Fix _mentioned_agents() to ignore reasoning tags in speaker selection

[veeceey]

Summary

Fixes #6891

When LLMs use reasoning tags like <thinking>, <reflection>, <planning>, etc., agent mentions inside these blocks should not affect speaker selection. This PR filters out common reasoning blocks before counting agent mentions.

Problem

The issue occurred because models like Qwen would mention agents in their internal reasoning (e.g., "<thinking>Maybe AgentA or AgentB could help</thinking>I suggest AgentC"), and the selector would incorrectly count those mentions, leading to wrong speaker selection.

Solution

• Added _strip_reasoning_blocks() method to remove common reasoning tags:
  • thinking, thought, reflection, reasoning
  • analysis, internal, scratch, planning
• Updated _mentioned_agents() to use cleaned content before counting
• Added comprehensive test case

Testing

• Manual unit tests confirm correct behavior:
  • Reasoning blocks are stripped before counting
  • Multiple tag types are handled
  • Case-insensitive matching
  • Tags with attributes are supported
• Added test case test_selector_group_chat_ignores_thinking_tags

Example

Before:

<thinking>Maybe agent1 or agent2 could help</thinking>I suggest agent3

Would count: agent1=1, agent2=1, agent3=1 (incorrect)

After:
Would count: agent3=1 only (correct)

[veeceey]

All checks passing, DCO signed, ready for merge

[veeceey]

Hey @ekzhu, I noticed your comments on #6891 about preferring structured output / JSON mode and the family: "r1" approach for handling reasoning tags. Those are definitely the better long-term solutions.

That said, this PR is meant to handle the case where users haven't configured those options -- when they're just using a model that happens to emit thinking tags (like Qwen with default settings), the speaker selector still picks up those mentions and makes bad choices. This is basically a fallback for the "unstructured" path.

Totally understand if you'd prefer to just point users toward structured output / family: "r1" instead of adding this. Happy to close this if that's the direction you want to go. Just wanted to offer it as a safety net for the common case.

[veeceey]

Friendly ping @ekzhu - just checking if you had a chance to see my earlier comment about the approach here. Totally understand if you'd prefer to point users toward structured output / family: "r1" instead -- happy to close this if that's the preferred direction, or make any changes you'd like. Thanks for your time!

[ekzhu]

@veeceey I don't work at Microsoft anymore.

Cc @victordibia

[veeceey]

Thanks for the heads up @ekzhu! @victordibia would appreciate a look at this whenever you get a chance.

[veeceey]

Rebased on latest main. @victordibia any thoughts on this one? happy to adjust the approach if you'd prefer a different direction.

[veeceey]

hey @victordibia, just a friendly bump on this - any thoughts on the approach?