openclaw - 💡(How to fix) Fix [Bug]: 2026.4.29: Control UI accepts prompts but renders no assistant output, even after clean reinstall (OpenAI Responses / gpt-5-mini) [2 comments, 2 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

In OpenClaw 2026.4.29 (a448042), the Control UI accepts prompts and shows brief typing activity, but no assistant reply ever appears.

I spent quite a while narrowing this down. The important part is that this still reproduces after a clean reinstall of the same version with a minimal plugin/config baseline, which makes this look like an OpenClaw runtime/output-path bug rather than bad local config.

Environment

• OpenClaw version: 2026.4.29 (a448042)
• Node: v22.22.1
• Install type: global npm install under /usr/lib/node_modules/openclaw
• Runtime OS: Linux VM
• Gateway: root-owned instance under /root/.openclaw
• Reverse proxy: nginx -> 127.0.0.1:18789
• Model used for repro: openai/gpt-5-mini
• Minimal plugin baseline after reinstall: openai + telegram

User-visible symptom

From the Control UI:

• prompt is accepted
• typing dots appear briefly
• then nothing is rendered
• no assistant text appears at all

In earlier phases of debugging, some hidden prompt-error events were recorded even when the UI looked blank.

Clean reinstall repro

I then did a clean reinstall/rebuild of 2026.4.29 and reproduced again.

Clean baseline after reinstall:

• rebuilt service as proper root systemd unit
• loopback bind behind nginx
• explicit gateway.controlUi.allowedOrigins
• gateway.trustedProxies = ['127.0.0.1', '::1']
• default model openai/gpt-5-mini
• timeout 300
• only openai + telegram enabled

After that clean rebuild:

1. open Control UI
2. send a tiny prompt like reply with exactly OK
3. UI shows activity but no visible assistant reply ever appears

What was ruled out

Not basic OpenAI/network failure

From the same VM, with the same API key/environment:

• curl to OpenAI worked
• direct authenticated responses API POST worked
• OpenClaw’s bundled OpenAI SDK worked outside OpenClaw
• both of these worked outside OpenClaw:
  • client.responses.stream(...)
  • client.responses.create({ ..., stream: true })

So the VM can talk to OpenAI fine, and streaming works outside OpenClaw.

Not stale poisoned session context

I reset the stuck main session and verified a fresh zero-context session on gpt-5-mini. The bug still reproduced.

Not primarily nginx / websocket access

There was an earlier access problem (1006) due to Control UI connection details, but that was fixed well enough to regain UI access. After that, the remaining bug persisted on the clean rebuild.

Not just plugin scan/event-loop starvation

I found and reduced a lot of bundled-extension scan churn during debugging. That improved responsiveness/startup behavior, but the blank-reply bug still reproduced afterward, including on the clean reinstall with a minimal plugin set.

Strongest technical clue

Inside OpenClaw’s runtime path, the request reaches the model layer, but the stream never seems to yield usable events/chunks to the assistant/output path.

During debugging on 2026.4.29, I instrumented the runtime and consistently saw behavior like:

• client.responses.create(...) returned
• OpenAI HTTP response status was 200
• content-type: text/event-stream; charset=utf-8
• response body existed as a ReadableStream
• but no first chunk / no first event was observed in the OpenClaw path within the expected window

At the same time, the equivalent OpenAI SDK calls worked correctly outside OpenClaw on the same machine.

That makes this look like a bug in OpenClaw’s runtime/stream integration or assistant output handling on this build, not raw provider connectivity.

Additional observations

In prior runs, session transcript artifacts recorded hidden errors such as:

• LLM idle timeout (120s): no response from model | ...
• request timed out | request timed out

even when the UI only showed “typing then nothing”.

I also found that provider:"openclaw", model:"gateway-injected" messages were present in session files, but those appeared to be gateway-generated status/transcript messages, not the actual selected model for the failing run.

Regression/workaround note

I eventually restored the VM from snapshot to bring the older pre-upgrade instance back online, and the older instance is working again. So for now the practical workaround is rollback.

Steps to reproduce

Repro

Before clean reinstall

This reproduced even after:

• forcing a fresh/low-context main session
• switching default model to openai/gpt-5-mini
• lowering thinking
• increasing timeouts
• resetting the wedged agent:main:main session state

Expected behavior

Assistant text should appear in the Control UI for trivial prompts.

Actual behavior

Prompt is accepted, brief typing indicator may appear, but no assistant message is rendered.

OpenClaw version

2026.04.29

Operating system

Ubuntu 24.02

Install method

npm

Model

chatgpt-5.4

Provider / routing chain

openclaw-->nginx-->LAN

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response