GoClaw/todo.md

# GoClaw Control Center TODO

- [x] Basic Dashboard layout (Mission Control theme)
- [x] Agents page with mock data
- [x] Nodes page with mock data
- [x] Chat page with mock conversation
- [x] Settings page with provider cards
- [x] Docker Stack integration
- [x] Fix Home.tsx conflict after upgrade
- [x] Fix DashboardLayout.tsx conflict after upgrade
- [x] Create server-side Ollama API proxy routes (tRPC)
- [x] Integrate real Ollama /v1/models endpoint in Settings
- [x] Integrate real Ollama /v1/chat/completions in Chat page
- [x] Add OLLAMA_API_KEY and OLLAMA_BASE_URL secrets
- [x] Write vitest tests for Ollama API proxy
- [x] Update Dashboard with real model data
- [ ] Add streaming support for chat responses
- [ ] Connect real Docker Swarm API for node monitoring
- [ ] Add authentication/login protection

## Phase 1: Agent Management UI
- [x] Connect Agents page to trpc.agents.list (load real agents from DB)
- [x] Create AgentDetailModal component for viewing agent config
- [x] Create AgentCreateModal component with form validation
- [x] Implement agent update mutation (model, temperature, maxTokens, systemPrompt)
- [x] Implement agent delete mutation with confirmation
- [x] Add start/pause/restart actions for agents
- [x] Add agent metrics chart (requests, tokens, processing time)
- [x] Add agent history view (recent requests/responses)
- [x] Write vitest tests for agent management components

## Phase 2: Tool Binding System
- [x] Design Tool Binding API schema
- [x] Create tool registry in database
- [x] Implement tool execution sandbox
- [x] Add tool access control per agent
- [x] Create UI for tool management

## Phase 3: Tool Integration
- [x] Implement Browser tool (HTTP fetch-based)
- [x] Implement Shell tool (bash execution with safety checks)
- [x] Implement File tool (read/write with path restrictions)
- [x] Implement Docker tool (container management)
- [x] Implement HTTP tool (GET/POST with domain whitelist)

## Phase 4: Metrics & History
- [x] AgentMetrics page with request timeline chart
- [x] Conversation history log per agent
- [x] Raw metrics table with token/time data
- [x] Stats cards (total requests, success rate, avg response time, tokens)
- [x] Time range selector (6h/24h/48h/7d)
- [x] Metrics button on agent cards
- [x] Navigation: /agents/:id/metrics route
- [x] Tools page added to sidebar navigation

## Phase 5: Specialized Agents

### Browser Agent
- [ ] Install puppeteer-core + chromium dependencies
- [ ] Create server/browser-agent.ts — Puppeteer session manager
- [ ] tRPC routes: browser.start, browser.navigate, browser.screenshot, browser.click, browser.type, browser.extract, browser.close
- [ ] BrowserAgent.tsx page — live browser control UI with screenshot preview
- [ ] Session management: multiple concurrent browser sessions per agent
- [ ] Add browser_agent to agents DB as pre-seeded entry

### Tool Builder Agent
- [ ] Create server/tool-builder.ts — LLM-powered tool generator
- [ ] tRPC routes: toolBuilder.generate, toolBuilder.validate, toolBuilder.install
- [ ] Dynamic tool registration: add generated tools to TOOL_REGISTRY at runtime
- [ ] Persist custom tools to DB (tool_definitions table)
- [ ] ToolBuilder.tsx page — describe tool → preview code → install
- [ ] Add tool_builder_agent to agents DB as pre-seeded entry

### Agent Compiler
- [ ] Create server/agent-compiler.ts — LLM-powered agent factory
- [ ] tRPC routes: agentCompiler.compile, agentCompiler.preview, agentCompiler.deploy
- [ ] AgentCompiler.tsx page — ТЗ input → agent config preview → deploy
- [ ] Auto-populate: model, role, systemPrompt, allowedTools from ТЗ
- [ ] Add agent_compiler to agents DB as pre-seeded entry

### Integration
- [ ] Add all 3 pages to sidebar navigation
- [ ] Write vitest tests for all new server modules
- [ ] Push to Gitea (NW)

## Phase 6: Agents as Real Chat Entities
- [ ] Remove unused pages: BrowserAgent.tsx, ToolBuilder.tsx, AgentCompiler.tsx
- [ ] Seed 3 agents into DB: Browser Agent, Tool Builder Agent, Agent Compiler
- [ ] Add tRPC chat endpoint: agents.chat (LLM + tool execution per agent)
- [ ] Update Chat UI to support agent selection dropdown
- [ ] Create /skills page — skills registry with install/uninstall
- [ ] Update /agents to show seeded agents with Chat button
- [ ] Update /tools to show tools per agent with filter by agent
- [ ] Add /skills to sidebar navigation
- [ ] Write tests for chat and skills endpoints

## Phase 6: Orchestrator Agent (Main Chat)
- [x] Fix TS errors: browserSessions/toolDefinitions schema exports, z.record
- [x] Seed 3 specialized agents into DB (Browser, Tool Builder, Agent Compiler)
- [x] Create server/orchestrator.ts — main orchestrator with tool-use loop
- [x] Orchestrator tools: shell_exec, file_read, file_write, http_request, delegate_to_agent, list_agents, list_skills, install_skill
- [x] Add trpc.orchestrator.chat mutation (multi-step tool-use loop with LLM)
- [x] Update /chat UI: show tool call steps, agent delegation, streaming response
- [x] Create /skills page with skill registry (install/remove/describe)
- [x] Add /skills to sidebar navigation
- [x] Update /agents to show seeded agents with Chat button
- [ ] Write tests for orchestrator

## Phase 7: Orchestrator as Configurable System Agent
- [x] Add isSystem + isOrchestrator fields to agents table (DB migration)
- [x] Seed Orchestrator as system agent in DB (role=orchestrator, isSystem=true)
- [x] Update orchestrator.ts to load model/systemPrompt/allowedTools from DB
- [x] Update /chat to read orchestrator config from DB, show active model in header
- [x] Update /agents to show Orchestrator with SYSTEM badge, Configure button, no delete
- [x] AgentDetailModal: orchestrator gets extra tab with system tools (shell, docker, agents mgmt)
- [x] Add system tools to orchestrator: docker_ps, docker_restart, manage_agents, read_logs
- [x] /chat header: show current model name + link to Configure Orchestrator

## Phase 8: Fix Orchestrator Chat
- [x] Fix: orchestrator uses model from DB config (minimax-m2.7, not hardcoded fallback)
- [x] Fix: real tool-use loop — execute shell_exec, file_read, file_list tools
- [x] Fix: show tool call steps in Chat UI (tool name, args, result, duration)
- [x] Fix: Chat.tsx shows which model is being used from orchestrator config
- [x] Fix: Streamdown markdown rendering for assistant responses
- [ ] Add: streaming/SSE for real-time response display

## Phase 9: Go Gateway Migration (Variant C)
- [x] Create gateway/ directory with Go module (git.softuniq.eu/UniqAI/GoClaw/gateway)
- [x] Implement config/config.go — env-based configuration
- [x] Implement internal/llm/client.go — Ollama API client (chat, models, health)
- [x] Implement internal/db/db.go — MySQL connection, agent/config queries
- [x] Implement internal/tools/executor.go — Tool Executor (shell_exec, file_read, file_write, file_list, http_request, docker_exec, list_agents)
- [x] Implement internal/orchestrator/orchestrator.go — LLM tool-use loop, config from DB
- [x] Implement internal/api/handlers.go — REST API handlers
- [x] Implement cmd/gateway/main.go — HTTP server with chi router, graceful shutdown
- [x] Go Gateway compiles successfully (10.8MB binary)
- [x] Create server/gateway-proxy.ts — Node.js proxy client to Go Gateway
- [x] Create docker/docker-compose.yml — local dev (control-center + gateway + ollama + db)
- [x] Create docker/docker-stack.yml — Docker Swarm production (2 replicas, rolling updates)
- [x] Create docker/Dockerfile.gateway — multi-stage Go build
- [x] Create docker/Dockerfile.control-center — multi-stage Node.js build
- [ ] Update server/routers.ts: replace orchestrator.ts calls with gateway-proxy.ts calls
- [ ] Write Go unit tests (gateway/internal/tools/executor_test.go)
- [ ] Write Go integration test for orchestrator chat loop
- [ ] Push to Gitea (NW)

## Phase 10: LLM Provider Configuration
- [x] config.go: default LLM_BASE_URL = https://ollama.com/v1 (Ollama Cloud)
- [x] config.go: support LLM_BASE_URL + LLM_API_KEY env vars (legacy OLLAMA_* aliases kept)
- [x] config.go: normaliseLLMURL() — auto-append /v1 for bare Ollama hosts
- [x] docker-compose.yml: ollama service commented out (GPU only), LLM_BASE_URL/LLM_API_KEY added
- [x] docker-stack.yml: ollama service commented out (GPU only), llm-api-key secret added
- [x] docker/.env.example: 4 LLM provider options documented (Ollama Cloud, OpenAI, Groq, Local GPU)

## Phase 11: Frontend → Go Gateway Integration
- [x] gateway-proxy.ts: fix getGatewayTools() — map OpenAI format {type,function:{name,...}} to GatewayToolDef
- [x] gateway-proxy.ts: add executeGatewayTool(), getGatewayAgent(), isGatewayAvailable() methods
- [x] routers.ts: orchestrator.getConfig — Go Gateway first, Node.js fallback
- [x] routers.ts: orchestrator.chat — Go Gateway first, Node.js fallback
- [x] routers.ts: orchestrator.tools — Go Gateway first, Node.js fallback
- [x] routers.ts: orchestrator.gatewayHealth — new endpoint for UI status
- [x] routers.ts: ollama.health — Go Gateway first, direct Ollama fallback
- [x] routers.ts: ollama.models — Go Gateway first, direct Ollama fallback
- [x] gateway/db.go: TLS auto-detection for TiDB Cloud (tidbcloud/aws/gcp/azure hosts)
- [x] server/gateway-proxy.test.ts: 13 vitest tests (health, config, tools, mapping)
- [x] End-to-end test: orchestrator.chat via tRPC → Go Gateway → Ollama (source: "gateway")
- [x] End-to-end test: tool calling — file_list tool executed by Go Gateway

## Phase 12: Real-time Nodes Page
- [ ] Add Docker API client in Go Gateway: /api/nodes endpoint with real node data
- [ ] Add /api/nodes/stats endpoint for CPU/memory per node
- [ ] Add tRPC nodes.list and nodes.stats procedures via gateway-proxy
- [ ] Update Nodes.tsx: real data from tRPC + auto-refresh every 5 seconds
- [ ] Show: node ID, hostname, status, role (manager/worker), availability, CPU, RAM, Docker version, IP
- [ ] Show live indicator (green pulse) when data is fresh
- [x] Deploy to server 2.59.219.61
- [x] Docker API client: /api/nodes, /api/nodes/stats
- [x] tRPC nodes.list, nodes.stats procedures
- [x] Nodes.tsx rewritten with real data + auto-refresh 10s/15s
- [x] 14 vitest tests for nodes procedures

## Phase 13: Seed Data for Agents & Orchestrator
- [x] Create server/seed.ts with default agents (orchestrator, coder, browser, researcher)
- [x] Create default orchestrator config seed
- [x] Integrate seed into server startup (idempotent — runs only when tables are empty)
- [x] Write vitest tests for seed logic (18 tests, all pass)
- [x] Commit to Gitea and deploy to production server
- [x] Verify seed data on production DB — 6 agents seeded successfully

## Phase 14: Auto-migrate on Container Startup
- [ ] Create server/migrate.ts — programmatic Drizzle migration runner
- [ ] Create docker/entrypoint.sh — wait-for-db + migrate + start server
- [ ] Update Dockerfile.control-center — copy entrypoint, set as CMD
- [ ] Write vitest tests for migrate logic
- [ ] Commit to Gitea and deploy to production server
- [ ] Verify auto-migrate on production (check logs)

## Phase 14 (Bug Fixes): Real Header Metrics + Seed Fix
- [x] Fix seed: agents not appearing on production after restart (check isSystem column query)
- [x] Fix header metrics: UPTIME/NODES/AGENTS/CPU/MEM show hardcoded data instead of real values
- [x] Connect header stats to real tRPC endpoints (agents count from DB, nodes/CPU/MEM from Docker API)
- [x] Write vitest tests for header stats procedure (82 tests total, all pass)
- [x] Commit to Gitea and deploy to production (Phase 14) — verified: nodes=6/6, agents=6, CPU=0.2%, MEM=645MB, gatewayOnline=true

## Phase 15 (Bug Fix): Agents Page Shows Empty List
- [x] Diagnose: find why /agents page shows no agents (userId=0 in seed vs SYSTEM_USER_ID=1 in router)
- [x] Fix agents tRPC query: getAllAgents() instead of getUserAgents(SYSTEM_USER_ID)
- [x] Update vitest tests (86 tests, all pass)
- [ ] Deploy to production (Phase 15)

## Phase 16: Bug Fixes & Code Quality (2026-03-21)

### Исправления ошибок
- [x] Fix AgentDetailModal: список моделей теперь загружается из реального API (trpc.ollama.models) с индикатором загрузки и фоллбэком на текущую модель агента при недоступности API
- [x] Fix AgentCreateModal: убрана некорректная фильтрация по провайдеру; список моделей берётся из API напрямую; disabled state во время загрузки
- [x] Fix Chat 401 unauthorized: Go Gateway теперь проверяет доступность модели из БД через API перед отправкой запроса; при отсутствии модели автоматически переключается на первую доступную (resolveModel)
- [x] Fix gateway/orchestrator.go: добавлено поле modelWarning в ChatResult для уведомления фронта о смене модели
- [x] Fix gateway-proxy.ts: добавлено поле modelWarning в GatewayChatResult
- [x] Fix Chat.tsx: добавлено отображение modelWarning в виде amber badge рядом с именем модели

### Замечания (технический долг) → закрыто в Phase 17
- [x] Dashboard.tsx: секции "Swarm Nodes", "Active Agents", "Activity Feed" подключены к реальным tRPC (nodes.list, nodes.stats, agents.list, dashboard.stats) — моки NODES/AGENTS/ACTIVITY_LOG удалены
- [x] server/index.ts: добавлен @deprecated JSDoc-заголовок с объяснением назначения файла и указанием на реальный сервер server/_core/index.ts
- [x] Phase 9 TODO: проверено — orchestrator.ts вызовы в routers.ts заменены на gateway-proxy.ts, пункт актуальности снят

## Phase 17: Technical Debt Closure (2026-03-21)

### Исправлено
- [x] Dashboard.tsx полностью переведён на реальные данные:
  - nodes.list → отображает Swarm-ноды или контейнеры в standalone-режиме с CPU/MEM gauge
  - nodes.stats → live CPU% и MEM для каждого контейнера
  - agents.list → реальные агенты с isActive/isSystem/model/role
  - dashboard.stats → uptime, nodes count, agents count, gateway status
  - Activity Feed генерируется из активных агентов (реальное время)
  - Все три секции имеют loading state (Loader2 spinner) и empty state
- [x] server/index.ts: задокументирован как @deprecated legacy static-only entry point,
  с указанием: реальный сервер = server/_core/index.ts; содержит предупреждение в console.log

### Архитектурные решения (ADR — не требуют реализации сейчас)

#### ADR-001: Streaming LLM responses
- **Статус**: ОТЛОЖЕНО (accepted: deferred)
- **Контекст**: ответы LLM приходят целиком (non-streaming). Chat UI показывает индикатор "Thinking..." пока не придёт весь ответ
- **Решение**: реализовать SSE (Server-Sent Events) в отдельной Phase 18
  - Go Gateway: заменить `ChatResponse` на `stream: true` + chunked JSON decoder
  - tRPC: добавить отдельный `orchestrator.chatStream` subscription (или REST SSE endpoint)
  - Chat.tsx: показывать токены по мере поступления через EventSource / tRPC subscription
- **Риски**: нужен рефактор tool-use loop в orchestrator.go для поддержки промежуточного стриминга
- **Приоритет**: средний — UX улучшение, не блокирует работу

#### ADR-002: Authentication / Authorization
- **Статус**: ПРИНЯТО как внутренний инструмент (accepted: internal tool)
- **Контекст**: все tRPC endpoints используют `publicProcedure` — нет аутентификации
- **Решение**: приемлемо для внутреннего инструмента, доступного только в закрытой сети
  - Если нужна защита: добавить `protectedProcedure` с JWT middleware в server/_core/context.ts
  - OAuth уже частично реализован (server/_core/oauth.ts, OAUTH_SERVER_URL env var)
  - При активации: заменить `publicProcedure` на `protectedProcedure` во всех роутерах
- **Риски**: текущая архитектура позволяет любому в сети вызывать shell_exec, file_write
- **Приоритет**: высокий — если сервис будет доступен публично