076-Source-claim grounding is a promoted skill

Type: ../types/adr.md · Status: accepted

Status: accepted Date: 2026-08-25

Context

ADR 073 made one procedure responsible for deciding what source evidence a claim needs: read the ingest's ## Quotes, then either declare them sufficient, retain the minimum verbatim excerpts through cp-skill-ingest's quote_append_request, or declare snapshot required. That procedure was a plain instruction addressed by file path.

Its only inbound routes were the stop-and-report handoff inside cp-skill-write and cp-skill-write-multistage, ADR 073 itself, and the rollout workshop. A request that does not start inside a writer — an operator asking to retain a quote that supports a discussion — had no route at all. Because instructions are addressed by path, the writers' handoff also had to carry two literal routes, one for the source checkout and one for an installed project.

The unrouted request did not stay unrouted; it matched the wrong surface. cp-skill-ingest advertised "mechanically append verified verbatim quotes", so an agent holding that request matched the ingest skill, invoked it with a natural-language target, and the skill selected quotes anchored to no claim. That happened on 2026-08-25 and forced the choice: one request had two advertised surfaces, and the one it reached did not own the decision.

Two standing claims decide where the procedure belongs. Capability placement should follow autonomy readiness: a bounded procedure that has already run repeatedly and autonomously — the rollout cohorts — is promotion-ready. Skills are instructions plus routing and execution policy: promotion changes discovery and invocation, not the procedure itself.

Decision

Grounding is the promoted skill cp-skill-ground. It lives at kb/instructions/cp-skill-ground/SKILL.md, is installed by commonplace-init, and is user-invocable with arguments Target and Claim needed. Its body and result protocol — quotes sufficient, quotes added, snapshot required — are unchanged from the instruction it replaces.

cp-skill-ground is the single advertised entry for deciding which quotes an ingest retains, and the only procedure that writes an ingest's ## Quotes section. cp-skill-ingest ingests a URL or local snapshot into a new .ingest.md and executes bounded re_ingest_requests; it has no quote-append path at all. One request, one matching surface.

Selection and verification stay separate without a second skill, because the independent check is code. commonplace-validate resolves every Source extract (verbatim) against the name-paired snapshot and fails the ingest on any extract that does not occur there, so the agent that chose a passage never certifies its own transcription.

Writers still stop when Quotes are insufficient. Their handoff becomes a single route addressed by skill name — invoke cp-skill-ground with Target and Claim needed — instead of a path branched on install layout.

Considered alternatives

Keep the instruction and add routing pointers. An AGENTS.md pointer, or a broader cp-skill-ingest description, would make the procedure findable without promoting it. Rejected: it leaves two advertised surfaces for one request, which is the failure that motivated the decision. The capability-placement rule puts an autonomy-ready procedure on the skill surface rather than in a control-plane inventory.

Promote it and let writers invoke it automatically. This is the candidate direction of the proposal Source-claim grounding skill and evidence-retention boundary. Deferred, not rejected: a request to write one artifact would then mutate a second, and that mutation authority is a separate decision. Promotion for explicit use carries no new mutation footprint, so it need not wait on that question.

Keep the append as a mechanical path inside cp-skill-ingest. This is the ADR 073-era shape: grounding selects the passages and hands a quote_append_request to the ingest skill, which performs the mutation. Rejected: it makes one skill two things — create a source record, and mutate an existing one — and keeps a competing description trigger alive for exactly the request this decision routes. Its claimed benefit, a mutation owner independent of the selector, is already supplied by the validator's deterministic quote resolution.

Move retained evidence out of the ingest first. The same proposal leaves the evidence-retention boundary open. Left open here: the skill's result protocol does not change if the append target later moves, so promoting now forecloses nothing.

Consequences

Operativity path: grounding and quote-retention requests reach cp-skill-ground through harness skill discovery, with no competing match on cp-skill-ingest; the force is routing, not validation. Writers report one route, and installed projects need no path-branched handoff.

cp-skill-ingest has one job and carries no structured mutation request other than re-ingest. A bad quote append is now diagnosed by commonplace-validate's source quote check against the pinned snapshot rather than by a second agent's recheck of the first agent's selection.

The proposal retains only its undecided parts — automatic invocation from writers, and the evidence-retention boundary.

Caveat on the expected benefit: on this harness a context: fork skill inherits the invoking conversation, so promotion does not by itself isolate source reading from target prose. Any later claim that grounding runs in a fresh context must name the launch mode that actually provides it.

Relevant Notes: