feat(codex): wire cross-model verification into verify, ship, fix, nuclear-review#59
Merged
Merged
Conversation
…clear-review Slots the Codex bridge into the four skills where Claude judges its own output — the self-preferential-bias surface cross-model review exists to break. Each insertion is gated and fails open: if the bridge is unavailable, the skill runs Claude-only. - verify: adds a non-Claude finder (codex-verifier) in parallel with the Claude Finder, merged into the superset before the Adversary stage — a different model family in an otherwise all-Claude panel. - ship: parallel cross-model review at Step 6, right before commit; a Claude/Codex disagreement on a Critical becomes a human gate before Step 7. - fix: cross-model review alongside the reviewer for security-sensitive fixes (auth, crypto, permissions, input validation). - nuclear-review: a whole-codebase 'ask' pass (NOT diff-scoped review — there is no diff in a repo-wide audit) folded into Phase 3 synthesis as a second opinion. Verifier uses the codex-verifier agent for diff-based skills; nuclear-review uses codex-run.ts ask directly since it audits the whole tree, not a change.
Codex is the roomy pool (600–3000 msgs / 5h on Pro) — quota was never the constraint, so the two high-frequency review skills get the cross-model pass as well. review runs as the reviewer agent (Bash, no Agent tool), so it calls codex-run.ts review directly; refactor is in orchestration mode and uses the codex-verifier agent in parallel — refactors are a top source of subtle regressions, exactly where a second model family pays off. Both gated, fail open.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this does
Wires the Codex bridge into the six skills where Claude reviews its own work —
verify,ship,fix,nuclear-review,review, andrefactor. In each, an independent review from a different model family (Codex) runs alongside Claude's, so a blind spot Claude shares with itself gets a second pair of eyes. It's opt-in by availability: if Codex isn't installed or logged in, every one of these skills runs exactly as before, Claude-only.The point is the self-preferential-bias surface — Claude judging Claude tends to wave its own output through. A different family doesn't share those blind spots, so this is where the bridge earns its keep. Codex is the roomy pool (600–3,000 messages / 5h on Pro), so quota is not a reason to be sparing — the high-frequency review skills are wired in too.
Summary
verify— adds a non-Claude finder (codex-verifier) in parallel with the Claude Finder, merged into the issue superset before the Adversary stage. The skill's whole reason to exist is breaking self-review bias; its three-agent panel was 100% Claude.ship— parallel cross-model review at Step 6, right before a commit lands. A Claude/Codex disagreement on a Critical/HIGH becomes a human gate before Step 7.fix— cross-model review alongside the reviewer for security-sensitive fixes (auth, crypto, permissions, input validation).nuclear-review— a whole-codebaseaskpass folded into Phase 3 synthesis. Deliberately not the diff-scopedreview— a repo-wide audit has no diff, so it usescodex-run.ts askpointed at the tree.review— runs as therevieweragent (hasBash, noAgenttool), so it callscodex-run.ts reviewdirectly and reconciles findings into the verdict (HIGH→Critical, MEDIUM→Warning, LOW→Suggestion).refactor—codex-verifierin parallel with the reviewer; refactors are a top source of subtle regressions, exactly where a second family pays off.Diff-based skills with
Agentaccess use thecodex-verifieragent;review(a forked reviewer) andnuclear-review(diff-less) callcodex-run.tsdirectly. Every insertion states the bridge is gated and fails open.Test Plan
bun run lint:skills— 35 skills, 0 errors/warningsnuclear-reviewusesask(correct for a diff-less audit) andreviewuses the directcodex-run.tscall (the reviewer fork has noAgenttool)