fix: extract <think> tags from streaming content in agent and team paths

[hztBUAA]

Summary

Fixes workflow steps producing empty content when using thinking models (e.g., qwen3 with enable_thinking=True) served via OpenAI-compatible APIs.

Root cause: In the streaming path, models that embed reasoning in <think>...</think> tags (qwen3, deepseek-r1 via OpenAI-compatible endpoints like vLLM, DashScope) accumulate raw tags in run_response.content. The non-streaming path already extracted these tags at the provider level (_parse_provider_response), but the streaming path had no post-streaming extraction. This caused downstream workflow steps to receive content with embedded <think> tags, confusing the next model and producing empty output.

Fix: Add extract_thinking_content() after the streaming loop completes in both the agent and team response handlers (sync + async = 4 locations). A guard condition ("</think>" in content) ensures zero cost for non-thinking models.

Fixes #6305

Changes

File	Change
agno/agent/_response.py	Extract <think> tags after streaming in handle_model_response_stream (sync) and ahandle_model_response_stream (async)
agno/team/_response.py	Same extraction in _handle_model_response_stream (sync) and _ahandle_model_response_stream (async)
tests/unit/workflow/test_thinking_content.py	19 unit tests: extract_thinking_content, _process_step_output, _prepare_message, streaming extraction simulation, end-to-end workflow flow
tests/unit/team/test_thinking_content.py	8 unit tests: team streaming extraction, propagation to TeamRunOutput, edge cases (None, non-string, incomplete tags, native reasoning preserved)

Type of change

• Bug fix

---

Checklist

• Code complies with style guidelines
• Ran format/validation scripts (./scripts/format.sh and ./scripts/validate.sh)
• Self-review completed
• Documentation updated (comments, docstrings)
• Tested in clean environment
• Tests added/updated (if applicable)

---

Testing

• Unit tests: 891 passed (agent + team + workflow), 0 failures
• Integration tests: Verified with real Ollama models (qwen3:4b via OpenAILike, native Ollama, llama3.2)
• Non-thinking models: Confirmed zero impact (guard condition skips entirely)
• format.sh / validate.sh: All pass, mypy 0 issues

Additional Notes

• Native Ollama API separates thinking into a dedicated thinking field (no tags in content). The fix targets OpenAI-compatible API endpoints where thinking is embedded as <think> tags.
• Existing reasoning_content from native model fields is never overwritten by tag-based extraction.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[hztBUAA]

Thanks for the review and feedback. I am following up on this PR now and will either push the requested changes or reply point-by-point shortly.

[hztBUAA]

Quick follow-up: I am reviewing the feedback and will update this PR shortly.

[harshsinha03]

LGTM! @hztBUAA Thanks for contributing

[Qxiaobei]

I'm using Qwen3 with OpenAILike. When I open the thinking mode using Qwen3's thinking model, the Agno documentation doesn't provide a mechanism to automatically remove the <think> tag. The <think> content returned by the model is treated as a regular assistant message. In the agent loop, Agno treats the content between <think> and the <think> tag as 'role': 'assistant', places it in the context, and sends it back to the model. This affects my model calls, sometimes causing function calls to frequently get stuck in a loop of thinking and tool calls. How can I quickly resolve this issue?