Android driver native thread exhaustion in TcpForwarder on macOS with physical device (reproduces on 2.3.0 and 2.4.0)

[Auri850]

Summary

Maestro Android execution on a physical device consistently fails with native thread exhaustion in the Android forwarding path.

This reproduces across hosts and across two adjacent CLI versions, and reproduces with direct maestro test invocation outside our wrapper script.

Classification

• Status: HARNESS_DEFECT
• Reason: maestro_native_thread_failure
• Signature: pthread_create_or_eagain

Environment

• Host OS: macOS (Darwin 24.6.0)
• Java: Temurin OpenJDK 17.0.18
• Maestro CLI tested: 2.3.0 and 2.4.0
• Device: Samsung SM_G981U
• Device serial: RFCN200GA3F

Reproduction

1. Ensure device is visible:
   • adb devices -l shows RFCN200GA3F device
2. Apply reverse bindings:
   • adb -s RFCN200GA3F reverse tcp:7001 tcp:7001
   • adb -s RFCN200GA3F reverse tcp:8081 tcp:8081
3. Run direct Maestro command (outside wrapper):
   • maestro --device RFCN200GA3F test .maestro/_harness-admission-smoke.yaml --no-reinstall-driver

Also reproduced with isolated 2.4.0 binary from release artifact:

• /tmp/maestro-cli-2.4.0/maestro/bin/maestro --device RFCN200GA3F test .maestro/_harness-admission-smoke.yaml --no-reinstall-driver

Observed failure signatures

• pthread_create failed (EAGAIN)
• OutOfMemoryError: unable to create native thread
• stack includes dadb.forwarding.TcpForwarder
• gRPC DEADLINE_EXCEEDED

Representative stack frames include:

• dadb.forwarding.TcpForwarder.handleForwarding(TcpForwarder.kt:63)
• maestro.drivers.AndroidDriver.deviceInfo(...)
• maestro.orchestra.Orchestra.runFlow(...)

Additional observations

• Reproduces cross-host (Mini + laptop).
• Reproduces outside harness wrapper (direct Maestro invocation), so not wrapper-only.
• Device visibility and reverse bindings are healthy before failure.
• Constrained JVM probe did not resolve issue:
  • JAVA_TOOL_OPTIONS='-Xss256k -XX:ActiveProcessorCount=2'
  • Failure shape remained materially the same.

Limits snapshot from failing environment

• ulimit -u: 1392
• ulimit -n: 1048575
• launchctl limit maxproc: 2048 soft / 2500 hard
• launchctl limit maxfiles: 256 soft / unlimited hard

Expected behavior

Maestro should execute the flow on the connected physical Android device without exhausting native threads in forwarding setup.

Actual behavior

Execution starts, then fails during Android forwarding/runtime initialization with native thread creation failures and gRPC timeout.

[linear]

MAE-438

[Fishbowler]

No idea what the wrapper is. This is your own thing?

Also, what's the "Apply reverse binding" step? That's not something I remember from the Maestro docs.