signal-bot/DOCS-add-photo.md

# Adding a photo from the chat: `!add_photo`

How a member of an album's group proposes a photo for it by sending one,
how the operator approves it, and what keeps that path from leaking anything.
The album itself — where it lives, what its manifests look like, how several
albums coexist — is `DOCS-photo-pack.md`.

Send a photo captioned `!dodaj alfa wakacje2019` in one of the album's groups.
`alfa` resolves to that person's key; `wakacje2019` matches nobody and travels
as a *proposed* label. The photo is stripped and lands in the album's
`pending/`, and the reply is `czeka na zatwierdzenie: 1`. It is **not in the
album yet**: `!alfa` and `!losowe wakacje2019` cannot find it until somebody
approves it.

### Approval: the management page, and nothing else

`./web.sh` shows a **Waiting for approval** section above the tagger whenever
`pending/` is not empty (`web/src/views/pending.tsx`). Each queued photo is a
card with the album's people as checkboxes — the ones the sender named are
pre-ticked — and one checkbox per proposed label. **Approve** moves the file
into `media/`, writes `tags.json`, mints any ticked label in `custom.json`,
records the sender in `adds.json` (with `approved_at`), and asks the bot for
`POST /_reload`. From that moment `!losowe wakacje2019` answers with it, with
no restart. **Reject** deletes the file and its sidecar; nothing else to undo,
because nothing else was written.

The queue is the only door in. The bot's own process never approves anything;
it writes `pending/<content-hash>.<ext>` plus a `.json` sidecar (`keys`,
`labels`, `by`, `at`, `group`) and reads the result back at the reload. The
page can only approve a label the sender actually proposed (anything else is
`400`), and it can keep fewer — an unticked label is simply not minted. If the
bot then refuses the reload (a stale alias, say) the page **reports** it and
does not roll back: the photo is in the album with the tags the approver
chose, and the bot's reason is about something else.

What holds:

- **No typed label travels.** Labels become keys in the registry's
  `arg_resolver`, at the dispatcher, *before* the line that logs the command
  name; a new word becomes a per-message token (`new1`, `new2`) and is claimed
  back by its own handler, so concurrent frames cannot swap labels. The reply
  is counts only — `czeka na zatwierdzenie: 2 · 1 już było` — and the proposed
  word exists in exactly one place, the sidecar inside the album's `0700`.
  `scripts/check_personal_data.sh` reads those sidecars too, so a proposed
  name is on the forbidden list before it is ever approved.
- **No cap.** No cooldown, no size limit, no daily limit. `!random` takes the
  ordinary pack cooldown; `!add_photo` takes no decorator at all.
- **At most six labels**, the same ceiling `!alfa beta …` has.
- **The attachment must be on the same message**, never a quoted one: a reply
  carries the *quoted* message's media in `quote.attachments`, and taking it
  would let anyone re-file somebody else's photo.
- **A duplicate changes nothing.** The stored name is a content hash, so a
  photo already in the album is `już było`, one already waiting is `już
  czeka` — the same photo twice in one message counts as waiting too — and a
  mixed batch queues the new files and says how many were which.
- **Approval is whole or not at all.** Media, `tags.json`, `custom.json` and
  `adds.json` are written, the album is loaded again, and only a load that
  succeeds keeps them. A load that fails rolls all four back and leaves the
  photo in the queue with a `400` naming the reason.
- **The second copy goes.** signal-cli keeps its own copy of every inbound
  attachment outside the album's `0700`. `!add_photo` deletes it once the file
  is queued, and also when the upload was refused.
- **Stripping happens before the wait.** The file in `pending/` is already
  the stripped one; `scripts/e2e.sh` re-reads the queue with exiftool before
  approving anything, and the album after.

### Stripping: exiftool, with a positive check

`ffmpeg -map_metadata -1 -c copy` was tried first and **fails silently on
photos**: given a JPEG with a real APP1 Exif segment it returned a
byte-identical file, GPS marker intact, and exited 0. `-c copy` passes the
compressed frame through as an opaque packet and a JPEG's Exif lives inside
it. Do not reintroduce ffmpeg for this.

So the gate is a **positive check on the output**, never an exit code:

```sh
exiftool -all= -time:all= -o <dst> <src>   # never --overwrite_original
exiftool -G0 -s -all <dst>                 # then read it back
```

`-time:all=` is **not** redundant with `-all=`. `media_tags.ZEROED_DATE_TAGS`
allows the QuickTime date tags "by value" on the grounds that `-all=` zeroes
rather than deletes them. Measured 2026-08-31: it does not. `-all=` alone left
`QuickTime:CreateDate` at its real value, and the consequence was that **every
mp4 was refused** — the code documented behaviour it did not get. `-time:all=`
is what actually zeroes them.

and every surviving tag must be **structural**. Measured against exiftool
13.55, the plan's original "the surviving set must be empty" would have
rejected every format except JPEG — a valid PNG always reports its IHDR under
`PNG`, a GIF its screen descriptor under `GIF`, an MP4 some fifty `QuickTime`
tags that *are* the container. So the allowlist is per **tag name**
(`signal_bot/media_tags.py`), which keeps the property that matters: an atom
nobody anticipated arrives as a name that is not on the list, and the upload
is refused.

Allowlist, not blocklist, for the reason a naive gate misses: an MP4 geotag is
a QuickTime `udta`/`©xyz` atom reported under **`QuickTime`**, not `GPS`.

### Tracks: ffmpeg, for what no tag allowlist can see

A tag allowlist is structurally blind to one class, and it is the worst one.
`IMG_6924.mp4` in the live album carries two `mebx` timed-metadata tracks, and
after a *complete* `-all= -time:all=` pass its bytes still contain:

```
com.apple.quicktime.detected-face
com.apple.quicktime.detected-face.bounds
com.apple.quicktime.detected-face.face-id
com.apple.quicktime.detected-face.roll-angle / .yaw-angle
```

Per-frame face bounding boxes and a face identity number. exiftool is not
broken for missing them: `mebx` samples live in `mdat` as opaque payload, not
as tags, so **there is no tag to refuse**. The only answer is to drop the
track, which `signal_bot/media_tracks.py` does before exiftool runs:

```sh
ffmpeg -v error -i <src> -map 0:v -map "0:a?" -c copy \
       -map_metadata -1 -map_chapters -1 -y <dst>
ffprobe -v error -show_entries stream=codec_type -of csv=p=0 <dst>
```

and every surviving stream must be `video` or `audio`.

**This is not the ffmpeg mistake above.** That one was ffmpeg replacing
exiftool for *tag* removal, where it silently failed on JPEG. Here ffmpeg
removes *tracks*, exiftool still does every tag removal afterwards, and the
gate is a positive read-back of which streams survived rather than an exit
code. `-c copy` is safe for this job precisely because nothing is being
claimed about the frames — the claim is about which tracks exist, and ffprobe
answers that directly. Do not collapse the two jobs back together.

`QuickTime:MetaFormat` is deliberately **absent** from the allowlist: a file
still reporting one has a metadata track the re-mux should have removed, and
refusing it is the correct outcome.

**Both entry points obey this.** `./web.sh --import` used to be a
verbatim `shutil.copy2`; since 2026-08-31 it runs the same stripper, so the
guarantee on this page covers the whole album and not just the chat path.

The list includes the **codec's own configuration** — `CompressorName`,
`VendorID`, and the HEVC parameter set (`HEVCConfigurationVersion`,
`GeneralProfileIDC`, `BitDepthLuma`, …). That was forced by measurement, not
chosen for convenience: with them off the list exiftool 13.55 could produce no
acceptable video *and* no acceptable HEIC, because `CompressorName` lives in
the `avc1` sample description and exiftool reports it as not writable at all
(`Sorry, QuickTime:CompressorName doesn't exist or isn't writable`), and a real
HEIC reports sixteen HEVC config tags. Since HEIC is the iPhone default, the
choice was codec parameters or no phone photos. They describe how the picture
is encoded, never where it was taken or who took it; every location and
identity tag is still refused, and `scripts/e2e.sh` proves a QuickTime GPS atom
does not survive on a file the album actually stored.

Formats are decided by **magic bytes**, never by the filename: jpg, png, gif,
webp, mp4, mov, m4v, heic, heif, avif, webm and svg. Every one of them must
also be in `pack.MEDIA_TYPES`, or a file that uploaded cleanly is refused by
the *next* load and takes startup down.

### ⚠️ .webm and .svg are stored unstripped

exiftool 13.55 can write neither (`exiftool -listwf` lists no `WEBM` and no
`SVG`; a strip attempt returns `Can't write SVG files`). Its SVG *reader* does
not surface `<metadata>` content either, so there is no detection to fall back
on.

kuhy's decision of 2026-08-31 is to keep both on the allowlist and store them
**as they arrive, metadata included**. The consequence, stated plainly: an
`.svg` carrying `<metadata>` or an embedded
`<image xlink:href="data:image/jpeg;base64,…">` keeps whatever is in it, and
every later `!alfa` re-sends that file verbatim. Every other format on the
list is stripped and verified.