[davior]

You can now paste a page from a music library into Add → "Or paste a page of music tracks" and press Find tracks. GAM lists the tracks on the page with their details. You filter the list, tick the tracks you want (up to 50 at a time) and import them. Each track becomes its own background job: it is downloaded, catalogued like an upload and given the page's details. The tempo, key and energy analysis from davior/gam#47 then fills in whatever the page left out.

How a page is read (app/ingest/page_tracks.py)

Sources are tried in order, most reliable first, and the review list says which one was used:

Source	Gives
incompetech	Its page builds the track list in the browser from pieces.json, so GAM reads that file directly. Title, genre and album (names taken from the lookup tables in the page's own script), feel (as moods), instruments, BPM, length, description, date, and the licence's exact CC BY attribution text as the credit line.
schema.org MusicRecording JSON-LD	Title, artist, genre, duration, date, licence. This is what sites publish for search engines.
mixkit (on top of its JSON-LD)	The moods, instruments and tags linked from each track's card, matched to the track by MP3 URL, plus the mood or instrument the listing is filtered to.
AI provider (fallback, can be turned off)	Reads the page text. It may only name audio URLs that actually appear on the page; any other URL it returns is dropped. Costs one completion, recorded in Usage.
bare links	Every linked or embedded audio file, titled by its link text.

Duplicates

Check	When	Catches	Then
URL	In the review list, and again in the job	A track already imported from this address	Shown as In library and can't be ticked. If an import of it runs anyway, the job points at the existing copy.
Checksum (SHA-256)	In the job, after downloading	The same file at another URL, or a file uploaded by hand	The download is discarded. The page's details only fill the existing copy's blanks, so a person's corrections stay. Its source URL is recorded, so the URL check catches it next time.
Title + artist	In the review list	The same piece downloaded elsewhere as a different file	Shown as Maybe in library. It can still be ticked, because titles collide, but Tick shown skips it.

A track ticked twice in one import becomes a single job.

Downloading (app/ingest/web.py)

• An SSRF check runs on every redirect hop.
• Audio is accepted only uncompressed and is capped at 200 MB while it streams in.
• A response that isn't audio (a login wall, an interstitial page) is refused rather than catalogued.
• Requests identify themselves as GAM, and downloads run one at a time on the enrichment queue.
• Audio URLs that come back from the browser are checked again before any job is queued, and dates are normalised.

What lands on the asset

The page's details are stamped embedded:

• creator, publisher, album, licence and credit line;
• source URL and retrieval time;
• genres, moods, instruments and BPM;
• the page's tags, attached as library tags.

Checked against your saved pages

• Mixkit: "Carefree" (13 tracks) and "Strings" (10 tracks), each with artist, duration, genre, moods, instruments, tags and licence.
• incompetech: the saved page confirms it loads pieces.json. I couldn't reach the file itself, so the adapter reads exactly the fields the page's own script uses, and the test fixture follows that shape.
• The committed fixtures are trimmed, hand-made copies of each page's structure. No third-party page is committed.

Design notes are in docs/page-import.md.

Testing

• pytest -q: 1370 passed. tests/test_page_import.py covers:
  • each adapter;
  • the AI fallback, including that an invented URL is dropped;
  • bare links;
  • the import job;
  • all three duplicate checks, including that corrections on the existing copy survive;
  • SSRF on a redirect, a non-audio response, the size cap, and the per-import limit.
• npm run format:check, npm run lint, npm test (425 passed) and npm run build all clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DCL4jm6ogZBBhnjEfBBaXU