Hermes Ops
LiveSignature capability
The safety rules here are code, not prompt instructions, so they hold regardless of which model is running.
What it is
Hermes is an agent on a cloud host that runs scheduled reporting and operations jobs. This is the tested core underneath it: a pure functional core of nine modules with zero I/O, and one adapter module that is the only thing touching the outside world. The scheduled jobs run with the LLM out of the loop entirely for any number that gets reported.
Highlights
- The arm interlock is fail-closed: only the exact byte 1 means armed. Missing, empty, oversized, binary or merely truthy-looking values such as "true", "yes" and "01" all resolve to DISARMED. Uncertainty never permits trading.
- A partial read yields no total. If any capital leg fails to read, the function returns None, so there is no number to print. This is not a rule the renderer is asked to follow, it is an absence it cannot fill.
- It proved itself in production on 2026-08-11 during a real Arbitrum RPC failure: the brief printed PARTIAL READ instead of reporting half the capital as the total.
- Untrusted text never reaches a shell: the SSH helper returns an argv list, rejects shell strings outright and allowlists hosts.
- Mutation-checked: fail-open arm, truthy arm, partial-total, flipped hedge band and dropped SSH BatchMode were each injected and each caught by the suite.
- The README documents its own known limit, that the agent runs as root and can still write its own arm file, tracked as an open ticket.
Tech · Python 3, pure-core architecture with a single I/O adapter module, flock single-flight, pytest with property and security suites, mutation testing, cron.