dsh-context-compactor: A Context Compaction & Summarization Plugin for DeepSeek Harness

What Does It Do?

dsh-context-compactor is a DSH plugin for DeepSeek Harness that compacts and summarizes conversation context to manage token usage. It solves the problem of requests failing when the conversation approaches the model's context window limit, by summarizing older messages into a detailed checkpoint while preserving core information. Key capabilities include automatic triggering at 80% usage, post-compaction token verification, context-overflow recovery, oversized tool result pruning, and quick-action buttons above the input box.

Core Features

• Automatic summary compaction: at 80% of the model window, it uses an LLM to compress earlier history into a detailed Chinese checkpoint, keeping only the most recent 16% of raw conversation.
• Post-compaction guarantee: verifies totalTokens strictly decreases after each compaction, with additional attempts if needed.
• Context-overflow auto-recovery: when the provider reports context length exceeded, it summarizes, compacts, and retries the request, up to 2 times by default.
• Tool result pruning: trims tool outputs over 8192 characters to head and tail, avoiding harsh truncation of conversation history.
• Input toolbar: shows real-time context usage percentage, with 'Compress Summary' and 'Prompt Enhance' buttons for one-click actions.
• Manual commands: /compact, /enhance-prompt, and /context-status for immediate compression or status checks.

How to Use This Plugin?

After enabling it in DSH, the plugin automatically listens for agent creation and mounts the compaction engine, working without extra configuration. When context usage reaches 80%, it triggers detailed summarization automatically; you can also click the 'Compress Summary' button above the input box or type /compact to compress manually. Use /context-status to view current token usage, window size, and threshold. To adjust trigger ratio or retention ratio, modify thresholdRatio and retainRatio in the config file, or set modelPolicies for specific models.