Jev for Developers: Verdicts in Your CI, PRs, and Agent Loops

Developers were the first audience Jev found, but most of the early builds pointed it outward: at emails, games, browsers. This page points it back at the workflow itself, because a developer's day is secretly a queue of small closed-set judgments (merge or not, flaky or real, urgent or backlog, mine or someone else's) and closed-set judgments at volume are the exact shape this model eats. The receipts below come from the directory; the pattern generalizes further than any single one.

PR and diff verdicts

The canonical entry: paste a PR link, get a structured merge-readiness verdict. Not "AI code review" in the essay-writing sense (that's frontier-model territory and honestly still human territory), but the checkable layer: Does the diff touch files outside the stated scope? Are tests added or modified alongside logic changes? Does the description match what the diff does? Is anything in the diff a secret, a debug print, a commented-out block? Each is a judge question with a yes/no answer, and a stack of them per PR costs effectively nothing at reported verdict prices. The output isn't a replacement reviewer; it's a pre-reviewer that makes the human review start at the interesting part.

CI gates that judge, not just assert

Test suites assert; they're bad at judging. A decision layer in CI covers the mushy middle: is this new error message actually user-comprehensible, does this changelog entry describe a breaking change, does this API response sample still match the documented shape semantically. These live beautifully as eval-style checks on every commit, because per-commit is affordable now, and per-commit is where regressions get caught cheap.

Test-failure and log triage

The 2 a.m. skill nobody lists on their resume: staring at a red build and ruling flaky-or-real. It decomposes: Does the failure reproduce a known flaky signature? Is the failing assertion in code touched by this diff? Is the error network-shaped or logic-shaped? The same machine runs incident log triage (which of these thousand error lines are the same error; which ones are new since the last deploy), which is classification wearing a pager. Cascade the ambiguous slice to a human per the routing pattern; route the obvious 80 percent automatically and reclaim the night.

Inside coding agents: the compaction trick

The most technically interesting dev-tool build in the catalog: instant context compaction for coding agents, using Jev to rule, at speed, on what in a bloated context window is still load-bearing (build, behavior as reported by the author). Compaction is normally a slow, expensive summarization pause; as a stream of keep/drop verdicts it becomes nearly free and nearly instant. The general lesson is bigger than the trick: agent housekeeping is decision-shaped, and the agent-verification pattern (judge the claim before trusting the step) plus cheap keep/drop/route verdicts is what makes long-running coding agents behave like tools instead of toddlers.

Where it doesn't belong

Writing code, explaining code, architectural judgment: generation and deep reasoning, not verdicts; see the limitations page before wiring anything. And keep verdicts advisory or gate-with-override in anything developer-facing: a bot that blocks merges on a nondeterministic ruling will be dead by Friday, ceremonially. Confidence thresholds and human override aren't caution theater; they're why these systems survive their first false positive.

Frequently asked questions

Can Jev review my code?

It can judge checkable properties of a diff (scope, tests, secrets, description accuracy) as structured verdicts; it cannot write review prose or reason about architecture. Pre-reviewer, not reviewer.

How do I add decision checks to CI?

Same as any check: a script that sends the artifact plus your questions and fails or warns on verdicts. Keep questions operational, version them like code, and start advisory before you gate.

Is it reliable enough to gate merges?

Gate with confidence thresholds and a human override lane, per the cascade pattern; never hard-block on a single low-confidence verdict. Teams reporting success run advisory for a week, measure, then promote the stable checks.

What's the best first developer use?

Whichever queue annoys you most at volume: PR pre-review, flaky-test triage, or log dedup. A few hundred historical cases scored against your own judgment, per the getting-started ritual, tells you in an afternoon.