Bug Description

When using the OpenAI-compatible POST /v1/responses endpoint with:

{
  "stream": true,
  "store": true
}

the server currently only persists the response to ResponseStore after the streaming run completes successfully.

If the client disconnects mid-stream — for example app killed, network interruption, foreground/background transition, or SSE connection dropped — Hermes interrupts and cancels the agent task. The partially generated response is not stored, and subsequent:

GET /v1/responses/{response_id}

returns 404 Response not found.

This makes it impossible for clients to implement ChatGPT/Claude-like recovery after stream interruption, even though the client already received a response.created event with a valid response_id.

Steps to Reproduce

1. Start a streaming response:

curl http://127.0.0.1:8642/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $API_KEY" \
  -d '{
    "model": "hermes-agent",
    "input": "Do something that takes a while or calls tools.",
    "stream": true,
    "store": true
  }'

2. Wait until the client receives:

event: response.created
data: {"type":"response.created","response":{"id":"resp_xxx", ...}}

3. Optionally wait until some text deltas or tool events arrive:

response.output_text.delta
response.output_item.added
response.output_item.done

4. Disconnect the client before response.completed.

5. Try to retrieve the response:

curl http://127.0.0.1:8642/v1/responses/resp_xxx \
  -H "Authorization: Bearer $API_KEY"

Expected Behavior

When store=true, after the server emits response.created, the response should be queryable via:

GET /v1/responses/{response_id}

At minimum, the store should contain a recoverable snapshot with a meaningful status:

{
  "id": "resp_xxx",
  "object": "response",
  "status": "in_progress",
  "output": []
}

During streaming, the stored response should be updated as output arrives:

• response.output_text.delta
• response.output_text.done
• response.output_item.added
• response.output_item.done
• response.failed
• response.completed

If the client disconnects before completion, the server should continue running in background.

For store=true, do not cancel the agent task on SSE client disconnect. Let the response continue to completion and persist the final response to ResponseStore.

This gives ChatGPT/Claude-like recovery:

1. client disconnects
2. server continues generation
3. client relaunches
4. client calls GET /v1/responses/{response_id}
5. final response is available once completed

Actual Behavior

GET /v1/responses/{response_id} returns:

{
  "error": {
    "message": "Response not found: resp_xxx",
    "type": "invalid_request_error",
    "param": null,
    "code": null
  }
}

The client cannot recover:

• partial assistant text
• completed tool calls
• tool outputs
• final assistant message

The response id previously emitted by response.created is therefore not useful for recovery unless the stream reaches response.completed.

Affected Component

Gateway (Telegram/Discord/Slack/WhatsApp)

Messaging Platform (if gateway-related)

No response

Debug Report

Report       https://paste.rs/veGtu
  agent.log    https://paste.rs/3eRBA
  gateway.log  https://paste.rs/BJmsi

Operating System

ubuntu 26.04

Python Version

3.11.15

Hermes Version

0.11.0

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

No response

Proposed Fix (optional)

For store=true, decouple the agent task from the SSE connection:

• SSE connection is only a live transport
• response execution continues independently
• ResponseStore is the source of truth
• GET /v1/responses/{response_id} can return:
  • queued
  • in_progress
  • completed
  • failed
  • incomplete

Client disconnect should not necessarily cancel response execution if the user requested store=true.

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR