Skip to content

README.md

Latest commit

 

History

History
165 lines (130 loc) · 7.25 KB

README.md

File metadata and controls

165 lines (130 loc) · 7.25 KB
Raw

AiCodingAgentTeam

🌐 Languages / 语言

English · 简体中文 · 日本語 · 한국어 · Français · Deutsch · Русский · Español · Italiano

A Golang-based AI coding orchestration platform. It owns no LLM itself; instead it dispatches external AI coding CLIs (Codex & OpenCode as real drivers; Claude-Code & DeepSeek-DSH as stubs), simulates a 9-role software-development team that collaborates over the A2A protocol, enforces a deterministic quality gate with governance auditing, ships in containers, and exposes four protocols externally — gRPC, MCP, ACP, and A2A — alongside a TypeScript TUI client.

CI Go License Docs

Core Features

  • Orchestration vs execution separation — The Coordinator only orchestrates and never writes code; host CLIs only execute and never decide.
  • Container-as-role — 9 team roles (PM, Architect, Frontend, Backend, QA, Security, DevOps, …), each in its own container, independently scalable and replaceable.
  • A2A protocol collaboration — Structured inter-agent communication, supporting InProc (development) and Redis Pub/Sub (containerized).
  • Deterministic quality gate — Hard-executed by golangci-lint + go vet + go test; never relies on model self-evaluation.
  • RAG knowledge base + project memory — BM25 retrieval injected into context; the memory store accumulates failure lessons and historical solutions.
  • Four external protocols — gRPC (TUI communication), MCP (external tool integration), ACP (standard Agent client), A2A (cross-instance mesh).
  • Local-first — Code never leaves the container; no API keys are held; all auth is delegated to the underlying host CLI.

Architecture Overview

User → TypeScript TUI (Ink) → gRPC → Coordinator
                                       ├── Router (intent routing)
                                       ├── Planner (DAG plan)
                                       ├── Scheduler (role dispatch)
                                       │    ├── A2A Bus → Reviewer Agents (parallel)
                                       │    └── Single-Writer Lock → Writer Agents (serial)
                                       ├── Knowledge (BM25 RAG retrieval)
                                       ├── Memory (facts/pitfalls/lessons)
                                       └── Quality Gate (deterministic checks)
                                            ↓
Host CLIs: Codex (real) | OpenCode (real) | Claude (stub) | DSH (stub)

See docs/04-系统架构图.md for the full architecture diagrams.

Quick Start

Single-binary mode

# Build
go build -o bin/aicodingagentteam ./cmd/aicodingagentteam

# Initialize a project
./bin/aicodingagentteam init

# Run a full pipeline
./bin/aicodingagentteam run "Build a REST API" --backend codex

# Quick edit
./bin/aicodingagentteam quick "Fix the login button style"

# Quality verification
./bin/aicodingagentteam verify

# RAG knowledge base demo
./bin/aicodingagentteam knowledge demo

# Start the coordinator service (gRPC + A2A HTTP)
./bin/aicodingagentteam serve

TUI client

cd tui && npm install && npm run build

# Start the coordinator
cd .. && go run ./cmd/aicodingagentteam serve

# Launch the TUI (in another terminal)
cd tui && node dist/cli.js
# Demo mode (no coordinator required)
node dist/cli.js --demo

Containerized deployment

docker compose -f deploy/compose/docker-compose.yml up -d

Command Reference

Command Description
init Initialize project configuration
run "requirement" Run the full pipeline
quick "description" Lightweight quick edit
verify Execute quality verification
govern [--ci] [path] Governance scan
report Emit compliance report
serve Start coordinator (gRPC + A2A)
knowledge index [dir] Index a directory into the knowledge base
knowledge search "query" Retrieve top-5 knowledge chunks
knowledge demo End-to-end RAG + memory demo
version Print version

Project Structure

agent_team/
├── cmd/aicodingagentteam/     # Go binary entry
├── internal/
│   ├── a2a/                   # A2A protocol (InProc + Redis)
│   ├── acp/                   # Agent Client Protocol
│   ├── agent/                 # 9-role registry
│   ├── audit/                 # Audit log
│   ├── config/                # Configuration loader
│   ├── coordinator/           # Orchestration core (5-layer flow)
│   ├── governance/            # Governance rules (113+ rules)
│   ├── host/                  # Host drivers (codex/opencode/claude/dsh)
│   ├── knowledge/             # RAG retrieval engine (BM25)
│   ├── mcp/                   # Model Context Protocol
│   ├── memory/                # Project memory (facts/pitfalls/lessons)
│   ├── planner/               # DAG plan builder
│   ├── qualitygate/           # Quality-gate engine
│   ├── router/                # Intent router
│   ├── scheduler/             # Role scheduler + single-writer lock
│   └── types/                 # Shared types
├── pkg/
│   ├── api/                   # gRPC proto + server
│   ├── contracts/             # Front/back-end contracts
│   └── runtime/               # Host driver trait
├── tui/                       # TypeScript TUI client
├── deploy/                    # Containerized deployment
├── proto/                     # gRPC proto definitions
└── docs/                      # Documentation (requirements/architecture/plans/ADR/spec/plan)

Quality Baseline

Dimension Threshold Actual
Go test coverage ≥ 80% (core ≥ 90%) router 100% · scheduler 96% · governance 96% · codex 94.9%
Lint 0 new warnings golangci-lint + eslint
Build ≤ 5 minutes CI full flow ~3 minutes
Security 0 high-severity CVEs govulncheck + npm audit

Development

make all      # lint + vet + test + build
make test     # Full test suite
make lint     # golangci-lint
make run      # build + serve

Documentation

License

MIT