A photo is now decoded only when its header says it is within 40 megapixels,
and bytes TREK cannot shrink, such as a WebP, are sent only up to 3.5 MB. A
scanned PDF page is drawn by its long edge, and pdf.js skips an embedded image
above the same pixel cap instead of decoding it.
The receipt reading is held to bounds in the shared contract, keeps the minus
sign of a discount and leaves such lines out so the Ticket split can save, and
its prompt states the JSON shape, so a server without json_schema still answers
something the parser reads. The model lookup sends the configured key, the
import error names the photo formats, the scanned photo is attached only for
someone who may upload files, and the phone settlement reloads after a receipt
is saved from the trip sheets. Test ids that clashed after the rebase are
renumbered, and the wiki describes the image setting, photo import and receipt
scan.
Whether the AI Parsing model reads images is an admin setting now, and it
lands with the two features that send a photo, so it has something to
control.
The setting is `vision: 'auto' | 'on' | 'off'` on the addon config. On
Automatic a local Ollama is asked what the model can do (`/api/show`), and a
cloud provider means no; Yes and No are applied as set. The old `multimodal`
flag is not read: the admin screens wrote it false on every save, so a
stored false meant nothing. The per-user toggle keeps its on/off meaning and
is read by the extraction now. The lookup is server-side only: capped read,
validated body, never throws, cached per server and model.
GET /api/llm/capabilities answers it for the signed-in caller, and both
photo pickers below ask it.
The booking import takes JPG, PNG and WEBP when the model reads images.
The photo goes as an image to every provider (image_url, an Anthropic image
block, Ollama's native chat with `images`), capped at 1600px, and skips KDE
Itinerary. A scanned PDF, which used to reach the model as empty text, has
its first two pages drawn by pdf-parse and sent the same way. The rest is
the import as it was: background job, review, the file attached to the
booking, the linked cost. On a real qwen3.5:4b a train ticket came back as
bare clock times until the router gave the model every type's field
meanings, and its price as an object inside a string, which read as 4900;
both are handled.
Costs gets Scan receipt: a dialog laid out like the booking import's takes
one photo, read on the import job queue, then reviewed in the existing
expense editor pre-filled with merchant, total, currency, day and lines,
the photo waiting in Receipts & Invoices to be uploaded on save. No table,
no migration, no category guess, no document-type detection.
These are the only two test files that call request.formData() inside an MSW
handler, and they hold exactly the three tests that failed on CI while passing
locally. The body axios' XHR adapter produces under jsdom is one Node 24's
multipart parser rejects, so the handler threw before reading a field and the
component reported a failed upload. Reading the raw payload asserts the same
thing without depending on the parser.
The helpers built a one-byte File and then redefined size to 100, so the
multipart body and the length it announced disagreed. Node 24 rejects that
while 22 lets it through, which is why only CI saw it: every test that
uploaded such a file and expected success failed, the one expecting an error
passed for the wrong reason, and the KML case with a truthful File was fine.
Maurice