Repository navigation
Description
What is true
The sensitive tier has held real data since 2026-10-02 (build-the-sensitive-tier-host.md:642: Vaultwarden and Paperless). Its secrets and its volume backups open with one key, trinity's.
.sops.yaml:89-91: thesecrets/sensitiverule lists a single recipient,age1cpdxl….:82-83says so directly: "The technical second's recipient (ADR-0024, Add a second age recipient so decryption does not depend on one person #294) is still owed here; the tier has held real data since that day."grep -c recipient: secrets/sensitive.sops.yamlreturns 1.scripts/backup-volumes.sh:258-259,1564-1567encrypts the volume sets to the same recipient list. Sovaultwarden-data,paperless-db-dataand the rest open only with trinity's key.docs/runbooks/restore-the-sensitive-tier.md:93: the second recipient "has to be on thesensitiverule before the first real item goes in."
#294 ("Add a second age recipient…") is closed. It landed before the sensitive rule existed (2026-09-28, #688), so the new rule was born with one key.
The Immich and Paperless library sets are partly covered, because household.recipients adds the technical second. The volume sets and the secrets file are not.
Why it matters
If trinity's disk dies, Vaultwarden's database and every backup of it die with the only key that opens them. That is the failure ADR-0024 was written to prevent.
soc.sops.yaml and lab.sops.yaml also have one recipient each. Those stakes are lower, and the lab is #671.
What closes it
make secrets-add-recipient STACK=sensitive PUBKEY=<technical second's age1…>, run on trinity.updatekeysre-wraps the data key; nothing is re-encrypted by hand.- One
make backup STACK=sensitive, so the newest volume sets carry both recipients. make secrets-verify-backupproves the second key opens both the secrets file and the newest volume set.- Decide whether soc gets the same treatment now or under The lab secrets are encrypted to one key, and it lives only on alexander #671's approach.
- A guard so this cannot recur silently.
SecretsKeyBackupUnproven(backup.rules.yaml:182) checks only the recipients that already exist. Add a check, in CI (scripts/check_sops_rules.py) or as a gauge plus rule, that fails or fires when a stack marked as holding real data has fewer than two recipients on its rule.
Done when
- The secrets file and the newest volume set both decrypt with the technical second's key on a machine that is not trinity.
- The guard exists and fails once, on purpose, against a one-recipient fixture.
Refs #294, #688, #671. Found in the 2026-10-03 repository review.
Activity
fix(secrets): give check_sops_rules repair advice that cannot recreat…
fix(secrets): two keys on the sensitive tier, and CI refuses one (#835)…
fix(secrets): name the stack in every secrets-verify-backup hint (#882)
Gerrrt commentedon Oct 4, 2026
Done, on both counts in Done when.
The guard.
scripts/check_sops_rules.py(#866, merged in f154884) reads each committed file's ownsops:recipients. It fails if they differ from the file's.sops.yamlrule, or if a data-guarding file (observability, sensitive, tofu) opens with fewer than two keys. Its first CI run failed against the real one-recipientsecrets/sensitive.sops.yaml, as intended. Ten--self-testcases cover the verdicts and the repair advice. Review then found that the advice could recreate the drift on the shared catch-all rule, and that was fixed before merge.The re-key. Run on trinity on 2026-10-03 with
make secrets-add-recipient STACK=sensitive PUBKEY=age19mkg…..sops.yamland the re-keyed file were committed together (6490d49, merged via #866). Only the data-key stanzas changed; no encrypted value did.Proven on a machine that is not trinity, with the technical second's key alone:
make secrets-verify-backup STACK=sensitive KEY=…passed.20261004T020639Z, the first taken after converge picked up fix(secrets): two keys on the sensitive tier, and CI refuses one (#835) #866): its MANIFEST recordsrecipient age19mkg76v0…,age1cpdxl6p4…. Every*.tar.gz.agein it was decrypted withage -d -i <second key>and streamed whole throughtar -tzf, all ok. No plaintext was written to disk.What stays true. Volume sets taken before 2026-10-04T02:06Z still open with trinity's key alone. The runbooks now say so.
Found along the way:
add-recipient.sh,remove-recipient.shandbootstrap.shprinted their verify hint withoutSTACK, so a key on the catch-all rule would be proved againstobservability. Fixed in #882 (e95e475).🤖 Posted by Claude Code
test(converge, backup): fixture suites for converge, backup-volumes a…