dsh-assistant-optimization: DSH web plugin for collapsible model output and enhanced prompts

# Introduction After starting a service using the web profile of DSH (DeepSeek Harness), several issues are encountered almost daily: - Some OpenAI-compatible APIs lack a dedicated `reasoning_content` channel; the Chain of Thought is mixed directly into the content stream, and DSH renders the entire block as plain text, burying the answer within a large chunk of thinking. - Once a turn is complete, the screen is cluttered with lines of thinking, tool calls, and intermediate explanations; what is often truly relevant is only the final reply. - If the process is manually stopped or the session encounters an error, resuming requires manually reconstructing the input. - When the model outputs Mermaid charts, the interface displays only a block of code text. dsh-assistant-optimiz