Swarm Coordinator

Orchestrate multiple Claude Code sessions or agents working on the same project through a shared task list.

Overview

The swarm coordinator uses CLAUDE_CODE_TASK_LIST_ID to enable multiple Claude Code instances to share progress state. Each instance can:

• •See all tasks via TaskList
• •Claim unblocked tasks
• •Update task status
• •Coordinate without conflicts

When to Use

• •Large implementation with many independent tasks
• •Want true parallelism (multiple terminal windows)
• •Tasks have clear file ownership (no conflicts)
• •Implementation plan includes Parallel: yes tasks

Setup Process

Step 1: Generate Shared Task List ID

TASK_LIST_ID="swarm-$(date +%Y%m%d-%H%M%S)"
echo "export CLAUDE_CODE_TASK_LIST_ID=$TASK_LIST_ID"

Save this ID - all workers need it.

Step 2: Create Task DAG from Plan

If IMPLEMENTATION_PLAN.md exists:

1. •Read plan and parse tasks manually (agent maintains ID mapping)
2. •Create all tasks with TaskCreate, recording plan ID → system ID mapping
3. •Set up dependencies with TaskUpdate using mapped system IDs

Output the task graph for user review.

Step 3: Launch Worker Sessions

Guide user to open additional terminal windows:

# Terminal 2 (worker)
export CLAUDE_CODE_TASK_LIST_ID=[same-id]
cd [project]
claude

# Terminal 3 (worker)
export CLAUDE_CODE_TASK_LIST_ID=[same-id]
cd [project]
claude

Step 4: Worker Protocol

Each worker follows this loop:

1. TaskList to see available work
2. Find unblocked pending task with no owner
3. TaskUpdate to claim (set owner to session ID)
4. TaskUpdate status to in_progress
5. Implement the task
6. Run tests
7. TaskUpdate status to completed
8. Repeat

Step 5: Orchestrator Monitoring

The coordinator (this session) monitors progress:

- TaskList every few minutes to see global state
- Identify stuck tasks (in_progress too long)
- Handle blockers and conflicts
- Run final integration tests when all tasks complete

Task Claiming Protocol

To prevent race conditions:

1. •Worker calls TaskUpdate with owner: "session-<id>" for a pending task
2. •If another worker claimed it first, the task will have a different owner
3. •Worker checks TaskGet to verify ownership before proceeding
4. •Only the owner can transition to in_progress and completed

Conflict Resolution

If two workers modify the same file:

1. •First to complete wins
2. •Second worker must rebase/resolve conflicts
3. •If conflict is complex, coordinator arbitrates

Completion

When all tasks are completed:

1. •Coordinator runs full quality gate suite
2. •Integration tests across all changes
3. •Final code review via /requesting-code-review
4. •Merge/squash if using worktrees

Example Session

Coordinator: "Starting swarm for auth-system implementation"

TaskList shows:
#1 [pending] Create user model - ready
#2 [pending] Create auth middleware - ready
#3 [pending] Create login endpoint - blocked by #1, #2
#4 [pending] Create register endpoint - blocked by #1
#5 [pending] Add password hashing - blocked by #1

Coordinator assigns:
- Worker 1: Task #1 (user model)
- Worker 2: Task #2 (auth middleware)

Workers complete, TaskList now shows:
#1 [completed] - by worker-1
#2 [completed] - by worker-2
#3 [pending] - now unblocked!
#4 [pending] - now unblocked!
#5 [pending] - now unblocked!

Coordinator assigns next batch...

Limitations

• •Works best with clear task boundaries
• •File conflicts require manual resolution
• •All workers must have same env setup
• •Shared task list requires same CLAUDE_CODE_TASK_LIST_ID