name: codex description: > Use codex exec for large multi-file implementation (5+ files) and code review. Escalation from Worker when task scope exceeds 1-3 files. Codex cannot run project tooling — Worker handles tests and builds.

Codex

Codex runs in a Windows sandbox with workspace write access. Use it when a task genuinely needs broad file context — typically 5+ files or a cross-cutting refactor. Default to Worker for anything under 5 files.

No outbound network. No uv/pytest. Implementation and review only.

When to Use Codex Over Worker

• Multi-file refactors touching 5+ files
• Large feature where full file context across modules matters
• Changes requiring reading many files to understand dependencies
• When Worker would need 4+ sequential invocations for one logical change

When to Use Worker Instead

• Single-file changes
• Mechanical fixes (lint, types, formatting)
• Test writing for a specific module
• Small bug fixes with clear scope
• Anything touching 1-3 files

Sandbox Modes

Review Mode

Codex has full filesystem read access in the sandbox. Pass file paths in the prompt — do NOT inline file content as shell arguments. Large diffs or files passed via $(cat ...) will hit the OS argument length limit and fail.

git diff HEAD > /tmp/codex-review-diff.patch

codex exec \
  --sandbox read-only \
  --output-last-message /tmp/codex-review-output.txt \
  "Review the diff at /tmp/codex-review-diff.patch. Focus on:
1. Bugs and logic errors (P0)
2. Security vulnerabilities (P0)
3. Unhandled edge cases (P1)
4. Performance issues when material (P1)

Skip style nits, formatting, naming preferences.

For each finding: file, line range, priority (P0/P1), description."

Clean up: rm /tmp/codex-review-diff.patch /tmp/codex-review-output.txt

Implement Mode

codex exec \
  --sandbox workspace-write \
  --output-last-message /tmp/codex-implement-output.txt \
  "<fully scoped implementation prompt>"

Prompt must include: exact goal, scope boundaries, explicit exclusions, required output format. Same structure as Worker prompts (Goal, Scope, Approach, Constraints) but broader scope.

Rules

• Review: --sandbox read-only always
• Implement: --sandbox workspace-write always
• One Codex pass only — no iteration loop
• Prompts must be fully self-contained
• Always use --output-last-message and read the file after
• After 3 failed attempts on same goal, stop and surface to Conductor

Verification

Codex runs a different model with different context. Treat all output as input to Conductor's judgment:

• Read every changed file before committing (use line ranges for large files)
• Cross-check against original intent
• Flag anything outside requested scope
• Have Worker run tests after Codex implements