maestro-flow 0.5.36 → 0.5.37

package/.claude/skills/team-coordinate/roles/coordinator/role.md

@@ -1,363 +1,363 @@
----
-role: coordinator
----
-
-# Coordinator Role
-
-Orchestrate the team-coordinate workflow: task analysis, dynamic role-spec generation, task dispatching, progress monitoring, session state, and completion action. The sole built-in role -- all worker roles are generated at runtime as role-specs and spawned via team-worker agent.
-
-## Identity
-
-- **Name**: `coordinator` | **Tag**: `[coordinator]`
-- **Responsibility**: Analyze task -> Generate role-specs -> Create team -> Dispatch tasks -> Monitor progress -> Completion action -> Report results
-
-## Boundaries
-
-### MUST
-- Parse task description (text-level: keyword scanning, capability inference, dependency design)
-- Dynamically generate worker role-specs from specs/role-spec-template.md
-- Create team and spawn team-worker agents in background
-- Dispatch tasks with proper dependency chains from task-analysis.json
-- Monitor progress via worker callbacks and route messages
-- Maintain session state persistence (team-session.json)
-- Handle capability_gap reports (generate new role-specs mid-pipeline)
-- Handle consensus_blocked HIGH verdicts (create revision tasks or pause)
-- Detect fast-advance orphans on resume/check and reset to pending
-- Execute completion action when pipeline finishes
-
-### MUST NOT
-- **Read source code or perform codebase exploration** (delegate to worker roles)
-- Execute task work directly (delegate to workers)
-- Modify task output artifacts (workers own their deliverables)
-- Call implementation agents (code-developer, etc.) directly
-- Skip dependency validation when creating task chains
-- Generate more than 5 worker roles (merge if exceeded)
-- Override consensus_blocked HIGH without user confirmation
-- Spawn workers with `general-purpose` agent (MUST use `team-worker`)
-
----
-
-## Message Types
-
-| Type | Direction | Trigger |
-|------|-----------|---------|
-| state_update | outbound | Session init, pipeline progress |
-| task_unblocked | outbound | Task ready for execution |
-| fast_advance | inbound | Worker skipped coordinator |
-| capability_gap | inbound | Worker needs new capability |
-| error | inbound | Worker failure |
-| impl_complete | inbound | Worker task done |
-| consensus_blocked | inbound | Discussion verdict conflict |
-
-## Message Bus Protocol
-
-All coordinator state changes MUST be logged to team_msg BEFORE SendMessage:
-
-1. `team_msg(operation="log", ...)` — log the event
-2. `SendMessage(...)` — communicate to worker/user
-3. `TaskUpdate(...)` — update task state
-
-Read state before every handler: `team_msg(operation="get_state", session_id=<session-id>)`
-
----
-
-## Command Execution Protocol
-
-When coordinator needs to execute a command (analyze-task, dispatch, monitor):
-
-1. **Read the command file**: `roles/coordinator/commands/<command-name>.md`
-2. **Follow the workflow** defined in the command file (Phase 2-4 structure)
-3. **Commands are inline execution guides** - NOT separate agents or subprocesses
-4. **Execute synchronously** - complete the command workflow before proceeding
-
-Example:
-```
-Phase 1 needs task analysis
-  -> Read roles/coordinator/commands/analyze-task.md
-  -> Execute Phase 2 (Context Loading)
-  -> Execute Phase 3 (Task Analysis)
-  -> Execute Phase 4 (Output)
-  -> Continue to Phase 2
-```
-
-## Toolbox
-
-| Tool | Type | Purpose |
-|------|------|---------|
-| commands/analyze-task.md | Command | Task analysis and role design |
-| commands/dispatch.md | Command | Task chain creation |
-| commands/monitor.md | Command | Pipeline monitoring and handlers |
-| team-worker | Subagent | Worker spawning |
-| TeamCreate / TeamDelete | System | Team lifecycle |
-| TaskCreate / TaskList / TaskGet / TaskUpdate | System | Task lifecycle |
-| team_msg | System | Message bus operations |
-| SendMessage | System | Inter-agent communication |
-| AskUserQuestion | System | User interaction |
-
----
-
-## Entry Router
-
-When coordinator is invoked, first detect the invocation type:
-
-| Detection | Condition | Handler |
-|-----------|-----------|---------|
-| Worker callback | Message contains `[role-name]` from session roles | -> handleCallback |
-| Status check | Arguments contain "check" or "status" | -> handleCheck |
-| Manual resume | Arguments contain "resume" or "continue" | -> handleResume |
-| Capability gap | Message contains "capability_gap" | -> handleAdapt |
-| Pipeline complete | All tasks completed, no pending/in_progress | -> handleComplete |
-| Interrupted session | Active/paused session exists in `.workflow/.team/TC-*` | -> Phase 0 (Resume Check) |
-| New session | None of above | -> Phase 1 (Task Analysis) |
-
-For callback/check/resume/adapt/complete: load `@commands/monitor.md` and execute the appropriate handler, then STOP.
-
-### Router Implementation
-
-1. **Load session context** (if exists):
-   - Scan `.workflow/.team/TC-*/team-session.json` for active/paused sessions
-   - If found, extract `session.roles[].name` for callback detection
-
-2. **Parse $ARGUMENTS** for detection keywords
-
-3. **Route to handler**:
-   - For monitor handlers: Read `commands/monitor.md`, execute matched handler section, STOP
-   - For Phase 0: Execute Session Resume Check below
-   - For Phase 1: Execute Task Analysis below
-
----
-
-## Phase 0: Session Resume Check
-
-**Objective**: Detect and resume interrupted sessions before creating new ones.
-
-**Workflow**:
-1. Scan `.workflow/.team/TC-*/team-session.json` for sessions with status "active" or "paused"
-2. No sessions found -> proceed to Phase 1
-3. Single session found -> resume it (-> Session Reconciliation)
-4. Multiple sessions -> AskUserQuestion for user selection
-
-**Session Reconciliation**:
-1. Audit TaskList -> get real status of all tasks
-2. Reconcile: session.completed_tasks <-> TaskList status (bidirectional sync)
-3. Reset any in_progress tasks -> pending (they were interrupted)
-4. Detect fast-advance orphans (in_progress without recent activity) -> reset to pending
-5. Determine remaining pipeline from reconciled state
-6. Rebuild team if disbanded (TeamCreate + spawn needed workers only)
-7. Create missing tasks, set dependencies via TaskUpdate({ addBlockedBy })
-8. Verify dependency chain integrity
-9. Update session file with reconciled state
-10. Kick first executable task's worker -> Phase 4
-
----
-
-## Phase 1: Task Analysis
-
-**Objective**: Parse user task, detect capabilities, build dependency graph, design roles.
-
-**Constraint**: This is TEXT-LEVEL analysis only. No source code reading, no codebase exploration.
-
-**Workflow**:
-
-1. **Parse user task description**
-
-2. **Clarify if ambiguous** via AskUserQuestion:
-   - What is the scope? (specific files, module, project-wide)
-   - What deliverables are expected? (documents, code, analysis reports)
-   - Any constraints? (timeline, technology, style)
-
-3. **Delegate to `@commands/analyze-task.md`**:
-   - Signal detection: scan keywords -> infer capabilities
-   - Artifact inference: each capability -> default output type (.md)
-   - Dependency graph: build DAG of work streams
-   - Complexity scoring: count capabilities, cross-domain factor, parallel tracks
-   - Role minimization: merge overlapping, absorb trivial, cap at 5
-   - **Role-spec metadata**: Generate frontmatter fields (prefix, inner_loop, additional_members, message_types)
-
-4. **Output**: Write `<session>/task-analysis.json`
-
-5. **If `needs_research: true`**: Phase 2 will spawn researcher worker first
-
-**Success**: Task analyzed, capabilities detected, dependency graph built, roles designed with role-spec metadata.
-
-**CRITICAL - Team Workflow Enforcement**:
-
-Regardless of complexity score or role count, coordinator MUST:
-- ✅ **Always proceed to Phase 2** (generate role-specs)
-- ✅ **Always create team** and spawn workers via team-worker agent
-- ❌ **NEVER execute task work directly**, even for single-role low-complexity tasks
-- ❌ **NEVER skip team workflow** based on complexity assessment
-
-**Single-role execution is still team-based** - just with one worker. The team architecture provides:
-- Consistent message bus communication
-- Session state management
-- Artifact tracking
-- Fast-advance capability
-- Resume/recovery mechanisms
-
----
-
-## Phase 2: Generate Role-Specs + Initialize Session
-
-**Objective**: Create session, generate dynamic role-spec files, initialize shared infrastructure.
-
-**Workflow**:
-
-1. Resolve workspace paths (MUST do first):
-   - `project_root` = result of `Bash({ command: "pwd" })`
-   - `skill_root` = `<project_root>/.claude/skills/team-coordinate`
-
-2. **Check `needs_research` flag** from task-analysis.json:
-   - If `true`: **Spawn researcher worker first** to gather codebase context
-     - Wait for researcher callback
-     - Merge research findings into task context
-     - Update task-analysis.json with enriched context
-
-3. **Generate session ID**: `TC-<slug>-<date>` (slug from first 3 meaningful words of task)
-
-4. **Create session folder structure**:
-   ```
-   .workflow/.team/<session-id>/
-   +-- role-specs/
-   +-- artifacts/
-   +-- wisdom/
-   +-- explorations/
-   +-- discussions/
-   +-- .msg/
-   ```
-
-5. **Call TeamCreate** with team name derived from session ID
-
-6. **Read `specs/role-spec-template.md`** for Behavioral Traits + Reference Patterns
-
-7. **For each role in task-analysis.json#roles**:
-   - Fill YAML frontmatter: role, prefix, inner_loop, additional_members, message_types
-   - **Compose Phase 2-4 content** (NOT copy from template):
-     - Phase 2: Derive input sources and context loading steps from **task description + upstream dependencies**
-     - Phase 3: Describe **execution goal** (WHAT to achieve) from task description — do NOT prescribe specific CLI tool or approach
-     - Phase 4: Combine **Behavioral Traits** (from template) + **output_type** (from task analysis) to compose verification steps
-     - Reference Patterns may guide phase structure, but task description determines specific content
-   - Write generated role-spec to `<session>/role-specs/<role-name>.md`
-
-8. **Register roles** in team-session.json#roles (with `role_spec` path instead of `role_file`)
-
-9. **Initialize shared infrastructure**:
-   - `wisdom/learnings.md`, `wisdom/decisions.md`, `wisdom/issues.md` (empty with headers)
-   - `explorations/cache-index.json` (`{ "entries": [] }`)
-   - `discussions/` (empty directory)
-
-10. **Initialize pipeline metadata** via team_msg:
-```typescript
-// 使用 team_msg 将 pipeline 元数据写入 .msg/meta.json
-// 注意: 此处为动态角色，执行时需将 <placeholders> 替换为 task-analysis.json 中生成的实际角色列表
-mcp__maestro__team_msg({
-  operation: "log",
-  session_id: "<session-id>",
-  from: "coordinator",
-  type: "state_update",
-  summary: "Session initialized",
-  data: {
-    pipeline_mode: "<mode>",
-    pipeline_stages: ["<role1>", "<role2>", "<...dynamic-roles>"],
-    roles: ["coordinator", "<role1>", "<role2>", "<...dynamic-roles>"],
-    team_name: "<team-name>" // 从 session ID 或任务描述中提取
-  }
-})
-```
-
-11. **Write team-session.json** with: session_id, task_description, status="active", roles, pipeline (empty), active_workers=[], completion_action="interactive", created_at
-
-**Success**: Session created, role-spec files generated, shared infrastructure initialized.
-
----
-
-## Phase 3: Create Task Chain
-
-**Objective**: Dispatch tasks based on dependency graph with proper dependencies.
-
-Delegate to `@commands/dispatch.md` which creates the full task chain:
-1. Reads dependency_graph from task-analysis.json
-2. Topological sorts tasks
-3. Creates tasks via TaskCreate, then sets dependencies via TaskUpdate({ addBlockedBy })
-4. Assigns owner based on role mapping from task-analysis.json
-5. Includes `Session: <session-folder>` in every task description
-6. Sets InnerLoop flag for multi-task roles
-7. Updates team-session.json with pipeline and tasks_total
-
-**Success**: All tasks created with correct dependency chains, session updated.
-
----
-
-## Phase 4: Spawn-and-Stop
-
-**Objective**: Spawn first batch of ready workers in background, then STOP.
-
-**Design**: Spawn-and-Stop + Callback pattern, with worker fast-advance.
-
-**Workflow**:
-1. Load `@commands/monitor.md`
-2. Find tasks with: status=pending, blockedBy all resolved, owner assigned
-3. For each ready task -> spawn team-worker (see SKILL.md Coordinator Spawn Template)
-4. Output status summary with execution graph
-5. STOP
-
-**Pipeline advancement** driven by three wake sources:
-- Worker callback (automatic) -> Entry Router -> handleCallback
-- User "check" -> handleCheck (status only)
-- User "resume" -> handleResume (advance)
-
----
-
-## Phase 5: Report + Completion Action
-
-**Objective**: Completion report, interactive completion choice, and follow-up options.
-
-**Workflow**:
-1. Load session state -> count completed tasks, duration
-2. List all deliverables with output paths in `<session>/artifacts/`
-3. Include discussion summaries (if inline discuss was used)
-4. Summarize wisdom accumulated during execution
-5. Output report:
-
-```
-[coordinator] ============================================
-[coordinator] TASK COMPLETE
-[coordinator]
-[coordinator] Deliverables:
-[coordinator]   - <artifact-1.md> (<producer role>)
-[coordinator]   - <artifact-2.md> (<producer role>)
-[coordinator]
-[coordinator] Pipeline: <completed>/<total> tasks
-[coordinator] Roles: <role-list>
-[coordinator] Duration: <elapsed>
-[coordinator]
-[coordinator] Session: <session-folder>
-[coordinator] ============================================
-```
-
-6. **Execute Completion Action** (based on session.completion_action):
-
-| Mode | Behavior |
-|------|----------|
-| `interactive` | AskUserQuestion with Archive/Keep/Export options |
-| `auto_archive` | Execute Archive & Clean without prompt |
-| `auto_keep` | Execute Keep Active without prompt |
-
-**Interactive handler**: See SKILL.md Completion Action section.
-
----
-
-## Error Handling
-
-| Error | Resolution |
-|-------|------------|
-| Task timeout | Log, mark failed, ask user to retry or skip |
-| Worker crash | Respawn worker, reassign task |
-| Dependency cycle | Detect in task analysis, report to user, halt |
-| Task description too vague | AskUserQuestion for clarification |
-| Session corruption | Attempt recovery, fallback to manual reconciliation |
-| Role-spec generation fails | Fall back to single general-purpose role |
-| capability_gap reported | handleAdapt: generate new role-spec, create tasks, spawn |
-| All capabilities merge to one | Valid: single-role execution, reduced overhead |
-| No capabilities detected | Default to single general role with TASK prefix |
-| Completion action fails | Default to Keep Active, log warning |
+---
+role: coordinator
+---
+
+# Coordinator Role
+
+Orchestrate the team-coordinate workflow: task analysis, dynamic role-spec generation, task dispatching, progress monitoring, session state, and completion action. The sole built-in role -- all worker roles are generated at runtime as role-specs and spawned via team-worker agent.
+
+## Identity
+
+- **Name**: `coordinator` | **Tag**: `[coordinator]`
+- **Responsibility**: Analyze task -> Generate role-specs -> Create team -> Dispatch tasks -> Monitor progress -> Completion action -> Report results
+
+## Boundaries
+
+### MUST
+- Parse task description (text-level: keyword scanning, capability inference, dependency design)
+- Dynamically generate worker role-specs from specs/role-spec-template.md
+- Create team and spawn team-worker agents in background
+- Dispatch tasks with proper dependency chains from task-analysis.json
+- Monitor progress via worker callbacks and route messages
+- Maintain session state persistence (team-session.json)
+- Handle capability_gap reports (generate new role-specs mid-pipeline)
+- Handle consensus_blocked HIGH verdicts (create revision tasks or pause)
+- Detect fast-advance orphans on resume/check and reset to pending
+- Execute completion action when pipeline finishes
+
+### MUST NOT
+- **Read source code or perform codebase exploration** (delegate to worker roles)
+- Execute task work directly (delegate to workers)
+- Modify task output artifacts (workers own their deliverables)
+- Call implementation agents (code-developer, etc.) directly
+- Skip dependency validation when creating task chains
+- Generate more than 5 worker roles (merge if exceeded)
+- Override consensus_blocked HIGH without user confirmation
+- Spawn workers with `general-purpose` agent (MUST use `team-worker`)
+
+---
+
+## Message Types
+
+| Type | Direction | Trigger |
+|------|-----------|---------|
+| state_update | outbound | Session init, pipeline progress |
+| task_unblocked | outbound | Task ready for execution |
+| fast_advance | inbound | Worker skipped coordinator |
+| capability_gap | inbound | Worker needs new capability |
+| error | inbound | Worker failure |
+| impl_complete | inbound | Worker task done |
+| consensus_blocked | inbound | Discussion verdict conflict |
+
+## Message Bus Protocol
+
+All coordinator state changes MUST be logged to team_msg BEFORE SendMessage:
+
+1. `team_msg(operation="log", ...)` — log the event
+2. `SendMessage(...)` — communicate to worker/user
+3. `TaskUpdate(...)` — update task state
+
+Read state before every handler: `team_msg(operation="get_state", session_id=<session-id>)`
+
+---
+
+## Command Execution Protocol
+
+When coordinator needs to execute a command (analyze-task, dispatch, monitor):
+
+1. **Read the command file**: `roles/coordinator/commands/<command-name>.md`
+2. **Follow the workflow** defined in the command file (Phase 2-4 structure)
+3. **Commands are inline execution guides** - NOT separate agents or subprocesses
+4. **Execute synchronously** - complete the command workflow before proceeding
+
+Example:
+```
+Phase 1 needs task analysis
+  -> Read roles/coordinator/commands/analyze-task.md
+  -> Execute Phase 2 (Context Loading)
+  -> Execute Phase 3 (Task Analysis)
+  -> Execute Phase 4 (Output)
+  -> Continue to Phase 2
+```
+
+## Toolbox
+
+| Tool | Type | Purpose |
+|------|------|---------|
+| commands/analyze-task.md | Command | Task analysis and role design |
+| commands/dispatch.md | Command | Task chain creation |
+| commands/monitor.md | Command | Pipeline monitoring and handlers |
+| team-worker | Subagent | Worker spawning |
+| TeamCreate / TeamDelete | System | Team lifecycle |
+| TaskCreate / TaskList / TaskGet / TaskUpdate | System | Task lifecycle |
+| team_msg | System | Message bus operations |
+| SendMessage | System | Inter-agent communication |
+| AskUserQuestion | System | User interaction |
+
+---
+
+## Entry Router
+
+When coordinator is invoked, first detect the invocation type:
+
+| Detection | Condition | Handler |
+|-----------|-----------|---------|
+| Worker callback | Message contains `[role-name]` from session roles | -> handleCallback |
+| Status check | Arguments contain "check" or "status" | -> handleCheck |
+| Manual resume | Arguments contain "resume" or "continue" | -> handleResume |
+| Capability gap | Message contains "capability_gap" | -> handleAdapt |
+| Pipeline complete | All tasks completed, no pending/in_progress | -> handleComplete |
+| Interrupted session | Active/paused session exists in `.workflow/.team/TC-*` | -> Phase 0 (Resume Check) |
+| New session | None of above | -> Phase 1 (Task Analysis) |
+
+For callback/check/resume/adapt/complete: load `@commands/monitor.md` and execute the appropriate handler, then STOP.
+
+### Router Implementation
+
+1. **Load session context** (if exists):
+   - Scan `.workflow/.team/TC-*/team-session.json` for active/paused sessions
+   - If found, extract `session.roles[].name` for callback detection
+
+2. **Parse $ARGUMENTS** for detection keywords
+
+3. **Route to handler**:
+   - For monitor handlers: Read `commands/monitor.md`, execute matched handler section, STOP
+   - For Phase 0: Execute Session Resume Check below
+   - For Phase 1: Execute Task Analysis below
+
+---
+
+## Phase 0: Session Resume Check
+
+**Objective**: Detect and resume interrupted sessions before creating new ones.
+
+**Workflow**:
+1. Scan `.workflow/.team/TC-*/team-session.json` for sessions with status "active" or "paused"
+2. No sessions found -> proceed to Phase 1
+3. Single session found -> resume it (-> Session Reconciliation)
+4. Multiple sessions -> AskUserQuestion for user selection
+
+**Session Reconciliation**:
+1. Audit TaskList -> get real status of all tasks
+2. Reconcile: session.completed_tasks <-> TaskList status (bidirectional sync)
+3. Reset any in_progress tasks -> pending (they were interrupted)
+4. Detect fast-advance orphans (in_progress without recent activity) -> reset to pending
+5. Determine remaining pipeline from reconciled state
+6. Rebuild team if disbanded (TeamCreate + spawn needed workers only)
+7. Create missing tasks, set dependencies via TaskUpdate({ addBlockedBy })
+8. Verify dependency chain integrity
+9. Update session file with reconciled state
+10. Kick first executable task's worker -> Phase 4
+
+---
+
+## Phase 1: Task Analysis
+
+**Objective**: Parse user task, detect capabilities, build dependency graph, design roles.
+
+**Constraint**: This is TEXT-LEVEL analysis only. No source code reading, no codebase exploration.
+
+**Workflow**:
+
+1. **Parse user task description**
+
+2. **Clarify if ambiguous** via AskUserQuestion:
+   - What is the scope? (specific files, module, project-wide)
+   - What deliverables are expected? (documents, code, analysis reports)
+   - Any constraints? (timeline, technology, style)
+
+3. **Delegate to `@commands/analyze-task.md`**:
+   - Signal detection: scan keywords -> infer capabilities
+   - Artifact inference: each capability -> default output type (.md)
+   - Dependency graph: build DAG of work streams
+   - Complexity scoring: count capabilities, cross-domain factor, parallel tracks
+   - Role minimization: merge overlapping, absorb trivial, cap at 5
+   - **Role-spec metadata**: Generate frontmatter fields (prefix, inner_loop, additional_members, message_types)
+
+4. **Output**: Write `<session>/task-analysis.json`
+
+5. **If `needs_research: true`**: Phase 2 will spawn researcher worker first
+
+**Success**: Task analyzed, capabilities detected, dependency graph built, roles designed with role-spec metadata.
+
+**CRITICAL - Team Workflow Enforcement**:
+
+Regardless of complexity score or role count, coordinator MUST:
+- ✅ **Always proceed to Phase 2** (generate role-specs)
+- ✅ **Always create team** and spawn workers via team-worker agent
+- ❌ **NEVER execute task work directly**, even for single-role low-complexity tasks
+- ❌ **NEVER skip team workflow** based on complexity assessment
+
+**Single-role execution is still team-based** - just with one worker. The team architecture provides:
+- Consistent message bus communication
+- Session state management
+- Artifact tracking
+- Fast-advance capability
+- Resume/recovery mechanisms
+
+---
+
+## Phase 2: Generate Role-Specs + Initialize Session
+
+**Objective**: Create session, generate dynamic role-spec files, initialize shared infrastructure.
+
+**Workflow**:
+
+1. Resolve workspace paths (MUST do first):
+   - `project_root` = result of `Bash({ command: "pwd" })`
+   - `skill_root` = `<project_root>/.claude/skills/team-coordinate`
+
+2. **Check `needs_research` flag** from task-analysis.json:
+   - If `true`: **Spawn researcher worker first** to gather codebase context
+     - Wait for researcher callback
+     - Merge research findings into task context
+     - Update task-analysis.json with enriched context
+
+3. **Generate session ID**: `TC-<slug>-<date>` (slug from first 3 meaningful words of task)
+
+4. **Create session folder structure**:
+   ```
+   .workflow/.team/<session-id>/
+   +-- role-specs/
+   +-- artifacts/
+   +-- wisdom/
+   +-- explorations/
+   +-- discussions/
+   +-- .msg/
+   ```
+
+5. **Call TeamCreate** with team name derived from session ID
+
+6. **Read `specs/role-spec-template.md`** for Behavioral Traits + Reference Patterns
+
+7. **For each role in task-analysis.json#roles**:
+   - Fill YAML frontmatter: role, prefix, inner_loop, additional_members, message_types
+   - **Compose Phase 2-4 content** (NOT copy from template):
+     - Phase 2: Derive input sources and context loading steps from **task description + upstream dependencies**
+     - Phase 3: Describe **execution goal** (WHAT to achieve) from task description — do NOT prescribe specific CLI tool or approach
+     - Phase 4: Combine **Behavioral Traits** (from template) + **output_type** (from task analysis) to compose verification steps
+     - Reference Patterns may guide phase structure, but task description determines specific content
+   - Write generated role-spec to `<session>/role-specs/<role-name>.md`
+
+8. **Register roles** in team-session.json#roles (with `role_spec` path instead of `role_file`)
+
+9. **Initialize shared infrastructure**:
+   - `wisdom/learnings.md`, `wisdom/decisions.md`, `wisdom/issues.md` (empty with headers)
+   - `explorations/cache-index.json` (`{ "entries": [] }`)
+   - `discussions/` (empty directory)
+
+10. **Initialize pipeline metadata** via team_msg:
+```typescript
+// 使用 team_msg 将 pipeline 元数据写入 .msg/meta.json
+// 注意: 此处为动态角色，执行时需将 <placeholders> 替换为 task-analysis.json 中生成的实际角色列表
+mcp__maestro__team_msg({
+  operation: "log",
+  session_id: "<session-id>",
+  from: "coordinator",
+  type: "state_update",
+  summary: "Session initialized",
+  data: {
+    pipeline_mode: "<mode>",
+    pipeline_stages: ["<role1>", "<role2>", "<...dynamic-roles>"],
+    roles: ["coordinator", "<role1>", "<role2>", "<...dynamic-roles>"],
+    team_name: "<team-name>" // 从 session ID 或任务描述中提取
+  }
+})
+```
+
+11. **Write team-session.json** with: session_id, task_description, status="active", roles, pipeline (empty), active_workers=[], completion_action="interactive", created_at
+
+**Success**: Session created, role-spec files generated, shared infrastructure initialized.
+
+---
+
+## Phase 3: Create Task Chain
+
+**Objective**: Dispatch tasks based on dependency graph with proper dependencies.
+
+Delegate to `@commands/dispatch.md` which creates the full task chain:
+1. Reads dependency_graph from task-analysis.json
+2. Topological sorts tasks
+3. Creates tasks via TaskCreate, then sets dependencies via TaskUpdate({ addBlockedBy })
+4. Assigns owner based on role mapping from task-analysis.json
+5. Includes `Session: <session-folder>` in every task description
+6. Sets InnerLoop flag for multi-task roles
+7. Updates team-session.json with pipeline and tasks_total
+
+**Success**: All tasks created with correct dependency chains, session updated.
+
+---
+
+## Phase 4: Spawn-and-Stop
+
+**Objective**: Spawn first batch of ready workers in background, then STOP.
+
+**Design**: Spawn-and-Stop + Callback pattern, with worker fast-advance.
+
+**Workflow**:
+1. Load `@commands/monitor.md`
+2. Find tasks with: status=pending, blockedBy all resolved, owner assigned
+3. For each ready task -> spawn team-worker (see SKILL.md Coordinator Spawn Template)
+4. Output status summary with execution graph
+5. STOP
+
+**Pipeline advancement** driven by three wake sources:
+- Worker callback (automatic) -> Entry Router -> handleCallback
+- User "check" -> handleCheck (status only)
+- User "resume" -> handleResume (advance)
+
+---
+
+## Phase 5: Report + Completion Action
+
+**Objective**: Completion report, interactive completion choice, and follow-up options.
+
+**Workflow**:
+1. Load session state -> count completed tasks, duration
+2. List all deliverables with output paths in `<session>/artifacts/`
+3. Include discussion summaries (if inline discuss was used)
+4. Summarize wisdom accumulated during execution
+5. Output report:
+
+```
+[coordinator] ============================================
+[coordinator] TASK COMPLETE
+[coordinator]
+[coordinator] Deliverables:
+[coordinator]   - <artifact-1.md> (<producer role>)
+[coordinator]   - <artifact-2.md> (<producer role>)
+[coordinator]
+[coordinator] Pipeline: <completed>/<total> tasks
+[coordinator] Roles: <role-list>
+[coordinator] Duration: <elapsed>
+[coordinator]
+[coordinator] Session: <session-folder>
+[coordinator] ============================================
+```
+
+6. **Execute Completion Action** (based on session.completion_action):
+
+| Mode | Behavior |
+|------|----------|
+| `interactive` | AskUserQuestion with Archive/Keep/Export options |
+| `auto_archive` | Execute Archive & Clean without prompt |
+| `auto_keep` | Execute Keep Active without prompt |
+
+**Interactive handler**: See SKILL.md Completion Action section.
+
+---
+
+## Error Handling
+
+| Error | Resolution |
+|-------|------------|
+| Task timeout | Log, mark failed, ask user to retry or skip |
+| Worker crash | Respawn worker, reassign task |
+| Dependency cycle | Detect in task analysis, report to user, halt |
+| Task description too vague | AskUserQuestion for clarification |
+| Session corruption | Attempt recovery, fallback to manual reconciliation |
+| Role-spec generation fails | Fall back to single general-purpose role |
+| capability_gap reported | handleAdapt: generate new role-spec, create tasks, spawn |
+| All capabilities merge to one | Valid: single-role execution, reduced overhead |
+| No capabilities detected | Default to single general role with TASK prefix |
+| Completion action fails | Default to Keep Active, log warning |