Claude builds. Codex checks. Then switch.

Keep the native tools native.

Install and authenticate Claude Code and Codex CLI using their providers’ normal setup. PapAgent launches the installed commands and coordinates their work; it does not translate both tools into a lowest-common-denominator API.

• Both CLIs must be installed and usable from the local shell.
• The repository must be a Git repository with a clean, known integration branch.
• Each card needs enough context and acceptance criteria to be evaluated independently.
• Provider access, usage limits, model behavior, and billing remain subject to each provider’s current terms.

Two patterns worth using.

Independent building produces the greatest time savings when tasks truly do not overlap. Builder–reviewer work is often safer for a single high-risk change because the second model approaches an existing diff instead of producing a competing implementation.

A concrete builder–reviewer flow.

1. Write the card. State the user-visible outcome, relevant files or interfaces, constraints, and commands that should verify completion.
2. Assign the builder. Let Claude Code or Codex work in the card’s isolated worktree. The choice should follow the task and your own observed results, not a permanent model ranking.
3. Collect evidence. Preserve the changed-file list, full diff, commits, test output, and any unresolved concerns from the builder.
4. Assign the other model to review. Give it the original acceptance criteria and the actual diff. Ask for findings tied to files and behavior—not a general opinion.
5. Make the human decision. Fix material findings, rerun verification, and inspect the final diff before merge. Model agreement is not proof of correctness.

The review handoff should be specific.

A useful cross-model review request separates facts from instructions and asks for actionable findings:

Task outcome:
Users can revoke an active session from account settings.

Acceptance criteria:
- Revocation requires the current user.
- The revoked token stops working immediately.
- Existing session-list behavior remains unchanged.

Review the branch diff against those criteria.
Prioritize correctness, authorization gaps, regressions,
and missing tests. Cite the file and behavior for every finding.
If no material issue is present, say so explicitly.

Do not ask only “Is this code good?” The reviewer needs the intended outcome, the evidence, and a clear severity bar.

Cross-model review is a second perspective, not an oracle.

Claude and Codex can repeat the same mistaken assumption, overlook behavior outside the diff, or approve code that passes narrow tests while breaking integration. The workflow improves coverage of reasoning paths; it does not remove the need for deterministic checks and accountable human review.

• Keep linting, type checks, tests, and security checks deterministic.
• Review generated migrations, permissions, billing, and destructive operations with extra care.
• Rotate builder and reviewer roles when useful; model names are not job titles.
• Track which findings were valid so your workflow improves from evidence.

For broader task decomposition, see the parallel AI coding guide. For the isolation layer, see the git worktree guide.