hermes-jev (DoGMaTiiC)

jev-judge is a pre_tool_call gate on terminal, write_file, and patch calls, asking whether a call is destructive, an exfiltration, or high-impact, and it also provides a jev_ask tool. jev-skill-router runs at pre_llm_call, asking Jev for at most one skill from the live roster and injecting a single skill_relevance line.

TypeSafe Jev directly with TYPESAFE_API_KEY (preferred), the Vercel AI Gateway with AI_GATEWAY_API_KEY, or OpenRouter via OPENROUTER_API_KEY (router only). The default backend: auto picks the direct route when its key is present, else the gateway, else OpenRouter.

Both plugins fail open: nothing breaks and nothing is sent. That is the README's shared convention, alongside thresholds living in code and one JSONL decision line per judgment, so an unreachable provider never blocks your Hermes turn.

Clone the repository once, then symlink the plugin you want into ~/.hermes/plugins (the README shows a ln -s for jev-skill-router) and run hermes plugins enable jev-skill-router. Put your keys in ~/.hermes/.env. Catalogue installs are noted as coming once the entries land.

No. The judge runs in shadow mode by default: it judges and logs gated calls, and that state leaves the machine, but it never blocks them. The router is off until switched on, and while off it sends nothing at all. Both are explicitly opt-in.

Each plugin directory has offline tests run with python3 tests/test_offline.py, and the README lists hermes plugins validate plugins/jev-judge plus hermes plugins doctor plugins/jev-skill-router --ci as verification commands. Each plugin's own README documents privacy notes and exactly what leaves the machine.

It is a community plugin set published by the dogmatiic GitHub account. The README closes with MIT - see LICENSE, with each plugin keeping its own copy. The digest shows no accuracy benchmarks for either plugin, so treat the design guarantees as documented behavior rather than measured outcomes.