@rynfar/meridian-plugin-hermes-scrub

A Meridian plugin that strips Hermes Agent's coding-harness fingerprint from the system prompt before it reaches Claude.

Why

Hermes Agent (by Nous Research) can talk to an Anthropic-compatible endpoint via api_mode: anthropic_messages, so you can point it at Meridian and use your Claude Max subscription. Its system prompt contains self-management tool identifiers such as session_search and skill_manage.

Those identifiers have triggered Extra Usage metering in live Hermes requests. Once Extra Usage is depleted, affected requests fail with:

400 invalid_request_error: You're out of extra usage. Add more at claude.ai/settings/usage and keep going.

The plugin changes the distinctive underscored identifiers to spaced prose (session_search → session search). Live parent and subagent checks showed the requests succeeding under Max after that change. The prompt's guidance and structure remain intact:

• the user-editable persona (# Hermes Agent Persona, "You run on Hermes Agent…")
• the finishing-the-job, memory, session-recall, and skill guidance
• mid-turn user steering instructions
• the full <available_skills> list
• tools, guidelines, and any user- or harness-appended content

The scrub is content-scoped (it runs on every adapter and changes only the listed identifiers) and idempotent (running it twice is a no-op). Prompts without those identifiers are unchanged.

Install

Option 1: Local clone (recommended for dev)

git clone https://github.com/rynfar/meridian-plugin-hermes-scrub.git ~/repos/meridian-plugin-hermes-scrub
cd ~/repos/meridian-plugin-hermes-scrub
npm install
npm run build

Then point Meridian's plugin config at the built file:

mkdir -p ~/.config/meridian
# add this entry to ~/.config/meridian/plugins.json:
#   { "path": "/Users/YOU/repos/meridian-plugin-hermes-scrub/dist/index.js", "enabled": true }

Restart Meridian (or curl -X POST http://localhost:3456/plugins/reload).

Verify at http://localhost:3456/plugins — you should see hermes-scrub listed as active.

Option 2: Drop-in file

Symlink or copy dist/index.js into ~/.config/meridian/plugins/ for auto-discovery:

ln -s ~/repos/meridian-plugin-hermes-scrub/dist/index.js ~/.config/meridian/plugins/hermes-scrub.js

Point Hermes at Meridian

In ~/.hermes/config.yaml:

model:
  provider: custom
  base_url: http://localhost:3456     # your Meridian instance
  api_mode: anthropic_messages
  default: claude-opus-5-5            # or any model Meridian maps
  api_key: meridian-local            # any non-empty value; Meridian uses your Max auth

Behavior

Input	Output
No system prompt	unchanged
System prompt without the targeted identifiers	unchanged (idempotent)
Hermes' default or subagent system prompt	Targeted tool identifiers changed to spaced prose; persona, guidance, steering, and skills preserved

Test

bun test

License

MIT