v0.6.0 — typed handoff between coding agents and Jev

Hand the steps of an agent loop that produce no text — is the build done, which element do I click, is this shell command safe, keep or drop this message — to Jev; the LLM takes over wherever something has to be written.

npx -y jev-use install    # Claude Code, Codex, pi — whichever it finds

In this release

• check / pick / rate question builders and verdicts read as answers.<key>.answer — plain names over SDK vocabulary.
• Backends: TypeSafe direct, OpenRouter, Vercel AI Gateway, and a keyless mock.
• Claude Code plugin with the PreToolUse command gate, Codex plugin manifest, pi extension, MCP server with two tools.
• Zero runtime dependencies on the judgment path.

Measured, first-party, all re-runnable from bench/examples/

• 454 judgments over real states: 82.2% agreement with a claude-opus-5 reference, 89.5% among non-escalated verdicts, against a 68.7% majority-class baseline. The family where it scores below a constant answerer is published too.
• Gating every shell command through the hook costs zero LLM tokens.
• The latency headline is corrected: enum-constrain the LLM baselines and the gap is 3x, not 14x. What survives the fair comparison is cost and the answer landing inside the option set.

Numbers, methodology and caveats: bench/RESULTS.md