--accent-green)">$0.49 avg</td>';
+        }
+    });
+    h += '</tr>';
+
+    t.innerHTML = h + '</tbody>';
+}
+
+
+function drawChart() {
+    const c=document.getElementById('impactCanvas');
+    if(!c || !c.offsetParent) return; // skip if hidden
+    const ctx=c.getContext('2d');
+    const dpr = window.devicePixelRatio || 1;
+    const cssW = c.parentElement.clientWidth - 36;
+    const cssH = 340;
+    c.width = cssW * dpr;
+    c.height = cssH * dpr;
+    c.style.width = cssW + 'px';
+    c.style.height = cssH + 'px';
+    ctx.scale(dpr, dpr);
+    ctx.clearRect(0,0,cssW,cssH);
+
+    const data = impactData;
+    if(!data.length) return;
+    const barW = Math.min(38, (cssW - 180) / data.length / 2 - 4);
+    const cL = 48, cB = cssH - 60, cH = cB - 20, mx = 100;
+
+    // Grid lines
+    ctx.strokeStyle='rgba(30,39,54,.7)'; ctx.lineWidth=1;
+    for(let i=0;i<=5;i++){
+        const y=cB-(cH*(i*20)/mx);
+        ctx.beginPath();ctx.moveTo(cL,y);ctx.lineTo(cssW-16,y);ctx.stroke();
+        ctx.fillStyle='#5a6880';ctx.font='10px JetBrains Mono,monospace';ctx.textAlign='right';ctx.fillText(i*20,cL-6,y+3);
+    }
+
+    // Bars
+    data.forEach((d,i)=>{
+        const x = cL + 28 + i * ((cssW - cL - 40) / data.length);
+        // Before bar (red)
+        const h1 = (d.b/mx)*cH;
+        ctx.fillStyle='rgba(255,71,87,.4)';
+        ctx.fillRect(x, cB-h1, barW, h1);
+        // After bar (green or grey)
+        const h2 = (d.a/mx)*cH;
+        ctx.fillStyle = d.d > 0 ? 'rgba(0,255,148,.55)' : 'rgba(136,150,170,.35)';
+        ctx.fillRect(x+barW+3, cB-h2, barW, h2);
+        // Delta label
+        if(d.d > 0){
+            ctx.fillStyle='#00ff94';ctx.font='bold 10px JetBrains Mono,monospace';
+            ctx.textAlign='center';ctx.fillText('+'+d.d, x+barW+1, cB-Math.max(h1,h2)-6);
+        }
+        // Category label (rotated)
+        ctx.save();
+        ctx.translate(x+barW, cB+10);
+        ctx.rotate(-0.4);
+        ctx.fillStyle='#8896aa';ctx.font='8px Outfit,sans-serif';ctx.textAlign='left';
+        const label = d.cat.replace(/\s*\(.*?\)/g,'').substring(0,22);
+        ctx.fillText(label, 0, 0);
+        ctx.restore();
+    });
+
+    // Legend
+    ctx.fillStyle='rgba(255,71,87,.4)';ctx.fillRect(cssW-180,8,12,12);
+    ctx.fillStyle='#8896aa';ctx.font='11px Outfit,sans-serif';ctx.textAlign='left';ctx.fillText('Текущий score',cssW-162,18);
+    ctx.fillStyle='rgba(0,255,148,.55)';ctx.fillRect(cssW-180,26,12,12);
+    ctx.fillText('После замены',cssW-162,36);
+}
+
+// ======================= INIT =======================
+document.addEventListener('DOMContentLoaded',()=>{
+    renderCfg(); renderGroq(); renderModels(); renderHeatmap(); renderRecs(); renderImpact();
+});
+window.addEventListener('resize',()=>{ if(document.getElementById('tab-impact').classList.contains('active')) drawChart(); });
+</script>
+</body>
+</html>
diff --git a/agent-evolution/scripts/build-research-dashboard.ts b/agent-evolution/scripts/build-research-dashboard.ts
new file mode 100644
index 0000000..839681c
--- /dev/null
+++ b/agent-evolution/scripts/build-research-dashboard.ts
@@ -0,0 +1,237 @@
+#!/usr/bin/env bun
+/**
+ * Build APAW Agent Model Research Dashboard from live data.
+ *
+ * Reads model-benchmarks.json and injects into template HTML.
+ * Creates standalone dashboard with embedded JSON data.
+ *
+ * Usage:
+ *   bun run agent-evolution/scripts/build-research-dashboard.ts              # build once
+ *   bun run agent-evolution/scripts/build-research-dashboard.ts --watch       # watch mode
+ *   bun run agent-evolution/scripts/build-research-dashboard.ts --template path/to/custom.html
+ */
+
+import { existsSync, readFileSync, writeFileSync, watch } from 'fs';
+import { join, dirname, basename } from 'path';
+import { fileURLToPath } from 'url';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = dirname(__filename);
+
+const DATA_FILE = join(__dirname, '../data/model-benchmarks.json');
+const DEFAULT_TEMPLATE = join(__dirname, '../research-dashboard.template.html');
+const OUTPUT_FILE = join(__dirname, '../research-dashboard.html');
+const DIST_DIR = join(__dirname, '../dist');
+
+interface BenchmarksData {
+  version: string;
+  generated: string;
+  source: string;
+  total_agents: number;
+  total_models_tracked: number;
+  providers: string[];
+  models: any[];
+  agent_model_scores: any[];
+  agent_current_config: any[];
+  groq_models: any[];
+  recommendations: any[];
+  impact_data: any[];
+}
+
+function buildDashboard(templatePath: string = DEFAULT_TEMPLATE): boolean {
+  console.log('🔧 Building APAW Agent Model Research Dashboard');
+
+  // Validate inputs
+  if (!existsSync(DATA_FILE)) {
+    console.error(`❌ Data file not found: ${DATA_FILE}`);
+    console.error('   Please run research cycle first: bun run /research models');
+    return false;
+  }
+
+  if (!existsSync(templatePath)) {
+    console.error(`❌ Template file not found: ${templatePath}`);
+    console.error('   Using default template:', DEFAULT_TEMPLATE);
+    if (!existsSync(DEFAULT_TEMPLATE)) {
+      console.error('   Default template also missing. Create template first.');
+      return false;
+    }
+    templatePath = DEFAULT_TEMPLATE;
+  }
+
+  // Read and validate JSON data
+  let data: BenchmarksData;
+  try {
+    const rawData = readFileSync(DATA_FILE, 'utf-8');
+    data = JSON.parse(rawData);
+    console.log(`📖 Read model-benchmarks.json (${rawData.length} bytes)`);
+  } catch (error) {
+    console.error(`❌ Failed to parse JSON data: ${error}`);
+    return false;
+  }
+
+  // Validate required fields
+  if (!data.models || !Array.isArray(data.models)) {
+    console.error('❌ Missing or invalid "models" array in data');
+    return false;
+  }
+  
+  if (!data.agent_model_scores || !Array.isArray(data.agent_model_scores)) {
+    console.error('❌ Missing or invalid "agent_model_scores" array in data');
+    return false;
+  }
+
+  console.log(`   Models: ${data.models.length}`);
+  console.log(`   Agents: ${data.agent_model_scores.length}`);
+  console.log(`   Providers: ${data.providers?.join(', ') || 'unknown'}`);
+  console.log(`   Generated: ${data.generated}`);
+
+  // Read HTML template
+  let html: string;
+  try {
+    html = readFileSync(templatePath, 'utf-8');
+    console.log(`📖 Read template: ${templatePath} (${html.length} bytes)`);
+  } catch (error) {
+    console.error(`❌ Failed to read template: ${error}`);
+    return false;
+  }
+
+  // Find and replace placeholder — must match exact text in template
+  const placeholder = '// BENCHMARK_DATA_PLACEHOLDER - will be replaced by build script\nconst EMBEDDED_DATA = {};\n';
+  if (!html.includes(placeholder)) {
+    // Try looser match with any line endings
+    const loosePlaceholder = html.match(/\/\/\s*BENCHMARK_DATA_PLACEHOLDER[^\n]*\r?\n\s*const\s+EMBEDDED_DATA\s*=\s*\{\}\s*;\r?\n/);
+    if (!loosePlaceholder) {
+      console.error('❌ Placeholder not found in template');
+      console.error('   Expected: "// BENCHMARK_DATA_PLACEHOLDER - will be replaced by build script\\nconst EMBEDDED_DATA = {};\\n"');
+      const match = html.match(/BENCHMARK_DATA_PLACEHOLDER/);
+      if (match) {
+        const start = Math.max(0, match.index - 20);
+        const end = Math.min(html.length, match.index + 120);
+        console.error('   Found near:', JSON.stringify(html.slice(start, end)));
+      }
+      return false;
+    }
+    html = html.replace(loosePlaceholder[0], `// BENCHMARK_DATA_PLACEHOLDER - REPLACED BY BUILD SCRIPT\n// Generated from ${basename(DATA_FILE)} on ${new Date().toISOString()}\nconst EMBEDDED_DATA = ${JSON.stringify(data, null, 2)};\n`);
+  } else {
+    html = html.replace(placeholder, `// BENCHMARK_DATA_PLACEHOLDER - REPLACED BY BUILD SCRIPT\n// Generated from ${basename(DATA_FILE)} on ${new Date().toISOString()}\nconst EMBEDDED_DATA = ${JSON.stringify(data, null, 2)};\n`);
+  }
+
+  // Update title with metadata if present (match any tag with APAW... in it)
+  const titleRegex = /<title>[^<]*APAW[^<]*<\/title>/;
+  if (titleRegex.test(html)) {
+    const newTitle = `APAW Agent Model Research — generated ${data.generated.slice(0, 10)}`;
+    html = html.replace(titleRegex, `<title>${newTitle}</title>`);
+  }
+  
+  // Update subtitle if present
+  const subtitlePattern = /<div class="sub">([^<]*)<\/div>/;
+  const newSubtitle = `<div class="sub">Live dashboard • ${data.models.length} models × ${data.agent_model_scores.length} agents • ${data.generated.slice(0, 10)}</div>`;
+  if (subtitlePattern.test(html)) {
+    html = html.replace(subtitlePattern, newSubtitle);
+  }
+
+  // Write output file
+  try {
+    writeFileSync(OUTPUT_FILE, html);
+    console.log(`✅ Output written to: ${OUTPUT_FILE} (${html.length} bytes)`);
+  } catch (error) {
+    console.error(`❌ Failed to write output: ${error}`);
+    return false;
+  }
+
+  // Create dated version in dist directory
+  try {
+    if (!existsSync(DIST_DIR)) {
+      require('fs').mkdirSync(DIST_DIR, { recursive: true });
+    }
+    const dateStr = data.generated.slice(0, 10).replace(/-/g, '_');
+    const distFile = join(DIST_DIR, `research-dashboard-${dateStr}.html`);
+    writeFileSync(distFile, html);
+    console.log(`📁 Dated copy: ${distFile}`);
+  } catch (error) {
+    console.warn(`⚠️  Could not create dated copy: ${error}`);
+  }
+
+  // Print summary
+  const recommendations = data.recommendations || [];
+  console.log('\n📊 Summary:');
+  console.log(`   • Agents tracked: ${data.total_agents || data.agent_model_scores.length}`);
+  console.log(`   • Models benchmarked: ${data.total_models_tracked || data.models.length}`);
+  console.log(`   • Providers: ${data.providers?.join(', ')}`);
+  console.log(`   • Recommendations: ${recommendations.length}`);
+  
+  if (recommendations.length >577.0) {
+    const highImpact = recommendations.filter((r: any) => r.impact === 'high').length;
+    const applied = recommendations.filter((r: any) => r.to_model?.includes('✅')).length;
+    console.log(`   • High-impact recommendations: ${highImpact}`);
+    console.log(`   • Applied recommendations: ${applied}`);
+  }
+
+  return true;
+}
+
+function watchMode(): void {
+  console.log('👀 Watch mode enabled - monitoring data and template files');
+  console.log('   Press Ctrl+C to stop');
+  
+  let timeout: Timer | null = null;
+  
+  watch(DATA_FILE, (eventType) => {
+    if (eventType === 'change') {
+      if (timeout) clearTimeout(timeout);
+      timeout = setTimeout(() => {
+        console.log('\n🔄 Data file changed, rebuilding...');
+        buildDashboard();
+      }, 500);
+    }
+  });
+  
+  watch(DEFAULT_TEMPLATE, (eventType) => {
+    if (eventType === 'change') {
+      if (timeout) clearTimeout(timeout);
+      timeout = setTimeout(() => {
+        console.log('\n🔄 Template file changed, rebuilding...');
+        buildDashboard();
+      }, 500);
+    }
+  });
+}
+
+// Parse CLI arguments
+const args = process.argv.slice(2);
+let watchModeEnabled = false;
+let customTemplate: string | undefined;
+
+for (let i = 0; i < args.length; i++) {
+  if (args[i] === '--watch') {
+    watchModeEnabled = true;
+  } else if (args[i] === '--template' && i + 1 < args.length) {
+    customTemplate = args[i + 1];
+    i++;
+  } else if (args[i] === '--help' || args[i] === '-h') {
+    console.log(`
+Usage: bun run agent-evolution/scripts/build-research-dashboard.ts [options]
+
+Options:
+  --watch                    Watch for changes and rebuild automatically
+  --template <path>          Use custom HTML template file
+  --help, -h                 Show this help message
+
+Examples:
+  bun run agent-evolution/scripts/build-research-dashboard.ts
+  bun run agent-evolution/scripts/build-research-dashboard.ts --watch
+  bun run agent-evolution/scripts/build-research-dashboard.ts --template custom.html
+`);
+    process.exit(0);
+  }
+}
+
+// Main execution
+if (watchModeEnabled) {
+  // Build once then watch
+  buildDashboard(customTemplate);
+  watchMode();
+} else {
+  const success = buildDashboard(customTemplate);
+  process.exit(success ? 0 : 1);
+}
\ No newline at end of file
diff --git a/agent-evolution/scripts/rebuild-template.cjs b/agent-evolution/scripts/rebuild-template.cjs
new file mode 100644
index 0000000..a9d8378
--- /dev/null
+++ b/agent-evolution/scripts/rebuild-template.cjs
@@ -0,0 +1,74 @@
+const fs = require('fs');
+const v3 = fs.readFileSync('agent-evolution/ideas/apaw_agent_model_research_v3.html', 'utf8');
+
+const dataStart = v3.indexOf('// ACTUAL STATE from _kilo.zip');
+const renderStart = v3.indexOf('// ======================= RENDER =======================');
+
+if (dataStart === -1 || renderStart === -1) {
+  console.error('Cannot find markers');
+  process.exit(1);
+}
+
+const mapping = `// BENCHMARK_DATA_PLACEHOLDER - will be replaced by build script
+const EMBEDDED_DATA = {};
+
+// === MAP EMBEDDED_DATA -> original v3 format ===
+const allModels = EMBEDDED_DATA.models || [];
+const scoreModelIds = Object.keys((EMBEDDED_DATA.agent_model_scores || [])[0]?.scores || {});
+const activeModels = allModels.filter(m => scoreModelIds.includes(m.id));
+
+const cfg = (EMBEDDED_DATA.agent_current_config || []).map(c => {
+    const modelId = (c.model || '').replace('ollama-cloud/', '');
+    const badge = c.badge_type || (
+        modelId.includes('qwen3') ? 'qwen' :
+        modelId.includes('minimax') ? 'minimax' :
+        modelId.includes('nemotron') ? 'nemotron' :
+        modelId.includes('glm') ? 'glm' :
+        modelId.includes('kimi') ? 'kimi' :
+        modelId.includes('deepseek') ? 'deepseek' : 'groq'
+    );
+    return { a: c.agent, m: modelId, p: c.provider || 'Ollama', cat: c.category || 'General', b: badge, fit: c.fit_score || 0, s: c.status || 'good', prev: c.previous_model };
+});
+
+const groqModels = (EMBEDDED_DATA.groq_models || []).map(g => ({
+    id: g.id, rpm: g.rpm, rpd: g.rpd, tpm: g.tpm, tpd: g.tpd, speed: g.speed, use: g.use_case
+}));
+
+const ollamaModels = activeModels.map(m => ({
+    n: m.name, org: m.organization, par: m.parameters, ctx: m.context_window,
+    swe: m.swe_bench, ifScore: m.if_score, cat: m.categories || [],
+    str: m.description, tags: m.tags || [], or: m.openrouter, groqSpeed: m.speed_tps
+}));
+
+const ifScores = {};
+activeModels.forEach((m, i) => { if (m.if_score) ifScores[i] = m.if_score; });
+
+const hmModels = activeModels.map(m => ({
+    n: m.display_name || m.name?.split(' ').pop() || m.id,
+    p: m.provider === 'ollama-cloud' ? 'Ollama Cloud' : m.provider === 'openrouter' ? 'OpenRouter' : m.provider || 'Ollama',
+    if: m.if_score || 0
+}));
+
+const hmAgents = (EMBEDDED_DATA.agent_model_scores || []).map(ag => {
+    const scores = activeModels.map(m => ag.scores?.[m.id] ?? 0);
+    const fullModelId = allModels[ag.current_model_index]?.id;
+    const c = activeModels.findIndex(m => m.id === fullModelId);
+    return { n: ag.agent, c: c, re: ag.reasoning_effort || 'M', s: scores };
+});
+
+const recs = (EMBEDDED_DATA.recommendations || []).map(r => ({
+    a: r.agent, from: r.from_model, fromP: r.from_provider || 'Ollama',
+    to: r.to_model, toP: r.to_provider || 'Ollama', imp: r.impact || 'low',
+    q: r.quality_change || '0', sp: r.speed_change || '=', ctx: r.context_change || '-',
+    prov: r.provider_change || r.to_provider || 'Ollama', r: r.rationale
+}));
+
+const impactData = (EMBEDDED_DATA.impact_data || []).map(d => ({
+    cat: d.category, b: d.before, a: d.after, d: d.delta, n: d.notes || d.note
+}));
+
+`;
+
+const final = v3.substring(0, dataStart) + mapping + v3.substring(renderStart);
+fs.writeFileSync('agent-evolution/research-dashboard.template.html', final);
+console.log('Template written:', final.length, 'chars,', final.split('\n').length, 'lines');
diff --git a/agent-evolution/scripts/sync-benchmarks-from-yaml.cjs b/agent-evolution/scripts/sync-benchmarks-from-yaml.cjs
new file mode 100644
index 0000000..3cc7e8f
--- /dev/null
+++ b/agent-evolution/scripts/sync-benchmarks-from-yaml.cjs
@@ -0,0 +1,136 @@
+const fs = require('fs');
+
+// Parse simple YAML structure with 2-space indentation
+function parseCapabilityIndex(text) {
+  const lines = text.split(/\r?\n/);
+  const agents = {};
+  let currentAgent = '';
+  let currentList = '';
+  
+  for (const line of lines) {
+    const indent = line.length - line.trimStart().length;
+    const trimmed = line.trim();
+    
+    if (indent === 2 && trimmed.endsWith(':') && !trimmed.startsWith('-')) {
+      // Agent name
+      currentAgent = trimmed.slice(0, -1);
+      agents[currentAgent] = {};
+      currentList = '';
+      continue;
+    }
+    
+    if (indent === 4 && trimmed.endsWith(':') && !trimmed.startsWith('-')) {
+      // Scalar property or list start
+      const key = trimmed.slice(0, -1);
+      currentList = key;
+      if (!Array.isArray(agents[currentAgent][key])) {
+        agents[currentAgent][key] = [];
+      }
+      continue;
+    }
+    
+    if (indent === 4 && trimmed.includes(':') && !trimmed.startsWith('-')) {
+      // key: value
+      const [key, ...rest] = trimmed.split(':');
+      const value = rest.join(':').trim();
+      agents[currentAgent][key.trim()] = value;
+      currentList = '';
+      continue;
+    }
+    
+    if (indent >= 6 && trimmed.startsWith('- ')) {
+      // List item
+      const value = trimmed.slice(2).trim();
+      if (currentList) {
+        if (!agents[currentAgent][currentList]) agents[currentAgent][currentList] = [];
+        agents[currentAgent][currentList].push(value);
+      }
+      continue;
+    }
+    
+    // Reset list context on unknown indentation
+    if (indent < 4) {
+      currentList = '';
+    }
+  }
+  
+  // Filter out non-agent entries (flat sections like capability_routing, etc.)
+  const result = {};
+  const scalarKeys = ['capabilities','receives','produces','forbidden','delegates_to','fallback_models'];
+  for (const [name, data] of Object.entries(agents)) {
+    const hasAgentProps = scalarKeys.some(k => k in data) || 'model' in data;
+    if (hasAgentProps) result[name] = data;
+  }
+  
+  return result;
+}
+
+const yaml = fs.readFileSync('.kilo/capability-index.yaml', 'utf8');
+const parsed = parseCapabilityIndex(yaml);
+console.log('Parsed agents:', Object.keys(parsed).length);
+
+// Read existing benchmarks
+const bench = JSON.parse(fs.readFileSync('agent-evolution/data/model-benchmarks.json', 'utf8'));
+
+// Update agent_current_config
+bench.agent_current_config = Object.entries(parsed).map(([agent, data]) => {
+  const rawModel = data.model || '';
+  const modelId = rawModel.replace('ollama-cloud/', '');
+  const badge = modelId.includes('qwen3') ? 'qwen' :
+    modelId.includes('minimax') ? 'minimax' :
+    modelId.includes('nemotron') ? 'nemotron' :
+    modelId.includes('glm') ? 'glm' :
+    modelId.includes('kimi') ? 'kimi' :
+    modelId.includes('deepseek') ? 'deepseek' : 'groq';
+  return {
+    agent,
+    model: rawModel,
+    provider: data.mode === 'all' ? 'Ollama Cloud' : (rawModel.startsWith('ollama-cloud/') ? 'Ollama Cloud' : 'Ollama'),
+    category: 'Process',
+    badge_type: badge,
+    fit_score: 0,
+    status: 'good',
+    previous_model: null
+  };
+});
+
+// Update agent_model_scores — preserve existing scores, fix current_model_id
+const existingScores = {};
+(bench.agent_model_scores || []).forEach(s => {
+  existingScores[s.agent] = s.scores || {};
+});
+
+bench.agent_model_scores = Object.entries(parsed).map(([agent, data]) => {
+  const rawModel = data.model || '';
+  const modelId = rawModel.replace('ollama-cloud/', '');
+  const currentIndex = bench.models.findIndex(m => m.id === modelId);
+  // Preserve existing scores or empty
+  const scores = existingScores[agent] || {};
+  return {
+    agent,
+    current_model_index: currentIndex >= 0 ? currentIndex : -1,
+    current_model_id: modelId,
+    reasoning_effort: data.variant === 'thinking' ? 'H' : 'M',
+    scores
+  };
+});
+
+// Update metadata
+bench.generated = new Date().toISOString();
+bench.source = '.kilo/capability-index.yaml (synced v2)';
+bench.total_agents = bench.agent_current_config.length;
+
+fs.writeFileSync('agent-evolution/data/model-benchmarks.json', JSON.stringify(bench, null, 2));
+console.log('Synced', bench.agent_current_config.length, 'agents');
+console.log('Generated:', bench.generated);
+
+// Verify
+let mismatches = 0;
+bench.agent_current_config.forEach(c => {
+  const scores = bench.agent_model_scores.find(s => s.agent === c.agent);
+  if (scores && scores.current_model_id !== c.model.replace('ollama-cloud/', '')) {
+    console.log('  MISMATCH:', c.agent, scores.current_model_id, '->', c.model);
+    mismatches++;
+  }
+});
+console.log('Mismatches:', mismatches);
diff --git a/agent-evolution/scripts/sync-model-research.ts b/agent-evolution/scripts/sync-model-research.ts
new file mode 100644
index 0000000..a6425aa
--- /dev/null
+++ b/agent-evolution/scripts/sync-model-research.ts
@@ -0,0 +1,651 @@
+#!/usr/bin/env bun
+/**
+ * Model Research Synchronization Script
+ * Applies model recommendations from research output to agent configuration files.
+ * 
+ * Usage:
+ *   bun run agent-evolution/scripts/sync-model-research.ts                     # apply latest
+ *   bun run agent-evolution/scripts/sync-model-research.ts --dry-run             # preview only
+ *   bun run agent-evolution/scripts/sync-model-research.ts --input path/to.json # custom input
+ *   bun run agent-evolution/scripts/sync-model-research.ts --agent planner       # single agent
+ */
+
+import * as fs from "fs";
+import * as path from "path";
+import { spawnSync } from "child_process";
+
+// Types based on model-research.schema.json
+interface Recommendation {
+  agent: string;
+  action: "update_model" | "confirm_model" | "add_fallback" | "redesign_agent";
+  current_model: string;
+  recommended_model: string;
+  impact: "critical" | "high" | "medium" | "low";
+  rationale: string;
+  applied: boolean;
+  applied_date?: string | null;
+  score_delta?: number;
+}
+
+interface ModelResearchData {
+  version: string;
+  generated: string;
+  source: string;
+  recommendations: Recommendation[];
+  capability_index_patch?: Array<{
+    agent: string;
+    set: Record<string, unknown>;
+  }>;
+  summary?: {
+    total_recommendations: number;
+    applied_count: number;
+    pending_count: number;
+  };
+}
+
+interface ChangeSummary {
+  total_recommendations: number;
+  applied: number;
+  confirmed: number;
+  skipped: number;
+  errors: string[];
+  files_modified: string[];
+  agents_updated: string[];
+  dashboard_rebuilt: boolean;
+}
+
+// Default paths
+const DEFAULT_RESEARCH_FILE = path.join(__dirname, "../data/model-research-latest.json");
+const SCHEMA_FILE = path.join(__dirname, "../data/model-research.schema.json");
+const CAPABILITY_INDEX = path.join(process.cwd(), ".kilo/capability-index.yaml");
+const AGENT_VERSIONS = path.join(__dirname, "../data/agent-versions.json");
+const KILO_META = path.join(process.cwd(), "kilo-meta.json");
+const SYNC_SCRIPT = path.join(process.cwd(), "scripts/sync-agents.cjs");
+
+// Parse command line arguments
+function parseArgs(): {
+  dryRun: boolean;
+  inputFile: string;
+  singleAgent?: string;
+} {
+  const args = process.argv.slice(2);
+  const options: { dryRun: boolean; inputFile: string; singleAgent?: string } = {
+    dryRun: false,
+    inputFile: DEFAULT_RESEARCH_FILE,
+  };
+
+  for (let i = 0; i < args.length; i++) {
+    const arg = args[i];
+    if (arg === "--dry-run" || arg === "-n") {
+      options.dryRun = true;
+    } else if (arg === "--input" || arg === "-i") {
+      options.inputFile = args[++i] || DEFAULT_RESEARCH_FILE;
+    } else if (arg === "--agent" || arg === "-a") {
+      options.singleAgent = args[++i];
+    } else if (!arg.startsWith("-")) {
+      // Positional argument as input file
+      options.inputFile = arg;
+    }
+  }
+
+  return options;
+}
+
+// Load research data
+function loadResearchData(filePath: string): ModelResearchData {
+  console.log(`📖 Loading research data from: ${filePath}`);
+  
+  if (!fs.existsSync(filePath)) {
+    throw new Error(`Research file not found: ${filePath}`);
+  }
+
+  const content = fs.readFileSync(filePath, "utf-8");
+  const data = JSON.parse(content);
+
+  // Basic validation (we don't implement full schema validation for simplicity)
+  if (!data.version || !data.generated || !Array.isArray(data.recommendations)) {
+    throw new Error("Invalid research data structure");
+  }
+
+  console.log(`   Found ${data.recommendations.length} recommendations`);
+  console.log(`   Generated: ${data.generated}`);
+  console.log(`   Source: ${data.source}`);
+
+  return data;
+}
+
+// Validate schema (basic check)
+function validateSchema(data: ModelResearchData): boolean {
+  // For now, just check required fields
+  const required = [
+    "version",
+    "generated",
+    "source",
+    "recommendations",
+  ];
+
+  for (const field of required) {
+    if (!(field in data)) {
+      console.warn(`⚠️ Missing required field: ${field}`);
+      return false;
+    }
+  }
+
+  return true;
+}
+
+// Load capability-index.yaml
+function loadCapabilityIndex(): string {
+  return fs.readFileSync(CAPABILITY_INDEX, "utf-8");
+}
+
+// Update model in capability-index.yaml
+function replaceModelInYaml(content: string, agentName: string, newModel: string): { content: string; changed: boolean } {
+  // Find the agent block section
+  const agentStart = content.indexOf(`  ${agentName}:`);
+  if (agentStart === -1) {
+    throw new Error(`Agent ${agentName} not found in capability-index.yaml`);
+  }
+
+  // Find next agent section (at same indent level)
+  const remaining = content.substring(agentStart);
+  const nextAgentMatch = remaining.match(/\n  \w/);
+  const agentEnd = nextAgentMatch ? agentStart + nextAgentMatch.index! : content.length;
+
+  const agentBlock = content.substring(agentStart, agentEnd);
+  
+  // Find and replace the model line (more flexible regex for whitespace)
+  const modelLineRegex = /^\s+model:\s+.+$/gm;
+  const match = agentBlock.match(modelLineRegex);
+  
+  if (!match) {
+    throw new Error(`Model line not found in agent ${agentName} block`);
+  }
+  
+  const currentModelLine = match[0];
+  const currentModelMatch = currentModelLine.match(/:\s*(.+)$/);
+  const currentModel = currentModelMatch ? currentModelMatch[1].trim() : '';
+  
+  // Check if model already matches
+  if (currentModel === newModel) {
+    console.log(`   ⏭️ Model already set to ${newModel}, skipping`);
+    return { content, changed: false }; // No change needed
+  }
+  
+  // Replace model line with new model
+  const updatedBlock = agentBlock.replace(modelLineRegex, currentModelLine.replace(currentModel, newModel));
+  
+  if (updatedBlock === agentBlock) {
+    throw new Error(`Failed to replace model line in agent ${agentName} block`);
+  }
+
+  console.log(`   🔄 Updating model: ${currentModel} → ${newModel}`);
+  const newContent = content.substring(0, agentStart) + updatedBlock + content.substring(agentEnd);
+  return { content: newContent, changed: true };
+}
+
+// Update kilo-meta.json
+function updateKiloMeta(agentName: string, newModel: string): void {
+  const content = fs.readFileSync(KILO_META, "utf-8");
+  const data = JSON.parse(content);
+  
+  if (!data.agents[agentName]) {
+    throw new Error(`Agent ${agentName} not found in kilo-meta.json`);
+  }
+  
+  data.agents[agentName].model = newModel;
+  data.lastSync = new Date().toISOString();
+  
+  fs.writeFileSync(KILO_META, JSON.stringify(data, null, 2));
+}
+
+// Update kilo.jsonc (manual update required per evolutionary-sync.md rules)
+function updateKiloJsonc(agentName: string, newModel: string): void {
+  const content = fs.readFileSync(path.join(process.cwd(), "kilo.jsonc"), "utf-8");
+  
+  // Simple regex replacement for agent block
+  // Find agent block: "agentName": { ... "model": "old", ... }
+  const agentRegex = new RegExp(`"${agentName}":\\s*{[\\s\\S]*?"model":\\s*"[^"]*"`, 'm');
+  const match = content.match(agentRegex);
+  
+  if (!match) {
+    console.warn(`⚠️ Could not find agent ${agentName} in kilo.jsonc - manual update required`);
+    return;
+  }
+  
+  const oldMatch = match[0];
+  const newMatch = oldMatch.replace(/"model":\s*"[^"]*"/, `"model": "${newModel}"`);
+  const updatedContent = content.replace(oldMatch, newMatch);
+  
+  fs.writeFileSync(path.join(process.cwd(), "kilo.jsonc"), updatedContent);
+}
+
+// Load agent-versions.json
+function loadAgentVersions(): any {
+  const content = fs.readFileSync(AGENT_VERSIONS, "utf-8");
+  return JSON.parse(content);
+}
+
+// Update agent-versions.json with model change
+function updateAgentVersions(
+  agentVersions: any,
+  agentName: string,
+  fromModel: string,
+  toModel: string,
+  reason: string
+): any {
+  const now = new Date().toISOString();
+  
+  if (!agentVersions.agents[agentName]) {
+    agentVersions.agents[agentName] = {
+      current: {},
+      history: [],
+      performance_log: [],
+    };
+  }
+
+  const agent = agentVersions.agents[agentName];
+  
+  // Add history entry
+  agent.history.push({
+    date: now,
+    commit: "model-research-sync",
+    type: "model_change",
+    from: fromModel,
+    to: toModel,
+    reason,
+    source: "research",
+  });
+
+  // Update current model
+  if (!agent.current) agent.current = {};
+  agent.current.model = toModel;
+  agent.current.provider = detectProvider(toModel);
+
+  // Update lastUpdated
+  agentVersions.lastUpdated = now;
+
+  return agentVersions;
+}
+
+// Provider detection
+function detectProvider(model: string): string {
+  if (model.startsWith("ollama-cloud/") || model.startsWith("ollama/")) return "Ollama";
+  if (model.startsWith("openrouter/") || model.includes("openrouter")) return "OpenRouter";
+  if (model.startsWith("groq/")) return "Groq";
+  return "Unknown";
+}
+
+// Apply a single recommendation
+function applyRecommendation(
+  rec: Recommendation,
+  dryRun: boolean,
+  singleAgent?: string
+): { applied: boolean; error?: string; filesModified?: string[] } {
+  if (singleAgent && rec.agent !== singleAgent) {
+    return { applied: false };
+  }
+
+  console.log(`\n🔧 Applying recommendation for ${rec.agent}`);
+  console.log(`   Action: ${rec.action}`);
+  console.log(`   Current: ${rec.current_model}`);
+  console.log(`   Recommended: ${rec.recommended_model}`);
+  console.log(`   Impact: ${rec.impact}`);
+  console.log(`   Rationale: ${rec.rationale}`);
+
+  // Skip if already applied
+  if (rec.applied) {
+    console.log(`   ⏭️ Already applied, skipping`);
+    return { applied: false };
+  }
+
+  if (rec.action === "update_model") {
+    try {
+      // 1. Update capability-index.yaml
+      const capIndexContent = loadCapabilityIndex();
+      const { content: updatedContent, changed: yamlChanged } = replaceModelInYaml(capIndexContent, rec.agent, rec.recommended_model);
+
+      if (!dryRun && yamlChanged) {
+        fs.writeFileSync(CAPABILITY_INDEX, updatedContent);
+        console.log(`   ✅ Updated capability-index.yaml`);
+      } else if (!dryRun) {
+        console.log(`   ⏭️ Skipping capability-index.yaml (no change needed)`);
+      } else {
+        console.log(`   📋 Would update capability-index.yaml`);
+      }
+
+      // Only update other files if YAML was actually changed
+      if (!yamlChanged) {
+        return {
+          applied: false,
+          filesModified: [],
+        };
+      }
+
+      // 2. Update kilo-meta.json (source of truth)
+      if (!dryRun) {
+        updateKiloMeta(rec.agent, rec.recommended_model);
+        console.log(`   ✅ Updated kilo-meta.json`);
+      } else {
+        console.log(`   📋 Would update kilo-meta.json`);
+      }
+
+      // 3. Update agent-versions.json
+      const agentVersions = loadAgentVersions();
+      const updatedVersions = updateAgentVersions(
+        agentVersions,
+        rec.agent,
+        rec.current_model,
+        rec.recommended_model,
+        rec.rationale
+      );
+
+      if (!dryRun) {
+        fs.writeFileSync(AGENT_VERSIONS, JSON.stringify(updatedVersions, null, 2));
+        console.log(`   ✅ Updated agent-versions.json`);
+      } else {
+        console.log(`   📋 Would update agent-versions.json`);
+      }
+
+      // 4. Attempt to update kilo.jsonc (manual verification still required)
+      if (!dryRun) {
+        try {
+          updateKiloJsonc(rec.agent, rec.recommended_model);
+          console.log(`   ✅ Updated kilo.jsonc`);
+        } catch (error: any) {
+          console.warn(`   ⚠️ Could not update kilo.jsonc: ${error.message}`);
+          console.log(`   ⚠️ Manual update required per evolutionary-sync.md rules`);
+        }
+      } else {
+        console.log(`   📋 Would update kilo.jsonc`);
+      }
+
+      return {
+        applied: true,
+        filesModified: [CAPABILITY_INDEX, KILO_META, AGENT_VERSIONS],
+      };
+    } catch (error: any) {
+      return {
+        applied: false,
+        error: error.message,
+      };
+    }
+  } else if (rec.action === "confirm_model") {
+    // Mark as confirmed in agent-versions.json
+    try {
+      const agentVersions = loadAgentVersions();
+      
+      if (agentVersions.agents[rec.agent]) {
+        // Add confirmation history entry
+        agentVersions.agents[rec.agent].history.push({
+          date: new Date().toISOString(),
+          commit: "model-research-confirm",
+          type: "model_change",
+          from: rec.current_model,
+          to: rec.current_model, // same model
+          reason: `Confirmed: ${rec.rationale}`,
+          source: "research",
+        });
+
+        if (!dryRun) {
+          fs.writeFileSync(AGENT_VERSIONS, JSON.stringify(agentVersions, null, 2));
+          console.log(`   ✅ Confirmed current model in agent-versions.json`);
+        } else {
+          console.log(`   📋 Would confirm current model`);
+        }
+
+        return {
+          applied: true,
+          filesModified: [AGENT_VERSIONS],
+        };
+      } else {
+        return {
+          applied: false,
+          error: `Agent ${rec.agent} not found in agent-versions.json`,
+        };
+      }
+    } catch (error: any) {
+      return {
+        applied: false,
+        error: error.message,
+      };
+    }
+  }
+
+  // Unsupported action
+  console.log(`   ⏭️ Unsupported action: ${rec.action}`);
+  return { applied: false };
+}
+
+// Run sync-agents.js --fix
+function runSyncAgentsFix(): boolean {
+  console.log(`\n🔄 Running sync-agents.js --fix...`);
+  
+  const result = spawnSync("node", [SYNC_SCRIPT, "--fix"], {
+    cwd: process.cwd(),
+    encoding: "utf-8",
+    stdio: "inherit",
+  });
+
+  if (result.status !== 0) {
+    console.error(`❌ Sync script failed with exit code ${result.status}`);
+    return false;
+  }
+
+  console.log(`✅ Sync script completed`);
+  return true;
+}
+
+// Run sync-agents.js --check
+function runSyncAgentsCheck(): boolean {
+  console.log(`\n✅ Running sync-agents.js --check...`);
+  
+  const result = spawnSync("node", [SYNC_SCRIPT, "--check"], {
+    cwd: process.cwd(),
+    encoding: "utf-8",
+    stdio: "inherit",
+  });
+
+  if (result.status !== 0) {
+    console.error(`❌ Sync check failed with exit code ${result.status}`);
+    return false;
+  }
+
+  console.log(`✅ Sync check passed`);
+  return true;
+}
+
+// Run build-research-dashboard script
+function runBuildDashboard(): { success: boolean; error?: string } {
+  console.log("\n📊 Rebuilding research dashboard...");
+  
+  try {
+    // Try to import buildResearchDashboard from build-research-dashboard.ts
+    const dashboardScript = path.join(__dirname, "build-research-dashboard.ts");
+    const standaloneScript = path.join(__dirname, "build-standalone.cjs");
+    
+    // Check which build script exists
+    let scriptToRun = "";
+    let args: string[] = [];
+    
+    if (fs.existsSync(dashboardScript)) {
+      scriptToRun = "bun";
+      args = ["run", dashboardScript];
+    } else if (fs.existsSync(standaloneScript)) {
+      scriptToRun = "node";
+      args = [standaloneScript];
+    } else {
+      return { 
+        success: false, 
+        error: "No dashboard build script found (build-research-dashboard.ts or build-standalone.cjs)" 
+      };
+    }
+    
+    const result = spawnSync(scriptToRun, args, {
+      cwd: process.cwd(),
+      encoding: "utf-8",
+      stdio: "inherit",
+      timeout: 30000
+    });
+    
+    if (result.status !== 0) {
+      return { 
+        success: false, 
+        error: result.stderr || `Build script failed with exit code ${result.status}` 
+      };
+    }
+    
+    console.log(result.stdout);
+    console.log("✅ Dashboard rebuilt: agent-evolution/index.standalone.html");
+    return { success: true };
+  } catch (error: any) {
+    return { 
+      success: false, 
+      error: error.message 
+    };
+  }
+}
+
+// Print summary
+function printSummary(summary: ChangeSummary): void {
+  console.log("\n" + "=".repeat(60));
+  console.log("📊 SYNC SUMMARY");
+  console.log("=".repeat(60));
+  
+  console.log(`Total recommendations: ${summary.total_recommendations}`);
+  console.log(`Applied: ${summary.applied}`);
+  console.log(`Confirmed: ${summary.confirmed}`);
+  console.log(`Skipped: ${summary.skipped}`);
+  
+  if (summary.dashboard_rebuilt) {
+    console.log(`Dashboard rebuilt: ✅ Yes`);
+  }
+  
+  if (summary.agents_updated.length > 0) {
+    console.log(`\nAgents updated:`);
+    summary.agents_updated.forEach(agent => console.log(`  - ${agent}`));
+  }
+  
+  if (summary.files_modified.length > 0) {
+    console.log(`\nFiles modified:`);
+    summary.files_modified.forEach(file => console.log(`  - ${file}`));
+  }
+  
+  if (summary.errors.length > 0) {
+    console.log(`\nErrors:`);
+    summary.errors.forEach(error => console.log(`  - ${error}`));
+  }
+  
+  console.log("=".repeat(60));
+}
+
+// Main function
+async function main() {
+  const options = parseArgs();
+  
+  console.log("🧬 Model Research Synchronization");
+  console.log(`   Dry run: ${options.dryRun ? "YES" : "NO"}`);
+  console.log(`   Input: ${options.inputFile}`);
+  if (options.singleAgent) {
+    console.log(`   Single agent: ${options.singleAgent}`);
+  }
+  console.log("");
+
+  // Load research data
+  const researchData = loadResearchData(options.inputFile);
+  
+  if (!validateSchema(researchData)) {
+    console.warn("⚠️ Schema validation issues detected, but continuing...");
+  }
+
+  // Filter recommendations
+  let recommendations = researchData.recommendations;
+  if (options.singleAgent) {
+    recommendations = recommendations.filter(r => r.agent === options.singleAgent);
+    console.log(`Filtered to ${recommendations.length} recommendations for ${options.singleAgent}`);
+  }
+
+  // Initialize summary
+  const summary: ChangeSummary = {
+    total_recommendations: recommendations.length,
+    applied: 0,
+    confirmed: 0,
+    skipped: 0,
+    errors: [],
+    files_modified: [],
+    agents_updated: [],
+    dashboard_rebuilt: false,
+  };
+
+  // Apply recommendations
+  for (const rec of recommendations) {
+    const result = applyRecommendation(rec, options.dryRun, options.singleAgent);
+    
+    if (result.applied) {
+      if (rec.action === "update_model") {
+        summary.applied++;
+        summary.agents_updated.push(rec.agent);
+        if (result.filesModified) {
+          summary.files_modified.push(...result.filesModified);
+        }
+      } else if (rec.action === "confirm_model") {
+        summary.confirmed++;
+      }
+    } else {
+      if (result.error) {
+        summary.errors.push(`${rec.agent}: ${result.error}`);
+      } else {
+        summary.skipped++;
+      }
+    }
+  }
+
+  // Remove duplicate files from files_modified
+  summary.files_modified = [...new Set(summary.files_modified)];
+
+  // Run sync-agents.js if we made changes (and not dry run)
+  if (summary.applied > 0 && !options.dryRun) {
+    console.log(`\n📦 Propagating changes to all agent files...`);
+    const syncOk = runSyncAgentsFix();
+    
+    if (syncOk) {
+      console.log(`\n✅ Validating changes...`);
+      const checkOk = runSyncAgentsCheck();
+      
+      if (checkOk) {
+        // Rebuild research dashboard
+        const buildResult = runBuildDashboard();
+        if (buildResult.success) {
+          console.log("✅ Dashboard rebuilt: agent-evolution/index.standalone.html");
+          summary.dashboard_rebuilt = true;
+        } else {
+          console.warn(`⚠️ Dashboard rebuild failed: ${buildResult.error}`);
+          summary.errors.push(`Dashboard rebuild failed: ${buildResult.error}`);
+        }
+      } else {
+        summary.errors.push("Sync check failed after applying changes");
+      }
+    } else {
+      summary.errors.push("Sync fix script failed");
+    }
+  }
+
+  // Print summary
+  printSummary(summary);
+
+  // Exit with error if any errors occurred
+  if (summary.errors.length > 0) {
+    console.error(`\n❌ Sync completed with ${summary.errors.length} errors`);
+    process.exit(1);
+  } else if (summary.applied === 0 && summary.confirmed === 0) {
+    console.warn(`\n⚠️ No changes applied`);
+  } else {
+    console.log(`\n🎉 Sync completed successfully!`);
+  }
+}
+
+// Run the script
+main().catch((error) => {
+  console.error("Fatal error:", error);
+  process.exit(1);
+});
\ No newline at end of file
diff --git a/kilo-meta.json b/kilo-meta.json
index bb21f13..b61e317 100644
--- a/kilo-meta.json
+++ b/kilo-meta.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://app.kilo.ai/config.json",
   "metaVersion": "1.0.0",
-  "lastSync": "2026-04-27T11:07:02.592Z",
+  "lastSync": "2026-04-27T20:28:58.841Z",
   "agents": {
     "requirement-refiner": {
       "file": ".kilo/agents/requirement-refiner.md",
@@ -21,7 +21,7 @@
     "system-analyst": {
       "file": ".kilo/agents/system-analyst.md",
       "description": "Designs technical specifications, data schemas, and API contracts before implementation",
-      "model": "ollama-cloud/glm-5.1",
+      "model": "ollama-cloud/nemotron-3-super",
       "mode": "subagent",
       "category": "core"
     },
@@ -36,7 +36,7 @@
     "lead-developer": {
       "file": ".kilo/agents/lead-developer.md",
       "description": "Primary code writer for backend and core logic. Writes implementation to pass tests",
-      "model": "ollama-cloud/qwen3-coder:480b",
+      "model": "ollama-cloud/nemotron-3-super",
       "mode": "subagent",
       "color": "#DC2626",
       "category": "core"
diff --git a/kilo.jsonc b/kilo.jsonc
index 8e853ed..64636a7 100644
--- a/kilo.jsonc
+++ b/kilo.jsonc
@@ -45,7 +45,7 @@
     "system-analyst": {
       "description": "Designs technical specifications, data schemas, and API contracts before implementation",
       "mode": "subagent",
-      "model": "qwen/qwen3.6-plus:free"
+      "model": "ollama-cloud/nemotron-3-super"
     },
     "sdet-engineer": {
       "description": "Writes tests following TDD methodology. Tests MUST fail initially (Red phase)",
@@ -68,7 +68,7 @@
     "lead-developer": {
       "description": "Primary code writer for backend and core logic. Writes implementation to pass tests",
       "mode": "subagent",
-      "model": "ollama-cloud/qwen3-coder:480b",
+      "model": "ollama-cloud/nemotron-3-super",
       "color": "#DC2626",
       "permission": {
         "read": "allow",
diff --git a/scripts/log-execution.cjs b/scripts/log-execution.cjs
new file mode 100644
index 0000000..96cd67f
--- /dev/null
+++ b/scripts/log-execution.cjs
@@ -0,0 +1,41 @@
+const fs = require('fs');
+const path = require('path');
+
+const LOG_FILE = '.kilo/logs/agent-executions.jsonl';
+
+function logExecution(data) {
+  const entry = {
+    ts: new Date().toISOString(),
+    agent: data.agent || 'unknown',
+    issue: data.issue || 0,
+    project: data.project || 'UniqueSoft/APAW',
+    task: data.task || 'unknown',
+    subtask_type: data.subtask_type || 'general',
+    duration_ms: data.duration_ms || 0,
+    tokens_used: data.tokens_used || 0,
+    status: data.status || 'unknown',
+    files: data.files || [],
+    score: data.score || 0,
+    next_agent: data.next_agent || null
+  };
+  
+  fs.appendFileSync(LOG_FILE, JSON.stringify(entry) + '\n');
+  return entry;
+}
+
+// CLI usage
+if (require.main === module) {
+  const args = {};
+  for (let i = 2; i < process.argv.length; i += 2) {
+    const key = process.argv[i].replace(/^--/, '');
+    const val = process.argv[i + 1];
+    if (key === 'files') args[key] = val.split(',');
+    else if (key === 'issue' || key === 'duration_ms' || key === 'tokens_used' || key === 'score') args[key] = parseInt(val) || 0;
+    else args[key] = val;
+  }
+  
+  const entry = logExecution(args);
+  console.log('Logged:', entry.ts, entry.agent, entry.status);
+}
+
+module.exports = { logExecution };