fischer-agentkit/AGENTS.md

# Fischer AgentKit — Project Context

## Rules

- Python >= 3.11, type hints required, `pydantic>=2.0` for all data models
- Ruff for lint + format: `ruff check src/ && ruff format src/` (target py311, line-length 100)
- Tests: `pytest` (asyncio_mode=auto), markers: `integration`, `redis`, `postgres`
- Never use `any` type — use proper Pydantic models or `Unknown`
- API key comparison must use `hmac.compare_digest` (constant-time)
- Expert names validated with `_EXPERT_NAME_RE = re.compile(r"^[a-zA-Z0-9_-]{1,64}$")`
- HandoffTransport queues bounded (`maxsize=1024`), close uses sentinel pattern
- Frontend: Vue 3 + TypeScript + Ant Design Vue, Pinia stores, no `require()` calls
- **Async generator safety**: Never use early `return` before the first `yield` in `async def` — use `return; yield` pattern instead (see `.trae/rules/project_rules.md`)

## Tech Stack

- **Backend**: Python 3.11+, FastAPI, Uvicorn, Pydantic v2, SQLAlchemy 2 (async)
- **Frontend**: Vue 3, TypeScript, Vite 5, Ant Design Vue 4, Pinia, Vue Router 4
- **Desktop**: Tauri 2.x (Rust shell + Python sidecar)
- **Infra**: Redis (bus/cache/state), PostgreSQL + pgvector (episodic memory)
- **CLI**: Typer + Rich
- **Exact versions**: see `pyproject.toml` (Python), `package.json` (Node)

## Commands

```bash
# Backend
pip install -e ".[dev]"                # Install with dev deps
agentkit gui --port 8002               # Web GUI (frontend + API)
agentkit serve --port 8001             # API-only server
agentkit chat                          # CLI interactive chat
agentkit init                          # Generate agentkit.yaml
agentkit version / doctor / usage      # Utility commands
agentkit task submit/status/list/cancel # Task management
agentkit skill list/load/info          # Skill management
agentkit pair --name X                 # Generate API key for external system
pytest                                 # Run all tests
pytest -m "not integration"            # Unit tests only
ruff check src/ && ruff format src/    # Lint + format

# Frontend
cd src/agentkit/server/frontend
npm install                            # Install deps
npm run dev                            # Vite dev server (proxy /api -> :8000)
npm run build:frontend                 # Production build -> ../static
npm run typecheck                      # TypeScript check

# Desktop
cd src/agentkit/server/frontend
npm run tauri dev                      # Tauri dev mode
npm run tauri build                    # Tauri production build

# Docker
docker-compose up -d                   # AgentKit + Redis + PostgreSQL
```

## Architecture

### Request Flow

```
User Input
  ├─ @board prefix -> BoardRouter (experts/board_router.py) -> BoardOrchestrator (multi-round discussion)
  ├─ @team prefix  -> ExpertTeamRouter (experts/router.py) -> TeamOrchestrator (pipeline collaboration)
  └─ otherwise     -> RequestPreprocessor (chat/request_preprocessor.py)
       Layer 0: @skill:xxx prefix -> explicit skill selection (SKILL_REACT or skill's configured mode)
       Layer 1: Trivial-input regex (~0ms, 0 tokens) -> DIRECT_CHAT
                (greetings, identity, factual Q&A, math, translation; guarded by _TOOL_CONTEXT_RE)
       Default: -> REACT (LLM decides tool usage autonomously in the agent loop)
  -> ExecutionMode: DIRECT_CHAT / REACT / SKILL_REACT / REWOO / REFLEXION / PLAN_EXEC / TEAM_COLLAB
     (chat handler currently supports DIRECT_CHAT, REACT, SKILL_REACT; others raise "not yet supported")
```

**Note**: The old 3-layer `CostAwareRouter` (with `RegexRules` / `HeuristicClassifier` / `SemanticRouter` / `Vickrey Auction`) has been replaced by `RequestPreprocessor`. The `IntentRouter` (`router/intent.py`) exists but is not wired into the chat flow. `AuctionHouse` with Vickrey auction lives in `marketplace/auction.py` (marketplace subsystem, not routing).

### Agent Hierarchy

```
BaseAgent (core/base.py) — abstract, execute() is final
  +-- ConfigDrivenAgent (core/config_driven.py) — YAML-driven, 3 task modes
  +-- ReActEngine (core/react.py) — Think->Act->Observe
  +-- ReflexionAgent (core/reflexion.py) — reflection-driven
  +-- ReWOOAgent (core/rewoo.py) — plan-without-observation
  +-- StandaloneAgent (core/standalone.py) — standalone runner
```

### Expert Team Mode (Pipeline)

```
ExpertConfig (extends AgentConfig) -> Expert (wraps ConfigDrivenAgent via AgentPool)
ExpertTeam: manages experts, shared workspace, team status (FORMING→PLANNING→EXECUTING→SYNTHESIZING→COMPLETED)
TeamOrchestrator: pipeline execution — Lead decomposes task into PlanPhase with depends_on, topological sort, parallel layers
PlanPhase: id, name, assigned_expert, task_description, depends_on, status (PENDING/RUNNING/COMPLETED/FAILED)
TeamPlan: phases with dependencies, topological_sort() returns execution layers (Kahn's algorithm)
ExpertTeamRouter: @team prefix routing, @team:dev_team template expansion, name validation, MAX_EXPERTS=10
HandoffTransport: InProcess (asyncio.Queue) + Redis Pub/Sub — used for event broadcasting only
```

**Pipeline Flow**:
1. `@team` prefix triggers team mode (or `@team:dev_team` for template, `@team:expert1,expert2` for explicit)
2. `ExpertTeam.create_team()` sets status to PLANNING
3. Lead Expert decomposes task into phases via LLM (fallback to single phase on failure)
4. `topological_sort()` arranges phases into layers (same-layer parallel, inter-layer serial)
5. Each phase creates an isolated `ConfigDrivenAgent` via `AgentPool.create_agent` (context isolation, KTD3)
6. Phase outputs passed via `SharedWorkspace` (`{plan_id}/phase/{phase_id}/output`)
7. Lead synthesizes results (BEST strategy)
8. On all-phases-fail: fallback to single agent mode

**Event Sequence**: `team_formed` → `plan_update` → `phase_started` → `expert_step` → `expert_result` → `phase_completed` → `team_synthesis` → `team_dissolved`

**Team Templates**: `configs/experts/dev_team.yaml` stores member list in `bound_skills` field (tech_lead, frontend_engineer, backend_engineer, qa_engineer, code_reviewer)

Lifecycle: FORMING -> PLANNING -> EXECUTING -> SYNTHESIZING -> COMPLETED -> DISSOLVED
On failure: fallback to single-agent mode (lead or first active expert).

### Module Map

| Layer | Modules | Purpose |
|-------|---------|---------|
| API | `server/`, `cli/` | FastAPI routes + Typer CLI |
| Auth | `server/auth/` | JWT + RBAC + terminal security (6-layer whitelist) |
| Service | `core/`, `chat/`, `skills/`, `experts/` | Agent engine, routing, skills, expert teams |
| Data | `memory/`, `session/`, `bus/` | Persistence, sessions, messaging |
| Utility | `llm/`, `tools/`, `evolution/`, `quality/`, `mcp/` | LLM gateway, tools, self-evolution, quality, MCP |
| Client | `client/` | ConfigSync, RemoteLLMProvider integration |

### Key Subsystems

- **LLM Gateway** (`llm/`): 6 providers (OpenAI/Anthropic/Gemini/Doubao/Wenxin/Yuanbao), fallback, semantic cache, usage tracking, RemoteLLMProvider (client→server proxy with 401 refresh retry)
- **Memory** (`memory/`): 4-layer (SOUL/USER/MEMORY/DAILY), WorkingMemory (Redis), EpisodicMemory (PG+pgvector), SemanticMemory (HTTP RAG)
- **Evolution** (`evolution/`): Reflector, PromptOptimizer (genetic), PitfallDetector, ABTester
- **Tools** (`tools/`): 21 built-in + MCP extension, composition (SequentialChain/ParallelFanOut/DynamicSelector)
- **Pipeline** (`orchestrator/`): PipelineEngine, SagaOrchestrator, DynamicPipeline, HandoffManager
- **Bus** (`bus/`): MemoryBus (in-process), RedisBus (distributed)
- **Auth** (`server/auth/`): JWT (access 15min + refresh 7d, HS256), API Key (constant-time compare), 3-level RBAC (member/operator/admin + permission bits), 6-layer terminal security (blocklist→shell-ops→builtin→global→user→session→danger), bcrypt password hashing (rounds=12)

### Server Routes (22 modules)

| Prefix | Module | Purpose |
|--------|--------|---------|
| `/api/v1/agents` | agents.py | Agent CRUD |
| `/api/v1/tasks` | tasks.py | Task submit/query/cancel |
| `/api/v1/skills` | skills.py | Skill register/list |
| `/api/v1/chat` | chat.py | Chat REST + WebSocket |
| `/api/v1/ws` | ws.py | WebSocket channel |
| `/api/v1/llm` | llm.py | LLM usage |
| `/api/v1/llm/chat` | llm_gateway.py | LLM gateway proxy (JWT auth, SSE streaming) |
| `/api/v1/health` | health.py | Health check |
| `/api/v1/metrics` | metrics.py | Metrics |
| `/api/v1/evolution` | evolution.py + evolution_dashboard.py | Self-evolution API |
| `/api/v1/memory` | memory.py | Memory management |
| `/api/v1/portal` | portal.py | Portal |
| `/api/v1/kb` | kb_management.py | Knowledge base |
| `/api/v1/skill-mgmt` | skill_management.py | Skill management |
| `/api/v1/workflows` | workflows.py | Workflows |
| `/api/v1/terminal` | terminal.py | Local terminal (client sidecar PTY) |
| `/api/v1/terminal/server` | terminal_server.py | Server terminal (server PTY + admin approval) |
| `/api/v1/terminal` | terminal_whitelist.py | Whitelist/blocklist/audit-log management |
| `/api/v1/settings` | settings.py | Settings |
| `/api/v1/auth` | auth.py | Login/refresh/logout/me |
| `/api/v1/system` | system.py | System resources (SYSTEM_CONFIG permission) |
| `/api/v1/config` | config_sync.py | Config version + sync (polling) |

### WebSocket Chat Protocol

Client -> Server: `message`, `reply`, `confirmation_reply`, `cancel`, `ping`
Server -> Client: `connected`, `token`, `thinking`, `step`, `final_answer`, `skill_match`, `confirmation_request`, `confirmation_result`, `ask_human`, `error`, `pong`
Expert Team events: `team_formed`, `expert_step`, `expert_result`, `plan_update`, `phase_started`, `phase_completed`, `phase_failed`, `team_synthesis`, `team_dissolved`

### Frontend Pages

- `/agent/chat` — Chat with Expert Team view
- `/agent/code` — Code/workflow
- `/agent/monitor` — Evolution dashboard
- `/computer-use` — Desktop control
- `/login` — Login page (JWT auth)
- Terminal panel — Local + server terminal with whitelist manager

### Configuration Priority

CLI args > `agentkit.yaml` > env vars (`${VAR:-default}`) > `.env` > hardcoded defaults

Config search: `--config` path > `./agentkit.yaml` > `~/.agentkit/agentkit.yaml`

## Conventions

- Skill configs: `configs/skills/*.yaml` (16 presets, unified as `SkillConfig`)
- Skill categories: `agent_template` (execution engines: react/direct/rewoo/reflexion/plan_exec/goal_driven) vs `business_skill` (domain skills). Classified via `_ENGINE_TEMPLATE_NAMES` in `server/routes/skill_management.py`. Frontend groups by `category` field — `SkillsView` two-column layout, `SkillCard`/`SkillsTab` show type tags (引擎/技能) and category-based icons
- LLM configs: `agentkit.yaml` llm section (unified with server config)
- Pipeline configs: `configs/pipelines/*.yaml`
- Expert templates: `configs/experts/*.yaml` (5 programming experts + dev_team team template), registered via `ExpertTemplateRegistry`
- Team templates: `bound_skills` field stores member list (e.g., `dev_team.yaml` lists tech_lead, frontend_engineer, backend_engineer, qa_engineer, code_reviewer)
- All Pydantic models use `model_config = ConfigDict(...)` not `class Config`
- Test files: `tests/unit/` and `tests/integration/`
- Frontend stores: Pinia, one per domain (chat, team, settings)
- Frontend components: `src/agentkit/server/frontend/src/components/`

## Boundaries

- Never modify `pyproject.toml` version without explicit request
- Never push to main directly — use feature branches
- Integration tests require Docker (Redis + PostgreSQL)
- Desktop builds require Rust toolchain + PyInstaller