Blog · · briefing

Put a permission bouncer between ask and --yolo — JEV Brief · Oct 3

Nothing TypeSafe-official moved the contract today. The useful noise is harness permission and cheap-classifier-first composition: Broderick-Westrope/anvil #38 (merged) adds a permission bouncer between ask-every-time and --yolo, and hipstersmoothie/standard-reader #192 keeps GlotLID on the hot path with Jev only as a confidence-gated tiebreak. If you only have four minutes, fail closed when the bouncer is unsure, then scan Top Repos.

  1. Put a permission bouncer between ask-every-time and --yolo — fail closed when unsure. anvil #38 (merged) adds a middle ground: explicit allow/deny rules always win; what rules leave open goes to a cheap System One classifier; timeouts, malformed answers, missing keys, and open circuit breakers fall through to the human (or to yolo exactly as today) — never open on error. Code computes facts (workspace bounds, hosts, protected paths); the model only answers narrow yes/no hazard questions. Live calibration on 55 cases: jev-1.13.0 separates routine vs dangerous; von-1.0.0 scores everything mid-band and never allows or denies. Under --yolo the bouncer’s allow/deny still stand (--yolo=full still bypasses). Quiet-day Pro move: pin a refuse path for unsure/error before you let a harness auto-act.
  2. Cheap local detector first; Jev only on the ~1% tiebreak — pin floors, keep Jev off the write path. standard-reader #192 ships GlotLID v3 (~0.3 ms/doc) for language tags, then Jev (pinned jev-1.13.0) only when GlotLID confidence is under 0.5, and only from the hourly sweep — never on the ingest write path — needing ≥0.8 to tag. Benchmark (1,049 docs): GlotLID + Jev tiebreak beats franc and either alone. Same shape as the bouncer: local/rules first, System One for residual judgment, hard floors.
  3. Adjacent doors and watchlist — not contract moves. dxos/dxos #13641 (merged) adds a code-index “System One zoom” that answers design questions with scored subgraphs and a judged diagram. awaku7/agentcli #140 (merged) extracts a shared goal-completion decision runtime (TypeSafe/OpenRouter Jev + Laya consistency). fbmoulin/pje-download #66 is docs-only research that flags wrong model ids and a LangGraph HITL loop that drops corrected prompts — useful refuse checklist, not a TypeSafe ship. New / watchlist: mavericksxx/jev-gateway (Claude API proxy that routes/escalates/trims with Jev; 0 stars — not Based yet).

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 (Laya / NanoJev / Kev / von) — star rank is not a swap for TypeSafe Jev. anvil’s live table also shows why: von scored everything mid-band and never allowed or denied on the bouncer harness.

What moved

Nothing TypeSafe-official that changes the contract, model pin, price, or channel door. Reddit collect returned 403 (empty). HN still carries the Luna Decision API press from prior days — competitive framing already covered Sep 30; not a fresh TypeSafe ship. Fresh-ish on HN: Decision Index 0.2.1 (70 open reproductions) — see Worth one click.

Worth one click

  1. anvil #38 — permission bouncer, decision log, rule triage — Explicit rules beat the classifier; unsure/error fails closed to the human; yolo still respects bouncer allow/deny.
  2. standard-reader #192 — GlotLID + Jev language tiebreak — Local detector on the hot path; pinned jev-1.13.0 only on low-confidence sweep with ≥0.8 floor.
  3. Decision Index 0.2.1 — 70 open reproductions of TypeSafe Jev’s Decision Model shape; useful for competitive reading, not a TypeSafe pin.

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.

Sources