Agent routing decisions don't need frontier models, and this repo proves it

Every agent wastes frontier-model tokens on decisions that don't require a frontier model. Which of 377 installed skills applies here? Are these retrieved passages worth reading? Which turns survive when the transcript must be cut? These are classification problems, not generation problems, and this repo treats them that way.

The mechanism is TypeSafe's Jev, a decision-only model that takes a state and typed questions — pick one, score this, yes or no — and returns a calibrated confidence. It never writes text. Hermes Jev Skills wires it into nine distinct agent operations: model routing, search ranking, memory filtering, turn selection, skill selection, triage, mailbox sorting, computer use, and browser use. Latencies run from roughly 0.4 seconds for routing and triage up to about 1.9 seconds for a full search round. Mailbox sorting costs two hundredths of a cent per message.

The handoff skill is the most honest entry in the table. The README reports that a handoff summary written from Jev's keep/summarize/drop digest actually recalled less than one written from the plain transcript — 58.7% versus the baseline's 37.5% alone, 75.0% versus 68.3% with one search assist. So what ships is the whole dialogue, capped at 1,200 words, with a path back to the old session. The repo measured its way to a worse result and shipped the better one anyway. That's a useful signal about the project's epistemic standards.

The privacy section is unusually detailed. Routing sends the user's turn, redacted and capped at 2,500 characters — never history, tool results, or files. Memory passages are truncated to 900 characters each, with store paths replaced by anonymous labels. Mail bodies are decoded before screening, because tracking links encode recipient addresses in base64 and a plain-text redactor would miss them. Profiles listed in private_profiles send only coarse features: length, whether code is present, whether risk words appear.

Fail-open behavior is explicit and specific. No key, timeout, rate limit, or low-confidence answer blocks a turn. Routing keeps the current model, memory returns the original list unfiltered, compaction drops nothing. Risk words like "production," "delete," and "payment" are hardcoded to never route to the cheapest tier regardless of what Jev returns.

The nine skills ship as plain SKILL.md files, making them readable by Claude Code, Codex, or anything else that consumes skill files. The installer detects whichever of those environments exist on the machine and installs for each.