Feedback

Codex agent backend: let me use my own Codex config (custom model and provider)

K
katepowlerinFeedback
8 days ago

I use Talat with an OpenAI-compatible gateway on localhost. Summaries work well through a custom LLM profile, but the Codex agent backend can't reach the same gateway.

From what I can see, Talat runs Codex with --ignore-user-config and only allows models from a built-in list (GPT-6 Astra/Sol/Luna, GPT-5.6, GPT-5.5). Any other model ID is rejected in settings, or falls back to the default at runtime.

So the agent always goes to OpenAI on the codex login account, even when ~/.codex/config.toml defines another provider.

That rules out the agent for anyone whose company routes model traffic through an approved gateway or proxy, or who wants to run a local or self-hosted model. I understand that its an experimental feature but it seems at odds with Talat being local-first, especially since the custom LLM profiles already support this for summaries.

I believe any one of these would solve it:

1. Respect the user's Codex config: add an opt-in setting such as "Use my ~/.codex/config.toml" that drops --ignore-user-config, so model_provider and model come from the user's config.

2. Custom model ID: accept a free-text model name for the Codex backend in addition to the built-in list.

3. Custom provider: let the Codex backend use an existing custom LLM profile (base URL and model) by passing it to Codex with -c model_provider=… / -c model_providers.<id>.base_url=….

Option 1 is probably the smallest change. Options 2 and 3 would also help users who don't maintain a Codex config.

  1. In Review
  2. Planned
  3. In Progress
  4. Completed
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...
Loading content, please wait...