[P1] Recall-policy prompt instructs the model to call the unregistered tool lcm_expand_query

[100yenadmin]

Severity: P1 · Confidence: 95% · Found by: hermes-lcm architecture review (slice S6).

The bug

recall_policy.py builds LOSSLESS_RECALL_POLICY_PROMPT, injected into the context every turn, which instructs the model to call lcm_expand_query. That tool was deferred (ADR-012) and is not registered in tools/__init__.py's TOOL_SCHEMAS. describe.py / grep.py schema prose may also reference it. The model is told every turn to use a tool it cannot see.

Fix

Remove or conditionally-gate every lcm_expand_query reference in model-facing prose so v0.1.1 never advertises an unregistered tool. ADR-016's "byte-verbatim prose" rule yields here to "don't reference unregistered tools" — add an # ADR-012 provenance comment at each edit site. Add a test asserting every tool named in the recall-policy prose is in TOOL_SCHEMAS.

Target: v0.1.1 patch

[100yenadmin]

Scoping note (v0.1.1 / PR #138): the per-turn directive in recall_policy.py — the harmful vector, since it injects every turn — is fixed in v0.1.1. The lcm_expand_query mentions that remain in the grep.py/describe.py/expand.py tool-schema descriptions were deliberately left: they are under a byte-identical verbatim lint (SHA-256 snapshots, ADR-016) and are 'see also' prose inside already-registered tools, not a per-turn directive. Those resolve when the v0.2.0 lcm_expand_query decision lands (register it in-process per the S6 review, or formally drop it + re-snapshot the fixtures). Tracked.