Jev Lab — pasblinn

An open lab for using Jev correctly next to Claude Code. It is independent — not affiliated with TypeSafe, Anthropic or the jev-router project it builds on — and focuses on where a typed decision belongs, where it does not, and how to notice when routing silently stopped working.

A local proxy, jev-router, sends Jev a Choice over models with the prompt context; a policy with a cache guard and a confidence rule picks haiku, sonnet or opus. In the lab's test — same prompt, same Jev answer — the stock router kept Opus in two situations where the patch routed to haiku.

With stock jev-router 0.3.0, an empty folder in -p mode showed the status line and rewrote the model, yet everything still went to Opus with no decision recorded; with a 145 KB CLAUDE.md it declined as downgrade-not-worth-cache-rebuild. Both were silent; the patch fixes them and the doc relates the upstream PRs.

jev-health asks Jev one real question before the session opens, passing only on HTTP 200. If it fails, the session runs plain Claude Code with alert hooks, so a routing failure never blocks the session but is made visible.

Over 7 days, 78% of one developer's tokens were read by 6 giant sessions, while sessions where a per-turn router makes decisions that stick added up to 1%. The lab's conclusion: routing works, it is just not where the spend is.

It needs a Jev key to route, and the README's working figure is about US$ 0.00001 per decision. The patch's root cause was measured with JEV_DUMP, and docs/03-testing.md recommends verifying by transcript, never by the status line.

It labels itself a lab, not production tooling. docs/06-risks.md is marked read-before-serious-work, the fallback covers only what its health gate checks, and the pasted README truncates inside the how-it-works section, so some steps were not reviewed for this entry.