Distill
Route coding-agent work, reasoning effort, and context retention through bounded Jev decisions.
What Jev decides
Which configured model and effort should handle a call, what evidence deserves attention, whether another attempt is useful, and what context should be retained.
Distill uses Jev as a decision layer around a coding-agent harness: it can select a model and effort, identify useful files and evidence, assess retries, and decide what context survives compaction.
- Jev role
- Make frequent bounded routing, quality, retrieval, and retention judgments around the main coding model.
- Input
- The user request, current turn phase, recent steps, supported model choices, and bounded file, log, search, or tool-result payloads.
- Output
- Typed choices, scores, and yes-or-no probabilities consumed by the harness policy.
- Decision frequency
- At enabled routing and evaluation points during a coding-agent turn and during context reduction.
- Fallback
- Keep the normal session model or execution path when a decision fails, times out, or falls below its confidence threshold; reject invalid utility results.
How it works
Overview
Distill is a Rust coding-agent harness and terminal UI with separate reasoning, worker, and utility model tiers. Jev receives small structured states and chooses only among candidates supplied by the harness; it does not invent permissions or approve tool calls.
Its documented routes cover per-call model and effort selection, candidate and evidence filtering, retry and quality judgments, and context retention. Utility tasks receive bounded payloads instead of the full conversation, and their results are checked before use. Individual routes have confidence floors and per-turn failure limits so an unavailable decision endpoint does not repeatedly block the main agent.
- Language
- Rust
- License
- Apache-2.0
- Author
- samuelfaj
Sources
Related projects
jev-model-router
Route Claude Code model tiers and reasoning effort with one parallel Jev classification.
What Jev decides
Which fast, balanced, or deep tier fits the work, how much reasoning it needs, and whether production risk requires escalation.
Jev Review
A code-review pipeline assembled from small typed judgments instead of one long reviewer prompt.
What Jev decides
What is risky, which files and evidence matter, what could fail, how severe it is, and who should review it.
typesafe-computer-use
A native macOS computer-use loop built from local OCR, accessibility data, and bounded Jev actions.
What Jev decides
Which computer action, visible item, site, or off-screen accessibility control to use next.