Coordinator Helper

Manage the coordinator daemon for autonomous task delegation, approval workflows, and monitoring agent work.

Quick Start

Most common usage:

# User says: "Delegate this bug fix to an agent"
# This skill will:
# 1. Check if coordinator daemon is running
# 2. Send the task via ailang messages
# 3. Monitor the task status
# 4. Guide you through approval when complete

# User says: "What tasks are pending?"
# This skill will:
# 1. Run ailang coordinator list --pending
# 2. Show interactive task explorer
# 3. Let you review diffs, logs, and approve/reject

When to Use This Skill

Invoke this skill when:

• User asks to "delegate a task" or "send to coordinator"
• User wants to "check task status" or "see what's running"
• User asks to "review agent work" or "approve/reject tasks"
• User says "start the coordinator" or "stop the daemon"
• User wants to "clean up worktrees" or manage coordinator state

Available Scripts

scripts/check_daemon.sh

Check if the coordinator daemon is running and show status.

scripts/delegate_task.sh <type> <title> <description>

Send a task to the coordinator for autonomous execution.

scripts/quick_status.sh

Show a quick summary of pending, running, and completed tasks.

Core Commands

Starting/Stopping the Daemon

# Start coordinator + server (recommended)
make services-start

# Or just coordinator
ailang coordinator start

# Check status
ailang coordinator status

# Stop all
make services-stop

Delegating Tasks

# Send a task
ailang messages send coordinator "Fix the null pointer bug in parser.go" \
  --title "Bug: Parser NPE" --from "claude-code" --type bug

Monitoring & Approving

# Interactive task list
ailang coordinator list

# Filter by status
ailang coordinator list --pending
ailang coordinator list --running

# Approve from list: select task, press [a]
# Or directly: ailang coordinator approve <task-id>

Task Lifecycle

pending → queued → running → pending_approval → completed
                          ↘ failed
                          ↘ rejected → [feedback] → pending (iteration 2) → running → ...

Feedback Loop (v0.6.4+): When rejecting, the task can be re-triggered with feedback up to 3 iterations. Claude uses --resume to continue with full conversation context.

Unified Approvals (v0.6.5+)

When an agent has trigger_on_complete configured with auto_approve_handoffs: false, approvals are combined:

| Approval Type | Description | On Approve | |--------------|-------------|------------| | merge | Simple merge only | Merges code to dev branch | | merge_handoff | Combined merge + handoff | Merges code AND triggers next agent |

CLI display shows:

⏳ [1] [merge+handoff] → sprint-planner  task-12345678
       Title: Agent completed work on: Fix parser bug

What happens on approve:

1. Code is merged to dev branch
2. Handoff message is sent to next agent's inbox with session_id for continuity
3. Worktree is cleaned up

What happens on reject:

1. Worktree is preserved
2. Feedback is sent to same agent's inbox
3. Agent resumes with --resume <sessionId> (same context, same worktree)
4. Iteration counter increments (max 3 attempts)

GitHub-Driven Workflow (v0.6.2+)

For tasks linked to GitHub issues, the coordinator supports a fully GitHub-native approval workflow.

How It Works

GitHub Issue
    ↓ (import)
DESIGN STAGE → posts design doc to GitHub → needs-design-approval label
    ↓ (human adds: design-approved)
SPRINT STAGE → posts sprint plan to GitHub → needs-sprint-approval label
    ↓ (human adds: sprint-approved)
IMPLEMENTATION → posts file changes → needs-merge-approval label
    ↓ (human adds: merge-approved)
Changes merged, issue auto-closed

GitHub Labels Reference

| You Add This Label | What Happens | |--------------------|--------------| | design-approved | Advances to sprint planning | | sprint-approved | Advances to implementation | | merge-approved | Merges changes, closes issue | | needs-revision | Pauses pipeline for changes |

Quick Commands for GitHub Workflow

# Import GitHub issues as tasks
ailang messages import-github

# Check which issues are being watched
tail -100 ~/.ailang/logs/coordinator.log | grep -i "watching issue"

# Fallback: approve locally if labels aren't detected
ailang coordinator approve <task-id>

# Check pending approvals
ailang coordinator pending

Why Use GitHub Workflow?

• Review in GitHub UI - See design docs and diffs alongside issue discussion
• Mobile-friendly - Approve from GitHub mobile app
• Team collaboration - Multiple reviewers can discuss in comments
• Audit trail - All approvals tracked in issue history

Workflow

1. Delegate a Task

1. Describe clearly - Be specific about what needs to be done
2. Choose type - bug, feature, docs, research, refactor, test
3. Send message - ailang messages send coordinator "..." --type bug

2. Review Completed Work

1. Open explorer - ailang coordinator list
2. Select task - Enter task number
3. Review:
   • [c] View chat history (turn-by-turn conversation with tool calls)
   • [d] View diff
   • [f] Browse files
   • [l] View logs
4. Decide:
   • [a] Approve - merge changes to dev branch
   • [r] Reject - prompt for feedback, re-trigger task with context

3. Task Routing

| Type | Executor | Use Case | |------|----------|----------| | bug-fix | Claude Code | Code fixes | | feature | Claude Code | New functionality | | docs | Gemini | Documentation | | research | Gemini | Investigation | | script | Shell | Deterministic workflows (v0.6.4+) |

4. Script Agents (v0.6.4+)

For deterministic tasks that don't need AI inference:

# In ~/.ailang/config.yaml
coordinator:
  agents:
    - id: echo-demo
      inbox: echo-demo
      invoke:
        type: script
        command: "./scripts/coordinator/echo_payload.sh"
        env_from_payload: true
        timeout: "1m"
      output_markers:
        - "ECHO_COMPLETE:"

Test the demo:

ailang messages send echo-demo '{"model": "gpt5", "benchmark": "fizzbuzz"}' \
  --title "Echo test" --from "user"

What happens:

• JSON {"model": "gpt5"} → env var MODEL=gpt5
• Nested JSON {"db": {"host": "x"}} → env var DB_HOST=x
• Auto-injected: AILANG_TASK_ID, AILANG_MESSAGE_ID, AILANG_WORKSPACE
• Cost: $0.00 (no AI inference)

Chain Execution Monitoring (v0.8.1+)

ailang chains is the canonical CLI for examining multi-agent workflows. Works offline (direct SQLite).

Find a Chain

# Find by coordinator task ID
ailang chains find --task-id task-29404032

# Find by message UUID
ailang chains find --message-id 29404032-74b3-40c6-acc3-23d6bbe14b68

# Find by GitHub issue
ailang chains find --github sunholo-data/ailang#131
ailang chains find --github sunholo-data/ailang#131 --json

Filter Chains

# By agent
ailang chains list --agent design-doc-creator
ailang chains list --agent sprint-executor --since 7d

# By time
ailang chains list --since 24h        # Last 24 hours
ailang chains list --since 7d         # Last week
ailang chains list --since 2026-02-01 # Since specific date

# Combined
ailang chains list --agent sprint-executor --status failed --since 7d --json

View Code Changes

# Git diff across all stages in a chain
ailang chains diff <chain-id>

# Diffstat summary only
ailang chains diff <chain-id> --stat

Recommended Timeouts by Agent Type

Two timeout types (v0.8.1+):

• timeout (hard ceiling): Max wall-clock time, regardless of activity. Default: 60m.
• idle_timeout: Kill if agent produces no output for this long. Default: 3m.

An agent actively writing code for 45 minutes stays alive. An agent that crashes or loops silently dies after 3 minutes.

Configure in ~/.ailang/config.yaml:

coordinator:
  agents:
    - id: design-doc-creator
      timeout: "20m"         # Hard ceiling
      idle_timeout: "3m"     # Kill if no output for 3m
    - id: sprint-planner
      timeout: "15m"
      idle_timeout: "2m"
    - id: sprint-executor
      timeout: "60m"         # Allow up to 1 hour
      idle_timeout: "5m"     # But kill if stuck for 5m

Auditing Agent Work

After a task completes, audit what the agent actually did before approving:

# View conversation per turn (shows agent reasoning + tool calls)
ailang coordinator logs <task-id> --limit 1000 --json | python3 -c "
import json, sys
data = json.load(sys.stdin)
events = data.get('events', [])
turns = {}; tools = {}
for evt in events:
    tn = evt.get('turn_num', 0); st = evt.get('stream_type', '')
    if st == 'text': turns.setdefault(tn, []).append(evt.get('text', ''))
    elif st == 'tool_use': tools.setdefault(tn, []).append(evt.get('tool_name', '?'))
for tn in sorted(turns.keys()):
    text = ''.join(turns[tn]).strip()
    if len(text) > 20:
        print(f'=== Turn {tn} (tools: {\", \".join(tools.get(tn, []))}) ===')
        print(text[:600]); print()
"

# View chain execution flow (recommended)
ailang chains view <chain-id> --spans
ailang chains tree <chain-id> --detailed

# View tool timeline with spans (requires server)
ailang dashboard spans --task-id <task-id> --limit 200

# View git changes
ailang coordinator diff <task-id>

Audit checklist:

• [ ] Did the agent modify internal/ code or just create examples/docs?
• [ ] What model was used? (Check executor.model in spans - Haiku may be too weak)
• [ ] Did it run ailang run (runtime test) or just ailang check (compile test)?
• [ ] Did it mark tasks as "already working" without verifying the specific bug scenario?

Per-agent model config (v0.8.0+): Set model: opus in agent config for complex coding tasks:

agents:
  - id: sprint-executor
    model: opus

Troubleshooting

Daemon won't start: Check ailang coordinator status, then make services-stop && make services-start

Task stuck: View logs with [l] in task explorer

Worktree limit: git worktree list then git worktree remove <path> --force

GitHub labels not detected: The ApprovalWatcher may not be detecting labels. Use CLI fallback:

ailang coordinator pending    # List tasks waiting for approval
ailang coordinator approve <task-id>   # Approve locally (syncs label to GitHub)

No logs from ApprovalWatcher: Check coordinator logs for "GitHub approval watcher started". If missing, verify ~/.ailang/config.yaml has github_sync.enabled: true.

Resources

See resources/reference.md for complete CLI reference and advanced options.

Notes

• Coordinator uses isolated git worktrees per task
• Worktrees auto-cleanup after approval
• Events stream to dashboard at http://localhost:1957
• State stored in ~/.ailang/state/coordinator.db