jevpipe
ActiveA Unix filter that lets coding agents program their own System 1: fast, cheap typed decisions over thousands of files, lines or records in one shell command.
Why
When I came across Jev, TypeSafe AI’s calibrated decision model, it clicked that this is a really interesting building block for combining a fast, cheap System 1 with a slow, deliberate System 2 reasoning model. I immediately wanted to connect the two. jevpipe is a small experiment in that direction, and I expect AI labs and other model providers to push the decision model frontier much further.
jevpipe puts System 1 filtering and decisions in the hands of a System 2 model. Instead of grepping or spending expensive reasoning and subagents on thousands of small calls, the agent programs its own System 1 for ad-hoc decisions: it frames the question, pipes the items through Jev, and only reasons about what comes back. In an independent study, Jev classified about as accurately as small LLMs at a fraction of the cost.
Highlights
- A decision per item: pipe in commits, log lines, issues or file paths and ask each one a question, e.g.
git log --format=%s | jevpipe filter "Is this a new feature?" - Typed answers:
mapasks several questions at once (yes/no, choice, score) and returns calibrated probabilities and confidence per record - Cheap and fast: about a cent per 1,000 commit messages, up to 100 records in parallel, with a
--max-costbudget cap - Agent-ready: ships as a skill that teaches your coding agent when and how to use it
jevpipe is an independent project, not affiliated with TypeSafe AI. Jev is their model, accessed through OpenRouter.