HN fixates on the HERMES.md billing bug, then on what it says about trust

Hacker News did not treat this as a funny parser glitch for long. The hook was absurd enough on its own: according to a reproducible GitHub issue, if recent git history contains the exact case-sensitive string HERMES.md in a commit message, Claude Code can route requests to paid extra usage instead of the included Max plan quota. But the HN thread quickly moved past the novelty. What people cared about was the trust boundary. If commit-message text can silently change billing behavior, users have almost no chance of spotting the cause before money disappears.

The report is unusually concrete. The issue says a Max 20x subscriber on Claude Code v2.1.119 burned $200.98 in extra credits while weekly plan capacity still showed mostly unused headroom. The trigger was not a file named HERMES.md on disk. It was the string in commit history. The repro was simple: commit with add HERMES.md, run Claude Code, and hit an extra-usage error; commit with lowercase hermes.md, run the same command, and it works. The reporter's read is that Claude Code includes recent commits in the system prompt and some server-side routing path changes when that exact token appears.

That alone would have been enough for HN, but the support response made the discussion much hotter. Early replies in the issue said compensation was not available for technical errors that caused incorrect billing routing, and HN commenters immediately treated that as the real scandal. Community discussion noted that Anthropic engineer Thariq Cheriyan later stepped into the thread and said affected users would get a full refund plus extra credits equal to a monthly subscription. Even then, the tone on HN stayed sharp. The takeaway was not “strange bug, glad it is fixed.” It was that support, billing, and engineering escalation looked badly out of sync at exactly the moment users most needed observability.

That is why the story landed harder than a normal SaaS billing complaint. Claude Code is a developer tool that reads repo context by design, so the line between “helpful context” and “context that changes backend behavior” matters a lot. HN read this bug as a warning that agent tooling still has hidden control surfaces ordinary software rarely exposes: prompt contents, routing layers, plan enforcement, and support automation all stacked together. When one invisible string can flip a request onto a different meter, the community stops asking whether the model is smart and starts asking whether the system around it is safe to trust.