loop-engineering
Run the repository's end-to-end delivery loop from idea or legacy-project audit to plan, implementation, verification, documentation, and release preparation. Use when the user wants structured interactive guidance followed by persistent execution with minimal unnecessary stops.
- License 许可
- MIT
- 最近更新
- 2026-06-26
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- loop-engineering
- description
- Run the repository's end-to-end delivery loop from idea or legacy-project audit to plan, implementation, verification, documentation, and release preparation. Use when the user wants structured interactive guidance followed by persistent execution with minimal unnecessary stops.
03 SKILL.md 正文
loop-engineering
本 skill 是專案主工作流。
讀取順序
AGENTS.mddocs/session_loop_contract.mddocs/interactive_project_flow.mddocs/project_architecture_best_practices.mddocs/project_lifecycle_automation.mddocs/subagent_dispatch.md
兩種模式
A. 新專案
- 必要時先用
interview-me - 用
idea-refine收斂想法 - 用
spec-driven-development建 spec - 技術棧定了就先用
project-config-generation - 用
planning-and-task-breakdown切 task - 再進實作 loop
B. 舊專案改善
- 先看入口檔、設定、測試、部署
- 先提改善與重構切片
- 使用者確認後才做大改
預設流程
這是預設路徑,不是固定步數。
- 先判斷現在是新專案還是舊專案
- 只問 1-3 組決策題
- 先跑
node scripts/inspect_agent_capabilities.js - 先讀:
docs/loop_maturity_model.mddocs/engineering_phase_loop.mddocs/loop_circuit_breaker.md
- 根據情境讀:
prompts/interactive_tech_stack_prompt.mdprompts/interactive_api_key_prompt.mdprompts/legacy_project_improvement_prompt.md
- 若缺
.loop/*,先用goal-loop或node scripts/init_session_loop.js . --goal "<objective>" - 技術棧確認後,先用
project-config-generation落設定檔 - 先給短版方案
- 視任務大小決定:
- 小任務:直接切 task 開工
- 中大型任務:再補完整計畫與 task 切片
- 完成必要確認後,進入 build-test-fix-doc loop
- 每輪記錄
maturity_level、retry_count、blocked_reason、convergence_state - 若是 JS / TS 專案且有 Biome,啟用 quality loop:
npx @biomejs/biome check --write <changed-paths>- 必要時再跑測試
- 進 merge / ship 前補:
code-review-and-qualityobservability-and-instrumentationshipping-and-launch
初始化
若是新專案或缺少核心檔案,先執行:
node scripts/init_project_workspace.js <targetDir> --name <projectName> --idea "<idea>"
再依情境推薦 skills / tools:
node scripts/auto_skill_setup.js --project-type <type> --ui-style <style> --deployment <target> --language <lang> --database <db>
實作 loop
每輪都做:
- 選一個最小 task
- 最小必要修改
- 跑對應驗證
- 失敗就修根因
- 看 circuit breaker 是否該停
- 更新:
.loop/PLAN.md.loop/STATE.json.loop/CHECKPOINTS.md.loop/EVIDENCE.md- 必要時記
tier、orientation、time_budget
- 若流程穩定可重用,參考
docs/skill_crystallization_loop.md - 更新:
docs/TASKS.mddocs/DEBUG_NOTES.mddocs/STATE.mdREADME.md
技能路由
- 想法模糊:
idea-refine - 要
/goal類跨平台 loop:goal-loop - 缺 spec:
spec-driven-development - 工作太大:
planning-and-task-breakdown - 技術棧已定但設定檔未落地:
project-config-generation - 開始實作:
incremental-implementation - 行為變動:
test-driven-development - framework 問題:
source-driven-development - 前端:
frontend-ui-engineering - API 邊界:
api-and-interface-design - runtime 驗證:
browser-testing-with-devtools - 除錯:
debugging-and-error-recovery - JS / TS 品質 loop:
biome-quality-automation - 合併前:
code-review-and-quality - 上線前觀測:
observability-and-instrumentation - 收尾文件:
documentation-and-adrs - 發版:
shipping-and-launch
工具路由
- 大 repo / 省 token:優先
Semble - 快速頁面驗證:優先
agent-browser - 可重現瀏覽器流程:優先
playwright-cli - 低 token MCP 探查:優先
mcp-cli - 沒有外部工具時,退回平台原生工具
子代理
- 平台支援 subagents 時,預設可自動分派
- 角色看
docs/subagent_dispatch.md - 不支援時,改為單代理順序執行
停止條件
只有以下情況才停:
- 缺憑證、權限、付款、登入
- 需要使用者做不可逆選擇
- 外部服務不可用且無替代方案
- 同一錯誤連修 3 次仍失敗
自治等級
- 預設
L1 report-only - 有 verifier 才升
L2 - 有 scope / budget / audit 才升
L3
規則
- 不吹噓未驗證能力
- 不假設所有平台都支援同一套 MCP / subagents
- 高風險決策採保守可逆預設,並寫入
docs/ADRS.md - 沒有驗證,不算完成
04 同仓库更多 Skills
website-best-practices
Build websites with strong semantic HTML, accessibility, performance, content structure.
using-agent-skills
Meta skill entrypoint. Select the right workflow skill for the current project phase, initialize required docs, and keep execution aligned with the project lifecycle.
ui-ux-designer
Create interface designs, design systems, and UX improvements with accessibility-first guidance.
typescript-best-practices
Apply TypeScript best practices for strict mode, safe narrowing, tsconfig safety, robust error handling.
tool-discovery-and-installation
Detect missing skills, MCP servers, CLIs, and sandbox tools; choose safe install paths; install or guide tool setup with explicit authorization and verification.
text-to-lottie
Generate or refine production-ready Lottie animations for web and app UI work.