Files
Claude-Code-Workflow/.codex/skills/team-iterdev/SKILL.md
catlog22 88ea7fc6d7 refactor: deep Codex v4 API conversion for all 20 team skills
Upgrade all team-* skills from mechanical v3→v4 API renames to deep
v4 tool integration with skill-adaptive patterns:

- list_agents: health checks in handleResume, cleanup verification in
  handleComplete, added to allowed-tools and coordinator toolbox
- Named targeting: task_name uses task-id (e.g. EXPLORE-001) instead
  of generic <role>-worker, enabling send_message/assign_task by name
- Message semantics: send_message for supplementary cross-agent context
  vs assign_task for triggering work, with skill-specific examples
- Model selection: per-role reasoning_effort guidance matching each
  skill's actual roles (not generic boilerplate)
- timeout_ms: added to all wait_agent calls, timed_out handling in
  all 18 monitor.md files
- Skill-adaptive v4 sections: ultra-analyze N-parallel coordination,
  lifecycle-v4 supervisor assign_task/send_message distinction,
  brainstorm ideator parallel patterns, iterdev generator-critic loops,
  frontend-debug iterative debug assign_task, perf-opt benchmark
  context sharing, executor lightweight trimmed v4, etc.

60 files changed across 20 team skills (SKILL.md, monitor.md, role.md)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-27 22:25:32 +08:00

8.1 KiB

name, description, allowed-tools
name description allowed-tools
team-iterdev Unified team skill for iterative development team. Pure router — all roles read this file. Beat model is coordinator-only in monitor.md. Generator-Critic loops (developer<->reviewer, max 3 rounds). Triggers on "team iterdev". spawn_agent(*), wait_agent(*), send_message(*), assign_task(*), close_agent(*), list_agents(*), report_agent_job_result(*), request_user_input(*), Read(*), Write(*), Edit(*), Bash(*), Glob(*), Grep(*)

Team IterDev

Iterative development team skill. Generator-Critic loops (developer<->reviewer, max 3 rounds), task ledger (task-ledger.json) for real-time progress, shared memory (cross-sprint learning), and dynamic pipeline selection for incremental delivery.

Architecture

Skill(skill="team-iterdev", args="task description")
                    |
         SKILL.md (this file) = Router
                    |
     +--------------+--------------+
     |                             |
  no --role flag              --role <name>
     |                             |
  Coordinator                  Worker
  roles/coordinator/role.md    roles/<name>/role.md
     |
     +-- analyze -> dispatch -> spawn workers -> STOP
                                    |
                    +-------+-------+-------+
                    v       v       v       v
               [architect] [developer] [tester] [reviewer]
              (team-worker agents, each loads roles/<role>/role.md)

Role Registry

Role Path Prefix Inner Loop
coordinator roles/coordinator/role.md
architect roles/architect/role.md DESIGN-* false
developer roles/developer/role.md DEV-* true
tester roles/tester/role.md VERIFY-* false
reviewer roles/reviewer/role.md REVIEW-* false

Role Router

Parse $ARGUMENTS:

  • Has --role <name> → Read roles/<name>/role.md, execute Phase 2-4
  • No --roleroles/coordinator/role.md, execute entry router

Delegation Lock

Coordinator is a PURE ORCHESTRATOR. It coordinates, it does NOT do.

Before calling ANY tool, apply this check:

Tool Call Verdict Reason
spawn_agent, wait_agent, close_agent, send_message, assign_task ALLOWED Orchestration
list_agents ALLOWED Agent health check
request_user_input ALLOWED User interaction
mcp__ccw-tools__team_msg ALLOWED Message bus
Read/Write on .workflow/.team/ files ALLOWED Session state
Read on roles/, commands/, specs/ ALLOWED Loading own instructions
Read/Grep/Glob on project source code BLOCKED Delegate to worker
Edit on any file outside .workflow/ BLOCKED Delegate to worker
Bash("ccw cli ...") BLOCKED Only workers call CLI
Bash running build/test/lint commands BLOCKED Delegate to worker

If a tool call is BLOCKED: STOP. Create a task, spawn a worker.

No exceptions for "simple" tasks. Even a single-file read-and-report MUST go through spawn_agent.


Shared Constants

  • Session prefix: IDS
  • Session path: .workflow/.team/IDS-<slug>-<date>/
  • CLI tools: ccw cli --mode analysis (read-only), ccw cli --mode write (modifications)
  • Message bus: mcp__ccw-tools__team_msg(session_id=<session-id>, ...)

Worker Spawn Template

Coordinator spawns workers using this template:

spawn_agent({
  agent_type: "team_worker",
  task_name: "<task-id>",
  fork_context: false,
  items: [
    { type: "text", text: `## Role Assignment
role: <role>
role_spec: <skill_root>/roles/<role>/role.md
session: <session-folder>
session_id: <session-id>
requirement: <task-description>
inner_loop: <true|false>

Read role_spec file (<skill_root>/roles/<role>/role.md) to load Phase 2-4 domain instructions.` },

    { type: "text", text: `## Task Context
task_id: <task-id>
title: <task-title>
description: <task-description>
pipeline_phase: <pipeline-phase>` },

    { type: "text", text: `## Upstream Context
<prev_context>` }
  ]
})

After spawning, use wait_agent({ targets: [...], timeout_ms: 900000 }) to collect results, then close_agent({ target }) each worker.

Model Selection Guide

Iterative development uses Generator-Critic loops. Developer and reviewer need high reasoning for code generation and review quality.

Role reasoning_effort Rationale
architect high Design decisions require careful tradeoff analysis
developer high Code generation needs precision, inner loop role
tester medium Test execution follows defined verification plan
reviewer high Critical review quality drives GC loop effectiveness

User Commands

Command Action
check / status View execution status graph
resume / continue Advance to next step

Session Directory

.workflow/.team/IDS-<slug>-<YYYY-MM-DD>/
├── .msg/
│   ├── messages.jsonl          # Team message bus
│   └── meta.json               # Session state
├── task-analysis.json          # Coordinator analyze output
├── task-ledger.json            # Real-time task progress ledger
├── wisdom/                     # Cross-task knowledge accumulation
│   ├── learnings.md
│   ├── decisions.md
│   ├── conventions.md
│   └── issues.md
├── design/                     # Architect output
│   ├── design-001.md
│   └── task-breakdown.json
├── code/                       # Developer tracking
│   └── dev-log.md
├── verify/                     # Tester output
│   └── verify-001.json
└── review/                     # Reviewer output
    └── review-001.md

Specs Reference

v4 Agent Coordination

Message Semantics

Intent API Example
Queue supplementary info (don't interrupt) send_message Send design context to running developer
Assign GC iteration round assign_task Assign revision to developer after reviewer feedback
Check running agents list_agents Verify agent health during resume

Agent Health Check

Use list_agents({}) in handleResume and handleComplete:

// Reconcile session state with actual running agents
const running = list_agents({})
// Compare with task-ledger.json active tasks
// Reset orphaned tasks (in_progress but agent gone) to pending

Named Agent Targeting

Workers are spawned with task_name: "<task-id>" enabling direct addressing:

  • send_message({ target: "DEV-001", items: [...] }) -- send design context to developer
  • assign_task({ target: "DEV-001", items: [...] }) -- assign revision after reviewer feedback
  • close_agent({ target: "REVIEW-001" }) -- cleanup after GC loop completes

Generator-Critic Loop via assign_task (Deep Interaction Pattern)

The developer-reviewer GC loop uses assign_task for iteration rounds:

// Round N: Reviewer found issues -> coordinator assigns revision to developer
assign_task({
  target: "DEV-001",
  items: [
    { type: "text", text: `## GC Revision Round ${N}
review_feedback: <reviewer findings from REVIEW-001>
iteration: ${N}/3
instruction: Address reviewer feedback. Focus on: <specific issues>.` }
  ]
})

After developer completes revision, coordinator spawns/assigns reviewer for next round. Max 3 rounds per GC cycle.

Error Handling

Scenario Resolution
Unknown command Error with available command list
Role not found Error with role registry
GC loop exceeds 3 rounds Accept with warning, record in shared memory
Sprint velocity drops below 50% Coordinator alerts user, suggests scope reduction
Task ledger corrupted Rebuild from TaskList state
Conflict detected Update conflict_info, notify coordinator, create DEV-fix task
Pipeline deadlock Check blockedBy chain, report blocking point