Pi Jev — madeye

No. The author states it does not accelerate inference; it targets end-to-end latency with focused excerpts and avoided repeat calls, and every uncached Jev call adds latency of its own.

It pins @earendil-works/pi-coding-agent 0.85.1 and the 0.85 series, and older @mariozechner Pi releases are untested. Register it as an extension: pi -e ./src/index.ts for one session, pi install for global settings, or pi install -l with an absolute path to write .pi/settings.json for a single project. The key is read from the environment at load time, so without TYPESAFE_API_KEY or TYPESAFE_BASE_URL, jev_search still runs on local retrieval while hosted ranking, adaptive thinking, and the tool-output filter stay off.

Ranking returns the original passages with an explicit failure status, tool output passes through unchanged, and bounded deadlines with no automatic retries keep failures local.

Keep them opt-in until measured: adaptive thinking only touches Qwen-family local endpoints, and output filtering keeps or withholds excerpts by confidence thresholds with explicit omission markers.

Run npm ci, then pi install followed by the checkout path for a global install or pi install -l /absolute/path for a project one; you can also pin a git install with pi install git:github.com/madeye/[email protected].

Choice confidence of 0.8 or above for skill advice, and passage promotion needs Score confidence of 0.8 or above with a score of 1.5 or more on a 0 to 2 rubric; these are untuned experimental defaults.

Identical successful requests are cached for 60 seconds in a 32-entry in-memory LRU keyed by model, full state, and questions, and after two consecutive failures requests fall back for 30 seconds; nothing is cached on disk.