Write the menu before Jev steers Claude Code — JEV Brief · Sep 23
Nothing TypeSafe-official moved the contract today. The new desk pattern is still the right shape: a plugin that lets Jev pick reasoning effort, a named subagent, and a named skill for Claude Code. The costly miss is letting Jev invent that menu — or write the prompt. If you only have four minutes, write the closed options, pin jev-1.13.0 on the router, then scan Top Repos.
- Write the menu first. Effort levels, named subagents, and named skills must be a closed list you authored. Jev’s job is Choice over those options. If the plugin can mint a new skill name or a free-form instruction, you have given it the LLM job.
- Jev picks; the LLM drafts. Rank, route, and gate with Jev. Let Claude Code (or any LLM) write the patch, the justification, or the typed string after the Choice lands. Do not ask Jev to compose the prompt it is supposed to score.
- Pin the router. Prefer
jev-1.13.0overjev-latestanywhere the Choice can auto-act. A floating alias on a Claude Code plugin is the same drift as a floating alias on a support triage bot. - Schema-safe ≠ correct. A constrained Choice still picks the wrong subagent; it just cannot invent a fourth one. Keep irreversible tool calls and arithmetic in code, with an abstain / low-confidence branch before the executor runs.
Top Repos today
Live board: https://jev.pro/top-repos. Stars are public rank, not a quality award. Metrics still sit on the 2026-09-21 snapshot (midnight refresh has not moved), so 1-day deltas are flat and the tables show stars only.
Jev-Based
| # | Repo | Stars |
|---|---|---|
| 1 | browser-use/jev-ultrafast | 14,008 |
| 2 | tamaratran/fast-jev-compaction | 5,763 |
| 3 | jarrodwatts/jev-trader | 1,690 |
| 4 | typesafe-ai/skills | 1,382 |
| 5 | Sac-Y/Jev-cu | 520 |
| 6 | thruwire/foreman | 450 |
| 7 | devagrawal09/jev-review | 437 |
| 8 | superagents-lab/jev-search | 354 |
| 9 | dabit3/jev-experiments | 352 |
| 10 | kerpopule/hermes-jev-skills | 340 |
Jev Integrations
| # | Repo | Stars |
|---|---|---|
| 1 | OpenByteInc/QuantDinger | 11,879 |
| 2 | 0xPlaygrounds/rig | 8,691 |
| 3 | ax-llm/ax | 2,935 |
| 4 | Arize-ai/openinference | 1,226 |
| 5 | reticlehq/reticle | 783 |
| 6 | awlevin/typesafe-computer-use | 693 |
| 7 | AgentiLoop/Agent | 618 |
| 8 | agentjido/req_llm | 580 |
| 9 | fhshaik/typesafe-mario | 314 |
| 10 | rokbenko/quackd | 222 |
OSS alternatives — not TypeSafe Jev
| # | Repo | Stars |
|---|---|---|
| 1 | mizorewww/laya-mlx | 2,255 |
| 2 | TianyuCodings/NanoJev | 1,688 |
| 3 | jaredpalmer/kev | 1,483 |
| 4 | bespokelabsai/nimble | 1,385 |
| 5 | vinnylarouge/jevlike | 1,141 |
| 6 | featherless-ai/simple-jev | 430 |
| 7 | wfzyx/von | 296 |
| 8 | ekzhang/openjev-sglang | 244 |
| 9 | razorback16/openjev | 242 |
| 10 | Heman10x-NGU/openJev-verdict-2.0 | 221 |
Ultrafast still leads Jev-Based; QuantDinger still leads Integrations. Treat the alternatives track as local-door reading (Kev / NanoJev / von) — star rank is not a swap.
What moved
Nothing TypeSafe-official that changes the contract, model pin, price, or channel door.
Worth one click
- Akramovic1/jev-pilot — Claude Code plugin that asks Jev to pick effort, subagent, and skill. Use it as a pattern check: is the menu closed, or is Jev writing the prompt?
- uehaj/jev-semgrep — Score every line against a meaning, then AND/OR/NOT the hits. A clean Score use; still not a reason to skip your own labels.
Read it here
Read on the blog. Also on jev.pro: the live board at https://jev.pro/top-repos. The 09:00 PT email is this same article in this same order.