dsh-context-lens

When an AI agent confronts a multi-file repository or a wall of test and build output, the context window fills up with boilerplate function bodies, passing test lines, and build chatter. The result is inflated inference cost, model distraction, and frequent rate-limit hits. dsh-context-lens exists to fix exactly this problem.

The plugin delivers three layers of optimization. First, a multi-language AST code skeletonizer covering TypeScript, JavaScript, Python, Go, Rust, and Java: it preserves class definitions, function signatures, type declarations, and export statements while stripping internal implementations, loops, and repetitive boilerplate. Second, a fast O(n) heuristic log condenser targets outputs from Jest, Vitest, Pytest, Go test, Maven, Cargo, and more, retaining stack traces and assertion failures while filtering out PASS markers and build notices, with raw, balanced, and aggressive modes for different noise levels. Third, an active path focus mechanism lets the agent designate the files and directories it is currently editing; everything outside that focus list is automatically presented as a lightweight AST skeleton. Real-time token savings tracking, cumulative efficiency percentage badges, and a budget guard that warns at 90% completion round out the experience, cutting overall context consumption by up to 85% while keeping 100% of essential architectural interfaces and failure traces intact.

If you are driving DeepSeek Harness agents through cross-package architecture reviews, multi-file collaborative refactoring, or large-scale test debugging, and you want the model to stay locked onto key interfaces and failure paths without loading every line of implementation, dsh-context-lens is a lightweight context-optimization layer worth adding to your workflow.