feat: implement thinking loop detection with automatic handler

[M0dEx]

Issue for this PR

#12716 - "Doom loop is not caught when during reasoning or output"
The doom loop guard previously only caught tool call loops, not reasoning/output loops. This branch adds detection for repetitive reasoning blocks.
#14343 - "Kimi K2.5 frequently gets in a reiteration loop"
Kimi K2.5 specifically gets stuck repeating the same reasoning patterns. The thinking loop detector catches these repetitions and intervenes.
#15271 - "Models get stuck in reasoning loop - making changes only in thinking, not actually executing tool calls"
Models plan changes in their thinking block but never actually call tools. The detector identifies when reasoning becomes repetitive and sends a reminder or compacts context.
#3743 - "Loop in certain models"
General issue with models (including GLM) getting stuck in loops.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Some models (especially GLM-5 and Kimi K2.5) get stuck in reasoning loops where they repeat the same thoughts over and over without actually doing anything. This PR adds detection for that behavior.

When the model starts repeating a block of text (20-1000 chars) 3+ times in a row, we:

1. Send a system reminder telling it to stop looping and take action
2. If it keeps looping, compact the context to reset it
3. If that fails, abort with an error

This is configurable via the experimental.thinking_loop config section. It's on by default since models that loop waste tokens and never complete tasks.

How did you verify your code works?

Tested on some workflows that caused the models to loop without it and saw successful nudging breaking the loops.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Found a potential related PR:

• PR fix: Doom-loop guard for repeated reasoning/output #12623 - "fix: Doom-loop guard for repeated reasoning/output"
  • This PR appears to address similar functionality for detecting doom loops in reasoning/output, which is closely related to the thinking loop detection feature in PR feat: implement thinking loop detection with automatic handler #15852. Both are focused on catching loops in model reasoning patterns and preventing models from getting stuck in repetitive thought cycles.

[bogzbonny]

Nice. Def want doc updates. Also It would be nice to see a comprehensive test which demonstrated the full workflow, I looked through the tests it didn't really look like that existed.

[M0dEx]

Nice. Def want doc updates. Also It would be nice to see a comprehensive test which demonstrated the full workflow, I looked through the tests it didn't really look like that existed.

Replaced by #21112, which implements this for text output as well, with integration tests.