1Decision
Complete Task 9 of the lynx OTA update server plan for real: a genuinely newer candidate (2026082611), built, signed, published, and installed end to end through the Updater app itself — the first time this project has ever completed a real, app-driven OTA install. Fixed a real, previously-undiscovered bug in tools/signing/publish-ota.sh found along the way.
2Why and evidence
Built candidate 2026082611 (OFFICIAL_BUILD=true, fresh BUILD_NUMBER/BUILD_DATETIME, m vendorbootimage vendorkernelbootimage target-files-package otatools-package), confirmed Updater/Auditor/PdfViewerGOS/Seedvault all present in the real target_files.zip directly (unzip -l, not installed-files.txt, which this build target doesn't produce).
Staged lynx-target_files.zip/lynx-otatools.zip to <build-root>/unsigned/2026082611/; the owner ran sign-ota.sh 2026082611 on the signing host with both hashes independently re-computed on the build host, matching this project's established dual-hash-gate bar — ceremony succeeded, vault closed, machine powered off automatically.
The first publish attempt (publish-ota.sh 2026082611 20260826111823, using sign-ota.sh's own internally-generated release timestamp as the second argument, matching the only precedent this project had) let the app download and independently verify the OTA package's whole-package signature successfully (real, positive confirmation the otacerts/ADR-049 fix works end to end) but then failed with GeneralSecurityException: incremental does not match server metadata from app.seamlessupdate.client.Service.onDownloadFinished — the first time any real device check in this project's history had ever reached that code path, since every earlier attempt failed earlier (otacerts) or exited early ("up to date").
Root-caused by reading the real Service.java source directly rather than guessing: the app compares the OTA package's own embedded post-build-incremental metadata against the server's published metadata field, and separately constructs the download filename as <device>-ota_update-<that incremental>.zip — both must equal the real Android ro.build.version.incremental (2026082611), not sign-ota.sh's unrelated internal release-build timestamp (20260826111823) that only exists to name the signed file on the signing host's side.
publish-ota.sh had conflated these two distinct numbers into one CLI argument since it was first written; never caught before because no real installation attempt had ever exercised this check.
Fixed publish-ota.sh to read the real incremental directly from target_files.zip's SYSTEM/build.prop (already opened for ro.build.date.utc) and use it for both the served filename and the metadata's first field, keeping the CLI argument scoped to locating the signed file by its real on-disk name.
Re-deployed and re-published; the app then genuinely completed a full download, signature verification, payload application to the inactive A/B slot, and — after the owner triggered a real reboot — a real slot switch (ro.boot.slot_suffix confirmed _b→_a) into the new build, confirmed via direct device evidence: ro.build.version.incremental=2026082611, ro.build.date.utc=1787742365 (exact match), 0 Dropbox entries, the Updater app's own waiting_for_reboot flag correctly reset to false by BootReceiver on the real boot, and otacerts.zip still CN=GrapheneOS.
A false-positive "installed" report was caught mid-session: the app's own UI/behavior after payload application (before any reboot) reads as complete, but waiting_for_reboot remaining true in the app's real persisted preferences proved no reboot had actually occurred yet — the fix was simply asking the owner to reboot, not a real defect.
3Consequences
Task 9, and with it the entire docs/superpowers/plans/2026-08-23-lynx-ota-update-server.md plan (all 9 tasks), is complete for the first time: a real signed OTA, built from this project's own source, delivered over the private LAN-only server, and installed through the device's normal update mechanism, with a fully closed trust chain (AVB, whole-package release signing, and now a confirmed real install) all verified on physical hardware.
publish-ota.sh's fix is durable and applies to every future publish, not just this one. A UI/state-only "installed" claim from the app is not sufficient evidence of a completed update on its own — check the app's own persisted waiting_for_reboot preference (or equivalent) and the real active A/B slot before trusting it.
4Review when
If a future device or build ever needs an actual vendor.img change delivered via this same pipeline (this candidate's OTA did not rebuild vendor.img itself, only vendorboot/vendorkernelboot), re-verify publish-ota.sh's incremental-reading logic still resolves correctly against whichever partition's build.prop the device's live property system prioritizes — this fix reads directly from the target-files zip's SYSTEM/build.prop, sidestepping the vendor-vs-system property-merge quirk already documented in the rollback-corruption incident, but that quirk should be re-confirmed non-blocking if the delivery mechanism itself ever changes.
Read from docs/governance/decision-log.md, entry ADR-050.