DeepSeek Harness Plugin Introduction: dsh-context-lens Semantic Compression and Skeletonization
`dsh-context-lens` is a semantic compression plugin for DeepSeek Harness, designed to reduce token consumption through structured code skeletons and intelligent log compression. It supports multilingual AST skeleton generation (preserving type definitions while removing implementation details) and provides three compression modes for logs from various testing/build tools. The plugin offers the `context_lens_code` and `context_lens_log` tools and supports path focus management to concentrate on the editing tasks in the current session. All logic is implemented using pure regular expressions, without relying on heavyweight parsers.
Read Moredsh-context-lens: View changes between adjacent model requests and cache reuse
dsh-context-lens is a pure observer plugin developed specifically for DeepSeek Harness, designed to continuously track and analyze context changes in real LLM requests within DSH sessions. By reading session logs (without injecting content or interfering with model calls), it generates a record for each request, displaying the submitted request context (such as model, system prompt, tool set, etc.) and cache reuse metrics, while automatically comparing differences with the previous request and alerting on potential drops. Its cache analysis is strictly based on the provider's `usage buckets` field to avoid inferring internal implementation. The plugin offers bilingual views (Chinese and English), retains the last 100 request records, emphasizes that "correlation does not imply causation," and explicitly does not guarantee determining the provider's internal cache key construction or harness.
Read Moredsh-context-lens: DeepSeek Harness Read-only Context Observation Plugin
`dsh-context-lens` is a read-only context observability plugin for DeepSeek Harness (DSH) used to debug the provider-neutral requests assembled during each `AgentLoop` of the agent. It records system, messages, tools, and invocation configurations, and adds a new "Context" tab to the Web interface, supporting the visualization of request volume, source attribution, search, and comparison of differences between different attempts. The plugin does not modify LLM requests, does not register tools, and does not intercept network traffic, ensuring it remains non-intrusive in production environments. Core features include recording request details, displaying UTF-8/UTF-16 bytes, and estimating token counts (based on a 4-character/token estimation, not precise values), as well as exporting diagnostics.
Read More