APAW/AGENTS.md

Kilo Code Agents Reference

This file configures AI agent behavior for the APAW project - a self-improving code pipeline with Gitea logging.

Pipeline Workflow

The main workflow is /pipeline - use it to process issues through all agents automatically.

User: /pipeline 42
Agent: Runs full pipeline for issue #42 with Gitea logging

Commands (Slash Commands)

Command	Description	Usage
/pipeline <issue>	Run full agent pipeline for issue	/pipeline 42
/status <issue>	Check pipeline status for issue	/status 42
/evolve	Run evolution cycle with fitness scoring	/evolve --issue 42
/evaluate <issue>	Generate performance report	/evaluate 42
/plan	Creates detailed task plans	/plan feature X
/ask	Answers codebase questions	/ask how does auth work
/debug	Analyzes and fixes bugs	/debug error in login
/code	Quick code generation	/code add validation
/research [topic]	Run research and self-improvement	/research multi-agent
/evolution log	Log agent model change	/evolution log planner "reason"
/evolution report	Generate evolution report	/evolution report
/web-test <url>	Visual regression testing in Docker	/web-test https://bbox.wtf
/e2e-test <url>	E2E browser automation tests	/e2e-test https://my-app.com

Pipeline Agents (Subagents)

These agents are invoked automatically by /pipeline or manually via @mention:

Core Development

Agent	Role	Model	Variant	Can Call
@requirement-refiner	Converts ideas to User Stories	glm-5.1	thinking	history-miner, system-analyst
@history-miner	Finds duplicates in git	nemotron-3-super	—	(read-only)
@system-analyst	Designs specifications	glm-5.1	thinking	sdet-engineer, orchestrator
@sdet-engineer	Writes tests (TDD)	qwen3-coder:480b	thinking	lead-developer, orchestrator
@lead-developer	Implements code	qwen3-coder:480b	thinking	code-skeptic, orchestrator
@frontend-developer	UI implementation	qwen3-coder:480b	—	code-skeptic, orchestrator
@backend-developer	Node.js/Express/APIs	qwen3-coder:480b	—	code-skeptic, orchestrator
@go-developer	Go backend services	qwen3-coder:480b	—	code-skeptic, orchestrator
@flutter-developer	Flutter mobile apps	qwen3-coder:480b	—	code-skeptic, orchestrator

Quality Assurance

Agent	Role	Model	Variant	Can Call
@code-skeptic	Adversarial review	minimax-m2.5	—	the-fixer, performance-engineer, orchestrator
@the-fixer	Fixes issues	minimax-m2.5	—	code-skeptic, orchestrator
@performance-engineer	Performance review	nemotron-3-super	—	the-fixer, security-auditor, orchestrator
@security-auditor	Security audit	nemotron-3-super	—	the-fixer, release-manager, orchestrator
@visual-tester	Visual regression + bbox extraction + console/network errors	qwen3-coder:480b	—	the-fixer, orchestrator
@browser-automation	E2E testing	qwen3-coder:480b	—	orchestrator

DevOps & Infrastructure

Agent	Role	Model	Variant	Can Call
@devops-engineer	Docker/K8s/CI-CD	nemotron-3-super	—	code-skeptic, security-auditor, orchestrator
@release-manager	Git operations, releases	glm-5.1	—	evaluator

Meta & Process

Agent	Role	Model	Variant	Can Call
@evaluator	Scores effectiveness	glm-5.1	thinking	prompt-optimizer, product-owner, orchestrator
@pipeline-judge	Objective fitness scoring	glm-5.1	—	prompt-optimizer
@prompt-optimizer	Improves prompts	glm-5.1	instant	(edits files)
@product-owner	Manages issues/tracking	glm-5.1	—	(read-only)

Analysis & Design

Agent	Role	Model	Variant	Can Call
@capability-analyst	Analyzes task coverage	glm-5.1	—	agent-architect, orchestrator
@agent-architect	Creates new agents	glm-5.1	thinking	capability-analyst, requirement-refiner, system-analyst
@workflow-architect	Creates workflows	glm-5.1	thinking	(edits files)
@markdown-validator	Validates Markdown	nemotron-3-nano:30b	—	orchestrator

Cognitive Enhancement

Agent	Role	Model	Variant	Can Call
@planner	Task decomposition	nemotron-3-super	—	(read-only)
@reflector	Self-reflection	nemotron-3-super	—	(read-only)
@memory-manager	Memory systems	nemotron-3-super	—	(read-only)

Workflow State Machine

[new] 
  ↓ @requirement-refiner
[planned] 
  ↓ @capability-analyst → (gaps?) → @agent-architect → create new agents
  ↓ @history-miner
[researching] 
  ↓ @system-analyst
[designed] 
  ↓ @sdet-engineer (writes failing tests)
[testing] 
  ↓ @lead-developer (makes tests pass)
[implementing] 
  ↓ @code-skeptic (review)
[reviewing] ──[fail]──→ [fixing] ──→ [reviewing]
  ↓ @review-watcher → (auto-validate) → create fix tasks
  ↓ [pass]
[perf-check] 
  ↓ @performance-engineer
[security-check] 
  ↓ @security-auditor
[releasing] 
  ↓ @release-manager
[evaluated] 
  ↓ @evaluator (subjective score 1-10)
  ├── [score ≥ 7] → [@pipeline-judge] → fitness scoring
  └── [score < 7] → @prompt-optimizer → [@evaluated]
        ↓
    [@pipeline-judge] ← runs tests, measures tokens/time
        ↓
    fitness score
        ↓
┌──────────────────────────────────────┐
│ fitness >= 0.85                      │──→ [completed]
│ fitness 0.70-0.84                    │──→ @prompt-optimizer → [evolving]
│ fitness < 0.70                      │──→ @prompt-optimizer (major) → [evolving]
│ fitness < 0.50                      │──→ @agent-architect → redesign
└──────────────────────────────────────┘
        ↓
[evolving] → re-run workflow → [@pipeline-judge]
        ↓
    compare fitness_before vs fitness_after
        ↓
    [improved?] → commit prompts → [completed]
              └─ [not improved?] → revert → try different strategy

Capability Analysis Flow

When starting a complex task:

[User Request]
      ↓
[@capability-analyst] ← Analyzes requirements vs existing capabilities
      ↓
[Gap Analysis] ← Identifies missing agents, workflows, skills
      ↓
[Recommendations] → Create new or enhance existing?
      ↓
[Decision]
  ├── [Create New] → [@agent-architect] → Create component → Review
  └── [Enhance] → [@lead-developer] → Modify existing
      ↓
[Integration] ← Verify new component works with system
      ↓
[Complete] ← Task can now be handled

Gitea Integration

Status Labels

Pipeline uses Gitea labels to track progress:

• status: new → status: planned → status: researching → ...
• Agents add/remove labels automatically

Performance Logging

Each agent logs to Gitea issue comments:

## ✅ lead-developer completed

**Score**: 8/10
**Duration**: 1.2h
**Files**: src/auth.ts, src/user.ts

### Notes
- Clean implementation
- Follows existing patterns
- Tests passing

Efficiency Tracking

Scores saved to .kilo/logs/efficiency_score.json:

{
  "version": "1.0",
  "history": [
    {
      "issue": 42,
      "date": "2024-01-02T10:00:00Z",
      "agents": {
        "lead-developer": 8,
        "code-skeptic": 7,
        "the-fixer": 9
      },
      "iterations": 2,
      "duration_hours": 1.5
    }
  ]
}

Fitness Tracking

Fitness scores saved to .kilo/logs/fitness-history.jsonl:

{"ts":"2026-04-06T00:00:00Z","issue":42,"workflow":"feature","fitness":0.82,"tokens":38400,"time_ms":245000,"tests_passed":45,"tests_total":47}
{"ts":"2026-04-06T01:30:00Z","issue":43,"workflow":"bugfix","fitness":0.91,"tokens":12000,"time_ms":85000,"tests_passed":47,"tests_total":47}

Manual Agent Invocation

// Use Task tool to invoke subagent
Task tool with:
  subagent_type: "lead-developer"
  prompt: "Implement authentication for issue #42"

Or via @mention:

@lead-developer implement authentication flow

Environment Variables

Required for Gitea integration:

GITEA_API_URL=https://git.softuniq.eu/api/v1
GITEA_TOKEN=your-token-here

Self-Improvement Cycle

1. Pipeline runs for each issue
2. Evaluator scores each agent (1-10) - subjective
3. Pipeline Judge measures fitness objectively (0.0-1.0)
4. Low fitness (<0.70) triggers prompt-optimizer
5. Prompt optimizer analyzes failures and improves prompts
6. Re-run workflow with improved prompts
7. Compare fitness before/after - commit if improved
8. Log results to .kilo/logs/fitness-history.jsonl

Evaluator vs Pipeline Judge

Aspect	Evaluator	Pipeline Judge
Type	Subjective	Objective
Score	1-10 (opinion)	0.0-1.0 (metrics)
Metrics	Observations	Tests, tokens, time
Trigger	After workflow	After evaluator
Action	Logs to Gitea	Triggers optimization

Fitness Score Components

fitness = (test_pass_rate × 0.50) + (quality_gates_rate × 0.25) + (efficiency_score × 0.25)

where:
  test_pass_rate = passed_tests / total_tests
  quality_gates_rate = passed_gates / total_gates (build, lint, types, tests, coverage)
  efficiency_score = 1.0 - clamp(normalized_cost, 0, 1)

Architecture Files

File	Purpose
AGENTS.md	This file - main config
.kilo/agents/*.md	Agent definitions with prompts
.kilo/commands/*.md	Workflow commands
.kilo/rules/*.md	Custom rules loaded globally
.kilo/skills/	Skill modules
src/kilocode/	TypeScript API for programmatic use

Using the TypeScript API

import { 
  PipelineRunner, 
  GiteaClient, 
  decideRouting 
} from './src/kilocode/index.js'

const runner = await createPipelineRunner({
  giteaToken: process.env.GITEA_TOKEN
})

await runner.run({ issueNumber: 42 })

Agent Evolution Dashboard

Track agent model changes, performance, and recommendations in real-time.

Access

# Sync agent data
bun run sync:evolution

# Open dashboard
bun run evolution:dashboard
bun run evolution:open
# or visit http://localhost:3001

Dashboard Tabs

Tab	Description
Overview	Stats, recent changes, pending recommendations
All Agents	Filterable agent cards with history
Timeline	Full evolution history
Recommendations	Priority-based model suggestions
Model Matrix	Agent × Model mapping with fit scores

Data Sources

Source	What it tracks
.kilo/agents/*.md	Model, description, capabilities
.kilo/kilo.jsonc	Model assignments
.kilo/capability-index.yaml	Capability routing
Git History	Model and prompt changes
Gitea Comments	Performance scores

Evolution Data Structure

{
  "agents": {
    "lead-developer": {
      "current": { "model": "qwen3-coder:480b", "fit_score": 92 },
      "history": [{ "type": "model_change", "from": "deepseek", "to": "qwen3" }],
      "performance_log": [{ "issue": 42, "score": 8, "success": true }]
    }
  }
}

Recommendations Priority

Priority	When	Example
Critical	Fit score < 70	Immediate model change required
High	Model unavailable	Switch to fallback
Medium	Better model available	Consider upgrade
Low	Optimization possible	Optional improvement

Code Style

• Use TypeScript for new files
• Follow existing patterns
• Write tests before code (TDD)
• Keep functions under 50 lines
• Use early returns
• No comments unless explicitly requested