mk:task-queue

What This Skill Does

Manages task assignment and file ownership during parallel agent execution. Tracks tasks in session-state/task-queue.json with a claiming protocol that prevents file conflicts between agents.

When to Use

• Parallel execution phases where multiple agents work simultaneously on different files
• Automatically invoked by mk:team-config when setting up parallel agent teams
• NOT for: single-agent tasks, sequential workflows, tasks without file ownership boundaries

Core Capabilities

• Task lifecycle: pending → claimed → in_progress → completed (or blocked)
• Claiming protocol: lowest-ID task where status=pending AND all blocked_by tasks are completed
• Ownership enforcement: each task declares file globs; agents can only write to owned files; overlapping claims are rejected
• Race condition prevention: orchestrator is the sole claim serializer — agents request through it, never self-claim
• Audit trail: completed tasks remain in queue until the parallel phase ends

Workflow

1. Orchestrator creates session-state/task-queue.json with tasks and ownership globs
2. Agent requests next available task from orchestrator
3. Orchestrator checks for ownership conflicts with other in_progress tasks
4. If no conflict: assigns task, sets status=in_progress, records owner
5. Agent completes task, sets status=completed
6. After all tasks complete, orchestrator runs integration test, then cleans up

Usage

# Invoked automatically by mk:team-config during parallel setup
# Not typically called directly

Example Prompt

Set up parallel task queue: 3 tasks — API endpoints (src/api/*), UI components (src/components/*), tests (tests/*). No overlapping ownership.

Common Use Cases

• Parallel agent execution during harness builds
• Multi-file feature implementation with clear file boundaries
• Coordinated refactoring across independent modules

Pro Tips

• Ownership globs must not overlap. src/api/* and src/api/auth/* overlap — the more specific glob must be in the same task.
• Queue is ephemeral. Lives in session-state/, cleared on new session. For persistent task tracking, use the plan's task system instead.
• Agents never self-claim. Always request through the orchestrator to prevent race conditions on parallel reads of task-queue.json.

Canonical source: .claude/skills/task-queue/SKILL.md