FAQ

Why only AGENTS.md and SKILL.md?

CodexOpt is intentionally narrow. These are the repo-local instruction assets that developers most commonly maintain for Codex workflows.

Does CodexOpt execute full task simulations?

Not yet. Task and issue files currently shape scoring and feedback rather than running end-to-end agent executions.

Do I need GEPA?

No. Heuristic mode works without GEPA and is the default.

Do I need prompt-learning as a dependency?

No. CodexOpt does not depend on the prompt-learning library. It uses some of the same ideas around natural-language feedback, but the implementation is native to CodexOpt.

Which models can I use with GEPA?

CodexOpt's GEPA path is model-agnostic. You can use OpenAI, Gemini, local models, or other GEPA / LiteLLM-compatible providers.

What if a GEPA run fails?

CodexOpt falls back to heuristic optimization and records that fallback in artifacts and reports.