dsh-frecency

In long DeepSeek Harness sessions, the built-in grep and glob tools spawn a fresh ripgrep process and scan from scratch on every single call. When several parallel subagents trigger the same searches dozens of times, those repeated cold starts become a noticeable bottleneck. dsh-frecency takes over both tools and keeps one resident in-memory index per working directory (powered by the Rust fff engine), so repeated content searches hit hot memory in single-digit milliseconds.

Results are ranked by frecency — files that are opened often or modified recently surface first, letting the model focus on the most relevant code without any extra prompting. Glob faithfully mirrors the built-in discovery semantics (hidden and ignored files included, VCS metadata excluded, modification-time ordering) and degrades gracefully to the resident index when ripgrep is unavailable. Grep output additionally carries isDefinition and per-file gitStatus annotations, so the model can spot definition lines and modified files at a glance without re-reading. The switch is completely transparent: tool names and schemas stay the same, requiring zero prompt changes, and if the native engine fails to load the built-in ripgrep tools simply keep working.

If you frequently run long harness sessions, coordinate multiple subagents, or perform heavy, repetitive file searches against the same codebase, this plugin turns each cold scan into a memory-level hit — freeing the model to spend its budget on reasoning instead of waiting on I/O.