pneumo

M4: Managed-device end-to-end smoke against a real server (ENG-268) #31

Merged

demi.dev 1w ago

targeting main from stoopkidddd/eng-268-m4-managed-device-end-to-end-smoke-against-a-real-server

Merged opened by and assigned to

demi.dev 1w ago targeting main from stoopkidddd/eng-268-m4-managed-device-end-to-end-smoke-against-a-real-server

Linear: https://linear.app/engineereddev/issue/ENG-268

What's in this PR (three commits):

Harness scripts/e2e/android.ts (now mise run e2e:android): harness MinIO and the built server on a random port (PUBLIC_URL=http://10.0.2.2:<port>, proxy mode), dev-client setup + pair-uri, a second cli sender session, credentials handed to the device through an app-private file (not runner args), the managed-device task, adb Doze control, a secret scan over Gradle output, instrumentation results, server log and logcat, and full teardown.
RealServerSmokeTest (UiAutomator + Compose): deep-link pair → Confirm → Pair; Connected; note → Copy → clipboard read from the foreground MainActivity; link → Open (ACTION_VIEW via espresso-intents) and Copy; 1 MiB file → Download → Open/Share → SHA-256 via MediaStore; Home composer send-back confirmed by the harness; Home + forced Doze + note + unforce → notification within 30 s. Skips cleanly without harness arguments, so the plain Gradle task still runs the other 14 tests.
Production fix (separate commit, bug from ENG-266): the 1 MiB download made 101 notification posts in one second, Android shed the terminal "Download complete" post, and Open/Share never appeared. Upload/download progress posts now need a changed percentage and ≥ 500 ms spacing; initial and terminal posts bypass the throttle. TransferProgressThrottle + JVM tests.
Test-only dependencies: UiAutomator 2.3.0, Espresso Intents 3.7.0. mise.toml e2e:android task and app/build.gradle.kts runner settings changed.

Verify (lead, on 8176869):

mise run check:android → exit 0
./gradlew testDebugUnitTest --rerun → exit 0 (350 tests, 0 failures)
mise run e2e:android → exit 0, twice in a row; each run: "PASS managed-device suite including real-server smoke", "Secret scan (Gradle, instrumentation results, server, logcat): 0 matches", "PASS emulator, server, MinIO and credential cleanup"
Afterwards: no pneumo-e2e- containers, no MinIO or emulator listeners.

Deviations: proxy mode only (permitted); ATD has no interactive shade, so the test fires the real posted notification's PendingIntents; clipboard read from the foreground activity (no debug probe); credentials via an app-private file.

🤖 Generated with Claude Code

Commits 4

M4: Share transfer progress notification gate (ENG-268)

+187 -45

M4: Throttle transfer progress notifications (ENG-268)

+148 -11

M4: Reuse Espresso's transitive Test Rules dependency (ENG-268)

+4 -5

M4: Add real-server Android managed-device smoke (ENG-268)

+1337 -9

Comments

v0 1w ago

demi.dev 1w ago

pneumo-reviewer: round 1, REQUEST_CHANGES #

The harness and real-server smoke are solid: every issue step runs against the built server and MinIO; teardown works on success and on SIGINT; the secret scan is real (independent counts: 0 live k=, 0 Bearer headers, 0 token/mk fields); the existing 14 tests pass and the new test skips cleanly without the harness. Send-back race closed correctly; Doze force/unforce restored in finally and teardown.

FIX (score ≥ 5)

often=3 bad=4 score=12 — transfer/TransferProgressThrottle.kt:5-24: the per-transfer 500 ms gate isn't enough. WorkManager 2.12.0 turns each setForegroundAsync into two enqueues (notify + startForeground), and Android drops updates above 5/s per package. Modelled with the real constants, the terminal Open/Share post is still dropped ~2–8% for 3–10 s downloads, ~17% for 10–30 s, and 16–65% for two overlapping downloads (a normal flow). The e2e can't see it: 1 MiB finishes in < 500 ms. Fix: one process-wide gate at ≥ 1 s (0% loss in the model), and a JVM test that two throttles share it.

FILE (score < 5, recorded on ENG-275)

score=4 — the download side of the throttle has no JVM test that fails on the old code (Robolectric has no rate limiter). Round 2 adds one.
score=2 — the coordinator listens on 0.0.0.0 (LAN-reachable unauthenticated endpoints for ~3 min); 127.0.0.1 works since 10.0.2.2 maps to host loopback.
score=1 — scripts/ci/check.sh doesn't type-check scripts/e2e/android.ts (clean today).

Spec §17, §14.3, §14.4, §14.8 OK; §14.6/§14.9 at risk only through finding 1. Next: omp round for finding 1.

v1 1w ago

demi.dev 1w ago

pneumo-reviewer: round 2 (head 3e1396a), APPROVE #

The round resolves the round-1 FIX; only the six expected files changed (no notify/**, no hot files).

FIX resolved: upload and download share one synchronized, process-wide 1 s gate (TransferProgressThrottle.Gate via the companion sharedGate); percent de-duplication stays per transfer; initial and terminal posts never go through the gate. The reviewer's rate model (AOSP 14 constants, two enqueues per WorkManager update) now shows 0% terminal loss for fast, long, overlapping and back-to-back downloads.
Download-side test resolved: downloadForegroundUpdatesArePacedAndTerminalActionsBypassTheGate decrypts 1 MiB with a virtual clock and checks pacing, bytes and terminal Open/Share; with the round-1 callback restored it fails ("First progress post must wait for the gate").
Fresh e2e on the new head: 15/15, 0 "Shedding" lines in logcat, secret scan 0 matches.
No new findings; the two round-1 FILE findings stay on ENG-275.

Verify (reviewer and lead): check:android exit 0; testDebugUnitTest --rerun 353/0 failures; mise run e2e:android exit 0. CI 4/4 on 3e1396a.

Versions

pneumo-reviewer: round 1, REQUEST_CHANGES #

pneumo-reviewer: round 2 (head 3e1396a), APPROVE #