squeeze — read long material without filling the chat

A plugin for DeepSeek Harness (DSH) that adds one tool, squeeze.

中文简介：squeeze 工具让 Agent 把长网页、多个网页、文本文件或 PDF 交给你选定的辅助模型阅读，只把答案带回对话，原文不占用上下文。

When an agent needs something from a long web page, a batch of pages, a big text file or a PDF, it calls squeeze with a specific question. squeeze downloads or reads the material, hands it to a helper model you choose, and returns only the answer. The material never enters the main conversation, so the agent's context stays small and cheap.

Typical uses: pull figures out of a long report, find one section in a large log or code file, compare a few web pages, summarise a transcript.

Example

The agent asks about a 17,000-character government PDF. Only this comes back into the chat (a few hundred characters instead of the whole document):

squeeze  url: https://tax.gov.ae/…/Cabinet Decision No. 75 of 2023 and amendments - For Publishing.pdf
         prompt: What penalty (AED) applies for failing to submit a corporate tax registration application on time?
                 Quote the item number.

Item 14 in the Table of Violations and Administrative Penalties: "Failure of the Taxable Person to submit a Tax
Registration application within the timeframe specified by the Authority …" — penalty AED 10,000.
(Footnote: added by Cabinet Decision No. 10 of 2024, effective 1 March 2024.)

squeeze: ok | model quick (Qwen Flash Next) | 6.9s | source 16,856 chars | thinking medium

(A real call from 3 Nov 2026, answered by a local model; the answer is shortened here.)

The status line tells the agent which model answered, how long it took and how big the source was. Failures say what went wrong and what to do next (for example squeeze SOURCE PROBLEM: … this PDF has no text layer (probably a scan or photo) — it needs OCR).

Settings page

Features

• Models from your DSH model list — pick a default model and up to two extra ones on the plugin's own settings page. Nothing to type twice: whatever providers and models DSH already knows are offered.
• Nicknames and "when to use" notes — give each model a short nickname (quick, careful, …) and a one-line note. The agent sees them in its tool description and picks a model by nickname.
• Fixed thinking level per model — you decide how hard each model thinks; the agent cannot change it.
• Size limit follows the model — each model's limit is worked out from its context window (default: the material may use 80% of it, the rest is left for thinking and the answer). The agent sees the limit per model.
• Optional fallback — when switched on, a model that is unreachable, returns a server error, has left the model list, or is too small for the material is skipped and the next model in your list is tried. The answer's status line says which model answered and why the others were skipped. Time-outs do not fall back.
• Clear errors — every failure starts with a kind the agent can act on (REQUEST PROBLEM, SOURCE PROBLEM, TOO BIG, MODEL UNAVAILABLE, NO ANSWER, TIMED OUT, CANCELLED) and ends with the next step.
• Batches of pages — up to 20 addresses per call (configurable), downloaded four at a time; dead pages are skipped and listed instead of failing the whole call.
• PDFs — PDF links and PDF files with a text layer are converted to text automatically when the free pdftotext program (poppler; brew install poppler on a Mac) is installed. Scanned PDFs are refused with advice to use OCR.
• Sites with huge headers — some government portals send more header data than Node's built-in downloader accepts (16 KB); squeeze retries those with a larger limit.
• Short pages — a page with almost no text (often a JavaScript-only or blocked page) is handed back to the agent as-is without a model call (or refused, if you prefer).
• One call at a time per provider — calls to the same provider queue, so a single local model is never flooded; calls to different providers (say a local box and a cloud model) do not wait for each other.
• No Python, no extra packages: the tool uses DSH's own model connections and logins.

Install

Requires DSH 0.2.1-rc.5 or newer. For PDF support also install pdftotext (brew install poppler on a Mac).

From the GitHub release — ask the DSH agent:

Install the bundle https://github.com/q8atnight/dsh-squeeze/releases/download/v2.1.0/dsh-squeeze-2.1.0.tgz with the plugin manager.

(It uses plugin_manager → install_bundle with that address. You can also download the .tgz file and give the agent its path instead.)

From a folder — clone or unpack this repository somewhere permanent (for example ~/.dsh/plugins/dsh-squeeze) and ask the agent to install the bundle at that folder.

Then restart DSH if the plugin manager says a restart is needed, and open Plugins → squeeze → Configure to choose your models.

Without any model chosen, squeeze uses the DSH default chat model.

Settings

Setting	Default	Meaning
Helper models	—	Default + up to 2 extra: model, nickname, thinking level, "when to use" note. The same model can be listed only once.
Fallback	off	Try the next model when one is down or too small.
Context fill (%)	80	Share of each model's context window the material may use.
Output budget (tokens)	16,000	Thinking + answer, capped at the model's own maximum.
Time limit per model call (s)	600	Longer calls are stopped and reported.
Time limit per page download (s)	30
Web addresses per call	20
Longest answer (characters)	6,000	Longer answers are cut at a line break and marked ANSWER CUT.
Minimum page length (characters)	300	Below this a single page is treated as "probably JavaScript-only or blocked".
Pages under the minimum length	hand over as-is	Or refuse with an error.

Changes apply to the next tool call; open sessions see the new tool description from their next request.

The tool (what the agent sees)

Parameters: prompt (required), and one source — url, urls (list) or file (path; file_path is accepted too) — plus model (a nickname) when more than one model is set up.

The result is the answer followed by a status line, for example:

squeeze: ok | model careful (GLM-5.3-Flash) | 41.2s | source 152,300 chars | thinking high

Limits

• Text, HTML and text PDFs only (PDFs need pdftotext). Convert Word or Excel files to text first; scans and photos need OCR.
• Pages that only render with JavaScript cannot be read — use a browser tool for those.
• Size limits are estimates (characters → tokens). If a model still reports "too long", the call is reported as TOO BIG (or falls back, if fallback is on).
• Answers come from a helper model: treat them as leads and check important figures against the source.
• Privacy: the material is sent to whichever model answers. If you mix local and cloud models and switch fallback on, material meant for a local model can end up with a cloud model when the local one is down.

Files

• index.js — the tool (host side)
• client.js — the settings page
• cordis.patch.yml — switches the plugin on when the bundle is installed
• locale/en.json, icon.svg — name, description and icon shown on the Plugins page
• assets/, screenshots.json — README pictures and the screenshot list used by plugin storefronts
• test/ — offline checks: ELECTRON_RUN_AS_NODE=1 "<DSH app binary>" --import ./test/hooks.mjs test/unit.mjs

Changes

• 2.1.0 (3 Nov 2026) — PDF links and files are read (via pdftotext); sites with more than 16 KB of headers are read; batch pages download in parallel; downloads without a size header are capped too; over-long answers are cut at a line break with status ANSWER CUT (out-of-budget answers keep ANSWER INCOMPLETE); the queue is per provider instead of global.
• 2.0.0 — settings page, up to three models with nicknames, fallback, size limits from the context window.

Licence

MIT — see LICENSE.