ai-engineering-team 发布：用多角色智能体流水线把需求做成可审查代码

An AI engineering team for your repository. Give it one feature request; it takes the feature from requirements to reviewed, tested code — the way a real engineering org would:

 PM → Developer → Checks → QA ──bugs──► Bug Fixer ──┐
                             ▲                      │
                             └──────────────────────┘
                     QA pass → Security → Code Review ──reject──► Developer → Checks → QA → …
                                                   approve → Release gate → APPROVED ✅ / REJECTED ❌

npm install -g ai-engineering-team
cd your-project
ai-team init
ai-team build "Add a health check endpoint"

Real output from a run on a small Node API (agent activity lines trimmed):

[1] Product Manager
  ✓ Requirements: 7 functional, 7 acceptance criteria
  ✓ 8 edge cases identified
  ! 4 open question(s)
[2] Clarifications
  ✓ Requirements finalized
[3] Developer
  ✓ 2 file(s) changed, 10 test(s) added
[4] Checks (lint / typecheck / tests / build)
  ✓ test (2.0s)
[5] QA Engineer
  ✓ 17 test scenarios
  ✗ 1 issue(s): BUG-001        ← flaky latency test, reproduced in 3 of 23 runs
  Round 1/5 → Bug Fixer
[6] Bug Fixer
  ✓ BUG-001 fixed
[7] Checks (lint / typecheck / tests / build)
  ✓ test (1.5s)
[8] QA Engineer
  ✓ 18 test scenarios
  ✓ QA PASSED
[9] Security Reviewer
  ✓ Security review APPROVED (1 low-severity note(s))
[10] Code Reviewer
  ✓ Code review APPROVED (4 low-severity note(s))
[11] Release Manager
  ✓ 4 lesson(s) added to engineering memory

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
             RELEASE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Requirements:    PASS — 7 acceptance criteria, 8 edge cases
Implementation:  PASS — 2 code revision(s)
Checks:          PASS — test
QA:              PASS — 18 scenarios
Security:        PASS — 1 low-severity note(s)
Code Review:     PASS — 4 low-severity note(s)
Iterations:      PASS — 1/5
Cost:            $1.81

STATUS: APPROVED ✅
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Why it works better than "Claude, build this feature"

• Separate roles with separate mindsets. The PM finds ambiguity; the developer builds exactly what was specified; QA assumes bugs exist and tries to prove it; the security reviewer thinks like an attacker; the code reviewer asks "would I approve this PR?"
• Agents don't trust each other. QA verifies the developer's claims by running things. Reviewers read the real diff, not summaries.
• Least privilege. PM, Security and Code Review are read-only. QA can run commands but not edit files. Only the Developer and Bug Fixer can change code.
• Typed contracts, not vibes. Every agent must return JSON matching a schema (requirements, bug reports, findings with severities). The orchestrator — not the model — decides what blocks.
• Deterministic gates. Lint/typecheck/tests/build are run by the orchestrator itself; an agent can't claim they pass. The release gate is plain code: every verdict must be a pass and must have been made against the current code version, so a stale approval can never ship a later change.
• Bounded loops. Fix/rework rounds are capped (maxIterations); if the team can't converge, the feature is rejected with reasons.
• Resumable. State is saved after every step. Crash, Ctrl+C, network error → ai-team resume.
• Engineering memory. Reviewers record durable project lessons into .ai-team/memory/lessons.md; every agent reads them on the next feature.

Requirements

• Node.js 20+
• Access to Claude: either you're logged in to Claude Code on this machine (run claude once), or ANTHROPIC_API_KEY is set. The Claude Code runtime is bundled via the Claude Agent SDK — no separate install needed.

Check your setup with ai-team doctor.

Commands

build options: --yes (don't ask the PM's open questions, accept proposed defaults), --max-iterations <n>, --model <id>, --git (branch + commit per step), --quiet (hide live agent activity).

When the PM finds real ambiguity it asks you (e.g. "Q-2: Can a coupon be combined with other discounts? [Enter = no]"). Your answers become part of the approved spec.

What gets created

.ai-team/
├── config.json          # workflow settings + checks
├── rules.md             # your project constitution — every agent must follow it
├── agents/              # role prompts — customize freely
│   ├── pm.md  developer.md  qa.md  bug-fixer.md  security.md  reviewer.md
├── memory/
│   └── lessons.md       # learned conventions, grows over time
└── runs/
    └── 20261008-1412-add-course-coupons/
        ├── request.md
        ├── 01-requirements.md
        ├── 02-implementation.md
        ├── 03-checks-v1.md
        ├── 04-qa-round1.md
        ├── 05-fix-round1.md
        ├── …
        ├── 11-release.md
        └── state.json   # resumable workflow state

Every run leaves a full audit trail: what was asked, what was specified, what was built, what broke, how it was fixed, and why it was approved.

Configuration — .ai-team/config.json

Agents also load your project's CLAUDE.md and Claude Code settings, so existing project instructions apply.

Git integration (--git): creates ai-team/<feature> and commits after each implementation/fix step, producing a reviewable history. It only activates on a clean working tree, so your uncommitted work is never mixed in. It never pushes.

Severity rules

• QA: any bug, or any failing check, sends the work to the Bug Fixer.
• Security & Code Review: any CRITICAL, HIGH or MEDIUM finding rejects — even if the agent itself wrote "APPROVED". LOW findings are recorded but don't block.

Programmatic use

import { Orchestrator, ClaudeCodeProvider, loadConfig } from 'ai-engineering-team';

const root = process.cwd();
const outcome = await new Orchestrator({ root, config: loadConfig(root), provider: new ClaudeCodeProvider() })
  .build('Add a health check endpoint');
console.log(outcome.state.status); // 'approved' | 'rejected' | 'failed'

Providers implement a single method (run(request) → { output, costUsd, turns }), so other models can be plugged in. MockProvider is included for testing your own workflows.

Roadmap

• v0.2 — GitHub: push branch, open PR with the release report, watch CI and fix failures.
• v0.3 — Architect agent for large features; parallel QA/security/review.
• v0.4 — ai-team dashboard: local web UI for runs, artifacts, costs.
• v0.5 — More providers (OpenAI, Gemini, local models).

License

MIT