Media and views

A capture has two derived layers on top of the canonical JSON: the localised media, and the rendered views. The media is controlled by the capture depth; the views are controlled at capture time and can both be rebuilt later from the JSON.

How media is localised

Everything downloaded lands under media/, bucketed by type:

media/thumb/   <vid>__<h6>.jpg     the picked thumbnail
media/avatar/  @mkbhd__<h6>.jpg
media/banner/  @mkbhd__<h6>.jpg
media/video/   <vid>__<fmt>.mp4    only at --depth media
media/audio/   <vid>__<fmt>.m4a    --depth audio, or -x

Each file's name is the source key plus a short hash of its source URL, which makes two things true: two thumbnails of one id never collide, and a channel avatar referenced across a thousand videos resolves to a single file on disk, downloaded once. The rendered pages rewrite their src attributes to these local paths, so the archive opens with no network.

Thumbnails, avatars, and banners are fetched at every depth, through the same client transport as the record reads, so they share its rate limiter, retry, and cache. Streams are fetched only at media or audio depth; their filename encodes the format selection, so two different selections of the same video coexist and a re-run with the same selector finds its file already present.

The manifest is honest about what made it: a media or stream fetch that fails is recorded, and the record still renders with a "media unavailable" or "stream not archived" placeholder. A missing file degrades a page; it never aborts the capture.

The HTML watch page

The HTML view follows kage's philosophy: the output is inert. No <script>, no on* handlers, no remote fonts, no analytics. index.html is the repo home (a channel header and a grid of videos, or for a single video, the watch page itself), and html/<vid>.html renders one video as a watch page: the title, the channel row, the metrics, the full description with linkified timestamps, the chapters, the inline transcript, and, when captured, the top comments.

At media depth the player region is a passive HTML5 <video> element over the local file, with no autoplay JS. At meta depth it is the localised thumbnail with a "stream not archived" caption. Either way it is a photograph that happens to play, not a program.

Choosing views

--view selects which rendered views to write. JSON and the sidecars are always written regardless:

kura archive @mkbhd --view html       # HTML only (default)
kura archive @mkbhd --view md         # Markdown only
kura archive @mkbhd --view html,md    # both

The Markdown view gives you a README.md index plus per-video Markdown under md/, each with a front-matter block, the description, the thumbnail as a relative-path image, the chapter list, and the full transcript inline, so the Markdown archive is a genuine greppable corpus.

Re-rendering offline

Because the views are derived, you can rebuild them from the stored JSON at any time with no network:

kura render $HOME/data/kura/youtube/@mkbhd

This is how you add a view to an archive, or replay a renderer improvement over an old capture:

# Add Markdown to an archive that only had HTML
kura render $HOME/data/kura/youtube/@mkbhd --view md

kura render reads only the stored JSON and sidecars. It never touches the network and never re-downloads media. Pass --date to fix the footer stamp for reproducible output.

Next

• The CLI reference lists every flag.
• Repository layout maps what lands on disk.