skill
yibi-stack
Agentic skill stack for Claude Code: bash hygiene, Spectra/OpenSpec methodology, PR review workflows, TDD, and productivity tools.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/heyu-ai/yibi-stack itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 25, 2026yibi-stack
English
What is yibi-stack?
yibi-stack is a Claude Code skill and plugin stack for engineering teams who use Claude as their primary development tool. It transforms Claude Code from a capable-but-undisciplined assistant into a workflow-aware engineering partner with built-in guardrails.
The stack ships in two layers:
- Plugins (
plugins/) — installable viaclaude plugin install, add hooks and skills that activate automatically in every session - Skills (
skills/) — SKILL.md runbooks that tell Claude how to approach specific tasks (TDD, PR review, spec writing, release workflow)
Why yibi-stack?
Without guardrails, Claude Code tends to:
- Write bash commands that trigger Claude Code's internal confirmation dialogs (nested quotes, stateful
cd, inline multi-language) - Generate specs too shallow to build from
- Lose work context when conversations compress
- Skip test discipline under time pressure
yibi-stack layers three forms of enforcement on top of Claude Code:
- Automated hooks — catch problems before execution, not after
- Methodology skills — runbooks that embed spec-writing, test-design, and review discipline into Claude's workflow (test-first is enforced by the PR cycle's red-first gate, not a skill)
- Persistent tooling — mycelium, scheduler, and port registry that survive across conversations
Benefits
| Benefit | Mechanism |
|---|---|
| Safer shell execution | harness plugin intercepts 3 anti-pattern categories (AP1: overcomplex single commands / AP2: unicode in shell strings / AP3: stateful cd) via PreToolUse hooks — before they cause silent failures |
| Spec-first development | sdd plugin + spectra-amplifier skill expands thin requirements into 5-layer specs (Context → Behavior → Data → Constraints → Acceptance Criteria), compatible with OpenSpec change management |
| Multi-model PR review | pr-cycle-deep orchestrates Claude + Codex + Gemini in parallel independent review → cross-model debate → aggregate, catching issues no single model would flag |
| Persistent work memory | mycelium skill auto-handovers before context compression and restores on next session start — no more losing track of multi-day work |
| Release discipline | bump-version, protect-push, and ci-triage skills codify release workflow so Claude doesn't push to main without explicit intent |
| Test methodology | pr-cycle-deep Step 1.7 red-first gate — when the repo provides scripts/red-first-check.py (otherwise reported as unchecked — not a pass, shown to the human; yibi-stack itself ships none), feat/fix tests must fail against the base production code and refactor/perf tests must still pass — plus qa-test-design (6 test design techniques) |
Architecture
plugins/ Claude Code plugin packs (installable via claude plugin install)
harness/ Harness readiness eval + PreToolUse hook enforcement + anti-pattern guide
sdd/ Spec-Driven Development: Spectra + OpenSpec methodology
growth/ Knowledge extraction: mycelium, learn, PR retro/control log, CLAUDE.md prune
dev-cycle/ Development flow: PR cycles, CI triage, setup/handover, local port manager, debug
3rd-tools/ Third-party AI: Codex, Antigravity CLI (agy/Gemini), model verification
methodology/ Portable event-storming, problem-frames, QA test design methodology
skills/ Agent execution layer -- SKILL.md runbooks (installed via make install)
<skill-name>/ Each skill is a flat directory with a SKILL.md runbook (or a symlink)
scope:global -- works in any repo (methodology, cross-project tools)
scope:project -- requires this repo's Python tasks
tasks/ Python implementation (CLI, service, models, SQLite DB)
commands/ Claude Code slash commands (symlinked to ~/.claude/commands/)
scripts/ CI and lint tooling
Plugins vs Skills — what's the difference?
Plugins (plugins/harness, plugins/sdd, plugins/growth, plugins/dev-cycle, plugins/3rd-tools, plugins/methodology)
are proper Claude Code plugins with package.json manifests. They install hooks, rules, and bundled skills. Installable via claude plugin install without cloning.
Skills (skills/*/SKILL.md) are runbook files — not plugins. They're installed as symlinks into ~/.claude/skills/ via make install.
They tell Claude how to approach a workflow; no hooks are involved. Skills are not individually installable via claude plugin install.
Install
The two user installation tracks are complementary: Track 1 installs the Claude Code layer, while Track 2 installs the Python CLIs required by tasks-backed skills.
Track 1 — Claude Code plugins
Use this track for the skills, commands, rules, and hooks bundled in each plugin. It is lightweight and does not require a clone.
claude plugin marketplace add heyu-ai/yibi-stack
claude plugin install harness@yibi-stack
claude plugin install sdd@yibi-stack
claude plugin install growth@yibi-stack
claude plugin install dev-cycle@yibi-stack
claude plugin install methodology@yibi-stack
Track 2 — tasks-backed CLI tools
Use this track to install the mycelium, pr-orchestrator, and portman console scripts consumed by tasks-backed skills:
uv tool install "yibi-stack @ git+https://github.com/heyu-ai/yibi-stack@v1.23.2"
Track 2 is required by these six skills, grouped by their plugin pack:
growth:pr-control-log,pr-retrospective,mycelium, andlearndev-cycle:pr-cycle-fastandlocal-port-manager
Without Track 2, those skills fail loud before doing any work and show the same command above as the install hint.
Developer/contributor path — local checkout
Use a local checkout when developing or contributing to yibi-stack, or when installing repo-managed skills, hooks, and the scheduler. This path is separate from the two user tracks above. Tasks-backed harness skills still require the checkout setup below for their Python implementation.
# 1. Install plugins (pre-execution hooks + rules)
claude plugin marketplace add heyu-ai/yibi-stack
claude plugin install harness@yibi-stack sdd@yibi-stack growth@yibi-stack dev-cycle@yibi-stack methodology@yibi-stack
# 2. Clone and install skills + hooks + scheduler
git clone https://github.com/heyu-ai/yibi-stack
cd yibi-stack
make install-all
Verify install status (shows only this repo's skills, excludes gstack/external):
make status-own
Key Skills
| Skill | What it does |
|---|---|
spectra-amplifier | 5-layer spec expansion via sdd plugin |
pr-review-cycle | Full PR lifecycle: create -> parallel review (4 subagents) -> fix -> CI -> merge |
pr-cycle-fast | Fast lifecycle orchestrator: Python state machine, 1 reviewer, resumable |
pr-cycle-deep | Deep lifecycle: mob review (Claude + Codex + Gemini) + SDD amplifier-verifier |
bash-anti-patterns | AP1/AP2/AP3 detection guide + shell quoting hygiene reference |
qa-test-design | 6 test design techniques (equivalence, boundary, decision table...) |
mycelium | Cross-session work handover and insight collection |
bump-version | Version bump (Flutter/Python/Node/Go) + CHANGELOG + git tag |
protect-push | Hook to prevent accidental pushes from worktree branches to main |
ci-triage | CI failure triage funnel (Lint -> Type -> Security -> Tests) |
learn | Browse, search, prune, and export lessons learned |
pr-retrospective | 5-question PR retro, routes lessons to .claude/rules/ or CLAUDE.md |
claude-md-prune | Prune CLAUDE.md bloat: route gotchas to .claude/rules/ files, stay under 200-line soft cap |
agy | Antigravity CLI (Gemini) second opinion: lightweight single-model review / challenge mode |
harness-eval | 11-dimension harness readiness assessment with PASS/WARN/FAIL checklist and priority TODO |
See skills/README.md for the full index.
Plugins
| Plugin | Install | Description |
|---|---|---|
harness | claude plugin install harness@yibi-stack | Harness readiness evaluation, pre-execution bash anti-pattern detection, worktree push protection, and fleet usage burn-rate guarding. |
sdd | claude plugin install sdd@yibi-stack | Spectra + OpenSpec spec-amplifier methodology |
growth | claude plugin install growth@yibi-stack | Knowledge extraction and retention: mycelium, learn, PR retro/control log, CLAUDE.md prune |
dev-cycle | claude plugin install dev-cycle@yibi-stack | Development workflow: PR cycles, CI triage, setup/handover, local port manager, debug |
3rd-tools | claude plugin install 3rd-tools@yibi-stack | Codex, Antigravity CLI (agy/Gemini), model verification |
methodology | claude plugin install methodology@yibi-stack | Portable methodology: event storming, problem frames, QA test design. |
繁體中文
yibi-stack 是什麼?
yibi-stack 是一套專為以 Claude Code 作為主力開發工具的工程師設計的 skill 與 plugin 集。它讓 Claude Code 從一個能力強但缺乏紀律的助手,升級為有工作流程意識、有自我約束力的工程夥伴。
這個 stack 分兩層:
- Plugins(
plugins/)— 透過claude plugin install安裝,每次 session 自動啟用 hook 與 skill - Skills(
skills/)— SKILL.md runbook,告訴 Claude 如何處理特定任務(TDD、PR review、規格撰寫、發版流程)
為什麼需要 yibi-stack?
沒有護欄的 Claude Code 容易出現這些問題:
- 寫出觸發 Claude Code 確認框的 bash 指令(巢狀引號、stateful
cd、內嵌多語言) - 產生粒度太粗、無法直接實作的規格
- 對話壓縮後遺失工作脈絡
- 在時間壓力下跳過測試紀律
yibi-stack 在 Claude Code 之上疊加三層約束:
- 自動化 hook — 在執行前攔截問題,而不是事後除錯
- 方法論 skill — 把規格撰寫、測試設計、PR 審閱的紀律嵌入 Claude 的工作流程(先寫測試由 PR 流程的 red-first gate 執行,不是靠 skill)
- 持久化工具 — 跨對話的 session 記憶、定期排程器、port 登記表
主要好處
| 好處 | 機制 |
|---|---|
| 更安全的 shell 執行 | harness plugin 透過 PreToolUse hook,在執行前攔截三類反模式(AP1 過複雜單行 / AP2 bash 字串 Unicode / AP3 stateful cd),避免靜默失敗 |
| 規格先行的開發 | sdd plugin + spectra-amplifier skill 把薄需求展開為五層規格(情境→行為→資料→約束→驗收條件),兼容 OpenSpec 變更管理框架 |
| 多模型 PR 審閱 | pr-cycle-deep 讓 Claude + Codex + Gemini 並行獨立審閱再交叉辯論,捕捉單一模型漏掉的問題 |
| 持久化工作記憶 | mycelium skill 在對話壓縮前自動交班,下次 session 開啟時自動恢復工作上下文,多日開發不斷線 |
| 發版紀律 | bump-version + protect-push + ci-triage 讓 Claude 不會在沒有明確意圖的情況下推上 main |
| 測試方法論 | pr-cycle-deep Step 1.7 的 red-first gate:repo 有提供 scripts/red-first-check.py 時才執行(沒有就回報 unchecked:不是通過,交給人確認;yibi-stack 本身沒有附),feat/fix 的測試在 base 產品碼上必須失敗、refactor/perf 必須仍通過;另有 qa-test-design(六大測試設計技術) |
架構
plugins/ Claude Code plugin packs(可透過 claude plugin install 安裝)
harness/ Harness 就緒度評量 + PreToolUse hook 防線 + 反模式修法指南
sdd/ Spec-Driven Development:Spectra + OpenSpec 方法論
growth/ 知識萃取與留存:mycelium、learn、PR 回顧/審計、CLAUDE.md 精簡
dev-cycle/ 開發流程:PR cycles、CI 診斷、工作區設定/交班、本機 port 管理、debug
3rd-tools/ 第三方 AI:Codex、Antigravity CLI(agy/Gemini)、模型驗證
methodology/ 可攜方法論:event-storming、problem-frames、qa-test-design
skills/ Agent 執行介面層(SKILL.md runbook,透過 make install 安裝)
<skill-name>/ 每個 skill 是一個目錄(或指向 plugins/ 的 symlink)
scope:global -- 跨專案可用(方法論、通用工具)
scope:project -- 本 repo 限定(需要 tasks/ Python 實作)
tasks/ Python 實作(CLI、service、models、SQLite DB)
commands/ Claude Code slash commands(symlink 到 ~/.claude/commands/)
scripts/ CI 與 lint 工具腳本
Plugin 與 Skill 的差別?
Plugin(plugins/harness、plugins/sdd、plugins/growth、plugins/dev-cycle、plugins/3rd-tools、plugins/methodology)
是有 package.json manifest 的正式 Claude Code plugin,會安裝 hook、rules 和隨附 skill,不需 clone 即可用 claude plugin install 安裝。
Skill(skills/*/SKILL.md)是 runbook 檔案,不是 plugin。透過 make install 以 symlink 安裝到 ~/.claude/skills/,告訴 Claude 如何執行特定工作流程。Skills 無法透過 claude plugin install 個別安裝。
安裝
兩條使用者安裝軌彼此互補:Track 1 安裝 Claude Code 層,Track 2 安裝 tasks-backed skill 所需的 Python CLI。
Track 1 — Claude Code plugin
這條安裝軌提供各 plugin 隨附的 skill、command、rule 與 hook,輕量且不需 clone repo。
claude plugin marketplace add heyu-ai/yibi-stack
claude plugin install harness@yibi-stack
claude plugin install sdd@yibi-stack
claude plugin install growth@yibi-stack
claude plugin install dev-cycle@yibi-stack
claude plugin install methodology@yibi-stack
Track 2 — tasks-backed CLI 工具
這條安裝軌提供 tasks-backed skill 使用的 mycelium、pr-orchestrator 與 portman console script:
uv tool install "yibi-stack @ git+https://github.com/heyu-ai/yibi-stack@v1.23.2"
以下六個 skill 需要 Track 2,依所屬 plugin pack 分組如下:
growth:pr-control-log、pr-retrospective、mycelium、learndev-cycle:pr-cycle-fast、local-port-manager
若未安裝 Track 2,這些 skill 會在執行任何工作前明確輸出 [FAIL],並顯示與上方完全相同的安裝提示。
開發者/貢獻者路徑 — local checkout
開發或貢獻 yibi-stack,或需要安裝 repo 管理的 skill、hook 與 scheduler 時,請使用 local checkout。 這條路徑獨立於上方兩條使用者安裝軌。 Tasks-backed harness skills 的 Python 實作仍需要下方 checkout 設定。
# 1. 安裝 plugin(pre-execution hook + 規則)
claude plugin marketplace add heyu-ai/yibi-stack
claude plugin install harness@yibi-stack sdd@yibi-stack growth@yibi-stack dev-cycle@yibi-stack methodology@yibi-stack
# 2. Clone 並安裝 skill + hook + scheduler
git clone https://github.com/heyu-ai/yibi-stack
cd yibi-stack
make install-all
確認安裝狀態(只顯示本 repo 的 skill,排除 gstack / 外部安裝):
make status-own
主要 Skills
| Skill | 功能 |
|---|---|
spectra-amplifier | 五層規格展開(透過 sdd plugin) |
pr-review-cycle | 完整 PR 生命週期:建立 → 並行 review(4 subagent)→ 修正 → CI → merge |
pr-cycle-fast | 快速 lifecycle orchestrator:Python state machine,可中斷 resume |
pr-cycle-deep | 深度 lifecycle:mob review(Claude + Codex + Gemini)+ SDD amplifier-verifier |
bash-anti-patterns | AP1/AP2/AP3 偵測指南 + shell 引號衛生參考 |
qa-test-design | 六大測試設計技術(等價類別、邊界值、決策表……) |
mycelium | 跨對話工作交班與洞察收集 |
bump-version | 版本 bump(Flutter/Python/Node/Go)+ CHANGELOG + git tag |
protect-push | 防止 worktree branch 意外推上 main 的 hook |
ci-triage | CI 失敗快速診斷漏斗(Lint → Type → Security → Tests) |
learn | 瀏覽、搜尋、修剪、匯出教訓記錄 |
pr-retrospective | PR 收尾五問回顧,路由 lesson 到 .claude/rules/ 或 CLAUDE.md |
claude-md-prune | CLAUDE.md 精簡:把 gotcha 路由到 .claude/rules/ 子檔,維持 200 行軟上限 |
agy | Antigravity CLI(Gemini)第二意見:輕量單一模型 review / 對抗模式 |
harness-eval | 11 維度 harness 就緒度評量,附 PASS/WARN/FAIL 清單與優先改善 TODO |
完整索引見 skills/README.md。
Plugins
| Plugin | 安裝指令 | 說明 |
|---|---|---|
harness | claude plugin install harness@yibi-stack | Harness 就緒度評量(11 維度)+ 執行前 bash 反模式偵測與自動修法指引 + worktree push 保護。 |
sdd | claude plugin install sdd@yibi-stack | Spectra + OpenSpec 規格展開 |
growth | claude plugin install growth@yibi-stack | 知識萃取與留存:mycelium、learn、PR 回顧/審計、CLAUDE.md 精簡 |
dev-cycle | claude plugin install dev-cycle@yibi-stack | 開發流程:PR cycles、CI 診斷、工作區設定/交班、本機 port 管理、debug |
3rd-tools | claude plugin install 3rd-tools@yibi-stack | Codex、Antigravity CLI(agy/Gemini)、模型驗證 |
methodology | claude plugin install methodology@yibi-stack | 可攜方法論:event storming、problem frames、QA test design。 |
License
MIT