jevgrep Files

Fix runaway searches

Jevgrep now bounds speculative exploration when many previews look weakly relevant but no useful source has been confirmed. It reuses early source judgments, drains work already running before stopping, and reports incomplete discovery clearly. Relationship checks judge each directory or whole-file preview using its own source.

• Default provider-attempt ceiling: 1,000; configurable with --max-requests.
• Default total stdout ceiling: 256,000 bytes; configurable with --max-output-bytes (0 disables the output ceiling). Truncated packets report omitted bytes and retain closed source fences and the final marker.
• Default source allocation, parser, and public Jevgrep skill remain unchanged.

Fixes [#41].

Verification

The retained ten-task development evaluation preserved all eight original release solves. Combined reported coding-agent and Jev cost was $5.7890 versus the matched release's $6.0776–$6.1174, at least 4.7% lower. These tuned tasks do not establish performance on unseen repositories. The release differs from the evaluated archive in version identity and generated source-path comments.

All eleven native macOS Apple Silicon installation checks passed locally. The release workflow verifies the package, exercises the exact archive in its Node-only Linux runtime, and verifies the published registry bytes.

Research record

Upgrade

:::sh
npm install -g @dzhng/jevgrep@latest