Team Lifecycle v4

Unified team skill: specification -> implementation -> testing -> review. Optimized from v3 with inline discuss subagent and shared explore utility, halving spec pipeline beats from 12 to 6.

Architecture

+---------------------------------------------------+
|  Skill(skill="team-lifecycle-v4")                  |
|  args="task description" or args="--role=xxx"      |
+-------------------+-------------------------------+
                    | Role Router
         +---- --role present? ----+
         | NO                      | YES
         v                         v
  Orchestration Mode         Role Dispatch
  (auto -> coordinator)      (route to role.md)
         |
    +----+----+-------+-------+-------+-------+
    v         v       v       v       v       v
 coordinator analyst writer planner executor tester
                                         ^        ^
                                on-demand by coordinator
                              +---------+ +--------+
                              |architect| |reviewer|
                              +---------+ +--------+
                              +-------------+ +-----+
                              |fe-developer | |fe-qa|
                              +-------------+ +-----+

  Subagents (callable by any role, not team members):
    [discuss-subagent]  - multi-perspective critique
    [explore-subagent]  - codebase exploration with cache

Role Router

Input Parsing

Parse $ARGUMENTS to extract --role. If absent -> Orchestration Mode (auto route to coordinator).

Role Registry

COMPACT PROTECTION: Role files are execution documents. After context compression, role instructions become summaries only — MUST immediately Read the role.md to reload before continuing. Never execute any Phase based on summaries.

Subagent Registry

Dispatch

1. Extract --role from arguments
2. If no --role -> route to coordinator (Orchestration Mode)
3. Look up role in registry -> Read the role file -> Execute its phases

Orchestration Mode

When invoked without --role, coordinator auto-starts. User just provides task description.

Invocation: Skill(skill="team-lifecycle-v4", args="task description")

Lifecycle:

User provides task description
  -> coordinator Phase 1-3: requirement clarification -> TeamCreate -> create task chain
  -> coordinator Phase 4: spawn first batch workers (background) -> STOP
  -> Worker executes -> SendMessage callback -> coordinator advances next step
  -> Loop until pipeline complete -> Phase 5 report

User Commands (wake paused coordinator):

Shared Infrastructure

The following templates apply to all worker roles. Each role.md only needs to write Phase 2-4 role-specific logic.

Worker Phase 1: Task Discovery (all workers shared)

Each worker on startup executes the same task discovery flow:

1. Call TaskList() to get all tasks
2. Filter: subject matches this role’s prefix + owner is this role + status is pending + blockedBy is empty
3. No tasks -> idle wait
4. Has tasks -> TaskGet for details -> TaskUpdate mark in_progress

Resume Artifact Check (prevent duplicate output after resume):

• Check if this task’s output artifacts already exist
• Artifacts complete -> skip to Phase 5 report completion
• Artifacts incomplete or missing -> normal Phase 2-4 execution

Worker Phase 5: Report + Fast-Advance (all workers shared)

Task completion with optional fast-advance to skip coordinator round-trip:

1. Message Bus: Call mcp__ccw-tools__team_msg to log message
   • Params: operation=”log”, team=, from=, to=”coordinator”, type=, summary=”[] “, ref=
   • team must be session ID (e.g., TLS-my-project-2026-02-27), NOT team name. Extract from task description Session: field → take folder name.
   • CLI fallback: ccw team log --team <session-id> --from <role> --to coordinator --type <type> --summary "[<role>] ..." --json
2. TaskUpdate: Mark task completed
3. Fast-Advance Check:
   • Call TaskList(), find pending tasks whose blockedBy are ALL completed
   • If exactly 1 ready task AND its owner matches a simple successor pattern -> spawn it directly (skip coordinator)
   • Otherwise -> SendMessage to coordinator for orchestration
4. Loop: Back to Phase 1 to check for next task

Fast-Advance Rules:

Fast-advance failure recovery: If a fast-advanced task fails (worker exits without completing), the coordinator detects it as an orphaned in_progress task on next resume/check and resets it to pending for re-spawn. Self-healing, no manual intervention required. See monitor.md Fast-Advance Failure Recovery.

Worker Inner Loop (同前缀多任务角色)

适用角色：writer (DRAFT-)、planner (PLAN-)、executor (IMPL-*)

当一个角色拥有同前缀的多个串行任务时，不再每完成一个就 spawn 新 agent，而是在同一 agent 内循环处理：

Inner Loop 流程：

Phase 1: 发现任务 (首次)
  │
  ├─ 找到任务 → Phase 2-3: 加载上下文 + Subagent 执行
  │                │
  │                v
  │          Phase 4: 验证 (+ Inline Discuss if applicable)
  │                │
  │                v
  │          Phase 5-L: 轻量完成 (Loop variant)
  │                │
  │                ├─ TaskUpdate 标完成
  │                ├─ team_msg 记录
  │                ├─ 累积摘要到 context_accumulator
  │                │
  │                ├─ 检查：还有同前缀待处理任务？
  │                │   ├─ YES → 回到 Phase 1 (内循环)
  │                │   └─ NO → Phase 5-F: 最终报告
  │                │
  │                └─ 异常中断条件？
  │                    ├─ consensus_blocked HIGH → SendMessage → STOP
  │                    └─ 错误累计 ≥ 3 → SendMessage → STOP
  │
  └─ Phase 5-F: 最终报告 (Final)
       ├─ SendMessage (含全部任务摘要)
       └─ STOP

context_accumulator (主 agent 上下文中维护，不写文件):

每个 subagent 返回后，主 agent 将结果压缩为摘要追加到 accumulator：

context_accumulator = []

# DRAFT-001 subagent 返回后
context_accumulator.append({
  task: "DRAFT-001",
  artifact: "spec/product-brief.md",
  key_decisions: ["聚焦 B2B 场景", "MVP 不含移动端"],
  discuss_verdict: "consensus_reached",
  discuss_rating: 4.2
})

# DRAFT-002 subagent 返回后
context_accumulator.append({
  task: "DRAFT-002",
  artifact: "spec/requirements/_index.md",
  key_decisions: ["REQ-003 降级为 P2", "NFR-perf 新增 SLA"],
  discuss_verdict: "consensus_reached",
  discuss_rating: 3.8
})

后续 subagent 调用时，将 accumulator 摘要作为 CONTEXT 传入，实现知识传递。

Phase 5-L vs Phase 5-F 区别：

中断恢复：

如果 Inner Loop agent 在 DRAFT-003 崩溃：

1. DRAFT-001, DRAFT-002 已落盘 + 已标完成 → 安全
2. DRAFT-003 状态为 in_progress → coordinator resume 时检测到无 active_worker → 重置为 pending
3. 重新 spawn writer → Phase 1 找到 DRAFT-003 → Resume Artifact Check:
   • DRAFT-003 产物不存在 → 正常执行
   • DRAFT-003 产物已写但未标完成 → 验证后标完成
4. 新 writer 从 DRAFT-003 开始循环，丢失的只是 001+002 的隐性摘要（可从磁盘重建基础信息）

恢复增强 (可选)：在每个 Phase 5-L 后将 context_accumulator 写入 <session>/shared-memory.json 的 context_accumulator 字段，crash 后可读回。

Inline Discuss Protocol (produce roles: analyst, writer, reviewer)

After completing their primary output, produce roles call the discuss subagent inline:

Task({
  subagent_type: "cli-discuss-agent",
  run_in_background: false,
  description: "Discuss <round-id>",
  prompt: <see subagents/discuss-subagent.md for prompt template>
})

Discussion round mapping (which role runs which discuss round):

The discuss subagent writes its record to discussions/ and returns the verdict. The calling role includes the discuss result in its Phase 5 report.

Consensus-blocked handling (produce role responsibility):

SendMessage format for consensus_blocked:

[<role>] <task-id> complete. Discuss <round-id>: consensus_blocked (severity=<HIGH|MEDIUM>)
Divergences: <divergence-summary>
Action items: <top-3-items>
Recommendation: <revise|proceed-with-caution|escalate>

Coordinator response (see monitor.md Consensus-Blocked Handling for full flow):

• HIGH -> revision task (max 1 per task) or pause for user decision
• MEDIUM -> proceed with warning, log to wisdom/issues.md
• DISCUSS-006 HIGH -> always pause for user (final sign-off gate)

Shared Explore Utility

Any role needing codebase context calls the explore subagent:

Task({
  subagent_type: "cli-explore-agent",
  run_in_background: false,
  description: "Explore <angle>",
  prompt: <see subagents/explore-subagent.md for prompt template>
})

Cache: Results stored in explorations/ with cache-index.json. Before exploring, always check cache first. See subagents/explore-subagent.md.

Wisdom Accumulation (all roles)

Cross-task knowledge accumulation. Coordinator creates wisdom/ directory at session init.

Directory:

<session-folder>/wisdom/
+-- learnings.md      # Patterns and insights
+-- decisions.md      # Architecture and design decisions
+-- conventions.md    # Codebase conventions
+-- issues.md         # Known risks and issues

Worker load (Phase 2): Extract Session: <path> from task description, read wisdom files. Worker contribute (Phase 4/5): Write discoveries to corresponding wisdom files.

Role Isolation Rules

Coordinator additionally prohibited: directly write/modify code, call implementation subagents, directly execute analysis/test/review.

Pipeline Definitions

Spec-only (6 tasks, was 12 in v3)

RESEARCH-001(+D1) -> DRAFT-001(+D2) -> DRAFT-002(+D3) -> DRAFT-003(+D4) -> DRAFT-004(+D5) -> QUALITY-001(+D6)

Each task includes inline discuss. (+DN) = inline discuss round N.

Impl-only / Backend (4 tasks)

PLAN-001 -> IMPL-001 -> TEST-001 + REVIEW-001

Full-lifecycle (10 tasks, was 16 in v3)

[Spec pipeline] -> PLAN-001(blockedBy: QUALITY-001) -> IMPL-001 -> TEST-001 + REVIEW-001

Frontend Pipelines

FE-only:       PLAN-001 -> DEV-FE-001 -> QA-FE-001
               (GC loop: QA-FE verdict=NEEDS_FIX -> DEV-FE-002 -> QA-FE-002, max 2 rounds)

Fullstack:     PLAN-001 -> IMPL-001 || DEV-FE-001 -> TEST-001 || QA-FE-001 -> REVIEW-001

Full + FE:     [Spec pipeline] -> PLAN-001 -> IMPL-001 || DEV-FE-001 -> TEST-001 || QA-FE-001 -> REVIEW-001

Cadence Control

Beat model: Event-driven, each beat = coordinator wake -> process -> spawn -> STOP.

Beat Cycle (single beat)
======================================================================
  Event                   Coordinator              Workers
----------------------------------------------------------------------
  callback/resume --> +- handleCallback -+
                      |  mark completed   |
                      |  check pipeline   |
                      +- handleSpawnNext -+
                      |  find ready tasks |
                      |  spawn workers ---+--> [Worker A] Phase 1-5
                      |  (parallel OK)  --+--> [Worker B] Phase 1-5
                      +- STOP (idle) -----+         |
                                                     |
  callback <-----------------------------------------+
  (next beat)              SendMessage + TaskUpdate(completed)
======================================================================

  Fast-Advance (skips coordinator for simple linear successors)
======================================================================
  [Worker A] Phase 5 complete
    +- 1 ready task? simple successor? --> spawn Worker B directly
    +- complex case? --> SendMessage to coordinator
======================================================================

Pipeline beat view (v4 optimized):

Spec-only (6 beats, was 12 in v3)
-------------------------------------------------------
Beat  1         2         3         4         5         6
      |         |         |         |         |         |
    R1+D1 --> W1+D2 --> W2+D3 --> W3+D4 --> W4+D5 --> Q1+D6
    ^                                                     ^
  pipeline                                            sign-off
   start                                               pause

R=RESEARCH  W=DRAFT(writer)  Q=QUALITY  D=DISCUSS(inline)

Impl-only (3 beats, with parallel window)
-------------------------------------------------------
Beat  1         2              3
      |         |         +----+----+
      PLAN --> IMPL --> TEST || REVIEW    <-- parallel window
                         +----+----+
                           pipeline
                            done

Full-lifecycle (9 beats, was 15 in v3)
-------------------------------------------------------
Beat 1-6: [Spec pipeline as above]
                                    |
Beat 6 (Q1+D6 done):      PAUSE CHECKPOINT -- user confirms then resume
                                    |
Beat 7      8           9
 PLAN --> IMPL --> TEST || REVIEW

Fullstack (with dual parallel windows)
-------------------------------------------------------
Beat  1              2                    3                4
      |         +----+----+         +----+----+           |
      PLAN --> IMPL || DEV-FE --> TEST || QA-FE -->  REVIEW
              ^                ^                   ^
         parallel 1       parallel 2          sync barrier

Checkpoints:

Checkpoint Output Template (QUALITY-001 completion):

Coordinator reads <session>/spec/readiness-report.md, extracts gate + dimension scores, displays:

[coordinator] ══════════════════════════════════════════
[coordinator] SPEC PHASE COMPLETE
[coordinator] Quality Gate: <PASS|REVIEW|FAIL> (<score>%)
[coordinator]
[coordinator] Dimension Scores:
[coordinator]   Completeness:  <bar> <n>%
[coordinator]   Consistency:   <bar> <n>%
[coordinator]   Traceability:  <bar> <n>%
[coordinator]   Depth:         <bar> <n>%
[coordinator]   Coverage:      <bar> <n>%
[coordinator]
[coordinator] Available Actions:
[coordinator]   resume              → Proceed to implementation
[coordinator]   improve             → Auto-improve weakest dimension
[coordinator]   improve <dimension> → Improve specific dimension
[coordinator]   revise <TASK-ID>    → Revise specific document
[coordinator]   recheck             → Re-run quality check
[coordinator]   feedback <text>     → Inject feedback, create revision
[coordinator] ══════════════════════════════════════════

Gate-specific guidance:

• PASS: All actions available, resume is primary suggestion
• REVIEW: Recommend improve/revise before resume, warn on resume
• FAIL: Recommend improve/revise, do not suggest resume (user can force)

Stall detection (coordinator handleCheck):

Task Metadata Registry

Coordinator Spawn Template

标准 Worker (单任务角色: analyst, tester, reviewer, architect)

When coordinator spawns workers, use background mode (Spawn-and-Stop):

Task({
  subagent_type: "general-purpose",
  description: "Spawn <role> worker",
  team_name: <team-name>,
  name: "<role>",
  run_in_background: true,
  prompt: `You are team "<team-name>" <ROLE>.

## Primary Instruction
All your work MUST be executed by calling Skill to get role definition:
Skill(skill="team-lifecycle-v4", args="--role=<role>")

Current requirement: <task-description>
Session: <session-folder>

## Role Guidelines
- Only process <PREFIX>-* tasks, do not execute other role work
- All output prefixed with [<role>] tag
- Only communicate with coordinator
- Do not use TaskCreate to create tasks for other roles
- Before each SendMessage, call mcp__ccw-tools__team_msg to log (team=<session-id> from Session field, NOT team name)
- After task completion, check for fast-advance opportunity (see SKILL.md Phase 5)

## Workflow
1. Call Skill -> get role definition and execution logic
2. Follow role.md 5-Phase flow
3. team_msg(team=<session-id>) + SendMessage results to coordinator
4. TaskUpdate completed -> check next task or fast-advance`
})

Inner Loop Worker (多任务角色: writer, planner, executor)

Task({
  subagent_type: "general-purpose",
  description: "Spawn <role> worker (inner loop)",
  team_name: <team-name>,
  name: "<role>",
  run_in_background: true,
  prompt: `You are team "<team-name>" <ROLE>.

## Primary Instruction
All your work MUST be executed by calling Skill to get role definition:
Skill(skill="team-lifecycle-v4", args="--role=<role>")

Current requirement: <task-description>
Session: <session-folder>

## Inner Loop Mode
You will handle ALL <PREFIX>-* tasks in this session, not just the first one.
After completing each task, loop back to find the next <PREFIX>-* task.
Only SendMessage to coordinator when:
- All <PREFIX>-* tasks are done
- A consensus_blocked HIGH occurs
- Errors accumulate (≥ 3)

## Role Guidelines
- Only process <PREFIX>-* tasks, do not execute other role work
- All output prefixed with [<role>] tag
- Only communicate with coordinator
- Do not use TaskCreate to create tasks for other roles
- Before each SendMessage, call mcp__ccw-tools__team_msg to log (team=<session-id> from Session field, NOT team name)
- Use subagent calls for heavy work, retain summaries in context`
})

Session Directory

.workflow/.team/TLS-<slug>-<date>/
+-- team-session.json           # Session state
+-- spec/                       # Spec artifacts
|   +-- spec-config.json
|   +-- discovery-context.json
|   +-- product-brief.md
|   +-- requirements/
|   +-- architecture/
|   +-- epics/
|   +-- readiness-report.md
|   +-- spec-summary.md
+-- discussions/                # Discussion records (written by discuss subagent)
+-- plan/                       # Plan artifacts
|   +-- plan.json
|   +-- .task/TASK-*.json
+-- explorations/               # Shared explore cache
|   +-- cache-index.json        # { angle+keywords_hash -> file_path }
|   +-- explore-<angle>.json
+-- architecture/               # Architect assessments + design-tokens.json
+-- analysis/                   # Analyst design-intelligence.json (UI mode)
+-- qa/                         # QA audit reports
+-- wisdom/                     # Cross-task knowledge
|   +-- learnings.md
|   +-- decisions.md
|   +-- conventions.md
|   +-- issues.md
+-- .msg/                       # Team message bus logs (messages.jsonl)
+-- shared-memory.json          # Cross-role state

Session Resume

Coordinator supports --resume / --continue for interrupted sessions:

1. Scan .workflow/.team/TLS-*/team-session.json for active/paused sessions
2. Multiple matches -> AskUserQuestion for selection
3. Audit TaskList -> reconcile session state <-> task status
4. Reset in_progress -> pending (interrupted tasks)
5. Rebuild team and spawn needed workers only
6. Create missing tasks with correct blockedBy
7. Kick first executable task -> Phase 4 coordination loop

Shared Spec Resources

Error Handling