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 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".

Dual-Verify (Codex)

Run a full code audit of the current branch vs develop, produce structured findings, and coordinate with the Claude Code deep review.

Platform note: examples use PowerShell syntax. For bash, replace 2>$null with 2>/dev/null.

Role boundary: Codex handles code logic, style, edge cases, metrics completeness, and Windows compat. TypeScript compilation (tsc --noEmit) is Claude Code's responsibility — do not attempt to run it here. Rescue subagent and MCP both operate in headless sandbox; file-reading audit is fully supported.

When

• Before marking any PR as ready for merge.
• When explicitly asked to run dual-verify or parallel review.
• Invoked as rescue subagent in a Claude Code session.

Step 1 — Collect diff

git diff develop...HEAD --stat
git diff develop...HEAD -- packages/orchestrator/src/

Step 2 — Codex audit

Read and analyze changed files. Check:

1. Code style: naming conventions, consistency, no dead code or stray debug artifacts.
2. Edge cases: null/undefined handling, empty collection paths, missing guard clauses.
3. Error handling: all async paths have .catch() or try/catch; no silently swallowed errors except explicitly labeled best-effort.
4. Metrics completeness: verify all four OpenSpace-compatible metric paths are wired:
   • recordSelection() — called when skill is retrieved for dispatch
   • recordApplied() — called at receipt submission (proxy for "agent used skill")
   • recordCompletion() — called on acceptance PASS
   • recordFallback() — called on terminal failure (maxReworks exceeded or BLOCKED)
5. Memory leak: injectedSkillsByTask Map entries deleted on ALL terminal acceptance paths (PASS, FAIL-maxReworks, BLOCKED). Rework cycles intentionally retain entries until terminal state.
6. Windows/PowerShell compat: no Unix-only shell assumptions; paths use join() not string concat.
7. Package accuracy: new import statements — verify package name and version match codebase expectations.

Step 3 — Produce results

## Codex Review Results
Branch: <branch-name>
Critical: N  High: N  Medium: N  Low: N

### Findings:
- [CRITICAL] <description>
- [HIGH] <description>
- [MEDIUM] <description>

Overall: PASS | NEEDS-CHANGES

Step 4 — Report back

Return results to the Claude Code session for consolidation. Do not mark review complete unilaterally.

Evidence

When included in receipt:

dual-verify/codex: PASS (logic audit clean, metrics wired, memory leak fixed on all paths)

Notes

• Report findings even on PASS so Claude's consolidation has full data.
• Codex-only findings (Windows compat, edge cases, metrics gaps) should be flagged clearly.
• TypeScript type errors are Claude Code's responsibility — flag only if apparent upon inspection.

Fallback

If Claude Code is unavailable:

• Use the auto-verify skill (Codex built-in quality gate) as the sole reviewer.
• Document in the PR description that dual-verify was attempted but Claude Code was unavailable.
• High-risk PRs (orchestrator core, auth, schema changes) should wait for Claude Code availability.