Evidence
What an agent can leave behind to show what happened, and how it looks inside an app.
record with action start (file, maxSeconds, default 1,800) and stop, one recording per device. It
outlives the process that started it, so appdyno record start and appdyno record stop work as two commands. A session’s
end stops a recording it left running.
stop reads the file, and a file that does not play is a refusal saying why (a recorder killed before it ended leaves no
index). The answer gives how long the file plays, its frames and size.
| how | what to know | |
|---|---|---|
| Android | screenrecord on the device, 3 minutes a run, 4 Mbps |
runs are joined into one video by copying their frames, with no dependency; a moment is lost at each join. Runs an encoder change makes unjoinable stay as name-1.mp4, name-2.mp4, and the answer says so |
| simulator | simctl io recordVideo --codec=h264, a QuickTime .mov |
a frame is written only when the screen changes, so a still screen gives few frames; the time the answer gives is how long the file plays, past its last frame |
| iPhone | its own stream into a movie, with no second encode | HEVC 4:4:4 at the phone’s own size, from a key picture on; a build without the stream uses devicectl device capture screen-record, which not every phone offers |
Report
Section titled “Report”report start (dir) keeps every call from then on: its arguments, its answer and, after anything that is not a plain read,
the screen and an image of it. stop writes index.html in that folder, with the video recorded meanwhile, the app’s log, its
crashes and the measurements taken. So does the end of a session. write writes it at any moment, and with no start it is a report
of the screen as it is.
The page loads nothing: it has no script, and its content policy allows only its own folder. Listed secrets and what looks like a credential are hidden in all text. Typed text, and any argument not known to name something, is shown by its length. An image is painted over where an element holds a secret or is a password field, and left out if it cannot be painted over. Text drawn with no element behind it cannot be told, and a video is not painted over.
Timeline
Section titled “Timeline”report timeline, then the journey, then report stop writes timeline.html beside the report: the video, touches, frames, requests, the log and memory on one clock, in one file to send. report flight keeps the last seconds of an app and saves them on a crash or a stall. From a terminal: appdyno timeline and appdyno flight. See One timeline in the performance reference, which also says what the file holds before you send it.
Dashboard
Section titled “Dashboard”appdyno perf survey --app A [--device S] # every measure on that device in turn, then the pageappdyno perf dashboard [--app A] [--out dir] [--open] # the page alone, from what is keptOne page of an app’s performance on all your devices: what needs attention, what Google Play would count as a slow or frozen session (one session here, never a rate across users), each measure by device (its newest value, marked if it is an estimate or was taken under noise, or why that device cannot read it), and a trend of each with its baseline’s band and budget.
It says that an emulator’s or a simulator’s numbers are relative to this Mac, and that real frame rates, frozen frames, hitches and energy need a phone. Where a number means something else on a device, it says what: a simulator’s frames are main-thread pacing, and its energy is this Mac’s. With no phone measured for an app, it gives the command for one.
The page is <measurements dir>/dashboard/index.html, rewritten after each kept run, and an open page reloads itself. There is no server: one file with no script, which loads nothing. perf_compare with action: dashboard writes it too, and Open Dashboard in the Mac app opens it. See the dashboard in the performance reference.
More to read, each in the performance reference:
perf_snapshotalso: surfacereads the frames of aSurfaceViewor a game, whichgfxinfonever sees, against Google Play’s slow sessions.perf_launchwithbaselineProfile: truesays of an Android release build whether its Baseline Profile is in the APK, whether it is applied, and what it is worth in cold starts with it and without: A Baseline Profile.perf_snapshotalso: batterygives an Android app’s battery signals: wake locks by tag, wakeups and jobs, with no power model.perf_snapshotalso: anonMemorygives its memory as Google Play judges it: anon RSS and swap by process state, and bitmaps.perf_snapshotalso: startInfosays why it last started, from Android 15.perf_tracesays why slow frames and hangs were slow.perf_compareaction: functionssays which functions moved between two profiles.
Network
Section titled “Network”network paths says what each way sees on this device, and why not. start picks one, read and stop give a bounded list.
| way | where | sees | bodies |
|---|---|---|---|
proxy |
Android: settings put global http_proxy, through adb reverse for one on this Mac, put back at stop or the session’s end. Simulator: follows the Mac’s proxy, and this tool does not change that |
every HTTP request of an app that keeps the system proxy; for HTTPS only the host unless your proxy decrypts it | only in the proxy’s export, only where it decrypted |
probe |
simulator: an NSURLProtocol injected into the app, which restarts |
URLSession requests: method, URL, status, time, size. Not background sessions, WKWebView or Network.framework |
no |
bytes |
Android: dumpsys netstats |
an app’s bytes and packets; not the loopback | no |
packets |
Android 14+: Perfetto’s android.network_packets |
an app’s packets by remote port and direction, read at stop |
no |
The proxy is one you run. read lists what it saw: by the proxy’s own export where this tool can call it, and from any proxy by a file,
har, its session exported as HAR. A list holds what went through the proxy from any client, since start. No
certificate is ever installed: to see HTTPS through a proxy, install its certificate yourself. An iPhone has only read with har: its proxy is in its Wi-Fi settings.
Web views
Section titled “Web views”webview: pages, elements (the DOM at tap coordinates, to read where the accessibility tree does not), click (a CSS
selector), eval (a script). On Android it speaks the DevTools protocol to the webview_devtools_remote_<pid> socket of a debuggable
app, through adb forward. On a simulator it speaks to the socket Safari’s Web Inspector uses, and from iOS 16.4 a page shows only if the app set
isInspectable. An iPhone is not reached this way. One connection to a simulator’s inspector serves a call and the calls in the minute after it, as the inspector serves a new one only some 10 s after the last one closed; a new one is waited for as long as the inspector takes, up to 15 s and once more on a fresh connection, and a timeout names the inspector. On the command line each call is a process of its own, so a call within 10 s of the last can take up to those 10 s. On a simulator, pages marks the tab that shows as (showing), by what each page’s own document says, and webview and screen read that one; on Android it lists each page once. With no page to inspect it says that a page still loading is listed once it has loaded, and to ask again.
screen marks what a web view shows as web, and adds what the page holds and the accessibility tree does not, by coordinates and
selector, waiting at most 3 s for the page, and says when it left it unread. An app that is not debuggable still shows what its web view tells accessibility.
Flutter
Section titled “Flutter”A Flutter app’s semantics are in the accessibility tree, so screen, selectors and taps work on it. The richer tree of the Dart VM
service is not built yet: this tool ships only what it has tested, and there is no Flutter test app to test it against.
It would go as webview does: find the Observatory or Dart VM service line in the log of a debug build, forward its
port, and call ext.flutter.inspector.getRootWidgetSummaryTree over its WebSocket.
React Native
Section titled “React Native”react_native talks to Metro on this Mac and needs no device: targets, logs (the JS console over the Hermes
runtime’s DevTools protocol, for ms, secrets hidden) and reload (Metro’s route, else the runtime’s). Tested against a stand-in for Metro only, not a real app. Left out: the
component tree and JS profiles. On Android, a device reaches Metro through your own adb reverse tcp:8081 tcp:8081. metro is localhost, 127.0.0.1 or [::1] and a port, so this Mac calls no other host, and a runtime’s socket that Metro names on another host or port is not opened.
Visual comparison
Section titled “Visual comparison”visual baseline saves the screen as name under dir (default ./appdyno-baselines, to commit with your project). compare refuses
with CHECK_FAILED past tolerance (percent of pixels, default 0.1; a pixel counts if a channel differs by more than 16), and writes
name.diff.png (what differs in red, masks in blue) and name.actual.png. masks are x,y,w,h in tap units or text, id, role; the status bar is always left out.
A baseline holds for one size of screen, and the refusal says so.
This site counts visits with Cloudflare Web Analytics, without cookies.