dual-verify

name	dual-verify
description	Run parallel Claude Code deep-review and Codex code-audit, then consolidate findings before marking PR ready. Use instead of /code-review or auto-verify when doing pre-merge review. Trigger on: "dual verify", "dual-verify", "parallel review", "run dual verify", "双路验证", "双向验证", "并行review", "双路review".
user-invocable	true
allowed-tools	Bash, Read, Grep, Glob, Agent

Dual-Verify

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

When

• Before marking any PR as ready for merge.
• As a replacement for single-agent /code-review.
• Whenever CLAUDE.md requires code review before commit.

Division of responsibility

Responsibility	Owner
TypeScript tsc --noEmit	Claude Code
Architecture / logic / integration correctness	Claude Code
Code style / edge cases / error handling	Codex rescue subagent
Metrics completeness (all 4 paths wired)	Codex rescue subagent
Memory leak (Map cleanup on all terminal paths)	Codex rescue subagent
Windows/PowerShell compat	Codex rescue subagent

Codex is invoked via Agent tool (rescue subagent) — no manual terminal step required.

Step 1 — Launch parallel reviewers

Claude Code deep review (this session):

# Detect the base branch the current branch was cut from.
# Prefer develop if it exists (Mercury convention), else the repo default branch.
BASE=$(gh repo view --json defaultBranchRef --jq '.defaultBranchRef.name' 2>/dev/null || echo "master")
if gh api "repos/$(gh repo view --json owner --jq '.owner.login')/$(gh repo view --json name --jq '.name')/branches/develop" --silent 2>/dev/null; then
  BASE=develop
fi
git diff "origin/${BASE}...HEAD" --stat
git diff "origin/${BASE}...HEAD"

Check: language-appropriate correctness gates (e.g. tsc --noEmit for TypeScript, pnpm lint, pytest --collect-only for Python), logic correctness, integration points, schema compliance, missing branches in switch/if chains, resource leaks.

Codex audit (rescue subagent — launch via Agent tool with subagent_type: codex:codex-rescue):

Step 2 — Collect results

Each reviewer produces:

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

Step 3 — Cross-reference

Produce a consolidated report:

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

Agreed Issues: <list or none>
Claude-only: <list or none>
Codex-only: <list or none>

Final Verdict: PASS | NEEDS-CHANGES

Step 4 — Fix, verify, mark complete

1. Fix all Critical + High issues.
2. Run auto-verify (tsc --noEmit, scope, lint).
3. Set the review-passed flag:

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

Note: this repo uses .claude/state/ (repo-scoped) instead of .mercury/state/. The original path works only inside the Mercury repo; .claude/state/ is a convention-neutral location that any repo using these skills can adopt.

4. Commit and push.

Evidence

dual-verify: PASS (Claude: PASS, Codex: PASS, N issues fixed)

Rules

• Both reviewers must return PASS before proceeding to merge.
• Fix before merge — do not proceed on a split verdict.
• Codex surfaces Windows-specific and platform concerns that may not be visible in Claude's review.

Fallback

If Codex is unavailable or the session cannot be started:

• Use /code-review (Claude Code built-in) as the sole reviewer.
• Document in the PR description that dual-verify was attempted but Codex was unavailable.
• This fallback is acceptable for low-risk changes; high-risk PRs (orchestrator core, auth, schema changes) should wait for Codex availability.