maestro-flow 0.4.14 → 0.4.16

package/.agy/skills/maestro-ralph/SKILL.md

@@ -15,7 +15,7 @@ Closed-loop decision engine for the maestro workflow lifecycle.
 Reads project state → infers position → builds adaptive chain → delegates execution.
 
 Entry points:
-- **`/maestro-ralph "intent"`** — New session: infer → decompose → build → (decomposition → emit /goal prompt, STOP；否则 dispatch ralph-execute)
+- **`/maestro-ralph "intent"`** — New session: infer → decompose → build → emit /goal prompt（如有 decomposition）→ dispatch ralph-execute
 - **`/maestro-ralph continue`** — Wrapper; dispatches to ralph-execute（首选直接 `/maestro-ralph-execute` 推进 step）
 - **`/maestro-ralph status`** — Display session progress
 
@@ -23,10 +23,9 @@ Entry points:
 
 Initial decomposition (S_DECOMPOSE): boundary-clarified via ≤3 questions for broad intents (重构/全面/迁移/重写). 写入 status.json 的 `boundary_contract` / `execution_criteria` / `task_decomposition`，附 `/goal` prompt。
 
-Node types:
-- **internal** (default): ralph-execute `view_file({command_path})` 命令 .md，按内容内联执行
-- **decision**: 回 ralph 评估
-- **external**: 仅在显式声明时使用（`maestro delegate --to claude`）
+Step kinds:
+- **执行 step**: ralph-execute 调 `run_command("maestro ralph next")` 加载 command .md + required_reading 全文，按 stdout 内联执行
+- **decision step**: `step.decision` 字段非空；回 ralph 评估（CLI 只读分析）
 
 Key difference from maestro coordinator:
 - maestro: static chain → one-time selection → runs all steps
@@ -43,9 +42,9 @@ Mutual invocation with `/maestro-ralph-execute` forms a self-perpetuating work l
                               │           writes status.json
                               │           emits /goal prompt
                               ▼
-                       ralph-execute  ◀─┐ internal → view_file(command_path) inline
-                              │         │ decision → Skill("maestro-ralph")
-                              └─────────┘ writes step.completion_confirmed
+                       ralph-execute  ◀─┐ 执行 step → `maestro ralph next` + inline + `ralph complete`
+                              │         │ decision step → Skill("maestro-ralph")
+                              └─────────┘ CLI writes step.completion_confirmed
                        loop until all completion_confirmed | paused
 ```
 </purpose>
@@ -68,14 +67,14 @@ Remaining     → intent
 
 <invariants>
 1. **Ralph never executes steps** — only creates sessions and evaluates decisions
-2. **Handoff via Skill("maestro-ralph-execute")** — 仅当 session 无 `task_decomposition` 时在创建后自动 handoff；decomposition 路径 STOP 等用户输入。decision 评估后始终 handoff
+2. **Handoff via Skill("maestro-ralph-execute")** — 创建 session 后始终自动 handoff；decision 评估后始终 handoff
 3. **Decision delegates read-only** — `maestro delegate --role analyze --mode analysis`
-4. **Default type = internal** — `external` 仅显式标注时出现，build 不默认生成
+4. **执行 step 通过 `maestro ralph next` CLI 加载并内联执行**（详见 invariant 8）
 5. **status.json 是唯一真源** — 不生成 markdown 清单或侧文件
-6. **每个 step 必须 `completion_confirmed: true`** — 基于 `--- COMPLETION STATUS ---` 的 `STATUS: DONE`；缺失则视为未完成
-7. **command_path 在 A_BUILD_STEPS 解析** — 全局优先 `~/.claude/commands/{name}.md`，fallback 项目 `.claude/commands/{name}.md`，写入 status.json
-8. **Internal step 加载契约** — ralph-execute 读 `command_path` 后，必须解析并加载该命令 `<required_reading>` 引用的所有文件（"入口 + workflow"形式的核心），并把 `<deferred_reading>` 路径记录到 `step.deferred_reads`；加载完成后输出 `✓ skill {name} 加载完成`。ralph 在 build 阶段只解析路径，不读 .md 内容
-9. **Decomposition is outcome-oriented** — sub-goals 为可观测交付，禁止 lifecycle 复刻；`/goal` 用户绑定，ralph 只发提示词后 STOP，等用户输入
+6. **每个 step 必须 `completion_confirmed: true`** — 由 `maestro ralph complete N --status DONE`（或 DONE_WITH_CONCERNS）写入；CLI 是唯一合法写入路径
+7. **command_path 在 A_BUILD_STEPS 解析** — 通过 `maestro ralph skills --json --quiet` 预校验（project 覆盖 global），命中即写绝对路径到 status.json；未命中标 `command_scope = "missing"`
+8. **执行 step 加载契约** — 由 `maestro ralph next` CLI 在执行期完成：解析 frontmatter + `<required_reading>` + `<deferred_reading>`，自动读取 required 文件全文并拼入 prompt；缺失 required → 退出码 1（E007），pause session。ralph build 阶段只通过 `maestro ralph skills` 校验路径存在性，不读 .md 内容
+9. **Decomposition is outcome-oriented** — sub-goals 为可观测交付，禁止 lifecycle 复刻；`/goal` 用户绑定，ralph 输出提示词后继续 handoff，用户可在执行过程中随时输入 `/goal`
 10. **planning_mode governs arg granularity** — `unified` → skill args 无 `{phase}`；`independent` → 含 `{phase}`
 11. **task_decomposition 驱动 steps[] 动态生长** — `post-goal-audit` 按 unmet 子目标插入 scoped mini-loop；字段可选/累加，既有字段不删不改
 </invariants>
@@ -147,9 +146,8 @@ S_BUILD_CHAIN:
   → S_CREATE_SESSION DO: A_BUILD_STEPS
 
 S_CREATE_SESSION:
-  → END             WHEN: task_decomposition present        DO: A_CREATE_SESSION (emits Goal Prompt → STOP，等用户输入 /goal 后手动 /maestro-ralph-execute)
-  → S_CONFIRM       WHEN: not auto_confirm AND no decomposition DO: A_CREATE_SESSION
-  → S_DISPATCH      WHEN: auto_confirm AND no decomposition  DO: A_CREATE_SESSION
+  → S_CONFIRM       WHEN: not auto_confirm                   DO: A_CREATE_SESSION
+  → S_DISPATCH      WHEN: auto_confirm                       DO: A_CREATE_SESSION
 
 S_CONFIRM:
   → S_DISPATCH      WHEN: user selects "Proceed"
@@ -159,7 +157,7 @@ S_CONFIRM:
 S_DISPATCH:
   → END             DO: view_file(AbsolutePath="<agy-skills-dir>/maestro-ralph-execute/SKILL.md") + execute inline
 
-S_DECISION_EVAL:
+S_DECISION_EVAL: (decision 节点 == `step.decision` 非空，下述 gate 名取自该字段)
   → S_APPLY_VERDICT WHEN: quality-gate (post-verify, post-business-test, post-review, post-test)
                      DO: A_DELEGATE_EVALUATE
   → S_APPLY_VERDICT WHEN: goal-gate (post-goal-audit)
@@ -198,7 +196,7 @@ S_FALLBACK:
 
 1. Find latest ralph session (by created_at)
 2. Display: Session, Status, Position, Progress, Current step
-3. List steps: [✓] completion_confirmed, [▸] current, [ ] pending, [◆] decision；附 `command_scope`(global/project) + `command_path`
+3. List steps: [✓] completion_confirmed, [▸] current, [ ] pending, [◆] decision（`step.decision` 非空）；执行 step 附 `command_scope`(global/project) + `command_path`
 4. If `task_decomposition` present (absent → skip):
    ```
    Sub-goals  ({done}/{total})    source: {session_dir}/status.json#/task_decomposition
@@ -378,24 +376,26 @@ narrow → derive defaults from intent + codebase, skip questions.
 
 Generate steps from `session.lifecycle_position` to `milestone-complete`.
 
-| Stage | Skill (independent) | Skill (unified) | Type | Decision after | quality_mode |
-|-------|---------------------|-----------------|------|----------------|--------------|
-| brainstorm | `maestro-brainstorm "{intent}"` | *(same)* | internal | — | all |
-| blueprint | `maestro-blueprint "{intent}"` | *(same)* | internal | — | all |
-| init | `maestro-init` | *(same)* | internal | — | all |
-| analyze-macro | `maestro-analyze "{intent}"` | *(same)* | internal | `post-analyze-scope` | all |
-| roadmap | `maestro-roadmap --from analyze:{analyze_macro_id}` | *(same)* | internal | — | all |
-| analyze | `maestro-analyze {phase}` | `maestro-analyze` | internal | — | all |
-| plan | `maestro-plan {phase}` *(scope=phase)* / `maestro-plan --from analyze:{analyze_macro_id}` *(scope=standalone)* / `maestro-plan --from blueprint:{blueprint_id}` *(scope=standalone)* | `maestro-plan` | internal | — | all |
-| execute | `maestro-execute {phase}` | `maestro-execute` | internal | — | all |
-| verify | `maestro-verify {phase}` | `maestro-verify` | internal | `post-verify` | all |
-| business-test | `quality-auto-test {phase}` | `quality-auto-test` | internal | `post-business-test` | full only |
-| review | `quality-review {phase}` | `quality-review` | internal | `post-review` | all (quick: append `--tier quick`) |
-| test-gen | `quality-auto-test {phase}` | `quality-auto-test` | internal | — | full / standard if coverage<80% |
-| test | `quality-test {phase}` | `quality-test` | internal | `post-test` | full, standard |
-| milestone-audit | `maestro-milestone-audit` | *(same)* | internal | — | all |
-| goal-audit | *(decision-only)* | *(same)* | decision | `post-goal-audit` | all (only if decomposed) |
-| milestone-complete | `maestro-milestone-complete` | *(same)* | internal | `post-milestone` | all |
+| Stage | Skill (independent) | Skill (unified) | Decision after | quality_mode |
+|-------|---------------------|-----------------|----------------|--------------|
+| brainstorm | `maestro-brainstorm "{intent}"` | *(same)* | — | all |
+| blueprint | `maestro-blueprint "{intent}"` | *(same)* | — | all |
+| init | `maestro-init` | *(same)* | — | all |
+| analyze-macro | `maestro-analyze "{intent}"` | *(same)* | `post-analyze-scope` | all |
+| roadmap | `maestro-roadmap --from analyze:{analyze_macro_id}` | *(same)* | — | all |
+| analyze | `maestro-analyze {phase}` | `maestro-analyze` | — | all |
+| plan | `maestro-plan {phase}` *(scope=phase)* / `maestro-plan --from analyze:{analyze_macro_id}` *(scope=standalone)* / `maestro-plan --from blueprint:{blueprint_id}` *(scope=standalone)* | `maestro-plan` | — | all |
+| execute | `maestro-execute {phase}` | `maestro-execute` | — | all |
+| verify | `maestro-verify {phase}` | `maestro-verify` | `post-verify` | all |
+| business-test | `quality-auto-test {phase}` | `quality-auto-test` | `post-business-test` | full only |
+| review | `quality-review {phase}` | `quality-review` | `post-review` | all (quick: append `--tier quick`) |
+| test-gen | `quality-auto-test {phase}` | `quality-auto-test` | — | full / standard if coverage<80% |
+| test | `quality-test {phase}` | `quality-test` | `post-test` | full, standard |
+| milestone-audit | `maestro-milestone-audit` | *(same)* | — | all |
+| goal-audit | *(decision-only)* | *(same)* | `post-goal-audit` | all (only if decomposed) |
+| milestone-complete | `maestro-milestone-complete` | *(same)* | `post-milestone` | all |
+
+> 所有执行 stage 统一通过 `maestro ralph next` CLI 加载 + 内联执行；decision 节点单独作为独立 step 插入（见规则 4）。
 
 **Build rules (按顺序应用):**
 
@@ -403,18 +403,19 @@ Generate steps from `session.lifecycle_position` to `milestone-complete`.
 1. **起点**：从 `session.lifecycle_position` 开始
 2. **跳过已完成**：跳过当前 milestone+phase 下已有 completed artifact 的 stage（按 `session.phase` 过滤）；unified 按 milestone 过滤
 3. **quality_mode 过滤**：按 `session.quality_mode` 排除不匹配 stage
-4. **决策节点**：每个 Decision after 非空的 stage 之后插入 `{ type: "decision", decision: "<gate>", retry_count: 0, max_retries: 2 }`
+4. **决策节点**：每个 Decision after 非空的 stage 之后插入 `{ decision: "<gate>", retry_count: 0, max_retries: 2, command_scope: null, command_path: null }`
 5. **goal-audit 插入**：`task_decomposition` 存在时，在最后一个 evidence-producing stage（verify/review/test）之后、`milestone-complete` 之前插入 `decision:post-goal-audit`
 6. **终点硬约束**：chain 以 `milestone-complete` 结尾
 7. **goal_ref 传播**：`task_decomposition` 存在时，每个 step 按 `step.stage ∈ g.lifecycle` 匹配 `step.goal_ref = g.id`（多匹配取字典序最小）；decision 节点不打 goal_ref
 8. **占位符**：independent 保留 `{phase}` `{intent}`；unified 不带 `{phase}`
-9. **command_path 解析**（每个非 decision step）：
+9. **command_path 解析**（每个执行 step，decision 节点跳过）：
    - 取 skill 名（args 前的第一个 token）
-   - 全局优先：`~/.claude/commands/{name}.md` 存在 → `command_scope = "global"`
-   - Fallback：`.claude/commands/{name}.md` 存在 → `command_scope = "project"`
-   - 两者都缺 → `command_scope = "missing"`, `command_path = null`，A_CREATE_SESSION 报错 E006
-   - **不在 build 阶段读取 .md 内容**；`<required_reading>` / `<deferred_reading>` 解析与加载由 ralph-execute A_EXEC_INTERNAL 负责（保持入口/工作流分离）
-10. **每个 step 初始化** `completion_confirmed: false`, `completion_status: null`, `completion_evidence: null`, `deferred_reads: []`
+   - **预校验通过 `run_command("maestro ralph skills --json --quiet")`** 一次性拉取所有可用 commands + skills（global + project，project 覆盖 global），匹配 skill 名得到：
+     - 命中 commands → `command_scope = "global" | "project"`，`command_path = <绝对路径>`
+     - 命中 skills → 同上（type=skill）
+     - 未命中 → `command_scope = "missing"`, `command_path = null`，A_CREATE_SESSION 报错 E006
+   - **不在 build 阶段读取 .md 内容**；`<required_reading>` / `<deferred_reading>` 解析与加载由 `maestro ralph next` CLI 在执行期完成
+10. **每个 step 初始化** `completion_confirmed: false`, `completion_status: null`, `completion_evidence: null`, `deferred_reads: []`, `load: null`（由 `ralph next` 写入）
 11. **scope_verdict gating**（仅当 chain 起点 = `analyze-macro`）：
     - `scope_verdict ∈ {medium, small}` → 跳过 `roadmap` + `analyze` 两 stage；`plan` 选 standalone 列（`--from analyze:{analyze_macro_id}`），不带 `{phase}`
     - `scope_verdict == large` → 保留 `roadmap` + `analyze`；`plan` 选 phase 列（`{phase}`）
@@ -432,7 +433,7 @@ Generate steps from `session.lifecycle_position` to `milestone-complete`.
 1. Validate: 所有 step 的 `command_scope != "missing"`；否则 raise E006 + 列出缺失 skill
 2. Write `.workflow/.maestro/ralph-{YYYYMMDD-HHmmss}/status.json` (Appendix: Session Schema)
 3. Display chain overview：每步显示 `{index}. {skill} [{type}] [{command_scope}]`
-4. If `task_decomposition` present: display **Goal Prompt block** (Appendix) → STOP，等用户输入 `/goal`（auto_confirm 也不跳过）
+4. If `task_decomposition` present: display **Goal Prompt block** (Appendix)，不阻塞流程，继续 handoff
 
 ### A_DELEGATE_EVALUATE
 
@@ -593,6 +594,8 @@ Runs only when `task_decomposition` present.
 {
   "session_id": "ralph-{YYYYMMDD-HHmmss}",
   "source": "ralph", "status": "running",
+  "ralph_protocol_version": "1",   // CLI-driven; absent/0 → legacy inline ralph-execute
+  "active_step_index": null,       // CLI-managed; only one step held at a time
   "intent": "", "lifecycle_position": "",
   "phase": null, "phase_is_new": false,
   "milestone": "",                // D-007 反查结果，禁止读 current_milestone
@@ -607,13 +610,15 @@ Runs only when `task_decomposition` present.
     "analysis_dir": null, "brainstorm_dir": null, "blueprint_dir": null },
   "steps": [{
     "index": 0,
-    "type": "internal|external|decision",
-    "skill": "",
+    "skill": "",                  // 执行 step 有值；decision 节点为空字符串/null
     "args": "",
     "stage": "",                  // brainstorm|blueprint|init|analyze-macro|roadmap|analyze|plan|execute|verify|...
     "scope": null,                // "phase"|"standalone"|"milestone"|null（plan 等需要）
-    "command_scope": "global|project|missing|null",
-    "command_path": "~/.claude/commands/{name}.md | .claude/commands/{name}.md | null",
+    "decision": null,             // 非 null → decision 节点（值为 gate 名，如 "post-verify"）；null → 执行 step
+    "retry_count": 0,             // decision 节点专用
+    "max_retries": 2,             // decision 节点专用
+    "command_scope": "global|project|missing|null",  // 执行 step；decision 节点固定 null
+    "command_path": "<absolute path resolved by `maestro ralph skills --json --quiet`> | null",
     "milestone_id": null,         // D-007 反查注入；仅含 {phase} 占位符的 step 有
     "source_artifact_ref": null,  // "analyze:ANL-xxx" | "blueprint:BLP-xxx" | null
     "status": "pending|running|completed|skipped|failed",
@@ -622,7 +627,8 @@ Runs only when `task_decomposition` present.
     "completion_status": null,
     "completion_evidence": null,
     "completed_at": null,
-    "deferred_reads": []          // 由 ralph-execute A_EXEC_INTERNAL 解析 .md 时填充
+    "deferred_reads": [],         // 由 ralph next CLI 解析 .md 时填充
+    "load": null                  // { loaded_at, required_files[], deferred_files[], resolve_version } —— 由 ralph next 写入
   }],
   "waves": [], "current_step": 0,
 
@@ -644,7 +650,7 @@ Runs only when `task_decomposition` present.
 
 ### Fix-Loop Templates
 
-所有插入 step 默认 `type: "internal"`，按 A_BUILD_STEPS 规则 9 解析 `command_path` + `command_scope`。
+所有插入的执行 step 按 A_BUILD_STEPS 规则 9 解析 `command_path` + `command_scope`；`decision:*` 条目为 decision 节点（`step.decision` 字段）。
 
 **post-verify:**
 ```
@@ -706,12 +712,12 @@ decision:post-goal-audit {retry+1}
 链路概览后逐字显示（仅当 decomposition 已产出）：
 
 ```
-📋 任务分解完成。复制以下 /goal 设定终止条件，随后运行 /maestro-ralph-execute：
+📋 任务分解完成。可随时复制以下 /goal 设定终止条件（执行过程中输入即可）：
 
 /goal 直到 {session_dir}/status.json 的 task_decomposition[*] 与 steps[*] 全部 completion_confirmed=true 才停。每轮以 status.json 为唯一行动手册，通过 /maestro-ralph-execute 推进 step；decision 节点由其自动 handoff 回 ralph 评估。禁止手动执行 skill 或修改 boundary_contract.out_of_scope。
 ```
 
-`/goal` 由用户输入；ralph 只输出此提示词。判据以 status.json 为权威。
+`/goal` 由用户输入；ralph 输出提示词后继续 handoff，不阻塞。
 
 ### Error Codes
 
@@ -743,8 +749,8 @@ decision:post-goal-audit {retry+1}
 - [ ] quality_mode 由 A_DETERMINE_QUALITY_MODE 决定，过滤 build steps
 - [ ] Decomposition: broad intent ≤3 question clarify；narrow auto-derive
 - [ ] status.json 唯一真源：boundary_contract + execution_criteria + task_decomposition；无外部清单
-- [ ] 每个 step 默认 `type: "internal"`，含 `command_scope` + `command_path`（全局优先 fallback 项目）
-- [ ] Ralph build 阶段只解析路径，不读 .md 内容；`<required_reading>` 加载由 ralph-execute A_EXEC_INTERNAL 完成
+- [ ] 执行 step 含 `command_scope` + `command_path`（通过 `maestro ralph skills --json --quiet` 预校验，project 覆盖 global）；decision step 通过 `step.decision` 字段标识
+- [ ] Ralph build 阶段只通过 `ralph skills` 校验路径存在性，不读 .md 内容；`<required_reading>` 加载由 `maestro ralph next` CLI 完成
 - [ ] 每个 step 含 `completion_confirmed` + `completion_status` + `completion_evidence` + `deferred_reads`（初始 false/null/[]）
 - [ ] 每个 sub-goal 含 `completion_confirmed`（初始 false）
 - [ ] post-goal-audit decision 仅在 decomposed 时插入，位于 milestone-complete 之前

package/.agy/skills/maestro-ralph-execute/SKILL.md

@@ -27,32 +27,29 @@ Remaining  → session_id (if matches maestro-* or ralph-*)
 ```
 Also read `session.auto_mode` from status.json — if true, treat as `-y`.
 
-**Node types:**
+**Step kinds:**
 
-| Type | Execution | Flow after |
-|------|-----------|------------|
-| decision (ralph-only) | `Skill("maestro-ralph")` | Execution ends here |
-| internal (default) | `view_file({file_path: step.command_path})` + 内联解释执行 | Self-invoke next |
-| external (opt-in) | `maestro delegate --to claude --mode write` (STOP → callback) | Self-invoke next |
+| Kind | Identifier | Execution | Flow after |
+|------|-----------|-----------|------------|
+| decision step | `step.decision` 非空 | `Skill("maestro-ralph")` | Execution ends here |
+| 执行 step | `step.decision == null` | `run_command("maestro ralph next")` → 内联按其 stdout 执行 → `run_command("maestro ralph complete N --status ...")` | Self-invoke next |
 
 HARD RULES:
-- internal step：优先通过 `view_file({command_path})` 把命令 .md 加载进当前会话，再按内容执行；不要对 internal step 使用 `Skill({skill})` 调用
-- **必须遵循 `<required_reading>` / `<deferred_reading>` 标签**：命令 .md 通常采用"入口 + workflow"形式，主体逻辑放在 workflow 文件中并通过 `<required_reading>` 引用；不加载 required_reading 会导致命令执行不完整
-- decision 节点例外：A_EXEC_DECISION 必须使用 `view_file(AbsolutePath="<agy-skills-dir>/maestro-ralph/SKILL.md") + execute inline` 进行 handoff（这是 decision 节点的唯一允许用法）
-- `command_path` 由 ralph 在 A_BUILD_STEPS 写入 status.json；ralph-execute 不再自行解析（缺失 → 报错 E002）
-- external 仅在 `step.type == "external"` 显式声明时使用，并 always append `-y` 到 prompt args
-- 每个 step 必须产出 `--- COMPLETION STATUS ---` 块，否则视为 NEEDS_RETRY
+- 执行 step：**统一通过 `maestro ralph next` CLI 加载**。CLI 负责读 command_path、解析 `<required_reading>` + `<deferred_reading>`、拼接 prompt、写 `step.load.*` + `active_step_index` + `step.status="running"`。不要再在会话里手动 Read + 解析 required_reading
+- decision step：A_EXEC_DECISION 通过 `view_file(AbsolutePath="<agy-skills-dir>/maestro-ralph/SKILL.md") + execute inline` handoff 给 ralph 评估（不走 CLI）
+- `command_path` 由 ralph 在 A_BUILD_STEPS 写入 status.json（缺失 → ralph next 返回 E006/E007 并拒绝执行）
+- 每个 step 结束必须调用 `maestro ralph complete N --status <S>` 或 `maestro ralph retry N`。STATUS 仅 4 个合法值：`DONE | DONE_WITH_CONCERNS | NEEDS_RETRY | BLOCKED`（**`NEEDS_CONTEXT` 已废除**，context 容量由 harness 自动压缩处理）
 </context>
 
 <invariants>
-1. **Internal = Read + inline** — 通过 Read 读取 `step.command_path`，按其指令在当前 session 内执行
-2. **Required reading must be loaded** — 命令 .md 中的 `<required_reading>` 引用的所有文件必须立即 Read；缺一 → 视为加载失败，pause session（E007）
-3. **Deferred reading recorded only** — `<deferred_reading>` 列出的文件路径需记录，执行过程按需 Read；不在加载阶段读取
-4. **Skill loaded confirmation** — 所有 required_reading 加载完成后必须输出一行确认：`✓ skill {step.skill} 加载完成 (required: N, deferred: M)`
-5. **External = explicit only** — `step.type == "external"` 才走 delegate；默认绝不发起
-6. **必须显式 completion confirmation** — 每个 step 完成时需有 `STATUS: DONE` 且写入 `step.completion_confirmed = true`
-7. **Self-invocation chain** — 持续直到全部 `completion_confirmed` 或 paused
-8. **status.json 每步骤后写盘** — resume-safe
+1. **执行 = `ralph next` + inline + `ralph complete`** — 调 `maestro ralph next` 拿到 skill 内容，按 stdout 内联执行
+2. **Required reading 由 CLI 负责** — `ralph next` 自动展开 + 加载 `<required_reading>` 引用的所有文件，缺失 → 退出码 1（E007），不写 active_step_index，不进入执行
+3. **Deferred reading recorded only** — `<deferred_reading>` 路径由 CLI 记录到 `step.load.deferred_files`，执行阶段按需 Read
+4. **一致性取代锁** — 同一 session 同时最多一个 step 持 `active_step_index`；CLI 校验失败直接退出码 3，不静默推进
+5. **Completion 通过 CLI 调用** — 每个 step 末尾调 `maestro ralph complete N --status <S>` 或 `maestro ralph retry N`，由 CLI 写 `completion_*` + 清 `active_step_index`
+6. **Self-invocation chain** — 持续直到全部 `completion_confirmed` 或 paused
+7. **status.json 每步骤后由 CLI 原子写盘** — resume-safe
+8. **STATUS 枚举受限** — 仅 `DONE | DONE_WITH_CONCERNS | NEEDS_RETRY | BLOCKED`；`NEEDS_CONTEXT` 已废除
 </invariants>
 
 <state_machine>
@@ -78,14 +75,13 @@ S_RESOLVE_ARGS:
   → S_EXECUTE       DO: A_RESOLVE_ARGS
 
 S_EXECUTE:
-  → END             WHEN: step.type == "decision"           DO: A_EXEC_DECISION
-  → S_POST_EXEC     WHEN: step.type == "internal" + success DO: A_EXEC_INTERNAL
-  → S_HANDLE_FAIL   WHEN: step.type == "internal" + failure DO: A_EXEC_INTERNAL
-  → END             WHEN: step.type == "external"           DO: A_EXEC_EXTERNAL
-                     (STOP after background delegate; on callback → S_POST_EXEC or S_HANDLE_FAIL)
+  → END             WHEN: step.decision != null              DO: A_EXEC_DECISION
+  → S_POST_EXEC     WHEN: step.decision == null + ralph complete invoked with DONE|DONE_WITH_CONCERNS  DO: A_EXEC_STEP
+  → S_HANDLE_FAIL   WHEN: step.decision == null + ralph next exit≠0 OR ralph complete with NEEDS_RETRY|BLOCKED  DO: A_EXEC_STEP
 
 S_POST_EXEC:
-  → S_LOCATE        DO: A_MARK_COMPLETE + Skill("maestro-ralph-execute")
+  → S_LOCATE        DO: run_command("maestro ralph complete ...") + Skill("maestro-ralph-execute")
+                     NOTE: CLI 已写完 completion_*, status, active_step_index；无需额外写盘
 
 S_HANDLE_FAIL:
   → S_LOCATE        WHEN: auto + not retried               DO: A_RETRY
@@ -108,8 +104,8 @@ S_FALLBACK:
 
 1. If session_id provided → load `.workflow/.maestro/{session_id}/status.json`
 2. Else: scan `.workflow/.maestro/*/status.json`, filter `status == "running"`, sort DESC, take first
-3. Extract: session_id, source, steps[], current_step, phase, milestone, intent, auto_mode, context, cli_tool
-4. Find first step with `status == "pending"` → next step
+3. Extract: session_id, source, steps[], phase, milestone, intent, auto_mode, context, cli_tool, active_step_index
+4. **不在此处选 pending step**——pending 选择由 `maestro ralph next` CLI 内部完成；A_LOCATE_SESSION 只确认 session 存在且 running，由 A_EXEC_STEP 调 CLI 推进
 
 ### A_RESOLVE_ARGS
 
@@ -151,79 +147,41 @@ Write enriched args back to status.json.
 ### A_EXEC_DECISION
 
 1. Mark step running, write status.json
-2. Display: `[{index}/{total}] ◆ {decision} Retry: {retry}/{max}`
+2. Display: `[{index}/{total}] ◆ {step.decision} Retry: {retry}/{max}`
 3. `view_file(AbsolutePath="<agy-skills-dir>/maestro-ralph/SKILL.md") + execute inline` — ralph 评估 + handoff
 4. 执行在此结束
 
-### A_EXEC_INTERNAL
-
-1. Validate `step.command_path != null`；否则 raise E002，pause session
-2. Mark step running, write status.json
-3. Display: `[{index}/{total}] {step.skill} [internal · {step.command_scope}]`
-4. `view_file({ file_path: step.command_path })` — 把命令 .md 全文加载进当前会话（prefer Read over Skill for internal steps；decision 节点另行使用 Skill 见 A_EXEC_DECISION）
-5. **解析 reading 标签**（"入口 + workflow"形式核心步骤）：
-   - 抽取 frontmatter `argument-hint` / `allowed-tools`
-   - 抽取 `<required_reading>` 块的所有 `@path` 引用 → 立刻 `view_file({ file_path: <expanded path> })` 加载（`~/` / `@~/` 展开为用户主目录）；任一文件缺失或读取失败 → raise E007，pause session
-   - 抽取 `<deferred_reading>` 块的所有路径 → 仅记录到 `step.deferred_reads = [...]`，执行阶段按需 Read
-   - 抽取 `<purpose>/<context>/<state_machine>/<execution>/<actions>` 等指令块
-6. **加载完成确认**：required_reading 全部成功 Read 后，输出一行：
-   ```
-   ✓ skill {step.skill} 加载完成 (required: {N}, deferred: {M})
-   ```
-   其中 N = required_reading 引用数，M = deferred_reading 路径数（缺省块按 0 计）
-7. 计算 `effective_args`：`step.args` + auto flag（`auto ? (flag_map[step.skill] || "") : ""`）
-8. 按读到的指令在本会话中**内联执行**：调用允许的工具完成命令所规定的工作，不再发起 delegate；执行过程中如触发 deferred_reading 引用的资源 → 按需 Read
-9. 执行结束：要求最后一段必须包含 `--- COMPLETION STATUS ---` 块（见 A_MARK_COMPLETE）
-10. Return success / failure
-
-**Auto flag map**: 所有 lifecycle skill → `-y`; `quality-test` → `-y --auto-fix`; 未列出 → 无 flag
-
-### A_EXEC_EXTERNAL
+### A_EXEC_STEP
 
-仅当 `step.type == "external"` 时使用（默认链路不产生）。
+1. **Load** — `run_command("maestro ralph next")`
+   - 退出码 0 → 按 stdout 内联执行
+   - 退出码 2 → 交给 S_LOCATE
+   - 退出码 3 → active_step_index 已被占用
+   - 退出码 1 → pause session
+2. **Inline execution** — 按 stdout 执行；deferred_reading 按需 Read
+3. **Complete**:
+   - `run_command("maestro ralph complete N --status DONE [--evidence <path>]")`
+   - `run_command("maestro ralph complete N --status DONE_WITH_CONCERNS --concerns \"...\"")`
+   - `run_command("maestro ralph retry N")`
+   - `run_command("maestro ralph complete N --status BLOCKED --reason \"...\"")`
+4. **Propagate context signals** — 关键信号 (`PHASE: N` / `scratch_dir: path` / `BLP-xxx`) 写入 `status.json.context`
 
-1. Mark step running, write status.json
-2. Display: `[{index}/{total}] ⚡ {step.skill} [external]`
-3. 始终在 prompt 内追加 `-y`（delegate session 非交互）：`flag = flag_map[step.skill] || "-y"`
-4. Execute:
-   ```
-   run_command({
-     command: `maestro delegate "/${step.skill} ${effective_args}" --to claude --mode write`,
-     run_in_background: true, timeout: 600000
-   })
-   STOP — wait for callback.
-   ```
-5. On callback: 把回调输出视为 step 的执行结果 → S_POST_EXEC / S_HANDLE_FAIL
-
-### A_MARK_COMPLETE
-
-1. 从 step 输出中提取 `--- COMPLETION STATUS ---` 块（required）
-2. 解析并写入：
-   - `STATUS: DONE` → `step.status = "completed"`, `step.completion_confirmed = true`, `step.completion_status = "DONE"`
-   - `STATUS: DONE_WITH_CONCERNS` → `step.status = "completed"`, `step.completion_confirmed = true`, `step.completion_status = "DONE_WITH_CONCERNS"`, `step.concerns = <CONCERNS>`
-   - `STATUS: NEEDS_RETRY` → `step.status = "pending"`, `step.retried = true`, `step.completion_confirmed = false`, → S_HANDLE_FAIL
-   - `STATUS: BLOCKED` / `NEEDS_CONTEXT` → `session.status = "paused"`, `step.completion_status` 记录原因, `step.completion_confirmed = false`
-   - 缺失 `--- COMPLETION STATUS ---` 块 → 视为 NEEDS_RETRY（不允许 heuristic fallback）
-3. 写入 `step.completion_evidence`（artifact 路径 / 关键输出节选）
-4. 扫描输出抓取 context 信号：`PHASE: N` → session.phase；`scratch_dir: path` → context.scratch_dir；`BLP-xxx` → context.blueprint_session_id
-5. `step.completed_at = now`，写 status.json
-6. **Sub-goal evidence 校验**（task_decomposition 存在时）：若 `step.goal_ref` 对应子目标的 `lifecycle` 覆盖当前 stage 且 evidence artifact 已生成 → 暂不直接置 done，仍交由 post-goal-audit 决策；仅在 step 显式确认时更新 `task_decomposition[*].completion_confirmed = false` 占位（保持 pending）
-7. Display: `[{index}/{total}] ✓ {step.skill} completed (confirmed)`
+完成后 S_LOCATE 触发 `view_file(AbsolutePath="<agy-skills-dir>/maestro-ralph-execute/SKILL.md") + execute inline` 自调用。
 
 ### A_RETRY
 
-1. `step.retried = true`, `step.status = "pending"`, `step.error = null`, `step.completion_confirmed = false`
-2. Write status.json
+1. `run_command("maestro ralph retry N")` — CLI 设 `step.retried = true`, `step.status = "pending"`, `step.completion_confirmed = false`, 清 `active_step_index`
+2. Display: `[{index}/{total}] ↻ {step.skill} retry`
 
 ### A_SKIP_STEP
 
-1. `step.status = "skipped"`, `step.completion_confirmed = false`
-2. Write status.json
+跳过执行 step — 手动编辑 `status.json`：将该 step `status` 设为 `"skipped"`，`completion_confirmed` 设为 `false`，并清 `active_step_index`（若指向此 step）。
+（不提供 CLI 子命令；跳过是非常规操作，避免自动化误用。）
 
 ### A_PAUSE_SESSION
 
-1. `session.status = "paused"`, write status.json
-2. Display: `[{index}/{total}] ✗ {step.skill} 失败，会话已暂停。/maestro-ralph continue 恢复。`
+通常由 `ralph complete N --status BLOCKED --reason "..."` 触发，CLI 已写 `session.status = "paused"`。手动 pause 场景下直接编辑 status.json。
+Display: `[{index}/{total}] ✗ {step.skill} 失败，会话已暂停。/maestro-ralph continue 恢复。`
 
 ### A_COMPLETE_SESSION
 
@@ -238,14 +196,14 @@ Write enriched args back to status.json.
      Session:  {session_id} [{source}]
      Steps:    {completed}/{total}   confirmed: {confirmed}/{completed}
 
-     [✓] 0.   maestro-plan 1            [internal · global]
-     [✓] 1.   maestro-execute 1         [internal · project]
-     [✓] 2.   maestro-verify 1          [internal · global]
+     [✓] 0.   maestro-plan 1            [global]
+     [✓] 1.   maestro-execute 1         [project]
+     [✓] 2.   maestro-verify 1          [global]
      [✓] 3. ◆ post-verify               [decision]
      ...
    ============================================================
    ```
-   Icons: `✓` confirmed, `—` skipped, `✗` failed, `◆` decision, `⚡` external
+   Icons: `✓` confirmed, `—` skipped, `✗` failed, `◆` decision
 
 </actions>
 
@@ -258,31 +216,30 @@ Write enriched args back to status.json.
 | Code | Severity | Description | Recovery |
 |------|----------|-------------|----------|
 | E001 | error | No running session found | Suggest /maestro or /maestro-ralph |
-| E002 | error | step.command_path missing for internal step | Pause, ask ralph to rebuild step |
-| E003 | error | status.json corrupt | Show path, manual check |
-| E004 | error | Delegate failed + user abort | Mark paused, suggest resume |
-| E005 | error | COMPLETION STATUS block missing | Trigger NEEDS_RETRY |
-| E007 | error | required_reading file 缺失或读取失败 | List missing paths, pause session |
+| E006 | error | command_path missing/unreachable for 执行 step | `ralph next` 拒绝；编辑 status.json 或重 build |
+| E007 | error | required_reading 引用文件缺失 | `ralph next` 拒绝；CLI stderr 列出缺失路径 |
+| E008 | error | `ralph complete` idx ≠ active_step_index | 编辑 status.json 修正一致性 |
+| E009 | error | `ralph complete` step.status ≠ running | 重复 complete 或非法跳跃；编辑 status.json |
+| E010 | error | status.json schema 损坏 | `ralph check` 显示具体损坏字段 |
 | W001 | warning | Step completed with concerns | Log and continue |
-| W002 | warning | command .md 无 `<required_reading>` 标签 | 直接执行 .md 主体，跳过加载阶段 |
+| W005 | warning | active_step_index 指向已 completed step | `ralph next` 自动清理后继续 |
+| W007 | warning | step.skill ≠ command .md frontmatter.name | 提示但不阻塞 |
 
 ### Success Criteria
 
 - [ ] Session discovery covers maestro-* and ralph-*
-- [ ] `-y` parsed from args 或 session.auto_mode
+- [ ] `-y` parsed from args 或 session.auto_mode；auto=true 时透传 `-y` 到 skill args
 - [ ] Placeholders resolved；per-skill enrichment 正确
-- [ ] Decision 节点 Skill("maestro-ralph") handoff
-- [ ] Internal 节点通过 view_file({step.command_path}) 内联执行，禁止 Skill()
-- [ ] Internal 节点 Read 后必须解析并加载 `<required_reading>` 引用的文件；缺失 → E007 pause
-- [ ] `<deferred_reading>` 仅记录路径到 `step.deferred_reads`，执行阶段按需 Read
-- [ ] required_reading 加载完成后输出 `✓ skill {name} 加载完成 (required: N, deferred: M)`
-- [ ] External 仅在显式声明时走 delegate，prompt 必带 `-y`
-- [ ] 每个 step 强制 `--- COMPLETION STATUS ---`；缺失 → NEEDS_RETRY
-- [ ] step.completion_confirmed = true 仅在 STATUS: DONE/DONE_WITH_CONCERNS 时设置
-- [ ] step.completion_evidence 记录 artifact path / 输出节选
-- [ ] Context signals 传播 status.json
+- [ ] Decision 节点（`step.decision != null`）走 Skill("maestro-ralph") handoff（**不调 ralph next CLI**）
+- [ ] 执行 step 通过 `run_command("maestro ralph next")` 加载；CLI 返回拼好的 prompt + completion 协议
+- [ ] required_reading 由 CLI 自动加载并拼入 prompt；缺失 → CLI 退出码 1，pause session
+- [ ] `<deferred_reading>` 由 CLI 记录到 `step.load.deferred_files`，执行阶段按需 Read
+- [ ] 每个 step 末尾必须调 `maestro ralph complete N --status <S>` 或 `maestro ralph retry N`
+- [ ] STATUS 枚举仅 `DONE | DONE_WITH_CONCERNS | NEEDS_RETRY | BLOCKED`；CLI 拒绝 `NEEDS_CONTEXT`
+- [ ] active_step_index 一致性由 CLI 维护；E008/E009 直接退出，不静默推进
+- [ ] step.completion_evidence 通过 `--evidence` 传入并记录
+- [ ] Context signals 由执行 step 显式写回 status.json.context（非 ralph-execute 内嵌扫描）
 - [ ] Auto mode: retry 一次后 pause；interactive 提供 retry/skip/abort
 - [ ] 自调用持续到全部 completion_confirmed 或 paused
-- [ ] A_COMPLETE_SESSION 校验全部 step confirmed + sub-goal all_done
 
 </appendix>

package/.agy/skills/manage-knowledge-audit/SKILL.md

@@ -0,0 +1,90 @@
+---
+name: manage-knowledge-audit
+description: Audit and prune knowledge across spec / knowhow / artifact stores
+argument-hint: --scope <spec|knowhow|artifact|all> [--level P0|P1|P2] [--timeline T1..T6] [--since YYYY-MM-DD] [--milestone <name>] [--include-archive] [--interactive] [--mark|--delete|--purge] [--dry-run] [--report]
+allowed-tools:
+  - ask_question
+  - define_subagent
+  - grep_search
+  - invoke_subagent
+  - manage_subagents
+  - replace_file_content
+  - run_command
+  - send_message
+  - view_file
+  - write_to_file
+---
+<purpose>
+对称于 `manage-harvest`（写入入口）的知识淘汰入口。harvest 负责把 artifact 抽取为 spec/wiki/issue；audit 负责审查这三大存储中已积累的条目，识别矛盾、失效、老化、孤儿，并通过 keep/deprecate/delete 三态决策清理。
+
+覆盖 8 大场景类、28 子场景（显性/隐性矛盾、失效老化、元数据质量、Maestro 特化、时间线产物 T1-T6、knowhow 漂移、artifact 残留），定义见 workflow knowledge-audit.md。
+
+**闭环**：harvest 写入 → audit 审查 → 三态淘汰；与 `harvest --prune`（物理 GC）互补：audit 做语义层判定，且删除未抽取的 artifact 前反向触发 harvest 抢救。
+</purpose>
+
+<required_reading>
+@~/.maestro/workflows/knowledge-audit.md
+</required_reading>
+
+<deferred_reading>
+- @~/.maestro/workflows/harvest.md (audit 检测的 artifact 是 harvest 的产物源)
+- @~/.maestro/workflows/specs-add.md (deprecate 操作所需的 `<spec-entry>` 变形)
+</deferred_reading>
+
+<context>
+Arguments: $ARGUMENTS
+
+**Scope（必选）：** `spec` | `knowhow` | `artifact` | `all`
+
+**删除策略**默认 `--interactive`（三态面板逐项决策）；非交互模式 `--mark`（仅打标）/ `--delete`（软删到 `.trash/`）/ `--purge`（物理擦除，仅 artifact 且需双重确认）。
+
+Flag 全集、scope 对应的扫描路径、Stage 步骤、检测算法定义在 workflow knowledge-audit.md。
+</context>
+
+<execution>
+Follow `~/.maestro/workflows/knowledge-audit.md` Stages 1-8 in order.
+
+**Key invariants:**
+1. **Backup before mutate** — Stage 6 必须把待变更文件打包到 `.workflow/.trash/knowledge-audit-{timestamp}/`，备份失败禁止 Stage 7。
+2. **Deprecate over delete** — 文本存储（spec/knowhow）首选注入 `status="deprecated"` 保留历史；只有 artifact 物理残留才走 delete/purge。
+3. **Purge requires double confirmation** — `--purge` 仅作用于 artifact scope，且 Stage 5 必须显式 `[y/N]` 二次确认 + 输入 artifact id。
+4. **Rescue before delete** — 删除未抽取的 artifact 前（`harvest-log.jsonl` 无记录），强制提示 "是否先 `/manage-harvest`？"。
+
+Scope 路径、8 类检测算法、三态决策面板、报告 schema 定义在 workflow knowledge-audit.md。
+
+**Next-step routing on completion:**
+- 复审淘汰记录 → `.workflow/.knowledge-audit/audit-report-{date}.md`
+- 抢救未抽取 artifact → `/manage-harvest <artifact-id>`
+- 验证 spec 现状 → `/spec-load --role implement`
+- 周期巡检 → 每 milestone 结束跑 `--scope all --report`
+</execution>
+
+<error_codes>
+| Code | Severity | Condition | Recovery |
+|------|----------|-----------|----------|
+| E001 | error | `.workflow/` 未初始化 | 先跑 `/maestro-init` |
+| E002 | error | `--scope` 缺失或非法 | 提供 spec/knowhow/artifact/all |
+| E003 | error | `--purge` 与 `--dry-run` 同用 | 二选一 |
+| E004 | error | `--purge` 作用于非 artifact 范围 | purge 仅支持 artifact scope |
+| E005 | error | 备份失败（`.trash/` 写入异常） | 检查磁盘空间与权限，重试 |
+| W001 | warning | 检出冲突但用户选择 keep | 记入 report，不阻断 |
+| W002 | warning | 待删 artifact 无 harvest-log 记录 | 提示先跑 manage-harvest |
+| W003 | warning | 循环 supersedes 链 | 自动断环或交互选保留节点 |
+| W004 | warning | 检测耗时 >120s（大规模 spec 库） | 建议加 `--scope` 收敛或 `--since` 增量 |
+| W005 | warning | LLM detector 不可用 | 降级到正则+图算法子集，跳过 B/G 类语义场景 |
+</error_codes>
+
+<success_criteria>
+- [ ] Scope 正确解析，互斥标志校验通过
+- [ ] 三存储按 scope 加载完成，构建出统一 finding 池
+- [ ] Stage 3 时间线索引建立（mtime ↔ session/milestone 状态）
+- [ ] Stage 4 按 P0/P1/P2 输出 finding 列表
+- [ ] 如非 `--report`：用户对每项做出三态决策
+- [ ] 未 harvest 的 artifact 删除前触发抢救确认
+- [ ] Stage 6 backup tarball 生成于 `.workflow/.trash/`
+- [ ] `deprecate` 通过元数据注入完成（spec/knowhow 文件未被物理删除）
+- [ ] `delete` 移动至 `.trash/`，索引同步更新
+- [ ] `purge` 仅在双重确认通过后执行
+- [ ] `audit-report-{date}.md` + `audit-log.jsonl` 写入完成
+- [ ] 摘要展示三存储变更计数与下一步路由
+</success_criteria>