diff --git a/.out-of-scope/harness-specific-branches.md b/.out-of-scope/harness-specific-branches.md
new file mode 100644
index 0000000..2a80b0a
--- /dev/null
+++ b/.out-of-scope/harness-specific-branches.md
@@ -0,0 +1,19 @@
+# 已拒绝：为某一种 harness 单开分支
+
+- 状态：拒绝（长期有效）
+- 日期：2026-10-31
+
+## 想法
+
+为 Codex / DSH / Claude 某一个客户端的特例，单独复制一份技能或加一个 harness 分支。
+
+## 为什么拒绝
+
+1. 复制出来的两份技能会立刻开始漂移，评测也要做两遍；
+2. 客户端差异应交给安装器（目录、调用模式、字段映射）处理，不该写进技能内容；
+3. 特例需求通常是个人的工作流调整，属于 `AGENTS.md` / `CLAUDE.md` 或 fork 的范围。
+
+## 允许的替代路径
+
+- 客户端差异 → 安装器加 target / 字段映射；
+- 个人流程 → 个人目录或 fork。
diff --git a/.out-of-scope/new-skills.md b/.out-of-scope/new-skills.md
new file mode 100644
index 0000000..ac5132d
--- /dev/null
+++ b/.out-of-scope/new-skills.md
@@ -0,0 +1,20 @@
+# 已拒绝：新增「随手想到」的技能
+
+- 状态：拒绝（长期有效）
+- 日期：2026-10-31
+
+## 想法
+
+「我觉得应该有个技能做 X」——从想象出发新增技能，而不是从真实失败出发。
+
+## 为什么拒绝
+
+1. 技能目录是每次会话启动就要付的上下文预算，加一个就多一份；
+2. 没有真实失败做依据的技能，写出来是泛泛而谈，实际触发率极低；
+3. 技能之间会互相抢触发（description 相似），新增一个会稀释已有技能的准确率。
+
+## 允许的替代路径
+
+- 先用现有技能做一遍真实任务，记录卡点；
+- 确实组合不出来，再按 `SCOPE.md` 门槛提，附真实失败证据；
+- 只是个人偏好 → 放 `skills-personal/`，不进团队库。
diff --git a/.out-of-scope/personal-workflow-tweaks.md b/.out-of-scope/personal-workflow-tweaks.md
new file mode 100644
index 0000000..33b26f3
--- /dev/null
+++ b/.out-of-scope/personal-workflow-tweaks.md
@@ -0,0 +1,19 @@
+# 已拒绝：把个人工作流调优塞进团队技能
+
+- 状态：拒绝（长期有效）
+- 日期：2026-10-31
+
+## 想法
+
+把某个人的偏好（提交信息格式、分支命名、审查语气、工具习惯）直接写进团队技能。
+
+## 为什么拒绝
+
+1. 团队技能要能被所有人用，个人偏好会变成隐式规训；
+2. 一旦写进技能，改起来要走 PR，个人改不动；
+3. 这类偏好本来就该放个人 `AGENTS.md` / `CLAUDE.md`。
+
+## 允许的替代路径
+
+- 个人偏好 → 个人配置文件或 `skills-personal/`；
+- 如果是团队共识（不是个人偏好），走 `SCOPE.md` 门槛并附真实失败。
diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000..68b4d8a
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,21 @@
+# Changelog
+
+## Unreleased — 2026-10-31
+
+### 新增
+
+- 技能治理：`SCOPE.md` 收录门槛 + `.out-of-scope/` 拒绝档案。
+- 生命周期分桶（`promoted` / `beta` / `deprecated`）与调用模式（`model-invoked` / `user-invoked`）。
+- `scripts/validate-skills.py`：规范/大小/引用/调用模式/文档校验。
+- `scripts/generate-catalog.py`：生成可审阅的 `catalog.json`。
+- `scripts/scan-skill.sh` + `scripts/scan-allowlist.txt`：入库前安全扫描。
+- 新技能（beta）：`skill-creator`、`ai-proj-helper-entry`、`verification-before-completion`。
+- `install-skills.sh`：`--agent dsh`、`--link`、`--audit`、`--include-beta`、`--no-validate`。
+
+### 变更
+
+- `dev-review`：增加「接收与响应评审意见」。
+- `req-design`：增加「任务依赖边（blocking edges）」。
+- `agent-swarm`：增加「并行编排流程」。
+- `dev-test`：增加「TDD 循环」。
+- `generate-marketplace.py`：只发布 `promoted`，beta/deprecated 不随 marketplace 分发。
diff --git a/CLAUDE.md b/CLAUDE.md
index c1ae585..3ae2cea 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -52,3 +52,10 @@ skills:
 - `skills-personal/` 不被 Git 跟踪，用于存放个人配置和工具
 - 其余 `skills-*` 目录均由 Git 版本控制
 - 只有受 Git 跟踪的分类目录会被 `generate-marketplace.py` 扫描并加入公开 marketplace；`skills-personal/` 始终排除
+
+## 技能治理规则
+
+- 公开 marketplace 只发布 `status=promoted`；`beta`/`deprecated` 不进 marketplace（`generate-marketplace.py` 已按 status 过滤）。
+- 每个 `plugin.json` 建议带 `status`，每个 `SKILL.md` 在 `metadata.invocation` 标 `model-invoked` / `user-invoked`。
+- 任何技能增删改后，回改 `catalog.json` 与入口技能 `ai-proj-helper-entry`。
+- 提交前依次跑：`python3 scripts/validate-skills.py`、`./scripts/scan-skill.sh`、`python3 scripts/generate-catalog.py`。
diff --git a/README.md b/README.md
index ce150be..77de812 100644
--- a/README.md
+++ b/README.md
@@ -195,3 +195,17 @@ A: SSE 模式直连远程服务器，零依赖开箱即用（推荐）。stdio 
 - **管理员**：Donglin Lai (qiudl)
 - **邮箱**：qiudl@zhiyuncai.com
 - **仓库**：https://gitea.pipexerp.com/pipexerp/ai-proj-helper
+
+## 技能治理（2026-10-31 起）
+
+- **真相源**：本仓库。`~/.agents/skills`、`~/.claude/skills`、`$DSH_HOME/skills` 都只是安装产物。
+- **收录门槛**见 [SCOPE.md](./SCOPE.md)：只收「真实失败 + 组合不出来」，被拒的想法记在 `.out-of-scope/`。
+- **生命周期**：`promoted`（默认安装）/ `beta`（需 `--include-beta`）/ `deprecated`（不安装）；新技能一律先进 beta。
+- **调用模式**：`model-invoked`（占模型目录预算）vs `user-invoked`（只由人触发，不占目录预算）。
+- **安装**：
+  - Codex：`./install-skills.sh --agent codex`
+  - DSH：`DSH_HOME=<path> ./install-skills.sh --agent dsh --link`
+  - Claude（兼容）：`./install-skills.sh --agent claude`
+  - 看漂移：`./install-skills.sh --audit`
+- **校验 / 安全 / 清单**：`python3 scripts/validate-skills.py`、`./scripts/scan-skill.sh`、`python3 scripts/generate-catalog.py`
+- **建技能**：走 `skill-creator` 技能（收录门槛、分桶、调用模式、evals、四段式文档、路由回改），不要手写。
diff --git a/SCOPE.md b/SCOPE.md
new file mode 100644
index 0000000..9ff52bb
--- /dev/null
+++ b/SCOPE.md
@@ -0,0 +1,45 @@
+# Scope — ai-proj-helper 技能的收录门槛
+
+本仓库是团队每天在用的技能集，**不是技能收藏夹**。技能目录会在每次会话启动时占用模型的上下文预算，所以「多加一个」永远是有代价的。
+
+## 门槛
+
+新技能或改动**默认拒绝**，只有同时满足两条才进入评审：
+
+1. **真实失败**：能说清在哪个真实任务/会话里出了问题——跑了什么、期望什么、实际什么。假设性的「这样会更好」不算。
+2. **组合不出来**：用现有技能组合做不到。能组合的，改现有技能，不新增。
+
+规模不重要：一个词的修改和一次重写按同一标准审。
+
+## 分桶
+
+| 桶 | 含义 | 安装 | 进 README / plugin.json |
+| --- | --- | --- | --- |
+| `promoted` | 已验证、日常在用 | 默认 | 必须 |
+| `beta` | 公开试跑、收反馈 | `--include-beta` | 不进默认清单 |
+| `deprecated` | 不再使用 | 不安装 | 只留文档，标弃用 |
+
+新技能一律先进 `beta`；promoted 需要真实使用记录 + evals 通过。
+
+## 调用模式
+
+- `model-invoked`（默认）：进模型目录，agent 自己判断何时用。
+- `user-invoked`：只由人下命令触发，不占目录预算；必须写 `disable-model-invocation: true` + `metadata.invocation: user-invoked`。
+
+判定：这件事是不是「人下命令才开始」？是 → user-invoked。
+
+## 已经拒绝的想法
+
+每个被拒概念一个文件，写在 [`.out-of-scope/`](./.out-of-scope/) 下，先读再提。当前已有：
+
+- [`new-skills.md`](./.out-of-scope/new-skills.md)
+- [`harness-specific-branches.md`](./.out-of-scope/harness-specific-branches.md)
+- [`personal-workflow-tweaks.md`](./.out-of-scope/personal-workflow-tweaks.md)
+
+## 提交流程
+
+1. 先过 [`skill-creator`](./skills-core/skill-creator-plugin/skills/SKILL.md) 的收录门槛与质量门禁。
+2. `python3 scripts/validate-skills.py` 通过。
+3. `./scripts/scan-skill.sh` 无高危项。
+4. `python3 scripts/generate-catalog.py` 更新 `catalog.json`。
+5. 版本号 + CHANGELOG 更新，走 PR。
diff --git a/SYNC-GUIDE.md b/SYNC-GUIDE.md
index 6360413..5819b97 100644
--- a/SYNC-GUIDE.md
+++ b/SYNC-GUIDE.md
@@ -67,3 +67,24 @@ Claude Code 需要显式选择目标：
 **如何移除技能？**
 
 删除插件目录、重新生成 marketplace、提交变更。使用者随后执行 `./install-skills.sh --dry-run --cleanup`，确认后再去掉 `--dry-run`。
+
+## 安装矩阵（2026-10 更新）
+
+| agent | 安装目录 | 默认模式 | 状态文件 |
+| --- | --- | --- | --- |
+| `codex`（默认） | `~/.agents/skills` | copy | `~/.agents/.ai-proj-helper-installed-skills.json` |
+| `dsh` | `$DSH_HOME/skills` | copy（开发机建议 `--link`） | `$DSH_HOME/.ai-proj-helper-installed-skills.json` |
+| `claude`（兼容） | `~/.claude/skills` | copy | `~/.claude/.installed-skills.json` |
+
+```bash
+# 本机（DSH 开发机，改仓库即时生效）
+DSH_HOME="<your DSH_HOME>" ./install-skills.sh --agent dsh --link
+
+# 远程/其他机器
+./install-skills.sh --agent codex
+
+# 看谁落后了
+./install-skills.sh --agent codex --audit
+```
+
+`beta` 技能默认不装：需要时加 `--include-beta`。`promoted/beta/deprecated` 见 `SCOPE.md`。
diff --git a/catalog.json b/catalog.json
new file mode 100644
index 0000000..41803fa
--- /dev/null
+++ b/catalog.json
@@ -0,0 +1,683 @@
+{
+  "repo": "ai-proj-helper",
+  "generated_at": "2026-10-31T14:03:00Z",
+  "total": 61,
+  "counts": {
+    "promoted/model-invoked": 58,
+    "beta/user-invoked": 2,
+    "beta/model-invoked": 1
+  },
+  "skills": [
+    {
+      "name": "biz-contract",
+      "category": "biz",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "商务合同撰写。支持多种合同类型：软件订阅合同、软件定制开发合同、物流合同、销售服务合同、贸易合同等。当用户提到合同、协议、签约、合作协议相关任务时自动激活。",
+      "description_len": 78,
+      "path": "skills-biz/biz-contract-plugin/skills/SKILL.md",
+      "digest": "b8a94cb7dece26e4"
+    },
+    {
+      "name": "biz-ops",
+      "category": "biz",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "商务运营技能。支持商业计划书(BP)撰写和商务合同起草。当用户提到商业计划书、BP、融资计划、商业模式、合同、协议、签约等相关任务时自动激活。",
+      "description_len": 71,
+      "path": "skills-biz/biz-ops-plugin/skills/SKILL.md",
+      "digest": "c4959eef7dabfbe3"
+    },
+    {
+      "name": "biz-plan",
+      "category": "biz",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "商业计划书撰写。支持融资BP、战略规划、商业模式设计等。当用户提到商业计划书、BP、融资计划、商业模式、战略规划相关任务时自动激活。",
+      "description_len": 66,
+      "path": "skills-biz/biz-plan-plugin/skills/SKILL.md",
+      "digest": "4229b7a96501e0c2"
+    },
+    {
+      "name": "finance",
+      "category": "biz",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "财务对账技能。支持银行对账、应收对账、应付对账、内部往来对账等。通过自然语言实现账单导入、差异分析、对账报告生成。当用户提到对账、银行流水、应收应付、账单核对、财务核销相关任务时自动激活。",
+      "description_len": 94,
+      "path": "skills-biz/finance-plugin/skills/SKILL.md",
+      "digest": "c3f3eabd01ff7cac"
+    },
+    {
+      "name": "ai-proj-helper-entry",
+      "category": "core",
+      "version": "0.1.0",
+      "status": "beta",
+      "invocation": "user-invoked",
+      "description": "ai-proj-helper 技能库路由：告诉你有哪些技能、什么任务用哪个、以及「REQ released + PDV 通过才算交付」的强制口径。人在开始需求/开发/评审/发布任务前主动调用。",
+      "description_len": 96,
+      "path": "skills-core/ai-proj-helper-entry-plugin/skills/SKILL.md",
+      "digest": "41d39b76d4c71dbf"
+    },
+    {
+      "name": "ai-proj",
+      "category": "core",
+      "version": "2.0.1",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "ai-proj 任务与需求管理。用于创建/查询/更新需求、任务、文档。当用户提到 ai-proj、任务管理、需求管理、REQ-XXX 相关操作时自动激活。",
+      "description_len": 77,
+      "path": "skills-core/ai-proj-plugin/skills/SKILL.md",
+      "digest": "f2ef00fdebb29034"
+    },
+    {
+      "name": "pm-ask",
+      "category": "core",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "基于 ai-proj MCP + git log 真实数据的项目问答。禁止编造。当用户说\"/ask\"、\"问一下\"、\"统计一下\"、\"最近多少\"、\"谁做了多少\"、\"本月进度\"等项目相关问题时自动激活。",
+      "description_len": 98,
+      "path": "skills-core/pm-ask-plugin/skills/SKILL.md",
+      "digest": "e5b69a0b56985472"
+    },
+    {
+      "name": "pm-risk",
+      "category": "core",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "三维度项目风险扫描（需求层/代码层/流程层）。当用户说\"/risk\"、\"风险扫描\"、\"有什么风险\"、\"哪些需求停滞了\"、\"分支健康\"时自动激活。",
+      "description_len": 72,
+      "path": "skills-core/pm-risk-plugin/skills/SKILL.md",
+      "digest": "928cfec76018329f"
+    },
+    {
+      "name": "publish",
+      "category": "core",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "发布 ai-proj-helper 技能市场更新并通知飞书群。用户说 /publish 时激活。",
+      "description_len": 48,
+      "path": "skills-core/publish-plugin/skills/SKILL.md",
+      "digest": "222151d3dbf35099"
+    },
+    {
+      "name": "skill-creator",
+      "category": "core",
+      "version": "0.1.0",
+      "status": "beta",
+      "invocation": "user-invoked",
+      "description": "团队技能的标准流程：收录门槛、分桶生命周期、调用模式、SKILL.md 规范、evals 全量、四段式文档、路由回改。当要新建技能、大改现有技能、或评审一个技能是否合格时使用。",
+      "description_len": 88,
+      "path": "skills-core/skill-creator-plugin/skills/SKILL.md",
+      "digest": "18f18c43c3fa56a6"
+    },
+    {
+      "name": "agent-browser",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "浏览器自动化技能。用于网页交互、E2E冒烟测试、需求验收验证、前端开发验证、截图对比。基于 Vercel agent-browser CLI。",
+      "description_len": 71,
+      "path": "skills-dev/agent-browser-plugin/skills/SKILL.md",
+      "digest": "f382bffe3d1c84b2"
+    },
+    {
+      "name": "agent-swarm",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Multi-agent orchestration using OpenAI Swarm patterns. Coordinate specialized agents for complex development workflows with handoffs and context sharing.",
+      "description_len": 153,
+      "path": "skills-dev/agent-swarm-plugin/skills/SKILL.md",
+      "digest": "a0d8a329ab9468f5"
+    },
+    {
+      "name": "ai-chat",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "AI Chat 测试与管理。发送消息测试 AI Chat 工具调用链路，管理工具开关和 Provider 配置，支持 local/staging 环境切换。",
+      "description_len": 78,
+      "path": "skills-dev/ai-chat-plugin/skills/SKILL.md",
+      "digest": "97ce03993394605c"
+    },
+    {
+      "name": "ai-proj-cicd-release",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Execute and audit the AI-Proj service CI/CD flow across Gitea gates, immutable image builds, staging verification, approved production release, rollback, and evidence capture. Use for AI-Proj CI status, release PRs, staging or production deployments, failed-release diagnosis, deployed commit/schema verification, and delivery-chain repair. Do not use for unrelated repositories or macOS application packaging.",
+      "description_len": 410,
+      "path": "skills-dev/ai-proj-cicd-release-plugin/skills/SKILL.md",
+      "digest": "65adfbf96fc4b392"
+    },
+    {
+      "name": "ai-proj-macos-release",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Build, sign, notarize, publish, and verify the AI-Proj macOS Apple Silicon package through the repository's release chain. Use when asked to release, republish, update, or repair the downloadable macOS application, including Gatekeeper failures and download-manifest updates. Do not use for service deployments or unrelated applications.",
+      "description_len": 337,
+      "path": "skills-dev/ai-proj-macos-release-plugin/skills/SKILL.md",
+      "digest": "595fbd781968025e"
+    },
+    {
+      "name": "db-migration",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "数据库变更方案插件。Migration 脚本生成、数据迁移策略、回滚方案。挂载在 design 阶段，涉及数据库变更时由 req-design 推荐激活。",
+      "description_len": 77,
+      "path": "skills-dev/db-migration-plugin/skills/SKILL.md",
+      "digest": "f8d6ebbddbbf7a71"
+    },
+    {
+      "name": "defect-analysis",
+      "category": "dev",
+      "version": "1.1.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "系统性设计缺陷分析。对需求方案/代码架构进行多维度检查，发现隐藏的技术风险和设计漏洞。当用户提到缺陷检查、方案审查、设计审计时自动激活。",
+      "description_len": 68,
+      "path": "skills-dev/defect-analysis-plugin/skills/SKILL.md",
+      "digest": "23dfafda92d0c622"
+    },
+    {
+      "name": "deploy-rollback",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "回滚方案插件。部署后发现问题时的回滚策略和执行步骤。挂载在 deploy 阶段，部署出问题时激活。",
+      "description_len": 49,
+      "path": "skills-dev/deploy-rollback-plugin/skills/SKILL.md",
+      "digest": "75ebf66a3f74195a"
+    },
+    {
+      "name": "dev-android",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Android 开发插件。Kotlin + Jetpack Compose + Hilt 依赖注入。当涉及 Android 开发任务时按需加载。",
+      "description_len": 72,
+      "path": "skills-dev/dev-android-plugin/skills/SKILL.md",
+      "digest": "ca7442afd02c6129"
+    },
+    {
+      "name": "dev-arch",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "软件架构设计技能。用于系统设计、技术选型、架构评审、设计文档编写。当用户提到架构设计、系统设计、技术方案、API 设计相关任务时自动激活。",
+      "description_len": 69,
+      "path": "skills-dev/dev-arch-plugin/skills/SKILL.md",
+      "digest": "c9a419c83a307a11"
+    },
+    {
+      "name": "dev-cicd",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "CI/CD 流水线设计、优化与排查。适配 Gitea Actions + Go/Swift/Next.js/Docker 栈。当用户提到 CI、CD、流水线、pipeline、workflow、构建失败、runner 相关任务时自动激活。",
+      "description_len": 119,
+      "path": "skills-dev/dev-cicd-plugin/skills/SKILL.md",
+      "digest": "a763f0f140516411"
+    },
+    {
+      "name": "dev-coding",
+      "category": "dev",
+      "version": "2.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "软件编码开发技能。用于代码编写、功能实现、重构优化。集成 ai-proj CLI 进行任务管理和进度跟踪。核心支持 Go 后端 + Vue/React 前端开发。",
+      "description_len": 81,
+      "path": "skills-dev/dev-coding-plugin/skills/SKILL.md",
+      "digest": "1b754e668965132f"
+    },
+    {
+      "name": "dev-commit",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "智能 git commit — 分支保护 + 自动建功能分支 + Conventional Commits 生成。当用户说\"提交代码\"、\"commit\"、\"/commit\"、\"保存修改\"时自动激活。",
+      "description_len": 99,
+      "path": "skills-dev/dev-commit-plugin/skills/SKILL.md",
+      "digest": "1560658c4a30df45"
+    },
+    {
+      "name": "dev-deploy",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "应用部署技能。支持 iOS TestFlight、Docker 容器等多平台部署。当用户提到部署、发布、TestFlight、上架、build、archive 相关任务时自动激活。",
+      "description_len": 90,
+      "path": "skills-dev/dev-deploy-plugin/skills/SKILL.md",
+      "digest": "00873ab922fb1595"
+    },
+    {
+      "name": "dev-integration",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "前后端联调技能。API 契约验证、接口对接、联调报告生成。对应 req 流程 integration 阶段。纯后端需求自动跳过。",
+      "description_len": 64,
+      "path": "skills-dev/dev-integration-plugin/skills/SKILL.md",
+      "digest": "eff6be8eb57b7fef"
+    },
+    {
+      "name": "dev-ios",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "iOS 开发插件。Swift/SwiftUI + MVVM 架构，TestFlight 部署。当涉及 iOS 开发任务时按需加载。",
+      "description_len": 65,
+      "path": "skills-dev/dev-ios-plugin/skills/SKILL.md",
+      "digest": "364efd9f6f8158c2"
+    },
+    {
+      "name": "dev-mcp",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "MCP Bridge 开发插件。TypeScript MCP 服务开发，Token 管理，HTTP 客户端模式。当涉及 mcp-task-bridge 开发时按需加载。",
+      "description_len": 84,
+      "path": "skills-dev/dev-mcp-plugin/skills/SKILL.md",
+      "digest": "83d5912212fcfa4a"
+    },
+    {
+      "name": "dev-pda",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "PDA 应用开发插件。Android 原生 + 扫码枪集成 + 离线优先。当涉及 PDA/手持终端开发时按需加载。",
+      "description_len": 57,
+      "path": "skills-dev/dev-pda-plugin/skills/SKILL.md",
+      "digest": "e93c3ab182398571"
+    },
+    {
+      "name": "dev-review",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "代码评审技能。六视角对抗性扫描法（含 Karpathy Scope 审计），用于 PR 代码审查、安全评审、质量检查。当执行 /req cr 或独立 PR review 时自动激活。",
+      "description_len": 91,
+      "path": "skills-dev/dev-review-plugin/skills/SKILL.md",
+      "digest": "a8a1ec79b85e3227"
+    },
+    {
+      "name": "dev-scaffold",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "模块脚手架插件。新建模块时自动生成分层代码骨架。挂载在 dev 阶段，新建模块时激活。",
+      "description_len": 43,
+      "path": "skills-dev/dev-scaffold-plugin/skills/SKILL.md",
+      "digest": "4044787b1e468302"
+    },
+    {
+      "name": "dev-test",
+      "category": "dev",
+      "version": "2.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "软件测试技能。用于单元测试、集成测试、E2E测试、测试用例设计。支持 Go、Vue、React、iOS、Android 等多平台测试。",
+      "description_len": 67,
+      "path": "skills-dev/dev-test-plugin/skills/dev-test/SKILL.md",
+      "digest": "a03536a478eabe6c"
+    },
+    {
+      "name": "executing-plans",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Use when you have a written implementation plan to execute in a separate session with review checkpoints",
+      "description_len": 104,
+      "path": "skills-dev/executing-plans-plugin/skills/SKILL.md",
+      "digest": "8991a50091af83a4"
+    },
+    {
+      "name": "finishing-a-development-branch",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Use when implementation is complete and all tests pass - verifies and creates PR",
+      "description_len": 80,
+      "path": "skills-dev/finishing-branch-plugin/skills/SKILL.md",
+      "digest": "1102047ccc060e60"
+    },
+    {
+      "name": "frontend-design",
+      "category": "dev",
+      "version": "1.0.1",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Create distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, or applications. Generates creative, polished code that avoids generic AI aesthetics.",
+      "description_len": 234,
+      "path": "skills-dev/frontend-design-plugin/skills/SKILL.md",
+      "digest": "6bc30a1a562bae8c"
+    },
+    {
+      "name": "karpathy-guidelines",
+      "category": "dev",
+      "version": "1.0.1",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Karpathy 四原则编码行为守则。减少 LLM 常见编码错误：过度实现、静默假设、顺手重构、无验证标准。在任意编码场景激活。",
+      "description_len": 64,
+      "path": "skills-dev/karpathy-guidelines-plugin/skills/SKILL.md",
+      "digest": "a68f347031110d6a"
+    },
+    {
+      "name": "pull-request",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Use when starting new development tasks, creating pull requests, reviewing code, or managing PR feedback cycles on Gitea. Triggers on /pr commands or when user mentions PR, pull request, code review, or branch creation.",
+      "description_len": 219,
+      "path": "skills-dev/pull-request-plugin/skills/SKILL.md",
+      "digest": "40dac0d290782bb5"
+    },
+    {
+      "name": "review-checklist",
+      "category": "dev",
+      "version": "1.1.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "项目级代码评审检查清单。按项目积累特定检查项，挂载在 dev-review 下自动加载。每次 CR 时触发。",
+      "description_len": 54,
+      "path": "skills-dev/review-checklist-plugin/skills/SKILL.md",
+      "digest": "5bae13b6a8d92650"
+    },
+    {
+      "name": "slark-cicd",
+      "category": "dev",
+      "version": "1.1.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Slark 仓库 staging、生产和 Desktop 安装包的端到端 CI/CD 可执行 runbook。覆盖本地预推快检、PR 与内网 ci/internal-gate 门禁、合并 origin/main、生产前强制 staging 验收、审批式生产发布，以及在 m5max 构建签名、公证并上传 macOS/Windows Desktop 包到 OSS。当用户要在 slark（qiudl/qiu-slark）里发布/上线、部署到生产或 staging、发布 Desktop 安装包、盯 CI/合并 PR、复现或验证运行时行为、或排查发布链/推送路由问题时使用。",
+      "description_len": 285,
+      "path": "skills-dev/slark-cicd-plugin/skills/SKILL.md",
+      "digest": "2919c19d8a38dd23"
+    },
+    {
+      "name": "slark-dsh-cicd",
+      "category": "dev",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "DSH 插件包与 Slark 宿主补丁的发布 runbook。覆盖 npm run pack 双 tgz 双资产、无上游写权限时的 fork 发布、干净单提交与树一致性 gate、改指向后必须重新打包、draft→publish 节奏、GitHub HTTPS/SSH 与 fetch 卡死处置、发布后 SHA 核对、app.asar 完整性约束、ai-proj 发布状态对齐。当用户要发布或升级 DSH 插件（dsh-worktable 等）、发布/更新 Slark 宿主补丁、push 或下载 GitHub 失败、插件「检查更新」看不到新版本、需要 fork 发布、或排查 asar 完整性/签名问题时使用。",
+      "description_len": 306,
+      "path": "skills-dev/slark-dsh-cicd-plugin/skills/SKILL.md",
+      "digest": "44a1322e15d42845"
+    },
+    {
+      "name": "data-excel",
+      "category": "integration",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "Excel 数据处理与 BI 集成。通过自然语言操作 Excel 文件的读取、编辑、转换，并支持导入到 BI 系统（Metabase）进行可视化分析。当用户提到 Excel、表格处理、数据导入、BI 分析相关任务时自动激活。",
+      "description_len": 112,
+      "path": "skills-integration/data-excel-plugin/skills/SKILL.md",
+      "digest": "1d6a88d967ef8eaf"
+    },
+    {
+      "name": "doubao-voice",
+      "category": "integration",
+      "version": "1.0.1",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "豆包语音API调用。支持语音合成(TTS)和唱歌。当用户提到语音合成、文字转语音、唱歌、豆包语音相关任务时自动激活。",
+      "description_len": 58,
+      "path": "skills-integration/doubao-voice-plugin/skills/SKILL.md",
+      "digest": "716037a762904fdd"
+    },
+    {
+      "name": "feishu-bitable",
+      "category": "integration",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "飞书多维表格操作。用于记录增删改查、批量操作、筛选排序、数据同步。当需要操作飞书多维表格时使用。",
+      "description_len": 48,
+      "path": "skills-integration/feishu-bitable-plugin/skills/SKILL.md",
+      "digest": "e9678f06e0e02c52"
+    },
+    {
+      "name": "feishu-docx",
+      "category": "integration",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "飞书云文档操作。用于创建、编辑云文档，插入内容块，会议纪要生成。当需要操作飞书云文档时使用。",
+      "description_len": 46,
+      "path": "skills-integration/feishu-docx-plugin/skills/SKILL.md",
+      "digest": "a90e9c689c2e914d"
+    },
+    {
+      "name": "feishu",
+      "category": "integration",
+      "version": "1.1.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "飞书文档与多维表格操作入口。当用户提到飞书、云文档、多维表格、Bitable 相关任务时自动激活。",
+      "description_len": 49,
+      "path": "skills-integration/feishu-plugin/skills/SKILL.md",
+      "digest": "36c8167f43a3df88"
+    },
+    {
+      "name": "siyuan",
+      "category": "integration",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "思源笔记 API 集成。通过自然语言创建、编辑、搜索笔记，管理笔记本和文档树。当用户提到思源笔记、SiYuan、笔记管理、知识库相关任务时自动激活。",
+      "description_len": 74,
+      "path": "skills-integration/siyuan-plugin/skills/SKILL.md",
+      "digest": "92b97573d6682e8b"
+    },
+    {
+      "name": "siyuan-to-feishu",
+      "category": "integration",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "将思源笔记导出为 PDF 并发送到飞书群。当用户提到\"发送笔记\"、\"导出PDF发飞书\"、\"/siyuan send\"、\"分享到飞书\"相关任务时自动激活。",
+      "description_len": 76,
+      "path": "skills-integration/siyuan-to-feishu-plugin/skills/SKILL.md",
+      "digest": "4ddc5e600c520f48"
+    },
+    {
+      "name": "wecom",
+      "category": "integration",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "企业微信集成。通过自然语言发送消息、管理群机器人、操作审批流程、管理通讯录。当用户提到企业微信、微信工作、群机器人、企业号、wecom相关任务时自动激活。",
+      "description_len": 77,
+      "path": "skills-integration/wecom-plugin/skills/SKILL.md",
+      "digest": "57eacac81d6f76ab"
+    },
+    {
+      "name": "req-audit",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "部署后审计。运行时日志检查 + 静态缺陷分析 + 设计偏移检测。可独立调用或由 /req done 自动触发。",
+      "description_len": 55,
+      "path": "skills-req/req-audit-plugin/skills/SKILL.md",
+      "digest": "ff2e6197bbb059c6"
+    },
+    {
+      "name": "req-compare",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "对比式需求分析插件。系统平移、竞品借鉴、版本升级时使用参考对象对比法编写 PRD。挂载在 analysis 阶段，需要对比分析时由 req-prd 推荐激活。",
+      "description_len": 79,
+      "path": "skills-req/req-compare-plugin/skills/SKILL.md",
+      "digest": "3428e3f37ef9e9e8"
+    },
+    {
+      "name": "req-design",
+      "category": "req",
+      "version": "2.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "需求开发设计技能。用于 PRD 到开发设计的转换：API 契约定义、数据模型变更方案、变更文件清单、开发任务拆分、技术风险评估。当用户执行 /req doc 或需要编写开发设计文档时自动激活。",
+      "description_len": 96,
+      "path": "skills-req/req-design-plugin/skills/SKILL.md",
+      "digest": "8ec85838666b5cc5"
+    },
+    {
+      "name": "req-dev",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "需求开发计划编写技能。用于 PRD 到代码转换、代码库分析、开发任务拆分、技术方案设计。当用户执行 /req doc 或需要编写开发计划时自动激活。",
+      "description_len": 74,
+      "path": "skills-req/req-dev-plugin/skills/SKILL.md",
+      "digest": "3b0dd9396df3c236"
+    },
+    {
+      "name": "req-lookback",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "回归测试。部署后自动验证变更涉及的功能是否正常。可独立调用或由 /req done 自动触发。",
+      "description_len": 47,
+      "path": "skills-req/req-lookback-plugin/skills/SKILL.md",
+      "digest": "cc6bff614a8f7bf4"
+    },
+    {
+      "name": "req",
+      "category": "req",
+      "version": "2.1.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "需求工作流管理。用于 /req 命令使用、需求全生命周期管理（创建→PRD→评审→开发→测试→部署→归档）。当用户提到需求管理、/req 命令、REQ-XXX 相关任务时自动激活。",
+      "description_len": 90,
+      "path": "skills-req/req-plugin/skills/SKILL.md",
+      "digest": "bdb1fab4d3d8c3d2"
+    },
+    {
+      "name": "req-prd",
+      "category": "req",
+      "version": "2.3.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "产品设计与需求管理。用于 PRD、需求分析、用户故事、功能设计和原型规划，并将正式需求文档双写到本地仓库与 ai-proj Task Document。",
+      "description_len": 76,
+      "path": "skills-req/req-prd-plugin/skills/SKILL.md",
+      "digest": "0e5b16fcef3105bd"
+    },
+    {
+      "name": "req-prototype",
+      "category": "req",
+      "version": "2.2.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "原型生成与关联。生成可交互 HTML、上传到 ai-proj OSS 并关联 Requirement/PRD iframe，也支持 Stitch AI 视觉探索。",
+      "description_len": 81,
+      "path": "skills-req/req-prototype-plugin/skills/SKILL.md",
+      "digest": "7e362f4f895fb50a"
+    },
+    {
+      "name": "req-research",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "需求调研插件。代码审计、数据库分析、现有功能调研、技术可行性评估。挂载在 analysis 阶段，需要深度调研时由 req-prd 推荐激活。",
+      "description_len": 71,
+      "path": "skills-req/req-research-plugin/skills/SKILL.md",
+      "digest": "a94924f53bd4c80e"
+    },
+    {
+      "name": "req-retro",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "复盘总结。自动采集数据、计算质量评分、跨需求模式识别、技能自动进化。可独立调用或由 /req done 自动触发。",
+      "description_len": 57,
+      "path": "skills-req/req-retro-plugin/skills/SKILL.md",
+      "digest": "636b7099d85e9549"
+    },
+    {
+      "name": "req-review",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "PRD 评审方法论。用于需求评审、PRD 文档审查、评审意见编写。当执行 /req review 或需要评审 PRD 文档时使用。",
+      "description_len": 65,
+      "path": "skills-req/req-review-plugin/skills/SKILL.md",
+      "digest": "cd1a5181d72f35f4"
+    },
+    {
+      "name": "req-test-gate",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "测试与质量门禁制度。覆盖需求级测试(Gates 1-5)、部署级验证(Deploy Gates)、持续回归(Regression)、Harness Engineering 工程约束方法论（Ratchet、约定建立、门禁层级）。当用户提到质量门禁、架构约束、约定检查、ratchet、harness 相关任务时自动激活。",
+      "description_len": 159,
+      "path": "skills-req/req-test-gate-plugin/skills/SKILL.md",
+      "digest": "0eba060fa456885d"
+    },
+    {
+      "name": "req-workflow",
+      "category": "req",
+      "version": "1.0.0",
+      "status": "promoted",
+      "invocation": "model-invoked",
+      "description": "需求完整工作流。用于从创建到归档的完整流程、Hook 自动同步、测试环境流程。当需要了解需求完整生命周期或同步策略时使用。",
+      "description_len": 61,
+      "path": "skills-req/req-workflow-plugin/skills/SKILL.md",
+      "digest": "87499da3817d162d"
+    },
+    {
+      "name": "verification-before-completion",
+      "category": "req",
+      "version": "0.1.0",
+      "status": "beta",
+      "invocation": "model-invoked",
+      "description": "汇报「完成/修好/通过」或关闭任务、提交 PR、宣布 REQ 完成之前，必须先运行验证命令并附证据；没有证据不得声称完成。当准备说完成、关闭任务、或汇报交付前使用。",
+      "description_len": 82,
+      "path": "skills-req/verification-before-completion-plugin/skills/SKILL.md",
+      "digest": "7c23649f4a0a8798"
+    }
+  ]
+}
diff --git a/generate-marketplace.py b/generate-marketplace.py
index af0db19..65b4d29 100644
--- a/generate-marketplace.py
+++ b/generate-marketplace.py
@@ -96,6 +96,14 @@ def scan_plugins(directory, source_prefix, disabled):
         with open(manifest_path) as f:
             manifest = json.load(f)
 
+        # Lifecycle buckets: only promoted plugins ship in the public
+        # marketplace. beta/deprecated must be installed with --include-beta.
+        import os
+        status = manifest.get("status", "promoted")
+        if status in ("beta", "deprecated") and os.environ.get("ALLOW_BETA") != "1":
+            print(f"  ⊘ {plugin_name} (status={status}, not shipped)")
+            continue
+
         category, keywords = get_category_and_keywords(plugin_name)
 
         plugins.append({
diff --git a/install-skills.sh b/install-skills.sh
index 32d93c9..db3ecef 100755
--- a/install-skills.sh
+++ b/install-skills.sh
@@ -5,7 +5,11 @@
 #   ./install-skills.sh [options]
 #
 # Options:
-#   --agent <agent>     Install target: codex (default) or claude
+#   --agent <agent>     Install target: codex (default), dsh, or claude
+#   --link              Symlink skills from the repo instead of copying (dev machines)
+#   --audit             Show installed vs repo versions for the selected target
+#   --include-beta      Also install skills whose plugin.json status=beta
+#   --no-validate       Skip scripts/validate-skills.py before installing
 #   --dry-run           Preview changes without writing anything
 #   --category <cat>    Only install plugins in dir_category=<cat>
 #                       Valid values: biz, core, dev, integration, personal, req
@@ -30,6 +34,10 @@ FORCE=false
 CLEANUP=false
 LIST_ONLY=false
 INSTALL_ACTION=false
+LINK_MODE=false
+AUDIT_ONLY=false
+INCLUDE_BETA=false
+NO_VALIDATE=false
 
 # ── Colour helpers ─────────────────────────────────────────────────────────────
 GREEN='\033[0;32m'
@@ -48,9 +56,13 @@ dry()     { echo -e "${YELLOW}[dry]${RESET}  $*"; }
 while [[ $# -gt 0 ]]; do
   case "$1" in
     --agent)
-      [[ $# -ge 2 ]] || { error "--agent requires codex or claude"; exit 1; }
+      [[ $# -ge 2 ]] || { error "--agent requires codex, dsh or claude"; exit 1; }
       AGENT_TARGET="$2"; shift ;;
     --dry-run)   DRY_RUN=true ;;
+    --link)      LINK_MODE=true ;;
+    --audit)     AUDIT_ONLY=true ;;
+    --include-beta) INCLUDE_BETA=true ;;
+    --no-validate)  NO_VALIDATE=true ;;
     --force)     FORCE=true ;;
     --cleanup)   CLEANUP=true ;;
     --list)      LIST_ONLY=true ;;
@@ -77,13 +89,21 @@ case "$AGENT_TARGET" in
     SKILLS_DIR="${AI_PROJ_HELPER_SKILLS_DIR:-${HOME}/.agents/skills}"
     STATE_FILE="${AI_PROJ_HELPER_STATE_FILE:-${HOME}/.agents/.ai-proj-helper-installed-skills.json}"
     ;;
+  dsh)
+    if [[ -z "${DSH_HOME:-}" ]]; then
+      error "--agent dsh requires DSH_HOME (or set AI_PROJ_HELPER_SKILLS_DIR explicitly)"
+      exit 1
+    fi
+    SKILLS_DIR="${AI_PROJ_HELPER_SKILLS_DIR:-${DSH_HOME}/skills}"
+    STATE_FILE="${AI_PROJ_HELPER_STATE_FILE:-${DSH_HOME}/.ai-proj-helper-installed-skills.json}"
+    ;;
   claude)
     SKILLS_DIR="${AI_PROJ_HELPER_SKILLS_DIR:-${HOME}/.claude/skills}"
     COMMANDS_DIR="${AI_PROJ_HELPER_COMMANDS_DIR:-${HOME}/.claude/commands}"
     STATE_FILE="${AI_PROJ_HELPER_STATE_FILE:-${HOME}/.claude/.installed-skills.json}"
     ;;
   *)
-    error "Unsupported agent: $AGENT_TARGET (expected codex or claude)"
+    error "Unsupported agent: $AGENT_TARGET (expected codex, dsh or claude)"
     exit 1
     ;;
 esac
@@ -281,8 +301,19 @@ install_plugin() {
     effective_install_type="skill"
   fi
 
+  local status
+  status="$(read_field "$json_path" status)"
+  status="${status:-promoted}"
+
   if [[ "$LIST_ONLY" == true ]]; then
-    echo "  [$dir_category] $effective_install_type:$install_name  v$version"
+    echo "  [$dir_category] $status $effective_install_type:$install_name  v$version"
+    return
+  fi
+
+  if [[ "$status" == "deprecated" ]]; then
+    return
+  fi
+  if [[ "$status" == "beta" && "$INCLUDE_BETA" != true ]]; then
     return
   fi
 
@@ -305,6 +336,21 @@ install_plugin() {
     return
   fi
 
+  # Dev machines may symlink instead of copying: repository edits apply at once.
+  if [[ "$LINK_MODE" == true && "$effective_install_type" != "command" ]]; then
+    if [[ "$DRY_RUN" == true ]]; then
+      dry "$install_name → symlink $target_path → $src_dir"
+    else
+      mkdir -p "$SKILLS_DIR"
+      rm -rf "$target_path"
+      ln -sfn "$src_dir" "$target_path"
+      state_set "$install_name" "$version" "symlink" "$(content_digest "$src_dir")"
+      ok "$install_name → symlink (v$version)"
+    fi
+    INSTALL_ACTION=true
+    return
+  fi
+
   # A recorded content digest distinguishes repository updates from user edits.
   # Legacy state is adopted automatically only when the target is missing or
   # already identical to the repository source.
@@ -415,6 +461,38 @@ cleanup_removed() {
   done < <(state_all_names)
 }
 
+# ── Audit ──────────────────────────────────────────────────────────────────────
+audit_target() {
+  python3 - "$REPO_DIR" "$STATE_FILE" "$AGENT_TARGET" "$SKILLS_DIR" <<'PYEOF'
+import json, sys, pathlib
+repo, state_file, agent, skills_dir = sys.argv[1:5]
+root = pathlib.Path(repo)
+state = {}
+if pathlib.Path(state_file).is_file():
+    try: state = json.loads(pathlib.Path(state_file).read_text())
+    except Exception: state = {}
+rows = []
+for pj in sorted(root.glob('skills-*/*-plugin/.claude-plugin/plugin.json')):
+    try: meta = json.loads(pj.read_text())
+    except Exception: continue
+    name = meta.get('install_name')
+    if not name: continue
+    status = meta.get('status', 'promoted')
+    want = meta.get('version', '')
+    got = state.get(name, {}).get('version', '')
+    itype = state.get(name, {}).get('install_type', '')
+    drift = '' if got == want else ('missing' if not got else 'stale')
+    rows.append((name, status, want, got or '-', itype or '-', drift))
+drift_rows = [r for r in rows if r[5]]
+print(f'audit agent={agent} skills_dir={skills_dir}')
+print(f'{"name":34} {"status":10} {"repo":8} {"installed":9} {"mode":8} drift')
+for r in rows:
+    print(f'{r[0]:34} {r[1]:10} {r[2]:8} {r[3]:9} {r[4]:8} {r[5]}')
+print(f'\n{len(rows)} 个技能，{len(drift_rows)} 个漂移（未安装/版本落后）')
+raise SystemExit(1 if drift_rows else 0)
+PYEOF
+}
+
 # ── Main ───────────────────────────────────────────────────────────────────────
 main() {
   if [[ "$LIST_ONLY" == true ]]; then
@@ -429,7 +507,19 @@ main() {
     return
   fi
 
+  if [[ "$AUDIT_ONLY" == true ]]; then
+    audit_target
+    return
+  fi
+
   info "Installing skills for $AGENT_TARGET from: $REPO_DIR"
+  if [[ "$NO_VALIDATE" != true && -f "$REPO_DIR/scripts/validate-skills.py" ]]; then
+    if ! python3 "$REPO_DIR/scripts/validate-skills.py" >/tmp/ai-proj-helper-validate.log 2>&1; then
+      error "validate-skills.py 未通过，修复后再装（或加 --no-validate 跳过）"
+      tail -20 /tmp/ai-proj-helper-validate.log >&2 || true
+      exit 1
+    fi
+  fi
   [[ "$DRY_RUN" == true ]] && warn "DRY RUN — no files will be written"
   [[ -n "$CATEGORY_FILTER" ]] && info "Category filter: $CATEGORY_FILTER"
 
diff --git a/scripts/generate-catalog.py b/scripts/generate-catalog.py
new file mode 100755
index 0000000..ea8474a
--- /dev/null
+++ b/scripts/generate-catalog.py
@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+"""Generate catalog.json — the reviewable list of every skill in this repo.
+
+Usage: python3 scripts/generate-catalog.py [--check]
+  (default) write catalog.json
+  --check   fail if catalog.json is out of date (for CI)
+"""
+from __future__ import annotations
+
+import hashlib
+import json
+import re
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+CATALOG = ROOT / "catalog.json"
+
+
+def frontmatter(text: str) -> dict:
+    if not text.startswith("---"):
+        return {}
+    end = text.find("\n---", 3)
+    if end == -1:
+        return {}
+    data: dict = {}
+    metadata: dict = {}
+    key = None
+    for line in text[3:end].splitlines():
+        if not line.strip():
+            continue
+        m = re.match(r"^([A-Za-z0-9_-]+):\s*(.*)$", line)
+        if m:
+            key = m.group(1)
+            data[key] = m.group(2).strip().strip('"').strip("'")
+            continue
+        m = re.match(r"^\s+([A-Za-z0-9_-]+):\s*(.*)$", line)
+        if m and key == "metadata":
+            metadata[m.group(1)] = m.group(2).strip().strip('"').strip("'")
+    if metadata:
+        data["metadata"] = metadata
+    return data
+
+
+def build() -> dict:
+    skills = []
+    for plugin_json in sorted(ROOT.glob("skills-*/*-plugin/.claude-plugin/plugin.json")):
+        plugin_dir = plugin_json.parent.parent
+        meta = json.loads(plugin_json.read_text(encoding="utf-8"))
+        skill_files = sorted((plugin_dir / "skills").rglob("SKILL.md"))
+        if not skill_files:
+            continue
+        skill_md = skill_files[0]
+        fm = frontmatter(skill_md.read_text(encoding="utf-8", errors="replace"))
+        desc = fm.get("description", "")
+        body = skill_md.read_text(encoding="utf-8", errors="replace")
+        skills.append(
+            {
+                "name": meta.get("install_name") or fm.get("name"),
+                "category": meta.get("dir_category"),
+                "version": meta.get("version"),
+                "status": meta.get("status", "promoted"),
+                "invocation": (fm.get("metadata") or {}).get("invocation", "model-invoked"),
+                "description": desc,
+                "description_len": len(desc),
+                "path": str(skill_md.relative_to(ROOT)),
+                "digest": hashlib.sha256(body.encode("utf-8")).hexdigest()[:16],
+            }
+        )
+    counts: dict = {}
+    for skill in skills:
+        key = f"{skill['status']}/{skill['invocation']}"
+        counts[key] = counts.get(key, 0) + 1
+    return {
+        "repo": "ai-proj-helper",
+        "generated_at": datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+        "total": len(skills),
+        "counts": counts,
+        "skills": skills,
+    }
+
+
+def main() -> int:
+    catalog = build()
+    text = json.dumps(catalog, ensure_ascii=False, indent=2) + "\n"
+    if "--check" in sys.argv:
+        if not CATALOG.is_file():
+            print("catalog.json 不存在，请运行 python3 scripts/generate-catalog.py")
+            return 1
+        current = json.loads(CATALOG.read_text(encoding="utf-8"))
+        current.pop("generated_at", None)
+        fresh = dict(catalog)
+        fresh.pop("generated_at", None)
+        if current != fresh:
+            print("catalog.json 已过期，请重新生成")
+            return 1
+        print("catalog.json 是最新的")
+        return 0
+    CATALOG.write_text(text, encoding="utf-8")
+    print(f"写入 {CATALOG.relative_to(ROOT)}：{catalog['total']} 个技能 {catalog['counts']}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/scan-allowlist.txt b/scripts/scan-allowlist.txt
new file mode 100644
index 0000000..35f462c
--- /dev/null
+++ b/scripts/scan-allowlist.txt
@@ -0,0 +1,5 @@
+# 安全扫描允许清单（人工确认过的误报）
+# 格式：<路径子串>::<匹配文本中的特征串>
+# 例：某技能文档把危险命令当作“禁止清单”引用，而非真的要执行。
+skills-dev/slark-cicd-plugin::NEVER_AUTO
+skills-dev/slark-cicd-plugin::destructive
diff --git a/scripts/scan-skill.sh b/scripts/scan-skill.sh
new file mode 100755
index 0000000..deed20c
--- /dev/null
+++ b/scripts/scan-skill.sh
@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+# scan-skill.sh — security scan before a skill enters the team library.
+#
+# Usage: ./scripts/scan-skill.sh [plugin-dir ...]
+#   With no arguments scans every skills-*/*-plugin directory.
+# Exit 0 = clean (warnings allowed), 1 = high severity finding.
+set -uo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+HIGH=0
+WARN=0
+
+# High severity: destructive, remote-code-execution, secret material.
+HIGH_PATTERNS=(
+  'rm[[:space:]]+-[a-zA-Z]*r[a-zA-Z]*[[:space:]]+/($|[[:space:]])'
+  'curl[^|]*\|[[:space:]]*(ba|z)?sh'
+  'wget[^|]*\|[[:space:]]*(ba|z)?sh'
+  'chmod[[:space:]]+777'
+  'BEGIN (RSA|OPENSSH|EC|DSA) PRIVATE KEY'
+  'sk-[A-Za-z0-9]{16,}'
+  'AKIA[0-9A-Z]{16}'
+  'ghp_[A-Za-z0-9]{20,}'
+)
+# Warnings: machine-specific paths, plain http endpoints.
+WARN_PATTERNS=(
+  '/Users/[A-Za-z0-9._-]+/'
+  'http://[A-Za-z0-9.-]+'
+  'password[[:space:]]*=[[:space:]]*[^$]'
+)
+
+targets=("$@")
+if [[ ${#targets[@]} -eq 0 ]]; then
+  while IFS= read -r dir; do
+    targets+=("$(dirname "$(dirname "$dir")")")
+  done < <(find "$ROOT" -path "*/skills-*/*-plugin/.claude-plugin/plugin.json" | sort)
+fi
+
+ALLOWLIST="$ROOT/scripts/scan-allowlist.txt"
+is_allowlisted() {
+  # is_allowlisted <abs_path> <matched_line>
+  local path="$1" line="$2" entry sub needle
+  [[ -f "$ALLOWLIST" ]] || return 1
+  while IFS= read -r entry; do
+    [[ -z "$entry" || "$entry" == \#* ]] && continue
+    sub="${entry%%::*}"
+    needle="${entry#*::}"
+    if [[ "$path" == *"$sub"* && "$line" == *"$needle"* ]]; then
+      return 0
+    fi
+  done < "$ALLOWLIST"
+  return 1
+}
+
+for dir in "${targets[@]}"; do
+  rel="${dir#"$ROOT"/}"
+  for pattern in "${HIGH_PATTERNS[@]}"; do
+    while IFS= read -r hit; do
+      abs="${hit%%:*}"
+      detail="${hit#*:}"
+      if is_allowlisted "$abs" "$hit"; then
+        echo "[allow] $rel: $detail"
+        continue
+      fi
+      echo "[HIGH]  $rel: $detail"
+      HIGH=$((HIGH + 1))
+    done < <(grep -rnE --include='*.md' --include='*.sh' --include='*.py' --include='*.js' --include='*.ts' --include='*.json' -- "$pattern" "$dir" 2>/dev/null | cut -c1-200)
+  done
+  for pattern in "${WARN_PATTERNS[@]}"; do
+    while IFS= read -r hit; do
+      echo "[warn]  $rel: ${hit%%:*}:${hit#*:}"
+      WARN=$((WARN + 1))
+    done < <(grep -rnE --include='*.md' --include='*.sh' --include='*.py' --include='*.js' --include='*.ts' -- "$pattern" "$dir" 2>/dev/null | cut -c1-200)
+  done
+done
+
+echo ""
+echo "扫描完成：${#targets[@]} 个插件，$HIGH 个高危，$WARN 个警告"
+if [[ "$HIGH" -gt 0 ]]; then
+  echo "存在高危项，禁止入库，除非逐条确认并记录原因。"
+  exit 1
+fi
+exit 0
diff --git a/scripts/validate-skills.py b/scripts/validate-skills.py
new file mode 100755
index 0000000..803ca96
--- /dev/null
+++ b/scripts/validate-skills.py
@@ -0,0 +1,161 @@
+#!/usr/bin/env python3
+"""Validate plugins and SKILL.md files against the Agent Skills spec + team rules.
+
+Usage: python3 scripts/validate-skills.py [--strict]
+Exit 0 = pass, 1 = errors found.
+"""
+from __future__ import annotations
+
+import json
+import re
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+ALLOWED_FRONTMATTER = {
+    "name", "description", "license", "compatibility", "metadata",
+    "allowed-tools", "disable-model-invocation", "version",
+}
+NAME_RE = re.compile(r"^[a-z0-9][a-z0-9-]{0,63}$")
+BODY_MAX_BYTES = 20 * 1024
+DESC_MIN, DESC_MAX = 30, 1024
+STATUSES = {"promoted", "beta", "deprecated"}
+INVOCATIONS = {"model-invoked", "user-invoked"}
+REQUIRED_PLUGIN_FIELDS = ("install_name", "install_type", "dir_category", "version")
+
+errors: list[str] = []
+warnings: list[str] = []
+
+
+def parse_frontmatter(text: str):
+    """Return (frontmatter dict, body text). Supports one level of nesting."""
+    if not text.startswith("---"):
+        return {}, text
+    end = text.find("\n---", 3)
+    if end == -1:
+        return {}, text
+    raw, body = text[3:end], text[end + 4:]
+    data: dict = {}
+    metadata: dict = {}
+    key = None
+    for line in raw.splitlines():
+        if not line.strip():
+            continue
+        m = re.match(r"^([A-Za-z0-9_-]+):\s*(.*)$", line)
+        if m:
+            key = m.group(1)
+            data[key] = m.group(2).strip().strip('"').strip("'")
+            continue
+        m = re.match(r"^\s+([A-Za-z0-9_-]+):\s*(.*)$", line)
+        if m and key == "metadata":
+            metadata[m.group(1)] = m.group(2).strip().strip('"').strip("'")
+        elif key:
+            data[key] = (data.get(key, "") + " " + line.strip()).strip()
+    if metadata:
+        data["metadata"] = metadata
+    return data, body
+
+
+def check_plugin(plugin_json: Path):
+    plugin_dir = plugin_json.parent.parent
+    rel = plugin_dir.relative_to(ROOT)
+    try:
+        meta = json.loads(plugin_json.read_text(encoding="utf-8"))
+    except Exception as exc:  # noqa: BLE001
+        errors.append(f"{rel}: plugin.json 解析失败: {exc}")
+        return
+
+    for field in REQUIRED_PLUGIN_FIELDS:
+        if not meta.get(field):
+            errors.append(f"{rel}: plugin.json 缺少 {field}")
+    status = meta.get("status", "")
+    if status and status not in STATUSES:
+        errors.append(f"{rel}: status 非法 {status!r}（应为 {sorted(STATUSES)}）")
+    install_name = meta.get("install_name", "")
+    if install_name and not NAME_RE.match(install_name):
+        errors.append(f"{rel}: install_name 不合规 {install_name!r}")
+
+    skills_dir = plugin_dir / "skills"
+    skill_files = sorted(skills_dir.rglob("SKILL.md")) if skills_dir.is_dir() else []
+    if not skill_files:
+        errors.append(f"{rel}: 没有找到 skills/**/SKILL.md")
+        return
+    for skill_md in skill_files:
+        check_skill(rel, skill_md, meta)
+
+
+def check_skill(rel: Path, skill_md: Path, meta: dict):
+    text = skill_md.read_text(encoding="utf-8", errors="replace")
+    fm, body = parse_frontmatter(text)
+    where = f"{rel}/{skill_md.name}"
+
+    name = fm.get("name", "")
+    if not name:
+        errors.append(f"{where}: frontmatter 缺少 name")
+    elif not NAME_RE.match(name):
+        errors.append(f"{where}: name 不合规 {name!r}（小写字母/数字/连字符，≤64）")
+
+    desc = fm.get("description", "")
+    if not desc:
+        errors.append(f"{where}: frontmatter 缺少 description")
+    elif not (DESC_MIN <= len(desc) <= DESC_MAX):
+        warnings.append(f"{where}: description 长度 {len(desc)}，建议 {DESC_MIN}-{DESC_MAX}（决定是否被触发）")
+
+    for key in fm:
+        if key == "metadata":
+            continue
+        if key not in ALLOWED_FRONTMATTER:
+            warnings.append(f"{where}: 非标准 frontmatter 字段 {key!r}，建议移入 metadata")
+
+    body_bytes = len(body.encode("utf-8"))
+    if body_bytes > BODY_MAX_BYTES:
+        warnings.append(f"{where}: 正文 {body_bytes} 字节，超过 20KB 预算，建议拆 references/")
+
+    metadata = fm.get("metadata") or {}
+    invocation = metadata.get("invocation", "")
+    if invocation and invocation not in INVOCATIONS:
+        errors.append(f"{where}: metadata.invocation 非法 {invocation!r}")
+    disable_model = str(fm.get("disable-model-invocation", "")).lower() == "true"
+    if invocation == "user-invoked" and not disable_model:
+        errors.append(f"{where}: user-invoked 技能必须设置 disable-model-invocation: true")
+    if invocation == "model-invoked" and disable_model:
+        errors.append(f"{where}: model-invoked 技能不应设置 disable-model-invocation: true")
+
+    status = meta.get("status", "")
+    if status == "promoted":
+        docs = ROOT / "docs" / "skills" / f"{meta.get('install_name', name)}.md"
+        if not docs.is_file():
+            warnings.append(f"{where}: promoted 技能缺少四段式文档 docs/skills/{docs.name}")
+
+    # Relative links must resolve.
+    for target in re.findall(r"\]\((\.[^)\s]+|[A-Za-z0-9_-]+/[^)\s]+)\)", body):
+        if target.startswith(("http://", "https://", "#")):
+            continue
+        if not (skill_md.parent / target).exists():
+            warnings.append(f"{where}: 引用的相对路径不存在: {target}")
+
+
+def main() -> int:
+    strict = "--strict" in sys.argv
+    plugins = sorted(ROOT.glob("skills-*/*-plugin/.claude-plugin/plugin.json"))
+    if not plugins:
+        print("未找到任何 plugin.json", file=sys.stderr)
+        return 1
+    for plugin_json in plugins:
+        check_plugin(plugin_json)
+
+    for warning in warnings:
+        print(f"[warn]  {warning}")
+    for error in errors:
+        print(f"[error] {error}")
+    print(f"\n检查 {len(plugins)} 个插件：{len(errors)} 个错误，{len(warnings)} 个警告")
+    if errors:
+        return 1
+    if strict and warnings:
+        print("--strict 模式下，警告视为失败")
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/skills-core/ai-proj-helper-entry-plugin/.claude-plugin/plugin.json b/skills-core/ai-proj-helper-entry-plugin/.claude-plugin/plugin.json
new file mode 100644
index 0000000..ddd6c96
--- /dev/null
+++ b/skills-core/ai-proj-helper-entry-plugin/.claude-plugin/plugin.json
@@ -0,0 +1,13 @@
+{
+  "name": "ai-proj-helper-entry-plugin",
+  "description": "技能库路由：列出技能、什么任务用哪个、交付终点口径。由人主动调用，不占模型目录 token。",
+  "version": "0.1.0",
+  "author": {
+    "name": "qiudl"
+  },
+  "install_name": "ai-proj-helper-entry",
+  "install_type": "skill",
+  "dir_category": "core",
+  "status": "beta",
+  "invocation": "user-invoked"
+}
diff --git a/skills-core/ai-proj-helper-entry-plugin/skills/SKILL.md b/skills-core/ai-proj-helper-entry-plugin/skills/SKILL.md
new file mode 100644
index 0000000..3f7c491
--- /dev/null
+++ b/skills-core/ai-proj-helper-entry-plugin/skills/SKILL.md
@@ -0,0 +1,77 @@
+---
+name: ai-proj-helper-entry
+description: ai-proj-helper 技能库路由：告诉你有哪些技能、什么任务用哪个、以及「REQ released + PDV 通过才算交付」的强制口径。人在开始需求/开发/评审/发布任务前主动调用。
+metadata:
+  invocation: user-invoked
+  owner: qiudl
+disable-model-invocation: true
+---
+
+# ai-proj-helper 入口（路由器）
+
+## 0. 这个技能是什么
+
+一个**纯人触发**的路由器：人问「该用哪个技能 / 从哪开始」，就读这里。它不进模型目录，所以不占每次会话的目录 token；代价是必须由人主动喊。
+
+> 路由规则：**任何技能增、删、改名、改行为后，必须回改本文件**。一个没提到新技能、或还指向已删技能的路由器，就是假的路由器（见 `skill-creator` §9）。
+
+## 1. 交付终点（强制，优先于路由）
+
+开发任务的交付终点是：需求按 REQ 流程推进到具备发布条件并完成上线（`delivery_stage=released` + PDV 验收通过）。
+
+- 写代码、过单测、提 PR、过 internal gate、合并代码，都是过程节点，**不得当作交付终点**；
+- 任务 `completed` ≠ 需求完成；REQ 未 `released` 不得报完成；
+- 汇报格式：**当前 REQ 阶段 + 下一个要做的节点**；
+- 只有必须由人拍板/授权的阻塞（生产发布授权、密钥权限、需求变更、无法自动修复的失败）才停下来问，并写清卡点与下一步。
+
+## 2. 主流程：需求 → 上线
+
+```
+讨论(req) → PRD(req-prd) → 评审(req-review) → 设计(req-design)
+→ 开发(dev-coding / karpathy-guidelines) → 测试(dev-test + req-test-gate)
+→ 代码评审(dev-review) → 发布(publish / slark-cicd / slark-dsh-cicd) → PDV 验收 → released
+```
+
+| 阶段 | 技能 |
+| --- | --- |
+| 讨论/调研 | `req`、`req-research`、`req-compare` |
+| PRD | `req-prd`、`req-prototype` |
+| 评审 | `req-review` |
+| 设计/拆任务 | `req-design`、`dev-arch`、`db-migration` |
+| 编码 | `dev-coding`、`karpathy-guidelines` |
+| 测试 | `dev-test`、`req-test-gate` |
+| 代码评审 | `dev-review`、`review-checklist` |
+| 发布 | `publish`、`slark-cicd`、`slark-dsh-cicd`、`deploy-rollback` |
+| 部署后 | `req-audit`、`req-lookback` |
+| 复盘 | `req-retro` |
+| 项目管理 | `project-management`、`pm-risk`、`pm-ask` |
+| 建/改技能 | `skill-creator` |
+| 汇报完成前 | `verification-before-completion` |
+
+## 3. 按任务类型的旁路
+
+- 「这个 bug 很难查」→ `diagnosing-bugs`（先建红循环再推理）
+- 「帮我审这个 PR」→ `dev-review`
+- 「发版」→ 对应 `*-cicd` runbook，不是通用 `dev-deploy`
+- 「不确定该用哪个」→ 读到这里就够了，别猜
+
+## 4. 红线
+
+- 不得自审批（提交人不能批自己的 REQ）
+- 不得绕过门禁（测试 Gate、发布 Gate、归档 Gate）
+- 不得把「某步骤完成」当结束
+- 不得无证据声称完成（见 `verification-before-completion`）
+- 不得把凭据/内网地址/绝对路径写进技能或文档
+
+## 5. 技能从哪来
+
+- 用户级：`~/.agents/skills/`（Codex）、`$DSH_HOME/skills/`（DSH）
+- 项目级：项目根 `.agents/skills/`
+- **真相源**：`ai-proj-helper` 仓库；本机目录是安装产物，不要直接改
+- 个人技能：`skills-personal/`（不入库、不发布）
+
+## 6. 维护检查
+
+- `./install-skills.sh --audit` 漂移数 > 0 → 先同步再干活
+- 技能有增删改 → 本文件 + `catalog.json` 必须同 PR 回改
+- 每季度人工过一遍本路由，删掉不再用的技能引用
diff --git a/skills-core/ai-proj-helper-entry-plugin/skills/evals/evals.json b/skills-core/ai-proj-helper-entry-plugin/skills/evals/evals.json
new file mode 100644
index 0000000..198478a
--- /dev/null
+++ b/skills-core/ai-proj-helper-entry-plugin/skills/evals/evals.json
@@ -0,0 +1,24 @@
+{
+  "skill_name": "ai-proj-helper-entry",
+  "skill_version": "0.1.0",
+  "evals": [
+    {
+      "id": 1,
+      "kind": "happy_path",
+      "prompt": "我要给 qiu-slark 做个新功能，从哪开始",
+      "expected_output": "给出 req → req-prd → req-review → req-design 的路径，并声明交付终点是 released + PDV，不是写完代码"
+    },
+    {
+      "id": 2,
+      "kind": "edge_case",
+      "prompt": "我已经把单测跑过了，这个需求可以算完成了吗",
+      "expected_output": "否定：单测只是过程节点，REQ 未 released 不算完成；给出当前阶段与下一节点"
+    },
+    {
+      "id": 3,
+      "kind": "failure_recovery",
+      "prompt": "你说的那个技能我机器上没装，怎么办",
+      "expected_output": "说明技能真相源在 ai-proj-helper 仓库、本机目录是安装产物；给出 install-skills.sh 安装/审计命令"
+    }
+  ]
+}
diff --git a/skills-core/skill-creator-plugin/.claude-plugin/plugin.json b/skills-core/skill-creator-plugin/.claude-plugin/plugin.json
new file mode 100644
index 0000000..159c420
--- /dev/null
+++ b/skills-core/skill-creator-plugin/.claude-plugin/plugin.json
@@ -0,0 +1,13 @@
+{
+  "name": "skill-creator-plugin",
+  "description": "团队技能创建与维护的标准流程：收录门槛、分桶生命周期、调用模式、evals、四段式文档、路由回改。",
+  "version": "0.1.0",
+  "author": {
+    "name": "qiudl"
+  },
+  "install_name": "skill-creator",
+  "install_type": "skill",
+  "dir_category": "core",
+  "status": "beta",
+  "invocation": "user-invoked"
+}
diff --git a/skills-core/skill-creator-plugin/skills/SKILL.md b/skills-core/skill-creator-plugin/skills/SKILL.md
new file mode 100644
index 0000000..e32e763
--- /dev/null
+++ b/skills-core/skill-creator-plugin/skills/SKILL.md
@@ -0,0 +1,127 @@
+---
+name: skill-creator
+description: 团队技能的标准流程：收录门槛、分桶生命周期、调用模式、SKILL.md 规范、evals 全量、四段式文档、路由回改。当要新建技能、大改现有技能、或评审一个技能是否合格时使用。
+metadata:
+  invocation: user-invoked
+  owner: qiudl
+disable-model-invocation: true
+---
+
+# 技能创建（团队标准流程）
+
+## 0. 一句话
+
+先有真实失败，再写技能；没有 evals 的技能不算完成；**技能目录是预算，不是收藏夹**。
+
+## 1. 何时用
+
+- 新建技能
+- 大改现有技能（改 description、改核心流程、改门禁）
+- 评审别人提上来的技能
+- 判断某个想法「该不该做成技能」
+
+何时不用：只改错别字/格式 → 直接改；一次性个人偏好 → 放个人技能，不进团队。
+
+## 2. 收录门槛（先过这一关，再谈怎么写）
+
+新技能**默认拒绝**，只有同时满足两条才立项：
+
+1. **真实失败**：能说出在哪个真实任务/会话里出了问题——跑了什么、期望什么、实际什么。假想改进（「这样会更好」）不算。
+2. **组合不出来**：用现有技能组合做不到；能组合的，就改现有技能，不新增。
+
+被拒的想法记进仓库 `.out-of-scope/<概念>.md`，写清理由，避免下次重复讨论。
+
+## 3. 分桶与生命周期
+
+| 桶 | 含义 | 是否安装 | 是否进 README / plugin.json |
+| --- | --- | --- | --- |
+| `promoted` | 已验证、日常在用 | 默认安装 | 必须出现 |
+| `beta` | 公开试跑，收反馈 | 需 `--include-beta` | 不进默认清单 |
+| `deprecated` | 不再用 | 不安装 | 留文档，标弃用 |
+
+- 新技能一律先进 `beta`，跑够真实使用 + evals 通过才 `promoted`。
+- 只有 `promoted` 能进模型目录；beta 不占目录预算。
+
+## 4. 调用模式（省目录 token 的关键）
+
+每个技能二选一，写进 frontmatter：
+
+- **`model-invoked`**（默认）：进模型目录，agent 自己判断何时用。
+- **`user-invoked`**：只由人下命令触发，**不进模型目录**；必须同时写 `disable-model-invocation: true` 和 `metadata.invocation: user-invoked`。
+
+判定口径：这件事是不是「人下命令才开始」？是 → user-invoked。典型：publish、retro、triage、发布 runbook、建技能。
+
+## 5. SKILL.md 硬约束（Agent Skills 规范）
+
+- 目录：`SKILL.md` + 可选 `scripts/`、`references/`、`assets/`
+- frontmatter 必填：`name`（小写字母/数字/连字符，≤64）、`description`（≤1024，写清「做什么 + 何时用」）
+- 可选：`license`、`compatibility`、`metadata`、`allowed-tools`
+- **非标准字段**一律进 `metadata`
+- 正文 ≤20KB（约 5000 token），细节拆 `references/`
+- 只写 agent 不知道的东西：项目约定、边界、坑；不写通用常识
+
+## 6. 正文结构模板
+
+```
+# 名称
+## 0. 一句话
+## 1. 何时用 / 何时不用
+## 2. 步骤（可执行、有顺序）
+## 3. 硬约束与红线
+## 4. 常见坑（gotchas）
+## 5. 验证清单
+## 6. 参考（references/ 链接）
+```
+
+## 7. evals（强制、全量）
+
+- 路径：`evals/evals.json`；至少 3 条：正常路径 / 边界输入 / 失败恢复
+- 跑「有技能 / 无技能」对比，结果存 `iteration-<N>/`，按 rubric 打分，PR 附结果
+
+## 8. 四段式文档（每个 promoted 技能一篇）
+
+`docs/skills/<name>.md`，四节固定顺序：
+
+1. **What it does**
+2. **When to reach for it**
+3. **Common questions**
+4. **It's working if**（必须给可验证证据，不许写「感觉对了」）
+
+beta 不写文档；弃用保留文档并标 archived。
+
+## 9. 路由回改（硬规则）
+
+所有技能增、删、改名、改行为后，**必须同步回改入口技能 `ai-proj-helper-entry` 和 `catalog.json`**。一个没提到新技能、或还在路由到已删技能的路由器，就是假的路由器。
+
+## 10. 质量门禁（提 PR 前逐条打勾）
+
+- [ ] 过了收录门槛（真实失败 + 组合不出来）
+- [ ] 桶与调用模式已标，`beta` 不得出现在默认清单
+- [ ] `name`/`description` 合规，description 含「何时用」
+- [ ] 正文 ≤20KB，细节已拆 references
+- [ ] evals ≥3 条，有「有/无技能」对比
+- [ ] 四段式文档已写（promoted 才需要）
+- [ ] 入口路由与 `catalog.json` 已回改
+- [ ] `plugin.json` 已登记 `install_name`/`install_type`/`dir_category`/`status`/版本号
+- [ ] `python3 scripts/validate-skills.py` 通过
+- [ ] `./scripts/scan-skill.sh` 无高危项
+- [ ] 版本号与 CHANGELOG 已更新
+- [ ] 无绝对路径、无凭据、无内网地址
+
+## 11. 反模式
+
+- 一次写 800 行「大全」，什么都想覆盖
+- description 写成「各种任务都能用」
+- 没有 evals 就提 PR；没有四段式文档就 promoted
+- 直接复制外部技能，不改写成我们的 REQ/ai-proj 语境
+- 把临时排障步骤固化成团队技能（这类放 personal）
+- 技能改名后忘了回改路由
+
+## 12. 评审别人技能时看什么
+
+1. 收录门槛过没过（有没有真实失败）
+2. description 能不能让模型判断「何时用」
+3. 桶与调用模式是否正确（该 user-invoked 的别占目录）
+4. 和现有技能是否重复（先查 `catalog.json`）
+5. evals 是否覆盖失败路径、文档是否有「It's working if」
+6. 是否触碰红线（自审批、绕过门禁、绝对路径、凭据）
diff --git a/skills-core/skill-creator-plugin/skills/evals/evals.json b/skills-core/skill-creator-plugin/skills/evals/evals.json
new file mode 100644
index 0000000..ee3b94b
--- /dev/null
+++ b/skills-core/skill-creator-plugin/skills/evals/evals.json
@@ -0,0 +1,24 @@
+{
+  "skill_name": "skill-creator",
+  "skill_version": "0.1.0",
+  "evals": [
+    {
+      "id": 1,
+      "kind": "happy_path",
+      "prompt": "我想加一个新技能，用来在每次部署后自动跑一遍冒烟",
+      "expected_output": "先问真实失败场景（哪次部署出过问题、跑了什么、期望什么），再判断现有 req-lookback/req-audit 能否组合；给出 bucket/调用模式/evals 计划"
+    },
+    {
+      "id": 2,
+      "kind": "edge_case",
+      "prompt": "直接帮我把这个想法写成 SKILL.md：一个能自动优化代码风格的技能",
+      "expected_output": "拒绝直接写，指出没有真实失败证据，要求先给案例；若无证据则建议放个人技能"
+    },
+    {
+      "id": 3,
+      "kind": "failure_recovery",
+      "prompt": "技能有 evals 但 validate-skills.py 报了 name 不合规，怎么处理",
+      "expected_output": "说明 name 规则（小写/数字/连字符、≤64），给出改名步骤，并提示要同步 install_name、catalog、入口路由"
+    }
+  ]
+}
diff --git a/skills-dev/agent-swarm-plugin/skills/SKILL.md b/skills-dev/agent-swarm-plugin/skills/SKILL.md
index 52c8f61..5a586ad 100644
--- a/skills-dev/agent-swarm-plugin/skills/SKILL.md
+++ b/skills-dev/agent-swarm-plugin/skills/SKILL.md
@@ -404,3 +404,14 @@ ssh melbourne "cd /Users/coolbuy-dev/coding/new-ai-proj && \
 - [OpenAI Swarm 文档](https://github.com/openai/swarm)
 - [Multi-Agent Systems 设计模式](https://arxiv.org/abs/2308.00352)
 - [Claude Code Skills 文档](https://docs.anthropic.com/claude-code/skills)
+
+## 并行编排流程
+
+把大任务拆给多个子 agent 时按这个顺序：
+
+1. **先划写入边界**：每个子 agent 拥有互不重叠的目录/文件；共享文件（锁文件、入口文件）由编排者最后统一改。
+2. **子 agent 必须新上下文**：不要 fork 父上下文，否则它继承了本该甩掉的包袱。
+3. **每个子任务给验收标准 + 回报格式**：三句话（结果 / 证据路径 / 下一步）。
+4. **子 agent 不做合并与发布**：只交分支或补丁，合并由编排者做。
+5. **汇总时先跑集成验证**，再回报；冲突由编排者解，不让子 agent 互改。
+6. 两个子 agent 不得写同一文件；不确定时先串行。
diff --git a/skills-dev/dev-review-plugin/skills/SKILL.md b/skills-dev/dev-review-plugin/skills/SKILL.md
index 6611403..ecbef4b 100644
--- a/skills-dev/dev-review-plugin/skills/SKILL.md
+++ b/skills-dev/dev-review-plugin/skills/SKILL.md
@@ -303,3 +303,13 @@ mcp__ai-proj__link_tasks_to_requirement({
 4. **严重度准确** — 不要所有问题都标 High，按实际影响分级
 5. **建议可操作** — 每个发现必须附带具体修复建议
 6. **关注变更** — 评审范围是 diff，不要对未变更的代码提意见（除非变更引入了对旧代码的新风险）
+
+## 接收与响应评审意见
+
+评审不是单向输出，收到意见后按这套来：
+
+1. **先分类**：阻断项（必须修）/ 建议项（可讨论）/ 澄清项（问题理解不一致）。
+2. **先复述再回应**：用自己的话重述意见，确认理解一致，再给证据或修改。
+3. **不辩解**：用证据说话（代码、测试、日志），不用语气词反驳。
+4. **逐条回评**：每条写「已修（commit）/ 不修（理由）/ 需澄清（具体问题）」，不得默默跳过。
+5. **阻断项没修完不得合并**；建议项不修要记录原因，进 PR 描述。
diff --git a/skills-dev/dev-test-plugin/skills/dev-test/SKILL.md b/skills-dev/dev-test-plugin/skills/dev-test/SKILL.md
index bf8ad79..aeb8778 100644
--- a/skills-dev/dev-test-plugin/skills/dev-test/SKILL.md
+++ b/skills-dev/dev-test-plugin/skills/dev-test/SKILL.md
@@ -215,3 +215,11 @@ AI 根据需求变更范围动态生成 Playwright spec，模板见 `templates/p
 若 git diff 显示同一功能的 `handlers/`、`middleware/`、`routes/` 文件均有变更（通过命名模式识别，如 `impersonation_handler.go` + `impersonation_middleware.go` + `impersonation_routes.go`），则除单元测试外**额外生成** `backend/tests/` 下的集成测试。
 
 识别方式：提取文件名中的功能前缀（如 `impersonation`），若在三个目录中均出现，则触发集成测试生成。
+
+## TDD 循环（红 → 绿 → 重构）
+
+1. **一次只做一个行为**：写一个小测试，先确认它**失败**，并确认失败原因就是你要修的那件事（不是编译错误/环境问题）。
+2. **最小实现让它变绿**，不做顺手重构。
+3. **绿了再重构**，重构期间测试保持绿。
+4. 禁止先写实现再补测试；补的测试要能证明它真的会红（临时改坏一次验证）。
+5. 交付时附：测试命令、通过/失败数、以及「曾经红过」的证据（日志或截图记录）。
diff --git a/skills-req/req-design-plugin/skills/SKILL.md b/skills-req/req-design-plugin/skills/SKILL.md
index bc24be2..3895cc8 100644
--- a/skills-req/req-design-plugin/skills/SKILL.md
+++ b/skills-req/req-design-plugin/skills/SKILL.md
@@ -474,3 +474,13 @@ PRD 说「做什么」（用户故事、业务规则、验收标准），设计
 | V1.0 | 2026-01-26 | 初始版本（原名 req-dev） |
 | V2.0 | 2026-04-06 | 重构为 req-design：移除架构指南和编码规范，聚焦 API 契约 + 任务拆分 |
 | V2.1 | 2026-04-16 | 新增 Memory 隔离规则（REQ-20260416-0017） |
+
+## 任务依赖边（blocking edges）
+
+拆子任务时，每个任务必须显式声明依赖，而不是只给层级：
+
+- 每个任务写 `blockers: [任务编号...]`，没有依赖就写 `blockers: []`；
+- **被阻塞的任务不得先开工**，即使它看起来更简单；
+- 依赖图落进需求的任务描述，供 ai-proj / PM 检查；
+- 出现循环依赖直接拒绝拆分方案，重新设计边界；
+- 验收时检查：所有 blockers 是否 completed，再开自己。
diff --git a/skills-req/verification-before-completion-plugin/.claude-plugin/plugin.json b/skills-req/verification-before-completion-plugin/.claude-plugin/plugin.json
new file mode 100644
index 0000000..bbd50f6
--- /dev/null
+++ b/skills-req/verification-before-completion-plugin/.claude-plugin/plugin.json
@@ -0,0 +1,13 @@
+{
+  "name": "verification-before-completion-plugin",
+  "description": "汇报完成/关闭任务/提交 PR 前必须先跑验证并附证据；没有证据不得声称完成。",
+  "version": "0.1.0",
+  "author": {
+    "name": "qiudl"
+  },
+  "install_name": "verification-before-completion",
+  "install_type": "skill",
+  "dir_category": "req",
+  "status": "beta",
+  "invocation": "model-invoked"
+}
diff --git a/skills-req/verification-before-completion-plugin/skills/SKILL.md b/skills-req/verification-before-completion-plugin/skills/SKILL.md
new file mode 100644
index 0000000..a808c35
--- /dev/null
+++ b/skills-req/verification-before-completion-plugin/skills/SKILL.md
@@ -0,0 +1,55 @@
+---
+name: verification-before-completion
+description: 汇报「完成/修好/通过」或关闭任务、提交 PR、宣布 REQ 完成之前，必须先运行验证命令并附证据；没有证据不得声称完成。当准备说完成、关闭任务、或汇报交付前使用。
+metadata:
+  invocation: model-invoked
+  owner: qiudl
+---
+
+# 完成前验证
+
+## 0. 一句话
+
+**没有证据的「完成」不算完成。**
+
+## 1. 何时用
+
+- 准备说「搞定了 / 修好了 / 通过了 / 可以发了」
+- 准备把任务标 `completed`、关 issue、提 PR
+- 准备宣布 REQ 完成或可发布
+- 收到「你确认一下」这类要求时
+
+何时不用：还在探索阶段、明确记录的是中间发现（这时要写「未验证」而不是「完成」）。
+
+## 2. 证据三件套
+
+1. **命令**：原样可复制，别人能重跑
+2. **输出关键行**：可读文本，不要只看截图
+3. **判定**：退出码 / 通过数 / 失败数
+
+格式：`验证：<命令> → <结果>（通过 N / 失败 M），证据：<日志路径或输出片段>`
+
+## 3. 按类型的验证清单
+
+| 改动类型 | 必须验证 |
+| --- | --- |
+| 代码改动 | 构建 + 单测 + 类型检查（有则跑） |
+| bug 修复 | 先复现（证明坏）→ 修复 → 同一用例再跑（证明好） |
+| 需求级交付 | 验收标准逐条打勾 + PDV 验收通过 |
+| 发布/部署 | 健康检查 + E2E + 回滚预案可用 |
+| 配置/文档 | 生效验证（如 `--dump-config` 结果、命令输出） |
+| 数据变更 | 变更前后数据抽查 + 回滚脚本验证 |
+
+## 4. 反模式
+
+- 「应该可以」「本地没问题」「理论上过了」
+- 只贴成功截图，不贴命令
+- 跳过失败用例不说明
+- 把「编译通过」当成「功能可用」
+- 把「测试跑了一部分」说成「全部通过」
+
+## 5. 与 ai-proj 对接
+
+- 任务标 `completed` 前，把证据挂到该任务文档；
+- REQ 未 `released` 时不得报需求完成（见 `ai-proj-helper-entry` 的交付终点）；
+- 证据缺失时，正确做法是状态写 `in_progress` 并说明「还差哪项验证」。
diff --git a/skills-req/verification-before-completion-plugin/skills/evals/evals.json b/skills-req/verification-before-completion-plugin/skills/evals/evals.json
new file mode 100644
index 0000000..a70fddf
--- /dev/null
+++ b/skills-req/verification-before-completion-plugin/skills/evals/evals.json
@@ -0,0 +1,24 @@
+{
+  "skill_name": "verification-before-completion",
+  "skill_version": "0.1.0",
+  "evals": [
+    {
+      "id": 1,
+      "kind": "happy_path",
+      "prompt": "我改完了登录接口，可以提交了吗",
+      "expected_output": "要求给出构建/单测命令与结果、失败数；附证据路径后再提交"
+    },
+    {
+      "id": 2,
+      "kind": "edge_case",
+      "prompt": "测试环境没有依赖装不上，先合了吧，本地我手动看过没问题",
+      "expected_output": "拒绝以「手动看过」作为证据；说明缺失依赖要记录为未验证项，状态保持 in_progress"
+    },
+    {
+      "id": 3,
+      "kind": "failure_recovery",
+      "prompt": "测试跑过了 10 个，还有 2 个没跑，能说全部通过吗",
+      "expected_output": "不能：必须写明通过 10 / 未跑 2 及原因，不得声称全部通过"
+    }
+  ]
+}