Hermes Agent Community Reports Plugin Loading Issue with Custom Directories; Gateway May Omit Traces

A Hermes Agent issue report published on October 30 describes a possible observability gap in multi-environment deployments: the command line shows a plugin as enabled, but the gateway that processes messages does not import it. The reported environment was macOS, with a source installation of v0.22.5+2381.g0431192; separate data directories were used for development and production, and the Langfuse plugin in production did not generate the expected traces. Issue report

According to diagnostic logs added by the author, startup initially read the production environment’s enabled-plugin list but later read settings from the default directory. The bundled plugin could still be found, making the problem harder to spot from the overall startup status. The author attributes the issue to a switch in the effective directory during initialization and believes that, in an asynchronous context, a directory override takes precedence over the environment variable. However, this remains the reporter’s diagnosis. As of verification, the issue was still open, with no public fix linked. Diagnostics and issue status

The official documentation explains that profiles use HERMES_HOME to separate settings, memory, conversations, and gateway state; it serves a different purpose from the operating system’s HOME. The plugin development documentation treats enablement settings, module imports, and registration through register(ctx) as distinct steps, and provides discovery logs and isolation tools for diagnosis. Therefore, a successful installation or an enabled status in the list does not by itself prove that the persistent gateway has run the plugin’s callback. Profile configuration, plugin development documentation

For teams that rely on traces to debug or evaluate agents, this kind of failure could leave tool-execution records incomplete. A practical engineering check is to verify the effective data directory, loading logs, and delivery of test events to the tracing backend from within the actual service process. This is a verification direction suggested by the documentation and case, not a confirmed fix. Follow-up should focus on whether maintainers provide a minimal reproduction and whether regression tests cover plugin discovery and callback execution across different profiles.