agile-coordinator
npx skills add https://github.com/jwynia/agent-skills --skill agile-coordinator
Agent 安装分布
Skill 文档
Agile Coordinator
Orchestrates multiple worker agents to implement groomed tasks from the backlog, handling task assignment, progress monitoring, merge coordination, and verification.
Core Principle
Coordinate, don’t implement. The coordinator assigns tasks to workers, monitors their progress, coordinates merges, and verifies results. Workers execute the actual implementation via the agile-workflow skill.
Quick Reference
When to Use
- Multiple ready tasks in the backlog need implementation
- You want autonomous batch execution of development work
- You need coordinated merges to avoid conflicts
- You want progress tracking across multiple tasks
Invocation
/agile-coordinator # Auto-discover and execute ready tasks
/agile-coordinator TASK-001 TASK-002 # Execute specific tasks
/agile-coordinator --dry-run # Preview execution plan only
/agile-coordinator --parallel # Run workers in parallel
/agile-coordinator --sequential # Run workers one at a time (default)
Flags
| Flag | Description | Default |
|---|---|---|
--sequential |
Execute tasks one at a time | Yes |
--parallel |
Execute tasks concurrently | No |
--max-workers N |
Maximum concurrent workers | 2 |
--dry-run |
Show plan without executing | No |
--autonomous |
Auto-continue at all checkpoints | Yes |
--supervised |
Pause after each task completes | No |
--verbose |
Show all worker updates | No |
--summary-only |
Show major milestones only | Yes |
Workflow Phases
Phase 1: Discovery
Read the backlog to find tasks ready for implementation.
Actions:
1. Read context/backlog/ for task files
2. Filter to status: ready
3. Parse task metadata (priority, size, dependencies)
4. Sort by priority (high â medium â low)
5. Present findings
Output: List of ready tasks with metadata
Phase 2: Planning
Create an execution plan based on task characteristics.
Actions:
1. Determine execution mode (sequential or parallel)
2. Check for task dependencies (A must complete before B)
3. Assign tasks to workers in priority order
4. Generate worker instructions
Output: Execution plan with task assignments
Checkpoint: TASKS_DISCOVERED
- Display: Ready tasks and proposed execution plan
- Auto-continue: If –autonomous flag and tasks found
- Options:
continue,reorder,exclude [TASK-ID],stop
Phase 3: Execution
Spawn and monitor worker agents.
For SEQUENTIAL mode:
for each task in queue:
1. Spawn worker with Task tool
2. Worker runs agile-workflow for the task
3. Monitor progress via file system
4. When complete: proceed to merge phase
5. On failure: handle error, decide continue/stop
For PARALLEL mode:
1. Spawn workers up to max_workers
2. Monitor all workers concurrently
3. As workers complete: queue their branches for merge
4. Spawn next worker if tasks remain
5. Continue until all tasks processed
Checkpoint: WORKER_COMPLETE (per worker)
- Display: Worker summary, branch name, next action
- Auto-continue: If successful and –autonomous
- Options:
continue,retry,skip,stop
Phase 4: Merging
Execute merges sequentially to avoid conflicts.
Actions:
1. For each completed task in merge queue:
a. git checkout main && git pull
b. Merge branch (git merge --squash)
c. Verify merge succeeded
d. Delete feature branch
2. If conflict: pause and alert user
Output: All branches merged to main
Phase 5: Verification
Verify system integrity after all merges.
Actions:
1. git checkout main && git pull --rebase
2. npm run build (verify build passes)
3. npm test (run full test suite)
4. Check for regressions
5. Generate verification report
Output: Verification status (PASSED/FAILED)
Checkpoint: VERIFIED
- Display: Test results, build status
- Auto-continue: If all tests pass
- Options:
done,investigate,revert
Phase 6: Persist Progress
Update source-of-truth documentation to reflect completed work.
Actions:
1. For each completed task:
a. Update task status in the backlog epic file (ready â complete)
b. Recalculate epic-level progress (e.g., "22/28 complete" â "24/28 complete")
c. Unblock dependent tasks (blocked â ready) if blockers are now satisfied
2. Update project status file (context/status.md):
a. Current project phase
b. Epic progress table
c. Recently completed work
d. Active/upcoming work summary
3. Commit and push documentation updates
Output: Backlog and project status files reflect actual progress
Why this phase exists: Internal tracking (.coordinator/state.json, worker progress files) is session-scoped and ephemeral. The backlog epic files and project status are the persistent source of truth that humans and future sessions rely on. Without this phase, completed tasks remain marked “ready” in backlog files â in one real-world case, 22 merged tasks were never updated in the backlog.
Phase 7: Summary
Generate comprehensive completion report.
Output:
- Tasks completed with PR numbers and commits
- Metrics (workers spawned, PRs merged, tests added)
- Verification status
- Documentation updates applied
- Remaining backlog tasks
Worker Protocol
Workers are spawned using Claude Code’s Task tool and run agile-workflow for their assigned task.
Worker Instruction Template
See templates/worker-instruction.md
Key requirements for workers:
- Run
agile-workflowwith autonomous mode - Write progress to
.coordinator/workers/{worker-id}/progress.json - Do NOT self-merge – signal ready-to-merge status instead
- Handle all agile-workflow checkpoints automatically
Progress Tracking
Workers report progress via file system:
// .coordinator/workers/worker-1/progress.json
{
"worker_id": "worker-1",
"task_id": "TASK-006",
"status": "in_progress|completed|failed|ready-to-merge",
"phase": "implement|review|merge-prep|merge-complete",
"commit": null,
"branch": "task/TASK-006-description",
"last_update": "2026-01-20T10:15:00Z",
"milestones": [
{"phase": "implement", "timestamp": "..."},
{"phase": "review", "timestamp": "..."}
]
}
State Tracking
The coordinator maintains state in .coordinator/state.json:
{
"session_id": "coord-2026-01-20-abc123",
"state": "EXECUTING",
"config": {
"execution_mode": "sequential",
"autonomy_level": "autonomous"
},
"tasks": {
"queued": ["TASK-008"],
"in_progress": ["TASK-007"],
"completed": ["TASK-006"],
"failed": []
},
"workers": [...],
"merge_queue": [],
"verification": null
}
See references/state-tracking.md for details.
Failure Handling
| Failure Type | Detection | Recovery |
|---|---|---|
| Worker fails | Progress shows failed |
Offer retry, skip, or abort |
| Tests fail | npm test exits non-zero | Worker retries up to 2x |
| Merge conflict | git merge fails | Pause, alert user |
| Verification fails | Post-merge tests fail | Offer investigate or revert |
See references/failure-handling.md for details.
Integration with agile-workflow
The coordinator spawns workers that execute agile-workflow:
agile-coordinator (orchestrator)
â
âââ Worker 1 â agile-workflow --task TASK-006
âââ Worker 2 â agile-workflow --task TASK-007
âââ Worker 3 â agile-workflow --task TASK-008
Key integration points:
- Workers run agile-workflow in autonomous mode
- Coordinator handles merges (workers create PRs but don’t merge)
- Progress synced via file system
- Coordinator runs verification after all merges
Anti-Patterns
AP1: Parallel Merges
Pattern: Merging multiple PRs simultaneously Problem: Creates merge conflicts, corrupts main branch Fix: Always merge sequentially, pull before each merge
AP2: Worker Self-Merge
Pattern: Workers merge their own branches Problem: Race conditions, uncoordinated merges Fix: Workers signal ready-to-merge, coordinator executes merges
AP3: No Verification
Pattern: Skipping post-merge verification Problem: Regressions go undetected Fix: Always run full test suite after merges complete
AP4: Skipping Documentation Updates
Pattern: Relying on internal tracking (.coordinator/state.json, worker progress files) without persisting to backlog epic files and project status
Problem: Completed tasks remain marked “ready” in the backlog; project status stays stale; future sessions see incorrect state and may attempt duplicate work
Fix: Always run Phase 6 (Persist Progress) to update backlog epic files, unblock dependent tasks, and update project status before generating the summary report
Example Interaction
User: /agile-coordinator
Coordinator: Discovering ready tasks...
Found 2 ready tasks:
1. TASK-006 - Persistent Message Status (medium priority, medium size)
2. TASK-007 - Add Unit Tests (high priority, large size)
Execution Plan:
- Mode: sequential
- Order: TASK-006 â TASK-007
Proceed? [continue/reorder/stop]
User: continue
Coordinator: Starting TASK-006...
[Spawns Worker 1]
**Milestone**: TASK-006 - Implementation complete, tests passing
Coordinator: Merging task/TASK-006-persistence...
**Milestone**: TASK-006 complete (commit abc123)
Coordinator: Starting TASK-007...
[Spawns Worker 2]
**Milestone**: TASK-007 - Implementation complete, tests passing
Coordinator: Merging task/TASK-007-tests...
**Milestone**: TASK-007 complete (commit def456)
Coordinator: Running verification...
- Build: PASSED
- Tests: 47/47 passing
- Coverage: 82%
## Summary
Tasks completed: 2
- TASK-006: merged (commit abc123)
- TASK-007: merged (commit def456)
Verification: PASSED
Integration Graph
Inbound (From Other Skills)
| Source Skill | Trigger | Action |
|---|---|---|
| requirements-elaboration | Tasks groomed | Coordinator can execute |
| backlog manager | Backlog ready | Coordinator discovers tasks |
Outbound (To Other Skills)
| This Action | Triggers Skill | For |
|---|---|---|
| Spawn worker | agile-workflow | Task implementation |
| Verification fails | research | Debug investigation |
Complementary Skills
| Skill | Relationship |
|---|---|
| agile-workflow | Workers execute this skill |
| context-network | Manages backlog this reads |
Design Constraints
- Requires Claude Code’s Task tool for spawning workers
- Workers cannot directly communicate with each other
- File system used for progress coordination
- Sequential merges only (parallel merges disabled)
- Assumes context network backlog structure
What You Do NOT Do
- Implement tasks directly (workers do this)
- Merge branches in parallel (always sequential)
- Skip verification (always verify after merges)
- Skip documentation updates (always persist progress to backlog epic files and project status)
- Rely solely on internal tracking files as source of truth (
.coordinator/state.jsonis ephemeral) - Continue after critical failures without user consent