Handbook
Local LLM worker (system)
Feature FW-WC-001. Status: published.
Updated
Intent
local_llm_worker loads a micro-pack, checks required context keys, issues one governed LCDL chat, and writes parseable findings plus a run log.
Behavior
| Item | Value |
|---|---|
| Runtime | forge-workcells CLI |
| Entry | forge-workcells run --workcell local_llm_worker --pack … --context … --out-dir … |
| Pack home | Kitchen Sink tools/forge-micro-agent/packs/ (UX pack ux-ai-rule-v1) |
| Inference | forge-lcdl profile from --llm-env |
| Outputs | micro-agent.log and model text for parse-ai-agent-findings.mjs |
| Gate | Advisory. The harness decides. |
Request shape is pack path plus context.json, not a full domain launch_pack. Envelopes still sit on forge.workcell_request.v1 / forge.workcell_result.v1 when a ForgeRun records the session.
Limits
No DecisionPack graph, no RAG, no MCP, no multi-rule campaign. A separate worker, pm_bet_cluster_llm_worker (FW-WC-006), emits product-strategy JSON instead of UX findings.