Context
During review of PR #130, the reviewer began substantive diff review before checking whether the newest signed author-side artifact was from the same model/version. The current review protocol already says to inspect signed provenance first, skip same-model PRs, and only proceed with an explicit user-confirmed self-review as a non-gate-satisfying audit artifact.
That means the immediate failure was operator error, not missing policy. But it also exposed a product gap: the policy lives in closure-and-review.md, while common entry points such as direct "review PR #N" requests and lightweight review flows do not force a provenance preflight before substantive review begins. Issue #132 addresses the hunt recommendation side of this problem; this issue covers the remaining execution-path gap.
Design Summary
- Treat provenance and reviewability resolution as a mandatory preflight before any substantive PR review work
- Keep the existing cross-model gate and self-review audit rules; this issue is about operationalizing them, not changing the policy
- Make the preflight visible in the review workflow so agents do not have to remember a buried paragraph in a long reference file
- Ensure the fallback for same-model or unknown-reviewability cases leads directly to the correct next action: stop, ask for confirmation for audit-only self-review, or emit a review contract
Approach
Add an explicit review-entry section to the shiplog review workflow that runs before diff inspection. That section should require the reviewer to resolve reviewability from signed artifacts, classify the PR as cross-model / same-model / unknown, and branch immediately:
cross-model: proceed to substantive review
same-model: stop and ask whether to record an audit-only self-review
unknown: stop and clarify provenance or generate a review contract
This should be documented in the main review protocol and any review-facing command or template that currently jumps straight from "user asked for review" to "inspect the diff."
Alternatives Considered
- Leave the policy as-is and rely on reviewer discipline: rejected because the existing wording was already correct, yet a same-model substantive review still happened.
- Fold everything into
hunt only: rejected because direct review requests can bypass hunt, so fixing recommendation logic alone does not close the execution gap.
- Ban audit-only self-review entirely: rejected because the current protocol intentionally allows it as a visible non-gate-satisfying artifact when explicitly confirmed.
Tasks
Open Questions
- Should shiplog add a dedicated review command or short review skill whose entire first phase is provenance preflight, instead of relying on a long shared reference doc?
- Should a same-model audit-only self-review be prohibited from including substantive findings unless the user explicitly requests the audit artifact?
- Should blocked review attempts be captured as a signed timeline artifact on the PR, or only surfaced to the user unless they choose audit-only self-review?
Related
Authored-by: openai/gpt-5.4 (codex, effort: high)
Captain's log entry created by shiplog
Context
During review of PR #130, the reviewer began substantive diff review before checking whether the newest signed author-side artifact was from the same model/version. The current review protocol already says to inspect signed provenance first, skip same-model PRs, and only proceed with an explicit user-confirmed self-review as a non-gate-satisfying audit artifact.
That means the immediate failure was operator error, not missing policy. But it also exposed a product gap: the policy lives in
closure-and-review.md, while common entry points such as direct "review PR #N" requests and lightweight review flows do not force a provenance preflight before substantive review begins. Issue #132 addresses thehuntrecommendation side of this problem; this issue covers the remaining execution-path gap.Design Summary
Approach
Add an explicit review-entry section to the shiplog review workflow that runs before diff inspection. That section should require the reviewer to resolve reviewability from signed artifacts, classify the PR as cross-model / same-model / unknown, and branch immediately:
cross-model: proceed to substantive reviewsame-model: stop and ask whether to record an audit-only self-reviewunknown: stop and clarify provenance or generate a review contractThis should be documented in the main review protocol and any review-facing command or template that currently jumps straight from "user asked for review" to "inspect the diff."
Alternatives Considered
huntonly: rejected because direct review requests can bypasshunt, so fixing recommendation logic alone does not close the execution gap.Tasks
T1: Add a mandatory review-entry preflight to the protocol
[tier-2]skills/shiplog/references/closure-and-review.md(modify)narrowT2: Wire the preflight into review-facing entry points
[tier-2]narrowT3: Define the blocked-path outputs for same-model and unknown cases
[tier-1]skills/shiplog/references/closure-and-review.md(modify), any linked review workflow references needed for consistencyfullsame-modelandunknowncases without leaving the reviewer in an ambiguous half-reviewed stateOpen Questions
Related
huntrecommendation gap for reviewability-aware triageAuthored-by: openai/gpt-5.4 (codex, effort: high)
Captain's log entry created by shiplog