Team PlanEx

Unified team skill: plan-and-execute pipeline for issue-based development. Built on team-worker agent architecture — coordinator orchestrates, workers are team-worker agents loading role-specific instructions from roles/<role>/role.md.

Architecture

Skill(skill="team-planex", 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
               [planner]                       [executor]
         (team-worker agent,            (team-worker agent,
          loads roles/planner/role.md)   loads roles/executor/role.md)

Role Registry

| Role | Path | Prefix | Inner Loop | |------|------|--------|------------| | coordinator | roles/coordinator/role.md | — | — | | planner | roles/planner/role.md | PLAN- | true | | executor | [roles/executor/role.md](roles/executor/role.md) | EXEC- | true |

Role Router

Parse $ARGUMENTS:

• Has --role <name> -> Read roles/<name>/role.md, execute Phase 2-4
• No --role -> roles/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, followup_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: PEX
• Session path: .workflow/.team/PEX-<date>-<slug>/
• 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_turns: "none",
  message: `## 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>
execution_method: <codex|gemini>

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

## Task Context
task_id: <task-id>
title: <task-title>
description: <task-description>
pipeline_phase: <pipeline-phase>

## Upstream Context
<prev_context>`
})

After spawning, use wait_agent({ timeout_ms: 1800000 }) to collect results. If result.timed_out, send STATUS_CHECK via followup_task (wait 3 min), then FINALIZE with interrupt (wait 3 min), then mark timed_out and close agents. Use close_agent({ target }) each worker.

Model Selection Guide

| Role | model | reasoning_effort | Rationale | |------|-------|-------------------|-----------| | Planner (PLAN-) | (default) | high | Solution planning requires deep code analysis | | Executor (EXEC-) | (default) | high | Code implementation needs precision |

Override model/reasoning_effort in spawn_agent when cost optimization is needed:

spawn_agent({
  agent_type: "team_worker",
  task_name: "<task-id>",
  fork_turns: "none",
  model: "<model-override>",
  reasoning_effort: "<effort-level>",
  message: "..."
})

User Commands

| Command | Action | |---------|--------| | check / status | View execution status graph | | resume / continue | Advance to next step | | add <issue-ids or --text '...' or --plan path> | Append new tasks to planner queue |

Session Directory

.workflow/.team/PEX-<slug>-<YYYY-MM-DD>/
├── .msg/
│   ├── messages.jsonl          # Message bus log
│   └── meta.json               # Session state
├── task-analysis.json          # Coordinator analyze output
├── artifacts/
│   └── solutions/              # Planner solution output per issue
│       ├── <issueId-1>.json
│       └── <issueId-N>.json
└── wisdom/                     # Cross-task knowledge
    ├── learnings.md
    ├── decisions.md
    ├── conventions.md
    └── issues.md

Specs Reference

• specs/pipelines.md — Pipeline definitions, task metadata registry, execution method selection

v4 Agent Coordination

Message Semantics

| Intent | API | Example | |--------|-----|---------| | Send plan updates to running executor | send_message | Queue planner solution details to EXEC- workers | | Not used in this skill | followup_task | No resident agents -- planner and executor are one-shot | | Check running agents | list_agents | Verify planner/executor health during resume |

Two-Phase Pipeline Pattern

Plan-and-execute is a Two-Phase pattern: planner creates solution plans (PLAN-), then coordinator spawns executors (EXEC-) for each planned issue. The planner may dynamically create EXEC- task entries in tasks.json.

// Phase 1: Planner runs, creates EXEC-* tasks in tasks.json
spawn_agent({ agent_type: "team_worker", task_name: "PLAN-001", ... })
wait_agent({ timeout_ms: 1800000 })  // 30 min — apply timeout cascade if timed_out
// Phase 2: Executors run per-issue, may run in sequence or parallel
// Inner loop: planner/executor handle multiple tasks internally

Agent Health Check

Use list_agents({}) in handleResume and handleComplete:

// Reconcile session state with actual running agents
const running = list_agents({})
// Compare with tasks.json active_agents
// 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: "EXEC-001", message: "..." }) -- queue plan solution to running executor
• close_agent({ target: "PLAN-001" }) -- cleanup by name after completion

Error Handling

| Scenario | Resolution | |----------|------------| | Unknown command | Error with available command list | | Role not found | Error with role registry | | Role spec file not found | Error with expected path (roles/<name>/role.md) | | team-worker agent unavailable | Error: requires .claude/agents/team-worker.md | | Planner issue planning failure | Retry once, then skip to next issue | | Executor impl failure | Report to coordinator, continue with next EXEC- task | | Pipeline stall | Coordinator monitors, escalate to user | | Worker no response | Report waiting task, suggest user resume |