hermes-jev-approvals
PoC: TypeSafe Jev as the reviewer for Hermes Agent smart command approvals.
Want the full Hermes Agent plus Jev setup, including skill routing and MCP? Read the Hermes Agent + Jev guide first.
What it does
PoC: TypeSafe Jev as the reviewer for Hermes Agent smart command approvals. 8.7x faster, 4.4x fewer prompts, measured on 153 real commands. Approvals only. Agent plumbing where probabilities beat regex on model output. Fork the pattern, not the prose.
Tags in the wild: agent tooling, agents. The repo has install steps, env vars, and recent commits if you are deciding whether to depend on it.
How Jev is used
Registers TypeSafe Jev (or OpenRouter jev-latest) as Hermes' auxiliary approval provider. Each flagged command triggers one parallel ask with six typed questions: verdict Choice (APPROVE/DENY/ESCALATE), policy Noul, blast radius Score, and safety nouls for self-advocacy, secrets, and outbound data.
Setup: Set approvals.mode smart, point auxiliary.approval.provider at typesafe-jev, and write approvals.smart_policy from your own traffic. TypeSafe direct setup is two menu clicks per README; OpenRouter needs manual config.
Caveats: Benchmarks used real mined commands on one machine; independent live-sandbox studies report smaller speed ratios. Without smart_policy, Jev may deny routine developer commands.
Why Hermes needs a typed reviewer
Hermes smart mode sends flagged shell commands to an auxiliary model that must answer with exactly APPROVE, DENY, or ESCALATE. A general chat model wastes tokens generating reasoning you throw away. Jev answers that shape natively with one Choice plus supporting nouls and scores.
The plugin does not modify Hermes core: auxiliary_client already resolves per-task providers, so you only swap the provider name in config.
Measured baseline (README)
On 156 real commands from session dumps, typesafe-jev averaged 405 ms per gate vs 3968 ms for a small auxiliary chat model, with human prompts dropping from 42 to 10. Only about 11% of commands reach the gate at all.
FAQ
- What is Hermes Jev?
Community shorthand for this Hermes plugin: Jev as the smart approval auxiliary for shell commands. Search terms hermes jev and jev hermes usually point here.
- Which Jev primitives does it use?
One Choice for verdict, Score for blast_radius, and four Noul questions (policy_allows, self_advocating, reads_secrets, sends_outbound) fired in parallel per README table.
- Is the 9.8x speedup universal?
No. README and docs/METRICS.md position it as a baseline against one configured auxiliary model. An independent v0.2.1 sandbox study reported roughly 1.24x reviewer-time ratio instead.
- Do I need operator policy text?
Yes for developer machines. Without approvals.smart_policy, benchmarks showed routine kills and git pushes escalated or denied. Policy clauses are trusted-channel text Hermes already supports.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.