jev-debug
Locates the file a bug report is about, or flags defects in a diff, by asking Jev yes/no over small source batches.
Overview
jev-debug finds which file a bug report refers to and reviews diffs using Jev. Since Jev answers questions about text but does not search a repository, the package does the search with a fixed algorithm and asks Jev only about small batches. It shortlists files by identifier overlap with no model call, packs them into windows of at most three files and 60k characters in two different groupings, and asks whether the code that must change is in each batch.
Each file is scored by how well 'the bug is in this file' predicts all the answers, then near-best files are asked about individually so innocent neighbours drop out. Code review uses diff hunks as batches and asks whether a reviewer should flag a logic, contract or security defect. A pilot on twelve SWE-bench Lite instances reports the shortlist alone placed the gold file in the top 10 for 11 of 12 cases.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is the code that must change in this batch? | yes · no |
| Yes / no | Should a reviewer flag a defect in this hunk? | yes · no |
How it works
- Shortlist files by overlap with identifiers in the bug report
- Batch shortlisted source in two groupings and ask Jev per batch
- Score each file by log-likelihood across answers
- Re-ask near-best files individually to finalise
Strengths
- Cheap localisation by limiting model calls to small batches
- Cross-checking in two groupings reduces false positives
- Same method reviews diffs
Caveats
- Pilot covers only twelve SWE-bench Lite instances
- Identifies files, not functions or fixes
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Locate a bug file, or flag a diff, by cross-checking small source batches with JEV.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Whether a batch of source contains the code that must change, or whether a diff hunk has a defect.
Does Jev search the repo?
No. A fixed algorithm shortlists and batches files; Jev only judges batches.
How well does it work?
In a 12-case SWE-bench Lite pilot the no-model shortlist alone had the gold file in the top 10 for 11 cases.
Related projects
Codex skill and MCP for advisory Jev decisions in Java/Maven backend work: file relevance, change risk, SQL risk and more.
MCP serverbetaTypeScriptClaude Code plugin that adds Jev as an independent judge for review findings, debugging hypotheses, design options and search hits.
PluginbetaSemantic code search and plain-English review rules that send repository chunks to Jev and return scored source locations.
ProjectbetaTypeScriptSix evidence-oriented Jev tools for pi, omp and any MCP client: ask about code, find files, locate ranges, check diffs and select tests.
ServicebetaTypeScriptSemantic grep for code, diffs and CSV rows that packs 16 chunks per Jev request and prints file:line hits with CI-friendly exit codes.
Agent skillbetaTypeScriptMCP server that finds relevant code with Jev Choice and Noul checks instead of grep sweeps, and audits diffs against task intent.
MCP serverprototypePython