maestro-flow 0.4.10 → 0.4.11

package/.claude/commands/maestro-ralph.md

@@ -21,7 +21,7 @@ Entry points:
 - **`/maestro-ralph continue`** — Resume via maestro-ralph-execute
 - **`/maestro-ralph status`** — Display session progress
 
-Initial decomposition (S_DECOMPOSE): broad intents (重构/全面/迁移/重写) are boundary-clarified via ≤3 questions, producing 执行准则 + 子目标清单 written into status.json, plus a `goal-checklist.md` and a copy-paste `/goal` prompt for the user to bind.
+Initial decomposition (S_DECOMPOSE): broad intents (重构/全面/迁移/重写) are boundary-clarified via ≤3 questions, producing 执行准则 + 子目标清单 written into status.json, plus a `goal-checklist.md` (a rendered view of status.json) and a copy-paste `/goal` prompt for the user to bind.
 
 Three node types:
 - **internal**: `Skill()` call (synchronous, lightweight)
@@ -34,6 +34,24 @@ Key difference from maestro coordinator:
 
 Session: `.workflow/.maestro/ralph-{YYYYMMDD-HHmmss}/status.json`
 Mutual invocation with `/maestro-ralph-execute` forms a self-perpetuating work loop.
+
+### 执行方式 / Execution Flow
+
+```
+ /maestro-ralph "intent" ─▶ ralph        infer → decompose → build chain
+                              │           writes status.json (truth)
+                              │           renders goal-checklist.md (view)
+                              │           emits /goal prompt
+                              ▼
+                       ralph-execute  ◀─┐ runs next step:
+                              │         │  internal → Skill()
+                              │         │  external → delegate (STOP→callback)
+                              │         │  decision → Skill("maestro-ralph")
+                              └─────────┘ updates status.json + re-renders checklist
+                       loop until done | paused
+```
+
+`status.json` 是唯一真源；checklist 是渲染视图。决策节点是链路生长点（`post-goal-audit` 插入按子目标的 mini-loop；`post-verify/review/test` 插入修复 loop）。评估只走 ralph，执行只走 ralph-execute，两者不互相替代。
 </purpose>
 
 <context>
@@ -68,8 +86,9 @@ Remaining     → intent
 S_PARSE_ROUTE     — 解析参数、路由入口                  PERSIST: —
 S_STATUS          — 显示 session 进度                   PERSIST: —
 S_CONTINUE        — 恢复执行                            PERSIST: —
-S_INFER           — 读 state.json、推断生命周期位置      PERSIST: session.lifecycle_position
-S_RESOLVE_PHASE   — 解析目标 phase                      PERSIST: session.phase
+S_RESOLVE_PHASE   — 解析目标 phase + 标记 phase_is_new   PERSIST: session.phase, session.phase_is_new
+S_INFER           — 基于已解析 phase 推断 lifecycle_position PERSIST: session.lifecycle_position
+S_QUALITY_MODE    — 决定质量管线模式                     PERSIST: session.quality_mode
 S_DECOMPOSE       — 边界澄清、写执行准则+子目标清单       PERSIST: session.boundary_contract, .execution_criteria, .task_decomposition
 S_BUILD_CHAIN     — 构建步骤链                           PERSIST: session.steps[]
 S_CREATE_SESSION  — 写 status.json                      PERSIST: session (全量)
@@ -86,7 +105,7 @@ S_PARSE_ROUTE:
   → S_STATUS        WHEN: intent == "status"
   → S_CONTINUE      WHEN: intent == "continue"
   → S_DECISION_EVAL WHEN: running session with decision step in "running" status
-  → S_INFER         WHEN: intent is non-empty
+  → S_RESOLVE_PHASE WHEN: intent is non-empty                  ← phase 必须先于 position
   → S_FALLBACK      WHEN: no intent AND no running session
 
 S_STATUS:
@@ -96,15 +115,18 @@ S_CONTINUE:
   → S_DISPATCH      WHEN: running session found
   → S_FALLBACK      WHEN: no running session               DO: display "无运行中的 ralph 会话"
 
-S_INFER:
-  → S_RESOLVE_PHASE WHEN: position resolved                 DO: A_INFER_POSITION
-  → S_FALLBACK      WHEN: cannot infer
-
 S_RESOLVE_PHASE:
-  → S_DECOMPOSE     WHEN: phase resolved or null            DO: A_RESOLVE_PHASE
+  → S_INFER         WHEN: phase resolved or null            DO: A_RESOLVE_PHASE
   → S_FALLBACK      WHEN: ambiguous
                      GUARD: auto_confirm does NOT skip phase ambiguity
 
+S_INFER:
+  → S_QUALITY_MODE  WHEN: position resolved                 DO: A_INFER_POSITION
+  → S_FALLBACK      WHEN: cannot infer
+
+S_QUALITY_MODE:
+  → S_DECOMPOSE     DO: A_DETERMINE_QUALITY_MODE
+
 S_DECOMPOSE:
   → S_BUILD_CHAIN   DO: A_DECOMPOSE_TASKS
                      GUARD: broad intent (重构/全面/重写/迁移/overhaul/migrate/rewrite) → MUST clarify boundary even if auto_confirm
@@ -162,7 +184,36 @@ S_FALLBACK:
 1. Find latest ralph session (by created_at)
 2. Display: Session, Status, Position, Progress, Current step
 3. List steps: [✓] completed, [▸] current, [ ] pending, [◆] decision
-4. If `task_decomposition` present: show `Sub-goals: {done}/{total}` and any unmet G-ids (graceful skip if field absent — backward compat)
+4. If `task_decomposition` present (graceful skip if absent — backward compat):
+   ```
+   Sub-goals  ({done}/{total})    source: {session_dir}/status.json#/task_decomposition
+   [x] G1 done_when={done_when}   evidence={evidence}
+   [ ] G2 done_when={done_when}   evidence={evidence} ◀ unmet
+   ...
+   Checklist view: {session_dir}/goal-checklist.md  (regenerated from status.json)
+   ```
+   Reads state directly from `status.json.task_decomposition[*].status` — checklist is just a view, status.json wins on conflict.
+
+### A_RESOLVE_PHASE
+
+**前置于 A_INFER_POSITION**——position 推断需要先知道 target phase 是否在 state.json 中已存在。
+
+**Priority (产出 `phase` + `phase_is_new` 二元组):**
+
+| Step | 行为 | phase_is_new |
+|------|------|--------------|
+| 1 | intent 匹配 `phase\s*(\d+)` 正则 → 取 state.json 中对应 phase | false |
+| 2 | intent 派生短语（如 "docs-site-redesign", "auth-refactor"）→ 在 state.json.milestones[*].phases / artifacts[*].path 中查找 | false (匹配) / true (无匹配) |
+| 3 | 未派生 → 取最新 in-progress artifact 的 phase | false |
+| 4 | 仍无 → state.json 首个 incomplete phase | false |
+| 5 | position 将是 brainstorm/init/roadmap → phase = null | n/a |
+| 6 | 仍模糊 → `AskUserQuestion`（新 phase / 已存在 phase 二选一） | 由用户回答确定 |
+
+**写入 session**: `phase`（字符串或 null）+ `phase_is_new`（bool）。`phase_is_new=true` 表示当前 milestone 下需要为这个 phase 创建全套生命周期（A_INFER_POSITION 据此强制起始位置）。
+
+**新派生 phase 时 milestone 处理**：
+- 若 state.json 当前 milestone 仍 active → 沿用当前 milestone，仅新增 phase
+- 若 intent 同时派生了新 milestone 名（如 "M1-visual-replication"）→ 写入 session 但**仅作标签**；state.json.milestones 由后续 `maestro-roadmap` / `maestro-milestone-release` 真实创建。session 层面禁止虚构 milestone 直接改 state.json
 
 ### A_INFER_POSITION
 
@@ -175,18 +226,22 @@ S_FALLBACK:
 | No `.workflow/` + no source files | `brainstorm` |
 | No `.workflow/` + has source files | `init` |
 | Has `.workflow/` but no state.json | `init` |
-| Has state.json | → artifact-based inference |
+| Has state.json | → phase-aware artifact inference |
 
-**Artifact-based inference:** Filter by current_milestone + target phase:
+**Phase-aware artifact inference** (使用 A_RESOLVE_PHASE 已写入的 `session.phase` + `session.phase_is_new`)：
 
-| Latest artifact type | Position |
-|---------------------|----------|
+| Condition | Position |
+|-----------|----------|
+| `phase_is_new == true` (intent 派生的新 phase, state.json 中无) | **`analyze`** (强制从头起) |
 | no milestones or no roadmap.md | `roadmap` |
-| none for phase | `analyze` |
-| analyze | `plan` |
-| plan | `execute` |
-| execute | `verify` |
-| verify | → refine from result files |
+| `phase == null` (brainstorm/init/roadmap 已由 override 决定) | n/a |
+| phase 已存在 + 无任何 artifact | `analyze` |
+| phase 已存在 + 最新 artifact = analyze | `plan` |
+| phase 已存在 + 最新 artifact = plan | `execute` |
+| phase 已存在 + 最新 artifact = execute | `verify` |
+| phase 已存在 + 最新 artifact = verify | → refine from result files |
+
+**关键不变量**：artifact 过滤必须用 `session.phase`（A_RESOLVE_PHASE 已写入），而不是 state.json.current_phase。当 `phase_is_new` 时跳过过滤直接走 `analyze`，避免错用其他 phase 的 artifact 推断。
 
 **Refine from verify results:**
 
@@ -199,9 +254,17 @@ S_FALLBACK:
 | uat.md: all passed | `milestone-audit` |
 | uat.md: has failures | `test-failed` |
 
-### A_RESOLVE_PHASE
+### A_DETERMINE_QUALITY_MODE
 
-Priority: 1) regex from intent 2) latest artifact's phase 3) first incomplete phase 4) null if brainstorm/init/roadmap 5) AskUserQuestion if ambiguous
+决定下游质量管线长度。读 `session.quality_mode_override`（CLI 标志 `--quality`），无则按规则推断：
+
+| Condition | Mode | Pipeline (verify 之后) |
+|-----------|------|-------------------------|
+| Has `specs/REQ-*.md` + 当前 phase 业务范围明确 | `full` | business-test → review → test-gen → test |
+| Default | `standard` | review → test-gen (当 coverage<80%) → test |
+| `--quality quick` | `quick` | review --tier quick |
+
+写入 `session.quality_mode`。A_BUILD_STEPS 据此过滤 stage（见下）。
 
 ### A_DECOMPOSE_TASKS
 
@@ -242,37 +305,52 @@ narrow → derive defaults from intent + codebase, skip questions.
 
 ### A_BUILD_STEPS
 
-Generate steps from lifecycle_position to milestone-complete:
-
-| Stage | Skill | Type | Decision after |
-|-------|-------|------|----------------|
-| brainstorm | `maestro-brainstorm "{intent}"` | external | — |
-| init | `maestro-init` | internal | — |
-| roadmap | `maestro-roadmap "{intent}"` | internal | — |
-| analyze | `maestro-analyze {phase}` | external | — |
-| plan | `maestro-plan {phase}` | internal | — |
-| execute | `maestro-execute {phase}` | external | — |
-| verify | `maestro-verify {phase}` | internal | `post-verify` |
-| business-test | `quality-auto-test {phase}` | internal | `post-business-test` |
-| review | `quality-review {phase}` | internal | `post-review` |
-| test-gen | `quality-auto-test {phase}` | internal | — |
-| test | `quality-test {phase}` | internal | `post-test` |
-| milestone-audit | `maestro-milestone-audit` | internal | — |
-| goal-audit | *(decision-only, no skill)* | decision | `post-goal-audit` |
-| milestone-complete | `maestro-milestone-complete` | internal | `post-milestone` |
-
-`goal-audit` row is inserted **only when `task_decomposition` present**, immediately before `milestone-complete`. It is a pure decision node (no skill step) — its job is to re-check the goal-checklist and **dynamically grow `steps[]`** for unmet sub-goals.
+Generate steps from `session.lifecycle_position` to `milestone-complete`（终点硬约束）。
+
+| Stage | Skill | Type | Decision after | quality_mode |
+|-------|-------|------|----------------|--------------|
+| brainstorm | `maestro-brainstorm "{intent}"` | external | — | all |
+| init | `maestro-init` | internal | — | all |
+| roadmap | `maestro-roadmap "{intent}"` | internal | — | all |
+| analyze | `maestro-analyze {phase}` | external | — | all |
+| plan | `maestro-plan {phase}` | internal | — | all |
+| execute | `maestro-execute {phase}` | external | — | all |
+| verify | `maestro-verify {phase}` | internal | `post-verify` | all |
+| business-test | `quality-auto-test {phase}` | internal | `post-business-test` | **full only** |
+| review | `quality-review {phase}` | internal | `post-review` | all (quick: append `--tier quick`) |
+| test-gen | `quality-auto-test {phase}` | internal | — | full / standard if coverage<80% |
+| test | `quality-test {phase}` | internal | `post-test` | full, standard |
+| milestone-audit | `maestro-milestone-audit` | internal | — | all |
+| goal-audit | *(decision-only, no skill)* | decision | `post-goal-audit` | all (only if decomposed) |
+| milestone-complete | `maestro-milestone-complete` | internal | `post-milestone` | all (chain 终点) |
 
 Type rationale: `internal` = Skill(), lightweight/interactive; `external` = delegate --to claude, context-isolated heavy computation
 
-Build rules: start from position, skip completed, insert decision nodes with `{ retry_count: 0, max_retries: 2 }`, args use placeholders resolved at execution time by ralph-execute. Steps dynamically inserted by `post-goal-audit` carry optional `goal_ref: "G{n}"` tracing which sub-goal they serve.
+**Build rules (按顺序应用):**
+
+1. **起点**：从 `session.lifecycle_position` 开始；不读 `state.json.current_phase`
+2. **跳过已完成**：跳过当前 milestone+phase 下已有 completed artifact 的 stage（artifact 过滤同样按 `session.phase`，不按 state.json 当前 phase）
+3. **quality_mode 过滤**：按 `session.quality_mode` 排除上表 `quality_mode` 列不包含该模式的 stage（如 standard 不跑 business-test、quick 不跑 test-gen/test）
+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 }`
+5. **goal-audit 插入**：当且仅当 `task_decomposition` 存在 → 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 必须以 `milestone-complete` 步骤结尾（除非 `lifecycle_position` 已是 `milestone-complete` 之后的状态）。生成器在收尾时不论 quality_mode 都必须 append 该 step
+7. **初始 goal_ref 传播**：若 `task_decomposition` 存在，遍历每个新建 step：
+   - 计算 `step.stage`（如 execute/verify/review）
+   - 对每个 `g ∈ task_decomposition`：若 `step.stage ∈ g.lifecycle` → `step.goal_ref = g.id`
+   - 多 G 匹配 → 取 id 字典序最小者；保证 verify/review 等共用 lifecycle 的 step 也有可追溯标签
+   - decision 节点不打 goal_ref（goal-audit 自身除外，它的 verdict 直接修改 task_decomposition）
+8. **占位符**：args 用 `{phase}` `{intent}` 等，由 ralph-execute 在执行时解析
+9. **动态插入**：`post-goal-audit` 触发 A_APPLY_GOAL_FIX 时插入的步骤同样按规则 7 打 `goal_ref`（追溯到触发它的子目标 G{n}）
 
 ### A_CREATE_SESSION
 
-1. Write `.workflow/.maestro/ralph-{YYYYMMDD-HHmmss}/status.json` (see Appendix: Session Schema) — include decomposition fields only if produced (additive)
-2. If `task_decomposition` present: write `{session_dir}/goal-checklist.md` (see Appendix: Goal Checklist Template) — stable within session, referenced verbatim by `/goal`
+1. Write `.workflow/.maestro/ralph-{YYYYMMDD-HHmmss}/status.json` (Appendix: Session Schema) — decomposition fields included only if produced (additive)
+2. If `task_decomposition` present:
+   - Set `session.goal_checklist_path = "{session_dir}/goal-checklist.md"` in status.json
+   - **Render** checklist from status.json (Appendix: Goal Checklist Template + Sync Rule). This is a one-way projection — status.json drives content; the file is never hand-edited
+   - Stable filename within session (so `/goal` condition string survives context compaction)
 3. Display chain overview with step list
-4. If `task_decomposition` present: display the **Goal Prompt block** (Appendix: Goal Prompt Template) — the copy-paste `/goal …` line binding the checklist as a Stop-hook target
+4. If `task_decomposition` present: display the **Goal Prompt block** (Appendix: Goal Prompt Template) — the copy-paste `/goal …` line binds status.json + checklist as a Stop-hook target
 
 ### A_DELEGATE_EVALUATE
 
@@ -311,20 +389,37 @@ Build rules: start from position, skip completed, insert decision nodes with `{
 
 ### A_GOAL_AUDIT_EVALUATE
 
-Re-checks the goal-checklist and decides whether `steps[]` must dynamically grow. Only runs when `task_decomposition` present.
+Re-checks sub-goals against `status.json` (source of truth) and decides whether `steps[]` must dynamically grow. Only runs when `task_decomposition` present.
 
-1. Read `session.task_decomposition` + `goal_checklist_path`
+1. Read `session.task_decomposition` from status.json (NOT from checklist — checklist is just a view)
 2. For each sub-goal `status != "done"`: resolve its `evidence` artifact (verification.json / review.json / uat.md / test path) under the current phase scratch dir
 3. Delegate read-only audit (run_in_background, STOP, wait):
    ```
-   maestro delegate "PURPOSE: 审计子目标达成情况，决定是否需要补充执行步骤
-   TASK: 逐个读取每个未完成子目标的 evidence 产物 | 对照 done_when 判定 met/unmet | 给出每个 unmet 子目标的差距
-   CONTEXT: @{goal_checklist_path} @{evidence artifacts} | 执行准则: {execution_criteria} | 边界: {boundary_contract}
-   EXPECTED: ---VERDICT--- STATUS(all_met|has_unmet) / UNMET=[{id:G2,gap:'...',target_phase:execute}] / CONFIDENCE_SCORE(0-100) ---END---
-   CONSTRAINTS: 只评估不修改 | 严格按 done_when 判定 | 不得超出 boundary_contract"
+   maestro delegate "PURPOSE: 审计未完成子目标，判定哪些已达成、哪些仍需补步骤
+   TASK:
+     1. 读取 status.json.task_decomposition 中每个 status!=done 的子目标
+     2. 打开其 evidence 产物，对照 done_when 严格判定
+     3. 输出 met / unmet 分类，unmet 给出差距 + 应回补的 target_phase
+   CONTEXT:
+     status.json   = {session_dir}/status.json
+     checklist     = {goal_checklist_path}        (人类可读视图，仅供参考)
+     evidence      = {evidence artifacts}
+     执行准则      = {execution_criteria}
+     边界契约      = {boundary_contract}
+   EXPECTED (单行 verdict 块，严格遵循):
+     ---VERDICT---
+     STATUS=all_met|has_unmet
+     UNMET=[{id:G2,gap:'...',target_phase:execute}, ...]   # 空数组当 STATUS=all_met
+     CONFIDENCE_SCORE=0-100
+     ---END---
+   CONSTRAINTS:
+     - 只评估，不修改任何文件
+     - 严格按 done_when 判定；evidence 缺失 → 视为 unmet
+     - 不得建议超出 boundary_contract 的修改
+   "
    --role analyze --mode analysis
    ```
-4. On callback: parse UNMET list. For each met sub-goal → set `task_decomposition[i].status="done"` + flip `[ ]→[x]` in goal-checklist.md
+4. On callback: parse UNMET list. **status.json is the write target** — for each met sub-goal: `task_decomposition[i].status="done"` + `task_decomposition[i].completed_at=now`. Then regenerate checklist view from status.json (see Sync Rule in Appendix).
 5. **Decision log**: append to `{session_dir}/decisions.ndjson` with `"type": "goal-gate"`, `unmet_count`, `unmet_ids`
 6. Verdict: `all_met` → A_APPLY_GOAL_DONE; `has_unmet` → A_APPLY_GOAL_FIX
    GUARD: retry_count >= max_retries AND still unmet → A_APPLY_ESCALATE (insert quality-debug, hand to human)
@@ -357,10 +452,10 @@ Re-checks the goal-checklist and decides whether `steps[]` must dynamically grow
 
 ### A_APPLY_GOAL_DONE
 
-1. Set all `task_decomposition[*].status="done"`, write status.json
-2. Append `ALL_GOALS_DONE` sentinel line to goal-checklist.md (satisfies the user's `/goal` Stop hook)
+1. Write status.json (source of truth): set every `task_decomposition[*].status="done"`, `completed_at=now`, plus top-level `task_decomposition_all_done=true`
+2. Regenerate `goal-checklist.md` from status.json (Sync Rule in Appendix) — all boxes flip to `[x]`, sentinel `ALL_GOALS_DONE` appended at file end
 3. Mark goal-audit decision completed; proceed to `milestone-complete`
-4. Display: ◆ Goal audit: 全部子目标达成 ✓ — checklist 已写入 ALL_GOALS_DONE
+4. Display: ◆ Goal audit: 全部子目标达成 ✓ — status.json + checklist 已同步 ALL_GOALS_DONE
 
 ### A_ADVANCE_MILESTONE
 
@@ -387,8 +482,9 @@ Re-checks the goal-checklist and decides whether `steps[]` must dynamically grow
   "session_id": "ralph-{YYYYMMDD-HHmmss}",
   "source": "ralph", "status": "running",
   "intent": "", "lifecycle_position": "",
-  "phase": null, "milestone": "",
-  "auto_mode": false, "quality_mode": "standard",
+  "phase": null, "phase_is_new": false, "milestone": "",
+  "auto_mode": false,
+  "quality_mode": "standard",   // "full" | "standard" | "quick" — 由 A_DETERMINE_QUALITY_MODE 写入
   "cli_tool": "claude", "passed_gates": [],
   "context": { "issue_id": null, "scratch_dir": null, "plan_dir": null,
     "analysis_dir": null, "brainstorm_dir": null },
@@ -404,15 +500,18 @@ Re-checks the goal-checklist and decides whether `steps[]` must dynamically grow
   "execution_criteria": [],
   "task_decomposition": [
     { "id": "G1", "goal": "", "boundary": "", "done_when": "",
-      "evidence": "", "lifecycle": [], "status": "pending|done" }
+      "evidence": "", "lifecycle": [], "status": "pending|done",
+      "completed_at": null }
   ],
-  "goal_checklist_path": ""
+  "task_decomposition_all_done": false,
+  "goal_checklist_path": "",
+  "goal_checklist_synced_at": null
 }
 ```
 
-> **Extensibility contract (two dimensions)**:
-> 1. **Schema-additive** — decomposition block fields are optional; absence = old behavior.
-> 2. **Step-dynamic** — `steps[]` is a *living array*: `post-goal-audit` (and existing fix/escalate/milestone decisions) **append/reindex steps at runtime** until sub-goals converge. The JSON "extends" primarily by growing `steps[]`, not by freezing a plan. `goal_ref` (optional, default null) traces dynamically-added steps back to the sub-goal that spawned them.
+**扩展约定**：
+- Schema 加字段 = 可选（缺省 = 旧行为）；不删/不改既有字段名。
+- `steps[]` 是活数组：`post-goal-audit` 与 `post-verify/review/test` 等决策节点按需追加+重排，子目标收敛即停。`goal_ref`（可选）回溯每条动态插入步骤所属子目标。
 
 ### Fix-Loop Templates
 
@@ -472,48 +571,76 @@ decision:post-goal-audit {retry+1}
 ```
 Notes: only unmet sub-goals' phases are re-run (no full-pipeline replay); inserted steps carry `goal_ref` for traceability; loop exits when audit returns `all_met` (→ A_APPLY_GOAL_DONE) or retry hits max (→ escalate to human). This keeps growth bounded.
 
-### Goal Checklist Template
+### Goal Checklist Template (status.json projection)
 
-Written to `{session_dir}/goal-checklist.md`. Stable within the session; never renamed (so the `/goal` condition string stays valid across context compaction).
+`{session_dir}/goal-checklist.md` 是 `status.json` 的只读投影：不要手改，永远从 JSON 重渲染。文件名在 session 内稳定，`/goal` 条件字符串跨上下文压缩仍可用。
 
 ```markdown
 # Ralph Goal Checklist — {session_id}
-> Intent: {intent}
+<!-- AUTO-GENERATED from status.json. Source: ../status.json#/task_decomposition -->
+
+> Intent    : {intent}
+> Source    : `{session_dir}/status.json` (authoritative)
+> Last sync : {ISO}    Phase: {phase}    Milestone: {milestone}
+
+## Resume / 恢复入口
+**不要直接执行 skill，调用入口：**
+```
+/maestro-ralph continue
+```
+ralph 评估下一步门控，ralph-execute 跑实际步骤。Stop hook 触发时也走这条路径。
 
 ## 执行准则 / Execution Criteria
+<!-- status.json#/execution_criteria -->
 - {criterion 1}
 - {criterion 2}
 
 ## 边界契约 / Boundary Contract
-- In scope: {in_scope}
-- Out of scope: {out_of_scope}
-- Constraints: {constraints}
-- Definition of Done: {definition_of_done}
-
-## 子目标 / Sub-goals
-- [ ] G1: {goal} — done when: {done_when} (evidence: {evidence})
-- [ ] G2: {goal} — done when: {done_when} (evidence: {evidence})
-- [ ] G3: ...
-
-<!-- ralph-execute flips [ ]→[x] when a sub-goal's evidence artifact confirms done;
-     appends the line `ALL_GOALS_DONE` once every box is [x]. -->
+<!-- status.json#/boundary_contract -->
+- In scope           : {in_scope}
+- Out of scope       : {out_of_scope}
+- Constraints        : {constraints}
+- Definition of Done : {definition_of_done}
+
+## 子目标 / Sub-goals    ({done}/{total})
+<!-- status.json#/task_decomposition -->
+- [ ] G1 — {goal}
+       done_when : {done_when}
+       evidence  : {evidence}
+       lifecycle : {lifecycle}
+       <!-- ref: status.json#/task_decomposition/0  status=pending -->
+- [x] G2 — {goal}
+       done_when : {done_when}
+       evidence  : {evidence}
+       <!-- ref: status.json#/task_decomposition/1  status=done  completed_at={ISO} -->
+- [ ] G3 — ...
+
+<!-- A_APPLY_GOAL_DONE 在全部 status=done 时在文件末尾追加 ALL_GOALS_DONE -->
 ```
 
-`maestro-ralph-execute` responsibility (additive, optional): after a step whose `lifecycle` covers a sub-goal, re-check that sub-goal's `evidence` artifact; if satisfied, set `task_decomposition[i].status="done"` in status.json AND flip its checkbox in goal-checklist.md. When all done → append `ALL_GOALS_DONE`. If decomposition fields absent → do nothing (old behavior).
+### Sync Rule (status.json ↔ goal-checklist.md)
+
+单向投影；status.json 为真源；冲突时直接重渲染，不合并。
+
+1. 只有 ralph / ralph-execute 写 `task_decomposition[*].status`；markdown 不可写入状态。
+2. `task_decomposition` / `execution_criteria` / `boundary_contract` 任一变化 → 立即重渲染 checklist。
+3. 全部 `status="done"` → 在文件末追加 `ALL_GOALS_DONE` 哨兵行。
+4. 检测到 checklist 缺失或与 JSON 漂移 → 直接重渲染覆盖；视 markdown 为可丢弃产物。
+5. ralph-execute（可选、向后兼容）：完成 `lifecycle` 覆盖某子目标的 step 后，校验 evidence；满足则置 `status="done"` + 重渲染。无 `task_decomposition` 字段 → 不动作。
 
 ### Goal Prompt Template
 
-Displayed verbatim after the chain overview (only when decomposition produced):
+链路概览后逐字显示（仅当 decomposition 已产出）：
 
 ```
-📋 任务分解完成。可选(推荐)：复制下面一行设定目标，让会话在所有子目标达成前不停：
+📋 任务分解完成。复制下面一行设定目标，会话在子目标全部达成前不停：
 
-/goal 当 {session_dir}/goal-checklist.md 中所有子目标复选框均为 [x] 且文件含 ALL_GOALS_DONE 哨兵时目标达成；否则依据文件内"执行准则"继续推进未完成子目标，不得超出边界契约范围
+/goal 目标达成条件: {session_dir}/status.json 中 task_decomposition[*].status 全部为 "done"（等价: {session_dir}/goal-checklist.md 末尾含 ALL_GOALS_DONE）。未达成时: 阅读 {session_dir}/goal-checklist.md 取得"执行准则/边界契约/子目标"作为行动手册, 然后调用 /maestro-ralph continue 推进下一步; 严禁手动执行 skill 或越界修改 status.json.boundary_contract.out_of_scope。
 
-然后运行 /maestro-ralph continue 开始执行。
+随后运行 /maestro-ralph continue 立即开始执行。
 ```
 
-> `/goal` 是 harness 命令、仅用户可输入 — ralph 只输出此提示词，不能自行注册。这与 status.json 的可扩展约定一致：绑定是用户侧、可选的增量能力。
+`/goal` 是 harness 命令，仅用户能输入；ralph 只能输出此提示词。判据以 status.json 为权威，哨兵为等价信号，避免视图漂移误判。
 
 ### Error Codes
 
@@ -531,14 +658,22 @@ Displayed verbatim after the chain overview (only when decomposition produced):
 
 ### Success Criteria
 
+- [ ] Phase 先于 position 解析（S_RESOLVE_PHASE → S_INFER → S_QUALITY_MODE）；phase_is_new 标记写入 session
+- [ ] phase_is_new=true 时 lifecycle_position 强制为 `analyze`（禁止用其他 phase 的 artifact 推断）
+- [ ] artifact 过滤始终按 session.phase（不读 state.json.current_phase）
+- [ ] quality_mode 显式由 A_DETERMINE_QUALITY_MODE 决定（full/standard/quick），过滤 build steps
 - [ ] State parsed, position inferred from bootstrap + artifacts + result files
 - [ ] Decomposition runs as initial step; broad intent boundary-clarified via ≤3 questions (ignores auto_confirm); narrow auto-derives
 - [ ] status.json enriched additively with boundary_contract + execution_criteria + task_decomposition; absent fields = old behavior preserved
-- [ ] goal-checklist.md generated with verifiable done_when mapped to ralph evidence + ALL_GOALS_DONE sentinel
-- [ ] Goal Prompt emitted for user to bind via /goal
+- [ ] status.json is single source of truth; goal-checklist.md is a regenerated projection (Sync Rule), never hand-edited
+- [ ] goal-checklist.md carries explicit `status.json#/task_decomposition[i]` JSON-pointer refs per entry + last-sync timestamp
+- [ ] Mutations to task_decomposition trigger checklist re-render; ALL_GOALS_DONE sentinel appended only when all entries status=done
+- [ ] Goal Prompt names status.json as the authoritative judgement source, with ALL_GOALS_DONE sentinel as equivalent signal
 - [ ] post-goal-audit decision node inserted before milestone-complete (only when decomposed)
 - [ ] Unmet sub-goals DYNAMICALLY grow steps[] via scoped per-goal mini-loops (goal_ref tagged), looping until all_met or max retries → escalate
-- [ ] Quality pipeline generated: verify → business-test → review → test-gen → test
+- [ ] Quality pipeline 按 quality_mode 生成（full=全管线 / standard=skip business-test / quick=仅 review --tier quick）
+- [ ] Chain 末端硬约束：必须以 `milestone-complete` 结尾（goal-audit decision 紧前）
+- [ ] 初始构建按 task_decomposition[*].lifecycle 给每个 step 打 goal_ref（verify/review 等共用 lifecycle 的 step 也有标签）
 - [ ] Decision nodes delegate-evaluated via maestro delegate --role analyze
 - [ ] Verdict parsed with confidence adjustment
 - [ ] Fix-loop templates applied with retry tracking

package/.claude/commands/maestro.md

@@ -138,18 +138,20 @@ S_FALLBACK:
 
 ### A_DECOMPOSE_TASKS
 
-Shares the decomposition contract with maestro-ralph `A_DECOMPOSE_TASKS` — **reference that spec; do not duplicate the table logic here.** Condensed:
-
-1. Classify intent breadth (broad: 重构/全面/重写/迁移/overhaul/migrate/rewrite; narrow: single file/function/bug). Skip entirely for narrow / single-step / {status,init,quick} chains
-2. Broad/medium → `AskUserQuestion` ≤3 rounds: Scope (in/out) | Constraints (compat/API/perf/test bar) | Definition of Done
-3. Derive `execution_criteria` (3-6 imperative rules) + `task_decomposition` (outcome sub-goals, each `done_when` objectively verifiable and mapped to a ralph evidence artifact: verification.json / review.json / uat.md / test path)
-4. Write `{session_dir}/goal-checklist.md` (same template as maestro-ralph) with `ALL_GOALS_DONE` sentinel; set `goal_checklist_path`
-5. Append a `{ type: "decision", decision: "post-goal-audit", retry_count: 0, max_retries: 2 }` node as the FINAL node — after the last evidence-producing step (verify/review/test), before a milestone-complete/close-out step if the chain ends with one. (Audit needs evidence artifacts to exist.) ralph-execute then dynamically grows `steps[]` for unmet sub-goals
-6. **Emit the `/goal` bind prompt:**
+与 maestro-ralph `A_DECOMPOSE_TASKS` 共享分解契约 —— 引用该规范，不重复表格。Condensed:
+
+1. 分类意图广度（broad: 重构/全面/重写/迁移/overhaul/migrate/rewrite；narrow: 单文件/函数/bug）。narrow / 单步 / `{status,init,quick}` 链直接跳过
+2. broad/medium → `AskUserQuestion` ≤3 轮：Scope（in/out）| Constraints（兼容/API/perf/test）| Definition of Done
+3. 派生 `execution_criteria`（3-6 条命令式规则）+ `task_decomposition`（outcome 子目标，`done_when` 客观可校验，绑定到 ralph 产物：verification.json / review.json / uat.md / test path）
+4. **status.json 为唯一真源**：写入 `boundary_contract` / `execution_criteria` / `task_decomposition` / `goal_checklist_path`；从 status.json 投影渲染 `{session_dir}/goal-checklist.md`（同 maestro-ralph 的 Sync Rule，含 Resume 区块指向 `/maestro-ralph continue`）
+5. 在链路末尾（evidence 产出步骤之后、milestone-complete/close-out 之前）追加 `{ type: "decision", decision: "post-goal-audit", retry_count: 0, max_retries: 2 }`。ralph-execute 在该节点动态生长 `steps[]` 以收敛未达成子目标
+6. **输出 `/goal` 绑定提示词（参考文档 + 内嵌 ralph 调用）：**
    ```
-   📋 任务分解完成。复制下面一行设定目标(推荐)：
-   /goal 当 {session_dir}/goal-checklist.md 中子目标全 [x] 且含 ALL_GOALS_DONE 时达成；否则按执行准则继续推进且不越边界契约
+   📋 任务分解完成。复制下面一行设定目标，会话在子目标全部达成前不停：
+
+   /goal 目标达成条件: {session_dir}/status.json 中 task_decomposition[*].status 全部为 "done"（等价: {session_dir}/goal-checklist.md 末尾含 ALL_GOALS_DONE）。未达成时: 阅读 {session_dir}/goal-checklist.md 取得"执行准则/边界契约/子目标"作为行动手册, 然后调用 /maestro-ralph continue 推进下一步; 严禁手动执行 skill 或越界。
    ```
+   `/goal` 仅用户可输入；判据以 status.json 为权威，哨兵为等价信号。
 
 ### A_CREATE_SESSION
 
@@ -207,7 +209,7 @@ Unlisted commands have no auto flags.
 
 - [ ] Intent classified with task_type, complexity, clarity_score
 - [ ] Broad lifecycle intents decomposed (S_DECOMPOSE, ≤3 boundary questions) sharing maestro-ralph contract; narrow/single-step skip
-- [ ] Decomposition writes additive block + goal-checklist.md + post-goal-audit node; emits /goal bind prompt (no self-call, no phantom create_goal)
+- [ ] status.json 为唯一真源；goal-checklist.md 为投影视图（Resume 区块内嵌 `/maestro-ralph continue`）；post-goal-audit 节点追加；/goal 提示词内含"读 checklist + 调 /maestro-ralph continue"双重指引
 - [ ] Chain selected and confirmed (or auto-confirmed)
 - [ ] Session dir created with status.json before execution; decomposition fields additive-only
 - [ ] Tracking via TodoWrite (Claude); status.json steps[] grown dynamically by ralph-execute post-goal-audit

package/.codex/skills/maestro/SKILL.md

@@ -251,18 +251,18 @@ Read `.workflow/state.json` and route by condition:
 
 ### A_DECOMPOSE_TASKS
 
-Shares the decomposition contract with maestro-ralph `A_DECOMPOSE_TASKS` — **reference that spec; do not duplicate.** Condensed:
+与 maestro-ralph `A_DECOMPOSE_TASKS` 共享分解契约 —— 引用该规范，不重复。Condensed:
 
-1. Classify intent breadth. Skip for narrow / single-step / {status,init,quick} chains
-2. Broad/medium → `AskUserQuestion` ≤3 rounds: Scope (in/out) | Constraints (compat/API/perf/test bar) | Definition of Done
-3. Derive `execution_criteria` (3-6 imperative rules) + `task_decomposition` (outcome sub-goals; each `done_when` objectively verifiable, mapped to a ralph evidence artifact: verification.json / review.json / uat.md / test path)
-4. Write `{session_dir}/goal-checklist.md` (same template as maestro-ralph) with `ALL_GOALS_DONE` sentinel; set `goal_checklist_path`
-5. Append `{ type: "decision", decision: "post-goal-audit", retry_count: 0, max_retries: 2 }` as the FINAL node — after the last evidence-producing step (verify/review/test), before a milestone-complete/close-out step if present (audit needs evidence artifacts) → dynamic step growth for unmet sub-goals
+1. 分类意图广度。narrow / 单步 / `{status,init,quick}` 链跳过
+2. broad/medium → `AskUserQuestion` ≤3 轮：Scope（in/out）| Constraints（兼容/API/perf/test）| Definition of Done
+3. 派生 `execution_criteria`（3-6 条命令式规则）+ `task_decomposition`（outcome 子目标，`done_when` 可校验，绑定到 ralph evidence 产物）
+4. **status.json 为唯一真源**：写入 `boundary_contract` / `execution_criteria` / `task_decomposition` / `goal_checklist_path`；从 status.json 投影渲染 `{session_dir}/goal-checklist.md`（同 maestro-ralph Sync Rule，含 Resume 区块指向 `$maestro --continue`）
+5. 链路末尾（evidence 产出步骤后、milestone-complete/close-out 前）追加 `{ type: "decision", decision: "post-goal-audit", retry_count: 0, max_retries: 2 }`。S_DECISION_EVAL 在该节点动态生长 `steps[]`
 6. **Register goal via `create_goal`:**
    ```
    create_goal({ objective: "Maestro {chain}: {intent} — converge {N} sub-goals within boundary",
      success_criteria: task_decomposition.map(g => `${g.id}: ${g.done_when}`),
-     constraints: [...execution_criteria, "stay within boundary_contract"] })
+     constraints: [...execution_criteria, "stay within boundary_contract; resume via $maestro --continue"] })
    ```
 
 ### A_CREATE_SESSION
@@ -277,9 +277,13 @@ Shares the decomposition contract with maestro-ralph `A_DECOMPOSE_TASKS` — **r
      "steps": [{ "index", "type": "skill|decision", "skill", "args", "status": "pending", "goal_ref": null }],
      "waves": [], "current_step": 0, "status": "running",
      "_comment": "↓ OPTIONAL additive block — present only if S_DECOMPOSE ran; absent → flat-chain behavior",
-     "boundary_contract": {}, "execution_criteria": [], "task_decomposition": [], "goal_checklist_path": "" }
+     "boundary_contract": {}, "execution_criteria": [],
+     "task_decomposition": [{ "id": "G1", "goal": "", "done_when": "", "evidence": "", "status": "pending|done", "completed_at": null }],
+     "task_decomposition_all_done": false,
+     "goal_checklist_path": "", "goal_checklist_synced_at": null }
    ```
    Decomposition fields written ONLY if A_DECOMPOSE_TASKS produced them (additive)
+4. goal-checklist.md 是 status.json 的投影：从 task_decomposition 渲染（同 maestro-ralph Sync Rule）；Resume 区块写入 `$maestro --continue`
 4. Initialize tracking:
    - If decomposed: goal already registered by A_DECOMPOSE_TASKS. Else: `create_goal({ objective: "Maestro {chain}: {N} steps [{skill list}]" })`
    - `update_plan({ plan: steps.map(step => ({ step, status: "pending" })) })`
@@ -310,13 +314,13 @@ Direct in-context skill invocation — **replaces the old spawn/wave/CSV mechani
 
 ### A_GOAL_AUDIT_EVALUATE
 
-Entry action of S_DECISION_EVAL — mirrors maestro-ralph `A_GOAL_AUDIT_EVALUATE` (reference that spec; do not duplicate). Condensed:
+S_DECISION_EVAL 入口；镜像 maestro-ralph `A_GOAL_AUDIT_EVALUATE`。Condensed:
 
-1. Read `session.task_decomposition` + `goal_checklist_path`
-2. For each sub-goal `status != "done"`: resolve its `evidence` artifact under current phase scratch dir
-3. Delegate read-only audit (`maestro delegate --role analyze --mode analysis`): for each unmet sub-goal, read evidence, judge against `done_when`, return `STATUS(all_met|has_unmet) / UNMET=[{id,gap,target_phase}]`
-4. For each met sub-goal → set `task_decomposition[i].status="done"` + flip `[ ]→[x]` in goal-checklist.md; persist
-5. Produce verdict (`all_met` / `has_unmet`) consumed by S_DECISION_EVAL transition. GUARD: retry >= max_retries AND still unmet → escalate
+1. 读 `session.task_decomposition`（status.json，真源）
+2. 对每个 `status != "done"` 的子目标：解析 `evidence` 产物
+3. `maestro delegate --role analyze --mode analysis` 读取 evidence、对照 done_when 判定，返回 `STATUS=all_met|has_unmet / UNMET=[{id,gap,target_phase}] / CONFIDENCE_SCORE`
+4. status.json 为写入目标：每个达成子目标 `status="done"` + `completed_at=now`；然后从 status.json 重渲染 checklist（Sync Rule）
+5. Verdict（`all_met` / `has_unmet`）由 S_DECISION_EVAL 消费。GUARD: retry >= max_retries AND still unmet → escalate
 
 ### A_APPLY_GOAL_FIX
 
@@ -324,9 +328,10 @@ Entry action of S_DECISION_EVAL — mirrors maestro-ralph `A_GOAL_AUDIT_EVALUATE
 
 ### A_APPLY_GOAL_DONE
 
-1. Set all `task_decomposition[*].status="done"`, persist; append `ALL_GOALS_DONE` to goal-checklist.md
-2. `update_goal({ status: "complete" })` — release decomposition goal
-3. Proceed to chain's terminal step
+1. status.json（真源）：全部 `task_decomposition[*].status="done"` + `completed_at=now` + `task_decomposition_all_done=true`
+2. 从 status.json 重渲染 `goal-checklist.md`（见 Sync Rule）；文件末追加 `ALL_GOALS_DONE`
+3. `update_goal({ status: "complete" })` —— 释放分解 goal
+4. 继续到 chain 的终结步骤
 
 ### A_FINALIZE
 
@@ -464,6 +469,7 @@ Entry action of S_DECISION_EVAL — mirrors maestro-ralph `A_GOAL_AUDIT_EVALUATE
 - [ ] Intent classified and chain resolved
 - [ ] Broad lifecycle intents decomposed (S_DECOMPOSE, ≤3 boundary questions) sharing maestro-ralph contract; narrow/single-step skip
 - [ ] Goal registered via built-in `create_goal`; status.json decomposition fields additive-only
+- [ ] status.json 为唯一真源；goal-checklist.md 是从 JSON 重渲染的投影视图（Resume 区块指向 `$maestro --continue`）
 - [ ] post-goal-audit node appended as final node (after evidence-producing steps); unmet sub-goals dynamically grow steps[] (goal_ref tagged)
 - [ ] Session dir initialized with status.json before first step
 - [ ] Every skill invoked DIRECTLY in-context — NO spawn_agents_on_csv, NO wave/CSV/worker