pi-gpt-cache

A Pi package that improves OpenAI Responses API prompt caching by adding prompt_cache_key to OpenAI provider requests.

What it does

This package registers a Pi extension that overrides the existing openai provider's openai-responses streaming implementation. It keeps Pi's built-in OpenAI models, delegates request handling to Pi's built-in streamSimpleOpenAIResponses, and injects prompt_cache_key into the final Responses API payload.

The cache key is Pi's options.sessionId, matching the Codex client's default behavior of deriving promptCacheKey from the session ID. Requests from the same Pi session can share OpenAI prompt-cache affinity. If another hook already sets prompt_cache_key, this extension leaves that value unchanged.

Installation

Install as a Pi package:

pi install npm:pi-gpt-cache

For a project-local install:

pi install -l npm:pi-gpt-cache

To test directly from a checkout before publishing:

pi -e ./extensions/openai-responses-prompt-cache.ts
# or
pi install ./path/to/pi-gpt-cache

Configuration

No configuration is required. The prompt cache key always comes from the Pi session id.

Repository layout

• extensions/openai-responses-prompt-cache.ts - Pi extension entrypoint included in the published npm package.

How it works

The extension calls:

pi.registerProvider("openai", {
  api: "openai-responses",
  streamSimple(...) {
    // delegates to streamSimpleOpenAIResponses(...)
    // and adds prompt_cache_key to the payload
  },
});

Because the provider override does not define models, Pi preserves its existing OpenAI model list and only replaces the streaming implementation.

Development

Preview the npm package contents:

npm run pack:dry-run

Publish when ready:

npm publish

Make sure the package name in package.json is available on npm, or rename it before publishing.

License

MIT