Handbook
Local LLM worker
Feature FW-WC-001. Workcell local_llm_worker.
Updated
Who this is for
Operators and harness authors who want one bounded local-model judgment, such as a single UX-audit rule, without opening a Cursor session.
Problem
A full coding agent is the wrong tool for one classified finding. This worker runs one prompt pack over an assembled context and returns advisory findings the harness can parse.
What you can do
- Run one pack against one context and read the findings.
- Compare those findings with the same parse gate the Cursor path uses.
- Treat the output as advice. The harness, not the model, decides what passes.
What you cannot do
- Approve a docs promotion, merge a pull request, or edit a product from this worker.
- Load a whole ruleset, search a vector index, or call MCP tools on this path.
- Replace day-to-day Cursor work with this worker.
Limits
One invocation, one pack, one primary model call. Context is capped by the pack budget (default 32K tokens). Findings stay advisory.