Infinite CoT (Chain of Thought) Loop / Failure to converge Claude Code
#26
by
Svyatoblood
- opened
Environment
- CLI:
Claude Code - Issue: Infinite CoT (Chain of Thought) Loop / Failure to converge
- Reference: News 2026-01-12 - Reasoning Support
Severity
🔴 Blocker - The model is effectively unusable with Reasoning enabled.
Description
The mimo-v2-flash model exhibits a critical failure in its reasoning termination logic. Unlike expected behavior where the model thinks and then answers, this model enters an infinite semantic loop inside the thinking block until the hard token limit is hit and the API kills the connection.
Crucially: This happens regardless of the MAX_THINKING_TOKENS or CLAUDE_CODE_MAX_OUTPUT_TOKENS settings.
- Set to 16k? -> Loops for 16k tokens -> API Error.
- Set to 128k? -> Loops for 128k tokens -> API Error.
Reproduction Case
- Send a request requiring logical analysis (e.g., code debugging).
- Enable "Thinking" mode (any budget).
- Observe the output stream.
Actual Result (The Loop of Death)
The model repeats the same diagnostic steps without ever attempting to write the final response.
[Thinking] ...Let's check hkslSetTagForFrame...
[Thinking] ...Maybe the flag is not set...
[Thinking] ...Let's check hkslSetTagForFrame... (Identical repetition)
[Thinking] ...Maybe the flag is not set... (Identical repetition)
...
[System] API Error: Claude's response exceeded the 128001 output token maximum. To configure this behavior, set the CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable.