Harness · openai/codex
Codex
OpenAI's own terminal coding agent, and the reference implementation of the sandboxed exec loop. Provider-agnostic in practice — it will drive a non-OpenAI model through a custom provider block.
AEQI builds one of the harnesses measured here. This board is in method preview and pre-registration — no performance scores are published yet, and we lead with our own losses, not our wins. Read the method & charter →
Result: not yet measured
The facts
- Maintainer
- OpenAI
- Source
- openai/codex
- License
- Apache-2.0
- Category
- Coding harness
- Headless run
- Yes — `codex exec` (non-interactive, `--json` event stream)
- Native OpenRouter
- — can be pinned to one shared model
- Determinism notes
- Pin the npm version. Points at any OpenAI-compatible base URL via `model_providers`, so one pinned model is straightforward.
Where it sits
Performance scores land here only with published traces, under the pre-registered method. Until then, this profile is the fair-run readiness check: can we start Codex headless and pin it to one shared model to isolate the scaffold?