Codex-Claude Engineering Loop

Workflow Overview

Plan (Claude) → Validate (Codex) → Implement (Claude) → Review (Codex) → Fix → Re-validate → Done

Environment Notice

Non-TTY environment: See codex-cli SKILL for CLI fundamentals. Key rule: Always use codex exec in Claude Code (not codex)

Phase 0: Pre-flight Check

Before executing any Codex command:

1. Create context directory:

mkdir -p .codex-loop

2. Check Git Repository Status

git rev-parse --git-dir 2>/dev/null && echo "Git OK" || echo "Not a Git repo"

3. If NOT in Git repository, ask user via AskUserQuestion:

   • "This directory is not a Git repository. Codex requires Git by default."
   • Options:
     • "Initialize Git repository here (git init)"
     • "Use --skip-git-repo-check flag (bypass check)"
     • "Cancel operation"

4. Apply user choice:

# Option 1: Initialize Git
git init

# Option 2: Use skip flag
codex exec --skip-git-repo-check -s read-only "prompt"

5. Ask user via AskUserQuestion:
   • Model preference (gpt-5.2-codex, gpt-5.2, gpt-5.1-codex-max, gpt-5.1-codex-mini)
   • Reasoning effort level (low, medium, high, xhigh)

Phase 1: Planning (Claude)

1. Create detailed implementation plan
2. Break down into clear steps
3. Document assumptions and risks
4. Save to .codex-loop/plan.md

Phase 2: Plan Validation (Codex)

Execute with timeout: 600000:

codex exec -m MODEL -c model_reasoning_effort=LEVEL -s read-only \
  "Review this plan: $(cat .codex-loop/plan.md) Check: logic, edge cases, architecture, security"

Save result: > .codex-loop/phase2_validation.md

Phase 3: Feedback Loop

If issues found in .codex-loop/phase2_validation.md:

1. Summarize Codex feedback to user
2. Ask via AskUserQuestion: "Revise and re-validate, or proceed?"
3. If revise → Update .codex-loop/plan.md → Repeat Phase 2

Phase 4: Implementation (Claude)

1. Implement using Edit/Write/Read tools
2. Execute step-by-step with error handling
3. Save summary to .codex-loop/implementation.md

Phase 5: Code Review (Codex)

Execute with timeout: 600000:

codex exec -m MODEL -s read-only \
  "Review: $(cat .codex-loop/plan.md) $(cat .codex-loop/implementation.md) Check: bugs, performance, security. Classify: Critical/Major/Minor/Info"

Save result: > .codex-loop/phase5_review.md

Claude response by severity:

• Critical → Fix immediately
• Architectural → Discuss with user
• Minor → Document and proceed

Phase 6: Iteration

1. Apply fixes from .codex-loop/phase5_review.md
2. Significant changes → Re-validate with Codex
3. Use codex exec resume for session continuity
4. Loop until quality standards met

Context Files

.codex-loop/
├── plan.md               # Implementation plan
├── phase2_validation.md  # Plan validation result
├── implementation.md     # Implementation summary
├── phase5_review.md      # Code review result
└── iterations.md         # Iteration history

Quick Reference

Always use timeout: 600000 (10 min) for all Codex commands.

Models, reasoning effort, options: See codex-cli SKILL

Error Handling

Full error reference: See codex-cli SKILL

Error Recovery Flow:

1. Non-zero exit → Stop and report error message
2. Summarize error via AskUserQuestion
3. Confirm with user before: architectural changes, multi-file mods, --skip-git-repo-check

Timeout Configuration

Approval modes: See codex-cli SKILL - valid values: untrusted, on-failure, on-request, never

Best Practices

• Always use codex exec in Claude Code environment (non-TTY)
• Always create .codex-loop/ directory at start
• Always save outputs to context files for traceability
• Always validate plans before implementation
• Never skip review after changes
• Default to -s read-only for all reviews
• Use resume for session continuity
• Check Git status before first Codex command
• Ask user permission before using --skip-git-repo-check
• Set 10-minute timeout for all Codex exec commands (timeout: 600000)

References

• Codex CLI fundamentals: See codex-cli SKILL (models, options, error handling)