OpenAI Responses WebSocket hangs after a top-level error event #50848

Description

The native OpenAI Responses WebSocket route emits a provider-error for a top-level error event but never completes the response stream. The same transport completes normally for response.failed.

packages/llm/src/protocols/openai-responses.ts handles both event types, but TERMINAL_TYPES only includes response.completed, response.incomplete, and response.failed. Because a Responses WebSocket remains open between requests, waiting for the socket to close leaves LLMClient.stream() and LLMClient.generate() pending indefinitely after an API error.

Expected: a top-level error event emits its provider-error and terminates the current response stream.

Plugins

None

OpenCode version

dev (18ef3cc7c5, package version 1.18.32)

Steps to reproduce

1. Configure the OpenAI Responses WebSocket route with an in-memory socket.

2. After response.create, send this frame and keep the socket open:

   {"type":"error","code":"rate_limit_exceeded","message":"Rate limit reached"}

3. Collect LLMClient.stream() with a timeout.

4. Observe provider-error(rate_limit_exceeded: Rate limit reached), followed by a timeout because the stream remains open.

5. Replace the frame with {"type":"response.failed","response":{"error":{"code":"server_error","message":"boom"}}} and observe that the stream ends immediately.

This reproduces without a network request.

Screenshot and/or share link

N/A

Operating System

Linux

Terminal

N/A (in-process reproduction)