Channel LLM call fails with "Already borrowed" when messages arrive during active branch/worker streams

Spacebot had about 5-6 tasks running simultaneously when these errors occured. I had the model stop them and then had the model run a two simultaneously and didn't get the error. I had Claude dig into the errors deeper and this is what Claude formulated.

When a user sends a message to the channel while branches or workers are actively streaming LLM
responses, the channel's own LLM call fails with:

CompletionError: ProviderError: OpenAI-compatible streaming error: Already borrowed

Environment

• Provider: OpenAI-compatible (Chutes AI, https://llm.chutes.ai)
• Channel model: Qwen/Qwen3-235B-A22B-Instruct-2507-TEE
• Worker model: zai-org/GLM-5-Turbo
• Messaging: Telegram

ERROR handle_message: spacebot::agent::channel: message=channel LLM call failed
channel_id=
error=CompletionError: ProviderError: OpenAI-compatible streaming error: Already borrowed

Notes

The error does not originate in spacebot's own code — it appears to come from a shared mutable reference
(RefCell) in the OpenAI-compatible streaming client path, likely in rig or the HTTP streaming layer.
Concurrent streaming requests seem to share state that should be isolated per-call.

Start by reproducing the failure with multiple branch or worker streams active while a channel message arrives. Inspect the channel LLM call and the OpenAI-compatible streaming client path, particularly shared mutable state in the streaming layer. Done means concurrent streams no longer produce an “Already borrowed” error and the behavior is covered by an appropriate regression test.