Codex Code Review for Local and Uncommitted Changes
Let Codex implement and self-check quickly, then use a separate finish pass when the completion claim needs stronger evidence.
Ask Codex to implement and run targeted checks. Before accepting completion on consequential work, capture the whole working tree and send the declared intent to an independent reviewer. Feed the structured result back to Codex for repair or handoff.
Self-check first, independent check second
Codex can inspect repository state, edit files, run commands, and review its own work. That is the shortest feedback loop and should stay. A second pass earns its cost when the failure could hide behind the same assumptions that shaped the patch and its tests.
Connect once, then call the finish pass
npm install --global @critiquedotsh/cli
critique login
critique integrate --agent allWrite an intent that can be disproved
“Review my code” invites generic commentary. “Reject a duplicate idempotency key without creating a second charge, and keep the existing retry response stable” gives the verifier something concrete to reproduce. State the invariant, failure condition, and unchanged behavior.
Review a real Codex change
At the finish boundary, ask Codex to summarize the intended behavior and the checks it already ran, then let the independent pass reconstruct that claim from the working tree. If the result identifies a defect, return the smallest repair context: finding, evidence, affected paths, and the check that must be rerun.
- 1IntentThe behavior that must exist and the behavior that must not regress.
- 2ScopeThe complete candidate working tree, including new and deleted paths.
- 3EvidenceCommands, fixtures, outputs, and any unavailable environment.
- 4Next actionA clean result, a bounded repair request, or an explicit human decision.