dsh-knit

Your agent cranks out thirty Markdown files, a few screenshots, and some screen recordings in a single session, and you cannot remember which one you were just talking about. Knit does one thing well: it scans the entire working folder for documents and media, then ranks them by relevance to what you are actually discussing right now. An architecture chat floats the architecture doc to the top; a competitor discussion surfaces the competitive brief. No manual file hunting, nothing to store, and it persists across restarts, new sessions, and days.

Under the hood it is BM25 scoring on local string operations. The most recent user messages are mined for keywords (exact English tokens, Chinese 2/3-grams, boundary-fragment filtering), and every document in the folder is scored from there. No embeddings, no model calls, no network requests, zero latency, zero cost. The panel offers three views (documents, images and video, all), grid thumbnails with first-frame video previews, inline expansion for full-text reading, and double-click for a new-tab render. On the agent side, a knit_docs tool hands back the pre-ranked list of paths and summaries, saving the token cost of globbing paths and reading them one by one.

If you maintain a growing project folder where file names do not always describe the content, or if you want your agent to reference project history without guessing file paths, Knit is a zero-dependency, zero-network add-on that earns its place the first time you stop searching and start sorting.