Review Fix Loop

The parent coordinates rounds. Orchestration owns model defaults, delegation, recovery, and the general handoff. Review assesses one snapshot; Review Sweep validates and fixes findings. The caller owns commits, publication, replies, and merging.

Prepare

Resolve the original base once using Review. Gather the complete snapshot, requirements, and verification evidence with the canonical review brief.

Choose coverage in this order, honoring explicit user requests:

For behavior or agent-procedure changes, add a separate Adversarial assessment when the combined change is substantial: it changes a shared contract or ownership boundary, introduces interacting behaviors across modules, redesigns a multi-step workflow, or alters security, data integrity, concurrency, retries, partial failure, migrations, or recovery. Broad implementation or instruction rewrites also qualify. Judge the changed behavior and consequences, rather than line count or file count alone. Record the decision and a one-sentence reason in the brief. Keep required coverage through fix rounds against the original base; expand it if fixes introduce a new trigger. When a plausible material consequence remains uncertain, include Adversarial.

Apply review assignments for the criteria, model mapping, additional draws, provider references, and skip policy. Finish preparation when scope, proof, and assignments are recorded.

Review, settle, repeat

1. Run each selected assessment as a T3-owned task through Orchestration, using the shared review brief. Await the required assessments and all additional reviews that have not been skipped.
2. Recheck the snapshot before accepting results. If it moved, cancel obsolete tasks, retain useful observations as unverified claims, and return to step 1 with a refreshed snapshot against the original base. End this step there; superseded assessments do not satisfy coverage. For a matching snapshot, check the reports' assigned-criteria coverage, including unread inputs or missing assessments. Use Review evidence to validate task/output identity, read receipts, and source citations, then archive the original report and receipt bytes before later writers run. Invalid or lost required coverage needs a fresh assessment under the selected failure policy. Give every result and partial finding to Review Sweep with its provenance.
3. Settle findings and check accepted fixes through Review Sweep. After code or substantive instruction fixes, refresh the full snapshot against the original base and start a new round. Ordinary prose and wording-only fixes use targeted checks instead.

Finish when the required assessments cover the final snapshot, accepted fixes and checks are settled, and code/substantive instructions have a final round with no fixes. Missing required coverage blocks completion. Reopen a settled finding only with new evidence; report stalled disputes rather than repeating equivalent rounds.

In the Orchestration handoff, map child tasks to review assignments, criteria, models, and rounds. Include the reviewed base/snapshot, findings/dispositions, coverage/diversity gaps, skipped additional reviews and reasons, and provider references used or unavailable.