Codex Problems

Cannot switch arbitrary custom model IDs in the TUI

Files:

  • codex-rs/tui/src/chatwidget/model_popups.rs
  • codex-rs/tui/src/slash_command.rs
  • codex-rs/tui/src/chatwidget/slash_dispatch.rs
  • codex-rs/tui/src/app/event_dispatch.rs
  • codex-rs/tui/src/app/thread_settings.rs
  • codex-rs/app-server/src/request_processors/turn_processor.rs
  • codex-rs/core/src/codex_thread.rs

Problem: /model only shows visible ModelPresets from the resolved model catalog (try_list_models() + show_in_picker). There is no /model <model_id> path and no free-text "Custom model…" picker entry.

The runtime path can already accept arbitrary model strings: AppEvent::UpdateModel(String) → thread/settings/update → Op::ThreadSettings → collaboration_mode.with_updates(model, effort, ...).

So the small missing piece is UI/command input for an arbitrary model ID.

Important limitation: This is model-string switching only. UpdateModel does not switch model_provider. Switching providers/profiles in-session is a separate, larger problem and may require a new/forked session to avoid provider/auth/history incompatibilities.

Impact: Custom/self-hosted model users cannot switch among uncataloged model IDs interactively. They must restart or preconfigure the desired model/catalog.

Workarounds:

  • Start with codex -m <model_name>.
  • Start with codex -p <profile> for provider+model.
  • Set model / model_provider in config.toml before launch.
  • Use startup-only model_catalog_json so custom models appear in /model.

Related upstream issues:

  • openai/codex#5841: dynamic custom model switching; closed not planned.
  • openai/codex#22160: expose profiles/custom aliases in pickers; closed not planned.
  • openai/codex#24659: custom models not loaded from models_cache.json; model_catalog_json workaround.
  • openai/codex#24612: cross-provider switching can fail due to incompatible history items.

Suggested minimal fix: Support inline model args:

/model <model_id>

Implementation:

  1. Add SlashCommand::Model to supports_inline_args().
  2. In slash_dispatch.rs, handle non-empty /model args.
  3. Dispatch UpdateModel(model) and PersistModelSelection { model, effort: None }.
  4. Warn/document that this changes only the model string for the current provider.

Optional UI fix: Add "Custom model…" to the /model picker and reuse the same helper as /model <model_id>.