Dual Verify

Dual-Verify: Parallel Claude + Codex Code Review

Run Claude Code deep review and Codex code audit in parallel, then consolidate findings before marking review complete.

When

• Before marking any PR as ready for merge.
• When the task is implementation-complete and you are preparing the final commit batch.
• As a replacement for single-agent code-review — whenever review is required by CLAUDE.md.
• On the same branch; no worktree isolation required.

Workflow

Division of responsibility

Codex is invoked as rescue subagent — fully automated, no manual terminal step required.

Step 1 — Launch both reviewers in parallel

Claude Code (deep review — TypeScript correctness, architecture, integration, OpenSpace schema):

git diff develop...HEAD --stat
git diff develop...HEAD
npx tsc --noEmit  # Claude Code's responsibility

Codex (logic audit — style, edge cases, metrics, memory leaks, Windows compat):

Invoke the Codex audit agent with this prompt:

Audit feat/<branch> vs develop. Check: code style, edge cases, metrics completeness (all 4 paths), memory leak cleanup on terminal paths, Windows compat. TypeScript typecheck is not required.

(Agent-specific invocation syntax is defined per-agent in .claude/skills/dual-verify or .agents/skills/dual-verify.)

Step 2 — Collect results

Both reviewers produce a structured result block:

## <Reviewer> Review Results
Critical: <count>  High: <count>  Medium: <count>  Low: <count>
<finding-1>
<finding-2>
Overall: PASS | FAIL | NEEDS-CHANGES

Step 3 — Cross-reference findings

Consolidate into a unified report:

## Dual-Verify Consolidated Report
Branch: <branch-name>
Claude: PASS | FAIL | NEEDS-CHANGES
Codex:  PASS | FAIL | NEEDS-CHANGES

### Agreed Issues (both flagged):
- <issue>

### Claude-only:
- <issue>

### Codex-only:
- <issue>

### Resolution Plan:
- <action-1>
- <action-2>

### Final Verdict: PASS | NEEDS-CHANGES

Step 4 — Fix and mark review

1. Fix all Critical and High issues.
2. Address or document Medium issues.
3. Run auto-verify to confirm TypeScript and scope pass.
4. Mark review flag:

mkdir -p .mercury/state && touch .mercury/state/review-passed

5. Commit and push.

Output Evidence

Record in implementationReceipt.evidence:

dual-verify: PASS (Claude: PASS, Codex: PASS, N critical fixed, M high fixed)

Notes

• Both reviewers must agree on PASS before the PR is marked ready.
• If one reviewer returns NEEDS-CHANGES, fix before proceeding — do not merge on split verdict.
• Codex may surface Windows/PowerShell-specific concerns not visible to Claude; always include Codex in the loop.
• Cross-referencing catches false positives: an issue flagged by only one reviewer should be investigated, not auto-dismissed.

Fallback

If one reviewer is unavailable, fall back to that reviewer's individual skill:

Document in the PR description that dual-verify was attempted but one side was unavailable. High-risk PRs (orchestrator core, auth, schema changes) should wait for both reviewers.

Source: 392fyc/Mercury — distributed by TomeVault.