pi-openai-codex-fast

Pi package that adds an openai-codex-fast provider backed by built-in openai-codex with serviceTier: "priority".

Behavior

openai-codex-fast is a separate selectable provider that delegates to Pi's built-in openai-codex implementation with the same model id and serviceTier: "priority". Normal openai-codex/<modelId> selections are left on the normal/default-tier path.

Currently exposed fast models:

• gpt-5.5
• gpt-5.4
• gpt-5.4-mini

Runtime behavior when openai-codex-fast/<modelId> is selected:

• Reuses existing openai-codex auth from Pi auth storage.
• Sends Codex requests through the built-in Codex response API with serviceTier: "priority".
• Stores all generated assistant messages canonically as built-in Codex, including normal replies, tool-calling replies, and setup/error/aborted replies:
  • provider: "openai-codex"
  • api: "openai-codex-responses"
• Does not rewrite stored assistant history back to openai-codex-fast or openai-codex-fast-responses.

Fast-mode recovery:

• No custom fast-mode session state is persisted.
• On any session_start reason (startup, reload, new, resume, or fork), the extension scans the current branch backward for the latest overall model_change.
• If that latest model_change is openai-codex-fast/<modelId>, it selects openai-codex-fast/<modelId> again.
• If the latest model_change is anything else, it does nothing and lets Pi's normal model recovery handle it.
• The extension does not handle session_tree, so branch switches do not trigger model reconciliation.

Install

Local path

pi install .

Temporary use

pi -e .

After install, log in to built-in Codex if needed:

/login openai-codex

Then select a fast model with /model, for example:

openai-codex-fast/gpt-5.5

Local development

npm install
npm test # typecheck + integration tests against real Pi runtime/local Codex server
npm run check
npm run lint
npm run fmt
npm run benchmark

Packaging

This package publishes these project files explicitly:

• index.ts
• README.md
• CHANGELOG.md
• LICENSE

Release helper:

npm run release:publish
npm run release:publish -- --execute

npm run release:publish runs npm run check first and defaults to an npm dry-run. Pass --execute to perform the real publish.

The publish helper does not create commits or tags. A real publish requires a clean Git worktree, the release commit pushed to the branch upstream, and a pushed vX.Y.Z tag pointing at the release commit.