pi-jev

How it uses Jev

Speed-focused Pi extension combining a local coding model with hosted TypeSafe Jev judgments. File retrieval reduces context sent to the generator; exact-result caching and an outage cooldown bound repeated network overhead. Qwen, DeepSeek, and other generators continue to write code and use Pi's tools.

Highlights

• Optional skill suggestion: enable with --jev-skills to have Pi's before agent start hook send the current prompt and loaded skill names/descriptions to Jev in one Choice request.…
• File retrieval: jev search takes a question and 1–16 known file paths. It reads bounded text files locally, selects up to 12 excerpts by word overlap, then asks Jev to prioritize…
• Evidence ranking: the jev rank tool takes a question and up to 24 candidate passages. One batched request scores their relevance. Confident direct evidence moves first; all other…
• Bounded failure: a 1.5-second deadline, no automatic retries, strict response validation, cancellation, and a 24 KB serialized-request limit. A failed ranking returns the original…
• Built-in tool output (opt-in): --jev-tools adds a tool result hook for Pi's grep , find , ls , and bash . Text output of at least 4 KB is split into the same 12-line/800-byte exce…
• Avoid repeated overhead: successful identical requests are cached for 60 seconds in a 32-entry in-memory LRU. The key includes the model, full supplied state, and questions; chang…