Community · E2 · artifact verified

Filter streams of text with typed batch judgments

jevpipe pipes thousands of lines, files or records through one question and gets a typed judgment per item - grep-style filtering where Jev decides - shipped as a Rust binary on PyPI with an agent skill that teaches coding agents when to reach for it.

01 · Role in the system

What Jev does here

Any stream becomes a judged stream: git log --format=%s | jevpipe filter "Is this a new feature?" asks the question once per line, and each answer comes back as a typed Choice or Noul over the systemone wire protocol rather than free text to parse. Judgments run in parallel over thousands of items with costs reported at the end; the README's own example classifies all 19 commits of a ripgrep release in one command. Distribution is built for agents: install via uv, pipx or npx skills add, so a coding agent can decide at scale over files, commits or records instead of reading everything itself. Calls route through the operator's chosen systemone-compatible gateway (OpenRouter by default, base-url overridable).

grep, but the predicate is a judgment

jevpipe's idea fits in one line: git log --format=%s piped through a filter asking is this a new feature, with each line answered as a typed Choice or Noul over the wire protocol rather than free text to parse. Thousands of items judge in parallel, costs report at the end, and the README's own example classifies all 19 commits of a ripgrep release in one command.

Distribution is built for the audience that will actually use it: installable via uv, pipx, or npx skills add, with an agent skill that teaches coding agents when to reach for it. The target user is partly not human — an agent deciding at scale over files, commits, or records instead of reading everything itself.

One question per item, and the limits of that

The structural limit is the flip side of the simplicity: judgment quality is bounded by the single question asked per item, with no multi-step reasoning — this is a filter, not an analyst. Default routing also goes through an OpenAI-compatible gateway rather than the native endpoint, so operators wanting the calibrated native path must override the base URL.

The project is young — zero stars at review time, benchmarked only by its own examples — and the record says so. What it contributes to the index is the shell-native form of the batch-judgment pattern: typed judgments as a composable stage in pipelines humans and agents already run.

02 · Control boundary

Where Jev sits

Jev as a stream filter in shell pipelines: one typed question mapped over thousands of items in parallel, structured answers instead of parsed text, and an agent skill so coding agents invoke it where a human would grep.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • Default routing goes through OpenRouter's gateway, not api.typesafe.ai directly; base-url must be overridden for the native endpoint.
  • Judgment quality is bounded by the single question asked per item; no multi-step reasoning.
  • Young project (0 stars at review time); benchmarked only by its own examples.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.