← Hacker News

Hermes.md in Git commit messages causes requests to route to extra usage billing

Hacker News · adunk · April 25, 2026

Detailed Analysis

A reproducible bug in Anthropic's Claude Code integration causes API requests to fail with a spurious 400 error citing quota exhaustion whenever the string "Hermes.md" appears in Git commit messages or related prompt content. Rather than reflecting genuine usage limits, the error is a misclassification within Anthropic's request-routing and content-filtering infrastructure. Affected users are incorrectly directed to billing settings, suggesting the system is conflating benign file references with some trigger condition tied to the extra_usage billing pool — a separate allocation bucket reserved for third-party OAuth-authenticated clients. The issue has been documented in open GitHub issues across both the NousResearch and Anthropic repositories, and as of late April 2026 remains unresolved.

The root cause appears to lie at the intersection of two distinct systems: Anthropic's OAuth proxy pathway and the content-filtering layer that screens requests before they reach the model. When tools like Hermes Agent — an open-source AI agent developed by Nous Research — authenticate via Anthropic's OAuth flow rather than a standard API key, their requests are routed to a distinct billing pool that is either empty by default or carries a significantly lower quota ceiling. Hermes Agent's system prompts frequently reference internal configuration files such as Hermes.md, SOUL.md, and MEMORY.md as part of its operational scaffolding. When those references pass through the OAuth proxy path, they appear to trip a misclassification that the system interprets as an extra-usage exhaustion event, generating false error messages even when the user's actual account limits are nowhere near exceeded.

The practical impact on developers is significant. Users integrating Hermes skills within Claude Code — Anthropic's CLI-based coding assistant — encounter blocking errors during routine operations like writing Git commit messages, a workflow step that should be entirely benign. The false error message actively misdirects users toward billing pages, introducing friction and eroding trust in the toolchain. Published workarounds include avoiding the literal string "Hermes.md" in any content routed through Anthropic proxies, switching from OAuth-based authentication to direct API key authentication to bypass the extra_usage pool entirely, and ensuring Hermes Agent installations are updated to the latest version via configuration files such as `~/.hermes/auth.json`.

This incident highlights a broader and increasingly relevant challenge in the AI tooling ecosystem: the compounding complexity that emerges when third-party agent frameworks are layered on top of foundation model APIs. As open-source projects like Hermes Agent build sophisticated agentic behaviors on top of Anthropic's infrastructure, they expose edge cases in authentication, billing architecture, and content moderation that neither party fully anticipates. The separation of OAuth-routed traffic into a distinct billing pool, while presumably designed for access-control or monetization purposes, creates a secondary quota surface that is poorly documented and easily exhausted — or, in this case, falsely reported as exhausted due to a filtering mismatch.

The bug also underscores ongoing tensions in how AI providers manage content filtering at the infrastructure level versus at the model level. A string like "Hermes.md" carries no inherently problematic semantic content, yet its presence in a request is sufficient to trigger a system-level rejection under specific routing conditions. As agentic AI systems grow more complex — passing richer context, referencing internal files, and operating across multi-step pipelines — the surface area for such misclassifications expands considerably. Anthropic and other foundation model providers will face mounting pressure to make their request-routing logic, billing pool structures, and content-filter triggers more transparent and developer-accessible, particularly as third-party ecosystems built on their APIs mature and scale.

Read original article →