Hermes Safe Mode for Isolated Troubleshooting

Safe mode helps answer a basic debugging question: is the failure in Hermes core or in this installation’s customization? It disables user configuration and injected rules, plus plugins, shell hooks, and MCP servers for one isolated run.

Official documentation: https://hermes-agent.nousresearch.com/docs/reference/cli-commands

Start with the smallest reproduction

Use hermes chat --safe-mode -q "<minimal reproduction>" for a bounded test. If the problem disappears, add layers back deliberately: provider configuration, project context, skills, plugins, hooks, then MCP servers. If it remains, collect the exact command, version, operating system, and error without dumping secrets.

Related flags such as --ignore-user-config and --ignore-rules isolate narrower layers. Safe mode combines them with plugin, hook, and MCP suppression, making it the stronger first bisect for customization-related failures.

Reproduction checklist

  • Reduce the failure to one command or prompt.
  • Record Hermes version and platform.
  • Run the same test normally and in safe mode.
  • Keep provider and model differences visible.
  • Avoid production mutations during diagnosis.
  • Add customization layers back one at a time.
  • Capture only redacted logs and stable error text.
  • Confirm the final fix outside safe mode.

Pitfalls

  • Calling safe mode a fix rather than a diagnostic control.
  • Comparing runs that use different providers or working directories.
  • Re-enabling several extensions at once and losing the signal.
  • Testing a destructive production flow merely because customization is disabled.
  • Sharing a debug bundle without reviewing it for private paths and credentials.

Verification steps

  1. Reproduce the issue in the ordinary session.
  2. Repeat the minimal case with --safe-mode.
  3. If it clears, test --ignore-user-config and --ignore-rules separately.
  4. Re-enable plugins, hooks, and MCP servers in controlled groups.
  5. Identify the first layer that restores the failure.
  6. Correct or remove that layer and restart normally.
  7. Re-run the original user path and its verification checks.