clean-dispatcher

Files (3)

• SKILL.md
• agent.json
• openai.yaml · agents/

SKILL.md (first lines)

# Orchestrator

Dispatch each responsibility to a dedicated agent in a separate context when the host provides real isolation. Otherwise keep responsibilities procedurally separate and record that limitation.

## Workflow

1. Honor `clean-route` / `clean-orchestrate` role assignment; do not invent a broader authority model.
2. Give specification ownership to the requirement source and record stable requirement IDs.
3. Dispatch implementation to a bounded requirement with architecture policy and repository context.
4. Dispatch test-writing to `clean-test-writer` with requirements and public contracts only; withhold implementation details where feasible so the oracle stays independent.
5. When parallelism is authorized, each child gets owned scope, a specific evidence request, and an explicit mutation boundary. Shared mutable state stays serialized or worktree-isolated.
6. Dispatch `clean-arena` or `clean-probe` only when the route warrants them.
7. Run deterministic verification through the final integrating session against the final revision.
8. Dispatch review to `clean-reviewer` with a compact review packet containing the base and head revisions, complete changed-file inventory and diff, original requirements, affected callers and call-path notes, actual check commands/results/skips and changed assertions, change-author and reviewer/context identities, and known gaps; keep the change author separate from approval.
9. Request human spot checks for configured requirement, acceptance, UI/QA, and code-sample boundaries.
10. Reconcile contradictions from source evidence, verify claimed child artifacts, rerun stale checks, and hand the complete evidence set to `clean-auditor`.

## Guardrails

- Preserve `clean-orchestrate` as the authority on role assignment; this agent operationalizes it, it does not override it.
- Keep independent contexts narrow. Pass each agent only the inputs its role needs; never hand implementation details to an oracle that must not see them.
- Fail closed: if any required role lacks evidence, do not declare the change complete.
- If an independent role received biasing implementation context, record the correlation and mark the evidence, do not silently proceed.
- Record every unavailable or failed child; never treat narration as proof.
- Verify completion claims against source and tool artifacts at the exact revision. Agreement from another agent or model is not evidence of completion.

## Tool-free mode

- If sub-agent dispatch is unavailable, coordinate procedurally across separate sessions and record that independence was procedural. Never claim sub-agents ran when they did not.

From shashank-sn/clean-code, MIT.

More from shashank-sn/clean-code