Team Coordinate

Universal team coordination skill: analyze task -> generate role-specs -> dispatch -> execute -> deliver. Only the coordinator is built-in. All worker roles are dynamically generated as lightweight role-spec files and spawned via the

team-worker

Architecture

+---------------------------------------------------+
|  Skill(skill="team-coordinate")                 |
|  args="task description"                           |
+-------------------+-------------------------------+
                    |
         Orchestration Mode (auto -> coordinator)
                    |
              Coordinator (built-in)
              Phase 0-5 orchestration
                    |
    +-------+-------+-------+-------+
    v       v       v       v       v
 [team-worker agents, each loaded with a dynamic role-spec]
  (roles generated at runtime from task analysis)

  CLI Tools (callable by any worker):
    ccw cli --mode analysis  - analysis and exploration
    ccw cli --mode write     - code generation and modification

Delegation Lock

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

Before calling ANY tool, apply this check:

spawn_agent

wait_agent

close_agent

send_message

followup_task

list_agents

request_user_input

mcp__ccw-tools__team_msg

Read/Write

.workflow/.team/

Read

roles/

commands/

specs/

Read/Grep/Glob

Edit

.workflow/

Bash("ccw cli ...")

Bash

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. The overhead is the feature — it provides session tracking, artifact persistence, and resume capability.

Shared Constants

TC

.workflow/.team/TC-<date>-<slug>/

team-worker

mcp__ccw-tools__team_msg(session_id=<session-id>, ...)

ccw cli --mode analysis

ccw cli --mode write

Role Router

This skill is coordinator-only. Workers do NOT invoke this skill -- they are spawned as

team-worker

Input Parsing

Parse

$ARGUMENTS

--role

Role Registry

Only coordinator is statically registered. All other roles are dynamic, stored as role-specs in session.

<session>/role-specs/<role-name>.md

CLI Tool Usage

Workers can use CLI tools for analysis and code operations:

Dispatch

Always route to coordinator. Coordinator reads

roles/coordinator/role.md

Orchestration Mode

User just provides task description.

Invocation:

Skill(skill="team-coordinate", args="task description")

Lifecycle:

User provides task description
  -> coordinator Phase 1: task analysis (detect capabilities, build dependency graph)
  -> coordinator Phase 2: generate role-specs + initialize session
  -> coordinator Phase 3: create task chain from dependency graph
  -> coordinator Phase 4: spawn first batch workers (background) -> STOP
  -> Worker executes -> callback -> coordinator advances next step
  -> Loop until pipeline complete -> Phase 5 report + completion action

User Commands (wake paused coordinator):

check

status

resume

continue

revise <TASK-ID> [feedback]

feedback <text>

improve [dimension]

Coordinator Spawn Template

v2 Worker Spawn (all roles)

When coordinator spawns workers, use

team-worker

spawn_agent({
  agent_type: "team_worker",
  task_name: "<task-id>",
  fork_turns: "none",
  message: `## Role Assignment
role: <role>
role_spec: <session-folder>/role-specs/<role>.md
session: <session-folder>
session_id: <session-id>
requirement: <task-description>
inner_loop: <true|false>

Read role_spec file 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 })

result.timed_out

close_agent({ target: <name> })

Inner Loop roles (role has 2+ serial same-prefix tasks): Set

inner_loop: true

Single-task roles: Set

inner_loop: false

Model Selection Guide

Roles are dynamically generated at runtime. Select model/reasoning_effort based on the generated role's

responsibility_type

Map each generated role's

responsibility_type

team-session.json#roles

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: "..."
})

v4 Agent Coordination

Message Semantics

send_message

followup_task

list_agents

Note: Since roles are dynamically generated, the coordinator must resolve task prefixes and role names from

team-session.json#roles

fork_turns Strategy

fork_turns: "none"

fork_turns: "all"

• Runtime analysis reveals the task requires deep familiarity with the full conversation context
• The dynamically-generated role-spec indicates the worker needs project-wide understanding
• The coordinator has already accumulated significant context about the codebase

This decision should be made per-task during Phase 4 based on the role's

responsibility_type

Agent Health Check

Use

list_agents({})

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

Named Agent Targeting

Workers are spawned with

task_name: "<task-id>"

• send_message({ target: "<TASK-ID>", message: "..." })

  -- queue upstream context without interrupting

• close_agent({ target: "<TASK-ID>" })

  -- cleanup by name

Completion Action

When pipeline completes (all tasks done), coordinator presents an interactive choice:

functions.request_user_input({
  questions: [{
    question: "Team pipeline complete. What would you like to do?",
    header: "Completion",
    multiSelect: false,
    options: [
      { label: "Archive & Clean (Recommended)", description: "Archive session, clean up team" },
      { label: "Keep Active", description: "Keep session for follow-up work" },
      { label: "Export Results", description: "Export deliverables to target directory, then clean" }
    ]
  }]
})

Action Handlers

Specs Reference

Session Directory

.workflow/.team/TC-<date>-<slug>/
+-- team-session.json           # Session state + dynamic role registry
+-- task-analysis.json          # Phase 1 output: capabilities, dependency graph
+-- role-specs/                 # Dynamic role-spec definitions (generated Phase 2)
|   +-- <role-1>.md             # Lightweight: frontmatter + Phase 2-4 only
|   +-- <role-2>.md
+-- artifacts/                  # All MD deliverables from workers
|   +-- <artifact>.md
+-- .msg/                       # Team message bus + state
|   +-- messages.jsonl          # Message log
|   +-- meta.json               # Session metadata + cross-role state
+-- wisdom/                     # Cross-task knowledge
|   +-- learnings.md
|   +-- decisions.md
|   +-- issues.md
+-- explorations/               # Shared explore cache
|   +-- cache-index.json
|   +-- explore-<angle>.json
+-- discussions/                # Inline discuss records
|   +-- <round>.md

team-session.json Schema

{
  "session_id": "TC-<date>-<slug>",
  "task_description": "<original user input>",
  "status": "active | paused | completed",
  "team_name": "<team-name>",
  "roles": [
    {
      "name": "<role-name>",
      "prefix": "<PREFIX>",
      "responsibility_type": "<type>",
      "inner_loop": false,
      "role_spec": "role-specs/<role-name>.md"
    }
  ],
  "pipeline": {
    "dependency_graph": {},
    "tasks_total": 0,
    "tasks_completed": 0
  },
  "active_workers": [],
  "completed_tasks": [],
  "completion_action": "interactive",
  "created_at": "<timestamp>"
}

Session Resume

Coordinator supports

resume

continue

1. Scan

   .workflow/.team/TC-*/team-session.json

   for active/paused sessions
2. Multiple matches -> request_user_input for selection
3. Audit task states -> reconcile session state <-> task status
4. Reset in_progress -> pending (interrupted tasks)
5. Rebuild team and spawn needed workers only
6. Create missing tasks, set dependencies
7. Kick first executable task -> Phase 4 coordination loop

Error Handling