jev
Reviewed
at commit 4115077
Caution: Pre-approves running the unpinned npm package jevctl through npx in its commands and skill, so each new version runs without asking.
Sends data to TypeSafe Jev (api.typesafe.ai)
Cheap, fast, typed AI judgments inside Claude Code via the jevctl CLI (verify, screen, find, classify, extract, rerank, match, route, ask, batch), plus…
0 upvotes
Log in to upvote or save- Made by
- Nasr Shaer
- Pinned commit
-
41150770a59f1fe596a5afe7641c31702879d344 - What we checked
-
Replaces Claude Code's compaction summary: it sends your conversation (prompts, replies and tool inputs, not tool output) to TypeSafe's Jev service, which decides which old tool results to cut, and starts compaction at 60% context. Slash commands and a skill have Claude run the jev CLI (npx jevctl) to send text, files or images and screenshots to TypeSafe for checks, including a new /locate command. It stores nothing.
Review details
Static review of the source at the pinned commit; not run. Delta review of the changes since 142f3cf; the rest carries over from the review at that commit. hooks.json loads one module, fast-jev.ts, with vendored code in compaction/ (compact.ts, request.ts, state.ts, types.ts); the module and compaction code did not change. Changes in 0.3.0: a new command commands/locate.md (allowed-tools Bash(jev:*), Bash(npx jevctl:*), Bash(sips:*), Read, Glob) that has Claude run jev locate on an image path to find where something is in it, with a macOS sips resize or conversion into a temp copy if the image is rejected; the jev skill documents image inputs (locate, ask/classify/screen --image, --box questions), a long-context flag and a note that screenshots can hold personal data and should be cropped before sending; the plugin.json description and the model example (jev-1.14.0) were updated. Images go to TypeSafe through the jevctl CLI, a separate npm package that was not reviewed. Carried over: session.compact (enabled by default via userConfig compaction=true) builds a state from the whole transcript (message text abridged in stages, tool inputs truncated to 1000/200/60 chars, tool results replaced by a length note, the last three user prompts as goal) and POSTs it with yes/no questions per old tool call to https://api.typesafe.ai/v1/systemone with a Bearer TYPESAFE_API_KEY via $.http.fetch, at most 4 requests in flight; it then drops or truncates stale tool calls and results, keeping the first message and newest 6 untouched, and returns the edited messages instead of a summary. On a missing key, an API error, or a reduction below 25% it falls back to the built-in summary. turn.complete reads $.session.usage() and calls $.session.compact() once context reaches compactAtPercent (default 60; 0 disables). The key comes from the plugin's sensitive userConfig apiKey, $.env.get('TYPESAFE_API_KEY') or the env block from $.settings.read(). Eight earlier commands (ask, classify, compact, extract, find, route, screen, verify) and the jev skill pre-approve Bash(jev:*) and Bash(npx jevctl:*) (no version pin), plus Bash(curl:*) and WebFetch in /screen, Write in /route and Bash(which:*) in the skill; they tell Claude to write content to temp files and pass it to the jev CLI. /compact reads the newest session transcript under ~/.claude/projects. The skill suggests npm install -g jevctl and warns not to send secrets. plugin/hooks/types/claude-code.d.ts is a type declaration file, truncated in the bundle and not read; it is not executed.
- What it uses
-
- Reads your prompts: yes
- Sees tool calls: yes
- Blocks or changes tool calls: no
- Reads or writes files: no
- Runs programs on your machine: no
- Calls the model (uses tokens): no
- Sends data over the network: yes
- Stores data: no
- Shows something on screen: yes
- Hooked events
Cheap, fast, typed AI judgments inside Claude Code via the jevctl CLI (verify, screen, find, classify, extract, rerank, match, route, ask, batch), plus verbatim Jev-guided context compaction that replaces the built-in summary.
Made this mod? Add a badge to your README
Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.
[](https://modsforclaude.com/mods/jev/)