jev-code and jev-dsh-decision: Decision Layers for Coding Agents

Name collisions are common enough in this ecosystem that wiring Jev into agents with MCP had to flag four different jev-mcp repos. This one is smaller but sharper: two entirely unrelated projects both live at a path called jev-code, and neither is aware of the other. Alongside them, jev-dsh-decision adds Jev-based tool and skill recommendation to a third kind of agent harness. None of the three should be confused for one another.

Key Takeaways

• devagrawal09/jev-code is actually "Stanley." The tool identifies itself in its own README as Stanley, a CLI that routes natural-language requests to ten built-in code-review workflows — the repo and future npm package are named jev-code/jev-code, but the product name is different.
• Stanley's npm package is a placeholder today. jev-code@0.0.1 on npm is explicitly a placeholder per the README; the real install path (npm install --global stanley-code) only activates after a proper release, and the current version is unreleased 0.1.0.
• rhighs/jev-code is a completely different, smaller project. An experimental CLI where Jev selects AST productions for Python or Bash generation, plus a separate command-line decision mode — 20 stars, no stated license, much thinner evidence.
• Stanley keeps decisions in code, not in the model's free text. Jev only picks a workflow from author-provided JSON routing metadata; fixed thresholds in Stanley's own code convert Jev's typed answers into findings.
• jev-dsh-decision serves three separate agent harnesses. A native DeepSeek Harness plugin plus an iPolloWork entry point that also serves OpenCode and Codex Harness — evidence here is the thinnest of the three projects.

devagrawal09/jev-code — really "Stanley," a code-review router

devagrawal09/jev-code (MIT, 113 stars as of September 2026) is, per its own README, called Stanley. You describe what you need in natural language, and Stanley routes the request to one of ten built-in deterministic workflows: finding relevant files, checking a change against its stated task, general correctness review, test-gap identification, change summarization, security audits, performance analysis, compatibility checks, triaging failing tests, and sorting review comments. Custom workflows can extend the set via TypeScript modules under .stanley/workflows/.

Jev's role is narrow and specific: it chooses one workflow from JSON routing metadata the workflow authors provide, and separately answers fixed multiple-choice questions about bounded evidence — a diff, a log excerpt, a code snippet. Stanley's own code then applies fixed thresholds to Jev's typed answers to produce findings. The model never writes a free-form review comment or makes the actual pass/fail call; per the README, "Jev chooses one built-in or repository workflow from author-provided JSON routing metadata," and the routing and thresholding stay in Stanley's code.

Two things are worth flagging before you install this expecting a finished tool. First, its npm package name really is jev-code, but jev-code@0.0.1 on npm is explicitly a placeholder — the intended install path, npm install --global stanley-code, only becomes real once the software ships a proper release; today it requires building from source (git clone, npm ci, npm run build) with Node.js 22.18+. Second, the README is direct about scope: Stanley doesn't replace tests, type checkers, linters, security tools or human review, and an empty findings list isn't approval. It also warns that trusted repository workflows run with the user's full privileges, so review any custom .stanley/workflows/ module the same way you'd review a CI script before trusting it. This lines up closely with the pattern covered in Jev code review and triage — advisory findings, fixed thresholds, no autonomous merge authority. On aiskill.market: /skills/jev-code-devagrawal09.

rhighs/jev-code — an unrelated, much smaller AST experiment

rhighs/jev-code (20 stars, no license stated) shares nothing with Stanley beyond a repo name. It's described as an experimental coding CLI where Jev selects among AST (abstract syntax tree) productions for Python or Bash, effectively choosing program structure piece by piece from a bounded set of syntactic options, alongside a separate command-line decision mode where Jev's judgments drive which local tools get invoked.

We reviewed the README and integration code at a fixed commit; this site has not run the project or independently verified its performance, security posture, or any claimed business outcome — there isn't much to verify yet, since the README itself doesn't publish benchmark numbers or a comparison against Stanley or anything else. At 20 stars and days old, this reads as a genuine experiment rather than a production tool, and the two projects sharing a name is coincidence, not lineage. Confirm the owner (rhighs vs. devagrawal09) before you clone either one — the aiskill.market pages are separately slugged for exactly this reason: /skills/jev-code-rhighs versus /skills/jev-code-devagrawal09.

jev-dsh-decision — recommendation across three agent harnesses

jev-dsh-decision (129 stars, no license stated) takes a broader integration approach than either jev-code project. It provides Jev-based structured decision support for what its own description calls "Agent Harness" — recommending tools, skills and agents, and returning judgments with probabilities attached — via a native DeepSeek Harness plugin plus an iPolloWork entry point that also serves OpenCode and Codex Harness.

Evidence here is the thinnest of the three projects in this article. Our review is based on repository metadata and the README, with a rule-based classification pending human review, and no independent runtime or performance verification has been done — the same caveat that applies to most brand-new, low-star projects in this radar. That doesn't make it uninteresting: it's targeting a real gap (recommendation of which tool/skill/agent to invoke next, rather than a single yes/no gate), but with 129 stars and no independent testing yet, treat any specific claim about its recommendation quality as unverified until you've run it against your own harness. On aiskill.market: /skills/jev-dsh-decision-devin-axis.

What to do next

1. Check the GitHub owner before you clone anything called jev-code. devagrawal09/jev-code (Stanley) and rhighs/jev-code are unrelated projects at different maturity levels.
2. Don't npm install jev-code expecting Stanley. The real package name is stanley-code, and it isn't live yet — build from source if you want to try it today.
3. Treat Stanley's findings as advisory, same as any other Jev-gated review tool. It says so itself: no replacement for tests, linters, or a human reviewer.
4. Review custom Stanley workflow modules before running them — they execute with your full local privileges.
5. If you're evaluating jev-dsh-decision for a DeepSeek Harness, OpenCode or Codex Harness setup, run your own verification first; nothing in the public record yet confirms its recommendation quality. See the full field at the Awesome Jev projects radar and read how to evaluate Jev projects honestly before trusting any single README's claims.