jev-lens

You ask Claude Code for a feature. It edits eight files and says done. Now what? You either read every line, or you skim and hope.

jev-lens answers the question you actually have: do I need to look at this, and where? When a session stops, it diffs what the agent did against the last thing you reviewed, asks TypeSafe Jev a handful of yes/no questions, and writes a verdict. Something like:

• Nothing needs you. Skip it.
• src/auth/session.ts changes behavior, have a look.
• src/app.ts has four leftover comments and a console.log. Want them stripped?

It never blocks Claude, never edits your files, and never says green unless it is sure. Pair it with jev-lens.nvim to see the verdict as a popup in Neovim.

Install

One line, from a terminal:

claude plugin marketplace add rashedInt32/jev-lens && claude plugin install jev-lens@jev-lens

Or from inside Claude Code:

/plugin marketplace add rashedInt32/jev-lens
/plugin install jev-lens@jev-lens

Then give it a TypeSafe key from console.typesafe.ai. Hooks do not see your shell profile, so put it in a file:

mkdir -p ~/.config/typesafe && printf '%s' "ts_..." > ~/.config/typesafe/key && chmod 600 ~/.config/typesafe/key

Restart Claude Code. That is it. Needs Node 20.12 or newer.

How it works

Four hooks, none of them slow, none of them talk to the API:

• SessionStart takes a snapshot of your repo. That is the baseline, the last state you are assumed to have reviewed.
• UserPromptSubmit remembers what you asked for.
• PostToolUse remembers which files the agent touched, including through shell commands.
• Stop kicks off the judge in the background and returns at once.

The judge diffs the baseline against now, drops lockfiles and secrets, and asks Jev per file: does this need a look, what kind of look, and which of your prompts asked for it. Comments, debug prints, TODOs, lint suppressions and any casts get one more question each: would a careful developer delete this? Everything comes back as a number, and thresholds decide what you see.

If jev-gates is installed, its proof gate leaves a result for the same repo at the same stop: which changes could alter behaviour and had nothing run, tested, or checked after the edit. The judge waits a moment for that file and lists those changes in the verdict as unverified. The lens then shows not just which files to look at, but which specific changes nobody has exercised yet. JEV_LENS_PROOF=0 turns the lookup off.

The result lands in ~/.claude/jev-lens/repos/<repo>/verdict.json. Anything can read it. jev-lens.nvim is the first thing that does.

When it stays quiet

Most stops should end in silence or one line, not a popup. So before spending a request, the judge skips:

• formatter runs, where nothing changed once whitespace is ignored
• diffs that only deleted comment lines
• diffs with no agent edit on record, which means it was your own typing

Diffs of three lines or fewer get judged for the log but are marked as not worth a popup. Generated output like coverage and .turbo is ignored along with lockfiles.

When the baseline moves

"Since you last reviewed" is the unit of review, however many stops it took. The baseline moves when you mark a verdict reviewed in nvim, or when a commit leaves the working tree clean. A commit is a review.

Trust it slowly

Start in shadow mode for a week. It judges and logs but nvim stays silent. Compare ~/.claude/jev-lens/decisions.log with what you would have said.

{ "env": { "JEV_LENS": "shadow" } }

Flip to active when the numbers agree with you.

Settings

All environment variables, all optional:

Var	Default	What it does
JEV_LENS	active	off, shadow, or active
JEV_LENS_GREEN	0.9	how sure "nothing needs you" has to be
JEV_LENS_UNSURE	0.6	below this the overall verdict is "look"
JEV_LENS_FILE_THRESHOLD	0.6	a file at or above this is flagged
JEV_LENS_KIND_CONFIDENCE	0.7	below this the category shows as a guess
JEV_LENS_DEBRIS_THRESHOLD	0.7	at or above this a comment or log is strippable
JEV_LENS_TINY_LINES	3	at or under this many changed lines, never a popup
JEV_LENS_REQUIRE_EDITS	1	0 judges diffs even with no agent edit on record
JEV_LENS_MAX_FILES	40	files judged per verdict; the rest are listed as not judged
JEV_LENS_MAX_CHARS	40000	state cap per request
JEV_LENS_IGNORE		colon-separated globs added to the ignore list
JEV_LENS_DIR	~/.claude/jev-lens	where verdicts live
JEV_LENS_MODEL	jev-latest

What leaves your machine

The diff, your recent prompts, and the rules from your CLAUDE.md. Lockfiles, build output, .env files, keys and credential stores are never sent, and anything key-shaped inside a patch is redacted first.

Judge by hand

node bin/judge.mjs --cwd .                   # judge now
node bin/judge.mjs --cwd . --reviewed        # move the baseline to now
JEV_LENS_VERBOSE=1 node bin/judge.mjs --cwd . # print the result

Tests

npm test

Runs against a local stand-in API, no key needed.

Related

• jev-lens.nvim shows the verdict in Neovim.
• jev-gates stops bad writes before they land. jev-lens ranks what did land. They work alone; together, the gates' proof result shows up in the lens as unverified changes.
• docs/SPEC.md has every design decision and the file formats.

License

MIT