mirror of
https://github.com/openai/codex.git
synced 2026-11-19 22:30:42 +00:00
add client-side ModelInfo overrides
This commit is contained in:
@@ -28,6 +28,32 @@ Codex can run a notification hook when the agent finishes a turn. See the config
|
||||
|
||||
The generated JSON Schema for `config.toml` lives at `codex-rs/core/config.schema.json`.
|
||||
|
||||
## `model_info_overrides` precedence
|
||||
|
||||
`model_info_overrides` lets you patch metadata for specific model slugs in `config.toml`.
|
||||
The patch is merge-only: only fields you set are overridden, and omitted fields keep their
|
||||
resolved value.
|
||||
|
||||
Precedence for model metadata is:
|
||||
|
||||
1. Resolved model metadata (remote `/models` when available, otherwise built-in fallback)
|
||||
2. `model_info_overrides[<slug>]`
|
||||
3. Existing top-level config overrides (for example `model_context_window`,
|
||||
`model_auto_compact_token_limit`, `model_supports_reasoning_summaries`, and
|
||||
`tool_output_token_limit`)
|
||||
|
||||
This means top-level overrides still win for their specific fields after per-model patching.
|
||||
|
||||
Example:
|
||||
|
||||
```toml
|
||||
[model_info_overrides.gpt-fake]
|
||||
display_name = "gpt-fake-dev"
|
||||
context_window = 400000
|
||||
supports_parallel_tool_calls = true
|
||||
base_instructions = "Custom model instructions"
|
||||
```
|
||||
|
||||
## Notices
|
||||
|
||||
Codex stores "do not show again" flags for some UI prompts under the `[notice]` table.
|
||||
|
||||
Reference in New Issue
Block a user