Question 1

Should I just use ripgrep instead of paying an API call per line?

Accepted Answer

If a regex can express what you want, the README itself says use grep or rg. It exists for questions a pattern cannot express: on its benchmark log it keeps a 503 access-log line, a Traceback, and an OOMKilled event, none of which contain the word error, while dropping error_rate=0.00, 0 failed payments, and error handler registered.

Question 2

Are those F1 and cost comparisons actually trustworthy?

Accepted Answer

Treat them as indicative. The benchmark is one hand-labelled 195-line sample with each model run once, so there is no variance estimate. Mean F1 is 0.904 for Jev against 0.873 for Claude Haiku 4.5 and 0.952 for Claude Sonnet 5, meaning the strongest model is still more accurate; the win is cost, at $0.0024 against $0.1113 and $0.4108.

Question 3

Will it keep up behind tail -f, or lag minutes behind?

Accepted Answer

It keeps up, with a measured lag. A match appears about 0.7 s after the line is written, which is a 300 ms idle timer plus one roughly 0.4 s request, and at most 0.82 s once warm. The first batch pays 1.0 to 1.5 s for connection setup, and output order is preserved no matter how many batches are in flight.

Question 4

Does any of my log content stay on my machine?

Accepted Answer

No. Every judged line is sent to TypeSafe or OpenRouter depending on which key is set and the provider setting, so nothing stays local. There is a caching consolation: identical lines are judged once per run from an LRU cache of 100,000 entries, and lines over 500 characters are truncated before judging though still printed in full.

Question 5

How do I install jevgrep?

Accepted Answer

uv tool install jevgrep-cli, or pipx install jevgrep-cli, or uv tool install 'jevgrep-cli[explain]' for --explain, then set OPENROUTER_API_KEY or TYPESAFE_API_KEY. uvx --from jevgrep-cli jevgrep runs it without installing.

Question 6

What are the license and exit codes?

Accepted Answer

MIT. Exit status follows grep: 0 if any line was selected, 1 if none, and 2 on errors. Ctrl+C exits 130 after printing --explain and --stats, and a closed pipe such as | head stops quietly.

Question 7

How do batching and caching work?

Accepted Answer

A batch is sent at --batch-size lines (default 20), after 300 ms without input, or once its oldest line has waited 1 s. Identical lines are judged once per run from a 100,000-line LRU cache, and lines over 500 characters are truncated before judging.

Jevgrep — allebee

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community