jev-governor is a Claude Code mod by roma-vibe that puts Jev behind decisions about model choice, reasoning effort, subagents, and context cleanup. The interesting constraint: changing models mid-conversation can invalidate the prompt cache, so a cheaper model may cost more for a short continuation.

Where Jev fits

  • hooks/lib/router.ts asks Jev about the task in both model-option orders, estimates effort, and flags risky work. Short “go on” replies keep the previous decision locally without an API call.
  • Moving the main chat from Opus to Sonnet needs enough expected savings to repay a fresh cache write; effort can change without a model switch. Subagents begin with empty context, making a model choice cheaper there.
  • The compaction hook asks Jev which old tool calls to prune while keeping remaining text verbatim. Removed calls can be archived and linked back; large build/test outputs are trimmed before they enter the conversation.
  • hooks/lib/jev.ts posts typed choice, score, and yes/no questions to a TypeSafe-compatible OpenRouter endpoint. The Jev call chooses from bounded actions; code still applies the routing thresholds and fallback rules.

What remains unproven

  • The README reports roughly 7–13% API-equivalent savings in one day of mixed work and about 60% on long A/B tasks. Those are the maintainer’s claims, not results I reproduced. hooks/lib/savings.ts separates ledger-derived estimates from effort savings based on a heuristic; even the “exact” counterfactual assumes what the same tokens would cost without the mod, not what an untouched session actually would have done.
  • The mod sends redacted recent history, commands, and output to OpenRouter. Its pattern-based redactor is inspectable, but cannot guarantee that every secret or sensitive detail is removed. An excluded-project setting sends nothing; function hooks run with user permissions, not in a sandbox.
  • On Jev failure the mod generally leaves Claude Code running with its fallback; failure to classify does not establish correct task routing. No independent quality comparison or privacy audit is provided.

Worth studying for its cache-aware routing and reversible pruning, not installing blindly for a savings promise. README, architecture, and implementation files inspected; plugin not installed or run. Related: Jev inside Pi and agent-stack’s policy boundary.