jev-crawlers
Unix-style tools where Jev learns a repo's decision norms and judges each node of a code lead as a bug, lead or clean.
Overview
jev-crawlers is a set of small command-line tools (seed, expand, judge, verify, report, norms) that pass JSON lines through pipes to hunt for bugs. Instead of reading a diff once, they follow a lead: who else calls this, what config changes it, where input becomes trusted. Jev, reached via the Vercel AI Gateway, gives a typed verdict on every node, so each step of the crawl gets judged rather than just each scan. A norms tool learns the repository's decision norms so past decisions can be adversarially judged against them.
It targets teams whose AI-written code outpaces review. The README lists claims each backed by mocked tests (46 tests across 9 files): credentials are redacted before state is built, excerpts are windowed around the hit line, fabricated or missing citations are demoted to unverified leads, and vendor and secret files are skipped.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is this node a bug, a lead worth expanding, or clean?(inferred) | — |
| Yes / no | Does this past decision violate the repository's learned norms?(inferred) | — |
How it works
- seed finds starting points in the repo
- expand follows callers, configs and co-committed files
- judge asks Jev for a typed verdict per node with redacted state
- verify checks cited lines exist and demotes fabricated ones; report summarises
Strengths
- Judges every node cheaply rather than one pass over a diff
- Credential redaction before any state leaves the machine
- Citations are verified so hallucinated lines cannot become bugs
Caveats
- Calls Jev through the Vercel AI Gateway, which needs its own key
- Cost claims come from the author's calibration runs
Getting started
Run `npm test` to execute the mocked test suite; tools are composed with pipes as JSON lines on stdin and stdout.At a glance
- Type
- Project
- Status
- beta
- Language
- JavaScript
- Runtime
- Node CLI tools
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Jev learns your repo's decision norms, then adversarially judges past decisions against them. Unix-style primitives (seed, expand, judge, verify, report, norms) with per-node typed judgments from typesafe-ai/jev.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev judge?
Each node a crawler reaches, giving a typed verdict, and past decisions against learned repo norms.
How does it avoid false bug reports?
A verify step checks cited lines and demotes fabricated or missing citations to unverified leads.
Does it send secrets?
Credentials are redacted before the judge state is built, as covered by tests.
Related projects
CLI that scans source, build and CI files with Jev for hidden, deceptive or data-stealing behavior before you run unfamiliar code.
CLIbetaTypeScriptRust code and security reviewer that asks Jev staged questions about diffs and posts findings with severity and owner routing.
ProjectbetaRustSemantic code linter that reads each method with its callers and callees and asks decision models about defects, kind, line and severity.
CLIbetaTypeScriptStaged code-review workflow for JS/TS where Jev screens risk, picks evidence, classifies mechanism, scores severity and routes reviewers.
AppprototypeTypeScriptProof-of-concept scanner that asks Jev hundreds of bug questions per Solidity file and writes a heat map of likely vulnerabilities.
ProjectprototypePython- jev-harness★ 30
Research proposal-review contract: an LLM proposes one action, Jev answers four narrow questions, code produces evidence, applies nothing.
LibraryresearchTypeScript