harnessed 1.0.0

package/routing/SCHEMA.md

@@ -0,0 +1,199 @@
+# routing/ — B+C Routing Frontmatter SSOT (v0.1 placeholder spec)
+
+> **决策来源**：[`PROJECT-SPEC.md` § 9](../PROJECT-SPEC.md)
+> **机器消费 artifact**：**未实现**（v0.1 phase 1.1 仅文档化字段；JSON Schema artifact + 真实 routing/*.md 由 v0.1 phase 1.4 落地）
+> **本文件定位**：B+C 混合路由方案的 yaml frontmatter 标准定义。B 层（skill description）与 C 层（hook script）共享**同一份 frontmatter** —— 改一处两层同步生效。
+
+---
+
+## 1. SSOT 模式（为什么 B 层 + C 层共享同一 frontmatter）
+
+`routing/<topic>.md` 是一份普通 markdown，但 frontmatter 同时被两层消费：
+
+| 层 | 读 frontmatter 哪块 | 干什么 |
+|---|---|---|
+| **B 层（skill description）** | `soft_hint.description_template` | 生成 skill 的 1 行描述，让 LLM 在 skill 路由时能读到"标准化、数据驱动的 UI 设计；优先于 frontend-design"这种 hint |
+| **C 层（hook script）** | `trigger.keywords` + `trigger.file_globs` + `hard_route.*` | hook 在用户输入触发关键词或匹配文件 glob 时强制路由到 `hard_route.primary` skill；override 条件支持差异化 |
+
+**好处**：
+1. 改一处两层同步生效（避免双轨维护腐烂）
+2. CI 校验 routing schema 合法性（v0.1 phase 1.4 起）
+3. 用户读 markdown 部分学习路由逻辑；引擎读 frontmatter 自动决策
+4. 跨 harness 时砍 `hard_route` 块即可，`soft_hint` 跨平台通用
+
+---
+
+## 2. 顶层结构
+
+```yaml
+---
+trigger:                          # 触发条件
+  keywords: [...]
+  file_globs: [...]?
+  contexts: [...]?
+
+hard_route:                       # C 层 hook 读这块
+  primary: <skill-name>
+  secondary: <skill-name>?
+  default: <skill-name>?
+  override:                       # 条件覆盖
+    condition: "<自然语言描述触发条件>"
+    primary: <skill-name>
+    secondary: <skill-name>?
+  branches:                       # 多分支（mutex 决策）
+    - condition: ...
+      route: <skill-name>
+      members: [...]?
+
+soft_hint:                        # B 层 skill description 自动生成
+  description_template: "<1-line hint, ≤ 500 char>"
+
+fallback:
+  on_hook_fail: <degrade_to_soft_hint|prompt_user>
+  on_both_fail: <prompt_user|halt>
+---
+
+# 人类可读的路由说明（markdown 部分）
+（详细规则、边界场景、历史决策……）
+```
+
+`additionalProperties: false`（v0.1 phase 1.4 落地）— 未声明字段被 reject。
+
+---
+
+## 3. trigger 字段表
+
+| 字段 | 必填 | 类型 | 说明 |
+|------|------|------|------|
+| `keywords` | ✅ | string[] | 至少 1 个；C 层 hook 在用户输入子串匹配时触发；中英文混合（`[ui, ux, frontend, layout, design, 前端, 界面, 设计稿]`） |
+| `file_globs` | optional | string[] | 仅当用户当前打开 / 修改这些 glob 文件时触发（`["**/*.tsx", "**/*.css"]`） |
+| `contexts` | optional | string[] | 上下文标签（`design-review` / `pr-review` / `bug-fix` 等），由 workflow phase 写入 `.harnessed/context.json` |
+
+至少要有 `keywords`；其他字段任选。所有 `keywords` / `file_globs` / `contexts` **OR** 关系（任一命中即触发）。
+
+---
+
+## 4. hard_route 字段表（C 层 hook 决策）
+
+| 字段 | 必填 | 类型 | 说明 |
+|---|---|---|---|
+| `primary` | optional | string | 默认主路由 skill name |
+| `secondary` | optional | string | 默认次路由（cooperative）skill name |
+| `default` | optional | string | 无 override / branches 命中时的兜底（兼容性） |
+| `override` | optional | object | 单一覆盖规则（见下） |
+| `override.condition` | required if override | string | 自然语言条件（如 `"用户明确说'做出风格'/'独特设计'"`），hook 用 LLM 子调用判定 |
+| `override.primary` | required if override | string | 覆盖后的主路由 |
+| `override.secondary` | optional | string | 覆盖后的次路由 |
+| `branches` | optional | object[] | 多分支决策（互斥，按顺序首个命中生效） |
+| `branches[].condition` | required if branches | string | 分支条件 |
+| `branches[].route` | required if branches | string | 该分支的 skill name |
+| `branches[].members` | optional | string[] | 群体路由成员（如 code-review-team） |
+
+**至少**要有 `primary` 或 `default` 或 `branches[0].route` 之一（保证至少有一个 fallback path）。
+
+---
+
+## 5. soft_hint 字段表（B 层 skill description 模板）
+
+| 字段 | 必填 | 类型 | 说明 |
+|---|---|---|---|
+| `description_template` | ✅ | string | 1 行 hint，≤ 500 char；写入 skill 的 description metadata，让 LLM 在 skill 路由时读到 |
+
+**好的 description 示例**：
+- `"标准化、数据驱动的 UI 设计；优先于 frontend-design 用于布局/规范类任务"`
+- `"通用 web 搜索（关键词、库文档、新闻）；性能、过滤器最丰富；默认选 Tavily"`
+
+**不好的示例**：
+- `"useful for UI"`（太泛）
+- `"description here"`（占位）
+
+---
+
+## 6. fallback 字段表
+
+| 字段 | 必填 | 类型 | 说明 |
+|---|---|---|---|
+| `on_hook_fail` | ✅ | enum | hook 异常 / 超时时行为：`degrade_to_soft_hint`（降级到 B 层）/ `prompt_user`（直接问用户） |
+| `on_both_fail` | ✅ | enum | B+C 双双失败时行为：`prompt_user` / `halt`（中止当前任务） |
+
+`degrade_to_soft_hint` 是默认推荐 — hook 失败时不卡用户；`halt` 仅用于安全敏感场景。
+
+---
+
+## 7. 完整示例（PROJECT-SPEC § 9 摘录）
+
+```yaml
+# routing/ui.md
+---
+trigger:
+  keywords: [ui, ux, frontend, layout, design, 前端, 界面, 设计稿]
+  file_globs: ["**/*.tsx", "**/*.vue", "**/*.css", "**/*.scss"]
+
+hard_route:                       # C 层 hook 读这块
+  primary: ui-ux-pro-max
+  secondary: frontend-design
+  override:
+    condition: "用户明确说'做出风格'/'要风格'/'独特设计'"
+    primary: frontend-design
+    secondary: ui-ux-pro-max
+
+soft_hint:                        # B 层 skill description 自动生成
+  description_template: "标准化、数据驱动的 UI 设计；优先于 frontend-design 用于布局/规范类任务"
+
+fallback:
+  on_hook_fail: degrade_to_soft_hint     # hook 失败时降级回 B 层
+  on_both_fail: prompt_user              # 都失败时直接问用户
+---
+
+# 路由说明
+ui-ux-pro-max 主导（数据驱动、可解释），frontend-design 在剩余维度补充。冲突时 ui-ux-pro-max 优先。
+**除非用户明确要求"做出风格"** —— 此时 frontend-design 主导。
+```
+
+---
+
+## 8. v0.1 落地预告：routing/search.md（web 搜索路由）
+
+依据用户全局 `CLAUDE.md` § "Web 搜索路由规则"，phase 1.4 将落地（示意，按 routing schema CI 校验）：
+
+```yaml
+# routing/search.md
+---
+trigger:
+  keywords: [search, 搜索, 查文档, find, "网络上"]
+  contexts: [research]
+
+hard_route:
+  primary: tavily-mcp                           # 默认 Tavily（关键词查询、库文档、新闻）
+  branches:
+    - condition: "描述式查询 / 学术论文 / 批量抓取多个 URL / 用户明说'研究'"
+      route: exa-mcp                            # 神经搜索
+    - condition: "抓整站 / 站点结构（crawl/map）"
+      route: tavily-mcp
+  override:
+    condition: "用户拿不准时"
+    primary: tavily-mcp
+    secondary: exa-mcp                          # 并行打两边对比
+
+soft_hint:
+  description_template: "通用 web 搜索；关键词查询 / 站点抓取走 Tavily，描述式 / 学术 / 批量 URL 走 Exa"
+
+fallback:
+  on_hook_fail: degrade_to_soft_hint
+  on_both_fail: prompt_user
+---
+```
+
+---
+
+## 9. 当前 status & 关联（v0.1 phase 1.1）
+
+- ✅ 字段定义已锁定（本文件 + SPEC § 9）
+- ❌ JSON Schema artifact / routing/*.md / validator 由 v0.1 phase 1.4 落地（TypeBox + Ajv 同 `src/manifest/schema/` 模式）；v0.3 验证 B+C 命中率 ≥ 85%
+
+**关联**：
+- 决策：`PROJECT-SPEC.md` § 9（B+C 混合方案 R4.1）
+- B+C 角色映射：用户全局 `CLAUDE.md` § "Web 搜索路由规则" + § "UI/UX / 前端相关子任务"
+- workflow 触发引用：`workflows/SCHEMA.md` § plan-feature reference（`02-brainstorm.conditional.if` 即来自 `routing/ui.md`）
+- manifest 上游引用目标：`manifests/SCHEMA.md`（路由结果是 manifest name）
+- ROADMAP 落地节奏：v0.1 phase 1.4 → routing schema + ≥ 1 份真实 routing/*.md

package/routing/decision_rules.yaml

@@ -0,0 +1,387 @@
+# yaml-language-server: $schema=../schemas/decision-rules.v1.schema.json
+#
+# decision_rules.yaml v2 — phase 1.5 T4.1 升级 (v1 frozen at phase 1.3 ship)
+#
+# IMPL NOTE (KICKOFF B1 + ASSUMPTIONS § C P0-1 + GRAY-AREA-1 § 2 真理 source):
+#   - 全局 rule-set (DMN Priority Hit Policy 风格) — phase 1.4+ routing engine 优先 load 本文件
+#   - 与 manifest.spec.decision_rules (per-manifest hint, ADR 0007 errata T2.1 加) schema 完全独立
+#   - W-6 v2 演化路径: 任何 v1 字段改动必走 ADR 0009 errata + scripts/migrate-decision-rules-v1-to-v2.mjs
+#   - B1 security: 本文件 string fields 不允许 shell metacharacters ($(...) / ${...} / backtick) —
+#     loadDecisionRules() 内 checkCmdString 二次过滤 (沿袭 phase 1.1.1 H7 mitigation pattern)
+#
+# IMPL NOTE — v1 → v2 升级 (phase 1.5 T4.1, ADR 0009 § Decision):
+#   - version bump 1 → 2; v1 的 12 rules 全部保留 byte-identical (additive-only — D1.5-10)
+#   - 新增 engineering category 5 specific rules (engineering-discuss-feature /
+#     -plan-architecture / -execute-tdd / -execute-debug / -verify-pr) — D1.5-3
+#   - 新增 mattpocock_phases: 段 — 4 phase (discuss/plan/execute/verify) × 21 unique
+#     skills × 23 trigger entries; 真理 source = 用户笔记 ~/.claude/CLAUDE.md 行 8 /
+#     38-39 / 148 (1:1 提取, 不创新 — D1.5-3)
+#   - engineering sub-rules 用 skills_overlay: { ref: mattpocock_phases.<phase>.skills }
+#     cross-link 而非内联 duplicate (D1.5-3 — 避免 skill list 漂移)
+#   - W-6 migration: scripts/migrate-decision-rules-v1-to-v2.mjs idempotent 升级 path
+
+version: 2
+hit_policy: P  # Priority — 多规则命中时按 priority 字段排序取最高
+
+# === L1 路由：关键词 + 优先级 ===
+rules:
+  # ─── design (2 rules) ────────────────────────────────────────
+  - id: ui-task-bold-style-override
+    priority: 100
+    domain: design
+    when:
+      task_type: ui-design
+      # Phase 2.3 W2 T2.1 — 9 keywords 词集校准 (RESEARCH § 3.3 D2.3-4):
+      # ADD: signature style / art direction / bold style (high-signal anchors)
+      # KEEP: 做出风格 / 独创 / 风格化 / 品牌调性 / design-led / distinctive / creative / experimental
+      # REMOVED: 独特 (false-pos "独特的搜索方案" 太高 — RESEARCH § 3.3)
+      override_keywords:
+        - 做出风格
+        - 独创
+        - 风格化
+        - 品牌调性
+        - design-led
+        - distinctive
+        - signature style
+        - art direction
+        - bold style
+        - creative
+        - experimental
+    decision:
+      primary_expert: frontend-design
+      secondary_expert: ui-ux-pro-max
+      conflict_policy: primary_wins
+      rationale: 用户明示 style-driven, 由 frontend-design 主导剩余维度
+
+  - id: ui-task-default
+    priority: 50
+    domain: design
+    when:
+      task_type: ui-design
+    decision:
+      primary_expert: ui-ux-pro-max
+      secondary_expert: frontend-design
+      conflict_policy: primary_wins
+      rationale: 数据驱动 + 标准化 + 可解释 优先 (CLAUDE.md UI 路由默认)
+      # Phase 2.3 W2 T2.1 — D-04 CD-3 negative-space + redirect (NEW optional fields).
+      # 若 task 含 style-led 信号则 reject 本规则,redirect 到 frontend-design (D-08 双向 redirect 一致性).
+      do_not_use_when:
+        - 做出风格
+        - 独创
+        - 风格化
+        - design-led
+      if_rejected_use: frontend-design
+
+  # ─── content (2 rules) ───────────────────────────────────────
+  - id: pptx-file-task
+    priority: 80
+    domain: content
+    when:
+      task_type: pptx-file-operation
+    decision:
+      primary_expert: anthropics-skills-pptx
+      rationale: PowerPoint 文件操作 — anthropics 官方 skill
+
+  - id: chinese-content-deck
+    priority: 70
+    domain: content
+    when:
+      task_type: slide-deck
+      language: zh
+    decision:
+      primary_expert: jimliu-baoyu-skills-baoyu-slide-deck
+      rationale: 中文 deck 场景, baoyu-skills license resolved 为 MIT-0 (D-05 / ADR 0010)
+      # Phase 2.3 W2 T2.1 — CD-3 negative-space: 通用 markdown 文档不走 deck-pack.
+      # (no if_rejected_use — content 段 default fallback 到 supervisor 而非定向 redirect)
+      do_not_use_when:
+        - 通用 markdown 文档
+        - generic markdown
+        - plain doc
+
+  # ─── testing (4 rules) ───────────────────────────────────────
+  - id: perf-a11y-memory
+    priority: 100
+    domain: testing
+    when:
+      task_type:
+        - performance
+        - a11y
+        - memory-leak
+        - Core Web Vitals
+        - LCP
+    decision:
+      primary_expert: chrome-devtools-mcp
+      forbidden_experts:
+        - playwright-cli
+        - playwright-test
+        - webapp-testing
+      rationale: 笔记硬规则 — perf/a11y/memory 不允许 playwright 系列做诊断
+
+  - id: e2e-with-python-backend
+    priority: 70
+    domain: testing
+    when:
+      task_type: e2e-test
+      backend_language: python
+    decision:
+      primary_expert: webapp-testing
+      rationale: setup 需调 Python 后端 (Tortoise ORM / pandas / backend utils)
+      # Phase 2.3 W2 T2.1 — CD-3 negative-space: 若 task 明示无 Python backend → redirect 到 playwright-test 默认。
+      do_not_use_when:
+        - pure-frontend
+        - no-Python-backend
+        - no python backend
+      if_rejected_use: playwright-test
+
+  - id: e2e-default
+    priority: 50
+    domain: testing
+    when:
+      task_type: e2e-test
+    decision:
+      primary_expert: playwright-test
+      rationale: 默认 E2E TS — frontend 是 TS/React/Vue, setup 仅浏览器 + HTTP API
+
+  - id: ai-explore-debug
+    priority: 50
+    domain: testing
+    when:
+      task_type:
+        - ai-explore
+        - interaction-debug
+        - element-ref
+    decision:
+      primary_expert: playwright-cli
+      rationale: Bash 一行命令 + token 最省 — 探查页面结构/调试一次性交互
+
+  # ─── search (2 rules) ────────────────────────────────────────
+  - id: search-academic-or-batch-or-token-sensitive
+    priority: 80
+    domain: search
+    when:
+      task_type: search
+      signals:
+        - 学术
+        - 论文
+        - describe
+        - 批量 URL
+        - token-sensitive
+    decision:
+      primary_expert: exa-mcp
+      rationale: 描述式查询/学术/批量抓取/token 敏感 — Exa highlights 密度高
+
+  - id: search-default
+    priority: 50
+    domain: search
+    when:
+      task_type: search
+    decision:
+      primary_expert: tavily-mcp
+      rationale: 默认 — 关键词查询/库 API 文档/新闻时效信息/生产 RAG
+
+  # ─── meta (2 rules) ──────────────────────────────────────────
+  - id: meta-create-skill
+    priority: 50
+    domain: meta
+    when:
+      task_type: skill-creation
+    decision:
+      primary_expert: anthropics-skills-skill-creator
+      rationale: meta 创建 skill — anthropics 官方 skill-creator
+
+  - id: meta-find-skill
+    priority: 50
+    domain: meta
+    when:
+      task_type: skill-discovery
+    decision:
+      primary_expert: vercel-labs-skills-find-skills
+      rationale: meta 搜索 skill — vercel-labs 生态发现入口
+
+  # ─── engineering (phase 1.5 T4.1 — 5 specific rules) ─────────
+  # NOTE: engineering category base layer 已装 (gstack/GSD/superpowers/karpathy/
+  #       mattpocock/ralph-loop/planning-with-files), 不需 install routing;
+  #       下列 5 rules 按 discuss/plan/execute/verify 分阶段调度 mattpocock 23 招式
+  #       (skills_overlay ref → mattpocock_phases 段, 见文件末). D1.5-3 真理 source:
+  #       用户笔记 ~/.claude/CLAUDE.md 工作流 4 阶段映射.
+  - id: engineering-discuss-feature
+    priority: 60
+    domain: engineering
+    when:
+      task_type: engineering
+      keywords:
+        - 新功能
+        - feature 启动
+        - discuss feature
+    decision:
+      workflow: gstack-decision-gate
+      skills_overlay:
+        ref: mattpocock_phases.discuss.skills
+      gstack_gates:
+        - office-hours
+        - plan-ceo-review
+      rationale: 新功能启动 — gstack 决策关卡先行 + discuss 阶段招式 overlay
+
+  - id: engineering-plan-architecture
+    priority: 60
+    domain: engineering
+    when:
+      task_type: engineering
+      keywords:
+        - 架构 plan
+        - architecture plan
+        - design
+    decision:
+      workflow: gsd-plan-phase
+      skills_overlay:
+        ref: mattpocock_phases.plan.skills
+      gstack_gates:
+        - plan-eng-review
+      rationale: 架构 plan — gsd plan-phase + 复杂架构 gstack eng review
+
+  - id: engineering-execute-tdd
+    priority: 60
+    domain: engineering
+    when:
+      task_type: engineering
+      keywords:
+        - TDD
+        - test first
+        - core logic
+        - algorithm
+    decision:
+      workflow: gsd-execute-task
+      skills_overlay:
+        ref: mattpocock_phases.execute.skills
+      triggers:
+        complexity_threshold: 5
+        category_match:
+          - core_business_logic
+          - algorithm
+          - high_reliability
+        tdd_required: true
+      rationale: 核心逻辑/算法 — gsd execute-task + TDD red-green-refactor 强制
+
+  - id: engineering-execute-debug
+    priority: 60
+    domain: engineering
+    when:
+      task_type: engineering
+      keywords:
+        - debug
+        - bug
+        - diagnose
+        - 排错
+    decision:
+      workflow: gsd-execute-task
+      skills_overlay:
+        ref: mattpocock_phases.execute.skills
+      primary_skills:
+        - diagnose
+        - zoom-out
+      rationale: 系统化排错 — gsd execute-task + diagnose/zoom-out 招式主导
+
+  - id: engineering-verify-pr
+    priority: 60
+    domain: engineering
+    when:
+      task_type: engineering
+      keywords:
+        - PR review
+        - code review
+        - ship
+        - release
+    decision:
+      workflow: gsd-verify-work
+      skills_overlay:
+        ref: mattpocock_phases.verify.skills
+      gstack_gates:
+        - review
+        - code-review
+      rationale: PR 审查/ship — gsd verify-work + gstack Paranoid Staff Engineer review
+
+# === mattpocock_phases — 23 招式 phase routing schema (phase 1.5 T4.1 / D4) ===
+# 4 phase × 21 unique skills × 23 trigger entries; 真理 source = 用户笔记
+# ~/.claude/CLAUDE.md 行 8 (招式定位) + 38-39 (Discuss/Plan/Execute/Verify 工作流) +
+# 148 (UI 路由). engineering category sub-rules skills_overlay.ref 指向本段.
+mattpocock_phases:
+  discuss:
+    skills:
+      - grill-with-docs
+      - to-prd
+      - grill-me
+      - explore
+    triggers:
+      - /grill-with-docs
+      - /to-prd
+      - /grill-me
+      - /explore
+      - 澄清规格
+      - 沉淀 PRD
+      - 拷问我
+      - 探索方案
+  plan:
+    skills:
+      - to-issues
+      - grill-me
+      - design-review
+    triggers:
+      - /to-issues
+      - /grill-me
+      - /design-review
+      - 拆任务
+      - design review
+  execute:
+    skills:
+      - tdd
+      - diagnose
+      - zoom-out
+      - caveman
+      - grill-with-docs
+      - playwright-cli
+      - improve-codebase-architecture
+    triggers:
+      - /tdd
+      - /diagnose
+      - /zoom-out
+      - /caveman
+      - /grill-with-docs
+      - /playwright-cli
+      - /improve-codebase-architecture
+      - TDD
+      - 诊断
+      - 陌生模块
+      - 架构健康
+  verify:
+    skills:
+      - qa
+      - review
+      - code-review
+      - cso
+      - security-review
+      - retro
+      - ship
+    triggers:
+      - /qa
+      - /review
+      - /code-review
+      - /cso
+      - /security-review
+      - /retro
+      - /ship
+      - QA
+      - Paranoid Staff Engineer
+      - code review
+      - 安全审查
+      - ship
+
+# === L2 路由：LLM Supervisor (复杂仲裁仅 fallback) ===
+fallback_supervisor:
+  trigger: L1 无 rule 命中 OR 多 rule 同 priority 冲突 OR 用户 prompt 含未列在 override_keywords 的语义
+  llm: claude-opus-4-7
+
+# === Deprecation 提示 ===
+deprecated:
+  - id: brave-search-mcp
+    reason: API 改 5 USD 月度信用 + 强制信用卡
+    fallback: tavily-mcp + exa-mcp

package/routing/plan-review-schema.yaml

@@ -0,0 +1,50 @@
+# routing/plan-review-schema.yaml (Phase 2.4 W2 D-02 ABSORB)
+# NEW SSOT for EE-4 plan-checker quantitative rubric.
+# Source: .planning/intel/omc-comparison.md L74-82 (omo 4 维原型 formalize; 含 detection / verification dimensions).
+# NOT vendor: 不复刻 vendor prompt 字眼 (含希腊神话 character naming), 仅借 4 维 schema 形态 (B-09 + intel L80).
+# Schema location: yaml-only SSOT (D-WP-2 (a)), NOT TypeBox regen (Karpathy YAGNI per B-10).
+# Thresholds: T2.0 baseline spike outcome RELAX (c8e4147) -- intel defaults
+#   (1.0/0.8/0.9/0) score 健康 ship phase-2.2 + phase-2.3 task_plan = 3/4 WARNING
+#   -> RELAX fr 1.0->0.80 (NEW-file 容忍) + ac 0.9->0.80 (heuristic 覆盖盲点).
+schemaVersion: 1
+apiVersion: harnessed/v1
+kind: PlanReview
+metadata:
+  name: ee-4-plan-review-schema
+  version: 1
+  source: intel/omc-comparison.md L74-82
+  baseline_spike: phase-2.4 T2.0 (commit c8e4147, 2026-05-16)
+spec:
+  dimensions:
+    file_references_verified:
+      description: "plan 提及的 file path 存在比 >= threshold (NEW file 容忍)"
+      threshold: 0.80
+      measurement: "regex files?(_modified|_created): -> token split -> fs.existsSync ratio"
+      rationale: "T2.0 spike: phase-2.2 fr=0.833, phase-2.3 fr=0.893; intel 1.0 过严"
+    reference_sources_real:
+      description: "plan 引用 .md/.yaml 路径存在比 >= threshold (anchor refs B-NN / ADR N / Phase X.Y 不参与 -- 按 structural assumed valid)"
+      threshold: 0.80
+      measurement: "regex (decision_source|see|ref): -> .md/.yaml token -> fs.existsSync ratio; no md ref -> 1.0"
+      rationale: "intel L74-82 默认, T2.0 confirm both plans sr=1.000"
+    concrete_acceptance:
+      description: "acceptance_criteria 子 bullet 含可量化 signal 比 >= threshold"
+      threshold: 0.80
+      measurement: "regex acceptance_criteria:\\n + indented sub-bullets; quant heuristic grep -c|wc -l|exit \\d+|<=N|>=N|==N|命中|absent|present|! grep|0 (增|副|残) etc."
+      rationale: "T2.0 spike: phase-2.2 ac=0.921, phase-2.3 ac=0.948; intel 0.9 紧贴, 留 false-neg 余地"
+    business_logic_assumptions:
+      description: "weasel words (presumably|likely|probably|maybe|should be|assumed) 计数 <= threshold; 白名单 (per D-NN) / (locked) / assumption (名词) 不计"
+      threshold: 0
+      measurement: "regex weasel words count; whitelist context window 80 chars"
+      rationale: "intel 0 默认, T2.0 confirm both plans weasel=0"
+  scoring:
+    pass_threshold: 4    # 4/4 -> PASS (APPROVED)
+    warning_threshold: 3 # 3/4 -> WARNING (APPROVED WITH CONDITIONS)
+    blocker_threshold: 2 # <=2/4 -> BLOCKER (REJECTED, manual rerun)
+  verdict_mapping:
+    pass: "APPROVED -- 4/4 dimensions met"
+    warning: "APPROVED WITH CONDITIONS -- 3/4 dimensions met, see warnings"
+    blocker: "REJECTED -- <=2/4 dimensions met, plan-phase rerun required (manual per B-12)"
+  enforcement:
+    ci_step: ".github/workflows/ci.yml step EE-4 plan-checker (Phase 2.4 W2 D-02 / T2.4)"
+    auto_spawn: false  # B-12: manual /gsd-plan-phase rerun; v0.3.0 absorbs auto-spawn
+    walker: "scripts/run-plan-checker.mjs (T2.2 ship)"

package/schemas/README.md

@@ -0,0 +1,33 @@
+# schemas/ — 可独立 publish 的 JSON Schema artifact
+
+`schemas/` 包含 harnessed 编译输出的**标准 JSON Schema 文件**，与 npm 包一起 publish 让社区在 IDE 中直接校验自己的 manifest / workflow / routing 文件。
+
+## 文件（v0.1 phase 1.1 起）
+
+- `manifest.v1.schema.json` — manifest schema（编译自 `src/manifest/schema/index.ts` TypeBox 定义）
+- `workflows.v1.schema.json` — phases schema（v0.3）
+- `routing.v1.schema.json` — B+C routing schema（v0.1 phase 1.4）
+
+## 使用（社区作者）
+
+在自己的 manifest yaml 文件顶部加一行：
+
+```yaml
+# yaml-language-server: $schema=https://harnessed.dev/schemas/manifest.v1.schema.json
+```
+
+VS Code redhat.vscode-yaml 插件会自动启用补全 + 实时校验。
+
+## 编译
+
+```bash
+pnpm run build:schema   # 调用 scripts/build-schema.ts
+```
+
+输出 `schemas/manifest.v1.schema.json`，由 CI 跑 `npx ajv compile -s schemas/manifest.v1.schema.json --strict=true` 验证产物自身合规。
+
+## 关联
+
+- 源 schema：`src/manifest/schema/`（TypeBox 定义）
+- 决策依据：`docs/adr/0001-manifest-schema-v1.md`
+- 校验器：Ajv 8.17 + TypeBox 0.34 LTS（见 GA-1）