Files
felhom.eu/documentation/audits/end-dates-list-2026-11-11/TABLE.md
T
admin 978666dde3
gates / gates (push) Successful in 2m2s
lists: the Clock list (70 jobs) and the End-dates list (39 things), 9 register rows, 4 decisions
Read-only review of every job on a timer and of everything that ends on a date,
with measurements from the three test boxes. Nothing ends in the next 14 days.
Register 236 -> 245 (R-1379..R-1387), pointer lines on twelve rows, D76-D79 in
STATUS.md. Part C: felhom-agent's clone address holds no token now; the other
clones wait for the operator's yes. R-1035's 4-week check moved to 2026-11-18
(operator read needed).

Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS
2026-11-11 11:54:42 +01:00

76 KiB

The End-dates list — the table

Built 2026-11-11 from the source of the four repositories, from read-only measurements on the three test boxes (Tester 1, demo-hp, the N100; host and guest) and from what any visitor can read with no login. One block for each thing that stops working, or stops being safe, on a date or after an age. tools/make-table.py writes this file from the helpers' answers (evidence/helpers/) and from tools/marks.py; do not edit it by hand.

NOT MEASURED, by the operator's ruling of 2026-11-10: DooPlex and ep0, and every account (Cloudflare, Hetzner, the mail service, the registrars). What this table says about them comes from the repository's own scripts and documents.

The eight questions: 1 what it is, who owns it, where it lives · 2 when it ends (a real date for each box where one was measured, else the rule) · 3 who renews it, and how long before · 4 what breaks, and for whom · 5 who is warned before the end · 6 a box that was off past the end date: does it come back by itself? · 7 what happens if the renewal fails · 8 how do we know (MEASURED, a test, SOURCE, a document, or nobody knows). A ninth line, outside the marks: is the hand step written in a runbook?

Marks: A answered, with a pointer · D decided, not built · K known gap, with its row · N nobody asked — it survived the searches in EV/searches.txt (S01…S23; controls in the same file). Every N names its class row: T1 = R-1379 an end exists and nobody is warned before it · T2 = R-1380 another party renews and our pin does not follow · T3 = R-1381 no document says when it ends, or where to read the date · T4 = R-1382 a box that was off past the date: nobody knows if it comes back.

Who read what. A cell whose mark is D, K or N, and every A that tools/marks.py carries, holds the main session's words: the pointer was opened and read. Every other cell holds a helper's words with the mark A: tools/resolve-helper.py opened each of its pointers (the file and the line exist, and the name the helper gave stands at that line). The main session read every helper answer to questions 4 to 7; it did not open each of their pointers. Treat those cells as a map, not as proof.

Pointers: AG/ = felhom-agent/ · CT/ = felhom-controller/controller/ · CAT/ = app-catalog-felhom.eu/ · HUB/ = felhom.eu/hub/internal/ · E/ = felhom.eu/ · AR/ = felhom.eu/documentation/architecture/ · RB/ = felhom.eu/documentation/runbooks/ · EV/ = this folder's evidence/. tools/check-pointers.py prints the text behind each one.

Named status (does an architecture document name this end date?): NAMED and MENTIONED with a pointer were read in the documents. Where the line says "a machine count", the status comes from a search of the architecture folder: 0 lines = UNNAMED, else MENTIONED.

The kinds, and where each thing is

Kind Things (worst mark) Named · mentioned · named nowhere
Certificate DB01 (N), DB02 (N), DB03 (N), DH20 (N) 1 · 1 · 2
Key DB06 (A), DB12 (K), DB13 (A), DB16 (N), DH08 (D), DH09 (A), DH10 (A), DH15 (A), DH22 (N), DH23 (N), DH25 (A), DH30 (N), DH31 (N), DH32 (N), DH33 (A) 5 · 10 · 0
Window DB04 (N), DB05 (K), DB08 (A), DB09 (A), DB11 (A), DB24 (N), DH01 (A), DH05 (A), DH14 (A) 1 · 8 · 0
Version DB18 (N), DB19 (N), DB20 (A), DB21 (K), DH19 (N), DH26 (A) 2 · 4 · 0
Account DH24 (N), DH27 (N), DH28 (N), DH29 (N) 0 · 4 · 0
OS DB17 (N) 0 · 1 · 0

Not in the table

Tiny (named once, not tabled): login sessions on the box and on the hub (hours to 7 days; a new login repairs it) · the one-use codes of the recovery ceremony (minutes) · the guest's token for the agent and the hub key (no date) · pinned build tools of the controller (a build fails loudly) · the hub's claim code and pairing code (no date; three resets a day) · operator actions in the queue (24 hours; a restart order 30 minutes) · the vouch alarm (2 hours) · the Debian template that the golden is built from (a bake fails loudly; the operator changes the default).

Left out for another reason:

  • The recovery passphrase — it stops on an event (a new one is made), not on a date.
  • Kept escrow packages — they have no end date by ruling; the 14-day countdown that removes one is DB04.
  • The off-site clean-up window (20 minutes, 6.5 days) — it is a timer, so it is on the Clock list.
  • The hub's view of signed jobs, signing keys, the golden's age, the agent versions, and OS support — the same things as DB11, DB12, DB21, DB20 and DB17.
  • WireGuard re-key on the box — the same thing as DH10.

DB01 — Web certificate of the household's address · Certificate · UNNAMED

architecture folder: 0 lines in 0 files match /acme\.json|Let.s Encrypt|letsencrypt/ (a machine count, not a reading) · helper section: dates-box „1. Web certificate of the household's address (Let's Encrypt through Traefik)”

# Answer Mark Pointer
1 The public TLS certificate for the dashboard and apps. Product-owned Traefik asks Let's Encrypt with the household's e-mail ACMEEmail; stored in acme.json. A CT/internal/stacks/infra.go:130 CT/internal/infra/templates/traefik.yml.tmpl:51
2 MEASURED 2026-11-11: the N100's certificate ends 2026-12-31, Tester 1's 2027-01-02, demo-hp's 2027-01-19. The life (90 days) is the issuer's rule, not ours. A EV/box/demo-felhom-guest-dates.txt EV/box/tester-1-guest-dates.txt EV/box/demo-hp-guest-dates.txt
3 Traefik renews by itself, by its own default lead time. MEASURED on demo-hp: it got a new certificate on 2026-10-21, 90 days before the end date 2027-01-19. A CT/internal/infra/templates/traefik.yml.tmpl:53 CT/internal/infra/templates/traefik.yml.tmpl:59
4 Browsers show a certificate error on every app address of the household. With no e-mail set Traefik serves a self-signed certificate, see comment at TraefikData. A CT/internal/infra/infra.go:111-118
5 Nobody. No part of the product reads the certificate's end date. If the renewal fails (the household's Cloudflare token ends, or the name check fails), the first sign is a browser warning on the end date. N CT/internal/infra/infra.go:157-159 · class T1 · S01
6 NOBODY KNOWS. A box that was off past the end date starts with a dead certificate. Traefik must get a new one at the start. Nobody has tested this. N CT/internal/infra/infra.go:157-159 · class T4 · S11
7 Traefik tries again by itself. The old certificate works until its end date. Nothing in the product reads the result, so a renewal that fails for weeks is silent. N CT/internal/infra/templates/traefik.yml.tmpl:51 · class T1 · S01
8 MEASURED: the three end dates, and one renewal by the product on demo-hp (the certificate was issued again on 2026-10-21) · SOURCE for the setup; NOBODY KNOWS for renewal timing, warning and the switched-off case. A EV/box/demo-hp-guest-dates.txt CT/internal/infra/templates/traefik.yml.tmpl:48-59

Hand step in a runbook? NO — no runbook says what to do when a household's certificate does not renew

DB02 — Proxmox's own certificate, pinned by the agent · Certificate · UNNAMED

architecture folder: 0 lines in 0 files match /pve-ssl|tls\.fingerprint|TLS pin/ (a machine count, not a reading) · helper section: dates-box „2. Proxmox web certificate pinned by the agent”

# Answer Mark Pointer
1 The agent talks to Proxmox on the host and trusts ONE certificate by fingerprint. The installer reads the served leaf once, fp=; the agent checks it in TLSConfig.build. Product-owned. A E/scripts/felhom-host-install.sh:3182-3185 AG/internal/proxmox/tls.go:39-59
2 MEASURED 2026-11-11: the certificate ends 2028-07-16 on the N100, 2028-08-19 on demo-hp, 2028-10-02 on Tester 1. On all three the agent's pin equals this certificate. The pin itself has no date. A EV/box/pin.txt AG/internal/proxmox/tls.go:40-58
3 Proxmox renews it by itself, about 14 days before the end (its daily update job). No agent code reads the new fingerprint. The installer writes the pin once. N EV/box/pveupdate-excerpt.txt AG/internal/proxmox/tls.go:40-58 E/scripts/felhom-host-install.sh:3182-3185 · class T2 · S02
4 After Proxmox makes the new certificate, every call from the agent to Proxmox fails with 'TLS pin mismatch'. That is: guest start and stop, whole-guest backups, restore tests, disk work. First box: the N100, about 2028-07-02. The other two follow within three months. N AG/internal/proxmox/tls.go:55 EV/box/pin.txt · class T2 · S02
5 Nobody before the day. No code compares the pin with the certificate's end date. N AG/internal/proxmox/tls.go:40-58 · class T2 · S02
6 No. Proxmox also renews a certificate that has ended, at its next daily job. The box comes back with a new certificate and the old pin. N EV/box/pveupdate-excerpt.txt · class T2 · S02
7 Here the renewal that works is the failure. The repair is by hand: write the new fingerprint into the agent's config and restart the agent. No code does it. N AG/internal/proxmox/tls.go:40-58 · class T2 · S02
8 MEASURED: the three end dates, pin equals certificate on all three, and Proxmox's own renewal rule read on demo-hp · TEST TestTLSConfig_Build for the pin check; NOBODY KNOWS for the renewal case. A EV/box/pin.txt EV/box/pveupdate-excerpt.txt AG/internal/proxmox/tls_test.go:28

Hand step in a runbook? NO — a search of the runbooks for 'tls.fingerprint', 're-pin' and 'pin mismatch' finds nothing

DB03 — The agent's local-API certificate, pinned in the guest · Certificate · NAMED

architecture: AR/03-host-agent.md:420-423 · helper section: dates-box „3. Agent local-API certificate and its pin in the guest”

# Answer Mark Pointer
1 The self-signed leaf the agent serves to the in-guest controller, made by generateLeaf. The controller pins its fingerprint in New. Product-owned, on the host. A AG/internal/localapi/cert.go:56 CT/internal/agentapi/client.go:105
2 certValidity = 10 * 365 * 24 * time.Hour, set as NotAfter. So a box installed in 2026 has a leaf that ends in 2036. A AG/internal/localapi/cert.go:25 AG/internal/localapi/cert.go:70
3 Nobody: EnsureLeaf loads the existing pair and never looks at its date. The comment says rotation is an operator action for a later slice. A AG/internal/localapi/cert.go:38-52 AG/internal/localapi/cert.go:21-24
4 At the end date nothing breaks: the controller compares the fingerprint only, not the date. A NEW certificate is what breaks the link. A CT/internal/agentapi/client.go:118-128 CT/internal/agentapi/client.go:125
5 Nobody. The certificate lives 10 years and nothing reads its end date. The first one ends about 2036. N AG/internal/localapi/cert.go:25 · class T1 · S01
6 Yes, by itself: the leaf is generated once and kept, and the reinstall path keeps it with --preserve-state-from. A AR/03-host-agent.md:420-423 E/scripts/felhom-host-install.sh:2749-2786
7 No renewal exists, so no failure state. A reinstall that makes a new leaf needs --allow-new-leaf and then the guest's pin is stale. A E/scripts/felhom-host-install.sh:2749-2786
8 TEST TestEnsureLeaf_StableFingerprintAcrossReload and TestClient_WrongPinFailsClosed; the 10-year end itself: SOURCE only. A AG/internal/localapi/cert_test.go:12 CT/internal/agentapi/client_test.go:55

Hand step in a runbook? no hand step

DB04 — Off-site copy after the household stops it: the 14-day countdown · Window · MENTIONED

architecture folder: 21 lines in 9 files match /14.day countdown|abandon/ (a machine count, not a reading) · helper section: dates-box „4. Off-site copy after the household stops it: 14-day countdown”

# Answer Mark Pointer
1 When the household abandons the off-site backup, the copies are kept for a grace period and then deleted. startAbandonCountdown. The household starts it; the product runs it. A CT/internal/backup/offbox_abandon.go:68-70
2 abandonGraceDays = 14. A CT/internal/backup/offbox_abandon.go:31
3 The household can stop or extend it; extending never shortens, test TestD1_ExtendNeverShortensTheCountdown. A CT/internal/backup/offbox_abandon_levers_d1_test.go:20
4 The off-site copies are gone for good; the household loses its only copy outside the house. The deleting side is the hub (see the hub helper); test name TestD1_ExtendRefusedWhileTheDeletionIsTheHubs. A CT/internal/backup/offbox_abandon_levers_d1_test.go:44
5 The household, at 5, 3 and 1 days left: AbandonRemindAtDays. A CT/internal/backup/offbox_abandon.go:40
6 NOBODY KNOWS for a box that is off during the whole countdown: the reminders at 5, 3 and 1 days come from the box. Nobody traced if the old store is removed with no reminder shown. N CT/internal/backup/offbox_abandon.go:31-40 · class T4 · S23
7 A failed hub cancel is reported, not hidden: TestD1_StopReportsFailureWhenTheHubCancelFails. A CT/internal/backup/offbox_abandon_levers_d1_test.go:69
8 TEST TestR241_ScenarioE_AbandonStartsAReversibleCountdown. A CT/internal/backup/offbox_abandon_r241_test.go:48

Hand step in a runbook? no hand step

DB05 — Off-site history: how far back a household can go · Window · MENTIONED

architecture folder: 13 lines in 5 files match /keepDaily|keep-daily|7 daily/ (a machine count, not a reading) · helper section: dates-box „5. Off-site backup history: how far back a household can go”

# Answer Mark Pointer
1 The restic off-site store keeps a thinning history. Policy in retentionPolicy. Product rule, household's data. A CT/internal/backup/offbox_window.go:81-82
2 keepDaily = 7, keepWeekly = 4, keepMonthly = 6. Older than about six months: gone. A CT/internal/backup/offbox_window.go:47 CT/internal/backup/offbox_window.go:48 CT/internal/backup/offbox_window.go:49
3 Nobody renews; each run forgets by the same rule, retentionPolicy. A CT/internal/backup/offbox_window.go:81-82
4 A file deleted or damaged more than six monthly points ago cannot be brought back from off-site. On-host copies are fewer: KeepLast, default 3 per comment. A AG/internal/config/config.go:436 AG/internal/config/config.go:409
5 Nobody, and by ruling: the window is the rule. A point leaves it with no warning. A CT/internal/backup/offbox_window.go:81-82 (the code path of question 1)
6 Yes. A box that was off simply has no points for that time; the next run applies the same rule. PBS copies are not pruned by the box by default, comment. A AG/internal/backup/runner.go:51-58
7 A clean-up that the guard refuses leaves more history than the rule says, not less. K R-1294
8 SOURCE. A CT/internal/backup/offbox_window.go:47-49

Hand step in a runbook? no hand step

DB06 — Off-site credentials: SSH key, host key, store password, PBS pin · Key · MENTIONED

architecture folder: 37 lines in 9 files match /host key|known_hosts|sub-account/ (a machine count, not a reading) · helper section: dates-box „6. Off-site credentials: SSH key, host key, store password, PBS pin”

# Answer Mark Pointer
1 The box reaches its off-site store with an SSH key and a pinned host key, StrictHostKeyChecking=yes, and a password file RESTIC_PASSWORD_FILE. PBS is pinned in pinnedTLS. A CT/internal/backup/offbox.go:345 CT/internal/backup/offbox.go:898 AG/internal/pbs/pin.go:17-30
2 NO END DATE in box code: no expiry field on the key, the host key or the password (grep expir in found none). The PBS pin ignores the date, VerifyPeerCertificate. The end comes from outside: the provider changes its host key or the hub ends the account. A CT/internal/backup/offbox.go AG/internal/pbs/pin.go:24
3 Nobody on a schedule. A new PBS descriptor is checked before use: probeFP. A AG/internal/pbsdr/manager.go:375-379
4 If the provider's host key changes, every off-site run fails at the SSH step, StrictHostKeyChecking=yes. The household has no off-site copy from that night on. A CT/internal/backup/offbox.go:895
5 Nobody, because there is no date. A failed night run is the sign. A CT/internal/backup/offbox.go:345 CT/internal/backup/offbox.go:898 (the code path of question 1)
6 Yes if nothing changed outside. A guest that was built again gets a new store password and cannot open the old store; the 14-day countdown (DB04) is the way out. A CT/internal/backup/offbox.go:910
7 PBS: state verify_failed and a retry at the next tick, probeFP. Restic: the run fails and the next night tries again with the same pins. A AG/internal/pbsdr/manager.go:375-379
8 SOURCE; provider-side dates: NOBODY KNOWS from the box. A CT/internal/backup/offbox.go:345

Hand step in a runbook? yes — RB/offsite-endpoint.md

DB08 — Claim code (72 hours) and the one-use gate token · Window · MENTIONED

architecture folder: 12 lines in 8 files match /claim code|claimCodeTTL|72 h/ (a machine count, not a reading) · helper section: dates-box „8. Claim code and the one-use gate token”

# Answer Mark Pointer
1 The one-time code a household uses to set its own password, and the short token that carries a pass between addresses. Product-made. claimCodeTTL; gateTokenLife. A CT/internal/web/claim.go:30 CT/internal/web/setup_gate.go:47
2 claimCodeTTL = 72 * time.Hour; gateTokenLife = 60 * time.Second, also stated in. A CT/internal/web/claim.go:30 CT/internal/web/setup_gate.go:47 AR/01-topology-and-trust.md:192
3 The operator sends a new code (hub side, see the hub helper). Nothing on the box mints one by itself. A CT/internal/web/claim.go:30 CT/internal/web/setup_gate.go:47 (the code path of question 1)
4 A household that opens the mail after three days cannot set its password until a new code arrives. Checked against the controller's clock s.claimNow. A CT/internal/web/claim.go:333
5 Nobody before the end. The household sees the dead code at the form, and the operator sends a new one. A CT/internal/web/claim.go:30 CT/internal/web/setup_gate.go:47 (the code path of question 1)
6 A box off for more than 72 hours comes back with a dead code; first failing step is the claim form,. A CT/internal/web/claim.go:333
7 Wrong tries lock the form for a time: claimLockoutWindow. A CT/internal/web/claim.go:33
8 TEST TestClaimSubmit_ExpiredCodeRefused. A CT/internal/web/claim_test.go:189

Hand step in a runbook? no hand step

architecture folder: 5 lines in 4 files match //s/<token>|share token|LauncherShareToken/ (a machine count, not a reading) · helper section: dates-box „9. Share link /s/<token>”

# Answer Mark Pointer
1 The link a household gives to guests to see its app tiles. Household-owned, stored as LauncherShareToken. A CT/internal/settings/settings.go:45
2 NO END DATE: the token has no expiry field; it ends only when the household rotates or disables it, SetLauncherShareToken. A CT/internal/settings/settings.go:1111
3 The household, by hand. No schedule. A CT/internal/settings/settings.go:45 (the code path of question 1)
4 Nothing ends. A link that was given away years ago still opens the page, until the household changes it. A CT/internal/web/share.go:31
5 Nobody, because there is no date. The link ends only when the household changes it. A CT/internal/settings/settings.go:45 (the code path of question 1)
6 Yes, unchanged. A CT/internal/settings/settings.go:45 (the code path of question 1)
7 No renewal to fail. A CT/internal/settings/settings.go:45 (the code path of question 1)
8 TEST TestShareGuest_RotateAndDisable. A CT/internal/web/share_test.go:256

Hand step in a runbook? no hand step

DB11 — Signed operator jobs: the validity window · Window · NAMED

architecture: AR/04-control-plane-authorization.md:139-150 · helper section: dates-box „11. Signed operator jobs: the validity window”

# Answer Mark Pointer
1 An order signed by the operator (update the agent, apply OS packages, destructive acts). Operator-owned; it carries a nonce and an expiry,. Checked by Verify and by the root wrapper. A AR/03-host-agent.md:268 AG/internal/authz/verifier.go:104
2 At most MaxWindow = 24 * time.Hour; the signer's default is 30 minutes, -ttl. Wrapper: MAX_WINDOW_S = 24 * 3600 AG/configs/felhom-os-apply:194. Design:. A AG/internal/authz/verifier.go:26 AG/cmd/felhom-opsign/main.go:57 AR/04-control-plane-authorization.md:78-84
3 The operator signs a new one. There is no renewal of a signed job. A AR/03-host-agent.md:268 AG/internal/authz/verifier.go:104 (the code path of question 1)
4 The job is refused and cleared; the executor is never called,. The box stays on its old agent or old packages. ErrExpired. A AG/internal/signedjobs/runner.go:163-170 AG/internal/authz/verifier.go:161-162
5 Nobody before. A job that ends with no result is what the operator sees. A new job is signed. A AR/04-control-plane-authorization.md:184
6 A box off for longer than the window never runs that job; it is checked against the BOX's clock, time.Now.UTC, and in the wrapper if now > exp or now < iss - 120 AG/configs/felhom-os-apply:676. A wrong box clock decides. A AG/internal/authz/verifier.go:90
7 The refused job is logged as a security event and removed, no retry. Used nonces are kept until their end, comment. A AG/internal/signedjobs/runner.go:163-170 AG/internal/authz/noncestore.go:47-50
8 TEST test_expired_signed_job_is_refused, test_a_window_longer_than_24h_is_refused, TestVerify_RejectsPerStage. A AG/configs/test_felhom_os_apply.py:833 AG/configs/test_felhom_os_apply.py:836 AG/internal/authz/verify_test.go:82

Hand step in a runbook? yes — the build skill's agent section

DB12 — Operator signing keys on each box · Key · NAMED

architecture: AR/04-control-plane-authorization.md:139-150 · helper section: dates-box „12. Operator signing keys on each box”

# Answer Mark Pointer
1 The two public keys every box trusts for signed jobs: OPERATOR_KEY_OPERATIONAL_ID and OPERATOR_KEY_RECOVERY_ID. Written to TRUST_SIGNERS AG/configs/felhom-os-apply:124 by _write_slow_lane_trust. Operator-owned. A E/scripts/felhom-host-install.sh:212 E/scripts/felhom-host-install.sh:214 E/scripts/felhom-host-install.sh:3424
2 NO END DATE: the keys carry no expiry; trust is by key material,. The key id is advisory,. A AG/internal/authz/verifier.go:122-126 AG/internal/authz/verifier.go:170-175
3 The keys have no end date. A change of key is designed as a signed order of its own. Only the name of that order exists in code. D AR/04-control-plane-authorization.md:139-150
4 If the operator loses both private keys, no box can be updated or repaired by a signed job again; households keep running on what they have. Design text. A AR/04-control-plane-authorization.md:139-150
5 Nobody; there is no date to warn about. A E/scripts/felhom-host-install.sh:212 E/scripts/felhom-host-install.sh:214 (the code path of question 1)
6 Yes, the keys are in the installer and on disk,. A E/scripts/felhom-host-install.sh:3452-3453
7 If the keys are lost, no box accepts a signed job again without a visit. The keys exist on DooPlex only. K R-924
8 DOC; for a real rotation: NOBODY KNOWS. A AR/04-control-plane-authorization.md:139-150

Hand step in a runbook? no — the design exists, a procedure does not

DB13 — The agent's Proxmox access token · Key · MENTIONED

architecture folder: 33 lines in 12 files match /pveum|API token|privsep/ (a machine count, not a reading) · helper section: dates-box „13. Proxmox API token of the agent”

# Answer Mark Pointer
1 The token the agent uses for every Proxmox call. Made by the installer in step_token. Product-owned, on the host. A E/scripts/felhom-host-install.sh:2405
2 MEASURED 2026-11-11 on the three boxes: the user and its token have no end date (expire 0). A EV/box/pin.txt E/scripts/felhom-host-install.sh:2448
3 Nobody on a schedule. A re-run of the installer reuses it or makes a new one,. A E/scripts/felhom-host-install.sh:2424-2445
4 Nothing ends by date. If someone sets an expiry by hand in Proxmox, every agent call fails that day. A E/scripts/felhom-host-install.sh:2405 (the code path of question 1)
5 Nobody, because there is no date. A EV/box/pin.txt
6 Yes. A E/scripts/felhom-host-install.sh:2405 (the code path of question 1)
7 No renewal to fail. A E/scripts/felhom-host-install.sh:2405 (the code path of question 1)
8 MEASURED: expire 0 on three boxes · SOURCE. A EV/box/pin.txt E/scripts/felhom-host-install.sh:2448

Hand step in a runbook? no hand step

DB16 — Package signing keys: Docker in the guest, Proxmox and Debian on the host · Key · MENTIONED

architecture: AR/11-os-updates.md:657 · helper section: dates-box „16. Package signing keys: Docker in the guest, Proxmox and Debian on the host”

# Answer Mark Pointer
1 The keys apt uses to trust packages. Docker's key is fetched once at golden bake into /etc/apt/keyrings/docker.asc. Proxmox's keyring path default is in align_apt_repos. Owned by Docker, Proxmox, Debian. A AG/configs/build-golden.sh:129 E/scripts/felhom-host-install.sh:2616
2 MEASURED 2026-11-11 on the three boxes: Docker's key never ends. The Debian keys in use end 2029-01-15 at the soonest. The Proxmox key for this series ends 2034-11-10. Two keys that end 2027-02-03 and 2027-04-12 are in the 'removed keys' file; nothing uses them. A EV/box/keyrings.txt
3 Debian and Proxmox keys arrive as packages through the normal update. Docker's key has no package: a change is "a slow-lane act for a person",. A AR/11-os-updates.md:657
4 The OS update is refused with code R7 when apt-get update fails, AG/configs/felhom-os-apply:1125-1127. The box keeps running but stops getting Docker security updates. Household unaware; operator sees the report. A AG/configs/build-golden.sh:129 E/scripts/felhom-host-install.sh:2616 (the code path of question 1)
5 Nobody before the end. After it the box reports a failed update. The first date is 2029-01-15. N AR/11-os-updates.md:657 EV/box/keyrings.txt · class T1 · S10
6 NOBODY KNOWS. A box that was off for years has old keys. The new keys come as a package, and the package list is checked with the old keys. N AR/11-os-updates.md:657 · class T4 · S11
7 The update fails cleanly and nothing is half-installed, the next signed job tries again and fails the same way until a person replaces the key. A AR/11-os-updates.md:657
8 MEASURED: every key's end date on three hosts and three guests · DOC; TEST for the refusal test_R7_not_downloadable_and_no_snapshot; key dates: NOBODY KNOWS. A EV/box/keyrings.txt AR/11-os-updates.md:657 AG/configs/test_felhom_os_apply.py:488

Hand step in a runbook? NO — the document says a key change is 'a slow-lane act for a person'; no procedure is written

DB17 — Debian 13 and Proxmox 9: the series the product is tied to · OS · MENTIONED

architecture folder: 9 lines in 2 files match /trixie|Debian 13|Proxmox (VE )?9/ (a machine count, not a reading) · helper section: dates-box „17. Debian 13 (trixie) and Proxmox 9: the series the product is tied to”

# Answer Mark Pointer
1 The OS series under everything. Written into source: suite default trixie, Docker repo trixie stable, base images CT/Dockerfile:52. Owned by Debian and Proxmox. A E/scripts/felhom-host-install.sh:2615 AG/configs/build-golden.sh:130
2 No document gives the day when security fixes stop for Debian 13 or for Proxmox 9. MEASURED: the three boxes run Debian 13.7 and Proxmox 9.2. N EV/box/demo-hp-host-dates.txt AR/11-os-updates.md:1 · class T3 · S09
3 No plan with a date. The move to the next series is not designed. N AR/11-os-updates.md:1 · class T3 · S09
4 After the series ends: no security updates for host and guest; a new install on a newer Proxmox is refused in byo mode and only warned in appliance mode,. A E/scripts/felhom-host-install.sh:2146-2152
5 Nobody. The dates are public; nothing in the product or in a document reads them. N AR/11-os-updates.md:1 · class T1 · S09
6 Yes while the archives still serve trixie. Pinned OS updates fall back to snapshot.debian.org in add_snapshot_sources AG/configs/felhom-os-apply:1451-1462; if that source stops, the update is refused. A E/scripts/felhom-host-install.sh:2615 AG/configs/build-golden.sh:130 (the code path of question 1)
7 Refused update, box unchanged, AG/configs/felhom-os-apply:1125-1127. A E/scripts/felhom-host-install.sh:2615 AG/configs/build-golden.sh:130 (the code path of question 1)
8 SOURCE; the dates: NOBODY KNOWS. A E/scripts/felhom-host-install.sh:2146-2152

Hand step in a runbook? NO

DB18 — Catalog app images and their recorded digests · Version · MENTIONED

architecture: AR/09-update-architecture.md:391 · helper section: dates-box „18. Catalog app images and their recorded digests”

# Answer Mark Pointer
1 65 app templates under CAT/templates/; 84 different image references; registries docker.io 47, ghcr.io 27, lscr.io 8, gitea.dooplex.hu 1, registry.gitlab.com 1. The update ladders record 149 different digests, e.g.. Owned by each app's publisher. A CAT/templates/actualbudget/.felhom.yml:196
2 NO END DATE in source. The end is the day a registry stops serving a tag or a digest. Moving tags are banned: BANNED_TAGS. A CAT/scripts/check-image-pins.py:28
3 The product, when it tests a new version and records its digest. reports how far behind a pin is and gates nothing. A CAT/scripts/check-currency.py:22
4 The box renders tag@sha256 in RenderWithLadderDigests; a digest the registry no longer serves makes the pull fail on every box,. A household cannot install or update that app; a running app keeps running. A CT/internal/stacks/digest.go:94 AR/09-update-architecture.md:391
5 Nobody. No check asks the registries if every recorded image is still served. The document speaks of the day a version is recorded, not of the days after. N AR/09-update-architecture.md:391 CAT/scripts/check-currency.py:22 · class T1 · S21
6 Running apps come back (images are local). The first failing step is the next install, update or restore that must pull,. A CT/internal/stacks/digest.go:10-23
7 The pull fails and the app stays on its current image; no automatic repair. Retry happens only when the household presses again. A CAT/templates/actualbudget/.felhom.yml:196 (the code path of question 1)
8 DOC; SOURCE; which digests are still served today: NOBODY KNOWS. A AR/09-update-architecture.md:391 CT/internal/stacks/digest.go:94

Hand step in a runbook? no hand step

DB19 — Images the controller itself starts (Traefik, tunnel client, file browser, Samba) · Version · MENTIONED

architecture folder: 41 lines in 15 files match /cloudflared|TraefikImage|traefik:v/ (a machine count, not a reading) · helper section: dates-box „19. Images the controller itself starts (Traefik, tunnel, file browser, Samba)”

# Answer Mark Pointer
1 Four fixed images in the controller binary: TraefikImage, CloudflaredImage, FileBrowserImage, SambaImage. Product-owned pins of other people's images. A CT/internal/infra/infra.go:33 CT/internal/infra/infra.go:34 CT/internal/infra/infra.go:35 CT/internal/infra/infra.go:50
2 No document says for how long Cloudflare accepts an old tunnel client. MEASURED on the N100: tunnel client 2026.10.3, Traefik v3.7.16. N EV/box/demo-felhom-guest-dates.txt · class T3 · S12
3 The product, by a controller release that names newer tags. A CT/internal/infra/infra.go:33 CT/internal/infra/infra.go:34 (the code path of question 1)
4 If Cloudflare stops accepting an old tunnel client, the household's address stops answering from outside while the box is healthy. If a registry drops a tag, a box that is built again cannot start that part. A CT/internal/infra/infra.go:33 CT/internal/infra/infra.go:34 (the code path of question 1)
5 Nobody. N EV/box/demo-felhom-guest-dates.txt · class T1 · S12
6 NOBODY KNOWS if the tunnel of a box that was off for a year still connects with its old client. Without the tunnel the box gets no update from outside. N EV/box/demo-felhom-guest-dates.txt · class T4 · S12
7 Not traced. A CT/internal/infra/infra.go:33 CT/internal/infra/infra.go:34 (the code path of question 1)
8 SOURCE; NOBODY KNOWS for the tunnel client's age limit. A CT/internal/infra/infra.go:33-50

Hand step in a runbook? no hand step

DB20 — Published agent versions: the newest five stay · Version · NAMED

architecture: AR/09-update-architecture.md §3, D50 · helper section: dates-box „20. Published agent versions: only the newest five stay downloadable”

# Answer Mark Pointer
1 The registry of agent binaries. Rule file generic_versions_kept. Operator-owned; he prunes by hand,. A AG/scripts/retention-policy.json:23 AG/scripts/retention-policy.json:7-10
2 Newest 5 versions plus two pins: hub_min_agent and day0_agent. Decision D50,. A AG/scripts/retention-policy.json:30 AG/scripts/retention-policy.json:31 AR/09-update-architecture.md:1370-1375
3 The session that records a vouch or a new floor updates the pins in the same commit,. A AG/scripts/retention-policy.json:12-15
4 A box told to install a version that was pruned cannot download it. A pin that names a missing version is a CI failure,. Deleting a git tag also ends the ability to install that version,. A AG/scripts/retention-policy.json:14-15 AG/scripts/retention-policy.json:20-21
5 The operator, through CI:. A AG/scripts/retention-policy.json:25
6 A box off for long is on an old agent; it updates only by a fresh signed job that names a version still kept. First failing step otherwise: the download in the wrapper, SELFUPDATE_OP AG/configs/felhom-os-apply:180. A AG/scripts/retention-policy.json:23 AG/scripts/retention-policy.json:7-10 (the code path of question 1)
7 Not traced on a failed download. A AG/scripts/retention-policy.json:23 AG/scripts/retention-policy.json:7-10 (the code path of question 1)
8 SOURCE; DOC. A AG/scripts/retention-policy.json:23-31 AR/09-update-architecture.md:1370-1375

Hand step in a runbook? yes — the build skill's agent section

DB21 — The golden image's age (7 days) and the waiver (14 days) · Version · NAMED

architecture: AR/09-update-architecture.md:1010-1018 · helper section: dates-box „21. Golden image age and the waiver file”

# Answer Mark Pointer
1 The ready-made guest image a new box is installed from, baked weekly,. Operator vouches. Gate:. A AR/09-update-architecture.md:1010-1018 E/scripts/golden_currency_gate.py:169
2 MEASURED 2026-11-11: the newest golden is 8 days old (baked 2026-11-03). The limit is 7. No waiver file exists. The gate fails today. A E/scripts/golden_currency_gate.py:169 EV/box/golden-gate.txt
3 The operator, weekly (Sunday night per). A AR/09-update-architecture.md:1010-1018
4 The gate goes red for the product team; no household is touched. A new install gets an older image and must update more on first start. A AR/09-update-architecture.md:1010-1018 E/scripts/golden_currency_gate.py:169 (the code path of question 1)
5 The operator, by the gate itself; no earlier warning found. A AR/09-update-architecture.md:1010-1018 E/scripts/golden_currency_gate.py:169 (the code path of question 1)
6 Not about a box. A AR/09-update-architecture.md:1010-1018 E/scripts/golden_currency_gate.py:169 (the code path of question 1)
7 A push of code to this repository is refused until a golden is baked or a dated waiver is written. A push of documents passes with a notice. K R-468
8 MEASURED: the gate's own output today · SOURCE; tests exist in (names not matched to this rule). A EV/box/golden-gate.txt E/scripts/golden_currency_gate.py:162-169 E/scripts/test_golden_currency_gate.py:1

Hand step in a runbook? yes — RB/RUNBOOK-manual-build.md

DB24 — Copies kept before a restore: deleted after 30 unused days · Window · MENTIONED

architecture folder: 18 lines in 4 files match /restore cop|restore-undo|D62/ (a machine count, not a reading) · helper section: dates-box „24. Kept copies made before a restore: deleted after 30 unused days”

# Answer Mark Pointer
1 The copy of an app's data the box keeps when the household restores an older state. Household's data, product rule D62. restoreCopyUnusedAge. A CT/internal/backup/restore_copy_sweep.go:29
2 restoreCopyUnusedAge = 30 * 24 * time.Hour. A CT/internal/backup/restore_copy_sweep.go:29
3 Nobody; using the copy is what keeps it. A CT/internal/backup/restore_copy_sweep.go:29 (the code path of question 1)
4 The household loses the way back to the state from before the restore. A CT/internal/backup/restore_copy_sweep.go:29 (the code path of question 1)
5 Nobody. The household is not told before the copy goes. The household can delete it sooner by hand. N CT/internal/backup/restore_copy_sweep.go:18-29 · class T1 · S05
6 A box that was off for more than 30 days deletes the copy at its first sweep after the start: the age is counted from the folder's time stamp by the wall clock. Nobody asked if the off time must count. N CT/internal/backup/restore_copy_sweep.go:21-29 · class T4 · S11
7 Not traced. A CT/internal/backup/restore_copy_sweep.go:29 (the code path of question 1)
8 SOURCE. · Mail relay's in-memory self-signed leaf, 10 years, remade at each start:. · Language cookie, 365 days:. · HSTS max-age one day: HSTSMaxAgeSeconds. · A backup counts as fresh enough for an update for 24 hours: DefaultUpdateBackupMaxAge. · Night-backup banner turns stale after 26 hours: StaleAfter. … A CT/internal/backup/restore_copy_sweep.go:29 CT/internal/mailrelay/server.go:386 CT/internal/web/i18n_web.go:478 CT/internal/infra/infra.go:108

Hand step in a runbook? no hand step

architecture folder: 32 lines in 7 files match /self-bind|selfbind|/bind// (a machine count, not a reading) · helper section: dates-hub „1. Self-bind (connect) link in the household's e-mail”

# Answer Mark Pointer
1 A one-use link that lets the household tie a new box to its account. Household uses it, product mints it. Stored hashed in the hub table selfbind_tokens; minted by selfBindTTL code. A HUB/store/store.go:796-806 HUB/web/selfbind_mint.go:20
2 7 days after minting: selfBindTTL, passed to the store at; Expired. Also locks at the 5th wrong answer, SelfBindMaxAttempts. A HUB/web/selfbind_mint.go:20 HUB/web/selfbind_mint.go:77 HUB/store/selfbind.go:39 HUB/store/selfbind.go:21
3 The household itself: the expired page has a button, handleBindResend, at most one new link an hour, bindResendEvery. No renewal before the end. The hub also mints at customer creation and RESET. A HUB/web/selfbind.go:384 HUB/web/selfbind.go:382 HUB/web/selfbind_mint.go:97
4 The household cannot bind the box from the mail; unknown and expired links show the same "expired" page. The box itself keeps waiting; nothing on it is lost. A HUB/web/selfbind.go:206
5 Nobody before, by design: the page of a dead link has a button that sends a new one. A HUB/web/selfbind_mint.go:1-110 HUB/web/selfbind_mint.go:93
6 A box first switched on after day 7: the link fails at the first step (opening it); the household presses the resend button. If an old host is still linked no link is sent and the operator gets eventBindOldHostLinked, logged at. A HUB/web/selfbind.go:440 HUB/web/selfbind.go:413
7 If the mail send fails at mint the token is deleted, so no dead link is left. No automatic retry; the household presses again after an hour. A HUB/web/selfbind_mint.go:82
8 TEST TestR719_AnExpiredLinkOffersAndSendsAFreshOne; SOURCE; DOC. A HUB/web/selfbind_resend_test.go:21 HUB/web/selfbind_mint.go:20 AR/05-hub-architecture.md:359

Hand step in a runbook? no hand step

DH05 — 'Switched off on purpose' hold (14 days) · Window · MENTIONED

architecture folder: 10 lines in 4 files match /deliberate.off|switched off on purpose/ (a machine count, not a reading) · helper section: dates-hub „5. "Switched off on purpose" hold (14 days of silence allowed)”

# Answer Mark Pointer
1 When a household switches the box off through the product, the hub stays quiet about the silence. Household act, hub rule: DeliberateOffHold. A HUB/store/deliberateoff.go:18
2 14 days after the switch-off: DeliberateOffHold, judged by DeliberateOffHeld. A HUB/store/deliberateoff.go:18 HUB/store/deliberateoff.go:21
3 Nobody renews. The hold ends by itself when the box reports again or at day 14. A HUB/store/deliberateoff.go:18 (the code path of question 1)
4 After day 14 the silence and missed-backup alarms start again for that household; the operator is told. A HUB/store/deliberateoff.go:18 (the code path of question 1)
5 The operator gets an event at switch-off time naming the moment, message built at. No warning before day 14 was found (grep DeliberateOffHold in hub). A HUB/api/deliberate_off.go:42
6 Yes, it comes back by itself: a box switched on at any time reports and the hold clears. TEST TestDeliberateOff_BackOnIsQuiet. A HUB/monitor/deliberate_off_test.go:138
7 Not applicable: nothing is renewed. A HUB/store/deliberateoff.go:18 (the code path of question 1)
8 TEST TestDeliberateOff_AlarmsAfterFourteenDays and TestDeliberateOff_NoSilenceAlarmDuringTheHold. A HUB/monitor/deliberate_off_test.go:112 HUB/monitor/deliberate_off_test.go:90

Hand step in a runbook? no hand step

DH08 — Storage Box sub-account password, and the hub's seal key · Key · MENTIONED

architecture folder: 27 lines in 8 files match /seal key|sub-account/ (a machine count, not a reading) · helper section: dates-hub „8. Storage Box sub-account password and the hub's seal key”

# Answer Mark Pointer
1 Each household's off-site store account at the provider, minted by the hub, and the hub key that seals stored secrets (OFFSITE_SECRET_KEY, named in). Operator and product. A E/manifests/hub.yaml:167-171
2 NO END DATE in code: no lifetime on the sub-account password. A clean reinstall's re-issue resets it, DOC; ResetBoxPassword. A AR/05-hub-architecture.md:274-279 HUB/hetznerapi/hetznerapi.go:297
3 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). The password changes when the operator presses Re-issue. A change of the seal key is 'not built'. D RB/secrets.md:159
4 If the seal key is lost or changed, the old sealed rows cannot be opened, DOC: operator loses stored off-site secrets and break-glass passwords until re-issued per household. A RB/secrets.md:159
5 Nobody, because there is no date. A HUB/hetznerapi/hetznerapi.go:1-320
6 A box off for a long time comes back by itself as long as its account was not re-issued; it uses its SSH key, the password is used only by the hub, DOC. A AR/05-hub-architecture.md:274-279
7 No renewal on a date exists, so none can fail on a date. A E/manifests/hub.yaml:167-171 (the code path of question 1)
8 DOC and; SOURCE. No test of a lifetime, because there is none. A RB/secrets.md:159 AR/05-hub-architecture.md:274-279 HUB/hetznerapi/hetznerapi.go:297

Hand step in a runbook? partly — RB/secrets.md:159 says the seal key change is not built

DH09 — PBS off-site token on ep0 · Key · NAMED

architecture: AR/05-hub-architecture.md:379-387 · helper section: dates-hub „9. PBS off-site token on ep0 and its descriptor on the hub”

# Answer Mark Pointer
1 The per-household token on the rented endpoint for the whole-guest off-site copy, and the hub record that must agree with it. Product mints, operator can re-issue. DOC. A AR/05-hub-architecture.md:373-387
2 NO END DATE stated: the lifecycle table has only events (provision, re-key, deprovision). One age rule: a staged secret not consumed in 15 minutes is wrong, unconsumedGrace. A AR/05-hub-architecture.md:379-387 HUB/pbsdrheal/reconciler.go:63
3 The hub re-keys on re-issue (delete and recreate). No scheduled renewal. A AR/05-hub-architecture.md:382
4 A re-key the box did not pick up leaves the box "serving a revoked credential": the household has no whole-guest off-site copy while both tiers look green, comment at. A HUB/pbsdrheal/reconciler.go:55-63
5 Operator only, after the fact: pbsdr_provision_failed on a final failure, DOC. A AR/05-hub-architecture.md:400-402
6 Host delete keeps the token; a box off and back uses the same token. After a re-issue while it was off, it must consume the new secret on its next report. A AR/05-hub-architecture.md:384
7 A failed self-check rolls back and is repeated 3 times, 7 s apart, then raises the operator event, DOC. A AR/05-hub-architecture.md:400-402
8 DOC; SOURCE. Lifetime of the token on the endpoint itself: NOBODY KNOWS from these sources (ep0 not read). A AR/05-hub-architecture.md:373-403 HUB/pbsdrheal/reconciler.go:63

Hand step in a runbook? yes — RB/offsite-endpoint.md

DH10 — WireGuard keys of the box's tunnel to the off-site endpoint · Key · NAMED

architecture: AR/06-offsite-connectivity.md:200-205 · helper section: dates-hub „10. WireGuard peers (the box's tunnel to the off-site endpoint)”

# Answer Mark Pointer
1 One peer per box on the endpoint, registered through the hub; the private key is in the household's sealed bundle. Product. DOC. A AR/06-offsite-connectivity.md:172-182
2 NO END DATE: the hub store only removes a peer on request, RemoveWGPeer. Scheduled key rotation is an open point, DOC. A HUB/store/wg.go:345 AR/06-offsite-connectivity.md:368
3 Nobody on a schedule. The agent makes a new key itself only after a handshake stamped in the future for 15 minutes, DOC. A AR/06-offsite-connectivity.md:205-207
4 Revocation: the box is off the off-site network at once, DOC; the household's off-site copies stop. A AR/06-offsite-connectivity.md:180-182
5 Nobody, because there is no date. A HUB/store/wg.go:1-360
6 Yes, a box off for months keeps its peer. A revoked box never re-registers on its own, DOC. A AR/06-offsite-connectivity.md:200-205
7 A failed fetch is never read as a teardown, DOC. Re-add is an operator call. A AR/06-offsite-connectivity.md:207-208
8 DOC; SOURCE. Box-side tests belong to the box helper. A AR/06-offsite-connectivity.md:200-208 HUB/store/wg.go:345

Hand step in a runbook? yes — RB/offsite-endpoint.md

DH14 — Recovery mode for a host (30 minutes to 4 hours) · Window · MENTIONED

architecture folder: 6 lines in 4 files match /recovery mode|defaultRecoveryTTL/ (a machine count, not a reading) · helper section: dates-hub „14. Recovery mode for a host (the door a rebuilt box uses)”

# Answer Mark Pointer
1 A switch the operator sets on a host so a rebuilt box may re-enrol and fetch its restore data. Operator, global key only, handleSetRecoveryMode. A HUB/api/dr.go:34
2 30 minutes by default, defaultRecoveryTTL; at most 4 hours, maxRecoveryTTL; clamp at. A HUB/api/dr.go:21 HUB/api/dr.go:22 HUB/api/dr.go:52
3 The operator sets it again. No lead time. A HUB/api/dr.go:34 (the code path of question 1)
4 A restore that takes longer than the window: the box's re-enrol call is refused and the household's recovery stops until the operator re-opens it. A HUB/api/dr.go:34 (the code path of question 1)
5 Nobody before. The operator who opened the window is present; he opens it again. A HUB/api/dr.go:34 (the code path of question 1)
6 No: a box that starts its recovery after the window fails at the re-enrol step. TEST TestReEnroll_GatedOnRecoveryMode. A HUB/api/dr_test.go:14
7 Nothing is left half-done by the expiry itself; the operator opens a new window. A HUB/api/dr.go:34 (the code path of question 1)
8 TEST TestRestoreDirective_GatedAndExpires; SOURCE. A HUB/api/dr_test.go:70 HUB/api/dr.go:21-22

Hand step in a runbook? yes — the recovery runbooks

DH15 — Break-glass password of a host · Key · MENTIONED

architecture folder: 42 lines in 13 files match /break-glass|host_recovery/ (a machine count, not a reading) · helper section: dates-hub „15. Break-glass credential (the host's emergency root password)”

# Answer Mark Pointer
1 A per-host emergency password, sealed in the hub, shown only to a logged-in operator. Operator. SaveHostRecoveryCredential; reveal. A HUB/store/host_recovery.go:28 HUB/web/hosts.go:983
2 NO END DATE: last write wins, day-0 sets it and a rotate re-sets it. Paper rule: re-print the sheet "when a key rotates, and yearly". A HUB/store/host_recovery.go:20 RB/break-glass-sheet.md:75
3 The operator, yearly by the paper rule; no code reminds him. A HUB/store/host_recovery.go:28 HUB/web/hosts.go:983 (the code path of question 1)
4 If the sealed copy cannot be opened (seal key, block 8) or the paper is stale, the operator has no way into a host whose agent is dead. A HUB/store/host_recovery.go:28 HUB/web/hosts.go:983 (the code path of question 1)
5 Nobody before. Each reveal writes the event recovery_credential_revealed. A HUB/web/hosts.go:1010
6 Yes, the password does not age on the box. A HUB/store/host_recovery.go:28 HUB/web/hosts.go:983 (the code path of question 1)
7 No renewal on a date exists, so none can fail on a date. A HUB/store/host_recovery.go:20-43
8 SOURCE; session-only route; DOC. A yearly re-print has no record: NOBODY KNOWS. A HUB/store/host_recovery.go:20-28 HUB/web/routes.go:130 RB/break-glass-sheet.md:75

Hand step in a runbook? yes — RB/break-glass-sheet.md:75 (print the sheet again each year)

DH19 — The installer ISO · Version · MENTIONED

architecture folder: 3 lines in 3 files match /ISO_VERSION|installer ISO|iso\.felhom\.eu/ (a machine count, not a reading) · helper section: dates-hub „19. The installer ISO (Proxmox base, installer version, first boot)”

# Answer Mark Pointer
1 The USB image a household or volunteer installs from. Product; built by, served from Cloudflare R2, DOC. A E/scripts/iso/build-felhom-iso.sh:51 RB/total-loss-of-dooplex.md:125
2 NO END DATE in code. ISO_VERSION is 1.30.0. The Proxmox ISO and its checksum are build ARGUMENTS, not pinned: the example names 9.2-1. A E/scripts/iso/build-felhom-iso.sh:51 E/scripts/iso/build-felhom-iso.sh:194 E/scripts/iso/README.md:40
3 No document says how old a public ISO may be, or when to build a new one. The ISO carries a Proxmox base and a first-boot program that age. N E/scripts/iso/build-felhom-iso.sh:51 RB/iso-release-gate.md:1 · class T3 · S22
4 An old ISO installs an old Proxmox base; the install script itself is fetched fresh at first boot (comment block at), so the hub decides the agent and golden. A E/scripts/iso/build-felhom-iso.sh:44-52
5 Nobody. The build warns only if golden and floor are not given: unenforced branch. A E/scripts/iso/build-felhom-iso.sh:79
6 A too-old ISO on first boot: registers, pairs, fetches the current install script; first failing step would be that fetch or the manifest. No "ISO too old" check found (grep too old/min.*version in bootstrap). A E/scripts/felhom-host-install.sh:472
7 A failed install is retried every 30 s; a failed script fetch is retried. A E/scripts/iso/felhom-bootstrap.sh:430 E/scripts/iso/felhom-bootstrap.sh:396
8 SOURCE (checksum of the given ISO is verified). How old an ISO still installs cleanly: NOBODY KNOWS. Build image is FROM debian:trixie, not pinned. A E/scripts/iso/build-felhom-iso.sh:222 E/scripts/iso/Dockerfile.assistant:10

Hand step in a runbook? yes for a build — RB/iso-release-gate.md; NO for 'when'

DH20 — The certificate that the hub's address serves · Certificate · MENTIONED

architecture folder: 1 lines in 1 files match /cert-manager|letsencrypt-prod/ (a machine count, not a reading) · helper section: dates-hub „20. The certificate the hub's address serves”

# Answer Mark Pointer
1 The TLS certificate for the hub's public name, used by every box and the operator. Operator's cluster issues it: annotation, stored under the secret name at. A E/manifests/hub.yaml:342 E/manifests/hub.yaml:360
2 Read from outside 2026-11-11, with no login: the certificate of the hub's address ends 2027-01-12. The download address ends 2026-12-23, the ISO address 2026-12-27, the Gitea address 2027-01-03, the website 2027-01-28. A EV/box/public-reads.txt
3 NOT MEASURED on DooPlex. cert-manager on DooPlex renews by itself. Read from outside: the certificate served today was issued on 2026-10-14, so a renewal worked. A E/manifests/hub.yaml:342 EV/box/public-reads.txt
4 If it lapses: boxes that pin or verify the hub cannot report or fetch config; households get no alarms and no updates; the operator page shows a browser warning. A E/manifests/hub.yaml:342 E/manifests/hub.yaml:360 (the code path of question 1)
5 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). In the Felhom repositories nothing warns before this date. If it ends, every box fails to report, and the hub then sees all boxes as silent. N E/manifests/hub.yaml:342 EV/box/public-reads.txt · class T1 · S01
6 Yes. The agent trusts the public issuers, not one certificate, so a renewed certificate is accepted. A AG/internal/hub/client.go:41-51
7 The old certificate serves until its end date. Nothing in the Felhom repositories warns (question 5). A E/manifests/hub.yaml:342
8 READ FROM OUTSIDE, no login · SOURCE and only. Dates: NOBODY KNOWS from files; needs a measurement. A EV/box/public-reads.txt E/manifests/hub.yaml:342 E/manifests/hub.yaml:360

Hand step in a runbook? NO

DH22 — Operator access tokens, among them CC's hub token · Key · NAMED

architecture: AR/05-hub-architecture.md §17 · helper section: dates-hub „22. Operator access tokens, among them CC's hub token”

# Answer Mark Pointer
1 Named, scoped tokens for scripts, shown once, stored hashed. Operator creates them in a browser session; CC holds one, with overview:read and floor:write, DOC. A AR/09-update-architecture.md:1022
2 90 days by default, tokenDefaultDays; at most 365, tokenMaxDays; "never" is not offered. End computed at. A HUB/web/tokens.go:29 HUB/web/tokens.go:30 HUB/web/tokens.go:254 HUB/web/tokens.go:180
3 The operator makes a new token and puts it in CC's store. There is no renew; no lead time is set anywhere. A AR/09-update-architecture.md:1022 (the code path of question 1)
4 CC's floor raise and overview reads get 401; a controller release cannot be carried to boxes by CC until the operator acts. Households wait for the update. A HUB/web/tokens.go:95
5 Nobody. A token lives 90 days by default and 365 at most. The token page shows the date; no monitor reads it. On the end date the release step that sets a floor answers 401. N HUB/web/tokens.go:29-30 HUB/web/tokens.go:180 · class T1 · S03
6 Not a box matter. A AR/09-update-architecture.md:1022 (the code path of question 1)
7 Not applicable; an expired token stays listed as expired. Column expires_at. A HUB/store/api_tokens.go:52
8 TEST TestToken_NoTokenUnknownExpiredRevoked401; DOC. The end DATE of CC's token is in no document (grep HUB_TOKEN in AR, RB, operations, STATUS): NOBODY KNOWS without the hub page. A HUB/web/tokens_test.go:310 AR/05-hub-architecture.md:655-669

Hand step in a runbook? partly — the build skill says '401 = token expired'; no step says to look at the date before

DH23 — A household's own Cloudflare token · Key · MENTIONED

architecture folder: 9 lines in 4 files match /Cloudflare token|CF_DNS_API_TOKEN/ (a machine count, not a reading) · helper section: dates-hub „23. A household's own Cloudflare token (hub side)”

# Answer Mark Pointer
1 A household that brings its own domain gives a Cloudflare API token; the hub checks what it can reach before saving. Household owns it.. A HUB/cloudflare/reach.go:14
2 The hub and the box do not know the end date that the household set at Cloudflare. No document tells the household to set none. N HUB/cloudflare/reach.go:14 · class T3 · S15
3 The household, at Cloudflare; then the operator saves the new token in the hub. A HUB/cloudflare/reach.go:14 (the code path of question 1)
4 If it lapses, the household's remote address stops being managed; candidate: first sign is a failed DNS or tunnel call on the box. A HUB/cloudflare/reach.go:14 (the code path of question 1)
5 Nobody. When the token ends, the web certificate stops renewing (see DB01), with no sign until the certificate's own end. N CT/internal/infra/infra.go:157-159 · class T1 · S15
6 The running apps and the tunnel do not use this token. The certificate renewal does. A CT/internal/infra/infra.go:157-159
7 The renewal is the household's hand step at Cloudflare. If it is not done, the web certificate stops renewing (DB01). A HUB/cloudflare/reach.go:14 (the code path of question 1)
8 SOURCE; the rest NOBODY KNOWS from hub files. A HUB/cloudflare/reach.go:14

Hand step in a runbook? NO

DH24 — The mail service key (all household and operator mail) · Account · MENTIONED

architecture: AR/_recovery-inventory-2026-07-28.md · helper section: dates-hub „24. The mail service key (all household and operator mail)”

# Answer Mark Pointer
1 The key the hub uses to send every mail: alarms, bind links, claim codes. Operator's account at the mail service; named in, value not opened. A E/manifests/hub.yaml:148-152
2 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). No document gives an end date for the key or for the account. N RB/secrets.md:33 · class T3 · S08
3 The operator re-mints at the service dashboard, DOC. A RB/total-loss-of-dooplex.md:47
4 All mail stops: no alarm reaches the operator or a household, no bind link, no claim code. Everyone. A E/manifests/hub.yaml:148-152 (the code path of question 1)
5 Nobody. No scheduled check proves that mail still goes out. If the key ends, no alarm mail can say so. N RB/secrets.md:33 · class T1 · S08
6 Not a box matter; boxes send through the hub. A E/manifests/hub.yaml:148-152 (the code path of question 1)
7 If the key ends and no new key is made, no mail goes out, alarm mail included (question 5). A HUB/web/selfbind_mint.go:82
8 DOC and. Whether the key or the plan has an end date: NOBODY KNOWS from files. A RB/secrets.md:33 RB/total-loss-of-dooplex.md:47

Hand step in a runbook? yes for a new key — RB/secrets.md:33

DH25 — The hub's global bearer key · Key · NAMED

architecture: AR/13-threat-model.md:70 · helper section: dates-hub „25. The hub's global bearer key”

# Answer Mark Pointer
1 The operator-tier key for the hub API (recovery mode, peer re-add). Operator. Named in; inventory row. A E/manifests/hub.yaml:157-161 AR/_recovery-inventory-2026-07-28.md:501
2 NO END DATE. Rotation is a supervised runbook. A RB/secrets.md:124
3 The operator, with an ordered procedure; no cadence stated. A RB/secrets.md:124
4 A rotation done out of order locks out whatever still holds the old key; DOC calls the order load-bearing. A RB/secrets.md:124
5 Nobody; no date. A E/manifests/hub.yaml:157-161 AR/_recovery-inventory-2026-07-28.md:501 (the code path of question 1)
6 No date drives it. Boxes use their own keys since 2026-10-12. What a box does that holds only the old global key after a change was not traced in this run. A AR/13-threat-model.md:70
7 No renewal on a date exists, so none can fail on a date. A E/manifests/hub.yaml:157-161 AR/_recovery-inventory-2026-07-28.md:501 (the code path of question 1)
8 DOC only. Two documents disagree on whether it was rotated: "rotated" at, "when last rotated" unknown at. Candidate for a doc fix. A AR/13-threat-model.md:70 AR/_recovery-inventory-2026-07-28.md:1030

Hand step in a runbook? yes — RB/secrets.md:124

DH26 — Pinned tool versions in builds and CI · Version · MENTIONED

architecture folder: 8 lines in 2 files match /security.watch|pinned by digest/ (a machine count, not a reading) · helper section: dates-hub „26. Pinned tool versions in builds and CI”

# Answer Mark Pointer
1 Versions of the tools that build and check the product. Operator and CC. Hub build image E/hub/Dockerfile:3, run image E/hub/Dockerfile:24, site publisher. A E/scripts/publish-site-cloudflare.sh:28
2 NO END DATE in code. Hub images are pinned by digest E/hub/Dockerfile:3; the ISO helper image is the moving tag at; the mailer image is latest at. A E/scripts/iso/Dockerfile.assistant:10 E/manifests/contact-mailer.yaml:54
3 Whoever bumps the pin. The nightly security watch runs at 03:40 UTC; nightly gates at 03:20. A E/.gitea/workflows/security-watch.yml:22 E/.gitea/workflows/gates.yml:27
4 A pinned digest removed upstream breaks the hub build for the operator; a moving tag changes the ISO build without a commit. A E/scripts/publish-site-cloudflare.sh:28 (the code path of question 1)
5 The security watch treats a release line with no fixed version as end of life, it reports to the operator. A E/scripts/security_watch.py:292
6 Not a box matter. A E/scripts/publish-site-cloudflare.sh:28 (the code path of question 1)
7 A CI job that does not run leaves no result. Whether a missed night run is seen is a question of the Clock list. A E/.gitea/workflows/gates.yml:27
8 SOURCE E/hub/Dockerfile:3 and. Support end dates of Go, Alpine, wrangler: no document states one. A E/scripts/security_watch.py:292

Hand step in a runbook? no hand step

DH27 — Domains and registrars · Account · MENTIONED

architecture folder: 4 lines in 3 files match /registrar|domain renew/ (a machine count, not a reading) · helper section: dates-hub „27. Domains, registrars and the home dynamic-DNS name”

# Answer Mark Pointer
1 NOT MEASURED. The product's two domains and the dynamic name the hub's address points to. Operator. Listed as row S13; the hub name is a CNAME to the dynamic name. A RB/total-loss-of-dooplex.md:46 RB/total-loss-of-dooplex.md:73
2 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). No document gives a registrar or a renewal date for any of the product's names. This run could not read the public records: no whois tool here, and the public lookups for .eu and .hu gave no answer. N EV/box/public-reads.txt · class T3 · S06
3 The operator, at the registrar. A RB/total-loss-of-dooplex.md:46 RB/total-loss-of-dooplex.md:73 (the code path of question 1)
4 If a domain or the dynamic name lapses, every box loses the hub (reports, alarms, updates, bind links) and the site and installer disappear. Every household. A RB/total-loss-of-dooplex.md:46 RB/total-loss-of-dooplex.md:73 (the code path of question 1)
5 Nobody that a document names. A name that ends takes down the hub address, the website, the mail and every household address under it. N EV/box/public-reads.txt · class T1 · S06
6 A box comes back by itself once the name resolves again; while it does not, the first failing step is the hub lookup. A RB/total-loss-of-dooplex.md:46 RB/total-loss-of-dooplex.md:73 (the code path of question 1)
7 The renewal is the operator's hand step at the outside account. If it is not done, the thing ends (question 4). A RB/total-loss-of-dooplex.md:46 RB/total-loss-of-dooplex.md:73 (the code path of question 1)
8 DOC. WHERE THE OPERATOR READS THE DATE: NOT FOUND. NOBODY KNOWS from files. A RB/total-loss-of-dooplex.md:46

Hand step in a runbook? NO

DH28 — Cloudflare account, the site publish token, the ISO store keys · Account · MENTIONED

architecture folder: 3 lines in 3 files match /CF_PAGES_TOKEN|R2 / (a machine count, not a reading) · helper section: dates-hub „28. Cloudflare account, the site publish token and the ISO store keys”

# Answer Mark Pointer
1 NOT MEASURED. The account that serves DNS for both domains, the site, the installer script and the ISO store. Operator. Row S11; site and ISO there since 2026-10-30. A RB/total-loss-of-dooplex.md:44 RB/total-loss-of-dooplex.md:125
2 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). No document gives an end date for the account, for the publish token or for the store keys. N E/scripts/publish-site-cloudflare.sh:12 · class T3 · S07
3 The operator at the Cloudflare dashboard. A RB/total-loss-of-dooplex.md:44 RB/total-loss-of-dooplex.md:125 (the code path of question 1)
4 Token lapse: CC cannot publish the site or installer script; the old one stays served. Account trouble: DNS, site, ISO downloads and households' tunnels, per row S11. A RB/total-loss-of-dooplex.md:44
5 Nobody. N E/scripts/publish-site-cloudflare.sh:12 · class T1 · S07
6 A new box needs the installer script from the site at first boot; it retries until it is served. A E/scripts/iso/felhom-bootstrap.sh:396
7 A failed publish leaves the previous site live (candidate; not stated in the script header). A E/scripts/publish-site-cloudflare.sh:12
8 DOC. WHERE THE OPERATOR READS THE DATE: NOT FOUND. A RB/total-loss-of-dooplex.md:44

Hand step in a runbook? NO

DH29 — Hetzner account: the off-site endpoint and the storage boxes · Account · MENTIONED

architecture: AR/_recovery-inventory-2026-07-28.md:1019 · helper section: dates-hub „29. Hetzner account, the rented endpoint and the storage boxes”

# Answer Mark Pointer
1 NOT MEASURED. The paid account behind ep0 (tunnel endpoint and off-site store) and the Storage Boxes. Operator. Row S8; API token named in. A RB/total-loss-of-dooplex.md:41 E/manifests/hub.yaml:222-226
2 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). No document gives a payment day, a contract end or the end of the API token. The recovery inventory asks the question and no document answers it. N AR/_recovery-inventory-2026-07-28.md:1019 · class T3 · S07
3 The operator; paid monthly by the provider's own cycle, not stated in any document read. A RB/total-loss-of-dooplex.md:41 E/manifests/hub.yaml:222-226 (the code path of question 1)
4 If the account is closed or unpaid: every household's off-site copies and the PBS tunnel stop; the hub cannot mint or reset sub-accounts. A RB/total-loss-of-dooplex.md:41 E/manifests/hub.yaml:222-226 (the code path of question 1)
5 Nobody before. After the end, the hub's 'off-site copy is old' alarm fires at 8 days. N HUB/monitor/deadline_tiers.go:1 · class T1 · S07
6 A box comes back by itself when the endpoint is back; first failing step is the off-site push. A RB/total-loss-of-dooplex.md:41 E/manifests/hub.yaml:222-226 (the code path of question 1)
7 The renewal is the operator's hand step at the outside account. If it is not done, the thing ends (question 4). A RB/total-loss-of-dooplex.md:41 E/manifests/hub.yaml:222-226 (the code path of question 1)
8 DOC. WHERE THE OPERATOR READS THE DATE: NOT FOUND. A RB/total-loss-of-dooplex.md:41

Hand step in a runbook? NO

DH30 — Tailscale node keys (the operator's remote way in) · Key · MENTIONED

architecture folder: 4 lines in 2 files match /tailscale|tailnet/ (a machine count, not a reading) · helper section: dates-hub „30. Tailscale node keys (operator's remote way in)”

# Answer Mark Pointer
1 NOT MEASURED. The operator's private network to DooPlex and the two demo boxes. Operator. Table. A E/documentation/operations/tailscale.md:11-13
2 DOCUMENT: demo-hp's key ends 2027-01-17; the other two nodes have no end. MEASURED 2026-11-11: the Tailscale service is not running on demo-hp and Tester 1; it runs on the N100. So the date could not be read on the box. A E/documentation/operations/tailscale.md:41-43 EV/box/pin.txt
3 The operator, in the admin console (Machines, Disable key expiry). No lead time set. A E/documentation/operations/tailscale.md:43
4 That node drops off the private network; the operator loses remote reach to it. No household is on it. A E/documentation/operations/tailscale.md:91
5 Nobody. The date is on the Machines page of the Tailscale console. N E/documentation/operations/tailscale.md:41-43 · class T1 · S04
6 The box keeps running; only the operator's remote path is gone until someone re-authenticates on site or over the LAN. A E/documentation/operations/tailscale.md:11-13 (the code path of question 1)
7 The renewal is the operator's hand step at the outside account. If it is not done, the thing ends (question 4). A E/documentation/operations/tailscale.md:11-13 (the code path of question 1)
8 DOC. Candidate: the section at says "both" nodes while the table has three. A E/documentation/operations/tailscale.md:41 E/documentation/operations/tailscale.md:89-92

Hand step in a runbook? yes — E/documentation/operations/tailscale.md:41-43

DH31 — Gitea and registry credentials · Key · MENTIONED

architecture folder: 2 lines in 1 files match /gitea.{0,40}(token|credential)/ (a machine count, not a reading) · helper section: dates-hub „31. Gitea and registry credentials”

# Answer Mark Pointer
1 NOT MEASURED. The login the cluster and the operator use to pull and push images and code. Operator. Row S15; the pull secret is a scoped token. A RB/total-loss-of-dooplex.md:48 RB/secrets.md:231
2 NOT MEASURED (DooPlex, ep0 and the accounts are outside this run). No document gives an end date for the scoped token or for CC's Gitea token. N RB/secrets.md:1 · class T3 · S13
3 The operator, in Gitea. A RB/total-loss-of-dooplex.md:48 RB/secrets.md:231 (the code path of question 1)
4 If the pull token lapses: the hub pod cannot be re-pulled on a restart, and boxes pulling the controller image through the same registry would fail (box helper's item). A RB/total-loss-of-dooplex.md:48 RB/secrets.md:231 (the code path of question 1)
5 Nobody. N RB/secrets.md:1 · class T1 · S13
6 Not a box matter: a box pulls its images with no login. A RB/secrets.md:1
7 The renewal is the operator's hand step at the outside account. If it is not done, the thing ends (question 4). A RB/total-loss-of-dooplex.md:48 RB/secrets.md:231 (the code path of question 1)
8 DOC and. WHERE THE OPERATOR READS THE DATE: NOT FOUND. A RB/secrets.md:231 RB/total-loss-of-dooplex.md:48

Hand step in a runbook? yes for a new token — RB/secrets.md

DH32 — CC's secret store: the tokens inside · Key · MENTIONED

architecture folder: 3 lines in 2 files match /cc-secrets|sec run/ (a machine count, not a reading) · helper section: dates-hub „32. CC's secret store: its key and the tokens inside (GitHub advisory token, hub token)”

# Answer Mark Pointer
1 NOT MEASURED. The store CC uses through sec, and the key that opens it. Operator owns; row S19; design in /mnt/5_hdd/felhom.eu/git/cc-secrets/ARCHITECTURE.md (lines 47-48 give the key location). A RB/total-loss-of-dooplex.md:52
2 The store holds a value and no end date. A token with an end at its issuer (the GitHub advisory token, the hub token) ends with no record here. N E/scripts/security_watch.py:1 · class T3 · S14
3 The operator stores a new value; only he may delete one. A RB/total-loss-of-dooplex.md:52 (the code path of question 1)
4 An expired GitHub token: the nightly security watch loses its source. An expired hub token: block 22. A E/.gitea/workflows/security-watch.yml:22
5 Nobody. The nightly security watch and the release step fail on the day. N E/scripts/security_watch.py:1 · class T1 · S14
6 Not a box matter. A RB/total-loss-of-dooplex.md:52 (the code path of question 1)
7 The renewal is the operator's hand step at the outside account. If it is not done, the thing ends (question 4). A RB/total-loss-of-dooplex.md:52 (the code path of question 1)
8 DOC. WHERE THE OPERATOR READS THE DATES: NOT FOUND. NOBODY KNOWS from files. A RB/total-loss-of-dooplex.md:52

Hand step in a runbook? NO

DH33 — SSH keys between the hub, DooPlex and the endpoint · Key · MENTIONED

architecture folder: 5 lines in 4 files match /ssh key|id_ed25519/ (a machine count, not a reading) · helper section: dates-hub „33. SSH keys between the hub, DooPlex and the endpoint”

# Answer Mark Pointer
1 NOT MEASURED. Keys the hub uses to reach the endpoint (three key files named at,) and the root key, row S9. Operator. A E/manifests/hub.yaml:191 E/manifests/hub.yaml:202 E/manifests/hub.yaml:209 RB/total-loss-of-dooplex.md:42
2 NO END DATE in any document. The only rule: rotate on an endpoint rebuild, DOC. A RB/offsite-endpoint.md:313
3 The operator, on rebuild. A E/manifests/hub.yaml:191 E/manifests/hub.yaml:202 (the code path of question 1)
4 A lost or changed key: the hub cannot add tunnel peers, mint off-site tokens or poke the endpoint; new households get no off-site tier. A E/manifests/hub.yaml:191 E/manifests/hub.yaml:202 (the code path of question 1)
5 Nobody; no date. A E/manifests/hub.yaml:191 E/manifests/hub.yaml:202 (the code path of question 1)
6 Existing boxes keep their own tunnel and token; only new provisioning fails. A E/manifests/hub.yaml:191 E/manifests/hub.yaml:202 (the code path of question 1)
7 No renewal on a date exists, so none can fail on a date. A E/manifests/hub.yaml:191 E/manifests/hub.yaml:202 (the code path of question 1)
8 DOC; row S9 says one authorized key's owner is "not recorded". NOBODY KNOWS who holds it. A RB/offsite-endpoint.md:313 RB/total-loss-of-dooplex.md:42

Hand step in a runbook? yes — RB/offsite-endpoint.md:313