fix: strip think blocks from reasoning field in extract_content_or_reasoning

[Ifkellx]

Summary

Fixes session titles showing raw think block content instead of the generated title.

Changes

• extract_content_or_reasoning: now strips <think> blocks from message.reasoning and reasoning_details before returning
• title_generator: uses extract_content_or_reasoning() instead of direct .content access for consistent handling
• 4 new tests covering think block stripping

Root cause

When the auxiliary model returned the title in the message.reasoning structured field instead of message.content, the fallback path returned raw reasoning text including <think> blocks without cleaning them.

Testing

python -m pytest tests/tools/test_llm_content_none_guard.py tests/agent/test_title_generator.py -v

[teknium1]

Thanks for identifying the structured-reasoning title path. The premise remains valid on current main: agent/title_generator.py:90 reads only message.content, while agent/auxiliary_client.py:7006-7026 already defines the structured-reasoning fallback.

Problems

• The added fallback at PR agent/auxiliary_client.py:3682 returns joined.strip() when tag removal yields an empty string. A reasoning field containing only `` therefore still leaks the raw block.
• The title fix is coupled to unrelated MiniMax Coding Plan vision, provider, and web-search work (agent/auxiliary_client.py:2435, tools/web_tools.py:998 in the PR diff). Current main has a user model-provider plugin discovery surface at providers/__init__.py:163-165.

Suggested changes

• Split out the focused title/extractor/test change.
• Return the cleaned structured-reasoning value even when empty, and test reasoning containing only a think block.
• Reuse the canonical scrubber behavior used by agent/title_generator.py:97-98.

Automated hermes-sweeper review.

[teknium1]

When joined contains only a paired think block, re.sub(...).strip() is empty and this fallback returns the original raw block. Return the cleaned value instead, and add coverage for reasoning that is only ``.

[teknium1]

Thanks for the contribution. Closing on policy: Hermes does not post-process or repair model output. Extracting titles out of the reasoning field when a model returns content=None is compensating for a model/provider that isn't following the request format — that's the inference engine's job to fix, not ours to paper over. The same policy applies to the think-block stripping on the reasoning field.

The bundled MiniMax Coding Plan vision client is a separate feature that shouldn't ride in a title-fix PR — if you want to pursue it, please open it as its own focused PR so it can be evaluated on its own merits.