mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-16 19:36:54 +00:00
32 lines
1.6 KiB
YAML
32 lines
1.6 KiB
YAML
name: local-system-one-hermes
|
|
title: Local System One
|
|
repo: https://github.com/NMTZ-z/system-one-benchmark-lab
|
|
sha: 823aff58d97b989f3468dbd2c2b12526c6e6936e
|
|
subdir: integrations/hermes/local-system-one-hermes
|
|
description: >-
|
|
Local System One / Laya Typed Decisions control plane for Hermes Agent. Sends task text to a
|
|
loopback sidecar for reversible OFF/Shadow/Canary routing decisions, with explicit Canary
|
|
acknowledgement and fail-open behavior. Disclosure — executes nothing and does not start the
|
|
sidecar. In Shadow/Canary mode it POSTs each turn's task text (last user message up to 4000 chars,
|
|
or the owned Kanban card title/body) plus up to 2000 chars of recent user/assistant messages to the
|
|
configured sidecar URL (default loopback 127.0.0.1:8787; user-configurable to any host; no auth
|
|
header). Its only request mutations, in acknowledged Canary mode on the first provider call of a
|
|
turn, are removing the web_search/web_extract tools and (off by default) lowering reasoning_effort
|
|
for gemini-3.8-flash-tiered; plugin state is metadata-only under plugin-data/. The sidecar (same
|
|
repo, separate pip install, Apple-Silicon MLX/CoreML) is a text classifier that downloads a pinned
|
|
Hugging Face model.
|
|
maintainer: NMTZ-z
|
|
tier: community
|
|
category: automation
|
|
requires_hermes: ">=0.21.4"
|
|
version: "0.5.2"
|
|
docs_url: https://github.com/NMTZ-z/system-one-benchmark-lab/tree/823aff58d97b989f3468dbd2c2b12526c6e6936e/integrations/hermes/local-system-one-hermes
|
|
readme: true
|
|
platforms: [macos]
|
|
capabilities:
|
|
provides_tools: []
|
|
provides_hooks:
|
|
- pre_llm_call
|
|
provides_middleware:
|
|
- llm_request
|
|
requires_env: []
|