A side panel for Claude Code and Codex CLI that shows the blind spots of each prompt you submit: what the request would have needed to consider and shows no sign of. Jev decides, in one call per prompt, whether the prompt is worth a second look; only then does the agent's own headless mode read the project and return the items. Nothing blocks the prompt, nothing is added to the session.
One gate call evaluates worth_checking, risk and possible gaps; local thresholds and duplicate suppression decide whether to start headless analysis.
Keeps the quick prompt gate separate from the deeper repository-reading agent and exposes why the panel remained quiet.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No hooks, trust settings or headless agents were installed. Advice still depends on what the gate state and analysis agent can inspect.
Decide when a submitted coding prompt merits a second, read-only inspection.
Material analysed：
thresholds.ts chooses a different worth threshold for high-risk prompts and suppresses duplicates. The panel is advisory; its scores do not grant the coding agent additional authority.
Inspect the supporting source ↗client.ts uses a hard timeout with SDK retries disabled and returns explicit no-key, timeout, HTTP or parse failures. Fake mode is a stub, not a live evaluation.
Inspect the supporting source ↗Useful as a prompt-review side panel when deeper analysis is costly. It should not be installed or trusted as an approval gate for destructive commands.
Read gate/client.ts and thresholds.ts together. Follow a duplicated prompt to the quiet branch, then compare a low-worth result with the analyze branch. JEV_FAKE mode can check wiring only, not actual Jev quality.