jev-codex-plugin and jev-rust-review: More Editor Integrations

The Jev code-review pattern — a fast typed classifier triaging findings before a language model spends tokens reasoning about them — is already covered at a general level in Jev review, code review triage. Two more projects extend the idea into specific tools: jev-codex-plugin wires Jev into Codex for general decision consultation and failure diagnosis, and jev-rust-review uses it as a semantic triage layer specifically for Rust, sitting between the compiler's own tools and Claude's reasoning.

Both are small — 3 and 2 stars respectively as of September 2026 — and both describe themselves candidly as unverified early software. That's worth reading in full rather than skimming past, because the honesty in these two READMEs is a useful model for how a tiny project should represent itself.

Key Takeaways

• jev-codex-plugin covers three jobs: decision classification across a taxonomy, failure diagnosis from command output, and evidence-based completion review — it explicitly recommends, and does not execute or grant permissions.
• jev-rust-review is a three-stage pipeline: rustc/Clippy/cargo-semver-checks first as deterministic fact, then Jev triage, then Claude's reasoning on the flagged subset.
• jev-rust-review's Jev layer answers 16 typed dimensions — correctness, async patterns, concurrency, ownership, unsafe code, and more — not free-text review comments.
• Both projects disclose their own limits plainly. jev-codex-plugin: automation is off by default. jev-rust-review: its evaluation is a 33-fixture synthetic corpus, not production code, and it can't see across files.
• Neither is an official TypeSafe or OpenAI product. jev-codex-plugin is built by an independent author (Romy Mondello, per its README) and installs via the Codex plugin marketplace.

jev-codex-plugin: consultation, not automation

jev-codex-plugin (MIT, 3 stars as of September 2026) packages three Jev-backed capabilities for Codex workflows. Decision classification lets a workflow compare real candidates against a taxonomy — the README's own framing is comparing "actual available candidates across eight domains or any custom taxonomy." Failure diagnosis takes selected evidence from a broken command and asks Jev to classify what went wrong. Completion review checks whether the evidence a task points to actually supports the claim that the task is done — a useful counter to the common failure mode where an agent reports success without confirming it.

The README is explicit about the boundary: the plugin "does not execute a decision, grant permissions, switch models automatically, or prove deployment." It recommends; the workflow author decides. That's the same code-owns-the-policy pattern seen throughout the guardrails article, applied here to Codex specifically rather than Claude Code. Setup needs Node.js 22+ and a TypeSafe API key, installed through the Codex plugin marketplace in two steps with no local npm install required. Automation is disabled by default on a fresh install, and hook coverage depends on what the host system exposes — a caveat worth checking before assuming full parity across different Codex environments. On aiskill.market: /skills/jev-codex-plugin-integrate-your-mind.

jev-rust-review: a narrow, disclosed pipeline

jev-rust-review (MIT, 2 stars as of September 2026) is a Claude Code plugin aimed at one language, and it earns that narrowness by going deep. The pipeline runs in three stages. First, standard Rust tooling — rustc, Clippy, and optionally cargo-semver-checks — produces deterministic findings, reported as facts rather than judgments. Second, Jev triages and verifies candidate findings by answering targeted yes/no and severity questions across 16 dimensions: correctness, error handling, async patterns, concurrency, ownership, unsafe code, type design, API compatibility, macros, serialization, idiom conformance, performance, security, and testing, among others. Third, Claude reviews only the code Jev flagged and writes the final findings with confidence assessments.

The value proposition is specific: the README targets bugs "no tool reports," giving the example of a select! branch that silently loses data when cancelled, or a check-then-act race across separate locks. That's a credible niche — Rust's own tooling is excellent at what it checks, and async/concurrency bugs are exactly the class of thing that slips past a compiler and a linter but shows up in a semantic read of the code.

The limitations section is unusually specific, which is a point in its favor. The tool can't reason across multiple files at once, so cross-file dependencies return uncertain verdicts rather than false confidence. It knows documented behavior for only 15 library facts, covering tokio, axum, serde and Dioxus. The README also notes that smaller models follow Jev's guidance more closely than larger ones — an interesting and slightly counterintuitive finding worth further testing rather than taking as settled. Most importantly: its evaluation results come from a 33-fixture synthetic corpus, not production code, so treat any accuracy claims as a starting benchmark, not a guarantee. It requires Claude Code 2.1+ on Linux or macOS, a TypeSafe API key, and Git and Cargo with Clippy installed. On aiskill.market: /skills/jev-rust-review-kindintelligence.

What to do next

1. Use jev-codex-plugin for general workflow decisions in Codex — failure triage and completion checks — not as a substitute for permissioning or deployment verification.
2. Reach for jev-rust-review only on async- or concurrency-heavy Rust code, where its 16-dimension triage targets bugs standard tooling misses; for straightforward code, cargo clippy alone still covers most ground.
3. Read both limitations sections before trusting either tool's output on real work. A 33-fixture synthetic corpus and a cross-file blind spot are the kind of detail that should shape how much weight you give a flagged finding.
4. Compare with the general code-review pattern in Jev review, code review triage if you need something language-agnostic instead.
5. Check the radar periodically. Small, single-purpose plugins like these appear weekly on the Awesome Jev projects list; see how to evaluate Jev projects honestly before adopting the next one.