Skip to content

Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.

2.5 KB
Raw Blame

Verification

The release gate is bun run verify: source tests, provider-protocol tests, parser differential tests, evaluation-harness tests, and installed Node-only journeys run alongside type and lint checks. Provider requests use local fixtures. No paid-provider or coding-agent task-quality claim follows from these checks.

node scripts/test-native.mjs exercises a fresh npm installation on Apple Silicon with only Node on its runtime PATH. It complements Linux Docker verification; neither environment silently supplies a Python executable or native compiler.

Archive validation must also run from a clean checkout after locked dependency installation and grammar preparation, matching the registry-verification job. Without preparation, canonical grammar comparison must fail; with preparation, the archive identity, bytes, notices, and integrity must match.

The frozen Python helpers are differential oracles. Real-source comparisons must distinguish retrieval regressions from documented syntax-version and class-header boundaries in the parser contracts. Golden fixtures cover each corrected regression and must fail when its fix is removed.

Recorded closeout — 2026-09-29

Integrated with main 6e9bc588. bun run verify passed: 204 source checks (119 Bun, 18 provider protocol, 33 parser, four release, 30 evaluation-harness) and 40 installed-package journeys. Typechecks passed; lint reported only the existing filesystem control-regex warning. Native smoke passed all ten journeys on darwin/arm64 with Node 24.14.0; Docker uses Node 22.

The native-installed archive and canonical archive validation shared SHA-256 3309d67d7ae8d5a9ca824eaf7d1d5a0c2193df6422998d18cd8f34d61c2cff38. Clean-checkout grammar preparation was verified with a failing missing-asset control and a successful canonical-byte, notice, and integrity comparison.

A 155-module Python 3.14 standard-library probe compared all four helpers (620 comparisons). Differences were the documented decorator header and Python 3.14 unparenthesized exception tuples: the prior embedded Python 3.11 runtime also rejects that newer syntax. Neighbourhood comparisons disregard ordering, as their consumer does. Go/Rust text fallback matched main on small and split Unicode fixtures. These probes do not establish task-quality gains.

Independent review findings were reproduced, fixed, and covered by tests that failed without the fixes. The review decisions explain their scope.