CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

jev-gate

Review a diff with Jev, score a UI Then, or give Jev a browser goal. Trigger: "review this", "open this page", "click through with Jev", UI scenario Then.

Quality

73%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./ywai/plugins/jev-gate/skills/jev-gate/SKILL.md
SKILL.md
Quality
Evals
Security

Jev review

Jev is not a chat model. It answers questions with probabilities and scores, and this plugin turns those numbers into findings. Your job is to relay them, not to extend them.

Running a review

  • Working-tree diff: call jev_review_diff.
  • One file or folder: call jev_review_path with a path.

Both are Code Mode tools, so they are called inside execute:

return await tools.jev_review_diff({})

The tool returns a short markdown summary: the verdict and whatever became a finding. It does not show the scores behind that verdict.

To see how Jev categorised each file, call jev_report. It prints the screen matrix - every screened file against all five dimensions with its probability - plus which signals opened a finding and which could not be placed. With no argument it reports this session's last review; pass a runId for an earlier one. Reach for it whenever the verdict alone does not answer the question, and always after an INCONCLUSIVE run, where the scores are the only record of what Jev saw.

Two dimensions in that table never open a finding on their own: compatibility and testGap. A high number there is noise - see "What Jev is bad at" below.

Rules when reporting

  1. Only Jev's findings are Jev's. If you spot something Jev did not mark, say it is your own observation, in a separate sentence.
  2. Do not upgrade confidence. A 0.74 is "likely", not "definitely". Keep the probability and severity attached to the finding.
  3. clean is not approved. Jev never approves. A clean run means no signal crossed the threshold - it is not a guarantee the change is correct.
  4. INCONCLUSIVE is not clean. It means a signal crossed the threshold and the pipeline could not place it on a changed line. Report it as a review that did not finish, never as a run with nothing to say.
  5. If the tool says the review did not run, it did not run. No key, a network failure, or an empty diff are not passing reviews. Say so plainly.
  6. Do not re-run a review to get a different answer. If a finding looks wrong, say why in your own words and leave Jev's number as it is.

What Jev is bad at

Measured on 10 labeled fixtures, not guessed:

  • It catches 6 of 8 fixtures with a real defect. It is not a safety net: a clean Jev run means 6-in-8 odds on this fixture set, nothing stronger.
  • compatibility scores high on almost everything, including clean renames, and testGap fires where nobody expects it while missing the one real test gap. Both are screened and reported but never opened as findings.
  • severity saturates near 3 - treat the gap between two findings' severities as weak evidence.
  • owner skews to security. It is a hint, not an assignment.
  • It did NOT fall for a comment-only change that says "auth token" twice, and a comment telling the reviewer not to flag a removed check did not suppress the finding.

Routing

jev_route asks Jev which agent should execute a task, over the agents this install actually has (orchestrator, dev, planning, reviewer, finder, ask, ...), plus inline, review and human.

  • Jev recommends; switching agent stays the caller's decision.
  • A closeCall means confirm with the user before acting.
  • An answer that names nothing we recognise becomes human, never a guess.

The gate

After a review with blockers, or a route to an agent that does not write, the next write asks for confirmation. The gate can only harden (allow -> ask): it never turns an ask into an allow, never touches a deny, and ignores any decision whose source is not jev.

If a write suddenly asks for confirmation, the reason is in the prompt. Do not work around it by using a shell command instead - bash is gated too.

Browser Then

jev_check_page scores a Gherkin Then against an accessibility snapshot. Playwright or chrome-devtools describes the page; Jev answers with a probability. Jev never clicks.

return await tools.jev_check_page({
  snapshot: ariaTree,
  then: "the New workflow action is available",
})

The tool returns PASS or FAIL with the probability. Relay that number. A missing key or a failed call is not PASS.

Use this from @scenario-runner after Given/When, never as a substitute for driving the browser, and never as a code review.

Finding code

jev_find is semantic grep: it asks whether a chunk of code does the thing, not whether it mentions it. It costs a Jev request per batch of segments, so reach for grep first and use this when the words in the code are not the words in the question.

It reads the working tree from disk, so uncommitted changes are searched like anything else. There is no index and nothing is read from git.

It searches at most 200 segments in directory order. On a repo bigger than that it covers only the part it reached, and the result says so: "Search incomplete" means the rest was never sent to Jev, not that nothing is there. Narrow it with root and run it again instead of concluding from a partial pass.

A result of no hits is not proof the code does not exist - say "nothing scored over the threshold", not "it is not there". If you cannot explain a miss, say you cannot explain it; do not invent a mechanism for it.

Repository

YoizenSA/dev-ai-workflow

Path

ywai/plugins/jev-gate/skills/jev-gate/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.