$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people
~/wiki/ideas

Repos: coding agents, orchestration, memory, browser and computer use, integrations

[ community tier ][ updated 2026-09-21 ][ confidence medium ][ jev-1.13.0 ]#community · repos · coding-agents · browser-use · integrations

TL;DR The agent-side half of the community repo catalogue. Index and the vetting checklist: Community repos: what people built and how they use Jev. Data, document, judging, real-time, market and replica repos: Repos: data, documents, judging, real-time, markets, business apps, replicas. Stars and languages are the 2026-09-20 capture (roundup repos 2026-09-21; see Community repos: what people built and how they use Jev on star drift). Unaudited code, most of it with shell, browser or editor reach.

Coding agents and developer workflow

repo stars/lang what it does, and how it uses Jev maps to
fast-jev-compaction 4918/TS Replaces Claude Code compaction. Conversation as state → 2 noul per unpinned call (keep call? keep result?) → drop or truncate below 0.5. Cookbook: Classifying RAG passages, Patterns: agent internals, context and coding agents P07
winnow 40/Py Claude Code function-hook sieve. Splits each large tool result into ~25-line blocks, asks one Noul per block ("needed now?") in one call, hides blocks below WINNOW_DROP 0.1 behind a recall stub, keeps the uncertain band 0.1–0.5 verbatim, and hides nothing if the judge says the output shows an error. A second hook ranks project memory files per prompt and injects the top 3 above 0.5. winnow replay scores a judge offline on your own transcripts (weak labels plus blind hand labels) and prints calibration (ECE) beside ordering (ROC AUC); the strict question set was overconfident — 23% of its bottom bin was needed. Falls back to TypeSafe's system-one-adapter on Haiku 4.5 with no key. Cookbook: Classifying RAG passages, system-one-adapter: LLM-backed drop-in for TypeSafeClient, Patterns: agent internals, context and coding agents P07
foreman 425/Py Supervises coding agents. Factory events → 10 nouls (worker_stuck, work_off_track, needs_human…) in one call → a deterministic policy picks STEER/STOP/VERIFY Confidence-gated routing, Patterns: agent internals, context and coding agents P04
Canny 19/TS Warden for Claude Code and Codex CLI: "facts go to code, judgments go to Jev, only facts can block." Two Nouls — "does this edit break a rule in CLAUDE.md/AGENTS.md" (a note above 0.9) and "does this message claim the work is done" (treated as not a claim below 0.1). Jev can only relax the done-gate; the block itself comes from an append-only ledger with no passing check since the last edit. Requests are cached by content hash, so canny replay re-derives verdicts offline. Cookbook: Guardrails for LLMs, Patterns: agent internals, context and coding agents P04
jev-review 400/TS Staged review. Noul risk matrix → Choice+Score file profiles → Choice evidence → Choice mechanism → Score severity → routing Composite scoring, Patterns: agent internals, context and coding agents P10
stanley-code 98/TS Bounded agent workflows. One choice over workflow metadata + cannot_tell, gated on confidence ≥ 0.6, probability ≥ 0.55, margin ≥ 0.15; then per-hunk choice Intent routing, Patterns: agent internals, context and coding agents P10
jev-codex-router 152/Py Per-call model and effort routing for Codex. Three independent Choice questions in one request: does the mandatory-Astra policy apply, the cheapest sufficient tier (luna/sol/astra), and the minimum thinking depth (low→max). Jev sees only a bounded dossier — task, step type, intent tail, tool name, tool-output tail, image flag — while the executing model gets Codex's full replay. Fail-open to astra/medium on any error, sentinel-file kill switch, shadow mode, decisions logged locally for calibration. Intent routing, Patterns: agent internals, context and coding agents P02
jev-router 245/JS Per-turn model routing. Prompt + tier + context tokens → four scored dimensions + confidence → fast or strong tier Intent routing, Patterns: agent internals, context and coding agents P02
compact-adviser 146/TS "Compact now?" Two one-sentence questions per request → one composed score against a floor sliding 0.90→0.50 as context fills. Confidence-gated routing, Patterns: agent internals, context and coding agents P07
pi-warden 102/TS Pi guardrails. Tool call + project Markdown rules → Jev judges irreversible calls, rule breaches, stuck loops and unverified "done", then steers rather than blocks. Questions live only in docs/guards.md Cookbook: Guardrails for LLMs, Patterns: agent internals, context and coding agents P03
jev-rules 43/JS Injects only the relevant standing rules. Prompt text → one noul per rule and map doc in one call → inject ≥ 0.6; shows all on error Cookbook: Skill suggestion, Patterns: agent internals, context and coding agents P08
jev-shell-history 69/TS zsh suggestions. Typed prefix + 100 id-tagged history entries → Choice over ids + Noul "does any complete this?" → show if top ≥ 0.3 and Noul ≥ 0.5 Cookbook: Line-by-line search
blink 38/TS Codebase search with no index and no embeddings. Jev scores file and folder names; likelier paths get more of N walkers, each descending to a file, and the result is the share of walkers that ended there (74%/16%/10% in the README's example). Every walker writes a trace with each choice and probability, so a search is auditable. Quotes $0.042/Mtok input with free output — verified. Cookbook: Line-by-line search, Patterns: judging, search, documents, real-time and markets P17
building-with-jev-skill 124/— Agent skill for writing Jev programs, not a caller: question design, state, thresholds; targets jev-1.13. The typesafe-ai agent skill and Claude Code plugin
jevify 37/— A prompt, not code (no licence): one brief you paste into a coding agent working in your repo. The agent reads the official docs plus the architecture probe study (Field reports: independent evaluations, critiques, open replicas), finds where the project pays for model calls, generates text only to parse a decision out of it, or falls back on brittle rules, then proposes three kinds of opportunity — direct savings, better outcomes, new capabilities — and names the assumptions that exist only because semantic computation was expensive. It also demands what most Jev pitches skip: a comparison against deterministic code, caching, embeddings, conventional classifiers or a small generative model; an evaluation designed to prove the idea wrong; vendor claims, measured results and hypotheses kept apart. Cited as a good example by @mathfax, who appears to work at TypeSafe. Consult guide: could Jev help this project?, The typesafe-ai agent skill and Claude Code plugin

Browser, computer and mobile use

repo stars/lang what it does, and how it uses Jev maps to
jev-ultrafast 11208/Py Browser agent, no screenshots. Indexed DOM element table → one request: operation (CLICK/TYPE_TEXT/SELECT/SCROLL/WAIT/DONE/BLOCKED) plus a speculative target question per operation. Speculative fan-out, Patterns: browser, computer use, voice and product UI P12
trycua/cua 25014/HTML Computer-use platform with a jev-use recipe. Not in the README; per the team's article, code builds candidate actions with ids, Jev picks one, code validates, executes, re-observes State: what you send Jev, Patterns: browser, computer use, voice and product UI P14
typesafe-computer-use 582/Py macOS computer use, ~$0.0002/step (author). OCR + accessibility tree → candidates → three Choices plus an offscreen question per request; a Noul verifies the typed field; aborts below 0.4 Speculative fan-out, Patterns: browser, computer use, voice and product UI P14
mobile-jev 264/JS Android agent, no ADB. Screen + installed apps → operation plus compatible-target questions per request; code rejects stale targets. Typed text is a span of the goal; DONE is not proof Speculative fan-out, Patterns: browser, computer use, voice and product UI P12
jev-voice-browser 148/JS Voice to Playwright, acting on partial speech. 9–11 questions per transcript: intent/target/site/spans (Choice), complete/is_command/destructive (Noul), scroll_amount (Score), each thresholded. Confidence-gated routing, Patterns: browser, computer use, voice and product UI P14
unclutter 143/TS Hides page clutter. DOM candidates → typed keep/ad/promotion/newsletter/social/cookie/uncertain choices → hides only when probability and confidence are ≥ 0.9 Cookbook: Guardrails for LLMs, Patterns: judging, search, documents, real-time and markets P18
WindTunnel 72/HTML WebMCP interface benchmark. Jev + Mercury 2.5 is 1 of 21 configs (49/49 with WebMCP, 25/49 with DOM controls); Jev decides, Mercury writes arguments. Runner in experiments/jev/ Workflow evals: how TypeSafe measures Jev, Patterns: browser, computer use, voice and product UI P13

vlad-terin/jev-browser, announced 2026-09-17 as a Codex browser adapter ("Codex plans, Jev selects, the browser moves"), 404'd at capture; an unrelated jkudish/jev-browser named in the jev-mcp README was not captured.

Agents, orchestration and memory

repo stars/lang what it does, and how it uses Jev maps to
skillbox 214/TS Self-hosted skills library over MCP. Task + skill descriptions → an uncalibrated 0–4 rubric score per skill (not a probability); returns ≥ 3, search fallback Cookbook: Skill suggestion, Patterns: agent internals, context and coding agents P08
agentic-stack-desktop 64/Py Knowledge graph across four coding agents. Not in the README; the linked post is about finding Jev-shaped decisions in a workflow Consult guide: could Jev help this project?
continual-memory-policy-model 5/Py Memory policy with delayed feedback. Not in the README; its WRITE/UPDATE/DELETE/LINK/COMPACT/NOOP action set is Jev-shaped, but it trains its own policy Primitives: Choice, Score, Noul, Patterns: agent internals, context and coding agents P09
hermes-jev-skills 387/Py Nine agent-agnostic SKILL.md skills (Hermes, Claude Code, Codex) handing small decisions to Jev: per-turn model routing over your own catalogue, which retrieved passages to read, which turns survive a cut, which of 377 installed skills a turn needs, message triage, mailbox lanes, and the next GUI or browser action from a table you already approved — Jev can only return an id you supplied. Author's numbers: ~0.4 s per routed turn, 377 skills in ~2.8 s, 71 turns in 0.95 s, $0.00006 per triage, $0.00002 per mail message, 60 passages per memory request. All of it fails open (no key, timeout or low confidence → model kept, nothing dropped). Honest negative: a Jev-digested handoff recalled less than the plain transcript, so they ship the transcript. The key goes to the OS secret store, never to the agent, and the README lists what leaves the machine per skill — a routed turn sends the user's turn redacted (2,500 chars) before the agent can object. MIT. Intent routing, Cookbook: Skill suggestion, Patterns: agent internals, context and coding agents P02, P07, P08
octavus-dual-brain 0/HTML Write-up placing Jev as System 1 before an orchestrator. intent (Choice) + risk_score (Score 0–3) in parallel; 10 filings in one call Intent routing

Integrations and catalogs

repo stars/lang what it does, and how it uses Jev maps to
jev-mcp 214/TS Ten MCP judgment tools: jev_verify, jev_screen, jev_find, jev_rerank, jev_classify, jev_decide, jev_compare, jev_extract, jev_review, jev_gate. The validation is the value — a returned choice must belong to the requested set and be a maximum-probability option; a distribution must carry exactly the expected keys, finite, in [0,1], summing to 1 within 0.01; a malformed answer fails closed for that item while valid siblings survive. jev_extract chooses among candidates a regex already found, so the model picks a value and never writes one. Thresholds (auto_accept 0.8, block_at 0.75, review_at 0.25) are cookbook starting points, not tuned. Providers: direct TypeSafe, OpenRouter (typesafe/jev-1.13), Cloudflare Workers AI, Vercel AI Gateway or any /v1/systemone-compatible endpoint — the gateway routes are unverified against first-party docs. HTTP API: POST /v1/systemone and GET /v1/models, Cookbook: Double-checking citations, Cookbook: Guardrails for LLMs
typesafe-mcp 133/Go MCP server with one evaluate tool. Passes state + a questions map ({type, instructions, criteria?}) to POST /v1/systemone, default jev-latest; validates criteria locally HTTP API: POST /v1/systemone and GET /v1/models
semdecide 18/Py "grep for meaning and jq for judgment" — typed decisions as Unix exit codes. is (Noul), choose (Choice over named options), score (ordered rubric), filter (JSONL, order preserved, _semdecide metadata added), and a guard recipe asking narrow signals (authorization, destructiveness, ambiguity, secret exposure, consequence) that deterministic code turns into allow/escalate/block. Exit codes keep false (1), uncertain (3) and provider failure (4) distinct; guard fails closed to escalate. Its --json sample shows "confidence": null for is — verified, Noul answers carry no confidence (HTTP API: POST /v1/systemone and GET /v1/models). --max-input-bytes defaults to 1,000,000, a local cap far above the documented 64k-token request budget. Cookbook: Guardrails for LLMs, Patterns: agent internals, context and coding agents P03
awesome-jev-by-typesafe 684/Py Curated use cases and starter code; a directory, not a caller. Quick start matches SDK 0.6.0; patterns mirror the official four Patterns overview
json-render 17848/TS Vercel Labs generative-UI framework — not Jev-specific; listed because @charliejhills' roundup links it. Its only Jev mention is an unreleased "Experimental Jev composition" path (experimental_composeSpec, experimental_createEvaluator). Nothing about the state, questions or thresholds it would send is in the capture. —

Related

Sources