Web links previously launched a separate Chrome/Edge kiosk process, which
caused the whole class of bugs in the tracker: the browser opening behind the
player, being handed off to an already-running instance and exiting instantly,
fighting for foreground/z-order, and leaking msedge.exe/chrome.exe processes
that were never closed.
WebView2 renders as a CHILD HWND of Kivy's own SDL window instead, so there is
no separate top-level browser to open behind the player, nothing to hand the
URL off to, no z-order contest, and no leaked browser process.
Windows/webview2_browser.py
- Environment -> controller -> navigate, driven through pythonnet.
- Async .NET Tasks are polled from Kivy's Clock. Calling GetAwaiter()
.GetResult() would deadlock: the continuation needs the same thread's
message pump.
- The controller is a .NET IntPtr, not a Python int (CreateAsync overloads
do not match otherwise).
- NavigationCompleted is tracked so a page that never loads can be told
apart from one that did. This matters on a closed network: an unreachable
host paints a Chromium error page, and without this the player would show
a blank/error screen for the item's whole slot instead of skipping it.
- is_alive() reports True while starting up. Start-up is async, so a
controller that does not exist yet is not a dead browser; treating it as
one made the first web link after a cold start be skipped instantly.
Windows/webview2_runtime.py
- Detects the Runtime (registry pv value, SDK probe as fallback) and
installs it silently when missing, unelevated, which produces a per-user
install and therefore never raises a UAC prompt on the signage display.
- Success is decided by RE-READING the installed version, not by the
installer exit code: Edge Update returns a non-zero HRESULT
(-2147219416) when the Runtime is already current, which is not a failure.
- On a closed network the online bootstrapper can never succeed, so it fails
fast with an actionable message instead of hanging for the full timeout.
- Failed attempts are cooldown-gated so a broken machine does not re-run an
installer on every start.
Offline hardening
- Browser arguments disable component updates, field trials, safe-browsing
list fetches, translate and other internet chatter. On an isolated LAN
each of those would otherwise have to time out, costing start-up latency.
Pages on the local server are unaffected.
Engine order (best first): WebView2 -> CEF -> Chrome/Edge subprocess. CEF has
no wheels past Python 3.9 so it is dormant on this build; the subprocess engine
remains only as a last resort.
Also fixes the reason the Windows adapters were never used at all:
SignagePlayer.__init__ assigned self.weblink_adapter_factory = None, which
shadowed the CLASS attribute that run_win.py injects. play_weblink() therefore
fell back to the generic adapter, whose find_browser() uses shutil.which() and
finds nothing on Windows because Chrome/Edge are not on PATH. The instance
attribute is now only set when the class attribute is absent.
Verified: windows/test_webview2_embed.py, test_webview2_navigation.py and
test_webview2_offline.py all pass (a locally served page renders with all
internet traffic disabled), and the packaged exe reports
"weblink_launch engine=webview2-embedded" -> "weblink_launched" on every cycle
with no leaked browser processes.
16 KiB
Web Link Playlist Items — Player Integration Guide
This document describes how the Kiwy-Signage player
(https://gitea.moto-adv.com/ske087/Kiwy-Signage.git) supports the weblink
playlist item type (display a live web page / URL instead of an uploaded media
file).
Status: implemented. The player supports
weblinkitems on both Raspberry Pi (chromiumsubprocess) and Windows (embedded WebView2, with the CEF and Chrome/Edge subprocess engines as fallbacks). Sections 1–4 describe the original design plan; section 6 documents the shipped architecture and the interaction model.
1. Background — how items flow
DigiServer API ──JSON──▶ player sync (get_playlists_v2.py) ──▶ playlist.json ──▶ main.py renders
/api/playlists downloads files to media/ by item type
Each playlist item the server returns currently looks like:
{
"id": 42,
"file_name": "promo.jpg",
"type": "image",
"duration": 10,
"position": 1,
"url": "https://server/digiserver/static/uploads/promo.jpg",
"edit_on_player": false
}
The player:
- Syncs (
src/get_playlists_v2.py→download_media_files()): downloadsurlinto the localmedia/directory, then rewrites each item keeping onlyfile_name,url(now a local relative path),duration,edit_on_player. Note: thetypefield is currently discarded here. - Renders (
src/main.py→play_current_media()): opens the local file and chooses a Kivy widget purely by file extension (.mp4/.avi/...→Video,.jpg/.png/...→AsyncImage). Unknown extensions are skipped as "unsupported".
A web link breaks all three assumptions: there is no file to download, no extension to switch on, and no widget that renders a web page.
2. New server contract (what DigiServer will send)
A web-link playlist item will look like this:
{
"id": 91,
"file_name": "weblink-3f9c1a2b",
"type": "weblink",
"duration": 30,
"position": 4,
"url": "https://example.com/dashboard",
"edit_on_player": false
}
Key differences vs. a file item:
| Field | File item | Web-link item |
|---|---|---|
type |
image / video |
weblink |
url |
Path to a file on the server | The web page to display (the link) |
file_name |
Real filename on disk | Synthetic id (weblink-<uuid>), no file exists |
The player must branch on type == "weblink" and treat url as the page to
open — never try to download it as a file.
3. Required player changes
3.1 Sync step — src/get_playlists_v2.py
Function: download_media_files(playlist, media_dir, ...)
- Skip download for web links. At the top of the per-item loop, detect
media.get('type') == 'weblink'and do not callsession.get()/ write any file for it. - Preserve
typeand the originalurl. Theupdated_mediadict that is appended toupdated_playlistcurrently dropstypeand rewritesurlto a local path. It must now carrytypethrough, and for web links keepurlas the original web address (do not convert to a local relative path).
Suggested shape of the per-item logic:
item_type = media.get('type', '')
if item_type == 'weblink':
# No file to download — pass the web link through unchanged.
updated_playlist.append({
'file_name': media.get('file_name', ''),
'type': 'weblink',
'url': media.get('url', ''), # the actual web page
'duration': media.get('duration', 10),
'edit_on_player': False,
})
continue
# ... existing download logic for file items ...
updated_playlist.append({
'file_name': file_name,
'type': item_type, # <-- now preserved
'url': os.path.relpath(local_path, os.path.dirname(media_dir)),
'duration': duration,
'edit_on_player': media.get('edit_on_player', False),
})
delete_unused_media()walksmedia/usingfile_name. Web links have no file, so they simply won't match anything on disk — no change strictly required, but make sure a missing local file for aweblinkitem does not trigger a re-download or an error elsewhere.
3.2 Render step — src/main.py
Function: play_current_media(self, force_reload=False)
The current logic builds media_path = os.path.join(self.media_dir, file_name)
and then does os.stat(media_path) — which will fail for a web link (no file).
Add a web-link branch before the file-existence check:
media_item = self.playlist[self.current_index]
file_name = media_item.get('file_name', '')
duration = media_item.get('duration', 10)
# NEW: handle web links before any file/path handling
if media_item.get('type') == 'weblink':
self.play_weblink(media_item.get('url', ''), duration)
return
# ... existing file existence check + extension branching ...
Then add a new method play_weblink(self, url, duration):
- Validate the scheme is
http/https(reject anything else, e.g.file://). - Open the page for
durationseconds, then advance withself.next_media(). - Wrap in
try/except; on failure incrementself.consecutive_errorsand callself.next_media(), matching the existing error-handling pattern. - Make sure the previous widget (
self.current_widget) is removed/stopped just like the image/video paths do.
Rendering approach (pick one)
Kivy has no production-grade embedded web view, especially on Raspberry Pi. Recommended options, in order of robustness:
-
Chromium kiosk overlay (recommended). Launch Chromium over the Kivy window for the item's duration, then close it and return to Kivy:
import subprocess, shutil from urllib.parse import urlparse from kivy.clock import Clock def play_weblink(self, url, duration): scheme = urlparse(url).scheme.lower() if scheme not in ('http', 'https'): Logger.warning(f"SignagePlayer: Refusing non-http(s) weblink: {url}") self.next_media() return try: browser = shutil.which('chromium-browser') or shutil.which('chromium') self._weblink_proc = subprocess.Popen([ browser, '--kiosk', '--app=' + url, '--noerrdialogs', '--disable-infobars', '--incognito', '--no-first-run', '--check-for-update-interval=31536000', ]) Clock.schedule_once(lambda dt: self._close_weblink_and_next(), duration) except Exception as e: Logger.error(f"SignagePlayer: Error opening weblink: {e}") self.consecutive_errors += 1 self.next_media() def _close_weblink_and_next(self): proc = getattr(self, '_weblink_proc', None) if proc and proc.poll() is None: proc.terminate() try: proc.wait(timeout=5) except Exception: proc.kill() self._weblink_proc = None self.next_media()Requirements / notes:
- Install Chromium on the player image (
chromium-browseron Raspberry Pi OS). - Ensure Chromium gets window focus over Kivy and is fully killed before the
next item, including on pause/stop/restart paths and on app shutdown
(
on_stop) so no stray browser window is left behind. - On Wayland/X11 the player already sets
SDL_VIDEODRIVER; verify Chromium launches on the same display/session.
- Install Chromium on the player image (
-
Embedded web view widget (
kivy_garden.webview, WebKit/GTK, or WebView2). Cleaner UX (stays inside the Kivy widget tree) but fragile and poorly supported on Pi/Wayland — only pursue if option 1 is unacceptable. -
Server-side screenshot fallback (no player change). If embedding a live browser is not desirable, DigiServer can periodically screenshot the URL and store it as a normal
imageitem; the player then needs no changes. This loses live/animated content. Documented here for completeness only.
4. Checklist for the player update
get_playlists_v2.py: skip download whentype == 'weblink'.get_playlists_v2.py: preservetypein the rewritten playlist items (fixes the current loss oftype).get_playlists_v2.py: keep the original weburlfor weblink items.main.pyplay_current_media(): branch toplay_weblink()before theos.stat()file check.main.py: implementplay_weblink(url, duration)(Chromium kiosk).main.py: validate scheme ishttp/https; reject others.- Kill/cleanup the browser process on next item, pause, stop, restart, and
on_stop. - Install Chromium on the player image / document it in the player README.
- Test: mixed playlist (image → video → weblink → image) cycles correctly
and respects per-item
duration.
5. Security notes
- Only allow
http/httpsschemes on both server and player; never openfile://,chrome://, etc. - The server validates and stores the URL when the operator adds it; the player should still re-validate the scheme before launching the browser (defence in depth).
- Consider running Chromium with
--incognito(no persistent cookies/cache) as shown above.
6. Shipped architecture (src/weblink_session.py)
The player-side implementation lives in one module, so launch, verification, timing and teardown have a single owner instead of being duplicated per platform:
| Piece | Responsibility |
|---|---|
WeblinkSession |
Owns one weblink item: validate → launch → verify → watch → teardown. Generation-tokened so stale callbacks are ignored, and atexit-safe. |
WeblinkAdapter |
The only platform-specific part: launch / wait for the window / is it alive / tear it down / pre-warm. extra_launch_args() lets a subclass add browser flags without copying launch(). |
ChromiumSubprocessAdapter |
Default engine (Raspberry Pi chromium; on Windows the Chrome/Edge fallback). |
InteractionWatcher |
Decides when the item is finished (see the interaction model below). |
WebInputSources |
Reads /dev/input/event* (Linux) and does a pointer-position tap (Windows, needed for embedded engines). |
webview2_browser.py |
Windows, preferred: embeds WebView2 as a child HWND of the Kivy window. |
webview2_runtime.py |
Windows: detects the WebView2 Runtime and installs it silently when missing. |
Platform wrappers inject their engines through
SignagePlayer.weblink_adapter_factory:
- Raspberry Pi / Linux — built-in Chromium subprocess adapter.
- Windows (
windows/run_win.py) — engines are tried in this order:Order Engine Renders Notes 1 WebView2 ( webview2_browser.py)child window inside Kivy Preferred. No subprocess, so no background/z-order/hand-off/leak problems. 2 CEF ( cef_browser.py)child window inside Kivy Dormant: cefpython3has no wheels past Python 3.9.3 Chrome/Edge subprocess separate window Last resort only; retains the old drawbacks.
WeblinkAdapter.extra_launch_args() is the hook subclasses use to add flags
without duplicating launch() — the Chrome adapter uses it for
--user-data-dir + --kiosk.
Do not give the factory a class-level
Nonedefault combined with an unconditional instance assignment.SignagePlayer.__init__originally setself.weblink_adapter_factory = None, which shadowed the class attribute the Windows wrapper installs — so the platform adapters were silently ignored and every weblink fell back to the generic adapter and failed. It now only sets the instance attribute when the class attribute is absent.
6.2 Windows: WebView2 Runtime
WebView2 is two separate things, and they ship differently:
- the SDK (
Microsoft.Web.WebView2.Core.dll,WebView2Loader.dll) — the API surface, bundled in the exe fromwindows\webview2_sdk\(~860 KB); - the Runtime (
msedgewebview2.exe) — the actual Chromium engine, shipped by Microsoft and verified/installed at start-up bywindows\webview2_runtime.py.
If the Runtime is absent the player runs an installer silently
(/silent /install) and unelevated, which produces a per-user install and
therefore never raises a UAC prompt on the signage display. A ~1.7 MB online
bootstrapper is bundled by default; a ~203 MB offline standalone installer can
be bundled instead (see windows\webview2_runtime\download_runtime_installers.ps1)
for machines with no internet.
Install success is decided by re-reading the installed version, not by the
installer exit code — Edge Update returns a non-zero HRESULT (e.g.
-2147219416) when the Runtime is already current, which is not a failure.
6.1 Interaction model — web links are not passive media
duration on a weblink is not a hard cut-off. The player advances only when
both conditions are true:
- the configured
durationhas elapsed; and - the viewer has not interacted with the page for
interaction_postponeseconds (default 10 s), measured from the most recent interaction.
Consequences:
- A viewer who taps, scrolls or navigates the page during the final seconds of the slot keeps the page on screen — the advance is pushed 10 s past that touch, and every further touch pushes it again. The link is never pulled out from under someone who is using it.
- An untouched page still advances on schedule, exactly like a media item.
- A multi-event burst (a drag, a page transition) counts as one interaction but the countdown is measured from the last event of that burst, so an item can never be cut off mid-gesture.
max_dwell(duration ×max_dwell_factor, at leastmin_max_dwell) is an absolute backstop so a wedged browser or a jammed touchscreen cannot park the playlist forever.
Pause/play does not apply to web links. A web link is an interactive
surface, so toggle_pause() is a no-op while one is on screen — the interaction
watcher owns its lifecycle. The pause button continues to work normally for
images and videos.
6.2 Verified start-up
Launching a browser is not the same as displaying a page. The session therefore does not report success immediately after spawning the process (that used to reset the error counter and leave a black screen for the whole duration). The watcher thread — never the Kivy main thread — waits for the browser window to appear, and if it never does the item is reported as failed and skipped.
6.3 Configuration
All timings are tunable in config/app_config.json under weblink:
"weblink": {
"engine": "auto",
"interaction_postpone": 10,
"interaction_debounce": 0.5,
"interaction_grace": 5.0,
"max_dwell_factor": 6.0,
"min_max_dwell": 300,
"launch_timeout": 15,
"prewarm": true
}
| Key | Meaning |
|---|---|
engine |
Preferred engine (auto, cef, subprocess). |
interaction_postpone |
Seconds the advance is postponed, measured from each interaction (default 10). |
interaction_debounce |
Logging/trace throttle for continuous drags (default 0.5). |
interaction_grace |
Settle window after the last raw event still counted as interacting (default 5). |
max_dwell_factor |
Hard ceiling = duration × factor. |
min_max_dwell |
Floor for that hard ceiling, in seconds. |
launch_timeout |
How long to wait for the browser window to appear. |
prewarm |
Pre-warm the next weblink (disabled on Windows). |
6.4 Diagnostics
The watcher traces structured events through playback_trace.py:
weblink_launch, weblink_visible, weblink_interaction, weblink_end
(with reason viewer_idle, browser_exited or max_dwell),
weblink_not_visible and weblink_failed.