Summary

The new API server is excellent for OpenAI-compatible chat clients, but there is still a gap for app-embedded integrations that need structured agent lifecycle events, not just final text.

Request

Please consider a first-class API mode for event-preserving integrations, including:

• structured streaming events for:
  • reasoning.available
  • step.started
  • tool.started
  • tool.completed
  • interaction.requested
  • interaction.resolved
  • message.completed
  • run.completed
  • run.failed
• resumable human-in-the-loop interactions:
  • surface clarify / approval as a pending interaction
  • allow the client to reply and resume the same run
• stable run_id / conversation_id
• per-run or per-conversation MCP/tool configuration:
  • tool allowlists
  • auth headers
  • ideally without forcing a custom fresh-process wrapper per turn

Why

This would make Hermes much easier to embed in operational apps and custom workflow UIs that need to:

• show live tool progress
• render tool outputs as native UI components
• pause for clarification/approval
• resume the same run after the user responds

Right now this seems to require a custom gateway/adapter around Hermes internals.

Possible shapes

No strong opinion on transport. Any of these could work:

• extend the current API server
• add a /runs API with SSE or WebSocket events
• extend the Responses-style API with structured intermediate events

Question

Is something like this already planned, especially for clarify / approval round-trips and per-conversation MCP isolation?