- Fix VIDEO-type port input (ComfyUI V3 VideoInput object) not being
recognized in _resolve_input_tensor, causing 'No video frames provided'
error. Now uses get_components().images to extract frames.
- Fix build_thinking_suppression receiving service_id instead of service
type, causing think:true param not injected for Ollama custom-named
services. Changed to pass service.get('type', provider) in vlm.py and
llm.py.
- Fix streaming response parser mistakenly appending Ollama's 'thinking'
field content into full_content, polluting the final answer.
Now only message.content is appended; thinking field is ignored.
- Add fallback: if filter_thinking_content returns empty string (e.g.
model wraps entire output in <think> tags), preserve original unfiltered
content to avoid false empty-result errors.
yawiii