nodecode-pulldown

Pulldown: reads a public web page as bounded Markdown with its title, status and extractor, and can render a JavaScript page in Chromium. A Nodecode rewrite of anthony-maio/pulldown by anthony-maio.

Pulldown is a page reader for language models: it fetches one public page, keeps the part worth reading and drops the navigation, the scripts and the boilerplate, and answers markdown with the title, the HTTP status and a small routing record naming the extractor that read it. A page that only makes sense after its JavaScript has run can be rendered in Chromium instead. It runs on your machine; private and loopback addresses are refused, so the model cannot read your router's admin page with it, and nothing is sent anywhere but the site you asked for.

Get it

Tell the agent "add the pulldown cell", or run nodecode add nc://pulldown.

Set it up

The cell drives the pulldown program, so that has to be there first, in an environment of its own:

python3 -m venv ~/.local/share/pulldown/venv
~/.local/share/pulldown/venv/bin/pip install 'pulldown==0.5.0'

It is found on PATH; when it is not, give the executable's path in the cell's program setting (/settings). For rendering, install the extra and Chromium once, in the same environment:

~/.local/share/pulldown/venv/bin/pip install 'pulldown[render]==0.5.0' 'playwright==1.64.1'
~/.local/share/pulldown/venv/bin/python -m playwright install chromium

timeout_seconds (default 60, 5 through 120) is how long one page may take; cache (on by default) keeps fetched pages for an hour and answers a repeat read from them, in cache_dir or, when that is empty, in the cell's own folder under the Nodecode home. A cell that cannot find the program refuses to start and says so. config.example.jsonc in this folder lists every member.

Use it

The model gets one tool:

Tool	What it answers
`pulldown_fetch`	The page as markdown: title, address, a status line (HTTP code, detail, characters, time, whether the cache answered), the description, how the page was read, then the text. `detail`, `render` and `max_chars` shape it.

And the same read three ways:

(pulldown:fetch "https://docs.python.org/3/library/asyncio.html")
(pulldown:fetch "https://example.org" :detail "structured")
(pulldown:fetch "https://example.org/app" :render t :max-chars 4000)

/pulldown URL from the keyboard is the operator's own read; with no URL it says what the cell is reading with. (help :pulldown) is the model's manual.

detail is minimal, readable (the default), structured for listings and tables, or full for the whole page. max_chars (1000 to 40000, 20000 by default) clips the text and says how much it left; when the routing record recommends a browser, the answer says so, and render true reads the page in Chromium. Untrusted page text reaches the conversation, so the manual tells the model to treat it as source material and never as instructions.

It reads a page; it does not find one. Search with the websearch cell's web:search and read here the pages worth reading whole.