Hunch (Kelbie)
Semantic code search and plain-English review rules that send repository chunks to Jev and return scored source locations.
Overview
Hunch by Kelbie lets a person or coding agent ask a plain-English question across a repository without guessing which identifiers the code uses. Each in-scope chunk is sent to a decision model and Hunch returns scored source locations to investigate. The same engine runs recurring review rules on diffs, locally, across whole files, or on pull requests via GitHub Actions or a GitHub App, with PR policy read from the immutable base commit.
It ships as an Agent Skill installable with npx skills add, plus a CLI. Providers are selectable per run: an AI gateway, TypeSafe Jev directly, or SemIf, an open model on your own GPU with no key. Search has modes such as condition questions (does this code retry a side effect?), JSON output for agents, and published rule packs. The README is careful that scores rank candidates and findings flag concerns; neither proves a bug, and a complete run means selected chunks were evaluated, not that everything was found.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this code turn a failed operation into a successful result? | yes · no |
| Score | How relevant is this chunk to the change or behavior being searched for?(inferred) | — |
| Choice | Which label applies to this finding under the configured rule?(inferred) | — |
How it works
- Install the agent skill or run the CLI with a question or rule
- Chunk in-scope files or the PR diff
- Send each chunk with the question to the chosen provider
- Rank locations or emit findings above threshold
- Person or agent reads the source to confirm
Strengths
- Finds behaviors that have many names, where grep fails
- Runs fully locally with --provider semif, sending code nowhere
- Small per-chunk model cost makes broad scans practical
- Works as a coding-agent skill and as PR CI
Caveats
- Scores do not prove the presence or absence of bugs
- Recall and latency need measuring per repository
- Code is sent to the provider unless SemIf is used
Getting started
`npx skills add Kelbie/hunch` for agents, or `npx -y --min-release-age=0 @kelbie/hunch auth login` then `@kelbie/hunch find "add cancellation to file uploads"`.At a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- Node CLI, Agent Skill, GitHub Action
- Has code
- yes
Use cases and integrations
How curated lists describe it
Code review: plain-English rules that Jev checks code against, locally or on every pull request, with Jev picking one label per finding.
yibie/awesome-jev
Frequently asked questions
What does Hunch decide?
Whether each code chunk matches a plain-English question or review rule, with a score used for ranking or flagging.
When should I use rg instead?
For exact symbols and text; Hunch is for behaviors that may have many implementations or names.
Can it run without sending code to a third party?
Yes, --provider semif uses an open model on your own machine.
Related projects
Codex skill and MCP for advisory Jev decisions in Java/Maven backend work: file relevance, change risk, SQL risk and more.
MCP serverbetaTypeScriptMCP server that finds relevant code with Jev Choice and Noul checks instead of grep sweeps, and audits diffs against task intent.
MCP serverprototypePythonCLI and agent commands auditing a git diff against YAML standards packs, scoring each hunk with Jev so agents fix only flags.
Agent skillbetaTypeScriptLocal MCP review engine where each rule becomes a Jev YES/NO/N/A Choice over agent-supplied diff chunks, with located evidence.
MCP serverbetaTypeScriptZero-dependency Python MCP server where Jev gates shell commands, audits patches and resolves decisions, with SQLite caching.
MCP serverbetaPythonSelf-hosted Jev server plus CLIs, agent skills and a Pi tool for typed code questions and per-file diff review triage.
Agent skillbetaPython