Usage observability for Hermes
quota-hermes
Expose quotas, windows, balances, context and qualified cost without sending provider credentials to the browser.
01
The problem
AI provider limits and costs are heterogeneous. Missing data must never become a fake zero, and provider keys must not flow into the desktop widget.
02
My role
Generalising an existing fork, adding providers and focused-session context, hardening distribution, testing installation and publishing the open-source release.
03
Key decisions
- Query providers from the local backend only; the widget reads a credential-free cache.
- Qualify cost as real, API-equivalent, subscription-included or unavailable.
- Degrade each provider independently on failure.
- Publish a clean history with upstream attribution and autonomous CI.
04
Demonstrable outcome
Version 1.0.4 is public, installable in single-host or split gateway/desktop modes and verified by CI. Updates stay deliberate: no silent auto-update mechanism is included.
05
Technologies
AI assistants accelerated parts of the design and development work. Network/security decisions, acceptance criteria, testing and validation remain human-led and human-verified.