ObliviOS Developer portal Project site

Home / Decisions / ADR-103

Sign candidate 2026092501 as a direct-flash image bundle via sign-baseline-images.sh, after three failed attempts on the same host and step.

Accepted, recorded 27 September 2026.

Decision
ADR-103
Date
2026-09-27
Status
AcceptedAccepted, fully closed — see ADR-105 for the completed second-machine verification.
Mentions
ADR-105
Mentioned by
ADR-099, ADR-105

1Decision

Sign candidate 2026092501 as a direct-flash image bundle via sign-baseline-images.sh, after three failed attempts on the same host and step.

2Why and evidence

Attempt 1 (2026-09-26): the live sign-baseline-images.sh on the signing host was a stale copy with a real regression in its otacerts test-key check (whitespace-insensitive match missing), fixed by replacing it with the repo copy via a diff staged on the shared unsigned/ mount.

Attempt 2: delta_generator (payload signing, part of GrapheneOS's own unconditional ota_from_target_files step inside generate-release.sh — confirmed against github.com/GrapheneOS/script branch 17 that this step cannot be skipped even when only an image bundle is wanted) crashed exit -11 (SIGSEGV on a null pointer) right after swap had been disabled for the key-in-swap-risk fix; kernel log showed __vm_enough_memory: ... delta_generator, no enough memory for the allocation.

Attempt 3: zram (2GB, RAM-only, no disk exposure) was tried as a safer substitute for disk swap but the OTA payload is dense, poorly-compressible data, so it provided little real headroom; delta_generator was OOM-killed for real (exit -9). Fix: restored the original disk swap (uncommented in /etc/fstab, matches the exact configuration all three earlier successful ceremonies — 2026081901, 2026082507, 2026082611 — ran under).

Attempt 4 succeeded: AVB verification passed, otacerts.zip subject in the signed target_files: subject=CN = GrapheneOS, published lynx-img-20260927140013.zip.

3Consequences

The theoretical unencrypted-swap key-exposure risk (see the engineering notes) is back in effect — no different from every ceremony this project ran before 2026-09-26, but explicitly re-accepted here rather than quietly reverted.

Independent re-verification (checklist step 9, "two separate machines") was not completed: the current release's AVB public key exists only inside the vault, not as a pre-extracted PEM on the build host, so avbtool verify_image there failed with "Embedded public key does not match" against the stale <build-root>/keys/alpha/alpha_key_pub.pem (a different, earlier key generation, confirmed by the mismatch itself).

What was checked from the build host instead: the bundle's own sha256, its 25-file/13-AVB-partition set matches the proven flash list exactly, rollback index 1788220800 is higher than the last accepted candidate's 1783209600, and otacerts.zip shows the real cert subject. The ceremony's own internal avbtool verify_image against the real vault key did pass, satisfying one of the two required independent checks, not both.

4Review when

Extract the vault's current AVB public key (PEM, not secret) to the shared signed/ mount during a future ceremony so the build host can do a real second-machine check without needing another full signing run; the flash fails or the device does not boot; a future candidate's payload is small enough to not need this swap workaround, in which case re-investigate whether swap can stay off.

Read from docs/governance/decision-log.md, entry ADR-103.