Anthropic fixes Claude Code harness bug tied to `HERMES.md` and `git status`

• Anthropic said a Claude Code bug in its third-party harness detection pulled git status into the system prompt, and according to trq212's reply, affected users are getting refunds plus another month of credits.
• The incident became concrete when Jeremy Howard's retweet amplified a case where the string HERMES.md in git commits allegedly helped drive a user through about $200 in one day, while Teknium's follow-up said Hermes-related users were later covered by the refund plan.
• mattpocockuk's finding showed Auto Mode was also injecting extra instructions to make Claude Code more AFK, and Anthropic's April 23 postmortem had already acknowledged recent quality issues tied to product-side changes rather than a base-model swap.
• According to the main HN thread on the postmortem and fresh HN follow-up comments, users kept separating Claude Code's harness behavior from the underlying API, pointing to Auto Mode defaults, cache loss after quota resets, and safety gating as the parts that changed perceived quality.

You can read Anthropic's postmortem, the big HN discussion on recent Claude Code quality reports, and an earlier HN analysis of the Claude Code source leak. The odd detail in this bug is that git status and commit metadata could change both behavior and spend, while mattpocockuk's Auto Mode report suggests prompt-side orchestration was already reshaping how Claude Code behaved in sessions.

Anthropic's public explanation came from trq212's reply, which described "a bug with the 3rd party harness detection" and said Claude Code was pulling git status into the system prompt.

The user-visible trigger that spread across X was more specific. In Jeremy Howard's retweet, the claim was that the string HERMES.md, a real convention in the Hermes Agent ecosystem, was enough to push Claude Code into the wrong path and rack up roughly $200 in charges in a day.

That makes the bug less about one unlucky repository name and more about how much hidden client logic can key off local project state. The official wording only mentions git status, but the Hermes reports imply commit and workspace context were enough to alter routing and billing behavior.

Before the refund thread, mattpocockuk's post had already pinned down another Claude Code behavior that lived in orchestration, not the model itself: Auto Mode was injecting instructions into the system prompt to make the agent more AFK.

Pocock's complaint was that those extra instructions were interfering with his existing skills. That lines up with the broader pattern in Anthropic's postmortem, where recent regressions were described as product and configuration changes around the model.

The most useful reporting outside X came from the HN thread on Anthropic's postmortem and the fresh follow-up summary, where commenters kept drawing the same line: Claude Code felt worse because the harness changed.

Three recurring complaints showed up there:

• Quota resets plus cache loss could wipe context, so the next message spent a large chunk of the new session budget just rebuilding state, according to Folcon's comment.
• Auto Mode had become sticky enough that switching out of Plan mode could drop users back into Auto, while safety checks could block basic Bash flows, according to bashtoni's comment.
• Users reported Claude Code forgetting workflow instructions, including remote investigation habits, which joshstrange's comment tied directly to Auto Mode behavior.

The sharpest formulation came from xpe's comment, which argued that calling the models directly by API had not changed, but the Claude Code harness had, and users were not told clearly.

Anthropic's remedy was unusually explicit for a consumer-facing bug. In the original reply, the company said it was reaching out to affected users with refunds and another month of credits, and it gave one concrete example of an extra $200.

Teknium's update added a wider scope: Hermes users, or people who had used Claude Code while developing Hermes Agent, were also being refunded or credited. That matters because it suggests Anthropic treated the issue as a misclassification problem that hit a recognizable cluster of repos and workflows, not a one-off support ticket.

The blowback went past the refund mechanics. Gergely Orosz's post framed the episode as a transparency problem for closed harnesses, while Teknium's later post argued that fully auditable agents make it easier to inspect what is being tracked or flagged behind the scenes.

The Hermes bug landed a month after an HN analysis of the Claude Code source leak had already turned hidden tool prompts, telemetry heuristics, session compaction, hooks, and anti-distillation behavior into a live governance story.

That older thread surfaced a detail with a different flavor but the same architectural theme: one top comment on "Undercover mode" said leaked prompts told Claude Code not to mention either "Claude Code" or the fact that it was AI in commit messages or PR descriptions. Another comment on DMCA takedowns said even forks that allegedly did not carry the leaked code were getting pulled down.

The new bug does not depend on the leak to matter. But the leak had already trained technical users to look for exactly this class of behavior: prompt-time policy, hidden heuristics, and client-side rules that are invisible until they fail.