Skip to content

Repository files navigation

Awsome-Jev-Router

Route repeated Agent judgments to reusable Jev practices.

Jev Catalog Offline-first License

English · 中文 · Catalog · Contributing

Awsome-Jev-Router is a searchable Jev practice catalog plus a local Agent skill. It maps recurring decision points in redacted logs to public projects, implementation notes, and source evidence. Use it to choose what to prototype; it does not run third-party projects or change your Agent automatically.

In 30 seconds

What is Jev?

Jev is TypeSafe AI's first System One Model: a model for fast, structured decisions that software can use directly. Give it state and a typed question; it returns a constrained result such as Choice, Score, or Noul, rather than a long chat response.

Read the official launch post for the design, API examples, and benchmark claims. Jev is early access; performance numbers in that post are TypeSafe's claims, not an independent benchmark from this repository.

Why is Jev getting attention?

Agent workflows often need a small, repeatable decision instead of another long-form generation call. Jev's early-access release, public SDK and skill ecosystem, and community experiments around routing, tool safety, context compaction, and browser actions made the idea easy to try and discuss.

That attention is not proof that Jev fits every workflow. Test whether a decision is stable enough to type, cheap enough to evaluate frequently, and safe to review when confidence is low.

A long-form LLM call A Jev-style decision point
Generates text or code that the caller must parse and validate Returns a constrained value the caller can branch on
Useful for open-ended reasoning and synthesis Useful for repeated, bounded choices in a workflow
Best when the output is inherently textual Can reduce general-model work when the choice is narrow
Agent logs → local redaction → recurring scenarios → Jev candidates → sources and evidence

The project keeps thresholds, escalation, and side effects in application code: a recommendation is a lead to review, never an automatic action.

Quick start

Run the local skill from the repository root:

python3 skill/jev-practice-recommender/recommend.py \
  skill/jev-practice-recommender/examples/sample-agent.jsonl \
  --output /tmp/jev-recommendations.json

The input accepts JSONL, JSON exports, and plain text logs. The output contains scenario frequency, trigger signals, redacted excerpts, line numbers, candidate projects, and source files.

Copy skill/jev-practice-recommender/ into a Codex, Claude Code, Pi, or custom Agent skill directory to mount it. Logs are read locally; the default run does not make network requests, install, or execute recommended projects.

Read the input contract and privacy boundary in SKILL.en.md · 中文 skill guide

Featured practices

The full catalog currently has 256 entries. The home page highlights one representative project per category; category pages keep the complete list.

Category Project What Jev decides Borrow it for
Classification & Routing jev-logtriage Runs Noul, Score, and Choice over Loki logs and maps them to suppress/watch/review/notify/page, sending low confidence to review. Turn logs into tiered action instead of summaries.
Verification & Guardrails jev-axi Scores shell commands for destructiveness, exfiltration, remote execution, and security weakening before tools run. Add a safety gate before Agent side effects.
Scoring & Ranking citation-verifier Claude locates evidence, Jev scores whether a paper supports a sentence, and a human keeps the final call. Separate machine scoring from human review.
Agent Decisions Jev Ultrafast Picks the next browser action and element from state, calling a language model only when text must be typed. Let the LLM understand and Jev choose frequent actions.
Data Labeling & Curation jev-align Makes typed judgments over CSV, Parquet, and JSONL rows, sends ambiguous cases to humans, and improves the saved definition from corrections. Build auditable human-in-the-loop labeling.
Evaluation & Benchmarking jevcal Fits confidence thresholds on labeled data, validates on a holdout set, and fails CI when model changes invalidate them. Turn escalation rules into regression tests.
Calibration & Research Laya Emits Choice, Score, and Noul probabilities in one forward pass. Study low-latency, private, offline decision models.
Infrastructure, SDKs & Integrations typesafe-ai/skills Teaches an Agent through an installable skill when to hand a judgment to Jev. Mount Jev into an existing Agent.
Game & Simulation jev-plays-pokemon-red Keeps route planning and arithmetic deterministic; Jev chooses only at branches and battles, scored with Brier metrics. Limit the model to small, measurable decisions.
Finance & Trading Jev X Sentiment Analysis Turns deduplicated post evidence into an entry-range, stop-loss, and target decision card without trading directly. Convert social signals into reviewable advice.
Compliance & Legal LegalForecast-MTD Forecasts motion-to-dismiss outcomes and evaluates probability quality with micro-Brier metrics. Separate legal advice from uncertainty.
Content Moderation mastra-jev-moderation Decides whether to block and which category applies, with timeouts, circuit breakers, and thresholds. Build a low-latency moderation path with fallback behavior.
Scientific Pipelines No core entry yet There is no direct experiment-gating or scientific-result validation project in the catalog yet. Keep the gap visible instead of padding the list.
Related Practices & Discussions Jev is a really smart switch statement Frames Jev as a decision layer that maps context to constrained branches. Get the Jev-versus-chat boundary quickly.

Each row answers “what does Jev decide?” and “what does the surrounding code do?” See the category pages and research sources for more projects, authors, and evidence.

Catalog

Tag Entries* Use it for
Classification & Routing 24 Intent, request, skill, model, and traffic routing
Verification & Guardrails 22 Tool calls, permissions, code, and supply-chain checks
Scoring & Ranking 20 Quality, relevance, risk, and candidate ranking
Agent Decisions 31 Browser, context, action, and workflow decisions
Data Labeling & Curation 5 Document, dataset, and content labeling
Evaluation & Benchmarking 16 Evaluation, regression, and reproducible experiments
Calibration & Research 22 Confidence, latency, model, and method research
Infrastructure, SDKs & Integrations 43 APIs, gateways, SDKs, deployment, and adapters
Game & Simulation 10 Games, world models, and simulated environments
Finance & Trading 4 Trading, portfolio, and risk decisions
Compliance & Legal 1 Compliance, contracts, and legal workflows
Content Moderation 4 Spam, ads, abuse, and safety judgments
Scientific Pipelines 0 Scientific data and experiment workflows
Related Practices & Discussions 54 X, blogs, interviews, and public signals without a standalone project

* Counts reflect the current category pages. Each project has one primary category. Chinese mirrors are in categories/zh-CN/.

Choose by problem

Your question Start with
“Which model or skill should handle this request?” Classification & Routing; Agent Decisions
“Should this tool call be allowed?” Verification & Guardrails; Compliance & Legal
“Which candidate is more relevant or safer?” Scoring & Ranking; Evaluation & Benchmarking
“How do I connect Jev or swap a provider?” Infrastructure, SDKs & Integrations
“What has been built or discussed in public?” Related Practices & Discussions; Research sources

Research sources

The research chapter records official material, open-source projects, technical writing, and public social discussions. It distinguishes runnable implementations, public opinions, and leads that still need verification. Prices, model aliases, performance, and platform capabilities change; inspect the original source before adoption.

Inclusion rules

  • The source is public, citable, and explicitly uses Jev or System One typed decisions.
  • The summary states the concrete scenario, judgment type, and implementation value.
  • Pure opinion, unverifiable promotion, and projects that only resemble Jev stay out of practice entries; they may appear in discussions with evidence strength labeled.
  • Inclusion is not an endorsement of code quality, security, stability, performance, or license suitability.

Found an outdated, duplicate, or weakly supported entry? Open an issue or pull request with the original link and verification evidence. See Contributing.

Verification

python3 -m py_compile skill/jev-practice-recommender/recommend.py
python3 skill/jev-practice-recommender/recommend.py \
  skill/jev-practice-recommender/examples/sample-agent.jsonl \
  --output /tmp/jev-recommendations.json
git diff --check

License

New code and documentation are MIT licensed; see LICENSE. Third-party projects, names, authors, links, and content in the catalog remain subject to their own licenses and rights.

About

A Jev practice catalog and local-first Agent skill router: map recurring logs to reusable decisions, tools, and workflows.

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages