the-own-lab/critique-dialogue

critique-dialogue

Purpose

Layer 3 of /review. Closes the adversarial loop: Main Agent must file a disposition for every Layer 2 finding, including Plausible Risks. No silent dismissals.

Inputs

• REVIEW.md Layer 2 section (buckets written by the reviewer agent).
• BRIEF.md — Main Agent's working context (unlike Layer 2, Layer 3 may use it).
• Git diff or code file:line referenced by each finding.

Outputs

• REVIEW.md Layer 3 section: one row per finding in the Critique Dialogue table with Disposition + Developer Response + Resolution.
• Any accepted-and-fixed findings produce code edits or new TODO items (written via update-docs, not here).
• needs-user-decision Budget counter updated. Hard cap: 2 per /review run.

Procedure

1. For every finding and plausible risk in Layer 2, assign exactly one disposition:
   • accepted — will fix; describe the fix or commit reference.
   • rejected — requires evidence, not preference; cite file:line or a spec clause.
   • deferred — accepted in principle, tracked as a new TODO or follow-up issue.
   • needs-user-decision — user must pick. Capped at 2 per run.
2. If needs-user-decision count would exceed 2, collapse the lowest-severity one into deferred with a follow-up TODO.
3. For every accepted, either produce the fix in-session (and re-run verify) or write the follow-up TODO.
4. For every rejected, include at least one piece of evidence (code cite, test cite, or explicit spec clause). "Not a real issue" is not a disposition.

Failure modes

• Any finding left with no disposition → FAIL; re-run until closed.
• needs-user-decision > 2 → FAIL; collapse per step 2.
• rejected without evidence → FAIL; either gather evidence or promote to deferred.

Does not do

• Does not edit Layer 2 findings; the reviewer agent owns Layer 2.
• Does not decide the final merge verdict; the Verdict line in REVIEW.md does.
• Does not run session-reflection; /review has no session-reflection step (ADR-001 D7).