Skip to content

FAQ

What files does CodexOpt improve?

CodexOpt targets the instruction files Codex users maintain most often:

  • AGENTS.md
  • .codex/skills/**/SKILL.md
  • .agents/skills/**/SKILL.md

What should I run first?

Start with a preview:

codexopt improve

Then run the live Codex loop when you want Codex-backed feedback:

codexopt improve --live

Does CodexOpt execute Codex tasks?

Yes, when you use live mode or Codex rollout tasks. CodexOpt can run codex exec --json in a temporary repo copy, parse the trajectory, and use that feedback for scoring and reflection.

Do I need GEPA installed?

No. The maintained reflective engine is implemented in CodexOpt and does not depend on the gepa package.

How is SkillOpt integrated?

CodexOpt uses SkillOpt-style controls in the optimization workflow:

  • train and validation splits
  • bounded edits
  • held-out validation gates
  • rollout reward when available
  • accepted and rejected candidate reporting

For most users, these controls are used through:

codexopt improve
codexopt improve --live

Which models can I use?

Use codex to run through codex exec, or use an OpenAI-compatible model such as openai/gpt-5-mini.

Configure models in codexopt.yaml:

reflective:
  optimizer_model: "codex"
  judge_model: "codex"

You can also use command flags:

codexopt improve --optimizer-model codex --judge-model codex

What happens when no live model is configured?

CodexOpt stays offline and uses deterministic cleanup plus static/verifier signals. It records notes in the report when it uses this weaker signal.

No. --engine gepa is a deprecated legacy path. Use --engine reflective or codexopt improve --live.