Squeeze: a DeepSeek Harness plugin that reads long pages, PDFs and files without filling the context

When my agent needs one fact from a 90-page PDF or a long government web page, I don't want the whole thing dumped into the context. So I built squeeze , a small tool plugin for DeepSeek Harness. The agent calls squeeze with a link (or up to 20 links, or a file) plus a specific question. squeeze reads the material, hands it to a helper model you choose, and returns only the answer . Your main context stays small and cheap. What it does: Your choice of helper models: pick up to 3 from your DSH model list, each with a nickname and a "when to use" note. They can be local models or cloud ones. Reads more than web pages: text PDFs, logs, CSV/JSON and code files. Automatic size limits: each model's limit is worked out from its context window. Optional fallback to your next model if one is down. Clear errors that tell the agent what to do next (dead page, JavaScript-only page, scanned PDF, too big, and so on). Settings page at Plugins → squeeze → Configure. No Python and no extra packages. Install: ask your DSH agent to install the bundle github.com/q8atnight/dsh-squeeze/releases/download/v2.1.0/dsh-squeeze-2.1.0.tgz with the plugin manager. You need DSH 0.2.1-rc.5 or newer. PDF support also needs pdftotext ( brew install poppler ). Repo (MIT licence): github.com/q8atnight/dsh-squeeze Feedback and bug reports welcome.