Five hot repos reviewed by where Jev sits in the loop
Instead of star rankings, the author reviews five repos by where Jev returns Choice/Score: compaction, skills, a desktop agent, computer-use, and code review.
Jev is a week old and already has a junk drawer of “hot repos.” Most lists just retype the READMEs. A better cut: ignore star rank and ask where the model actually sits in the loop. Jev does not write. It returns Choice / Score / Noul. These five repos are interesting because each one parks that judgment in a different layer. 1. Context, not summary tamaratran/fast-jev-compaction Claude Code’s default compact rewrites history. This plugin does not. It scores each tool call and result, drops or truncates the stale ones, and leaves user/assistant text verbatim. Created 17 Sep. About 6.3k stars in five days. The idea is small and sharp: keep the evidence, delete the noise. 2. Official primitives, not a sleeper hit typesafe-ai/skills This is TypeSafe’s own agent skill, not a community breakout. ~1.8k stars. It belongs on the list only as plumbing: it teaches an agent when a typed judgment should replace another prompt-and-parse step. Treat it as docs you can install, not as a product. 3–4. Two desktop agents, two sensors lahfir/agent-desktop ~1.4k stars Rust CLI. It drives macOS through the accessibility tree: stable refs, retryable actions, no screenshot-as-truth. Jev is not the core. jev-desktop landed in v0.9.2 on 17 Sep. The repo is a computer-use substrate that Jev can sit on. awlevin/typesafe-computer-use ~800 stars, created 16 Sep The opposite bet. OCR the screen, ask Jev for the next action, click. Author’s own numbers: ~$0.0002 and 0.13–0.38s per decision versus sending the screenshot to Opus. Cheap loop, weaker structure. Same noun (“computer use”), different claim. One trusts the OS tree. One trusts a 20¢-per-thousand-steps classifier. 5. Review as a pipeline, not a paragraph devagrawal09/jev-review Name collision warning: several repos are called jev-review. This is the staged one. ~500+ stars, created 16 Sep. Code owns the workflow. Jev only answers bounded questions: risk matrix, file profile, evidence, mechanism, severity, routing. Output is a local dashboard on 127.0.0.1, not a review essay. What the five make together If you stop treating them as a leaderboard: skills defines the primitive → compaction decides what the agent is allowed to remember → the two desktop tools close the act loop on a Mac, one via AX and one via OCR → jev-review gates a diff with typed scores instead of asking a chat model to “look at the PR.” Deliberately missing: browser-use/jev-ultrafast. It is the loudest Jev repo this week. It does not need another list. If you only clone one: compaction, if you live in long Claude Code sessions. If you only read one contrast: the two Mac tools. Stars will move tomorrow. That split will not.