| # Codex CLI |
|
|
| Facts from the Codex configuration reference (2026-10-15): `config.toml` keys `model`, `model_provider`, |
| `[model_providers.<id>]` (`base_url`, `env_key`, `wire_api`; `responses` is the only supported value), |
| `model_reasoning_effort` (values depend on the model: `low`, `medium`, `high`, `xhigh`, …), profile files |
| `$CODEX_HOME/<name>.config.toml` selected with `--profile`, and lifecycle hooks (`features.hooks = true`, event |
| `UserPromptSubmit`, command handlers). I did not find a documented way for a hook to change the model. |
|
|
| ## 1. Choose at launch (*router untested; `-m` and `-c` assumed from the Codex CLI*) |
| ```bash |
| python integrations/serve.py --model scombat/s1d1-router --policy integrations/policy.example.yaml & |
| python integrations/routed.py codex "fix the failing test in utils/" |
| ``` |
| runs `codex -m <model> -c model_reasoning_effort="<effort>" "<prompt>"`. Edit the `codex` profile in |
| `policy.example.yaml` first: put the model ids of your account in `models` (they are placeholders) and keep only the |
| efforts your models accept. If `-c` is not available in your version, use profile files (one per tier) and |
| `--profile`. |
|
|
| ## 2. Through a gateway |
| Point a custom provider at a LiteLLM proxy (`base_url = "http://localhost:4000/v1"`, `env_key`), with the Responses |
| API enabled on the proxy, and use the hook of `02-litellm.md`. *Untested: the Responses path of the hook is not |
| written; check that LiteLLM runs `async_pre_call_hook` on `/v1/responses` for your version.* |
|
|
