jev-harness
Decision harness for TypeSafe Jev, confidence gates, shadow mode, recipes, and evals.
What it does
Decision harness for TypeSafe Jev, confidence gates, shadow mode, recipes, and evals. Claude CLI 48.9s → Jev 1.3s on the same row-filter job. Numbers, harnesses, and comparisons you can rerun on your hardware instead of trusting a screenshot.
Tags in the wild: benchmarks, agents, eval. The repo has install steps, env vars, and recent commits if you are deciding whether to depend on it.
How Jev is used
Wraps POST /v1/systemone answers in a production-shaped loop: map Choice, Score, and Noul outputs through a policy function, enforce minConfidence with review or suppress paths, log shadow decisions without changing live behavior, and replay fixtures in the eval CLI.
Setup: npm install jev-harness, export TYPESAFE_API_KEY, then construct DecisionHarness and pass parallel questions plus a decide() policy. OpenClaw users install the jev-harness plugin that bundles these recipes for shell gates, tool allow lists, and browser next-action steps.
Caveats: Not affiliated with TypeSafe or OpenClaw Foundation. README wall-clock table (48.9 s vs 1.3 s on one row-filter job) is a maintainer measurement, not a universal SLA.
What the harness adds beyond one Jev call
Calling Jev once is easy. Shipping it usually needs a policy that turns answers into notify, suppress, queue, or skip actions, a confidence gate when probabilities are soft, shadow mode to log would-be actions, and reusable recipes for alerts, inbox triage, model cost routing, and agent handoffs.
The library stays small and TypeScript-first so OpenClaw plugins, Node services, and benchmarks can share the same DecisionHarness API.
Recipes and OpenClaw jev-harness
GitHub documents recipes for shell exec gates, tool allow and deny lists, stuck-agent recovery, and browser next-action classification. The OpenClaw plugin on openclawdir maps those recipes to plugin tools while generative work stays on Claude, Codex, or other providers.
FAQ
- Is this the same as jev-claw?
No. jev-claw focuses on multi-model routing labels inside OpenClaw. jev-harness focuses on pre-tool and pre-shell gates with recipes and shadow evals.
- Which primitives show up in recipes?
Most recipes combine Choice for the action enum, Score for severity or urgency, and Noul checks for human review or policy fit in one parallel request.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.