onev.cat / Prowl
native macOS codings agent orchestrator prowl.onev.cat
Star
Fork 0
Watch
Prowl MirrorClient Android docs validation.md
3.6 kB
Markdown
at agent-island-width

Android implementation validation #

Baseline: iOS 570a423, Prowl bbd8e6e4. Tests do not modify either repository, Ghostty, existing remote debug apps, or any physical device.

Executed locally (2026-09-12) #

  • Debug compilation and APK packaging succeeded.
  • JVM tests cover fixed binary vector/UTF-8/size rejection, pairing normalization and proof binding, Agent vs Shell routing, snapshot replacement/ACK, stale callbacks, draft revision, unknown receipt without replay, takeover, historical pagination, timeout overlap, Host-edit refusal and creation uncertainty after UI reopening.
  • A separate local Network.framework fixture using the actual Host TLS cipher accepted the Android transport: short-code enrollment → HMAC proof → save → device PSK reconnection → authenticated pane discovery. This proves library/API wire interoperability, not a complete real Agent workflow. Fixture is under ignored .local/ and its transient port is supplied only to the test process.
  • Phone emulator application launch and screenshot inspection succeeded.
  • 18 JVM tests passed with zero skips in the opt-in native fixture run: /tmp/android-mirror-close-tests.log and app/build/test-results/testDebugUnitTest/.
  • Two instrumented UI cases passed on the Android 16 phone emulator and again at 1280×800 / 160 dpi wide dimensions: pairing halves remain editable; explicit hardware Enter down/up twice submits, one Return only adds a newline. Logs: /tmp/android-mirror-close-tests.log, /tmp/android-mirror-tablet-tests.log. The original emulator display size/density were restored afterward.
  • Final APK/lint run: /tmp/android-mirror-final-build.log. Lint has zero errors; dependency age/target publishing and unused library JSSE warnings remain recorded.
  • Screenshot evidence: /tmp/android-mirror-phone2.png, /tmp/android-mirror-tablet.png. These are layout checks, not physical tablet validation.
  • Espresso was updated to 3.7.0 because Android 16 removed the reflective InputManager method used by older test runners. This addresses the documented AndroidX Test compatibility fix.
  • Final independent review confirmed the listed lifecycle fixes and found no remaining blocker in its reviewed paths; this is not a substitute for real devices.
  • Independent review compared Android to the current iOS implementation. Fixes: hardware Return key-up/repeat handling, explicit takeover, busy classification, per-request history timeout and lease reset, user-drag-only follow behavior, preservation of uncertain input on Host editing, 1 MiB text-frame bound, IME and selection cancellation, session-specific dialogs, session-owned launch state, and Unicode-safe detail chunking.

Not represented as passed #

  • Physical phone/tablet, vendor-specific IME, actual Bluetooth/USB keyboard.
  • Real Codex/Claude/Shell commands over a real Mac Host, cross-device takeover and network roaming. Client routing is independently tested; Host dispatch logic is not reimplemented or changed here.
  • Play Store signing, current target-SDK publishing requirements and release rollout.
  • App process death does not preserve output/drafts; credentials are retained.

Lint's remaining dependency-version notices reflect pinned versions. Bouncy Castle also ships unused JSSE trust-manager classes which lint flags; runtime uses only low-level PSK TLS with one allowed suite and a separate device-key challenge. These warnings are documented, not suppressed as an approval claim.