Jev Triage — Testing skill for Claude Code

jev-triage


**Millisecond-class test-failure triage for coding agents — with honest numbers.**

When a test fails, an agent normally has two options: re-run the suite blindly, or dump a few thousand lines of stack trace into an LLM and wait. `jev-triage` adds a third: classify the failure in one call and tell the agent what class of action it needs.

RETRY      flaky or transient — re-run unchanged, do not touch the source
FIX_CODE   a real defect in the code under test — fix the source, do not just re-run
FIX_ENV    missing dependency, unset env var, unreachable service — fix the setup

Works with pytest, vitest, jest and `go test`, as a CLI, as a Python library, and as a Claude Code `PostToolUse` hook.

**Read the numbers section before you adopt this.** Measured on 16 logs captured from real pytest and vitest runs: **16/16 correct**, **44% resolved locally at zero cost**, and **~48x cheaper** than handing the same logs to an LLM. Measured latency is **0.8–2.5 s per call** (median ~1.1 s) from our network path — not the 70–500 ms in the Jev launch material, but well below the multi-second wait for an LLM to read the same log.

---

What it actually does

test output ──► parsers ──► deterministic rules ──► Jev ──► verdict + advice
                (offline)   (offline, no cost)      (1 call)

1. Parse the runner output into structured failures (test name, error type, location, message, evidence). Pure text processing, no network, no cost.
2. Apply deterministic rules for symptoms that are unambiguous — ModuleNotFoundError, KeyError: 'DATABASE_URL', Address already in use, ECONNRESET, and about a dozen more. These resolve with zero API calls and zero cost.
3. Ask Jev about everything left over: the genuinely ambiguous failures. One batched request asks about a