consolidation-review

Consolidation Review

Review code integration, not the research conclusion. The approved reflection is authoritative and cannot be reopened from this gate.

Start the pinned review

Call agent.hello once first — this review is its own context window — and pass the returned agent_id in every Merv call that follows.

Use the assigned review_request_id with review.start. In an assigned agent session, pass reviewer_capability="assigned" and caller_session_id="assigned". Otherwise use the exact handoff values and a reviewer identity distinct from the producer.

Work from the returned proposal snapshot and consolidation.get. Do not edit, commit, submit artifacts, change research state, or advance Git. Running read-only checks and tests in the detached proposal checkout is allowed.

Check the proposal

Verify:

• The proposal SHA is the commit checked out and descends from, or equals, the declared central base.
• The proposal implements the approved reflection and change spec without silently changing their scientific decisions.
• Every experiment in the reflection corpus has exactly one decision with a concrete rationale.
• used_as_is and adapted decisions identify a real integration mechanism; reviewed_not_used and superseded carry no code, and superseded decisions name the covering experiment.
• The stated integration kind matches the diff and branch history. The runner will verify ancestry independently; do not call a rewrite or cherry-pick a merge.
• Important experiment changes were not omitted accidentally, and omitted work has an evidence-based reason.
• The consolidation is coherent as one central change: no duplicate implementations, unresolved conflicts, stray generated files, credentials, or experiment-only scaffolding.
• The submitted validation is credible. Rerun the smallest checks needed to verify risky or load-bearing changes. Separate observed checks from claims you could not verify, and inspect failure paths and unintended changes.