Hunch 🔮

Ask a plain-English question across a repository without guessing which words the code uses. Hunch sends each in-scope chunk to Jev and returns scored source locations for a person or coding agent to investigate. It also runs recurring review rules on diffs.

Use with your coding agent

npx skills add Kelbie/hunch

The Agent Skill explains when to search, how to phrase a question, and how to verify the results. No project config is needed for an ad hoc search. Requires Node 22+, Git, and provider authentication: run npx @kelbie/hunch auth login once and Hunch works in every directory (--vercel reuses a Vercel CLI login instead of a key).

These commands require 0.13.0. For unreleased main, run bun run hunch from this checkout.

# Gather implementation, callers, contracts, tests and precedents for a change.
npx @kelbie/hunch find "add cancellation to file uploads"

# Look for an existing behavior, regardless of its names.
npx @kelbie/hunch find "Does this code turn a failed operation into a successful result?" --mode condition

# Get structured evidence for an agent; include every above-threshold match.
npx @kelbie/hunch find "Does this code retry a side effect?" --mode condition --top 0 --reporter json

# Review a branch with a one-off rule.
npx @kelbie/hunch check --rule errors="Preserve failures that callers need to handle."

Use rg for exact symbols and text. Use Hunch when the behavior may have many implementations or names. Read the returned source and follow its dependencies before drawing a conclusion.

Recurring audits

Run npx @kelbie/hunch init to configure plain-English or typed questions. Run them locally with check, across whole files with check --all, or on pull requests through GitHub Actions or the GitHub App. PR policy comes from the immutable base commit.

Installation · Writing rules · Configuration · CLI reference

What the results mean

Search scores rank candidates; findings flag configured concerns. Neither proves a bug or the absence of one. A complete run means the selected chunks were evaluated, not that every relevant behavior was found. Reports identify omissions and distinguish result limits from search coverage.

Code is sent to the selected provider. Cost depends on input tokens, repeated context and questions; small model prices make broad scans practical, but latency and recall still need measurement.

Search workflow · Architecture · Research and evaluation design · Deployment