| name | ultraswarm |
| description | Orchestrate complex coding work through the ultraswarm v3 standalone runner. |
Ultraswarm v3 for small-harness
Use the repository's standalone runner as the only orchestration implementation. Do not recreate orchestration in the host, and do not implement delegated feature code directly.
Invocation: node ${ULTRASWARM_HOME:-$HOME/projects/ultraswarm}/bin/ultraswarm.mjs
Runner
Resolve the checkout from ULTRASWARM_HOME, this skill's real path, or ~/projects/ultraswarm. It must contain bin/ultraswarm.mjs and installed dependencies. Node >=22 is required.
Workflow
- Functionally verify the worker pool:
node <root>/bin/ultraswarm.mjs preflight. This runs a cached smoke test (not just --version) and prints a roster marking each CLI INSTALLED and FUNCTIONAL. Drop any worker shown UNUSABLE — do not pin a task to it — and confirm at least the policy minimum of functional workers remain.
- Inspect the target repository and prepare a validated JSON plan. Scope each task's
contract.allowed_paths to include coupled test files, not just the obvious source — flipping a value often has a test blast radius.
- Preview with
node <root>/bin/ultraswarm.mjs run --plan-file .ultraswarm-plan.json. Show tasks, waves, routing explanations, policy findings, and gates. Obtain explicit plan approval.
- Execute with
node <root>/bin/ultraswarm.mjs run --plan-file .ultraswarm-plan.json --approve-plan, keeping the runner's stderr visible and relaying its live progress to the user as it streams: wave headers, per-agent dispatch lines, gate results, and the periodic active/idle heartbeat that shows every worker's state.
- Report the run ID and present the runner's final human-readable summary — per-task outcome, which worker landed each task, and the tokens-saved estimate (the implementation work external CLIs did off your context). Integrated work remains off the checked-out branch.
