onfire7777/five-agent-dev-team

最近记录的来源活动: 2026年4月27日 07:04

SkillsMP 收录数据更新: 2026年11月9日 14:22

已收集 skills: 30

GitHub 星标: 0
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/onfire7777/five-agent-dev-team.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# Five-Agent Autonomous Software Development Team Local-first controller for a five-agent autonomous software development team. It runs continuously with Docker Compose, uses Temporal for durable workflows, Postgres for state, GitHub as the delivery surface, and a modern local dashboard for operators. The canonical build contract is [docs/implementation-spec.md](docs/implementation-spec.md). Automation queue items, cloud research, and implementation PRs should map work back to that spec before using README or narrower docs as supporting context. ## Agents 1. **Product & Delivery Orchestrator**: intake, priority, scope, acceptance criteria, routing, final summary. 2. **R&D, Architecture & Innovation Agent**: research, feasibility, prototypes, architecture decisions, build contract. 3. **Frontend & UX Engineering Agent**: UI states, accessibility, client logic, frontend tests. 4. **Backend & Systems Engineering Agent**: APIs, services, data, integrations, backend tests. 5. **Quality, Security, Privacy & Release Agent**: verification, security/privacy review, performance, rollback, autonomous go/no-go. ## Quick Start ```powershell Copy-Item .env.example .env npm install npm run check docker compose up --build ``` Dashboard: `http://127.0.0.1:5173` Controller API: `http://127.0.0.1:4310` For host-based development against already-running local Postgres and Temporal: ```powershell npm run dev ``` ## Configuration Connect GitHub from the dashboard first, then connect target repos from the same `Project` panel. The dashboard supports GitHub's OAuth device flow when `GITHUB_OAUTH_CLIENT_ID` is configured; env tokens and mounted `gh` config still work for advanced setups. Dashboard-managed tokens are stored locally in `.agent-team/github-auth.json`, which is gitignored, and the same account is injected into the GitHub CLI, Octokit SDK, and official GitHub MCP server paths. The GitHub account connection fans out into a single utility bundle instead of separate setup chores: CLI commands, Octokit API coordination, MCP dynamic toolsets, repo/code context, issues, pull requests, Actions/checks, releases, Projects, security context, and optional Copilot-agent toolsets. The dashboard shows which utilities are ready, available, or blocked by token scope, while actual tool use remains lazy and stage-gated. Create `agent-team.config.yaml` from `agent-team.config.example.yaml` only when you want a file-backed default repo. Each target repo must expose install, lint, typecheck, test, build, security, and release commands. The controller will not guess project-specific commands. Each connected repo is treated as an isolated project. Work items and permanent memory carry a project/repo scope so one repository's decisions, gotchas, and automation state are not mixed into another repository. Dashboard/API connections are stored in Postgres and mirrored to `.agent-team/projects/<project-id>.yaml` so controller and worker processes resolve the same per-project config instead of relying on a single mutable config file. Autonomous activities fail closed when `agent-team.config.yaml` is missing or invalid. Set `AGENT_TEAM_ALLOW_DEFAULT_CONFIG=true` only for local smoke tests where running against the controller checkout is intentional. ## Release Safety Invariant A completed autonomous loop must leave: - local target repo clean - local branch and remote branch synchronized - no duplicate automation branch for the same work item - PR checks passing in GitHub Actions - rollback plan present - secret scan passing - release created through the configured GitHub path - `CLOSED` loop summary promoted into repo-scoped latest-state memory - durable workflow claim released only after all parallel agent branches have settled Emergency stop is available in the dashboard and through `POST /api/emergency-stop`. ## ChatGPT Pro-Aware Operation The default `chatgpt_pro_assisted` mode is designed around ChatGPT Pro/Codex-style work: continuous event monitoring, fully autonomous workflow decisions, and controlled parallelism without wasteful always-on model loops. API-backed live mode remains available with `AGENT_EXECUTION_MODE=api_live` and separate OpenAI API billing. Default parallelism: - one complete work-item loop at a time inside each repo, so the next repo-local loop starts only after the current loop reaches `CLOSED` or `BLOCKED` - multiple five-agent teams can run across disjoint connected repos, bounded by `MAX_CONCURRENT_WORKFLOWS` - up to 5 active agent runs per team loop - 1 serialized repo-write/release lane per team policy - frontend and backend build stages run in parallel after contract lock Set `completeLoopBeforeNextWorkItem: false` only when cross-work-item parallelism is intentional and the work is disjoint. Internal stage parallelism remains enabled by default. ## Shared Memory Agents are context-aware through shared artifacts plus permanent memory. R&D decisions, contracts, release decisions, recurring risks, failures, and stable preferences are stored as typed memory records and injected into future agent runs when relevant. Every successful loop closure also upserts one repo-scoped `latest-loop` memory record. The next loop reads that memory before intake, so it starts from the latest completed codebase state rather than stale or cross-project context. The lean AutoMaker-inspired context pack lives in the target repo at `.agent-team/context/`. Put only durable rules, architecture notes, and gotchas there; the controller loads them into every relevant agent run without adding dashboard complexity. ## Lazy MCP, Skills, And Plugins Agents can be configured with MCP servers and capability packs, but they are not loaded into every run. The runner selects tools from the current work item, stage, agent role, risk-level keywords, and acceptance criteria, then closes MCP sessions after that run. This gives the team proactive access to browser automation, Chrome diagnostics, GitHub context, security scanners, database tools, Electron diagnostics, and specialized coding knowledge without bloating normal prompts. See `agent-team.config.example.yaml` and `docs/lazy-capability-model.md`. ## Model Policy Live agent runs default to the configured best coding model policy: `gpt-5.5` for coding, research, and review roles, with `gpt-5.4` only as an explicit fallback. Set `AGENT_MODEL` only when you intentionally want to override the per-role model policy. ## API - `GET /health` - `GET /api/status` - `GET /api/github/status` - `GET /api/projects` - `POST /api/projects` - `POST /api/projects/:id/activate` - `DELETE /api/projects/:id` - `GET /api/work-items` - `GET /api/memories` - `GET /api/events` - `GET /api/events/stream` - `GET /api/github/account` - `POST /api/github/device/start` - `POST /api/github/device/poll` - `POST /api/github/disconnect` - `GET /api/projects/:projectId/team-bus` - `GET /api/projects/:projectId/loop-runs` - `GET /api/projects/:projectId/direction` - `POST /api/projects/:projectId/direction` - `GET /api/projects/:projectId/opportunities` - `POST /api/projects/:projectId/opportunities/scan` - `POST /api/opportunities/:id/promote` - `GET /api/work-items/:id/proposal` - `POST /api/work-items/:id/proposal/accept` - `POST /api/work-items/:id/proposal/revise` - `POST /api/work-items/:id/proposal/reject` - `POST /api/work-items` - `POST /api/emergency-stop` - `POST /api/emergency-resume` ## Development Notes If `AGENT_LIVE_MODE=false`, agents use deterministic local templates so the workflow can be tested without an API key. Set `AGENT_LIVE_MODE=true` and `OPENAI_API_KEY` to use live OpenAI Agents SDK runs. ## Codex Automation Control This repo includes a guarded bulk-control utility for the Codex automations that build this project. It updates only the nine approved `five-agent-dev-team` automation folders under `%USERPROFILE%\.codex\automations`, preserves each automation's prompt/schedule/environment, and dry-runs unless `--apply` is provided. Use `--root=<automation-root>` when checking a different automation folder. Normal automation execution is cloud-first. GitHub Actions runs the recurring control plane: Codex Cloud Research creates scored queue issues, Codex Cloud Build Pipeline executes specialist lanes from those issues, Codex PR Review reviews pull requests, CI/CodeQL verify changes, Captain integrates ready PRs, Meta Health reconciles health, and Janitor clears safe stale cloud state. The desktop Codex automations remain local fallback/supervision workers unless a queue item explicitly uses `execution:local-fallback`. ```powershell Get-Content "$env:USERPROFILE\.codex\state\five-agent-dev-team-control.json" npm run automation:status -- --status=PAUSED npm run automation:pause npm run automation:verify-paused npm run automation:resume npm run automation:verify-active ``` See `docs/architecture.md` and `docs/target-repo-setup.md` for operational details. Security note: `npm audit` currently reports a Temporal transitive `uuid` advisory that cannot be fixed safely with non-breaking npm resolution. CodeQL is provided as an opt-in template because GitHub code scanning must be enabled first for this private repo. See `docs/security.md`.

正在加载仓库 README…

这个仓库中的 skills

软件开发工程师软件质量保证分析师与测试员项目管理专家信息安全分析师

5 个职业分类 · 已分类 100%

软件开发工程师

Update shared Zod schemas while preserving cross-lane API compatibility.

原文语言：英语

2026-04-27

已展示 30 / 30 个已收集 Skill。