D318 — The tools an agent is told to call are listed, and their answers parse

Decision. concept_batch, asset_delete and artifact_delete (registered only under allow_artifact_write) leave advancedToolNames: an agent-profile client sees them in tools/list. The responses an agent reads are made regular: lint on a clean scope returns the JSON shape of a dirty one; kb_repair with dry_run: false drops the planned array; kb_review reports limit_applied and, when it clamped the request, limit_capped and limit_max; gate_check accepts an empty changed_ids as the whole-KB gate (no commit gate); asset_delete without if_match says so instead of "stale write". A patch operation of concept_batch counts the edits it carries toward the 512 KiB aggregate, not the body that results. The gap between kb_status.conformance and lint counts is documented, not closed.

Why. The bundled kb-doctor skill prescribes concept_batch, and a descriptor-bound client cannot call by name what tools/list omits (D123): the skill and the profile contradicted each other. The other changes remove a parse failure (a bare string where JSON is expected), a silent truncation, a refusal with no use case, and a misleading error. The batch limit bounds the request line, so counting the post-patch body charged an agent for bytes it never sent.

Alternatives rejected. - Keep concept_batch unlisted and fix the skill: the skill is right, a multi-page refactor is a normal doctor session. - Make conformance.findings equal to the lint total: the subset is the structural debt the doctor can act on (D290); only the explanation was missing. - Server-side unescaping of old_string: the matching is a plain substring and correct; the fragility is the client's JSON escaping. - Hold the D285 budget at 23 KiB: after cutting concept_batch from 3 KiB to 1.4 KiB (the field detail moved to docs/control-plane.md) the catalogue was 1.8 KiB over, and the other descriptions were at their floor. The budget is raised to 25 KiB; artifact_delete (~0.3 KiB) is outside it, since it only appears with the per-KB flag.

Consequences. Callers that parsed lint's "Lint OK" string, or read planned after an apply, break; both were brittle or redundant. Every further tool in the agent profile is paid for against 25 KiB. kb_review's lint_judgement items are lint findings under another name: their totals must not be added.