Skip to content

Add Codex Ultrafast public beta support - #662

Draft
apollostreetcompany wants to merge 2 commits into
repoprompt:mainfrom
apollostreetcompany:codex/feat/ultrafast-catalog-beta
Draft

Add Codex Ultrafast public beta support#662
apollostreetcompany wants to merge 2 commits into
repoprompt:mainfrom
apollostreetcompany:codex/feat/ultrafast-catalog-beta

Conversation

@apollostreetcompany

@apollostreetcompany apollostreetcompany commented Jul 28, 2026

Copy link
Copy Markdown

What this adds

RepoPrompt now shows an Ultrafast — Public Beta option for Codex models that say they support it.

When someone chooses Ultrafast, RepoPrompt keeps that choice when:

  • starting a new session
  • reopening an existing session
  • sending later messages in the same session

The exact ultrafast setting is sent to Codex instead of being dropped or replaced.

What stays the same

  • Ultrafast only appears when the Codex model list says it is supported.
  • Models such as codex-auto-review do not get an Ultrafast option.
  • The existing Fast option still works as before.
  • Choosing no speed option still uses the normal Codex default.
  • If Codex rejects Ultrafast, RepoPrompt shows that error instead of quietly switching models.
  • Ultrafast keeps the model's recommended reasoning level unless the user chooses another one.

This PR does not add a setting for forcing Ultrafast on unsupported models. That is separate follow-up work.

Testing

The focused model-picker and Codex session tests pass. They cover:

  • which models show Ultrafast
  • matching choices across RepoPrompt's model pickers
  • new, reopened, and continuing sessions
  • unchanged Fast and default behavior
  • exact error handling without fallback
  • keeping the model's recommended reasoning level

Formatting, linting, the Codex schema check, the RepoPrompt build, test-ledger verification, repository guardrails, secret scanning, and contribution checks also pass.

The required GPT-5.6 Sol High review found one issue: the first version could forget the model's recommended reasoning level for Ultrafast. Follow-up commit 2dae4eaf72e083d3825f88909b79f658830611b9 fixes it, the focused tests pass, and the follow-up review is clean.

Known validation notes

  • A full local test run was stopped after more than 23 minutes because unrelated worktree and Context Builder tests were failing or hanging. No Ultrafast-focused test failed.
  • The non-disruptive live smoke check could not run because the local debug command-line tool is not installed. The app was not launched or stopped.
  • Hosted checks are running on the latest commit. This PR should remain a draft until those checks finish.

Secret Agent Mac added 2 commits July 28, 2026 15:23
Why: expose the Codex ultrafast service tier only when advertised by the shared catalog, preserve its exact wire value across CLI and app-server turns, and surface provider rejection without fallback.
Why: retain the catalog-advertised reasoning default on collapsed Ultrafast choices without marking the tiered variant as the provider-default model.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant