maestro-flow 0.5.45 → 0.5.47

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

@@ -39,8 +39,19 @@ $ARGUMENTS — natural language description, or flags.
 1. **Architecture specs**: Run `maestro load --type spec --category arch` to load architecture constraints. Use as context for node resolution — ensures workflow design respects documented patterns.
 2. **Coding specs**: Run `maestro load --type spec --category coding` to load coding conventions. Informs executor argument defaults and context injection.
 3. Optional — proceed without if unavailable.
+
+**Output boundary**: ALL file writes MUST target `~/.maestro/templates/workflows/` (template JSON + index.json) and `.workflow/templates/design-drafts/` (intermediate drafts) only. NEVER modify source code or `.claude/commands/` files.
 </context>
 
+<invariants>
+1. **Max 20 nodes** — workflow templates MUST NOT exceed 20 nodes; larger workflows MUST be split into sub-workflows
+2. **Acyclic DAG** — generated template MUST be a directed acyclic graph; cycles MUST be detected and rejected before persistence
+3. **No orphan nodes** — every node MUST be reachable from at least one edge; unreachable nodes MUST be flagged
+4. **Confirmation before write** — template JSON MUST be shown to user via AskUserQuestion before writing to `~/.maestro/templates/workflows/`; cancellation saves draft only
+5. **Draft preservation** — abandoning at any confirmation gate MUST save current progress to design-drafts directory; NEVER discard user's partial work
+6. **Deferred reading only** — node-catalog and template-schema MUST be read at their designated phases (P2, P5), NEVER loaded eagerly at startup
+</invariants>
+
 <state_machine>
 
 <states>

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

@@ -30,10 +30,35 @@ $ARGUMENTS — Parse subcommand and optional path argument.
 
 **Enforcement:** The `workflow-guard` hook (PreToolUse on Write/Edit) reads this config
 and blocks operations targeting files outside boundaries. Requires hooks level >= `full`.
+
+**Output boundary**: ALL file writes MUST target `.workflow/config.json` (guard section) only. NEVER modify hook files, `.claude/settings.json`, or source code.
 </context>
 
+<invariants>
+1. **Config-only mutation** — guard MUST only modify the `guard` section of `.workflow/config.json`; NEVER touch other config sections or files
+2. **Non-destructive** — `off` MUST preserve existing paths and mode; NEVER clear the path list when disabling
+3. **Mode switch confirmation** — switching between allow/deny mode MUST require AskUserQuestion confirmation when existing paths will be cleared
+4. **Hook dependency** — guard MUST warn when enabled but `workflow-guard` hook is not active (hooks level < full)
+5. **Path normalization** — all paths MUST use forward slashes with trailing slash for directories; NEVER store raw backslash paths
+</invariants>
+
 <execution>
 
+### Phase Gates (MANDATORY, BLOCKING)
+
+**GATE 1: Parse → Config Read**
+- REQUIRED: Subcommand parsed (on/off/status/allow/deny) or defaulted to `status`.
+- BLOCKED if: invalid subcommand provided.
+
+**GATE 2: Config Read → Execute**
+- REQUIRED: `.workflow/config.json` read successfully or initialized with empty guard section.
+- BLOCKED if: file unreadable and cannot be created (E001).
+
+**GATE 3: Execute → Confirm**
+- REQUIRED: Config mutation applied (for on/off/allow/deny) or status displayed (for status).
+- REQUIRED: Mode-switch AskUserQuestion answered (for allow↔deny transitions with existing paths).
+- BLOCKED if: user declines mode switch.
+
 **Step 1: Parse subcommand**
 
 Extract from $ARGUMENTS:

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

@@ -38,8 +38,18 @@ $ARGUMENTS — none for interactive mode, or `-y` with `@file` reference for aut
 
 **Load project state if exists:**
 Check for `.workflow/state.json` -- loads context if project already initialized.
+
+**Output boundary**: ALL file writes MUST target `.workflow/` (project.md, state.json, config.json, specs/) only. NEVER modify source code or files outside `.workflow/`.
 </context>
 
+<invariants>
+1. **Idempotent init** — re-running init on an already-initialized project MUST detect existing `.workflow/` and warn (E002); NEVER silently overwrite existing state
+2. **Scope guard** — init MUST only make initialization decisions; NEVER prejudge roadmap structure, plan scope, or implementation details
+3. **All artifacts required** — init MUST NOT report completion until project.md, state.json, and config.json all exist; missing artifacts MUST be created before exit
+4. **Template-driven** — deferred templates (project.md, state.json, config.json) MUST be read from `~/.maestro/templates/` and customized; NEVER generate from scratch without template
+5. **Interview writes back** — all interactive decisions MUST be written to project.md/config.json before proceeding to research or completion; NEVER leave decisions unrecorded
+</invariants>
+
 <interview_protocol>
 Follows @~/.maestro/workflows/interview-mechanics.md standard.
 
@@ -52,6 +62,24 @@ Follows @~/.maestro/workflows/interview-mechanics.md standard.
 </interview_protocol>
 
 <execution>
+
+### Phase Gates (MANDATORY, BLOCKING)
+
+**GATE 1: Pre-flight → Interview**
+- REQUIRED: `.workflow/` existence check completed.
+- REQUIRED: `--from` source validated (if provided).
+- BLOCKED if: E002 (greenfield conflict with existing `.workflow/`) unresolved.
+
+**GATE 2: Interview → Research**
+- REQUIRED: All interview decisions recorded in project.md and config.json.
+- REQUIRED: `.workflow/` directory created with initial structure.
+- BLOCKED if: interview decisions not yet written to files.
+
+**GATE 3: Research → Completion**
+- REQUIRED: All 3 required artifacts exist (project.md, state.json, config.json).
+- REQUIRED: `.workflow/specs/` initialized.
+- BLOCKED if: any artifact missing — write it before reporting completion.
+
 ### Pre-flight
 
 1. Check if `.workflow/` already exists — if so, load state and warn (E002 for greenfield conflicts)

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

@@ -34,8 +34,19 @@ Turn natural-language instructions into command overlays — JSON patch files th
 **Management** — listing and removing overlays is handled by `maestro overlay list` (ink TUI with interactive delete). This skill focuses solely on creation.
 
 **Available sections** (for `section:` in patches): `purpose`, `required_reading`, `deferred_reading`, `context`, `execution`, `error_codes`, `success_criteria`.
+
+**Output boundary**: ALL file writes MUST target `~/.maestro/overlays/` (overlay JSON + docs) only. Command file patching is handled by `maestro overlay add` — this skill NEVER modifies `.claude/commands/*.md` directly.
 </context>
 
+<invariants>
+1. **Non-invasive** — overlays MUST use hashed HTML-comment markers for injection; NEVER edit command file content directly outside the overlay system
+2. **Idempotent** — re-running `maestro overlay apply` with the same overlay JSON MUST produce no file changes
+3. **Creation only** — this skill MUST only create overlays; listing and removal are handled by `maestro overlay list` (ink TUI)
+4. **Pristine source preferred** — injection point analysis MUST read from `$PKG_ROOT/.claude/commands/` (untouched originals) first, fall back to `~/.claude/commands/` only if pristine unavailable
+5. **User approval before write** — overlay JSON MUST be shown and approved via AskUserQuestion before writing to disk; NEVER auto-install without confirmation
+6. **Chain skip option mandatory** — if a skill chain is configured, the injected content MUST include a "Skip" option in AskUserQuestion; NEVER force the user into a chain
+</invariants>
+
 <execution>
 ### 1. Parse user intent
 

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

@@ -67,8 +67,19 @@ $ARGUMENTS — template slug/path, or flags.
   "last_checkpoint": null
 }
 ```
+
+**Output boundary**: ALL file writes MUST target `.workflow/.maestro/player-*/` (session status.json + step outputs) only. Template files (`~/.maestro/templates/workflows/`) are read-only. NEVER modify source code or `.claude/commands/` files.
 </context>
 
+<invariants>
+1. **Resume-safe** — status.json MUST be updated after every step change; an interrupted session MUST be resumable from the last checkpoint via `-c`
+2. **One step at a time** — execution loop MUST process steps sequentially in topological order; parallel steps share a batch index but NEVER run concurrently within the player
+3. **CLI nodes async** — cli-type nodes MUST use `Bash(run_in_background: true)` + STOP pattern; NEVER block the main thread on delegate execution
+4. **Templates read-only** — player MUST NOT modify template JSON files; runtime state belongs in status.json only
+5. **Failure requires user decision** — on step failure without explicit `on_fail`, player MUST prompt via AskUserQuestion (Retry/Skip/Abort); NEVER silently skip or auto-abort
+6. **Reference resolution at runtime** — `{N-xxx.field}` and `{prev_*}` references MUST be resolved at execution time from status.json step results, NEVER pre-resolved during init
+</invariants>
+
 <state_machine>
 
 <states>

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

@@ -39,9 +39,38 @@ Parse for:
    - Browse: `maestro search --category coding`
    - Load task-relevant entries: `maestro load --type knowhow --id <id1> [id2...]`
 3. All are optional — proceed without if unavailable.
+
+**Output boundary**: ALL file writes MUST target `.workflow/scratch/` (task directory, plan.json, summaries) and modified source files as defined in plan.json tasks. State.json scratch entry is implicit workflow tracking.
 </context>
 
+<invariants>
+1. **Atomic commits** — each task execution MUST produce a commit with only the files modified by that task; NEVER stage unrelated files
+2. **Evidence-based summaries** — task summaries MUST include concrete evidence (files changed, tests run, commands executed); NEVER accept "task completed successfully" as a summary
+3. **Plan before execute** — plan.json MUST be written before any task execution begins; NEVER skip planning even for single-task workflows
+4. **Scratch isolation** — all workflow artifacts MUST live under `.workflow/scratch/{task-dir}/`; NEVER write workflow metadata outside this directory
+5. **Commit confirmation** — staged files and commit message MUST be shown via AskUserQuestion before committing (unless `-y`); NEVER auto-commit without user awareness
+</invariants>
+
 <execution>
+
+### Phase Gates (MANDATORY, BLOCKING)
+
+**GATE 1: Setup → Planning**
+- REQUIRED: Task description parsed and scratch directory created.
+- REQUIRED: Coding specs loaded (optional but attempted).
+- BLOCKED if: no task description (E001) or scratch directory creation failed (E002).
+
+**GATE 2: Planning → Execution**
+- REQUIRED: plan.json written with task definitions.
+- REQUIRED: --discuss decisions extracted and recorded (if flag set).
+- BLOCKED if: plan.json missing or empty.
+
+**GATE 3: Execution → Completion**
+- REQUIRED: All tasks executed with `.summaries/TASK-*-summary.md` written per task.
+- REQUIRED: Each summary contains concrete evidence of completion.
+- REQUIRED: --full verification passed (if flag set).
+- BLOCKED if: any task summary missing or lacking evidence.
+
 Follow '~/.maestro/workflows/quick.md' completely.
 
 ### Artifact Verification (before completion)

package/.claude/commands/maestro-ralph-v2.md

@@ -12,6 +12,7 @@ allowed-tools:
   - Skill
   - AskUserQuestion
   - Agent
+  - SendMessage
 ---
 <purpose>
 Adaptive lifecycle orchestrator: locate step → resolve args → load context → dispatch Agent(ralph-executor) per step (agent 调 `ralph next` + 执行) → extract signals → drift check → ralph complete → evaluate decision → next step → loop.
@@ -42,13 +43,16 @@ Remaining      → intent (amend_mode 时为 change_request)
 </context>
 
 <invariants>
-1. **Ralph-v2 owns the full loop** — locate step → resolve args → load context → dispatch agent → extract signals → drift → complete，全部在本命令内完成
-2. **One agent per step** — 每个执行 step 派发一个 executor agent，agent 执行完返回，主流程解析结果后决定下一步
-3. **Agent is a thin wrapper** — executor agent 调 `ralph next` 获取 skill prompt 并执行，返回输出；arg resolution、context loading、signal extraction、drift analysis、ralph complete 均由主流程完成
-4. **Synchronous dispatch** — Agent() 调用是同步的（等待返回），不需要 STOP/callback 模式
+1. **Ralph-v2 owns the full loop** — locate step → resolve args → load context → dispatch agent → wait for SendMessage → receive → extract signals → drift → complete，全部在本命令内完成
+2. **One agent per step** — 每个执行 step 派发一个 named executor agent，agent 通过 SendMessage 回传结果，主流程解析结果后决定下一步
+3. **Agent is a thin wrapper** — executor agent 调 `ralph next` 获取 skill prompt 并执行，通过 SendMessage 回传输出；arg resolution、context loading、signal extraction、drift analysis、ralph complete 均由主流程完成
+4. **Dual dispatch model** — 执行 Agent 与评估 Agent 使用不同派发模式：
+   - **执行 Agent**（A_STEP_DISPATCH）：`Agent({name: "exe-xxx"})` — named mailbox teammate，支持内部多 agent 编排，通过 `SendMessage({to: "main"})` 回传结果
+   - **评估 Agent**（A_AGENT_EVALUATE / A_AGENT_GOAL_AUDIT / A_AGENT_REGROUND）：`Agent()` 不传 `name` — 同步阻塞，直接返回结果（评估 agent 只读不需要多 agent）
+   - `agent_exec_name` 既用于 display/日志，也作为执行 Agent 的 `name` 参数
 5. **主流程调 `ralph complete`** — 每个 step 完成后由主流程调 `maestro ralph complete`，非 agent 上报
-6. **Decision evaluation inline** — decision 节点不 handoff，通过 `Agent()` 启动分析 Agent 在本循环内评估
-7. **No CLI delegation** — 本命令不使用 `maestro delegate`；执行和评估均通过 Agent() 完成
+6. **Decision evaluation inline** — decision 节点不 handoff，通过 Agent 或 CLI delegate 在本循环内评估
+7. **CLI delegation for evaluation only** — CLI delegate（`maestro delegate --mode analysis`）仅限评估环节；执行仍通过 executor Agent 完成
 8. **Decision delegates read-only** — 评估 Agent 通过 prompt 中的 CONSTRAINTS 约束为只读
 9. **执行 step 通过 `maestro ralph next` CLI 加载并内联执行**（由 execute Agent 完成）
 10. **status.json 是唯一真源** — 不生成 markdown 清单或侧文件
@@ -85,7 +89,8 @@ S_CONFIRM       — 用户确认
 
 S_STEP_LOCATE     — 找下一个 pending step                    PERSIST: —
 S_STEP_RESOLVE    — 解析占位符 + 丰富参数                    PERSIST: step.args (enriched)
-S_STEP_DISPATCH   — 组装上下文 + 派发 executor agent 执行单步  PERSIST: step.agent_exec_name, step.status = "running"
+S_STEP_DISPATCH   — 组装上下文 + 派发 named executor agent     PERSIST: step.agent_exec_name, step.status = "running"
+S_STEP_WAIT_MSG   — 等待 executor 的 SendMessage 回传          PERSIST: —
 S_STEP_ANALYZE    — 提取信号 + 组装 completion 参数            PERSIST: —
 S_STEP_DRIFT      — 产物 vs 目标偏离分析                      PERSIST: step.drift_score
 S_STEP_COMPLETE   — 调 `ralph complete` 上报                  PERSIST: step.completion_*
@@ -140,8 +145,12 @@ S_STEP_RESOLVE:
   → S_STEP_DISPATCH  DO: A_STEP_RESOLVE_ARGS
 
 S_STEP_DISPATCH:
-  → S_STEP_ANALYZE   WHEN: agent returned output               DO: A_STEP_DISPATCH
-  → S_HANDLE_FAIL    WHEN: agent returned error/null            DO: A_STEP_DISPATCH
+  → S_STEP_WAIT_MSG  WHEN: executor dispatched                  DO: A_STEP_DISPATCH
+
+S_STEP_WAIT_MSG:
+  → S_STEP_ANALYZE   WHEN: received agent-message with EXECUTOR_OUTPUT   DO: A_STEP_RECEIVE
+  → S_HANDLE_FAIL    WHEN: received agent-message with status=ERROR      DO: A_STEP_RECEIVE
+  → S_HANDLE_FAIL    WHEN: executor idle without SendMessage（崩溃检测）  DO: mark BLOCKED
 
 S_STEP_ANALYZE:
   → S_STEP_DRIFT     WHEN: STATUS == DONE|DONE_WITH_CONCERNS    DO: A_STEP_EXTRACT
@@ -321,14 +330,17 @@ Execution criteria: {session.execution_criteria joined by '; '}
    | debug | `dbg` |
    | Other | `run` |
 
-**4. Dispatch:**
+**4. Dispatch（named mailbox executor）:**
+
+> 执行 Agent 使用 `name` 参数派发为 named mailbox teammate，支持内部多 agent 编排。executor 通过 `SendMessage({to: "main"})` 回传执行结果。`resolved_agent_name` 同时用于 display 和 Agent `name` 参数。
 
 ```
 Agent({
-  subagent_type: "ralph-executor",
   name: "{resolved_agent_name}",
+  subagent_type: "ralph-executor",
   description: "执行 step {index}: {step.skill}",
   prompt: `Session: {session_id}
+Agent name: {resolved_agent_name}
 
 {goal_context 块，仅 protocol < 2 时}
 
@@ -337,11 +349,24 @@ Agent({
 })
 ```
 
-Agent 内部调 `maestro ralph next --session {session_id}` 获取 skill prompt 并执行。
-
 5. Write `step.agent_exec_name` to status.json
-6. Display: `[{index}/{total}] ⟶ {step.skill} → Agent:{name}`
-7. Agent 同步返回 → `agent_output` = 返回文本；返回 null → STATUS=BLOCKED
+6. Display: `[{index}/{total}] ⟶ {step.skill} → {resolved_agent_name}`
+7. Agent() 立即返回（named teammate 异步启动）→ 进入 S_STEP_WAIT_MSG 等待 executor 的 SendMessage
+
+### A_STEP_RECEIVE
+
+从 executor 的 `agent-message` 中提取执行输出。
+
+1. 接收到 `<agent-message from="{resolved_agent_name}">` 后，提取 message 内容
+2. 解析 `EXECUTOR_OUTPUT` 格式：
+   - `status: DONE` → agent_output = message 内容，继续 S_STEP_ANALYZE
+   - `status: DONE_WITH_CONCERNS` → agent_output = message 内容，继续 S_STEP_ANALYZE
+   - `status: ERROR` → STATUS=BLOCKED，转 S_HANDLE_FAIL
+3. idle_notification 处理（区分正常等待与崩溃）：
+   - 首次收到 `idle_notification`：executor 可能在等待 worker 回传，正常现象，继续等待
+   - 连续收到 2 次 `idle_notification` 且中间无 `agent-message`：向 executor 发送 `SendMessage({to: "{resolved_agent_name}", message: "请回报执行状态"})` 询问
+   - 询问后仍只收到 `idle_notification` 无 `agent-message`：STATUS=BLOCKED，转 S_HANDLE_FAIL
+4. `agent_output` = 提取的 message 内容
 
 ### A_STEP_EXTRACT
 
@@ -470,10 +495,12 @@ Agent 内部调 `maestro ralph next --session {session_id}` 获取 skill prompt
 
 ### A_AGENT_EVALUATE
 
-通过 Agent 在本循环内评估质量门。
+通过 Agent 和/或 CLI delegate 评估质量门。评估模式由 `step.evaluate_via` 决定。
+
+**1. Common setup:**
 
 1. Resolve artifact dir: `.workflow/scratch/{artifact.path}/` with fallback glob
-2. Parse decision metadata: `{ decision, retry_count, max_retries }`
+2. Parse decision metadata: `{ decision, retry_count, max_retries, evaluate_via }`
 3. Map result files:
 
    | Decision | Files |
@@ -484,33 +511,65 @@ Agent 内部调 `maestro ralph next --session {session_id}` 获取 skill prompt
    | post-test | uat.md, .tests/test-results.json |
    | post-frontend-verify | e2e-results.json |
 
-4. Dispatch evaluation Agent:
-   ```
-   Agent({
-     name: "eval-{decision}-{timestamp}",
-     description: "评估 {decision} 质量门",
-     prompt: "PURPOSE: 评估 {decision} 质量门结果
-   TASK: 读取以下结果文件 | 分析状态 | 评估严重性 | 给出建议
-   FILES: {result_file_paths}
-   SESSION: {session_dir}/status.json
-   EXPECTED: 输出以下格式：
-   ---VERDICT---
-   STATUS: PASS|FAIL|PARTIAL|BLOCKED
-   REASON: <一句话原因>
-   GAP_SUMMARY: <差距摘要>
-   CONFIDENCE: high|medium|low
-   CONFIDENCE_SCORE: 0-100
-   WEAKEST_DIMENSION: <最弱维度>
-   ---END---
-   CONSTRAINTS: 只评估不修改文件 | 置信度<60%倾向 fix | retry {n}/{max} 达上限必须 escalate"
-   })
-   ```
-5. On return: parse `---VERDICT---` block — STATUS must match strict enum `PASS|FAIL|PARTIAL|BLOCKED`; parse failure → fallback STATUS="fix", `parse_failed: true`, `confidence_score: 0` (invariant 18)
+4. `evaluate_via` 默认值：`"agent"`（未设置时）
+
+**2. Dispatch by mode:**
+
+**Mode: `agent`（默认）** — 同步 Agent 评估：
+
+```
+Agent({
+  description: "评估 {decision} 质量门（同步评估 Agent，不传 name）",
+  prompt: "PURPOSE: 评估 {decision} 质量门结果
+TASK: 读取以下结果文件 | 分析状态 | 评估严重性 | 给出建议
+FILES: {result_file_paths}
+SESSION: {session_dir}/status.json
+EXPECTED: 输出以下格式：
+---VERDICT---
+STATUS: PASS|FAIL|PARTIAL|BLOCKED
+REASON: <一句话原因>
+GAP_SUMMARY: <差距摘要>
+CONFIDENCE: high|medium|low
+CONFIDENCE_SCORE: 0-100
+WEAKEST_DIMENSION: <最弱维度>
+---END---
+CONSTRAINTS: 只评估不修改文件 | 置信度<60%倾向 fix | retry {n}/{max} 达上限必须 escalate"
+})
+```
+
+**Mode: `cli`** — CLI delegate 评估（异步后台）：
+
+```
+Bash({
+  command: `maestro delegate "PURPOSE: 评估 ${decision} 质量门结果\nTASK: 读取 ${result_file_paths} | 分析状态 | 评估严重性\nEXPECTED: ---VERDICT--- 格式（STATUS/REASON/GAP_SUMMARY/CONFIDENCE_SCORE）\nCONSTRAINTS: 只评估不修改文件" --mode analysis --rule analysis-review-code-quality`,
+  run_in_background: true
+})
+```
+等待 delegate 完成 → `maestro delegate output {exec_id}` 获取结果 → 解析 `---VERDICT---`
+
+**Mode: `dual`** — Agent + CLI 并行评估，交叉验证：
+
+1. 先派发 CLI delegate（`run_in_background: true`）
+2. 同时派发同步 Agent（阻塞等待）
+3. Agent 返回后，检查 CLI delegate 状态（`maestro delegate status {exec_id}`）
+4. 合并裁决：
+
+   | Agent 结果 | CLI 结果 | 合并策略 |
+   |-----------|---------|---------|
+   | 两者一致 | — | 采用共识，confidence_score 取较高值 |
+   | Agent=PASS, CLI=FAIL | — | 降级为 PARTIAL，confidence_score 取平均值 |
+   | Agent=FAIL, CLI=PASS | — | 维持 FAIL（保守策略） |
+   | CLI 未返回 | — | 使用 Agent 结果，标 `"cli_pending": true` |
+
+**3. Verdict parse + adjustment（所有模式通用）:**
+
+5. Parse `---VERDICT---` block — STATUS must match strict enum `PASS|FAIL|PARTIAL|BLOCKED`; parse failure → fallback STATUS="fix", `parse_failed: true`, `confidence_score: 0` (invariant 18)
 6. Confidence adjustment: <60 + proceed → fix; >95 + fix + retry>0 → suggest proceed
 7. **Decision log**: Append to `{session_dir}/decisions.ndjson`:
    ```json
    { "id": "DEC-{timestamp}", "timestamp": "{ISO}", "source": "ralph-v2",
      "node_id": "{step.decision}", "type": "quality-gate",
+     "evaluate_via": "{mode}", "cli_exec_id": "{exec_id|null}",
      "verdict": "{adjusted_verdict}", "confidence_score": {N},
      "parse_failed": false,
      "close_call": {N>=50 && N<=70}, "summary": "{REASON}" }
@@ -518,14 +577,13 @@ Agent 内部调 `maestro ralph next --session {session_id}` 获取 skill prompt
 
 ### A_AGENT_GOAL_AUDIT
 
-通过 Agent 审计子目标完成情况。
+通过 Agent 和/或 CLI delegate 审计子目标完成情况。支持 `evaluate_via` 三种模式（同 A_AGENT_EVALUATE）。
 
 1. Read `session.task_decomposition` from status.json
-2. Dispatch audit Agent:
+2. Dispatch audit（按 `evaluate_via` 模式，默认 `agent`）:
    ```
    Agent({
-     name: "goal-audit-{timestamp}",
-     description: "审计子目标完成情况",
+     description: "审计子目标完成情况（同步评估 Agent，不传 name）",
      prompt: "PURPOSE: 审计未完成子目标，判定 met / unmet
    TASK:
      1. 读取 {session_dir}/status.json 中 task_decomposition 的 status!=done 子目标
@@ -549,20 +607,19 @@ Agent 内部调 `maestro ralph next --session {session_id}` 获取 skill prompt
    })
    ```
 3. On return: parse verdict, update task_decomposition status
-4. Append `{session_dir}/decisions.ndjson`：`{ "type": "goal-gate", "unmet_count": N, "unmet_ids": [...] }`
+4. Append `{session_dir}/decisions.ndjson`：`{ "type": "goal-gate", "evaluate_via": "{mode}", "unmet_count": N, "unmet_ids": [...] }`
 5. Verdict routing: `all_met` + `INTENT_ALIGNED=true` → A_APPLY_GOAL_DONE；`all_met` + `INTENT_ALIGNED=false` → A_REGROUND_HALT；`has_unmet` → A_APPLY_GOAL_FIX
    GUARD: retry_count >= max_retries AND still unmet → A_APPLY_ESCALATE
 
 ### A_AGENT_REGROUND
 
-通过 Agent 执行意图保真检查。
+通过 Agent 和/或 CLI delegate 执行意图保真检查。支持 `evaluate_via` 三种模式（同 A_AGENT_EVALUATE）。
 
 1. Read status.json：intent, boundary_contract, completed steps, done goals
-2. Dispatch reground Agent:
+2. Dispatch reground（按 `evaluate_via` 模式，默认 `agent`）:
    ```
    Agent({
-     name: "reground-{timestamp}",
-     description: "意图保真检查",
+     description: "意图保真检查（同步评估 Agent，不传 name）",
      prompt: "PURPOSE: 意图保真检查 — 对照 intent 验证累积执行是否漂移
    TASK:
      1. 读取 intent + boundary_contract.definition_of_done
@@ -661,8 +718,7 @@ Agent 内部调 `maestro ralph next --session {session_id}` 获取 skill prompt
 **Phase 3 Agent prompt:**
 ```
 Agent({
-  name: "amend-grill-{timestamp}",
-  description: "Amend impact analysis",
+  description: "Amend impact analysis（同步评估 Agent，不传 name）",
   prompt: "PURPOSE: 评估目标修改对 running session 的影响
 TASK:
   1. 读取 {session_dir}/status.json 的 task_decomposition + boundary_contract + 已完成 steps
@@ -693,7 +749,10 @@ GUARD: 已完成（`status: "done"`）的目标不可 supersede（skip + warn）
 
 - **A_RETRY**: `Bash("maestro ralph retry {index}")` — CLI 设 `step.retried = true`, `step.status = "pending"`, 清 `active_step_index`
 - **A_PAUSE_SESSION**: `ralph complete N --status BLOCKED --reason "..."` — CLI 写 `session.status = "paused"`
-- **A_COMPLETE_SESSION**: 校验所有 step `completion_confirmed == true` + `task_decomposition_all_done == true`（若存在），通过后写 `session.status = "completed"`
+- **A_COMPLETE_SESSION**: 校验所有 step `completion_confirmed == true` + `task_decomposition_all_done == true`（若存在），通过后写 `session.status = "completed"`。完成后向所有 idle executor 发送 shutdown_request 清理：
+  ```
+  遍历 steps[].agent_exec_name → SendMessage({to: name, message: {type: "shutdown_request"}})
+  ```
 
 </actions>
 
@@ -760,6 +819,7 @@ Build rules 0-14 全部适用，包括 spec-setup 预检（rule 0.5）、grill a
     "stage": "",
     "scope": null,
     "decision": null,
+    "evaluate_via": "agent|cli|dual",
     "retry_count": 0,
     "max_retries": 2,
     "command_scope": "global|project|missing|null",
@@ -824,24 +884,27 @@ E001–E006, W001–W004 适用。Agent 新增：
 
 ### Success Criteria
 
-- [ ] ralph-v2 owns full step loop: locate → resolve → dispatch → extract → drift → complete → next
-- [ ] One agent per step — `Agent({ subagent_type: "ralph-executor" })` 每步派发一个
-- [ ] Agent 内调 `maestro ralph next` 获取 skill prompt 并执行，返回执行输出
+- [ ] ralph-v2 owns full step loop: locate → resolve → dispatch → wait → receive → extract → drift → complete → next
+- [ ] One agent per step — `Agent({ name, subagent_type: "ralph-executor" })` 每步派发一个 named executor
+- [ ] Executor 内调 `maestro ralph next` 获取 skill prompt 并执行，支持内部多 agent 编排
+- [ ] Executor 通过 `SendMessage({to: "main"})` 回传 `EXECUTOR_OUTPUT` 格式结果
 - [ ] 主流程调 `maestro ralph complete` 上报（非 agent 上报）
 - [ ] 主流程负责 arg resolution、context loading、signal extraction、drift analysis
-- [ ] Agent 返回执行输出文本，主流程从中提取信号和状态
-- [ ] Agent 崩溃（返回 null）→ STATUS=BLOCKED，转 S_HANDLE_FAIL
-- [ ] 无 STOP/callback 模式 — Agent() 同步等待返回
-- [ ] Decision evaluation 通过 Agent() 同步完成
+- [ ] A_STEP_RECEIVE 从 executor 的 agent-message 中提取执行输出
+- [ ] Executor 崩溃（idle 无 SendMessage）→ 询问 2 次后 STATUS=BLOCKED，转 S_HANDLE_FAIL
+- [ ] Dual dispatch: 执行 Agent 传 `name`（async + SendMessage），评估 Agent 不传 `name`（sync）或 CLI delegate
+- [ ] Decision evaluation 支持三种模式：agent（同步）、cli（CLI delegate）、dual（并行交叉验证）
+- [ ] `evaluate_via` 字段控制评估模式，默认 `"agent"`
+- [ ] dual 模式合并策略：一致取共识、分歧保守降级、CLI 未返回用 Agent 结果
 - [ ] Verdict 解析保持 `---VERDICT---` 格式，parse 失败 → fallback fix + parse_failed: true
 - [ ] decisions.ndjson 追加：source 字段为 `"ralph-v2"`
-- [ ] Session schema: `execution_mode: "agent"`，`agent_exec_name`（执行 Agent 名称标识），含 `artifacts_produced`
+- [ ] Session schema: `execution_mode: "agent"`，`agent_exec_name`（执行 Agent name + display），含 `artifacts_produced`
 - [ ] Chain building（S_RESOLVE_PHASE through S_BUILD_CHAIN）自包含执行
 - [ ] A_STEP_DISPATCH 含前序产出加载（滑动窗口 5 step + accumulated signals + stage-specific artifacts）
-- [ ] Agent name 含 stage prefix（grl/brn/anm/ana/pln/exe/rev/tst/dbg）
+- [ ] `agent_exec_name` 含 stage prefix（grl/brn/anm/ana/pln/exe/rev/tst/dbg）——同时用于 Agent name 参数和 display
 - [ ] `--summary` 在 DONE/DONE_WITH_CONCERNS 时为 MUST（动词开头，≤100 字）
 - [ ] CAVEATS 在 DONE_WITH_CONCERNS 时同时映射 --concerns
-- [ ] A_STEP_EXTRACT 从 agent 输出提取 artifact IDs、path signals、phase signals
+- [ ] A_STEP_EXTRACT 从 executor 输出提取 artifact IDs、path signals、phase signals
 - [ ] A_STEP_DRIFT_ANALYZE：ALIGNED/MINOR_DRIFT → complete；MAJOR_DRIFT+未重试 → retry；MAJOR_DRIFT+已重试 → DONE_WITH_CONCERNS
 - [ ] A_STEP_COMPLETE 将 context signals 写入 status.json.context
 - [ ] A_AMEND_GOAL：完整 5 步流程 + deferred_reading ralph-amend-goal.md + Agent mini grill 含完整 prompt

package/.claude/commands/maestro-swarm-workflow.md

@@ -43,6 +43,8 @@ Remaining        → intent
 | `wf-plan` | `{ context_dir?, from?, phase?, scope?, specs?, gaps?, quick? }` |
 | `wf-execute` | `{ plan_dir, specs?, codebase_context?, wiki_context?, auto_commit? }` |
 | `wf-milestone-audit` | `{ milestone?, is_adhoc? }` |
+
+**Output boundary**: ALL file writes MUST target `.workflow/scratch/{YYYYMMDD}-swarm-{script}-{slug}/` (results, ralph-compatible artifacts). When inside a ralph session, writes target the corresponding step's scratch directory. NEVER modify source code, workflow scripts (`~/.maestro/workflows/swarm/`), or `.claude/commands/` files.
 </context>
 
 <state_machine>
@@ -196,6 +198,29 @@ Workflow 返回 JSON 后：
 6. **结果必须展示** — Workflow 返回后必须向用户展示格式化摘要，不得静默完成
 </invariants>
 
+<execution>
+
+### Phase Gates (MANDATORY, BLOCKING)
+
+**GATE 1: Parse → Route**
+- REQUIRED: Intent text or `--script` flag parsed successfully.
+- BLOCKED if: no intent and no `--script` (E001).
+
+**GATE 2: Route → Context Assembly**
+- REQUIRED: Target script resolved unambiguously (single match or user-selected from candidates).
+- BLOCKED if: ambiguous routing without user resolution (E002).
+
+**GATE 3: Context → Dispatch**
+- REQUIRED: Args payload assembled with all required fields for target script.
+- REQUIRED: Script file exists at `~/.maestro/workflows/swarm/{script}.js`.
+- BLOCKED if: script file not found (E003) or required args missing.
+
+**GATE 4: Dispatch → Ingest**
+- REQUIRED: Workflow execution completed (success or partial results available).
+- BLOCKED if: Workflow execution failed entirely (E004) — offer `--resume`.
+
+</execution>
+
 <appendix>
 
 ### 与 Ralph 集成

package/.claude/commands/maestro-tools-execute.md

@@ -35,8 +35,37 @@ $ARGUMENTS — Tool name, keyword, or --category filter
 Empty arguments enters interactive mode: list all tools for user selection.
 </context>
 
+<invariants>
+1. **Confirmation before execution** — MUST AskUserQuestion before executing tool steps; NEVER auto-execute without user consent
+2. **Sequential step execution** — steps MUST be executed in defined order; NEVER skip or reorder steps unless user explicitly requests skip
+3. **Blocker escalation** — step failure MUST be reported to user with retry/skip/abort options; NEVER silently skip failed steps
+4. **Read-only tool definition** — tool execution MUST NOT modify the tool's knowhow document or spec entry; only the target codebase is modified per tool steps
+5. **Progress feedback** — each completed step MUST report `[Step N/M] done — <step_name>`; NEVER execute silently
+6. **Output boundary** — file writes are governed by the individual tool's step definitions. This command itself writes NO files beyond what the loaded tool prescribes
+</invariants>
+
 <execution>
 
+### Phase Gates (MANDATORY, BLOCKING)
+
+**GATE 1: Parse → Load**
+- REQUIRED: Tool name, keyword, or --category parsed from arguments (or empty for interactive mode).
+- BLOCKED if: invalid category value.
+
+**GATE 2: Load → Confirm**
+- REQUIRED: Exactly one tool resolved (direct match or user selection from candidates).
+- REQUIRED: Tool document loaded and steps extracted (ref entries expanded via `maestro load --type knowhow`).
+- BLOCKED if: E001 (no match found), E002 unresolved (multiple matches without user selection).
+
+**GATE 3: Confirm → Execute**
+- REQUIRED: User confirmed execution mode via AskUserQuestion (execute as-is / adjust / view only).
+- BLOCKED if: user selects "View only" — display steps and END without execution.
+
+**GATE 4: Execute → Report**
+- REQUIRED: All steps attempted (completed, skipped with user approval, or aborted by user).
+- REQUIRED: Results collected for each step (success/skip/fail).
+- BLOCKED if: user chose abort mid-execution — report partial results and END.
+
 ### Step 1: Load Tool
 
 **By name**:

package/.claude/commands/maestro-tools-register.md

@@ -35,8 +35,37 @@ $ARGUMENTS — Intent description
 ```
 </context>
 
+<invariants>
+1. **Schema validation** — tool knowhow document MUST include `tool: true`, `category`, `keywords`, and `summary` in YAML frontmatter; missing fields → reject write
+2. **No duplicate names** — tool title MUST be unique within its category; duplicate detection → E002 warning with overwrite/optimize confirmation
+3. **Category required** — every tool MUST declare exactly one category from: coding, test, review, arch, debug; empty category → E003
+4. **Confirmation gate** — MUST AskUserQuestion before writing knowhow document and spec ref entry; NEVER persist without user confirmation
+5. **Promote is in-place** — promote mode MUST update existing knowhow frontmatter via `maestro wiki update`; NEVER recreate the document
+6. **Output boundary** — ALL file writes MUST target .workflow/knowhow/ (tool documents) and .workflow/specs/ (ref entries via maestro spec add) only. NEVER modify source code or files outside these paths
+7. **Description format** — first line after `### Title` MUST state "Use when ..." (usage timing); this is critical for ref entry summary visibility in spec-load
+</invariants>
+
 <execution>
 
+### Phase Gates (MANDATORY, BLOCKING)
+
+**GATE 1: Parse → Gather**
+- REQUIRED: Mode determined (extract/generate/optimize/promote) from argument parsing.
+- REQUIRED: For optimize/promote modes, target tool/document exists and is loadable.
+- BLOCKED if: empty args without user response to AskUserQuestion.
+
+**GATE 2: Gather → Write**
+- REQUIRED: Tool name, category, and usage timing confirmed.
+- REQUIRED: Steps extracted or generated (extract: ≥1 step, generate: user-confirmed scope).
+- REQUIRED: Inline vs ref mode decided based on step count.
+- REQUIRED: User confirmed via AskUserQuestion (title, category, keywords, summary, step count).
+- BLOCKED if: E001 (.workflow/specs/ not initialized), E003 (no category), user cancels.
+
+**GATE 3: Write → Verify**
+- REQUIRED: Knowhow document written with `tool: true` frontmatter (or updated in-place for promote).
+- REQUIRED: Spec ref entry registered (if user confirmed).
+- BLOCKED if: write failed or spec add returned error.
+
 ### Step 1: Intent Detection
 
 Parse $ARGUMENTS to determine mode:

package/.claude/commands/maestro-ui-codify.md

@@ -33,8 +33,19 @@ Flags:
 - `--package-name <name>`: Package name for reference output (default: auto-generated from source directory)
 - `--output-dir <path>`: Output directory for reference package (default: `.workflow/reference_style`)
 - `--overwrite`: Allow overwriting existing package directory
+
+**Output boundary**: ALL file writes MUST target the `--output-dir` path (default: `.workflow/reference_style/`) for reference packages, and `.workflow/knowhow/` for knowledge assets (via `codify-to-knowhow`). NEVER modify the source directory being analyzed.
 </context>
 
+<invariants>
+1. **Source read-only** — the source path being analyzed MUST NOT be modified; extraction is purely read-only
+2. **Phase-sequential loading** — workflow files (ui-codify-extract, ui-codify-package, ui-codify-knowhow) MUST be read only when their phase starts; NEVER load all phases eagerly
+3. **User confirmation before knowhow** — Phase 3→4 gate MUST present AskUserQuestion before generating knowledge assets; NEVER auto-proceed to knowhow generation
+4. **Overwrite protection** — existing package directory MUST NOT be overwritten without `--overwrite` flag (E003)
+5. **Artifact completeness** — all 5 required artifacts MUST exist before reporting completion; NEVER skip artifact verification
+6. **Token-first extraction** — design-tokens.json MUST be generated before layout-templates.json; layout extraction depends on token foundation
+</invariants>
+
 <execution>
 ## 1. Load UI Specs