Bug type

Crash (process/app exits or hangs)

Beta release blocker

No

Summary

[Bug]: Telegram agent hangs with openai/gpt-5.4-mini; chat-completions fallback sends unsupported reasoning_effort

Environment:

• OpenClaw: 2026.4.29
• macOS: 26.3.1 arm64
• Node: 23.11.0
• Gateway: LaunchAgent, local loopback, Telegram channel in polling mode
• Telegram bot can receive updates and send direct sendMessage successfully
• OpenAI API key works via direct curl

Summary:
Telegram messages reach the OpenClaw gateway and route to the configured agent, but no Telegram reply is produced when using openai/gpt-5.4-mini. Direct OpenAI API calls work. The failure appears to be inside OpenClaw’s OpenAI transport/model routing rather than Telegram auth or the API key.

Observed failures:

• Built-in provider route openai/gpt-5.4-mini uses modelApi=openai-responses and streamStrategy=openai-websocket, even with transport=sse and openaiWsWarmup=false.
• Runs eventually fail with LLM idle timeout: no response from model.
• Custom provider route openai-chat/gpt-5.4-mini using openai-completions avoids the websocket path, but OpenClaw sends function tools with reasoning_effort and OpenAI returns:
  400 Function tools with reasoning_effort are not supported for gpt-5.4-mini in /v1/chat/completions. Please use /v1/responses instead.

Expected:
Telegram inbound messages should receive an assistant reply, or OpenClaw should honor transport=sse / avoid invalid reasoning_effort payloads for gpt-5.4-mini.

Workaround:
Switching the agent to a non-reasoning chat-completions model such as openai-chat/gpt-4.1 appears to avoid the bad path.

Relevant logs:

• [llm-idle-timeout] openai/gpt-5.4-mini produced no reply before the idle watchdog
• streamStrategy=openai-websocket while transport=sse
• 400 Function tools with reasoning_effort are not supported for gpt-5.4-mini in /v1/chat/completions
• Telegram provider starts and sendMessage works, so the channel itself appears connected

Steps to reproduce

1. Install OpenClaw 2026.4.29 on macOS arm64 with Node 23.11.0.
2. Configure the gateway as a LaunchAgent.
3. Configure Telegram channel in polling mode with a valid bot token.
4. Configure an OpenAI API key that works with direct curl calls to the OpenAI API.
5. Set the default/Telegram-routed agent model to openai/gpt-5.4-mini.
6. Set agents.defaults.params.transport = "sse" and agents.defaults.params.openaiWsWarmup = false.
7. Send a direct message to the Telegram bot.
8. Observe that the bot shows typing intermittently but no useful assistant reply arrives.
9. Inspect the OpenClaw trajectory/logs and observe streamStrategy: "openai-websocket" despite transport: "sse", followed by an LLM idle timeout.
10. Add a custom OpenAI-compatible provider using api: "openai-completions" and model gpt-5.4-mini.
11. Route the Telegram agent to openai-chat/gpt-5.4-mini.
12. Send another message or run an agent probe.
13. Observe OpenAI rejects the request with 400 Function tools with reasoning_effort are not supported for gpt-5.4-mini in /v1/chat/completions. Please use /v1/responses instead.

Expected behavior

Telegram inbound messages should receive an assistant reply, or OpenClaw should honor transport=sse / avoid invalid reasoning_effort payloads for gpt-5.4-mini.

Actual behavior

Built-in provider route openai/gpt-5.4-mini uses modelApi=openai-responses and streamStrategy=openai-websocket, even with transport=sse and openaiWsWarmup=false.

• Runs eventually fail with LLM idle timeout: no response from model.
• Custom provider route openai-chat/gpt-5.4-mini using openai-completions avoids the websocket path, but OpenClaw sends function tools with reasoning_effort and OpenAI returns:
  400 Function tools with reasoning_effort are not supported for gpt-5.4-mini in /v1/chat/completions. Please use /v1/responses instead.

OpenClaw version

OpenClaw: 2026.4.29

Operating system

macOS: 26.3.1 arm64

Install method

No response

Model

openai/gpt-5.4-mini

Provider / routing chain

Telegram channel (polling) -> OpenClaw gateway LaunchAgent -> binding routes channel=telegram/accountId=* to agent=clawbot -> OpenAI provider

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response