Emeric Commenge’s automodel puts Jev in front of Claude Code’s model and reasoning-effort selection. The interesting problem is not recognizing an easy prompt: it is recognizing that “also add a test” may continue a difficult job, while “what does HTTP 409 mean?” is a temporary aside.

I inspected the README, Jev client, routing policy, privacy transformation, proxy, and author’s integration notes at revision df503fd00abc936cf9ff631170871d397bd1b6ce. go test ./... passed locally. Those tests use fake Jev and upstream services; I did not install the integration, route a real Claude session, or reproduce model-quality or performance claims. This is not a security audit.

Keep a model of the work

  • Jev scores reasoning tiers and classifies the new prompt’s relation to unfinished work: continuation, extension, aside, wrap-up, or a separate task.
  • Ordinary code preserves a work-level floor, remembers a smaller detour, and restores the paused job when the user returns.
  • Explicit pins and repository bounds constrain routing; a short “yes” can restore the work’s tier without another model call.
  • On Jev errors, warm sessions generally retain their existing decision; other triggers fall back to a bounded default or the ongoing work’s level.

A cheaper model can make an expensive switch

The policy weighs expected tier cost, an application-selected penalty for insufficient capability, and the cost of rebuilding cached context. These are estimates—not measured probabilities that a coding task will succeed.

The author reports a cache-preserving effort change with 41k tokens read and 51 written, rather than a 13k-token rewrite. That result was not reproduced here. The inspected proxy implements deterministic effort-only messages on supported models, with a fallback when the upstream refuses them; compatibility depends on the client and provider behavior.

See the decision—and its data boundary

  • A local ledger records tier choices, confidence, fallbacks, usage, and why the router held or changed a decision.
  • Flagged mistakes become replayable labeled cases; shadow-model agreement is available but is not proof of correctness.
  • Normal routing sends prompt/reply excerpts, compaction summaries, and repository signals to Jev through OpenRouter.
  • Metadata mode replaces text with sizes and task hints; it reduces disclosure, not to zero, and discards useful context.

Test the integration separately from the economics

The installer changes Claude Code’s base URL, hooks, status line, and permissions, and enables automatic updates by default. A failed proxy can interrupt Claude access even when Jev routing itself has a fallback. Inspect those changes and the rollback path before installing.

The README’s subsecond decisions, roughly $0.00004 per call, and 98% cache-read share are author-reported and unreproduced. Its savings report is a counterfactual estimate, not a matched experiment demonstrating equal task quality or subscription-quota savings.

Typed output is not a correct decision. The useful contribution is the inspectable policy around Jev: work continuity, switching costs, user control, and evidence about what actually ran.