ObliviOS Developer portal Project site

Home / Decisions / ADR-064

Task 12's closing whole-branch review of the entire third plan.

Accepted, recorded 30 August 2026.

Decision
ADR-064
Date
2026-08-30
Status
AcceptedAccepted, qualified (this corrects, not merely supplements, ADR-063's "fully closed" framing — see Consequences)
Mentions
ADR-058, ADR-059, ADR-061, ADR-063
Mentioned by
ADR-067, ADR-079, ADR-086, ADR-096

1Decision

Task 12's closing whole-branch review of the entire third plan (dispatched to an independent reviewer, most capable model, after ADR-063 had already declared the plan complete) found a real, Critical gap every per-task review in this plan had missed: DecoyProfileHelper.setDecoyProfile never validated that the registered decoyUserId is an existing USER_TYPE_FULL_AEGISOS_DECOY user, only that it isn't the real user. Fixed (frameworks/base 0024) along with three lower-severity findings, rebuilt for physical lynx, signed, flashed, and the fix independently re-confirmed on real hardware via a genuine thrown exception.

2Why and evidence

Every Component B/C fix in 0019-0023 keys its identity-leak filtering off UserInfo.userType, not off the registered id — an ordinary secondary user mistakenly registered as "the decoy" (a real scenario given USER_TYPE_FULL_AEGISOS_DECOY's own confirmed max-count-of-one restriction, which forced Task 11 Phase B's own registration test to reuse the device's single existing decoy rather than create a second one) would still route and switch correctly via maybeRouteToDecoy, but silently receive none of the identity-leak or setup-complete protections this whole plan exists to provide — the exact "hiding" failure mode the plan's own threat model targets, undetected by any of Tasks 1-11's own reviews or by Task 11's own physical-hardware validation, because that validation's registered decoy (id 10) was already correctly typed.

The controller independently confirmed the gap against the real live branch (not the reviewer's own disposable-worktree copy, which the reviewer itself flagged as unverified given no SSH/shell tool in its environment) before fixing it.

Three lower-severity findings were fixed in the same patch: config_multiuserSkipKeyguardWhenSwitchingToUnlockedUsers reverted to false at the resource layer too (belt-and-suspenders — the aconfig flag alone, build/release 0004, is sufficient on this device's current release config, but a future rebase or different release config could otherwise silently reactivate ADR-061's regression); a third unscoped Wi-Fi call site (WifiUtils.java:387, packages/apps/Settings 0003) the design doc named but Task 8 missed; and a needless @UnsupportedAppUsage removed from a brand-new method with no prior reflective callers by construction.

One reviewer finding was investigated and rejected as a mischaracterization, not fixed: a claimed Settings ▸ Storage "System" bucket leak turned out, on reading the real getSystemSize() source, to be a fixed volume-vs-partition figure identical for every user on the device including the real owner — not scoped to the per-user loop at all, so there is no differential tell there; this predates the plan and is out of Component B's scope, not a regression.

Both patches were rebuilt for physical lynx (OFFICIAL_BUILD=true reconfirmed via a fresh installed-files*.txt grep for all four gated apps across /system and /product; only system.img needed regenerating, the other 11 partition images were byte-identical to the prior candidate's own output since none of their inputs changed — one fully coherent 12-partition set signed together, not a splice of new and old partitions), signed with the same alpha key and rollback index, 13/13 independently re-verified on two machines, flashed by the owner, and independently re-confirmed clean (matching fingerprint, unchanged boot state, zero Dropbox crash/ANR/tombstone entries).

The fix itself was then decisively confirmed on physical hardware, not just by inspection: a disposable ordinary user was created, the owner ran the real set-decoy-profile registration against it themselves (the controller never saw or handled the real credential, same handoff pattern as fastboot flash), and it failed with java.lang.IllegalArgumentException: decoyUserId must be an existing org.aegisos.usertype.full.DECOY user at exactly the new check's own line, full real call chain intact.

The device's actual, correctly-typed decoy (id 10) was confirmed unaffected throughout.

3Consequences

Full evidence: .superpowers/sdd/task-11-phaseB-report.md's successor investigation (this review-and-fix cycle was done directly by the controller, not a task-numbered SDD entry, given its cross-cutting nature); aegisos-platform/credential/duress-unlock-decoy/ SOURCE.md/README.md updated with frameworks-base 0024 and packages-apps-settings 0003, and corrected to name ADR-059 through this entry (README previously stopped citing at ADR-058 and still claimed physical hardware validation was "not started," both stale since ADR-059/062/063).

4Review when

This is the actual close of the third plan's Task 11/12 — ADR-063's "fully closed" language should be read as superseded by this entry, not as an independent, still-valid milestone: a Critical gap existed in what ADR-063 called complete, and it is this entry, not ADR-063, that reflects the plan's real final state. Task 12's remaining step (superpowers:finishing-a-development-branch) proceeds from here; the owner has already indicated the long-lived aegisos-duress-decoy-switch branch should be kept as-is across stages, matching every prior stage's precedent, rather than merged or discarded.

Read from docs/governance/decision-log.md, entry ADR-064.