ObliviOS Developer portal Project site

Home / Decisions / ADR-047

Resolve a project-wide gap found while executing Tasks 6-8 of the lynx OTA update server plan.

Accepted, recorded 25 August 2026.

Decision
ADR-047
Date
2026-08-25
Status
AcceptedAccepted
Mentions
ADR-046
Mentioned by
ADR-048, ADR-059, ADR-072, ADR-246

1Decision

Resolve a project-wide gap found while executing Tasks 6-8 of the lynx OTA update server plan: no lynx build this project has ever produced actually included GrapheneOS's real Updater app, because OFFICIAL_BUILD=true was never exported on any of them; fixed, re-signed, re-flashed, and Task 8 (real device confirms the private update server) now passes on real hardware.

2Why and evidence

Task 6's rebuild of system.img (carrying Task 5's private-URL Updater patch) produced a build that, on inspection of installed-files.txt, contained zero trace of the Updater app despite Task 5's source patch compiling cleanly and the build reporting #### build completed successfully ####.

Root-caused to build/make/target/product/media_system.mk's ifeq ($(OFFICIAL_BUILD),true) PRODUCT_PACKAGES += Updater endif gate — never set on any build this project has run (the 2026-08-04 alpha candidate, the Aug-16 duress-primitive build, and this first Aug-24 rebuild all share the same gap).

A bounded search of device/, vendor/, build/make/, and the Updater app's own tree found OFFICIAL_BUILD referenced in exactly one other place (vendor/adevtool's own state-collection build script, which already sets it as standard practice) — confirming this is expected GrapheneOS convention this project's own build invocations were silently missing, not evidence of a wider structural gap (e.g. no vendor/grapheneos overlay is absent by mistake; GrapheneOS patches AOSP source directly for this class of change).

Rebuilt with OFFICIAL_BUILD=true added (BUILD_NUMBER=2026082417), confirmed Updater.apk present in installed-files.txt before proceeding. Task 7 (re-sign system+vbmeta, flash) was then re-executed a second time against this corrected image, reusing the exact fix procedure ADR-046 established (11 unchanged partitions sourced from the device's real last-flashed state, i.e.

Task 3's Aug-16 files, not the 2026-08-04 candidate; rollback index 1783209600 preserved) — independently verified 13/13 on two machines before flashing, confirmed booting clean (0 Dropbox entries) with app.seamlessupdate.client now genuinely present via pm list packages.

Task 8's "up to date" check then hit two further, unrelated real blockers before passing: (1) an earlier test publish from Task 4 was still live on the server and, being older than Task 5's patch, would have silently reverted the Updater URL back to releases.grapheneos.org if installed — cleared, and republishing the metadata file as an outright absence produced a 404 the app misread as a hard failure rather than "up to date" (a real mistake made and corrected in the same session: an absent metadata file is not equivalent to a non-newer one, and the fix was to republish a non-newer entry, not delete the file); (2) a genuine device-side WiFi MAC-randomization/router-recognition outage, unrelated to this project's own infrastructure, confirmed via live-correlated adb logcat/dumpsys netpolicy/dumpsys connectivity evidence showing zero on-device policy restriction throughout the failure window — the blocking was entirely upstream of the device, at the household router, and resolved on the owner's own end.

Once real connectivity was restored, the check passed immediately with no further changes: fetching metadata for lynx-alpha / targetBuildDate: 1786724916 not higher than sourceBuildDate: 1786724916.

3Consequences

Every lynx build this project produced before this fix was missing the real Updater app — a durable, general finding now recorded in docs/research/build-and-signing-engineering-notes.md (the OFFICIAL_BUILD entry under Soong/Android build system behavior) and as a standing cross-session rule: verify the relevant installed-files-<partition>.txt for any app whose presence actually matters, rather than trusting a clean build exit.

The device now runs BUILD_NUMBER=2026082417's system.img + vbmeta-ota-server-signed-v3.img, independently verified 13/13, booting clean, with the Updater app functional against the private server end to end (Tasks 1-8 of 9 complete).

Also confirmed during this session, independent of the OFFICIAL_BUILD fix: the the signing host/the build host virtiofs mount read/write permissions are the opposite of what docs/superpowers/specs/2026-08-20-the signing host-ceremony-storage-design.md documents (<build-root>/unsigned is read-only on the signing host, <build-root>/signed is read-write there — confirmed via live mount output on both VMs rather than trusted from the older doc), which changed where scratch signing work had to be staged this session.

4Review when

Task 9 (build, sign, and publish a genuinely newer candidate from this same OFFICIAL_BUILD=true-corrected tree; a real end-to-end OTA install test) is in progress as of this entry, not yet complete — revisit this record once it finishes. Also revisit if a follow-up audit (requested by the owner, in progress as of this entry) finds another similarly-gated PRODUCT_PACKAGES conditional silently affecting a security- or function-relevant component beyond Updater.

Read from docs/governance/decision-log.md, entry ADR-047.