v0.2.0 — Project collection, guided setup & evaluation workflows

Pick a use case, give it to your agent

• 45 project/app/report/alternative entries, with evidence labels.
• 108 scenarios in both English and Chinese READMEs; all 15 + 22 + 39 supplied roundup entries mapped and attributed.
• 11 skills, including new jev-setup and jev-redteam.

Setup without dead ends

Prefer an existing OpenRouter account, or choose the official TypeSafe API with --provider typesafe and TYPESAFE_API_KEY. jev-decide setup is a read-only presence check. With no Jev key, the agent asks before using the current agent or another approved available model such as DeepSeek to simulate. No silent provider switches, retries or fabricated probabilities.

Evaluation examples

Authorized batch, multi-turn and team review protocols, three benign authored fixtures, and an offline JSONL request builder. These are not live jailbreak results or an attack runner. Existing 14 real API I/O examples and negative comparison results remain unchanged.

Validation

96 tests, 11 skill validators, fresh-wheel installation, 60 provider-specific copied-skill dry runs across three temporary host layouts, and offline request-builder checks. TypeSafe transport is mock-tested, not live-tested; no new paid model or target calls. Native host invocation is not established by packaging checks.

Agent installation guide · Full validation and source notes

Download the complete source ZIP for agent-first installation, or the skills-only ZIP for manual folder copying. The wheel installs the shared CLI, not the host skills. Verify downloads with SHA256SUMS.

v0.1.2 — Choose Jev or your current agent

No key? Choose A or B

All nine skills now stop and ask before making a missing-key judgment:

• A — Get an OpenRouter key and use real Jev.
• B — Let your current agent simulate the classification with the same input and criteria.

The agent must wait for your choice. B requires no Jev CLI or key and labels every result mode: agent_simulation, jev_called: false, with null probability/confidence. It does not fabricate Jev responses or calibration. Your host agent's usual costs and privacy terms still apply.

Both READMEs and the agent-first installer explain the two modes. This release also packages the 90-scenario catalog, community previews, 14 recorded input/output pairs, and copyable usage prompts added since v0.1.1.

Install

Give your agent the installation guide. It now pins v0.1.2.

• jev-skills-v0.1.2.zip: full source and all nine skills.
• jev-skill-v0.1.2.zip: standalone general jev skill.
• Wheel/source distribution: shared CLI package.
• SHA256SUMS: download checksums.

Validation

86 tests, nine skill validations and 27 copied-skill dry runs passed. The CLI is unchanged: a missing key returns an error, never a simulated response. No new paid API call was made. The A/B instructions were reviewed, not independently agent-tested or natively exercised in every client. Existing live evaluation results remain unchanged.

v0.1.1 — Context-first, parallel decisions

What's changed

• All nine skills now require enough decision-relevant context in every request: goals, rules, source evidence, relevant history and candidate meanings.
• Independent judgments should use shared-state batching or bounded host concurrency rather than unnecessary serial LLM calls. Dependent steps still wait for fresh state.
• Added a two-record, six-question batch template and a context/throughput guide.
• Updated both READMEs and the copy-to-agent installation guide to v0.1.1.

The CLI runtime is unchanged. This release adds skill instructions and examples, not a built-in parallel scheduler.

Install

Give your agent the prompt in the installation guide.

• jev-skills-v0.1.1.zip: complete source, all nine skills, examples and documentation.
• jev-skill-v0.1.1.zip: the self-contained general jev skill.
• Wheel: shared jev-decide CLI for the focused skills.
• Source distribution: runtime plus skills and reference material.
• Verify downloads against SHA256SUMS.

Validation

79 unit tests, all nine skill validators, and 27 copied-skill offline checks passed. The new batch template is synthetic and has not been live-evaluated. No new parallel speedup or quality result is claimed; previous live receipts remain unchanged.

Special thanks to LINUX DO.

Awesome Jev Skills v0.1.0

Awesome Jev Skills — first release

85 practical scenarios with editable prompts, templates and sources. English and Chinese READMEs show outputs from all 14 recorded API examples.

Included

• Eight focused agent skills plus the general jev skill.
• Shared Python 3.10+ jev-decide CLI, using OPENROUTER_API_KEY.
• Browser/desktop choices, routing, context advice, triage, documents, code review, code search and simulations.
• Community project comparisons and transparent evaluation receipts, including negative results.

Install

uv tool install git+https://github.com/wuyoscar/jev-skill.git@v0.1.0
npx skills add wuyoscar/jev-skill --skill jev-triage

Choose Codex, Claude Code or OpenCode in the installer. Set OPENROUTER_API_KEY in the agent's environment. See the README for other skills and examples.

Downloads

• jev-skills-v0.1.0.zip: complete source collection, all nine skills, docs, tests and evaluation receipts.
• jev-skill-v0.1.0.zip: just the general jev skill with its bundled script and references.
• jev_skill-0.1.0-py3-none-any.whl: shared CLI package, not host skills.
• jev_skill-0.1.0.tar.gz: Python source distribution including the skill collection.
• SHA256SUMS: checksums for these four files.

75 unit tests pass locally. Live API samples use synthetic input and do not establish end-to-end success in native clients. The source includes the paired agent pilot and calibration results for inspection.

Special thanks to LINUX DO.