Overview & result
A Pi extension combines local coding generation with optional Jev judgments for file retrieval, output focusing and exact-result caching. The author publishes positive context-compression results alongside negative latency tests and a local-model alternative.
From the original post
This does not accelerate model inference.
How Jev fits in the loop
- Retrieve focused file excerpts locally before sending context to the generator.
- Optionally use Jev to rank evidence or filter large tool outputs; reuse exact cached judgments when possible.
- Let Pi's generator write code, while fallback and outage cooldown bound network overhead.
How to reproduce
- Install the extension in Pi and configure a TypeSafe key or the documented compatible local endpoint.
- Use `/jev find` on a known repository question and inspect ranked source excerpts.
- Read VALIDATION.md and repeat the frozen retrieval and coding comparisons before claiming a speed benefit.
Why this build matters
Strong counter-evidence that a useful Jev decision layer can save context without necessarily making the full coding task faster.
Limitations
- The adopted author configuration uses a self-hosted DiffusionGemma-compatible endpoint, not always hosted TypeSafe Jev.
- In the author's small coding comparison, median completion time increased from 43.24s to 57.84s with Jev retrieval available; network timeouts occurred.