ObliviOS Developer portal Project site

Home / Decisions / ADR-055

Land Task 3 of 9 of the Duress decoy user switch plan.

Accepted, recorded 26 August 2026.

Decision
ADR-055
Date
2026-08-26
Status
AcceptedAccepted
Part
Duress Unlock
Mentions
ADR-054
Mentioned by
ADR-056

1Decision

Land Task 3 of 9 of the Duress decoy user switch plan (DecoyProfileState, decoy routing-state storage) under ADR-054's local-first discipline, and record two real process defects this task surfaced in that discipline itself.

2Why and evidence

DecoyProfileState.java (off/registered routing state via LockSettingsStorage, no new callers yet) was authored in a disposable local worktree (aegisos-platform/.worktrees/duress-decoy-task3/) and landed on the real aegisos-duress-decoy-switch branch with git am --3way, applied cleanly — real remote commit 204a46b432f3.

Two defects were found while doing this: (1) ADR-054's own citation and aegisos-platform/credential/duress-unlock-decoy/SOURCE.md's Task 2 entry had cited 591c7e526e6ee9de5941245f0d765defd0fee930 — the disposable worktree's own local commit SHA — as if it were the real branch commit.

Direct inspection of the real branch (git log on aegisos-duress-decoy-switch) found the real commit is c05c6b25fae8: git am reconstructs the commit against the real tree (different parent, different committer timestamp), so it produces a genuinely different SHA even for byte-identical patch content — this was assumed, not verified, when ADR-054 was written. Corrected in SOURCE.md.

(2) Task 3's patch was git am'd onto the branch while Task 2's own verification build (m systemimage, launched beforehand, ~8200 total actions) was still actively running.

That build's ninja graph was generated at invocation, before DecoyProfileState.java existed — a build that started before a new file lands cannot discover it (Soong only re-globs for new files at each fresh soong_ui invocation), so letting that build finish and pass proves nothing about Task 3, only about Task 2's three already-tracked, already-modified files (which were present when the build's ninja graph was generated, so that part of the verification is still valid).

3Consequences

Two durable rules added to AGENTS.md/docs/research/build-and-signing-engineering-notes.md: always confirm the real post-git am commit SHA via git log on the actual branch before citing it anywhere — never assume a disposable worktree's local commit SHA carries over; and never git am a new patch onto a build-server branch while a build against that same tree is actively running — new files added mid-build are silently excluded from that build's result.

DecoyProfileState.java is exported as aegisos-platform/credential/duress-unlock-decoy/patches/frameworks-base/0004.

Task 3 is not build-verified yet — a fresh build after Task 2's current one finishes is required, and per the batching decision made this session, that fresh build will be deferred until Tasks 4-6 also land, then run once for all of them together rather than once per task, given each full build's real cost (~100+ minutes observed for Task 2's).

4Review when

Do not treat Task 3 (or Tasks 4-6, once landed the same way) as build-verified until a build launched after the last of them lands actually completes clean. If a future task's patch must land while another build is active (e.g. time pressure), explicitly flag in that task's own record that verification is deferred, the same way this entry does — do not let a coincidentally-clean concurrent build imply false confidence.

Read from docs/governance/decision-log.md, entry ADR-055.