DEMO / RECORDING
Overview & result
Hunch sends in-scope code chunks to Jev to locate behavior by meaning rather than exact keywords, and can apply plain-English review rules to diffs. It returns source locations for a developer or coding agent to investigate.
From the original post
Ask a plain-English question across a repository without guessing which words the code uses.
How Jev fits in the loop
- Chunk the repository or diff within the selected scope.
- Ask Jev to score chunks against a semantic question or review rule.
- Return ranked source locations and findings; the agent reads the code and verifies the conclusion.
How to reproduce
- Install the Hunch Agent Skill or run the documented `npx` CLI with provider authentication.
- Ask a plain-English behavior question and inspect returned code locations in the source.
- Run a rule check in dry-run mode; follow dependencies before accepting a finding.
Why this build matters
A practical coding-agent retrieval and review case with original terminal demos and explicit uncertainty limits.
Limitations
- Semantic search scores do not prove correctness or absence of a behavior.
- The optional SemIf provider is an independent local model whose accuracy needs separate evaluation.