Think-token looping / raw (unrepaired) tool-call failures on DeepSeek V4.1 Flash & GLM 5.3 #967

Summary

Using Command Code with DeepSeek V4.1 Flash and GLM 5.3, I'm consistently seeing the model get stuck in a think-token loop and tool calls surface as raw failures instead of being repaired. The same models run correctly and quickly through opencode go with no looping and no tool-call issues, so this doesn't look like a client-side/config problem on my end.

Expected Behavior

Tool calls should be repaired the same way they are for other models served by Command Code, and the model should not loop indefinitely on think tokens.

Actual Behavior

• Model loops on think tokens.
• Tool calls fail and are surfaced raw, without going through the repair step.

Steps to reproduce the issue

1. In Command Code, select DeepSeek V4.1 Flash (or GLM 5.3) via /model.
2. Provide an article and ask the model to revise/rewrite its content.
3. Observe the model enters a think-token loop and any tool calls fail raw (not auto-repaired).

Command Code Version

1.73.2

Operating System

Windows

Terminal/IDE

Unknown

Shell

cmd.exe

Session file (optional)

No response

Fix prompt (optional)

No response

Additional context

• Ruled out stale session state — issue persists in a fresh session.
• Confirmed the same models run fine (fast, no looping, tool calls succeed) via opencode go, suggesting the issue is specific to how Command Code serves these two models rather than the models themselves.