Explain code call from codex disconnected from vllm server

[knyeoh]

Bug Title

codex disconnected to vllm server during tool call and did not process the server respond

Description

You can see the explain code being drop even we see the respond from vllm (10-03 18:57:08), open-responses-server does process the request at 2025-10-03 18:57:00,971. not sure what happen.

Steps to Reproduce

setup vllm v0.10.2 (window/linux) and run with this command:

vllm serve openai/gpt-oss-20b --max-model-len 130k --served-model-name gpt-oss-20b

start oct server (default vllm port is 8000 and oct default external port to 8080) :
oct start

setup codex with this configuration setting:

[model_providers.vllm]
name = "VLLM"
base_url = "http://localhost:8080/v1"

[profiles.gpt-oss-20b-vllm]
model_provider = "vllm"
model = "gpt-oss-20b"

start codex with this command:
codex --profile gpt-oss-20b-vllm

start in any code area, and ask codex to explain code (example):
explain COBSWebSerial.py

some time, it will process the request without dropping tool call, but most of the time it will. I am not sure how to debug, please advice.

Expected Behavior

we expect codex to wait for respond from vllm server, not just jump to next prompt.

Actual Behavior

codex did not wait for for the tool call to finish, but it just jump to next prompt input.

[knyeoh]

Not sure why, adding wireshark to track the http package seem to make the connection stable:

I try not using wireshark to capture the packet, it seem to back to random disconnect. why monitoring the port make the connection stable. I have no idea...

[knyeoh]

seem like change to unslot gpt-oss-20b model fix the problem, please close the ticket.

[OriNachum]

@knyeoh I prefer to keep it open and add observability.
You shouldn't have to open wiresharp to track it - thank you for reporting this.