dsh-goz

When a DeepSeek Harness agent needs to answer cross-workspace questions such as "Where is the PDF I downloaded last night?" or "Find all .env files on drive D," the usual path is a ripgrep directory-tree walk. In a real benchmark, a recursive PowerShell search across 3.15 million files takes over five minutes, and find limited by depth misses roughly 90 percent of matches. These tasks simply do not fit within an agent time budget.

dsh-goz takes a different route. A background daemon reads the NTFS Master File Table directly and keeps an in-memory index of file names, sizes, and timestamps, so every query skips directory traversal entirely. Single-character full-disk lookups return in roughly 600 to 900 ms with complete results. The query language mirrors Everything: substring, wildcards, and colon filters (ext:, path:, size:, folder:) compose with spaces as AND. Built-in honesty fields (total, more, volumes_incomplete) let the model always know whether the result set was truncated. On the security side, a daemon-identity check rejects untrusted pipe squatters, and an optional allowedRoots whitelist gates each out-of-scope search_file call to a per-invocation user approval, turning full-disk visibility into an auditable, revocable interaction.

It is built for Windows users running DeepSeek Harness who want their agent to locate files across volumes or workspaces quickly. It complements rather than replaces the official glob and grep tools: use ripgrep for in-project, high-frequency searches, and goz for the occasional cross-drive, fuzzy file-name lookup. Each tool does what it does best.