mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-06 05:17:15 +00:00
Vincent Attali
6dd515c45c
Removes the opt-in that could not be made safe and states the limitation plainly: an ambiguous 100 refuses, a false refusal is expected while a window sits at or under 1%, and the fix is upstream.
28 lines
3.1 KiB
YAML
28 lines
3.1 KiB
YAML
# Catalog entry for NousResearch/hermes-agent: plugin-catalog/usage-guard.yaml
|
|
# Pinned to tag v0.4.5 (release published 2026-10-26). Bump sha and version together.
|
|
name: usage-guard
|
|
repo: https://github.com/mister-ideas/usage-guard
|
|
sha: 2c3ebac27878cfa4b10836f2e38473d74e190446
|
|
description: "Keeps an agent inside its subscription allowance: refuses a provider dispatch once a subscription usage window is exhausted, so an account that allows paid extra usage stops at the wall instead of spending overage. Reads the usage Hermes already displays, holds no credentials, and never touches account or billing settings. Disclosure — refuses main-loop provider calls (including MoA turns) for the providers in `provider_keys` (Anthropic and OpenAI Codex by default) when a subscription usage window reaches `threshold_percent`, and by default (`fail_open: false`) also refuses when usage data is missing, stale, unreadable, or the guard's own config is malformed; usage is read through Hermes's own account-usage fetcher (no credentials held, at most one read per source every 180 seconds by default) and cached in $HERMES_HOME/plugin-data/usage-guard/usage.json; a `pre_llm_call` hook adds a short usage note to the next turn's context at 80% and 90% used by default, and an opt-in (off by default) `llm_request` middleware lowers an existing output-token budget near the usage limit; both read only the cache, never the network; auxiliary calls (compression, vision, titles, MoA advisors) are not gated. Borrowing is off by default: when `hydrate_from_paths` names files, the guard may read those JSON snapshots instead of reading the endpoint, validating every window and never writing them into its own cache."
|
|
image: https://raw.githubusercontent.com/mister-ideas/usage-guard/2c3ebac27878cfa4b10836f2e38473d74e190446/docs/banner.jpg
|
|
screenshots:
|
|
- https://raw.githubusercontent.com/mister-ideas/usage-guard/77c96af92bd3e144e4af9427c9b3772557f14205/docs/media/shot-refusal.png
|
|
- https://raw.githubusercontent.com/mister-ideas/usage-guard/77c96af92bd3e144e4af9427c9b3772557f14205/docs/media/shot-near-wall.png
|
|
- https://raw.githubusercontent.com/mister-ideas/usage-guard/77c96af92bd3e144e4af9427c9b3772557f14205/docs/media/shot-unavailable.png
|
|
- https://raw.githubusercontent.com/mister-ideas/usage-guard/77c96af92bd3e144e4af9427c9b3772557f14205/docs/media/shot-install.png
|
|
maintainer: mister-ideas
|
|
tier: community
|
|
category: automation
|
|
requires_hermes: ">=0.21.6"
|
|
version: "0.4.5"
|
|
known_issues:
|
|
- "Anthropic windows: Hermes's Anthropic fetcher scales any utilization <= 1 by 100, so a window at 0-1% can read as 100% and the guard refuses on it, while a window genuinely at 100% reads the same value. No in-plugin reading can separate the two without risking an exhausted window, so the guard refuses on the ambiguous value and a false refusal is expected while a window sits at or under 1% used; it clears as usage passes 1%. Fixed upstream by NousResearch/hermes-agent#139916 — this entry's requires_hermes rises to the release carrying it."
|
|
capabilities:
|
|
provides_tools: []
|
|
provides_hooks:
|
|
- pre_llm_call
|
|
provides_middleware:
|
|
- llm_execution
|
|
- llm_request
|
|
requires_env: []
|