JevGate turns one function at a time into a finding at 0.80

This topic was created 31 days ago, and the information it contains may have evolved or changed since then.

Nodehaven Online posted jevgate, a Rust review gate on crates.io as v0.5.1. Tree-sitter finds a function, a copy, or a test locally. Jev answers a short question. Code maps the probability to review, consider, note, clear, or uncertain at 0.80. The README says those probabilities are judgments, not a measured accuracy.

Nodehaven Online posted on September 23 that a review tool from Tech Byte Frontier was open sourced: cargo install jevgate --locked, with the crate at crates.io/crates/jevgate. The crate page lists v0.5.1, published the same day, owner Tauan BF, repository Tech-Byte-Frontier/jevgate. Minimum Rust is 1.90.0. The README license is MIT OR Apache-2.0.

JevGate parses the repository locally with tree-sitter. It builds a small unit, a function, a file outline, a renamed pair, or a test, and asks Jev a short question about that unit. Code turns the answers into a finding with a location, a probability, and a next step. Unchanged units are answered from a local cache. A follow-up request is sent only when an answer says the function should be split, and it asks once more with extra evidence such as callers.

The composition threshold is 0.80. An answer becomes review, consider, note, clear, or uncertain. Raw probabilities stay in the JSON report. An undecided answer is reported as uncertain. Maintainability is on by default: file organization, function simplification, shared logic, and hardcoded values. Tests are included with --include-tests. Security and documentation are opt-in. Security findings name a CWE and cover injection, sensitive data, and unsafe settings. The README limits that pass to one function plus at most one hop of callers. It does not cover SQL files, row-level policies, or access control. Secrets are out of scope, because judging them would mean uploading them.

Documentation rules read AGENTS.md, CLAUDE.md, GEMINI.md, and the rule files for Claude, Cursor, Copilot, Windsurf, and Cline, including hidden and gitignored copies. A Markdown file of 300 lines or more is judged from its headings. Staleness uses paths, scripts, tags, and deletions that Git and the manifests show. The README says paraphrases that share little wording are missed, and that token counts are estimates at four bytes per token.

upload_allow and upload_deny bound what can leave the machine. jevgate check --dry-run --show-requests prints the bodies with no network. --fail-on chooses which level fails the gate. Notes never fail it. Exit 0 is a pass, 1 is a fail, and 2 is an incomplete run or an error. jevgate baseline records the current findings so a later check fails only on new ones. Answers from a pinned model version do not expire.

The README’s sample report line is 42 files, 118 API requests, 263,410 input tokens, about $0.0111, with two consider findings. That block is the printed example of the output. Languages are Rust, Python, JavaScript, and TypeScript. Other files are listed as skipped. The limits section says the probabilities are model judgments, not a measured accuracy, and that the tool does not replace linters, type checkers, tests, or a security scanner. We did not run jevgate check.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

Primary post is @tbytefrontier, display name Nodehaven Online, September 23, 2026, 06:51 UTC. It says cargo install jevgate --locked and links crates.io/crates/jevgate. crates.io lists v0.5.1, published 2026-09-23, owner Tauan BF, repository Tech-Byte-Frontier/jevgate, minimum Rust 1.90.0.

README license MIT OR Apache-2.0. Composition threshold 0.80, labels review, consider, note, clear, uncertain. Maintainability is on by default. Tests need --include-tests. Security and documentation are opt-in.

Languages: Rust, Python, JavaScript, TypeScript. Security looks at one function plus at most one hop of callers and names a CWE. The sample report line in the README is 42 files, 118 API requests, 263,410 input tokens, about $0.0111.

Exit codes 0, 1, and 2. The README says the probabilities are model judgments, not measured accuracy. We did not run a check.

Compare

DocJev classifies and splits PDFs and prints 40/40 and 7/8 against Luna. JevGate reviews source and prints no agreement with a human reviewer. jev-gateway's chess bench is a tool-routing measurement, not a finding-level accuracy.

The 0.80 cut is this tool's own composition rule.

Terms

0.80
The probability at which JevGate maps an answer to review, consider, note, clear, or uncertain. The JSON report keeps the raw probability. An undecided answer is reported as uncertain.
evidence unit
The slice asked about in one request: a function, a file outline, a pair of copies, or a test. Tree-sitter finds it on the machine. The upload is that slice, bounded by upload_allow and upload_deny.

Sources

  1. Nodehaven Online, cargo install jevgate
  2. crates.io, jevgate 0.5.1
  3. Tech-Byte-Frontier/jevgate