Kerpopule's hermes-jev-skills wires TypeSafe Jev into nine agent decisions (routing, memory, compaction, skills, mail, computer and browser use) and reached 697 stars in five days under MIT

A GitHub repository named kerpopule/hermes-jev-skills, published on 18 September 2026, reached 697 stars in five days for a single idea: hand nine of an AI agent's small decisions to TypeSafe's Jev decision model, and leave the frontier model for the writing. The plugin is MIT-licensed and covers routing, memory filtering, transcript compaction, skill selection, mailbox sorting, triage, search, computer use and browser use.

What Jev decides in this plugin, and what the README measured

The core claim is that a modern agent burns frontier tokens on classification work: which model to call on this turn, which of 377 skills to load, which retrieved passages to read, which lane a message belongs in, which button to click next. Jev, TypeSafe's decision model, answers those with a calibrated confidence and returns no text.

The README carries per-skill numbers. Model routing decides which model is good enough for the turn in about 0.4 seconds. Skill selection scans 377 skills in about 2.8 s. Mailbox sorting hits 0.44 s p50 at about $0.00002 per message. Triage costs about $0.00006 per message. Search runs as a loop of two Jev calls per round (rank the results, then decide whether the evidence answers the question and which candidate query to try next), at about 1.9 s per round on the median of five live runs. Computer and browser use pick the next action in about 0.4 to 0.5 s.

Why the compaction skill deliberately does nothing by default

The most interesting entry in the table is the one that ships turned off. The team measured a handoff written from Jev's keep-summarize-drop digest against a handoff written from the plain transcript, and the plain transcript recalled more (58.7 percent alone and 75.0 percent with one search, against 37.5 and 68.3 for the earlier approach). So the shipping default sends the whole dialogue, 1,200 words, plus a link back to the old session, and Jev only compacts when a person opts in with HANDOFF_JEV=1.

That inversion is worth reading closely on any team already summarising transcripts by default: the plugin's own measurement recommended keeping the raw text, and the release honoured its measurement.

What the plugin is careful about, and what leaves the machine

Nine of the skills ship as plain SKILL.md files, so a team on Claude Code or Codex can use them without running Hermes. The hermes-jev plugin itself only uses public plugin seams (pre_llm_call, llm_request middleware, tools, and a slash command), so hermes update does not break it. Every skill fails open: on a Jev outage or timeout, routing keeps the current model, memory returns the original list, compaction drops nothing, and computer use returns reobserve, so a Jev incident costs at most the 2.5-second time budget and never blocks a turn.

The README is direct about what data leaves the machine: routing sends the redacted turn capped at 2,500 characters, memory sends the query and up to 900 characters per passage with store ids replaced by P0, P1, and a profile listed in private_profiles sends only coarse features and never the text.

For a team already routing between models by hand, the reason to run the shadow mode this week is that the numbers on real traffic will not match any vendor deck. A day of shadow decisions is cheap to log and answers the question a purchase does not: how many of the frontier turns your team pays for today are decisions in disguise.

Source

kerpopule, hermes-jev-skills on GitHub, published 18 September 2026, MIT licence, 697 stars on 23 September 2026.