Offline mode

The “offline” switch downloads all of a program’s materials to disk. After that the network is not needed: links inside topics open from the local archive.

How a material is fetched

A material’s type field already says how:

type	Strategy
docs	llms.txt / the .md variant first, otherwise mirror the section
article, post, guide	one self-contained HTML file plus the reader
spec, rfc	the same, usually clean HTML already
repo	a shallow clone
video	external yt-dlp, if it is present in the system
paper	direct PDF download

The liveness field is a ready-made filter. Materials marked paywall or login_required are not fetched at all — the result would be a login page, not the material — and are shown as “unavailable offline”, honestly.

What a saved page is

Archiving is done by monolith, used as a library: it inlines CSS, images, fonts and scripts as data URIs and returns a single HTML5 file. Pages are saved isolated and without JS — an archived article does not need scripts, and isolation guarantees the saved page will never reach out to the network.

monolith has no JS engine of its own, so a page that renders itself on the client would be saved empty. Bundling Chromium to fix that is unnecessary: Tauri already runs on the system WebView. A hidden window opens the URL, waits for load, hands the rendered HTML to monolith and closes. No new dependencies, no extra megabytes, and the window lives on a timeout so it cannot hang the queue.

The reader

Readable text is extracted by dom_smoothie, a Rust port of readability.js — the same algorithm browsers use for reader mode. It is a second layer on top of the archive, not a replacement: if extraction fails, the saved HTML is shown as it is rather than an empty screen.

Whole documentation sites

First we try what the site offers itself:

1. /llms.txt and /llms-full.txt — by now a de facto standard for developer documentation, and the ideal offline format: clean Markdown.
2. The .md variant of a page — many documentation generators serve one.
3. sitemap.xml — when a crawl is needed but the boundaries should be known upfront.

Only if none of that exists does the crawler run, with hard limits: the same domain, depth 2 by default, a page-count ceiling and robots.txt respected. Links between saved pages are rewritten to local ones.

Video

yt-dlp is not bundled. The app looks for yt-dlp and ffmpeg in PATH: if they are absent the button is disabled, with an explanation and the install command for the current OS next to it; if they are present they run as an external process with progress and a cancel button. PATH is parsed by the rules of its own OS — ; as the separator on Windows, : elsewhere.

The default resolution cap is 720p: a 40-minute lecture at 1080p is hundreds of megabytes, and slide legibility does not change.

The cache

<app-data>/cache/objects/<sha256[0:2]>/<sha256>
<app-data>/cache/index.sqlite

The index maps url → sha256, type, fetched_at, size, etag, last_modified. Content addressing gives deduplication: the same RFC referenced by three programs is stored once. The cache lives outside the bundle, because a bundle gets regenerated and the cache has to survive that.

Revalidation uses conditional requests with ETag / Last-Modified. The disk budget is configurable, eviction is LRU, and objects belonging to programs with offline enabled are protected from eviction.

The report

A download ends with an honest report, not with “done”:

Downloaded  34
Skipped      6   paywall (4), login_required (2)
Failed       3   timeout (2), 404 (1)

Failed items are retried by a separate action, without re-downloading the successful ones.