delegate-local-worker

Delegate to the local worker

Keep the primary model responsible for architecture, decomposition, scope, security judgment, patch review, and final acceptance. Offload mechanical, bounded implementation to local-llm-worker.

Decide before reading broadly

Inspect only enough repository structure, manifests, interfaces, or failing output to define the boundary. Do not load every implementation file into the primary context before delegating.

Delegate when all are true:

• The requested outcome and acceptance conditions are concrete.
• The change fits a narrow list of files or path globs.
• The returned diff can be reviewed completely.
• The project has local-llm-worker.toml.
• The propose_patch and apply_patch MCP tools are available.

Keep work in the primary model when it determines architecture, changes trust or authentication boundaries, performs an ambiguous migration, spans many subsystems, handles secrets, or cannot be accepted from a bounded diff.

Delegate

1. State the intended behavior and acceptance checks compactly.
2. Select the narrowest useful allowed_paths. Prefer exact files; use a small directory glob only when the target file is not yet known.
3. Call propose_patch with the absolute current repository path as workspace_root before independently implementing the change. Do not copy source files into supporting_context; the worker reads allowed files itself.
4. Review every returned line, touched path, warning, and statistic. Reject unrelated behavior, unexplained changes, or an unexpectedly large proposal.
5. If acceptable, call apply_patch with the returned proposal_id and exact sha256. Never reconstruct or substitute the digest.
6. Review validation results and inspect the resulting workspace diff. The primary model remains accountable for correctness.
7. On validation failure, either fix directly or request the one permitted repair using the failed proposal ID and identical path scope. Review the repair as a new proposal.