ObliviOS Developer portal Project site

Home / Decisions / ADR-085

Narrow the decision docs/superpowers/specs/2026-09-03-duress-decoy-user-switcher-visibility-design.md made.

Accepted, recorded 14 September 2026.

Decision
ADR-085
Date
2026-09-14
Status
AcceptedAccepted for Components A and B; plan status is paused, not closed
Part
Duress Unlock
Mentions
ADR-061, ADR-082
Mentioned by
ADR-086, ADR-089, ADR-213

1Decision

Narrow the decision docs/superpowers/specs/2026-09-03-duress-decoy-user-switcher-visibility-design.md made — that pre-unlock concealment was out of scope, only the decoy's own session mattered — to cover pre-unlock surfaces too; land Component A (pre-unlock concealment) and Component B (single-credential-entry decoy unlock, a three-grant authorization primitive) of docs/superpowers/plans/2026-09-07-duress-decoy-seamless-unlock.md in full; leave the plan's Task 15 (hardware closure) explicitly paused, not complete.

2Why and evidence

The owner asked (2026-09-07) to make Duress Unlock "perfect," naming three gaps the prior cycle's own design (2026-09-03-duress-decoy-user-switcher-visibility-design.md:136-142) had deliberately left open: the lock screen still advertised multiple users before any credential was entered; entering the decoy PIN demanded a second credential entry; the transition was visibly a user switch, not an unlock.

This design narrows that prior decision — the owner's own unlocked session remains untouched, but concealment now also applies to every surface reachable while the device is locked.

Component A (UserManagerService.isUserSwitcherEnabled's second early return, gated on a decoy existing so a decoy-free device stays bit-for-bit stock) landed as frameworks/base patches 0041+0044 and closed the one real chokepoint-bypassing surface Task 2's exhaustive re-audit found (the Multiuser home-screen-and-keyguard widget, reachable via Communal Hub, rendering with no isUserSwitcherEnabled() gate at all — packages/apps/Multiuser patch 0004).

Component B was built, tested, reviewed, and hardware-validated across five sub-cycles, each catching a real defect the previous one missed: Tasks 6-9 (patches 0045-0053) built and landed the original one-shot authorization and wired it into UserController's forced-lock-skip decision — deliberately not AOSP's own credential_capture/config_multiuserSkipKeyguardWhenSwitchingToUnlockedUsers condition, the exact generic mechanism ADR-061 already found was a device-wide keyguard bypass.

Task 11's real-touch-input testing (the first task in this whole cycle to drive the actual on-screen bouncer rather than the cmd lock_settings verify shell bypass every earlier task had used) found the single-credential promise was not actually delivered: a second, fully independent AOSP mechanism, KeyguardViewMediator.handleUserSwitchComplete(), still re-armed a live bouncer roughly 500ms after an authorized switch.

Tasks 11a-11b (patches 0058-0065) extended the authorization primitive to a second, independent single-use grant and wired it across the system_server/SystemUI process boundary to suppress that re-arm, closing the automatic prompt — verified with 3 clean real-touch trials.

That still left a swipe-revealed fresh bouncer, since nothing had ever told the newly-current decoy session's own keyguard state that it was already credential-verified; Task 11c (patches 0066-0072) added a third grant and reused a real, existing AOSP "already satisfied" dismiss chokepoint (the same one trust-agent and biometric unlocks already use) rather than suppressing anything — closing the single-PIN promise for real, confirmed 3/3 on real touch input, with the ADR-061 non-decoy regression case re-confirmed intact at every one of these five sub-cycles.

Task 12's consolidated build (all 32 frameworks/base patches plus the one Multiuser patch, reproduced byte-for-byte from real branch history) passed the ADR-061 hard gate on the emulator.

Task 13's systematic emulator sweep found two further real, disclosed (not fixed) gaps: Component B's failure-path fallback (DecoyRoutingFailureMarker.REASON_QUIESCE_FAILED) is dead code — quiesceUserApps cannot currently return false — and the owner's own "Select user" dialog can get stuck on stale data after a manual trigger, though the only reachable attack surface into it (the QS footer icon) is confirmed safe.

Task 14 built and alpha-key-signed a physical lynx candidate (fingerprint 265ade399a0f3ee65003e56daef26a8f4d70e6a4, 13/13 AVB, independently re-confirmed a third time by the controller after the dispatching session was itself interrupted mid-task and had to resume from confirmed-real build-server state rather than assume anything was lost).

The owner flashed it from their own workstation (files transferred and independently re-verified 13/13 before flashing, catching any transfer corruption first).

The first real physical-hardware trial of the decoy scenario confirmed the single-credential-entry fix holds on hardware, not just the emulator (no second "Verifying primary lockscreen credential" line in the real device's own logcat) — but surfaced two things real hardware use exposes that no emulator trial had: a swipe is still required between the decoy PIN and the decoy's home screen (the design's own "zero interaction, no lock screen at all" framing was never claimed as achieved, only "no second PIN prompt" was — this is that known, disclosed gap, now hardware-confirmed rather than newly found), and total transition latency (~6.5s) sits at the slow end of the emulator's own measured cold-attempt range.

The owner also reported, for the first time in this cycle, that the return path (decoy session back to the owner's own) requires two full password entries — a mechanism (ReturnToRealUserHelper) from an earlier project cycle that no task in this plan has read, tested, or touched.

3Consequences

Components A and B's core mechanisms are real, landed, reviewed, and repeatedly hardware/emulator-validated — do not treat either as unsafe or unbuilt.

But do not describe this plan as complete: Task 15's remaining steps (the ADR-061 gate re-run on physical hardware, Task 13's full checklist re-run there, and physical validation of the seven still-open carried-over patches from ADR-082) have not run, and two real, owner-confirmed usability gaps remain unfixed — the owner→decoy swipe requirement and the decoy→owner double-password return path.

The owner explicitly chose to pause here and resume in a future session rather than continue immediately; full detail for that session lives in docs/devices/lynx-duress-decoy-prelock-audit.md's "Task 15 (partial)" section, not repeated here.

4Review when

Before any future session describes this plan as closed: physical hardware must re-run the ADR-061 gate and Task 13's checklist, the seven ADR-082 carried-over patches need their own physical validation, and both open usability gaps need their own investigation-then-fix cycles with the same ADR-061-aware rigor Tasks 11a-11c required — do not assume either is a small fix without first tracing the real mechanism, the same discipline that turned what looked like one bug (Task 11's "second bouncer") into three distinct root causes across Tasks 11a-11c.

Read from docs/governance/decision-log.md, entry ADR-085.