fix(android): control emulators argent can't find over adb by blazejkustra · Pull Request #1397 · software-mansion/argent

blazejkustra · 2026-10-12T15:37:36Z

Problem

An Android emulator started outside argent under a different home directory (for example by another tool or user account) could not be screenshotted, tapped or swiped: argent did not find it, and each of those calls failed.

Approach

argent now falls back to controlling such an emulator over adb, the way it already controls physical phones. That path does not need to find the emulator first, so it works.

Tests

Live on an API 35 AVD started with HOME=<other dir> emulator -avd <avd> -read-only -no-window: on main, screenshot, gesture-tap and gesture-swipe failed (3 of 3); with this PR all three work (search opened, app drawer opened).
button and keyboard go through adb and work on both builds.
A plain emulator -read-only -no-window or token-less -grpc 8554 launch already works on main, so it is unaffected.
New blueprint unit test for the fallback; the full tool-server suite passes.
No docs change: the docs don't describe this limitation.
Not verified: Linux CI runners, Android Studio.

Screenshots

Before

android-emu-before.mp4

After

android-emu-after.mp4

…ontroller An emulator started outside argent under another HOME (a CI runner, a service account) has no gRPC discovery file where argent looks, so every screenshot, tap and swipe failed with 'not found among running emulators'. Retry such emulators with the screen-sharing (android_device) controller, which only needs adb.

coderabbitai · 2026-10-12T15:37:47Z

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

latekvo

[Diplomat, Opus 5.5]: Ran the PR on a real Pixel_6_API_34 emulator started under a separate HOME, comparing this PR with main and with a normally started emulator: screenshot, tap, swipe, pinch, two-finger gesture-custom, rotate, paste, keyboard, buttons, screen recording, stop-simulator-server respawn, and two tool-servers on one device. Ran the shipped Linux simulator-server in a container to check the discovery error paths. Also ran prettier, eslint, both tool-server typechecks and the full tool-server vitest suite (all green), checked that the new test fails on main, and mutation-tested the fallback guard. The core claim reproduces: on main, screenshot, tap and swipe fail on that emulator, and with this PR they work.

latekvo · 2026-10-13T21:52:15Z

+      `[sim ${device.id}] emulator not found through gRPC discovery; ` +
+        `falling back to the adb screen-sharing controller\n`
+    );
+    return spawnSimulatorServerProcess(device.id, "android_device");


[Diplomat, Opus 5.5]: Once an emulator falls back here, two-finger gestures report success but do nothing. Measured on a Pixel_6_API_34 started with HOME=<tmp> emulator -avd Pixel_6_API_34 -read-only -no-window, with Google Maps open:

gesture-pinch (centre 0.5,0.35, distance 0.1 -> 0.5) returned pinched:true.

gesture-custom with x2/y2 returned 200.

Neither changed the zoom in 3 attempts. A single-finger pan on the same server works.

The same two calls on main, against a normally started emulator on the android controller, zoom in and out. The device keeps kind: "emulator", so nothing downstream knows it is now behind the phone controller. Before this change these calls failed with SIMULATOR_SERVER_READY_EXITED; now they report a gesture that never happened.

Leaving open: making two-finger gestures work on the phone controller, or rejecting them for a fallen-back emulator (it keeps kind: "emulator"), is a design call for a human, not an unattended change.

latekvo · 2026-10-13T21:52:15Z

-      device.id,
-      subcommandForDevice(device)
-    );
+    const { proc, apiUrl, streamUrl } = await spawnWithEmulatorFallback(device);


[Diplomat, Opus 5.5]: A tap that is the first call to reach a fallen-back emulator is reported as done but does not land. Each trial: Settings open, search bar frame from uiautomator dump [42,591][1038,728] on a 1080x2400 screen, then stop-simulator-server, then gesture-tap at (0.5, 0.2747), then the top activity after 2 s.

Run Tap result Top activity

This PR, fallen-back emulator, tap as the first call tapped:true (~1.1 s) .Settings in 3/3

This PR, same, one screenshot before the tap tapped:true SearchActivity in 3/3

main, normal emulator, tap as the first call tapped:true SearchActivity in 3/3

Every respawn of this service is exposed: first use, after stop-simulator-server, and after the registry's connection-refused recovery. Each one hands back a server whose first touch is silently dropped.

Leaving open: the dropped first tap needs a warm-up or readiness fix verified on a real emulator, which I did not do unattended.

latekvo · 2026-10-13T21:52:15Z

+      `[sim ${device.id}] emulator not found through gRPC discovery; ` +
+        `falling back to the adb screen-sharing controller\n`
+    );
+    return spawnSimulatorServerProcess(device.id, "android_device");


[Diplomat, Opus 5.5]: Two tool-servers driving one fallen-back emulator evict each other. Each android_device start kills the screen-sharing agents already on the device, so the other server's simulator-server exits (status 143) and respawns on its next call, at 1.3-1.6 s per alternating call. The tap after such a respawn is the dropped first tap described on line 598. Two android servers on a normally started emulator coexist, with every frame fresh.

Leaving open: two tool-servers evicting each other comes from the android_device controller killing existing agents; fixing that is simulator-server work or a product decision on whether to accept it for this fallback.

latekvo · 2026-10-13T21:52:15Z

+      `[sim ${device.id}] emulator not found through gRPC discovery; ` +
+        `falling back to the adb screen-sharing controller\n`
+    );
+    return spawnSimulatorServerProcess(device.id, "android_device");


[Diplomat, Opus 5.5]: The phone controller leaves state on the emulator that the emulator controller never touches:

Each spawn adds an adb reverse for screen-sharing-agent-* that is never removed. There were 11 after one session of restarts.

The 9.5 MB agent stays in /data/local/tmp/.studio.

rotate sets accelerometer_rotation=0 and leaves it there. On main's emulator path it stays 1.

For an emulator someone else started, all of this persists after argent is gone.

Leaving open: cleaning up the adb reverse, the pushed agent and accelerometer_rotation on an emulator argent did not start needs a decision on what argent should own and undo.

latekvo · 2026-10-13T21:52:15Z

+      `[sim ${device.id}] emulator not found through gRPC discovery; ` +
+        `falling back to the adb screen-sharing controller\n`
+    );
+    return spawnSimulatorServerProcess(device.id, "android_device");


[Diplomat, Opus 5.5]: Screenshots through the fallback come from a 920x2044 stream instead of the emulator's 1080x2400. A default-scale screenshot is 230x511 here and 270x600 on the gRPC path of the same AVD. So screenshot-diff reports dimension_mismatch when one side was captured on the normal path and the other through this fallback.

Leaving open: the 920x2044 stream size comes from the phone controller; matching the gRPC path resolution (or documenting the mismatch) needs a human decision.

…ectory On Linux, a user with no avd/running directory gets 'Failed to find any running emulator' instead of 'not found among running emulators', so the fallback never triggered there. Match both, keep the emulator controller's error when the adb fallback fails too, and update the comments that said emulators always use the gRPC bridge.

blazejkustra changed the title ~~fix(android): drive emulators without a gRPC bridge through the adb controller~~ fix(android): control emulators argent can't find over adb Oct 12, 2026

latekvo reviewed Oct 13, 2026

View reviewed changes

blazejkustra added 2 commits October 22, 2026 00:21

Merge remote-tracking branch 'origin/main' into pr-1397-work

e728091

Run	Tap result	Top activity
This PR, fallen-back emulator, tap as the first call	`tapped:true` (~1.1 s)	`.Settings` in 3/3
This PR, same, one `screenshot` before the tap	`tapped:true`	`SearchActivity` in 3/3
main, normal emulator, tap as the first call	`tapped:true`	`SearchActivity` in 3/3

Conversation

blazejkustra commented Oct 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Approach

Tests

Screenshots

Uh oh!

coderabbitai Bot commented Oct 12, 2026

Draft PR not reviewed

Uh oh!

latekvo left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

blazejkustra commented Oct 12, 2026 •

edited

Loading