dsh-session-summarizer

Long conversations routinely outgrow a model's short-term memory, leaving it unable to recall key points, constraints, and decisions made earlier in the thread. dsh-session-summarizer was built to fix exactly this: it registers a session_summarize tool that the model can invoke on its own, pulling in a specified, current, or most-recent session and producing a structured Chinese context summary so the lost thread is handed back to the reasoning process.

The entire pipeline runs inside the official DSH runtime. Session content is read through @deepseek-ai/dsh-session (supporting both live derived messages and persisted event-log paths), and the summary is generated via the streaming API in @deepseek-ai/dsh-llm. No external services or third-party APIs are involved. For very long transcripts the plugin applies a keep-head-and-tail, truncate-middle strategy with a default budget of 12 000 characters; the optional lastTurns parameter lets the model focus on only the most recent N user turns for a tighter, more targeted summary.

If you regularly work through multi-turn, in-depth conversations in DSH and want the model to look back through context on its own instead of you manually pasting history, this plugin offers a lightweight, self-contained solution. It mounts into your profile as a bundle patch and is ready to use in-conversation with no extra setup or external dependencies.