@tylerho/pi-long-cache

Forces long prompt-cache retention in-process so providers without an explicit opt-in still get caching.

Install

pi install npm:@tylerho/pi-long-cache

---

Long cache

Sets process.env.PI_CACHE_RETENTION = "long" while pi loads, so every provider request in the process attempts the longest prompt cache the provider supports. Pi defaults to "short" retention, and without this, OpenAI-compatible endpoints other than OpenAI (deepseek, openrouter) send no prompt_cache_key at all, so no prompt caching is attempted for them.

How it works

The default export (_pi: ExtensionAPI) runs once at extension load, which includes startup and /reload, and ignores its argument. Setting the variable before the first provider request is what matters, and extension load happens early in the process.

Anthropic-style models (kimi-coding, anthropic) then send cache_control with ttl: "1h" instead of the 5-minute default. OpenAI-compatible models (deepseek, openrouter) send prompt_cache_key plus prompt_cache_retention: "24h" instead of in-memory retention. Providers without long-retention support are unaffected, because "long" is a request rather than a guarantee.

The extension covers pi's own process only. ~/.zshrc holds the same export (export PI_CACHE_RETENTION=long) for anything launched outside pi's process. Delete extensions/long-cache/index.ts, or the export line inside it, to disable the extension. No settings toggle exists.