Claude Code HERMES.md Bug Bills Extra Usage

This is the kind of billing bug that destroys trust faster than it costs money: users expect prompt content to affect model behavior, not whether they get billed on a subscription or metered API basis.

• –The reported trigger is alarming because `HERMES.md` is a normal convention in agent workflows, not a malicious payload.
• –The failure mode is especially bad because users see “out of extra usage” even when subscription headroom remains, which makes diagnosis and support loops painful.
• –Anthropic’s Claude Code docs already expose `CLAUDE.md`, memory files, and cost tracking, so this points to fragile server-side routing around context ingestion.
• –If you run Claude Code with Hermes-style project conventions, it’s worth auditing recent commits and prompt/context files until Anthropic clarifies the bug.