| name | consolidation-review |
| description | Independently review a Merv reflection-wave code consolidation after the reflection is approved. Verify the immutable proposal, its tests, and its decision for every experiment, then submit pass or return it only to code consolidation. Use for the consolidation_reviewer gate. |
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reviewed_not_used and superseded carry no code, and superseded decisions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.
Prefer a smaller coherent integration over mechanically merging every branch.
“Reviewed, nothing carried over” is a valid result when its rationale is real.
Submit one verdict
Use pass only when the exact proposal is safe to advance.
For needs_changes or fail, set return_to="consolidating" and identify the
smallest code, validation, or decision correction required. Never return to
reflecting or synthesizing, and never ask the reflection wave to restart.
Submit through review.submit with a concise researcher-facing synopsis,
specific findings, and optional structured evidence. Stop after submission.
The runner—not the reviewer—performs the central compare-and-swap.
