swarm-skill

Swarm — Task Decomposition & Parallel Execution

Environment Check

Run scripts/check_env.sh first. Handle by exit code:

• Exit 0: Ready. Proceed to skill discovery.
• Exit 1: Version too low. Tell user to run claude update. Stop.
• Exit 2: Agent Teams not enabled. Read ~/.claude/settings.json, add "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" to the "env" object (create if absent, preserve existing config), write back. Tell user to restart Claude Code. Stop.

Skill Discovery

Build the available skill inventory from two sources:

1. Session skills: Parse the <system-reminder> in current context for listed skills (name + description).
2. Filesystem commands: Run scripts/discover_skills.sh to scan ~/.claude/commands/ and .claude/commands/.

Merge into a flat list. Only use skills that actually appear — never assume.

Teammates auto-load the same skills as the lead. No extra config needed.

Workflow

Check env → Discover skills → Decompose → Create team + tasks → Spawn teammates → Monitor → Deliver

1. Decompose

Analyze the user's request. Split into subtasks that can run in parallel or pipeline.

Split when: parallelizable, different expertise, or pipelineable. Don't split: tightly coupled edits, single-file changes, heavy shared context.

See references/patterns.md for decomposition patterns and role-to-skill matching.

2. Create Team & Tasks

Create an Agent Team, then create Tasks. Each Task must be self-contained: clear goal, file paths, completion criteria, project conventions (from CLAUDE.md).

Use addBlockedBy for dependencies — blocked teammates wait automatically.

3. Spawn Teammates

1. Spawn independent teammates in parallel (single message, multiple Agent calls)
2. Dependent teammates also spawn immediately — they self-wait
3. All use run_in_background: true

Build prompts using the template in references/patterns.md. Only inject actually-discovered skills.

Agent type: default for code, "Explore" for research, "Plan" for design.

4. Monitor & Coordinate

• Verify output on completion; SendMessage corrections if needed
• Resolve conflicts if two teammates edit the same file
• On failure: analyze cause, guide retry via SendMessage

5. Commit (optional)

Create a committer teammate addBlockedBy all dev/test tasks. Inject commit skill if discovered; otherwise use git directly. See references/patterns.md for prompt templates.

6. Deliver

1. Run tests to verify changes
2. Summarize each teammate's output to the user
3. Shut down teammates and clean up the team