Jev Triage — ccai40359-wq

When a test fails, an agent normally re-runs the suite blindly or dumps thousands of lines of stack trace into an LLM. jev-triage adds a third option: classify the failure in one call and tell the agent what class of action it needs, from RETRY to FIX_CODE to FIX_ENV.

Test runner output is parsed offline into structured failures, deterministic rules resolve unambiguous symptoms for free, one batched Jev request covers what remains, and a verdict is emitted with calibrated confidence. Below the threshold it reports UNKNOWN and escalates to a full model or a human.

It works with pytest, vitest, jest and go test, and ships as a CLI, as a Python library and as a Claude Code PostToolUse hook.

On 16 logs captured from real pytest and vitest runs on 2026-09-21: 16/16 correct overall, 9/9 model-only, 44% resolved locally at zero cost, and about 48x cheaper than an LLM reading the same logs. Median Jev latency was 1,109 ms. Running three times on the same corpus returned 16/16 each time.

The 16 logs were written by the same person who wrote the prompt and sample the three categories the prompt was designed around, so it is not a random sample of real-world failures. The README explicitly says to read the numbers section before adopting and treats 100% as a pipeline check, not a guarantee.

Only the genuinely ambiguous failures reach Jev, and they are batched into a single request, so the cost per triage is roughly one small model call rather than one per failure. The deterministic rule layer resolves about a dozen common symptoms with no API calls at all; a TypeSafe key is required for the rest.

The repository is ccai40359-wq/jev-triage under the MIT license, powered by TypeSafe Jev and aimed at coding agents. The measured results come from the author's own live evaluation against the TypeSafe API, not from an independent third party.