New shared scripts/instructions_gate.py, registered in controller_gates.py and
agent_gates.py, never copied into a sibling repo (the reuse_refs_check.py
precedent). 20 fixture tests, all asserting the effect: exit code AND that the
message names the file and the reason.
It is a consistency gate, not a budget gate, and the failure message says so. A
/context reading measured the instruction files at 15k tokens against 869k free in
a 1M window -- space is not the constraint, and a future reader must not re-derive
the wrong reason. The 200-line ceiling is adherence guidance; a file nobody can
hold in their head is where contradictions hide, and five were found here.
Checks run against effective text (HTML comments stripped, because they are
stripped before injection): the line ceiling; every .claude/rules/*.md declares
paths: or an explicit unconditional: true; no component version literal; no
TEMPORARY block carrying a past date; and the workspace-root CLAUDE.md is
byte-identical to its versioned copy -- the live file sits outside any git repo,
so that copy is its only version-controlled record.
Two traps recorded so they are not reintroduced: a bare \d+\.\d+\.\d+ matches the
first three octets of every IPv4 (the gate excludes dotted quads, or it fails on
192.168.0.180 in the agent's own file); and unconditional: true is NOT a Claude
Code feature but this project's own marker.
Workspace-root CLAUDE.md 208 -> 182 lines (142 effective), copy kept identical.
The nine-instance invariant table moved into the felhom-testing skill, which
triggers when writing or reviewing a test; all three directive bullets stayed in
the core. felhom.eu/CLAUDE.md got surgical corrections only and is knowingly still
over the ceiling at 227 effective lines -- closing it needs the restructure R-229
defers, said plainly rather than quietly absorbed.
CONTEXT.md gains standing ruling S-35. OPEN-ITEMS.md gains R-229.
Docs only -- no Go, no version bump, nothing built or deployed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JJc8sAGRWmavP3rMtdpkr2
- OPEN-ITEMS: R-196 CLOSED; R-204 items 1-3 CLOSED with item 4 named and
its dependency stated. Header restates that R-202, the 1.2 GB ciphertext
deletion and R-198's still-unit-proven retention all REMAIN OPEN.
- capability map: the recovery row keeps its 'with a person present'
qualifier, names which crutch remains, and cites the three now gone.
- 07-backup-architecture: new 7.0 - what a customer can and cannot do
ALONE, the four steps in a table with status. This is the section a
future reader will use to answer that question.
- CONTEXT: standing ruling S-32, superseding S-31 steps 2-5.
- STATUS: rewritten to one screen per its own header; removes a corrupted
half-overwritten section left from the drill session.
- ROADMAP: R-196 and R-204 collapsed.
Read-only recon of the escrow -> recovery chain, from a dead node to an open
repository. No production code, no build, no version bump.
Headline: the hub's superseded-escrow retention does NOT retain the offsite
repository password. host_escrow_superseded has no identity_blob column and
demoteCurrentEscrowTx copies only the K-escrow blob, so what survives a
supersession is the PBS datastore key, not the restic repo password. The next
escrow ceremony -- which the system tells a rebuilt box's customer to run --
destroys the last copy. Both demo boxes crossed that line on 2026-08-04.
Also established:
- the hub's blob-serving endpoints (re-enroll / restore-directive) have zero
callers anywhere: agent, hub UI, scripts, runbooks (R-199)
- POST /backup/offbox/inject-password is routed and handled but no template
contains the form (R-200)
- nothing in the recovery path has ever been exercised; the one live
round-trip proof (2026-06-10) predates the ResticRepoPassword field (R-201)
- a fail-closed mint refusal IS implementable: the report ACK already carries
escrow{identity_blob_present, restic_pw_sha256} and the controller discards
it whenever no offbox target exists
Corrections: yesterday's spike annotated (candidate (b) overturned in part --
unattended recovery is impossible, customer-present is not); capability-map
retention claim struck through and replaced with what the code does.
Deliverable: documentation/audits/RECON-offsite-dr-chain-2026-08-04.md
Register: new R-198..R-201; R-193 and R-192 updated; STATUS.md refreshed.
The 20-minute latch expired at 10:20:29 and the hub logged degraded -> ok
(agent_capability_recovered) at 10:30:40. Final state on both boxes: agent
0.124.1, two ACL rows on /storage/felhom-backup.
Part 0's gate PASSED — ep0 prunes both namespaces daily since 2026-07-27 (18
tasks, all OK) — but three of my own queries said the opposite and all three were
broken instruments. Acting on them would have disabled the only pruning attempt
while filing a finding that nothing prunes.
Also records that v0.124.0's transition record failed in production with a green
test suite, that two red-proofs did not fail on the first attempt (one could not
compile, one asserted a helper rather than the path), and that two hollow tests
were caught in one file.
Four SCHEDULED runs, none triggered by hand: demo-felhom host 83.8s / offsite
540.4s; demo-hp host 109.3s / offsite 300.1s. Each restored into a scratch guest,
booted, verified and destroyed itself; zero 990000 guests or volumes afterwards
and both local-lvm figures returned to their pre-run values.
Both boxes had BOTH tiers due at once, so R-86's ordering was observed live for
the first time: never-proven sorted first, each box took its HOST tier, deferred
the offsite one, and picked it up on the next evaluation six hours later. The
host-tier proofs reached the hub through R-189's merge — demo-felhom's report
carries two tiers, and the local one can only have come from disk.
The capability map's optimistic half is cashed, with its scope stated: these two
boxes, not the fleet.
Surfaced and filed rather than fixed:
- R-190: a storage ACL that demonstrably worked at 04:44 was gone by 09:24, with
a reinstall, any logged pveum activity and any cluster-log entry ruled out.
- R-191: every weekly offsite backup uploads successfully and then fails the job
on a prune the box is deliberately not allowed to do (R-89 moved it
server-side; both boxes still arm keep_last=2).
Two corrections to yesterday's record: the R-185 drift DID surface as 403s on the
write path (six, with the hub raising whole_guest_backup_failed at the first), and
my earlier "no restore_test_* events" was produced by grepping a 404 page.
- OPEN-ITEMS: R-86 CLOSED with the trap in its own wording recorded (the literal
reading is never true on a daily tier); R-87 re-ranked UP because R-86 built
most of what it waited for; R-185 (the agent cannot list demo-felhom's host
backup tier — a missing storage ACL, pre-existing), R-186 (a released binary's
sha is not reproducible from its tag), R-187 (R-115's publish leg had never
actually run) filed. R-184 was the highest ID in use.
- ROADMAP: R-86 collapsed, keeping the reasoning and correcting the shape the row
itself proposed — which would have been the never-fires version.
- 07-backup-architecture: new contract section — restore-testing is per ARCHIVE
GENERATION, with the trap and what did not change (S-1).
- 00-capability-map: the unattended restore-proof row upgraded to PROVEN-LIVE on
the 635 s due-triggered offsite run, with the restart and teardown evidence.
- CONTEXT: S-17 (the rule, the trap, the config key, the hub's derivation) and
S-18 (ep0 is Tier 2 — extends D-d's protected list to three machines).
Numbered 17/18 because S-14 and S-15 were already duplicated in the file.
- STATUS: rewritten for the operator, trimmed back to one screen.
R-182 CLOSED (controller v0.194.0 + hub v0.90.0/.1), proven live on demo-hp. The
hub's notification_log for the run reads: two per-app failures RECORDED, one
digest SENT naming both, and the customer channel SKIPPED with operator_only.
Against the measured previous behaviour — two failures, one email naming one
app, one leaving no trace anywhere.
Scenario D proved itself on an event I had not planned: disk_critical alarmed on
two filesystems, the second was collapsed by the cooldown, and that collapse is
now visible WITH ITS KEY. Yesterday it would have left nothing at all.
A gap the spec did not anticipate is recorded with its fix: the per-app event
also fires from the periodic sweep, outside any run, so making it record-only
would have created a NEW silence. The sweep emits a digest too, with no run_id,
so it stays under the ordinary hourly cooldown.
ep0: MEASURED on the box — 7757 MB (8 GB), 4 vCPU, and the 4 GiB swapfile
SURVIVED the resize and is active (checked, because a resize is a stop/start).
The 40 GB local disk is UNCHANGED, so no disk figure was touched anywhere.
Five documents corrected — three of which the task's list did not name, found by
searching. Two audit/evidence documents ANNOTATED, body untouched: they record
what was true when written and that is their value.
R-90 CLOSED. R-86 unblocked and re-ranked, stated honestly: 8 GB is comfortable,
not unbounded — the original OOM was a 14.46 GB restore — so the restore-test
cadence should still be paced, just not by fear of the endpoint.
target-selection.md's "D-d did not name ep0 either way" is deliberately left
standing. It is the operator's question, not CC's.
STATUS.md 127 -> 83 lines, items rather than sentences.
R-182's direction REVERSED by Part 0's measurement. Filed yesterday as "the
reserve re-alerts on every status refresh" — too many alerts, seen at the
sending end. Measured at the receiving end: 9 events received today, 2 operator
emails sent. When two apps are refused in the same second the operator is told
about ONE; the other is dropped before LogNotification, so it leaves no row on
any channel and cannot be audited. The operator cooldown key is
customerID:eventType(+tier) and the capture-failed event carries `app` but no
`tier`, so the key has no app identifier. Same failure mode as R-97a, in a
second event type that never opted into the narrow fix. Nothing changed —
Part 0 was investigation only.
Correction owed: yesterday's report said "one recovery_unit_capture_failed per
app, HTTP 200". True of what the CONTROLLER pushed; a reader would take it as
"the operator was told about each app", which is false.
R-110 CLOSED (installer v1.23.0). Both channels moved. The spec's mechanism for
channel 2 rested on a factual error — the run-time fetches are sixteen, not
nine, and come from felhom-agent, not this repo — so no tag here could cover
them; pinned to the agent version being installed instead, on the operator's
ruling. Channel 3 needed no change: the URL never carried a ref, so no hub
change and no hub bump.
R-115 CLOSED. release-agent.sh builds, tags, publishes and verifies by an
independent download; check-published-versions.py refuses a tag with no package;
CI now runs the full gate set so it actually runs.
R-183 NEW+CLOSED: a fresh install fetched the vouched agent binary and its
sixteen config files from two different refs, and nothing compared them.
R-184 NEW: nothing stops the hub vouching a version that was never released.
The R-115 gate cannot see it — measured, the hub manifest and Gitea's package
listing are both 401 anonymously.
capability map: new PROVEN-LIVE row for the published installer channel.
STATUS.md 138 -> 127 lines.
R-181 CLOSED (controller v0.193.0 + v0.193.1) and proven live on demo-hp for
BOTH reserve terms. The reserve is now a per-app, per-run ADMISSION decision
taken before the app's first write and covering all three write legs, and it
gained a size term. The refusal's wording was not weakened; the behaviour moved
so it became true, verified by sha256 tree fingerprint.
R-156 CLOSED — papra's template mounts the app's own data root. Precondition
re-measured rather than inherited (both boxes were wiped today).
Part 4, documentation only, nothing built:
- R-110 WAITING-ON-OPERATOR -> READY. Ruling: option (b), the installer's
publish channel moves to a TAG. Recorded with the condition that decides
whether it works at all — it must cover BOTH the /scripts/ git-sync and the
nine files the installer fetches from raw/branch/main.
- R-115 WAITING-ON-OPERATOR -> READY. Ruling: mechanism (b), a build-side gate
refusing to deploy or vouch an unpublished version. The third instance (agent
v0.120.0) would have silently downgraded both demo boxes while succeeding.
R-182 NEW: the periodic status refresh has no admission scope, so a refused app
re-alerts on every poll (measured: a second alert pair 13s after the run's).
Pre-existing in v0.192.0; deliberately not fixed in the R-181 task.
capability map: the local-backup row moves to PROVEN-LIVE in BOTH halves.
ROADMAP: R-165 collapses to CLOSED; R-181 collapsed into it.
07-backup-architecture.md: the reserve's contract stated as what the code
provides (S-1 — an architectural contract changed in the same session).
STATUS.md trimmed 150 -> 111 lines, "What's broken" no longer holds shipped
work, and the stale "After:" line (pointing at work that shipped on 2 August)
is fixed.
Two boxes, two DIFFERENT supply paths, so the session proved the disk shape and
the delivery route rather than one of them twice.
demo-hp (layout proof, --golden <local volid>): mp0 at /var/lib/felhom,
backup=1, 70G, no mp1; /var/lib/docker and /mnt/sys_drive both real mounts of
its subdirectories via fstab; one df figure and one device id (64519) on all
three paths; reboots 3/3 with the binds surviving each.
demo-felhom (pipeline proof, --force-gitea-golden): fetch_verify succeeding
against the vouched manifest for BOTH artifacts -- 'verified sha256
54e2a4c431daf580... matches the hub manifest' for the golden, a7763d31... for
the agent. 250G single volume, grep -c '^mp1:' = 0, reboots 3/3.
Journey proven on both, endpoint-level: claim -> deploy -> back up -> restore,
with a planted marker returning byte-identical on each box. Ceiling measured
gone: 65 GiB and 233 GiB available to a recovery unit, against 19 and 45.
R-165 -> IMPLEMENTED, not PROVEN-LIVE, on the operator's ruling. B2, which that
row records as the bulkhead's replacement, fired live for the first time and
does refuse per app, delete nothing and alert -- but it is checked only in
captureAllRecoveryUnits while runVolumeDumps writes the bulk unguarded, and its
'the previous unit is untouched' claim was measured false (182,272 B dump
replaced by 2,147,666,432 B under a manifest still dated 06:34:26). -> R-181.
New: R-179 (uninstall leaves NAS network-storage units), R-180 (--archive-storage
not cross-checked against the ACL grant; 403 at step 8/8 after root@pam is
rotated), R-181. Third instance of R-115 recorded (agent 0.120.0 unpublished).
No code written, no version bumps -- this was a runbook.
07-backup-architecture.md gains §7.5.1 (S-1: the contract changed in the same
session): the ceiling §7.5 describes no longer exists for a box built from
golden >= 0.192.0, the bulkhead's replacement is recorded, and R-175 is FIXED
here rather than left standing — the bound is restated as a function of mp1
and scoped to split-layout boxes, naming all three real shapes.
Capability map: new row as IMPLEMENTED, deliberately NOT proven-live, with
the missing leg named — no box has been reinstalled from the golden, and
"the golden baked" is not "a box built from it works".
R-163 CLOSED: the ceiling it recorded stops existing. R-176(a) answered by
P1; (b) WITHDRAWN, since every node is reinstalled rather than migrated.
R-178 filed for the reinstalls, which were not done this session.
CONTEXT S-13 (the variant chosen on measurement; pruning rejected with its
reason) and S-14 (prove first, then vouch — the golden is published but
deliberately unvouched, because vouching is what makes a fresh install pick
up a layout no box has been proven from).
STATUS: plain-language section; both operator questions now answered, so the
waiting-on-you item is cleared. Two older entries trimmed so the page did
not grow.
R-167/R-158 shipped and proven live (controller v0.191.x, hub v0.89.0):
two new capability-map rows PROVEN-LIVE with live citations, and
07-backup-architecture.md §7.5's closing claim "nothing warns when an app
crosses the line" is now false and rewritten (S-1: an architectural
contract changed in the same session). §7.5 also gains the caveat that its
size bound is ONE BOX'S, not the fleet's.
Part 3 SPIKE (audits/SPIKE-r165-mp1-merge-2026-08-02.md): M1-M5 measured,
NO layout touched. Three findings the merge session must not re-derive:
"the layout" is not one thing (200G/50G vs 50G/20G vs 16G/8G); mp1 is a
BULKHEAD and not only a ceiling, so after the merge an overflow reaches
/var/lib/docker; the golden fails closed on the split in four places.
D-a's condition (1) is currently SATISFIED — no external box is in the
hub's register, and both demo boxes are Tier 0 and reinstallable.
Recommendation given, choice NOT made — it ends at the operator's ruling.
CONTEXT.md S-11 (D-c's routing, and why R-158's own backup_failed proposal
was overruled) and S-12 (the monitoring landed BEFORE the merge).
STATUS.md gains the plain-language section and the merge decision, with two
older entries trimmed so the page did not grow.
New rows R-174 (closed same session), R-175, R-176, R-177; each ID grepped
free before minting.
SPIKE-ci-runner-2026-08-02.md: all six probes with method, measurement and ruling; none
STOPped. P2 (stock image has git but no python3) and P6 (a runner that loses its state
re-registers and orphans the old record) changed the design; P5 (a failed run signals
NOTHING) is why the alarm exists at all.
R-168 SHIPPED with its evidence. R-29 CLOSED — on the demonstrated alarm and not on a green
run, as required: the class it opened is answered at both ends, the hook refusing locally and
CI catching a --no-verify bypass and emailing. R-161 noted: its automatic half now exists for
the STATIC gate, while its original scope, the runtime gate, is deliberately still not
automatic and should stay that way.
NEW R-169 (grep established R-168 was the highest in use): CI can only report, because there
is no gate in the road. Making it blocking needs branch protection plus a PR workflow, both
of which change how the operator works — so it is theirs to decide, and the row states the
cost honestly rather than recommending it.
CONTEXT gains S-8 (CI detects, does not block, and why that is structural), S-9 (a detector
that tells no one is not finished, plus the curl and Cloudflare-1010 traps), S-10 (the runner
is unprivileged because DooPlex is Tier 2), S-11 (CI reproduces the sibling layout).
CLAUDE.md gains the rule earned by red-proofing: a go test -run pattern that matches no test
prints ok and exits 0, and an instrument that can silently drop results is not a measurement.
REPORT overwritten per the standing rule; every red-proof, the core.hooksPath probe's four
measured outcomes, Scenario C's refusal-and-bypass, the hub deployment and the live Setup-tab
read are recorded there, plus three observations and two deliberate deviations from the spec
(a comment-only edit to felhom-host-install.sh, and __pycache__ in .gitignore).
STATUS: the 'check that needs a person to remember it' line is rewritten — the checks now run
themselves before every push, with both honest limits stated in plain words; and one entry
records the thirteen-check cleanup and the deleted installer version number.
Documentation only — no code, no box, no build.
STATUS.md (repo root, 652 words / 67 lines): what works · what's broken ·
what we're working on · waiting on you · changed since. A VIEW of
OPEN-ITEMS.md, holding nothing of its own; not CONTEXT.md, and both files
now say why they stay separate. No R-n is the subject of a sentence —
identifiers are bracketed pointers only.
CONTEXT.md S-5 records the four operator decisions taken 2026-08-02
(D-a … D-d), none of them implemented:
D-a merge mp1 into mp0 rather than resize it — before any external
install, and D-c ships in the same step → R-165
D-b desired/observed app state in its own store, with the state-store
safety rule verbatim → R-166 (BLOCKED)
D-c customer fill warning + operator backup-failure alert → R-167
D-d only DooPlex and Peti's box are protected → target-selection.md
R-163 RE-FRAMED, not closed: the sizing question is withdrawn rather than
answered; the row survives as the record of the constraint until R-165
lands. R-156's papra referral RESOLVED — deployed nowhere, so the template
fix strands nothing; the docker ps evidence is recorded with its
provenance and its scope limit.
target-selection.md: two protected machines, everything else disposable.
ep0 is no longer Tier 2 but is not scratch (it holds the only off-premises
copy of real customer data) — flagged for explicit operator confirmation.
The demo-box backup-target fence drops from prohibition to stated cost,
because D-d spends that reference anyway.
CLAUDE.md gains an End-of-session checklist carrying the STATUS.md
maintenance rule and "a finding goes in OPEN-ITEMS.md first".
The credential existed and was not reachable when it was wanted. Every box has
had a strong random root@pam password since TASK G1, vaulted in the hub at day 0
and used for real during the sshd incident — but the only way to read it back was
a hand-written curl carrying the global operator key, a secret kept out-of-band.
In practice the PVE web console on a demo box felt locked.
The host page grows a Console access card: presence + username + set_at by
default, Reveal fetches the plaintext on demand for 60 s with a Copy button.
Masking clears the JS variable, and also fires on a second click and on
visibilitychange. A host with nothing vaulted says so, and says why.
The secret is NEVER rendered into the page, and that constraint shapes the
change. The render path uses a new store.GetHostRecoveryMeta whose struct and
SELECT both omit the secret column, so it is structurally incapable of carrying
one. The plaintext crosses the wire only in the response to POST
/hosts/{id}/reveal-recovery-credential (Cache-Control: no-store, CSRF-gated at
the ServeHTTP level; POST precisely so that gate applies and so no secret is
retrievable by URL alone). Deliberately NOT the customer page's data-secret
widget, which embeds the plaintext on every load.
A delivered reveal writes one recovery_credential_revealed event on the host's
customer timeline (info, source hub, Hungarian) via SaveEvent alone — no
dispatcher, nobody emailed, the log_tail_requested shape. Two reveals write two
events: the register records accesses, not states. A 404 is not an access. An
unbound host reveals fine and writes no event; the [INFO] hub line, carrying the
username and a length only, is then the record.
The global-key API path is untouched by design — it is the route for when the
hub UI itself is broken, and coupling it to the session layer would delete the
independence that makes it a fallback.
Recorded as a real trade: the hub session password alone now unlocks console root
fleet-wide, where retrieval previously also needed the global key. Accepted for a
single-operator, HU-geo-fenced hub that already stores these passwords in
plaintext at rest (CONTEXT.md ruling S-4). The plaintext-at-rest half is filed as
R-133 — every hub DB backup is a fleet-wide console-credential dump.
Tests 550 -> 559; four red-proofs (page leak, audit event, CSRF gate, route
order) each run, observed failing, and reverted. The route-order proof is a seam
test driving ServeHTTP: a handler-level test cannot see that defect, because the
handler is correct and simply never runs.
Live: hub 0.81.0, agent 0.113.0, controller 0.185.1 on both demo boxes;
host-install 1.22.0 (script; no reinstall performed).
E-2a wrapper proven live as root on demo-hp: F-1 subdirectory refused, F-2
unmounted path refused, root device refused, idempotent re-apply is a no-op,
repointing refused -- 0 stray storages. The agent PVE role was NOT widened.
Scenario E proven live on BOTH boxes: healthy renders nothing, no message key.
Records three defects I introduced and caught: unreachable routes (mounted
outside /api/storage/, caught by the first live call), a hollow test exposed by
its own red-proof, and another gofmt-realignment no-op.
Not live-proven: the degraded banner and offer acceptance (both boxes healthy),
backup_target_absent end-to-end, Case A/B on a real install, drive-loss recovery.
hub 0.81.0, agent 0.112.0, controller 0.184.1 live on BOTH demo boxes.
E-2c: eject/decommission of the backup-target drive refused 409 on both boxes,
drives unmoved. E-2b: the never-called disconnect seam is wired, with the target
case raising the specific backup_target_absent.
Records the keying bug caught before deploy (a.Path is the GUEST path, so the
target branch was unreachable -- 0.184.0 superseded, never deployed) and states
plainly that backup_target_absent is NOT proven end-to-end live: proving it needs
a live enrolled drive to go absent.
Parts 2/3/4 and E-2a remain open; Peti risk stays parked.
Supervised operational run. No code, no version bump, nothing deleted.
Primary backup tier on both demo boxes moved from `local` (a dir storage on
/var/lib/vz -- the SAME physical device as the guest) to `felhom-backup`, a dir
storage on each box's secondary drive:
demo-hp /mnt/nvme-1tb uuid:91d2dc2d-... archive 2,256,044,492 B
demo-felhom /mnt/hdd_1 uuid:47a3361a-... archive 5,957,878,962 B
Both proven end to end via the real UI path: archive lands on the secondary
drive (df delta matches the archive byte-for-byte), restore-test auto-selects it
and passes with mount_parity: ok, and freshness survives an agent restart with
an empty in-memory store -- so the age can only have come from the new storage.
Phase 0: the target is CONFIGURATION, not converged (the sole writer of
agent.json touches only escrow.pbs_storage_id and preserves unknown keys), so
the runbook's STOP did not fire. No consumer hardcodes "local" on the backup path.
Findings:
- F-1 the storage path must BE the mountpoint; a subdirectory fails exactMount
and the target reports disconnected permanently (observe.go:321)
- F-2 --is_mountpoint 1 is load-bearing; proven live, an unguarded storage on a
non-mounted path reports active with the ROOT filesystem's free space and
had already created dump/ on pve-root -- a silent retarget onto the very
device this change escapes
- F-3 FelhomAgentStore is granted per storage path; without it every backup
403s. felhom-host-install.sh must issue it for new installs
- R-109 (new) the DR recipe records no backup target, and each box now carries
two content=backup dir storages, one live and one frozen
- R-105 narrowed and TRACED: dr_recipe drives was [] fleet-wide because the
enrolled drives were never PVE storages, so isUserDataDrive never saw
them. Both boxes now populate drives; SMART on the backup drives too
Absent-drive behaviour today is fail-loudly with no silent retarget (PVE half
live-proven; agent half source-traced). That is NOT the intended fall-back-and-
alarm design -- filed as E-2 with the honest single-drive label.
Reported in full in the record: the agent was restarted with a felhom-pbs backup
in flight, producing a spurious tier failure. The backup had in fact succeeded
(PVE task OK, 6,264,034,053 B snapshot) and the spurious failure reached no
channel -- R-84 ground truth superseded it.
Outstanding: full drive-loss recovery (needs physical access) and the agent half
of the absent-drive behaviour.
Phase 0 sized C9-F1 properly before anything was designed: 43 of the 53 catalog apps have NO
subtree the Tier-2 restore can read (not 2), 9 are covered only for their file legs and never
their database or volumes, 1 is stateless. The asymmetry is Tier-2's alone — Tier-1 and offsite
both restore the unit and replay volume dumps, so BookStack always had a working restore and
only this button lied.
Shipped: the restore refuses BEFORE stopping the app and names the action that does work; a run
that proceeds claims only what it EXAMINED and discloses that the database and volumes are not
covered. C9-F2 alarms after a 5-minute sustained-restarting threshold, set above the 120s deploy
timeout, Mealie's 60s start_period and R-97b's 180s grace; StateRestarting is deliberately NOT
added to IsDownState.
Live: silent through ten 30s samples then app_start_failed at 5m25s, heartbeat now reads
"1 currently down" where Campaign 9 recorded 0; a real deploy stayed silent; bookstack refused
with its uptime unbroken; paperless re-restored 43/43 byte-identical, 16/16 docs clean.
Filed, not fixed: C9-F1b (route to the Tier-1 restore — its own task because it puts a
destructive operation behind a non-destructive button) and C9-F4 (nothing reads the Tier-2
copy's recovery-unit/ mirror, so the second local copy that exists for drive loss is unreachable
by any customer action — potentially larger than C9-F1).
Phase A is the headline and it passed on live hardware, through the real endpoints the UI
posts to: a customer who deletes files — or their entire app data directory — gets everything
back byte-identical, and the app works afterwards (paperless served the restored bytes over
its own API at the exact pre-deletion sha256). A1's two non-destruction promises both hold.
Three defects, recorded not fixed:
C9-F1 (HIGH) the Tier-2 restore button is offered for apps it can never restore, takes a
real outage, and reports "nothing was missing" — indistinguishable from a
genuine result, while 156 MB of that app's data sits unread in the same copy.
C9-F2 (HIGH) an app in a crash loop never alarms on any channel; StateRestarting is in no
down-set, so F-OBS's own heartbeat printed "0 currently down" for 9 minutes.
C9-F3 (MEDIUM) an interrupted offsite run leaves a lock the self-heal cannot reach; the tier
is dead until a human unlocks and the operator is told "unknown reason".
This answers Phase C item 8.
Two candidates were deliberately NOT filed: a recovery-unit poisoning the catalog sync healed
in ~3 min, and a snapshot_id that is documented as logging-only. Reporting either would have
been reporting an artifact.
Stopped at the end of Phase B (plus D10), then full recovery — both boxes healthy, real
cadences, offsite tier proven working again, no leaked scratch guests, peti untouched.
D11's approved staleAfter compression turned out not to be a knob; reported, not worked around.
Seam sweep: TieredBackend was the FIRST, not the only one. BackupArchiveLister
has the identical silent-degrade shape and a worse blast radius (it degrades to
the pre-R-84 in-memory-only behaviour), and no compile-time witness existed in
production code anywhere in either repo. No defect found, so no version bump and
no deploy — the witnesses are guards, proven by breaking a signature and watching
go build fail where it previously passed.
Live outage: age_state=unknown captured on real hardware for the first time, with
demo-felhom's local tier genuinely due throughout — the controller deferred and
zero app stacks were stopped. The R-88 breaker did NOT arm and no
whole_guest_backup_failed travelled, because felhom-pbs was not due; recorded as
conditions-did-not-arise rather than claimed as coverage.
Post-boot: the volume changed device name (sdb->sda) across the reboot and the
mount survived only because fstab uses by-id. That was never tested before.
Corrects two wrong severity readings with evidence from the box and the code.
The PBS outage was ~15 min (07:00-07:18 UTC), caused by a global OOM at 06:58:12:
proxmox-backup-proxy peaked at 3.2G on a 3.8G box and a concurrent 1.9G rsync
tipped it over. Root SSH to that box works from DooPlex via the public IP, not
from felhom-pve via the tunnel IP — the documented path I failed to try first.
R-88: internal/quiesce has NO failure limiter, backoff or breaker; the loop
stopped after three cycles only because PBS recovered. Verified additionally that
scheduledRunAllowed (quiesce.go:476-478) returns true whenever lastAgeSecs is nil,
so the same missing value that makes every poll due also bypasses the time-of-day
gate — the cycles ran outside the [04:30,08:30) window. Fixing the due-verdict
without fixing the nil-age bypass would leave the hole open.
Recorded after the session report was written. The offsite PBS service stopped
listening on 8007 five minutes after this session's 14.46 GB restore-test read
from it; the box is up and the tunnel is healthy, but no SSH key to it exists so
the cause is unestablished — the restore load is a plausible mechanism on a cx23
and is recorded as correlation, not cause.
The agent restart then exposed R-88: an unreachable target reads as 'no backup
exists', so the offsite tier is perpetually due and the controller runs a full
quiesce cycle every ~5 min. Operator ruling: leave it running, it self-heals when
PBS returns and masking it would hide the fault.
The unattended offsite restore-test on demo-felhom passed: 14.46 GB archive,
duration_s=635.07 (10m35s), then it rotated to the local tier. Persisted state
confirms the credit: {"felhom-pbs": "2026-07-27T06:14:42Z"}.
CORRECTION: I estimated ~2 hours for this restore. It took 10m35s. I derived
the estimate from a download rate measured during the FAILED attempt, which was
running under contention; the real link does ~1.4 GB/min. I then used that wrong
figure to raise a design concern — that the heavy-op gate would block backups
for hours on this box — which at 10 minutes largely evaporates. An estimate
extrapolated from a degraded measurement is not a measurement.
The SPEC's closing risk note is corrected in place, with the original left
visible for the lesson.
R-86 (NEXT, operator ruling 2026-07-27): backup-ALIGNED restore-test scheduling
— test a tier ~1 day after ITS OWN backup. R-85 schedules on a free-running
interval, which cannot express 'the day after the PBS backup': any fixed offset
drifts, so alignment would be luck. Shape: trigger from the tier's own last
successful backup rather than a clock. Interim in force: 302400s (3.5d), which
lands each tier ~weekly — the cadence half of the ruling, not the alignment half.
R-87: the restic app-data offsite tier is NEVER restore-tested. R-85 covers
whole-guest vzdump tiers only; the agent has no restic surface. That is arguably
the tier that matters most — the only one that survives losing the box AND
carries the customer's app data, since the whole-guest snapshot excludes the
bind-mounted drives. Exactly the state PBS was in before R-85.
REPORT.md: the full R-80 -> R-85 arc, including a section on the seven mistakes
I made and the two recurring shapes behind them (inferring behaviour from an
artifact instead of the code that consumes it; reading a result without its exit
code). Records demo-felhom's restore-test as IN FLIGHT at close, with the manual
recovery step if the deferred restart watcher does not complete.
Hub gate green (17 packages, rc=0).
Restore round-trip on demo-hp: pass=true, verified=boot+running,
mount_parity=ok, source_tier=pbs (the v0.100.0 fix — the earlier attempt said
'local' and died at 600s), 4m5s restore+boot+verify+teardown, clean teardown
with no 403 and no leak. That last point confirms 06's reading that the
teardown 403 was a phantom, and corrects my earlier framing of it as a standing
privilege gap.
Multi-tier quiesce driven through the REAL UI endpoint (authed+CSRF):
exactly ONE stop/start pair with BOTH backups inside it, local-first/PBS-last,
app quiesced through the non-last tier, early resume on the last tier's
snapshot. Total downtime 1m27s for both tiers; app healthy after.
Capability map row upgraded IMPLEMENTED -> PROVEN-LIVE, kept distinct from the
DR-tier row above which proves ACTIVATION not ARRIVAL. Remaining gaps recorded:
the SCHEDULED restore-test still only selects the primary tier (manual path
proven, unattended not), and the hub infers cadence from storage type.
Slice D.1 — host-install 1.20.0: a FRESH box defaults to local-daily +
offsite-weekly (felhom-pbs, 604800s, keep_last=2). setdefault semantics proven
both ways: fresh gets the tier, an UPGRADE preserves the existing backup block
verbatim — so an in-place upgrade can never silently start writing to an
offsite datastore. Existing boxes are migrated explicitly.
Slice E:
- 07-backup-architecture.md: honest status header per CONTEXT ruling S-2, with
an explicit STALE-outside-the-PBS-tier verdict (the controller tiers were last
verified 41 controller versions ago). The PBS row claimed 'PBS on DooPlex'
(the retired spike store) with no cadence; it now names felhom-pbs ->
felhom-offsite on ep0 over wg-felhom, weekly, keep_last=2. NOT marked
ratified — that is Viktor's review of the section 10 list. Discharges R-83.
- 06-offsite-connectivity.md: the target-split remaining-work note collapsed
(shipped), and records HOW S4.1's tier-aware timeout silently regressed — the
mechanism was never removed, its INPUT changed when local_backup_target was
retargeted to 'local'. Also notes S4.1 already diagnosed the teardown 403 as a
phantom (a timeout consequence, not an ACL gap).
- capability map: new row for recurring offsite backups actually LANDING, as
distinct from the existing row proving ACTIVATION. IMPLEMENTED, not
PROVEN-LIVE — the restore round-trip has not completed under the fixed code.
- ROADMAP: R-82 SHIPPED with its remaining gate named, R-83 DISCHARGED, R-84
left open.
- CONTEXT + REPORT: the arc, including the mid-arc correction I had to make.
- B2 demo-hp + B3 demo-felhom migrated to the island (agent 0.96.0), apps
served throughout (0 container restarts), island /storage 200, LAN DNS pinned
to the LAN IP, hub reports 0.96.0. No rollback.
- capability-map 'site/network change' row PARTIAL -> PROVEN-LIVE
- ROADMAP R-50 -> SHIPPED (fleet-migrated); add R-74 (island on Peti's cluster)
- nodes.md: both boxes island-bound, agent 0.96.0
- A4: scratch guest provisioned from golden on the island-configured drill came
up with net1 automatically (zero manual edits) — the v0.96.0 provisioning path
proven live. Spike method caveat CLOSED.
- REPORT: Phase A done + vouched 0.96.0; B0/B1/A4 done; B2/B3 await operator go.
admin
Claude Code