Hermes adds a community skill router backed by an older TypeSafe experiment

Hermes has added a community catalog entry for a Jev-backed skill-selection package. The new adoption event is the catalog merge on September 24, not a newly run benchmark. The package offers one skill and a standard-input/output MCP server that helps choose which skill to inspect before loading it. It can also decide that none fits.

The distinction between distribution and evidence is important here. The submission cites a TypeSafe cookbook experiment using 182 Hermes skills and 488 requests. That cookbook says its published run used Jev 1.12 and a pinned Claude Haiku 4.5 model, rendered on July 31. Those numbers explain the design motivation; they do not establish today's performance on a different user's larger or edited roster.

The facts

- The catalog change merged on September 24 and identifies the package as community-maintained. - The final catalog points to ydmw74/jev-skill-router at revision 76ee09e, not the older revision cited in the initial submission prose. - The design ranks the roster, then rechecks three candidates using longer descriptions and instruction excerpts. - The cookbook reports wrong loads falling from 16.8% to 7.3% and needless loads from 9.8% to 4.0% in its pinned 488-request experiment. - The plugin reads the profile's live skills roster instead of baking in that benchmark roster. - The package and catalog warn that the portable loader does not interpolate the API-key variable in mcp.json; inherited environment is required there.

Why it matters

A growing skill library creates a selection problem before the agent performs any work. Short index descriptions make neighboring procedures look similar, and a forced choice can be worse than abstention. This integration turns that problem into an explicit optional decision step. The interesting product pattern is not simply adding another tool: it is separating retrieval of operating instructions from the larger model's execution of them, while leaving the final judgment with the agent.

Current

Checked September 24, 2026. The inspected release baseline is v2026.9.21. The primary source is merged 2026-09-24T00:53:09Z. Package availability and development-source state are separate: a tag does not establish that an open proposal is usable, and a merge does not establish that every installation has received it.

Evidence

Primary: NousResearch/hermes-agent PR #120081 (https://github.com/NousResearch/hermes-agent/pull/120081). Supporting context: Jev skill router — package and loader limits (https://github.com/ydmw74/jev-skill-router); TypeSafe — historical skill-suggestion experiment (https://docs.typesafe.ai/cookbooks/skill_suggestion.md). Source pages and available diffs were inspected; test and live-account results remain attributed to their authors, not independently repeated measurements.

Operator take

Treat the historical benchmark as a hypothesis worth testing on your own tasks, not an accuracy guarantee. Evaluate wrong selections, unnecessary selections, abstentions, extra request cost and added latency together. Inspect what request and skill context the decision service receives before enabling it for private work. Preserve the distinction between an MCP tool that can suggest a skill and a mandatory automatic router; the catalog entry does not establish that every Hermes turn is now routed through Jev.

Caveat

The cookbook results are vendor-produced, historical and model-specific. They were not independently reproduced for this report. The submission's earlier verification revision differs from the final catalog pin, and the loader limitation matters for packaging. This is an optional community integration, not a default replacement for Hermes skill policy.