agent (unreleased): a cut write of agent.json cannot leave a blind box — a journal of the old bytes, a repair at start (R-1380)

Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS

CHANGELOG.md

@@ -1,4 +1,318 @@
-## Unreleased (2026-10-28 night) — R-528: GET /guest/oom-kills, the kernel's OOM counter of each of the guest's containers, read on the host — commit only; the next agent release (needs the operator's vouch)
+## Unreleased (2026-11-15) — a cut write of `agent.json` cannot leave a blind box (R-1380; found before the release)
+
+- **`config.ReplaceInPlace` + `config.RecoverCutWrite`** (`internal/config/cutwrite.go`). `agent.json` is the agent's
+  file in a root-owned folder, so the trust move, the PBS seed and the backup-target move write it in place
+  (truncate, write). A power cut in between left an empty or half file and an agent that does not start. Now the old
+  bytes go into a journal in the state folder first (`agent.json.write-in-flight`, durable), the config is written,
+  synced and read back, and the journal is removed; the daemon's start repairs a config that does not parse from the
+  journal BEFORE it reads it. A selftest never repairs. No journal folder → no write. Tests: a cut staged at each of
+  four steps, the wiring in `main`, and a source pin that the three writers use the helper and open no file by hand.
+  Red-proofs: felhom.eu `documentation/audits/release-2b-3-2026-11-15/part0/`. Not proven: a real power cut.
+
+## Unreleased (2026-11-14 night) — release 3 („the hub tells you”) and the clock-step fix, the agent's parts, built and NOT released
+
+Commit only. Each part can be released alone; none needs release 2b. No `configs/` change in this section.
+
+- **Fixes after an independent review of the night's code (2026-11-14; felhom.eu
+  `documentation/audits/night-build-2026-11-14/review/`).** (1) `config.FileMu`: the three functions that read
+  `agent.json` whole and write it whole (the trust move, the PBS seed, the backup-target move) take one lock — two of
+  them could interleave and the second put back the document it read before the first one's change; pinned from the
+  source by `TestAgentJSONWritersTakeTheSharedLock`. (2) `pvetrust.Manager.Ensure` runs one at a time (start, timer,
+  a failed collect and a bundle install could overlap, and the second parked the already-moved file as the „before”
+  copy), and its root step has a 90-second deadline (it runs on the start path and inside the report loop). (3) **The
+  CA copy keeps its previous certificate behind the new one** (`configs/felhom-os-apply`, mode `pve-ca-copy`): the
+  copy was replaced before anything could be proven with the new root, and Proxmox serves its old certificate until
+  pveproxy is reloaded — an agent that restarted in that window had no root that verified it. **Only for the
+  changeover:** at every later run the step makes one handshake to the local Proxmox with the NEW root alone, and
+  drops the previous root as soon as that verifies, or when the previous root is no longer a CA that has not ended.
+  The copy is compared exactly, so nothing the step did not check stays in it. To drop the old root at once by hand
+  (a root changed because it leaked): delete the copy. `configs/` changes again, so the release needs both signed jobs.
+  (4) The profile-refresh stanza is built from the wrapper's ANSWER (its record of the last refresh comes back at
+  every ask); the record file is a bonus — its folder is root-only on a box where a signed job made it first (read in
+  source, not on a box). `rewritten` is absent when not known. (5) A watchdog give-up is forgotten when the guest is
+  no longer the watchdog's (onboot off, locked, destroyed) — it was reported for ever. (6) `/disks` sends both boot
+  tokens from ONE read of the guest's init (`GuestBootTokens`). 13 mutations seen red, two after a mutant that did
+  not compile was done again.
+- **The hub is told what the kernel clean-up did and what the profile refresh did (R-1383; felhom.eu `11` §5.11.1,
+  §5.12).** (1) The clean-up's report carries `kernel_cleanup` (`osupdate.CleanupDetail`: removed, kept, bytes_gained,
+  leftovers). It still has no `kernel` view and `release_id` stays `kernel-cleanup` — the hub reads only those two as
+  „the kernel this step was for” (the test asserts the wire form). (2) The host report carries `profile_refresh`
+  (`osupdate.ProfileRefreshStatus`, from `Leg.ProfileRefreshStanza`, wired in `main.go`): the root wrapper's own record
+  of the last refresh — so it outlives an agent restart — and how this process's last ask ended. One small file read,
+  no wrapper call, names checked and capped. No OS report is sent for the refresh (an unknown layer reads as a guest
+  step on the hub). Tests `TestKernelCleanup_NightOrderOnceAndNoNames` (extended), `TestProfileRefreshStanza_RecordAndLastAsk`,
+  `TestMainWiresPVETrust`; 9 mutations seen red (one mutant did not compile at first and was done again).
+- **A boot token that a clock step cannot move (R-1395).** `/disks` answers `guest_boot_key` =
+  `<kernel boot_id>-<guest init start>` beside the unchanged `guest_boot_id` (`<boot TIME>-<guest init start>`). The
+  kernel computes the boot time from the wall clock, so a step of the host's clock moved the old token and the
+  controller recreated every drive-backed app (the N100, 2026-11-14, +32.7 s). `GuestBinder.GuestBootKey`; `""` when
+  the guest does not run or the boot_id is unreadable. Only a controller that knows the key (built there, not
+  released) reads it. Tests `TestGuestBootKey_AClockStepDoesNotMoveIt`, `TestDisks_CarriesBothBootTokens`; 6 mutations
+  seen red (felhom.eu `documentation/audits/night-build-2026-11-14/C/redproof-agent-boot-key.txt`; two mutants did not
+  compile at first and were done again).
+- **A refused signed job and a watchdog give-up reach the hub (R-1383, D78, felhom.eu `09` §3 decision 299; `05`
+  §16.6).** (1) `signedjobs.Runner` clears a job with `hub.Client.CompleteJobWithOutcome`: `outcome` (done · rejected
+  · failed · malformed), a `reason` from the closed `signedjobs.ReasonCodes` (the two clock reasons `expired` and
+  `not_yet_valid` first — a box with a wrong clock refuses every job) and the op's name, in the `DELETE`'s query. No
+  signature, nonce or job parameter is sent. (2) The guest-power watchdog's give-up — one ERROR line in the journal
+  until now — rides the host report as `guest_power_gave_up` (`localapi.Server.GuestPowerGaveUpStanza`, wired in
+  `main.go`); the list empties when the guest runs. Tests `TestRunner_TellsTheHubHowEachJobEnded` (the real gate
+  and verifier), `TestReasonCode_ClosedList`, `TestCompleteJob_OutcomeRidesTheQuery`,
+  `TestGuestPower_GiveUpIsReportedUntilTheGuestRuns`, `TestMainWiresPVETrust`; 12 mutations seen red (felhom.eu
+  `documentation/audits/night-build-2026-11-14/B/redproof-agent-{job-outcome,guest-power}.txt`).
+- **The report carries the box's clock at the moment of sending (R-1384, D79, felhom.eu `09` §3 decision 300).**
+  `HostReport.SentAt` (`sent_at`, RFC3339 with nanoseconds) is stamped by `Client.Report` at every send, never by the
+  collector. `reported_at` could not serve: it is stamped when the collect starts. The hub (its own unreleased part)
+  compares it with its clock and alarms above one minute. Test `TestReport_StampsSentAtAtSending`; 3 mutations seen
+  red (felhom.eu `documentation/audits/night-build-2026-11-14/B/redproof-agent-sent-at.txt`).
+
+## Unreleased (2026-11-14 night) — release 2b, built and NOT released: trust in Proxmox by a copy of the box's own root CA (R-1380, D80, felhom.eu `09` §3 decision 303; design `03-host-agent.md` §12a)
+
+Commit only. No box ran any of this; the proof on a box (demo-hp, a forced new certificate, twice) is the next attended
+session's. `configs/` changes, so the release needs BOTH signed jobs.
+
+- **The root step that keeps the copy** (`configs/felhom-os-apply`, host-layer mode `pve-ca-copy`). It copies exactly
+  `/etc/pve/pve-root-ca.pem` to exactly `/etc/felhom-agent/pve-root-ca.pem` (0644 root), only if the source is ONE
+  certificate with `CA:TRUE` that has not ended; written whole or not at all; „same" when nothing changed. The plan may
+  name nothing (T1); no source (T2); not one CA certificate (T3). No appliance check: the copy is the agent's own trust
+  file and changes nothing of a BYO owner's host. Tests `PVECACopy` (9, one with the real openssl through the real
+  runner); 7 mutations seen red (felhom.eu `audits/night-build-2026-11-14/A/redproof-wrapper.txt`).
+- **The move of a box that exists, in a safe order** (`internal/pvetrust`, new; wired in `cmd/felhom-agent/main.go`
+  at the agent's start, after a config bundle install, on a timer — hourly until moved, daily after — and on every
+  call to Proxmox that fails on trust). (1) the root step makes the copy; (2) ONE real call to Proxmox with the copy
+  ALONE, while the pinned fingerprint is still the configured trust; (3) only then ONE write of `agent.json` sets
+  `proxmox.tls.ca_file` and clears `fingerprint` (the agent's own file, rewritten in place after a recovery copy
+  `agent.json.pre-pve-trust` was parked, read back and checked); (4) the RUNNING client takes the new trust
+  (`proxmox.Client.SetTLS`, new) — no restart. A failed copy, proof or write changes nothing: the fingerprint stays.
+  **`proxmox.tls` keeps its order (a set fingerprint wins over a CA file), so the fingerprint is cleared LAST.** An
+  operator's own CA file, a box with checks switched off and a config that is not a file are never moved.
+- **A pinned box whose certificate changed heals by itself:** the failed call starts the same order, the box moves to
+  the CA copy and the next call works. A refreshed copy (a renewed root CA) is proven and reloaded by the running
+  client.
+- **The cause is named** (`proxmox.TrustProblem`, new): the pinned certificate changed; the CA copy did not sign the
+  served certificate; a certificate ended; the copy is unreadable; the copy ends within 30 days; the root step gave no
+  copy; the proof failed. An ERROR line in the journal, and the new `proxmox_trust` stanza of the host report
+  (`mode`, `problem`, `detail`, `ca_not_after`, `ca_days_left`, `moved_at`). **The hub does not read the stanza yet**
+  (felhom.eu R-1396), and a box that is fully blind sends no host report at all — that case still shows only as
+  „host stale" on the hub.
+- Tests: `internal/pvetrust` (11, against a real TLS server that changes its certificate and renews its CA), the AST
+  wiring test `TestMainWiresPVETrust`; 12 + 3 mutations seen red (felhom.eu `audits/night-build-2026-11-14/A/redproof-agent.txt`;
+  one mutation is not a rule and says so).
+
+## v0.174.0 (2026-11-14) — the box removes its old kernels (R-1368); security profiles are refreshed after a kernel change (R-1388, R-1301); Go 1.27.4 (R-1391) — RELEASED 2026-11-14 (release-agent.sh from the clean working clone at `cc63dd5`, CI run id 2641; agent sha256 6efbb97809135483bc446d71d1321f9e2f7fc3268fe9b20e7fe742a2eb13db92, bundle sha256 7d87d0584dee79eee766f01c27b9e89273d39c89a14b0725a6e5564a55b3ec5f) and VOUCHED 2026-11-14 15:21:45 CET
+
+Release 2a of the build order (felhom.eu `09` §3 decision 306). Design: felhom.eu `11-os-updates.md` §5.11.1 and §5.12.
+`configs/` changed (the wrapper), so a box needs BOTH signed jobs: `agent_update`, then `agent_config_update`.
+
+**Part A — old kernels (`09` §3 decision 286).** Nothing removed old kernels; a ring-0 box got about one a week, 1 GB
+each. Now:
+
+- **The wrapper** (`configs/felhom-os-apply`, layer `kernel`, new mode `kernel-cleanup`): the hand-proven rule of
+  2026-11-10 as a program. Keep the running kernel, the step's `from`, `to` and `default_before`, both defaults, the
+  flag's kernel, an operator's boot pin and the newest installed; remove every other installed kernel image in ONE
+  `apt-get purge` by exact name. Refusals P1, P2, P3, K1, N1, S1, S2 (the design's table), C1 (the plan names
+  something — it may carry only `dry_run`), R9, R12. „Not run" (exit 0): no root-owned ring-0 mark, no kernel step yet,
+  the newest step did not end `good`, the box does not run its newest kernel, a step within the last 20 hours. K1 reads
+  the box a second time, by itself. Repair-first. After the purge it proves the default, the pinned file, the two
+  generators and the state file unchanged; a left-over folder is reported, never deleted. **Rule R4 and every update
+  plan are unchanged; `apt-get autoremove` is not used.**
+- **The agent** (`internal/osupdate/kernel.go`): the clean-up is the night leg's last call, after a healthy host step,
+  never in a night with a kernel step, never in a debug pass, never on a BYO host. The hub hears `cleaned`, `refused`
+  or `failed` (a kernel-layer report with NO kernel version — the hub would read that as a kernel step); „nothing to
+  remove" and „not run" are log lines. `--selftest=kernel-cleanup-dry-run` prints what it would do and changes nothing.
+- **No new sudo line.** The wrapper is reached by the one `--plan` line (pinned by `test_cleanup_needs_no_new_sudo_line`).
+
+**Part B — security profiles (spike first: felhom.eu `audits/profile-refresh-spike-2026-11-14/`).** After a kernel
+went back and forth a box can load an old compiled AppArmor profile; on the N100 this stopped the time service.
+
+- **The wrapper**, new host-layer mode `profile-refresh` (an appliance only; no package change; the plan names
+  nothing, F1): once per kernel the box starts on (a root-owned record), `apparmor_parser --replace --skip-read-cache
+  --write-cache -- /etc/apparmor.d`; then ONE restart for a `failed` service whose own program's cached profile the
+  rewrite changed (at most 3). Measured on demo-hp: 0.5 s, no guest or container touched.
+- **The agent** (`internal/osupdate/profilerefresh.go`) asks at every start and after a config bundle install, under
+  the heavy-op gate; nothing is sent to the hub.
+
+**Part C — Go 1.27.4** (GO-2026-6859; hygiene: the agent's one `MaxBytesReader` place does not match the advisory).
+
+**Tests, each seen red** (felhom.eu `audits/release-2a-2026-11-14/A/redproof*.txt`, `B/redproof*.txt`): 31 wrapper
+cases for the clean-up, one per acceptance scenario and one per kept-kernel kind (14 mutations, all red); 11 for the
+profile refresh (10 mutations); 5 + 4 agent tests (3 + 3 mutations). Two existing agent tests list the night's calls
+exactly and gained the clean-up call (`TestRing0_OneCallPerLayer`, `TestRing1_NoReleaseIsInventory`).
+
+**Delivered 2026-11-14 to the three test boxes by signed `agent_update` + `agent_config_update` each** (Tester 1,
+demo-hp, the N100): update committed, bundle installed (written 1 of 31), capability probe 77/77, the running binary
+reads `go1.27.4`. **On boxes** (felhom.eu `audits/release-2a-2026-11-14/D/`):
+
+- **Profile refresh, by the product's own path, on all three:** refused by the OLD wrapper at the agent's start
+  (R11, logged), then run right after the bundle install: 0.5 s each, nothing restarted on Tester 1 and demo-hp.
+  **On the N100 it repaired the time service:** the cached chrony profile was rewritten, `chrony.service` (failed
+  since 2026-10-16) was restarted and is active, the clock stepped +32.7 s; offset to DooPlex −32.55 s → +0.18 s. After
+  a restart of Tester 1 the agent answered „not needed" (the same kernel).
+- **Clean-up:** dry runs through the released agent on all three (demo-hp: nothing to remove; the N100: 6 of 8 will
+  go tonight; Tester 1: „not run", it has no ring-0 mark). Refusals P1, P2 and P3 by dry run on demo-hp against states
+  set and put back. **A real removal on Tester 1**, with the ring-0 mark set for the test on the operator's yes and
+  put back: `proxmox-kernel-7.0.14-20-pve-signed` removed, 1.12 GB gained, 8.5 s, the default, the pinned file, the
+  generators and the state file unchanged, no left-over; the second call removed nothing; one orderly restart came
+  back on the same kernel, healthy, a clean stop for the crash guard. **This real run was started by hand** (the
+  agent user, the one sudo line, the night leg's own plan), not by the night leg.
+- **Not proven:** the step by the agent at night (the N100's and demo-hp's own night, 2026-11-15); a box whose running
+  kernel is not its newest and a self-revert into the kept kernel (tests only); refusal K1 on a box (it guards a wrong
+  computation; tests only); a restart by the profile refresh of any program but chrony.
+- **Found:** a clock step makes the CONTROLLER read a new boot (its boot identity is `/proc/stat` btime): on the N100
+  it recreated its three drive-backed apps 9 s after the step (felhom.eu R-1395). The crash guard (kernel `boot_id`)
+  was not fooled.
+
+**Vouched by the operator 2026-11-14 15:21:45 CET** (the hub log: `Artifact manifest set: agent=0.174.0`, the bundle
+sha = the release; the hub's System page: „vouched agent: 0.174.0"). `pinned.day0_agent` = 0.174.0 (D50).
+
+## Unreleased (2026-11-11) — docs only: a repository address holds no credential (R-925, Part C of the clock/end-dates run)
+
+- `scripts/release-agent.sh` header: the tag push needs `GITEA_GIT_TOKEN` from the secret store (`sec run`). The origin
+  address of CC's clone on DooPlex held a login and a token until 2026-11-11; git now asks the helper
+  `felhom.eu/scripts/git-credential-sec.sh`. No code change, no version change, no release.
+
+## v0.173.0 (2026-11-10) — R-1368 part 2: the self-update folder keeps what it needs and no more — RELEASED and VOUCHED 2026-11-10 (release-agent.sh from a clean clone of `776025c`; agent sha256 752b2727fb6043417ae5f342acbda52330521b21bb799fda90760b960feef7ea, bundle sha256 51ebc1f4c66d87918c87c035b4297c0e40e10ed31e3250c3837dfeaa0002e1c1)
+
+- **The defect (measured 2026-11-10, space list place H12).** `Execute` downloads each new agent to
+  `<state dir>/selfupdate/felhom-agent-<version>` and removed that file only when the download or its checksum failed.
+  A good update left it for ever: 49 / 48 / 33 binaries = 740 / 724 / 509 MB on demo-hp, demo-felhom and Tester 1.
+- **What the folder is needed for (read first: `configs/felhom-selfupdate-guarded`).** The root wrapper reads the staged
+  file once, copies it to its own root-owned folder and flips from that copy. The rollback needs `pending.json` (in the
+  folder) and `/usr/local/bin/felhom-agent.prev` (outside it); the commit needs `pending.json`. No staged binary.
+- **The fix (`internal/selfupdate/prune.go`).** After the wrapper's `commit` returned success, `MaybeCommit` removes
+  every staged binary but the running version's. Nothing is removed while a pending marker exists or cannot be read,
+  at a plain start, after a rollback (the failed version's binary stays to be read), or when the running version is
+  not a version this code can name a file for. Only a regular file with exactly the name `Execute` writes is removed,
+  one `os.Remove` per name; a failed removal is logged and never fails the commit. A download and a prune take turns
+  (`stagingMu`).
+- **Tests `TestPrune_*` (6).** After six updates the folder holds one binary and every bystander (unknown names, a
+  folder and a symlink that carry a staged name); a pending or unreadable marker refuses; only a healthy commit prunes;
+  the name rule follows `versionRe`; the wrapper's `commit` and `rollback` blocks name no staged binary; and the
+  wrapper's REAL `commit` and `rollback` verbs, run from a sandboxed copy, still put the previous binary back after
+  four pruned updates. **Red seen three times:** no prune call; no pending refusal; no regular-file check.
+- **Not reached.** `/usr/local/bin/felhom-agent.bak-*` (place H13): root-owned, in a folder the agent cannot write, made
+  by a person. R-1368 part 4 stays open.
+- **Corrected 2026-11-10 (measured on three boxes).** This entry said: „A box gets the clean-up at its first good update
+  AFTER the one that brings this code". That was wrong. The commit runs in the NEW binary, so the box cleans the folder
+  at the commit of the update that brings the code.
+- **Delivered 2026-11-10 to the three test boxes by a signed `agent_update` each (no bundle job: `configs/` did not
+  change since v0.172.1).** The folder before → after, read on each box: demo-hp 49 entries / 722,596 KB → 1 / 16,028 KB
+  (commit 14:59:19 UTC, `removed=49`); demo-felhom 48 / 706,968 KB → 1 / 16,028 KB (commit 14:51:50 UTC, `removed=48`);
+  Tester 1 33 / 497,244 KB → 1 / 16,028 KB (commit 15:00:19 UTC, `removed=33`). After each: `pending.json` absent,
+  `/usr/local/bin/felhom-agent.prev` = 0.172.1, the unit active with no restart.
+- **A rollback still works after the clean-up (Tester 1, 2026-11-10 15:14 UTC).** A signed `agent_update` to 0.172.1
+  (the vouched version; it has no prune code) was installed with its pending marker. One `SEGV` in the 60-second dwell
+  → the unit's `OnFailure` → the wrapper's `rollback` → 0.173.0 ran again 9 seconds after the install, `pending.json`
+  gone, no failed unit. The failed version's staged file stays (by design, until the next good update): the folder
+  holds 2 files, 32,040 KB. Evidence: felhom.eu `documentation/audits/release-space-fixes-2026-11-10/agent/`.
+- **Vouched by the operator 2026-11-10 16:29:16 CET** (read from the hub log: `Artifact manifest set: agent=0.173.0`, the
+  bundle sha = the release). `pinned.day0_agent` = 0.173.0 (D50).
+
+## v0.172.1 (2026-11-07) — proxmox-grub joins the boot-chain lane: the nightly pve step stops being refused R6 — RELEASED and VOUCHED 2026-11-07 (release-agent.sh from `39db030`; agent sha256 505a0594a2fb117f359e329a0b7ae2c7a6902b74c51f52c1883145f3c2ac72f8, bundle 3e7acaee4500893bfaf567b1d427790ef89bf41f08cb6a1419ee528aeb40ddc3)
+
+- **The defect (measured on demo-felhom, 2026-11-05, 11-06, 11-07).** Proxmox published `proxmox-grub` 2.12-9+pmx4. It
+  depends on `grub-efi-amd64` of the same version. `HOST_SLOW_RE` matches `grub` only at the START of a name, so
+  `proxmox-grub` landed in the ring-0 pve plan; the simulation then pulled `grub-efi-amd64`, which is not in the plan,
+  and `felhom-os-apply` refused the WHOLE pve step (R6 „the plan would touch grub-efi-amd64") every night — 34 Proxmox
+  userspace upgrades held with it, and an `os_update_failed` error mail each morning. The same class as the 2026-10-09
+  `proxmox-secure-boot-support` fix. Read-only proof on the box: `apt-get -s install --only-upgrade proxmox-grub` →
+  `Inst grub-efi-amd64 …pmx4`.
+- **The fix.** `proxmox-grub` joins `HOST_SLOW_RE`: GRUB's meta-package now waits with the rest of the boot chain (the
+  slow lane, the operator's signed step), and the pve step runs without it.
+- **Test** `test_ring0_pending_pve_leaves_proxmox_grub_out`; red seen before the fix (proxmox-grub entered the plan), green
+  after; the wrapper suite 149 of 149.
+- **Delivered 2026-11-07 (operator, signed jobs):** demo-felhom, demo-hp, Tester 1 — `agent_update` committed
+  (23:05/23:13/23:13 CET), then the bundle: `BUNDLE DONE agent=0.172.1 written=1 same=30 self-check=ok`, capability probe
+  77/77 on each. `pinned.day0_agent` = 0.172.1 (D50).
+- **Delivery.** `configs/` changed, so a box needs BOTH signed jobs: `agent_update` 0.172.1, then `agent_config_update`
+  0.172.1 with the bundle sha. Not vouched by CC. Bundle vs 0.170.0 (what the boxes run): no path added or removed, only
+  `felhom-os-apply` changed — so no step bundle is needed (R-880). Both downloads re-checked by sha.
+
+## v0.172.0 (2026-11-03) — R-1302: the first grant of a tier is no loss — RELEASED and VOUCHED 10:18:10 CET (release-agent.sh from a clean clone of `ecbb0cc`; agent sha256 2c0222cd01713f65c627d6d17c18b4a8b2b1e0c39881db0df7ccb17de48e24c7, config bundle cc34b0bc3346ffd6a4ba80fe9269852dced8c56c9503c8b14a50d9ded2961cce; configs/ unchanged since 0.171.0, so the signed delivery was agent_update alone)
+
+- **R-1302 — the first grant of a tier is no loss.** A fresh install arms the whole-guest tier before its storage exists, so the
+  first probe found no grant, repaired it and logged ERROR „GRANT WAS MISSING … investigate the loss (R-190)" plus a DEGRADED
+  capability, on every new box (VM 9375, 2026-10-26). The agent now remembers which tiers have EVER held their grant
+  (`<state dir>/store-grant-seen.json`, written by a healthy probe). A first grant logs INFO „granted" and reports healthy; a
+  grant that existed and vanished keeps R-190's ERROR and its one degraded cycle. A doubt keeps the ERROR: an unreadable or
+  corrupt file, or no file path. A box updated to this agent records its tiers at the first healthy probe, so a loss in the
+  one probe before that would read as a first grant. Tests `TestR1302_*`; red seen with the first-grant branch removed and on
+  three more mutants.
+
+## v0.171.0 (2026-10-30) — R-1344 decided from the night's measurement: a controller that runs but does not answer is RESTARTED — RELEASED (release-agent.sh from a clean clone of `bd18347`; agent sha256 c888ca7ba22a1f689059081575814a725e2edcecbb6bcc56829e9d9de88202a5, config bundle 95f977f6dc827185f2b5a2aa9e1c829db41a9f425d49eb40f317286d13c83d2b; binary-only signed agent_update on Tester 1, demo-hp, demo-felhom; vouched by the operator 2026-10-30 10:47:36 CET; `pinned.day0_agent` = 0.171.0). Live on Tester 1: the controller frozen (SIGSTOP) 09:24:50 UTC → Docker „unhealthy" 09:26:36 → the alarm 09:29:21 → the restart 09:31:22 → healthy 09:31:24; the hub mailed `controller_unresponsive` and `controller_unresponsive_restart` (hub 0.178.0), not `controller_restarted_by_agent`
+
+- **The measurement (operator brief 2026-10-29 Part 2, read 2026-10-30).** Docker's own health record of the controller
+  on Tester 1 and demo-hp, every probe from 2026-10-29 11:12 to 2026-10-30 07:00 UTC: 4 733 probes through the night
+  chain (demo-hp's PBS vzdump, Tester 1's local vzdump, both DB dumps, tier-2, the off-site push, a restore test, two
+  controller updates). No failed probe, no 5-s timeout; slowest 303 ms (demo-hp, during the PBS vzdump). The only
+  „unhealthy" was the controller stopped by hand for a proof. A busy night never came near ONE failed probe.
+- **So the alarm becomes a restart.** At `controllerUnresponsiveConfirm` = 6 unhealthy sweeps the alarm is raised, as
+  before. At `controllerUnresponsiveRestart` = 10 (2 minutes later; ~6.5 minutes after the first failed probe) the agent
+  restarts the controller's bootstrap unit — unless the controller's busy file is fresh (a backup, a restore, an
+  update, an export) or cannot be read, or the brake refuses (3 restarts in 15 minutes → the crash-loop pause, which
+  the hub mails). A swap, a parked guest, a locked guest and a vzdump keep resetting the count, as before. A refusal
+  keeps the alarm and is logged once per reason. After a restart the count starts again from zero.
+- **The stanza** carries `unresponsive_restart_after_s` (120: the hub's alarm text says when the restart comes) and
+  `unresponsive_restart_at` (equals `last_restart_at` for that act; the hub mails `controller_unresponsive_restart`).
+- **Refactor, no behaviour change:** the not-running restart's brake and its record are now `supervisorBrakeAllows`
+  and `recordSupervisorRestart`, shared by both restarts. The busy-file stale log line no longer says „operator reboot".
+- Tests `TestSupervisor_UnresponsiveIsAlarmedThenRestarted`, `…Guards` (now also: never restarted),
+  `…RestartRefusals` (busy fresh, busy unreadable, the brake, a crash-loop pause), `…StaleBusyFileDoesNotShield` —
+  red first (build failure, then 3 behaviour failures); 7 mutants, each killed (one survivor — the count reset after a
+  restart — found and pinned).
+
+## v0.170.0 (2026-10-29) — „Reboot guest" from the hub (D65 = A): `reboot_guest` joins the agent's closed list — RELEASED (release-agent.sh from a clean clone of `ca15a31`; agent sha256 da8d12a05ef607c338b9ba307e7a81b18253a3188721bbd8421ba36658a22404, config bundle 03f2ff6442fc3ca600b1f25f26767ce4da466a770a258f3e362cf95f6e76b543; signed agent_config_update + agent_update on Tester 1, demo-hp, demo-felhom; vouched by the operator 2026-10-29 18:57:48 CET; `pinned.day0_agent` = 0.170.0)
+
+- **D65 = A (operator ruling 2026-10-29 15:50; felhom.eu `04` §6.1, `03`).** The agent's closed operator-action list is
+  now `restart_controller` and `reboot_guest` (arg: the vmid). No signature: a reboot destroys nothing. The act is the
+  existing guest-reboot path (`pct reboot <vmid>` of THAT guest over the existing exact grant) — never the host.
+- **Guards** (each refuses with the reason; nothing is called, no record is written): a guest this agent did not
+  provision, not running, locked, or parked; the host-wide heavy-operation gate held (a backup of any tier the agent
+  runs, a restore test, an OS step — the reboot HOLDS the gate until the guest is up again); `jobBlocker` (an agent
+  backup job, a controller swap, a drive or NAS job, an escrow ceremony, a vzdump of the guest); the CONTROLLER busy —
+  its `data/ctl/busy` file (a backup, a restore, an off-site run, an app update), fresh within 90 s; a stale file is a
+  controller that stopped writing, not a busy one; any read failure but „no such file" refuses (fail closed); at most
+  ONE reboot per hour per guest (`<guests>/<vmid>/operator-reboot.json`, on disk — it survives an agent restart); the
+  operator restart's brake (a refusal in force refuses a reboot; a reboot counts as one of the 3 restarts in 15 minutes).
+- **During the reboot** the guest-power watchdog and the controller supervisor leave the guest alone (`rebootInFlight`):
+  a racing start would fail the reboot's own start half.
+- **The result** says both times from the press: „rebooted; the guest was up after 41 s, the controller answered after
+  63 s" — or `failed`, said plainly. The record is written BEFORE the reboot; the new self-scoped
+  `GET /guest/operator-reboot` gives it to the controller for one day, for the household's line „Felhom support
+  restarted your box at <time>".
+- **One new sudo vector** (configs/ changed → the signed config bundle): `pct exec <vmid> -- cat
+  /var/lib/docker/volumes/felhom-controller-data/_data/data/ctl/busy` — read-only, one fixed path, in
+  `FELHOM_CONTROLLERSWAP`; capability `controller-busy-read` (not critical: without it, every reboot press is refused).
+- Tests `TestOpReboot_*` (the times + the record + the gate given back; 13 guard cases; a stale busy file; once per hour
+  across an agent restart; the brake counting a reboot; a failed reboot; the watchdogs leaving the guest alone; the
+  notice expiring after a day) and `TestOpActions_AgentClosedList` — each seen failing by a mutation (19 mutants).
+
+## v0.169.0 (2026-10-29) — „Restart controller" from the hub (the agent's own closed list); the unresponsive controller ALARM (R-1344) — RELEASED (release-agent.sh from a clean clone of `637c92c`; agent sha256 fc066de640007038ecfaed6a23e4301a27fc486f1d7851c36177acdd26c2df61, bundle 3ca33f9cbab3bdd4c1b7f8ac90aa21e43b64209f1cb45c77ac72b3405956041b; binary-only signed updates committed on demo-felhom, Tester 1, demo-hp) — **VOUCHED by the operator 2026-10-29 14:06:09 CET**; Day-0 pin set (D50). Live on Tester 1: four presses (one with the controller stopped, the fourth refused by the brake)
+
+- **Operator actions for the agent (operator brief 2026-10-29, Part 1; felhom.eu `04` §6.1).** The host-report reply may
+  carry `operator_actions: [{id, action, arg}]` (`hub.ControlEnvelope.OperatorActions`); the agent acts ONCE per id and
+  answers on the next report, `operator_action_results: [{id, outcome, message}]` (`hub.HostReport`), which it requests
+  at once through the cap-1 report trigger. CLOSED list: `restart_controller <vmid>` only — anything else refused, nothing
+  called. The act is the supervisor's (`systemctl restart felhom-controller-bootstrap.service` in the guest, existing
+  grants — no sudoers change, no config bundle) with its guards: provisioned guest, guest running, the guest list and the
+  vzdump state readable, not parked, no swap, not locked, no vzdump. The thrash brake counts operator restarts with the
+  supervisor's: 3 in 15 minutes → the press is refused, and presses stay refused for 30 minutes. The result says how long
+  the controller took to answer (Docker's health check: healthy), or `failed` when it did not within 3 minutes. One
+  `audit: operator action` line per decision. Works with the controller stopped. Tests `TestOpActions_*` (closed list,
+  unknown/bad arg refused, restart + answer time + report trigger + once per id + re-send until dropped, stopped
+  controller, every guard, the brake across supervisor and operator restarts, no healthy answer = failed) and
+  `TestMainWiresTheAgentOperatorActions` (AST) — each seen failing by a mutation.
+- **R-1344 — a controller that RUNS but does not answer: ALARM ONLY.** Each sweep also reads Docker's health status for a
+  running controller (the swap's `docker inspect -f` grant); 6 consecutive „unhealthy" sweeps (3 minutes, on top of
+  Docker's own ~90 s) start an episode: one ERROR line, the stanza's `unresponsive` / `unresponsive_since` (the hub raises
+  `controller_unresponsive`, operator-only). NOT restarted for it: the brief allows the restart only after a measurement
+  shows a busy night never reads like this, and that measurement runs through tonight (felhom.eu R-1344). Not counted
+  during a swap, a parked or locked guest, or a vzdump; healthy / starting / no health check / an unreadable status reset
+  the count. Tests `TestSupervisor_Unresponsive*` (red by mutation: the vzdump guard, the threshold).
+
+## v0.168.0 (2026-10-29) — R-528: the host's per-container OOM counter (`GET /guest/oom-kills`); R-1254: a lost or late escrow answer is asked again before it is worded — RELEASED (release-agent.sh from a clean clone of `91bd96c`; agent sha256 1e70196c51346cf10e3f04f51975de5899b5dc6baa5a45ec1d17bc248830d2ca, bundle 3415c034955d46c59ca3f7b284a648abf852b1257b77cc49bd6ea9994a7b8cb8; no configs/ change since v0.167.0 → binary-only signed updates to demo-felhom, Tester 1, demo-hp) — **VOUCHED by the operator 2026-10-29 10:21:56 CET**; Day-0 pin set (D50)
+
+### R-528
 
 - **Why:** inside a Felhom LXC guest Docker can leave `State.OOMKilled` false and send no `oom` event after a
   memory-limit kill (measured 2026-09-15 on 9202 and 2026-09-16 on a fresh install), so the controller's OOM line
@@ -11,7 +325,7 @@
   count is per container run. Tests `TestR528_*` (a fake cgroup tree; red-proved: a glob over every guest leaks another
   guest's container; parsing `oom` instead of `oom_kill` reads the wrong number).
 
-## Unreleased (2026-10-28 night) — R-1254 (the ceremony half, review round 2 HUNT-3 H2): the agent asks the hub which package it holds before it words a failed upload — commit only; the next agent release (needs the operator's vouch)
+### R-1254 (the ceremony half, review round 2 HUNT-3 H2)
 
 - **The defect:** the claim's (R-978) and the confirm's (R-1151) package upload can lose its answer, or answer late
   (30 s), AFTER the hub stored the new package. The agent then answered 502 „the previous code still works" — false: the

CONTEXT.md

@@ -1,5 +1,46 @@
 # CONTEXT — felhom-agent working state
 
+> **2026-11-14 (day, attended) — v0.174.0 RELEASED and VOUCHED (15:21:45 CET); on the three test boxes (agent + bundle, 77/77).** sha `6efbb978…`, bundle `7d87d058…`. Two new wrapper modes: `kernel-cleanup` (layer kernel; old kernel images by the hand-proven rule; ring-0 mark only; the plan names nothing; the night leg's last call, never with a kernel step) and `profile-refresh` (layer host; AppArmor profiles from a fresh compile once per kernel; one restart for a failed service whose cached profile was stale). Go 1.27.4. The N100's time service was repaired by the product path (+32.7 s step). A real kernel removal ran on Tester 1 by hand with a temporary ring-0 mark (operator's yes). **Next:** read the N100's and demo-hp's night (the clean-up by the agent at night is not proven yet); a ruling for ring-1 boxes; release 2b (trust in Proxmox by the root CA copy, D80). Report: felhom.eu `REPORT-release-2a-2026-11-14.md`.
+
+> **2026-11-10 (afternoon, attended) — v0.173.0 RELEASED and VOUCHED (16:29:16 CET); on the three test boxes.** sha `752b2727…`, bundle `51ebc1f4…` (`configs/` unchanged since 0.172.1, so no bundle job). Measured: a box cleans its self-update folder at the commit of the update that BRINGS the code (49 → 1, 48 → 1, 33 → 1 entries; about 16 MB stays). The line below („at the first good update AFTER that release") was wrong. A rollback after the clean-up was proven on Tester 1 with a signed update to 0.172.1 and one crash in the dwell. `pinned.day0_agent` = 0.173.0. Open: the `.bak-*` files (R-1368 part 4).
+
+> **2026-11-10 (day, operator brief) — COMMITTED, NOT RELEASED: the self-update folder cleans itself (R-1368 part 2).** After a healthy `commit` the agent removes every staged binary but the running version's (`internal/selfupdate/prune.go`; tests `TestPrune_*`, the wrapper's real `rollback` verb run in a sandbox). No version bump, no build, no vouch: it rides the next agent release. A box cleans its folder at the first good update AFTER that release. Not reached: `/usr/local/bin/felhom-agent.bak-*` (root-owned, R-1368 part 4). The old-kernel clean-up is a design only (felhom.eu `11-os-updates.md` §5.11.1, proven by hand on demo-hp); nothing of it is in this repo.
+
+> **2026-11-07 (attended) — v0.172.1 RELEASED and VOUCHED; on the three test boxes (agent + bundle, 77/77).** `proxmox-grub` joins `HOST_SLOW_RE` (the N100's ring-0 pve step was refused R6 three nights: proxmox-grub pmx4 pulls grub-efi-amd64). sha `505a0594…`, bundle `3e7acaee…` (only `felhom-os-apply` changed vs 0.170.0, no step bundle). Done: vouch, both signed jobs on the three boxes, `pinned.day0_agent` = 0.172.1. Next: the N100's pve step on its next night (positive line: `os-apply: DONE` for `layer=pve:9201`, no R6).
+
+> **2026-11-03 (day, attended) — v0.172.0 RELEASED and VOUCHED (10:18:10 CET), R-1302.** `2c0222cd…`, bundle `cc34b0bc…`, `pinned.day0_agent` = 0.172.0; on the three test boxes by signed `agent_update`. The appliance path pre-positions the grant, so a plain first start logs neither line; staged live both ways (felhom.eu `audits/release-2026-11-03/part3/`).
+>
+> **2026-10-30 (day, attended) — v0.171.0 RELEASED and VOUCHED (10:47:36 CET): R-1344 decided — a controller that runs
+> but does not answer is RESTARTED.** The night's measurement (4 733 probes, slowest 303 ms, no failure) licensed it:
+> alarm at `controllerUnresponsiveConfirm` = 6 unhealthy sweeps, restart at `controllerUnresponsiveRestart` = 10 —
+> refused by a fresh/unreadable busy file (`controllerBusy`, shared with D65's reboot) or the brake
+> (`supervisorBrakeAllows`, now shared with the not-running restart; `recordSupervisorRestart` records both). The stanza
+> carries `unresponsive_restart_after_s` and `unresponsive_restart_at` (hub 0.178.0 mails
+> `controller_unresponsive_restart`). Built from `bd18347` (`c888ca7b…`, bundle `95f977f6…`), binary-only signed
+> updates on the three test boxes; live on Tester 1 (SIGSTOP → alarm 4.5 min → restart 6.5 min → healthy).
+> `pinned.day0_agent` = 0.171.0. felhom.eu `REPORT-release-2026-10-30.md`.
+
+> **2026-10-29 (evening, attended) — v0.170.0 RELEASED and VOUCHED (18:57:48 CET): `reboot_guest` (D65 = A).**
+> `da8d12a0…`, bundle `03f2ff64…`, tag `v0.170.0`, from a clean clone of `ca15a31`; configs/ changed (one new read-only
+> sudo vector: `cat` of the controller's `data/ctl/busy`), so per box a signed `agent_config_update` + `agent_update`
+> (Tester 1, demo-hp, demo-felhom; capability probe 76/76). The agent's closed list is `restart_controller` +
+> `reboot_guest` (`opreboot.go`; guards in felhom.eu `03`/`04` §6.1). Proven on Tester 1: refused during an off-site run
+> (the controller's busy file), done in 67 s / 79 s, refused a second time inside the hour. `pinned.day0_agent` = 0.170.0.
+> **R-1344 decided 2026-10-30: RESTART** (see the 2026-10-30 entry above).
+
+> **2026-10-29 (afternoon, attended) — v0.169.0 RELEASED and VOUCHED (14:06:09 CET).** `fc066de6…`, bundle `3ca33f9c…`,
+> from `637c92c`; `pinned.day0_agent` 0.169.0 (`db25eb0`). The agent's own closed operator-action list:
+> `restart_controller <vmid>` (host-report reply → result on the next host report; the supervisor's guards; the brake
+> counts operator restarts, then `opBrakeUntil` refuses presses 30 min). R-1344: 6 unhealthy sweeps → `unresponsive_since`
+> → the hub's `controller_unresponsive` — ALARM ONLY until the night's health log is read (to 2026-10-30 07:00 UTC).
+> Proven live on Tester 1: stopped controller restarted, 4th press refused, one press on the release builds „done" in 16 s.
+
+> **2026-10-29 (attended release run) — v0.168.0 RELEASED and VOUCHED (10:21:56 CET).** `release-agent.sh` from a clean
+> clone of `91bd96c` (agent `1e70196c…`, bundle `3415c034…`, tag `v0.168.0`); no configs/ change → binary-only signed
+> `agent_update` to demo-hp, Tester 1, demo-felhom, each committed, self-check 76/76; `pinned.day0_agent` = 0.168.0 (D50).
+> R-528 proven on the release (Tester 1: one `app_oom` „seen by docker,cgroup", host counter 1). R-1254: the happy path ran
+> live (a ceremony on 0.168.0 + controller 0.347.0); the lost-answer branch has tests only.
+
 > **2026-10-28 night (unattended) — R-528 + R-1254 on `main` (`6f933d6`, `5ddd124`), NOT released (the next agent
 > release, the operator's vouch).** `GET /guest/oom-kills` (`internal/localapi/guest_oom.go`): the host-side `oom_kill`
 > of each Docker container of the token's guest (`Server.cgroupRoot` is the test seam); proven on Tester 1 with a test

REPORT.md

@@ -1,30 +1,31 @@
-# REPORT — two agent fixes on main, NOT released (night 2026-10-28 → 29, unattended)
+# REPORT — v0.174.0 released, vouched and delivered to the three test boxes (2026-11-14, attended)
 
-**Plain words first.** Two changes are on main and tested. Neither is released: an agent release needs the operator's
-vouch, and the night brief allowed no release. Both go on the morning list.
+**Plain words first.** A box now removes its old kernels by itself, and it reloads its security profiles after it
+starts on another kernel. Version 0.174.0 is released and runs on Tester 1, demo-hp and the N100. The operator vouched
+it at 15:21 CET. The N100's time service runs again; the product repaired it, with no hand step.
 
-## What changed
+The full report of the run is in `felhom.eu`: `REPORT-release-2a-2026-11-14.md`.
 
-- **R-528 — the OOM counter (`6f933d6`, CI run id 2473 / #284 green).** New local-API route `GET /guest/oom-kills`. It
-  reads the kernel's `oom_kill` count of each Docker container of the token's own guest, on the host, from
-  `/sys/fs/cgroup/lxc/<vmid>/ns/system.slice/docker-<id>.scope/memory.events` (mode 0444, readable by the non-root
-  agent user). Docker inside an LXC guest does not see these kills (`OOMKilled` stays false). The controller reads the
-  route every 30 s (controller `eed23e7`). Tests `TestR528_*`, red first by mutation (another guest's container served;
-  `oom` parsed for `oom_kill`).
-- **R-1254 — the escrow ceremony half (`5ddd124`, CI run id 2471 / #283 green).** After an escrow upload that fails, the
-  agent asks the hub which package it holds (the identity blob compared): the new one → success; another one → 502 and
-  „the previous code still works"; no answer → 504, which promises nothing. Tests `TestR1254_*`, red first.
+## What a reader must know
 
-## Live proof
+1. **The clean-up runs on a box with the root-owned ring-0 mark only.** Tester 1 has no mark, so the program answers
+   „not run" there. For the proof the operator allowed the mark on Tester 1 for some minutes; it is back as before.
+2. **The real removal on Tester 1 was started by hand**, with the agent's own command line. The night leg cannot be
+   started by day: it follows only the night's own backup (R-899). So „the step by the agent at night" is **not proven
+   yet**; the N100's and demo-hp's night of 2026-11-15 is the proof.
+3. **The profile refresh was spiked first** on demo-hp (felhom.eu `audits/profile-refresh-spike-2026-11-14/`). Three
+   things it found are built in: a stale copy is told by the cached file changing, not by the kernel's hash; one
+   profile never compiles the same twice; a refresh hides other kernel audit lines for a few seconds.
+4. **A clock step makes the controller read a new boot** (felhom.eu R-1395, a controller row): on the N100 it
+   recreated three apps once.
 
-- R-528, 2026-10-28 23:27 UTC on Tester 1 with a test build of this agent (put back after: 0.167.0, sha256 `28bdd282…`).
-  A Radicale capped at 128 MB, a memory hog inside it killed: the host counter 0 → 1, one `app_oom` „seen by
-  docker,cgroup", one mail to the operator with `"source":"docker,cgroup"`. Evidence: felhom.eu
-  `documentation/audits/night-2026-10-28/r528-oom/`.
-- R-1254: no live proof. It needs a hub answer lost on purpose; that is a hub change, and no hub deploy runs at night.
-
-## Next
-
-The agent release (the operator's vouch), then `pinned.day0_agent` in `scripts/retention-policy.json` (D50).
-The controller needs no MinAgent change for R-528: an old agent answers 404, and the controller then waits 10 minutes
-before it asks again.
+| | |
+|---|---|
+| Version | 0.174.0, sha256 `6efbb97809135483bc446d71d1321f9e2f7fc3268fe9b20e7fe742a2eb13db92`, bundle `7d87d0584dee79eee766f01c27b9e89273d39c89a14b0725a6e5564a55b3ec5f`, Go 1.27.4 |
+| Code commit, CI | `cc63dd5`, run id 2641 success |
+| Tests | wrapper 191 (42 new), agent `internal/osupdate` green; 30 mutations, each seen red (felhom.eu `audits/release-2a-2026-11-14/A`, `B`) |
+| Tester 1 | 3 kernels → 2, 1.12 GB gained, 8.5 s; second run: nothing; one restart, same kernel, healthy, clean stop |
+| demo-hp | nothing to remove; refusals P1, P2, P3 by dry run |
+| The N100 | time service repaired (offset −32.55 s → +0.18 s); tonight it removes 6 of 8 kernels |
+| Vouch | the operator, 2026-11-14 15:21:45 CET; `pinned.day0_agent` = 0.174.0 |
+| Not proven | the step at night by the agent; a fell-back box and a self-revert into the kept kernel on a box; K1 on a box |

REUSE.md

@@ -35,7 +35,7 @@
 | `SudoHostOps.ListCandidateDisks` | internal/storage/candidates.go | `ListCandidateDisks(ctx) ([]CandidateDisk, error)` | enroll-candidate discovery | Fail-safe: omits anything not provably unclaimed |
 | `antiRetargetResolveExpect` (+ `antiRetargetResolve`, `antiRetargetResolveBlank`) | internal/localapi/wipe_reresolve.go | `(durableID, expectDataBearing, resolve, derive, inspect) (device, err)` | pre-mkfs anti-retarget: resolve durable id → re-derive+match → re-inspect | AGENT-001 + audit D3; refuses path-only bindings; wired via `Server.reresolveWipe`/`reresolveBlank` (test-injectable) |
 | `signedjobs.WipeExecutor.Execute` | internal/signedjobs/wipe.go | `Execute(ctx, op, params) error` | operator-signed data-bearing wipe | Durable-id bound; nonce burned by gate BEFORE execute; refuses no-longer-data-bearing targets |
-| `selfupdate.Executor` / `selfupdate.Manager` | internal/selfupdate/{executor,commit}.go | `NewExecutor(Config)` / `NewManager(ManagerConfig)` | operator-signed agent self-update (D1): download+verify-vs-signed-sha → wrapper `apply`; startup dwell → `commit` | sha is the ONLY integrity root; wrapper (`felhom-selfupdate-guarded`) re-verifies as root + does the A/B flip; NEVER rolls back (systemd + wrapper do). `WrapperRunner` seam. Report seam `SelfUpdatePending()` |
+| `selfupdate.Executor` / `selfupdate.Manager` | internal/selfupdate/{executor,commit}.go | `NewExecutor(Config)` / `NewManager(ManagerConfig)` | operator-signed agent self-update (D1): download+verify-vs-signed-sha → wrapper `apply`; startup dwell → `commit` | sha is the ONLY integrity root; wrapper (`felhom-selfupdate-guarded`) re-verifies as root + does the A/B flip; NEVER rolls back (systemd + wrapper do). `WrapperRunner` seam. Report seam `SelfUpdatePending()`. **The staging folder cleans itself** (`prune.go`, R-1368): after a healthy `commit` only the running version's staged binary stays; a new file kept in `<state dir>/selfupdate/` must NOT be named `felhom-agent-<version>` or the prune removes it |
 | `Gate.AuthorizeStorageWipe` | internal/reconcile/gate.go | `AuthorizeStorageWipe(StorageWipeAuthz, *SignedOp) Decision` | tiered wipe authz | user-data ⇒ customer confirm bound to agent's DeviceDurableID; system/backup ⇒ operator signature only, `Confirmed` IGNORED by role |
 | `Gate.Authorize` | internal/reconcile/gate.go | `Authorize(Intent, *SignedOp) Decision` | every destructive intent | role-scoping (`roleAuthorizes`) + op-to-action binding; benign passes unsigned; audits every decision |
 | `storage.DeviceDurableID` / `ResolveDurableDevice` | internal/storage/durable_device.go | `DeviceDurableID(device) (string, error)` | WIPE-binding ids (`byid:`/`byuuid:`) | Single seam for /disks list AND gate (F20-BUG2); `ResolveDurableDevice` refuses bare paths |
@@ -62,6 +62,7 @@
 | `GuestBinder.EnsureSharedParent` | internal/localapi/intermediary.go | `EnsureSharedParent(ctx) error` | shared-parent bind + boot unit | make-private+make-shared ONLY on first bind — re-running orphans the guest's slave; F2-a: compares script AND unit for staleness |
 | `StablePathForRaw` / `DriveNameFromRaw` | internal/localapi/intermediary.go | `StablePathForRaw("/mnt/<n>") string` | raw host mount → stable guest path | Single source of truth both repos derive from |
 | `GuestBinder.GuestSeesMount` / `GuestBootID` | internal/localapi/intermediary.go | `GuestSeesMount(ctx, vmid, path) bool` | guest-visible (usable) signal; reboot detection | Host bind present ≠ guest sees it (non-recursive parent bind) |
+| `GuestBinder.GuestBootKey` (R-1395) | internal/localapi/intermediary.go | `GuestBootKey(ctx, vmid) string` | "is this a new boot of the guest?" — `<kernel boot_id>-<guest init start>` | **`GuestBootID` carries `/proc/stat` `btime`, which the kernel computes from the wall clock: a clock STEP moves it.** Use the key for any „same boot?” question; `""` = unreadable. Seams `hostBtimeFn` / `hostBootIDFn` / `procStarttimeFn` |
 | `SudoHostOps.EnsureNetworkMount` / `RemoveNetworkMount` / `ListNetworkMounts` | internal/storage/netmount.go | `EnsureNetworkMount(ctx, spec) error` | NAS automount pair | rm glob confined to `mnt-felhom*` units; NAS ≠ drive (no durable-id/SMART/wipe); RemoveNetworkMount doubles as the verify-fail rollback (idempotent) |
 | `NetworkMountedAt` / `NetworkEndpointReachable` | internal/storage/netmount.go | `NetworkMountedAt(where) bool` | verify mount-truth + the 2 s add pre-probe | /proc/mounts is the ONLY mount-success judge (autofs trigger ≠ mounted; readability ≠ mounted — SPIKE-nas-verify §8) |
 | `SudoHostOps.ReassertNetworkAutomounts` + `Server.ReassertNetworkMounts` + `guesthook.PostStartNetworkReassert` | internal/storage/netreassert.go, internal/localapi/netreassert.go, internal/guesthook/netreassert.go | `ReassertNetworkAutomounts(ctx) []NetReassertResult` | NAS guest-reboot heal (RCA fix 1): re-arm idle automount triggers (stop + enable --now) so the fresh mount event propagates into running guests | NEVER call from periodic health paths (an idle trigger is HEALTHY); active real mounts are never touched; hook leg runs as root (direct systemctl), daemon leg via sudo |
@@ -88,7 +89,8 @@
 | `EnsureLeaf` | internal/localapi/cert.go | `EnsureLeaf(certPath, keyPath, host) (cert, fingerprint, generated, err)` | pinned self-signed leaf | `generated=true` invalidates every issued bootstrap pin — log LOUD (B.1) |
 | `Server.RecoverStaleLockedGuests` | internal/localapi/stalelock.go | `RecoverStaleLockedGuests(ctx)` | startup stale vzdump-lock heal (F2-b) | Clears ONLY `backup`/`snapshot-delete`, only when no vzdump in-flight; A1 RESOLVED (v0.62.0): scan is pool-intersected (`ListLXC` ∩ `Client.Pool`), fail-safe skip on pool-read failure |
 | `ControllerSwapper.Swap` + `ValidControllerImage` | internal/localapi/controllerswap.go | `Swap(ctx, vmid, target) *ControllerSwapState` | agent-owned controller image swap + rollback | Strict image regex (repo + 3-part semver); state file written BEFORE swap; no-healthcheck images need `verifyDwell` |
-| `Server.ControllerSupervisorTick` + `ControllerParkedMarker` | internal/localapi/controllersupervisor.go | `ControllerSupervisorTick(ctx)` | R-523: restart a provisioned guest's not-running controller via its bootstrap unit | Two-sweep confirm; honours swapInFlight, the host-side park marker, guest lock + vzdump; 3 restarts/15 min → 30 min pause; record rides the report as `controller_supervisor` (the hub mints the events — the agent has no event channel) |
+| `Server.ControllerSupervisorTick` + `ControllerParkedMarker` | internal/localapi/controllersupervisor.go | `ControllerSupervisorTick(ctx)` | R-523: restart a provisioned guest's not-running controller via its bootstrap unit | Two-sweep confirm; honours swapInFlight, the host-side park marker, guest lock + vzdump; 3 restarts/15 min → 30 min pause (`supervisorBrakeAllows` — every supervisor restart passes it; `recordSupervisorRestart` puts one on the record + R-539's slow counter); R-1344 (2026-10-30): a RUNNING controller whose Docker health reads „unhealthy" is alarmed at 6 sweeps and restarted at 10 (`restartUnresponsive`: the controller's busy file, then the brake); record rides the report as `controller_supervisor` (the hub mints the events — the agent has no event channel) |
+| `Server.OnEnvelope` / `OperatorActionResults` (operator actions) | internal/localapi/opactions.go | `OnEnvelope(ctx, env)`, `OperatorActionResults() []hub.OperatorActionResult` | 2026-10-29: the operator's actions for THIS agent from the host-report reply — a CLOSED list (`restart_controller`, `reboot_guest` since D65 — `TestOpActions_AgentClosedList`), once per id, the result re-sent until the reply stops listing it, the report requested at once (`SetOperatorActionTrigger`). The restart is the supervisor's act with its guards; the 3-in-15-min brake counts operator restarts too (30-min refusal, `opBrakeUntil`). A new verb is an operator decision AND a hub `store.hostOperatorActionNames` entry in the same release |
 | `MemoryOps` + `Server.readMemoryBounds` | internal/localapi/guestmemory.go | `readMemoryBounds(ctx, vmid) (memoryBounds, err)` | guest RAM resize (v0.90.0, R-24): GET/POST /guest/memory | NEW narrow seam (never extend `GuestAPI` — it breaks every fake); the AGENT is the boundary — bounds recomputed FRESH per request (min 2048 / max host_total−2048 / shrink floor max(2048, usage+512)); §8 UNITS TRAP (config `memory`=MB, status/node=bytes); verify maxmem==target after `SetConfig` before claiming success; SetConfig NEVER called on a refusal path |
 
 ### Proxmox client / hub / PBS / provisioning
@@ -97,6 +99,10 @@
 |---|---|---|---|---|
 | `pvegate.Write` / `pvegate.Step` | internal/pvegate/pvegate.go | `Write(ctx) (release, waited, err)` / `Step(ctx) (end, err)` | R-812 option A: keep the agent's own /etc/pve writes out of a Proxmox package step (pmxcfs restarts) | Already wired at the two chokepoints — `Client.doBody` (every non-GET) and `ExecRunner.RunStdin` (`WritesEtcPVE`: pct config verbs, pvesm, pveum, felhom-pbs-apply create/reconcile). A new root CLI that writes /etc/pve goes into `WritesEtcPVE`, never its own lock. Never take `Step` around anything but the wrapper call (`Leg.runPVE`) — a `Write` inside a `Step` deadlocks until its context ends. |
 | `osupdate.KernelVerdict` / `Leg.KernelAfterBoot` | internal/osupdate/kernel.go | `KernelVerdict(before, after, tunnel, hubReached) (ok, why)` | R-836: THE one-shot-boot rule — the host health rule (`HostHealthVerdict`) AND the box reached the hub since the boot | The only judge of a new kernel. Never reboot the host from Go: every reboot is the wrapper's (`kernel-reboot`, `kernel-revert`), and only for a staged step. A new kernel-lane state lives in the wrapper's `/var/lib/felhom-kernel/state.json`, never in the agent's own files (the agent can write those). |
+| `Leg.runKernelCleanup` / `kernelCleanupSkip` / `Leg.KernelCleanupDryRun` + the wrapper's `Kernel.cleanup` | internal/osupdate/kernel.go; configs/felhom-os-apply (mode `kernel-cleanup`) | `kernelCleanupSkip(appliance, trigger, host, pve, kernel, blk, okStep) string` ("" = call it) | R-1368 (`11` §5.11.1): old kernel images are removed by the WRAPPER's rule; the agent decides only WHEN (the night leg, after a healthy host step, never in a night with a kernel step) | The plan names NO package (C1). Never `apt-get autoremove`. Whether a box may do it is the root-owned ring-0 mark, never the hub's ring. The hub report carries NO kernel view and no kernel version (the hub would read it as a kernel step). `--selftest=kernel-cleanup-dry-run` changes nothing. |
+| `Leg.ProfileRefresh` + the wrapper's `Apply.profile_refresh` | internal/osupdate/profilerefresh.go; configs/felhom-os-apply (mode `profile-refresh`) | `ProfileRefresh(ctx, why, gate, tries, wait) string` (the outcome) | R-1388 (`11` §5.12): load every AppArmor profile from a fresh compile once per kernel the box starts on; restart only a FAILED service whose own program's cached profile was stale | The plan names nothing (F1). A stale copy is told by the CACHED FILE changing at the rewrite — never by the kernel's loaded hash (it differs between a cache load and a fresh load for multi-profile files; `lxc-containers` never compiles the same twice). Called at agent start and after a bundle install, under the heavy-op gate. |
+| `pvetrust.Manager` (`Ensure`, `OnCallError`, `Status`) + `proxmox.Client.SetTLS` / `proxmox.TrustProblem` + `Leg.PVECACopy` + the wrapper's `pve_ca_copy` | internal/pvetrust/pvetrust.go; internal/proxmox/{client,trust}.go; internal/osupdate/pvecacopy.go; configs/felhom-os-apply (mode `pve-ca-copy`) | `Ensure(ctx, why) (changed bool)`; `TrustProblem(err) string` | R-1380 (felhom.eu `03` §12a): trust in Proxmox by a copy of the box's own root CA; the move from a pinned certificate; the cause of a trust failure | THE ORDER: the copy, a real call with the copy ALONE, then ONE config write that clears the fingerprint LAST (a set fingerprint wins over `ca_file` in `TLSConfig.build`). Never clear the fingerprint anywhere else. The wrapper's plan names nothing (T1). `agent.json` is written by the agent in place, with a parked recovery copy — the same way as the backup-target move. |
+| `config.ReplaceInPlace` / `config.RecoverCutWrite` (+ `config.FileMu`) | internal/config/cutwrite.go; internal/config/filelock.go | `ReplaceInPlace(path, prev, next) error`; `RecoverCutWrite(path) (note string, err error)` | R-1380 (felhom.eu `03` §12a item 7): the only way to write `agent.json` — the folder is root-owned, so no rename; the old bytes go to a journal in the state folder first, and the daemon repairs a cut write at start, before it reads the config | A new writer takes `FileMu`, calls `ReplaceInPlace`, opens no file by hand, and is added to the two source pins in `cmd/felhom-agent` (`…TakeTheSharedLock`, `…UseTheJournalledWrite`). A package whose tests write the config sets `config.CutWriteDir` in `TestMain` |
 | `Client.WaitTask` | internal/proxmox/task.go | `WaitTask(ctx, upid, opts) (TaskStatus, error)` | asserting EVERY mutating op | POST 200 ≠ success; authz can fail at task exec; `AllowWarnings` opt-in |
 | `Client.Pool` | internal/proxmox/query.go | `Pool(ctx, name) (PoolInfo, error)` | felhom-pool membership (the ownership registry, A1) | Needs `Pool.Audit` at `/pool/<name>` (host-install v1.9.0+); `Pool.Allocate` does NOT satisfy the read; members can be storages (type `storage`, vmid 0) — filter them |
 | `Client` mutate wrappers (`RestoreLXC/Vzdump/DestroyLXC/Snapshot/Rollback/SetConfig/ResizeLXC/Start/Stop`) | internal/proxmox/mutate.go | return `(upid, error)` | all API mutations | Async → always pair with WaitTask; route via gate/queue, not ad-hoc |

cmd/felhom-agent/config_cutwrite_test.go

@@ -0,0 +1,141 @@
+package main
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
+)
+
+// A box whose agent.json was cut in the middle of a write must start (R-1380): main repairs the file from the
+// journal BEFORE it reads it. The staging is the state a power cut leaves: a durable journal of the old bytes and a
+// half-written config.
+func TestMainRecoversACutConfigWrite(t *testing.T) {
+	dir := t.TempDir()
+	prev := config.CutWriteDir
+	config.CutWriteDir = filepath.Join(dir, "state")
+	t.Cleanup(func() { config.CutWriteDir = prev })
+	if err := os.MkdirAll(config.CutWriteDir, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	cfgPath := filepath.Join(dir, "agent.json")
+	old := `{"proxmox":{"node":"pve","tls":{"fingerprint":"AA:BB"}}}`
+	if err := os.WriteFile(filepath.Join(config.CutWriteDir, "agent.json.write-in-flight"), []byte(old), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(cfgPath, []byte(`{"proxmox":{"node":"pv`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := config.Load(cfgPath); err == nil {
+		t.Fatal("the staging is wrong: the cut config loads")
+	}
+	note, err := recoverCutConfigWrite(cfgPath)
+	if err != nil || note == "" {
+		t.Fatalf("note=%q err=%v", note, err)
+	}
+	cfg, err := config.Load(cfgPath)
+	if err != nil {
+		t.Fatalf("the agent does not start after the repair: %v", err)
+	}
+	if cfg.Proxmox.TLS.Fingerprint != "AA:BB" {
+		t.Fatalf("the repaired config lost the fingerprint: %+v", cfg.Proxmox.TLS)
+	}
+}
+
+// main must call the repair before config.Load, or the repair never runs on the box that needs it (a seam that is
+// built and not wired). Read from the source: the call's position in main's body.
+func TestMainCallsTheConfigRepairBeforeItLoadsTheConfig(t *testing.T) {
+	src, err := os.ReadFile("main.go")
+	if err != nil {
+		t.Fatal(err)
+	}
+	f, err := parser.ParseFile(token.NewFileSet(), "main.go", src, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var repair, load token.Pos
+	for _, d := range f.Decls {
+		fn, ok := d.(*ast.FuncDecl)
+		if !ok || fn.Name.Name != "main" || fn.Body == nil {
+			continue
+		}
+		ast.Inspect(fn.Body, func(n ast.Node) bool {
+			c, ok := n.(*ast.CallExpr)
+			if !ok {
+				return true
+			}
+			if id, ok := c.Fun.(*ast.Ident); ok && id.Name == "recoverCutConfigWrite" && repair == 0 {
+				repair = c.Pos()
+			}
+			if sel, ok := c.Fun.(*ast.SelectorExpr); ok && sel.Sel.Name == "Load" {
+				if pkg, ok := sel.X.(*ast.Ident); ok && pkg.Name == "config" && load == 0 {
+					load = c.Pos()
+				}
+			}
+			return true
+		})
+	}
+	if repair == 0 || load == 0 {
+		t.Fatalf("main: repair call found=%v, config.Load found=%v", repair != 0, load != 0)
+	}
+	if repair > load {
+		t.Fatal("main reads the config BEFORE it repairs a cut write — a cut config stops the agent at start")
+	}
+}
+
+// Every writer of agent.json goes through config.ReplaceInPlace and none truncates the file itself: a writer that
+// opens the config with O_TRUNC by hand is the unprotected write again.
+func TestAgentJSONWritersUseTheJournalledWrite(t *testing.T) {
+	for _, w := range []struct{ file, fn string }{
+		{"../../internal/pvetrust/pvetrust.go", "writeConfig"},
+		{"../../internal/pbsdr/manager.go", "seedEscrowStorageID"},
+		{"../../internal/localapi/backup_target.go", "setConfiguredBackupTarget"},
+	} {
+		src, err := os.ReadFile(w.file)
+		if err != nil {
+			t.Fatal(err)
+		}
+		fset := token.NewFileSet()
+		f, err := parser.ParseFile(fset, w.file, src, 0)
+		if err != nil {
+			t.Fatal(err)
+		}
+		found := false
+		for _, d := range f.Decls {
+			fn, ok := d.(*ast.FuncDecl)
+			if !ok || fn.Name.Name != w.fn || fn.Body == nil {
+				continue
+			}
+			found = true
+			body := string(src[fset.Position(fn.Body.Pos()).Offset:fset.Position(fn.Body.End()).Offset])
+			calls := 0
+			ast.Inspect(fn.Body, func(n ast.Node) bool {
+				if c, ok := n.(*ast.CallExpr); ok {
+					if sel, ok := c.Fun.(*ast.SelectorExpr); ok && sel.Sel.Name == "ReplaceInPlace" {
+						if pkg, ok := sel.X.(*ast.Ident); ok && pkg.Name == "config" {
+							calls++
+						}
+					}
+					if sel, ok := c.Fun.(*ast.SelectorExpr); ok && sel.Sel.Name == "OpenFile" {
+						t.Errorf("%s %s opens a file by hand — the config is written only by config.ReplaceInPlace", w.file, w.fn)
+					}
+				}
+				return true
+			})
+			if calls != 1 {
+				t.Errorf("%s %s: %d call(s) to config.ReplaceInPlace, want 1", w.file, w.fn, calls)
+			}
+			if strings.Contains(body, "O_TRUNC") {
+				t.Errorf("%s %s truncates a file itself", w.file, w.fn)
+			}
+		}
+		if !found {
+			t.Errorf("%s: function %s not found — the pin reads nothing", w.file, w.fn)
+		}
+	}
+}

cmd/felhom-agent/config_writers_lock_test.go

@@ -0,0 +1,62 @@
+package main
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"testing"
+)
+
+// Three functions in this process read agent.json whole, change one section and write it whole: the trust move, the
+// PBS seed and the backup-target move. Without one shared lock two of them interleave and the second puts back the
+// document it read before the first one's change (found by review, 2026-11-14). Each must take config.FileMu as its
+// FIRST statement and release it by defer — asserted from the source, because the race itself is not a thing a unit
+// test can stage reliably.
+func TestAgentJSONWritersTakeTheSharedLock(t *testing.T) {
+	for _, w := range []struct{ file, fn string }{
+		{"../../internal/pvetrust/pvetrust.go", "writeConfig"},
+		{"../../internal/pbsdr/manager.go", "seedEscrowStorageID"},
+		{"../../internal/localapi/backup_target.go", "setConfiguredBackupTarget"},
+	} {
+		f, err := parser.ParseFile(token.NewFileSet(), w.file, nil, 0)
+		if err != nil {
+			t.Fatalf("%s: %v", w.file, err)
+		}
+		found := false
+		for _, d := range f.Decls {
+			fn, ok := d.(*ast.FuncDecl)
+			if !ok || fn.Name.Name != w.fn || fn.Body == nil {
+				continue
+			}
+			found = true
+			isLockCall := func(e ast.Expr, name string) bool {
+				c, ok := e.(*ast.CallExpr)
+				if !ok {
+					return false
+				}
+				sel, ok := c.Fun.(*ast.SelectorExpr)
+				if !ok || sel.Sel.Name != name {
+					return false
+				}
+				inner, ok := sel.X.(*ast.SelectorExpr)
+				if !ok || inner.Sel.Name != "FileMu" {
+					return false
+				}
+				pkg, ok := inner.X.(*ast.Ident)
+				return ok && pkg.Name == "config"
+			}
+			if len(fn.Body.List) < 2 {
+				t.Errorf("%s %s: too short to hold the lock", w.file, w.fn)
+				continue
+			}
+			first, ok1 := fn.Body.List[0].(*ast.ExprStmt)
+			second, ok2 := fn.Body.List[1].(*ast.DeferStmt)
+			if !ok1 || !isLockCall(first.X, "Lock") || !ok2 || !isLockCall(second.Call, "Unlock") {
+				t.Errorf("%s %s: must start with config.FileMu.Lock() and defer config.FileMu.Unlock()", w.file, w.fn)
+			}
+		}
+		if !found {
+			t.Errorf("%s: function %s not found — a writer of agent.json was renamed; keep it under the shared lock and update this test", w.file, w.fn)
+		}
+	}
+}

cmd/felhom-agent/main.go

@@ -54,6 +54,7 @@ import (
 	"gitea.dooplex.hu/admin/felhom-agent/internal/poweroff"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/provision"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/pvetrust"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/reconcile"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/restorespace"
 	"gitea.dooplex.hu/admin/felhom-agent/internal/selfheal"
@@ -216,6 +217,15 @@ func main() {
 		return
 	}
 
+	// Only the daemon repairs: a selftest can run beside a live agent that is in the middle of a write.
+	var cutWriteNote string
+	var cutWriteErr error
+	if selftest.mode == "" {
+		cutWriteNote, cutWriteErr = recoverCutConfigWrite(cfgPath)
+		if cutWriteErr != nil {
+			fmt.Fprintln(os.Stderr, "config repair:", cutWriteErr)
+		}
+	}
 	cfg, err := config.Load(cfgPath)
 	if err != nil {
 		// A missing default config file is fine if env provides the values; only a
@@ -229,6 +239,11 @@ func main() {
 	// logRing is the always-DEBUG capture ring (v0.83.0 observability): served by the
 	// local API's GET /debug/logs and the heartbeat log-pull. Selftests ignore it.
 	logger, logRing := applog.New(cfg.LogLevel)
+	if cutWriteErr != nil {
+		logger.Error("config: the last write of the config was cut and could not be repaired", "err", cutWriteErr)
+	} else if cutWriteNote != "" {
+		logger.Warn("config: a write of the config was cut by a stop — repaired at start", "result", cutWriteNote)
+	}
 
 	switch selftest.mode {
 	case "":
@@ -251,6 +266,8 @@ func main() {
 		os.Exit(runSelftestOSUpdate(context.Background(), cfg, logger, vmid))
 	case "os-facts":
 		os.Exit(runSelftestFacts(context.Background(), cfg, logger, vmid))
+	case "kernel-cleanup-dry-run":
+		os.Exit(runSelftestKernelCleanupDryRun(context.Background(), cfg, logger, vmid))
 	case "live-restore":
 		os.Exit(runSelftestLiveRestore(context.Background(), cfg, logger, vmid))
 	case "pbs-verify":
@@ -521,6 +538,66 @@ type storeGrantRepairer struct {
 	mu       sync.Mutex
 	last     map[string]time.Time // target id → last ATTEMPT (success or failure)
 	repaired map[string]time.Time // target id → last CONFIRMED repair (drives the report latch)
+	// seenPath (R-1302) is the file that remembers which tiers have EVER held the grant: only those can LOSE it. "" = no
+	// memory, so every missing grant reads as a loss — the old, loud behaviour.
+	seenPath string
+}
+
+// everHeld reports whether this tier's grant was ever seen on this box (R-1302). The direction of a doubt is „yes": an
+// unreadable or corrupt file keeps the loss ERROR, so a real loss is never quietened by a broken record.
+func (r *storeGrantRepairer) everHeld(target string) bool {
+	if r == nil || r.seenPath == "" {
+		return true
+	}
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	seen, err := readGrantSeen(r.seenPath)
+	return err != nil || seen[target]
+}
+
+// noteHeld records that this tier holds the grant. A write that fails is logged and costs one false „first grant" at most
+// after a restart — the old file, or none, stands.
+func (r *storeGrantRepairer) noteHeld(target string) {
+	if r == nil || r.seenPath == "" {
+		return
+	}
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	seen, err := readGrantSeen(r.seenPath)
+	if err != nil {
+		seen = map[string]bool{} // a corrupt file is replaced by what is true now
+	}
+	if seen[target] {
+		return
+	}
+	seen[target] = true
+	b, _ := json.Marshal(seen)
+	tmp := r.seenPath + ".tmp"
+	if werr := os.WriteFile(tmp, b, 0o600); werr == nil {
+		werr = os.Rename(tmp, r.seenPath)
+		if werr == nil {
+			return
+		}
+		err = werr
+	} else {
+		err = werr
+	}
+	repairLogger(r).Warn("store-grant: could not record that this tier holds its grant (R-1302)", "target", target, "err", err)
+}
+
+func readGrantSeen(path string) (map[string]bool, error) {
+	b, err := os.ReadFile(path)
+	if os.IsNotExist(err) {
+		return map[string]bool{}, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	seen := map[string]bool{}
+	if err := json.Unmarshal(b, &seen); err != nil {
+		return nil, err
+	}
+	return seen, nil
 }
 
 // noteRepaired latches a confirmed repair so it is reported for storeGrantRepairReportWindow.
@@ -631,6 +708,7 @@ func storeGrantStatus(ctx context.Context, px *proxmox.Client, targetID string,
 		// host-report has certainly carried it. Without this latch the repairing probe may be a
 		// log-only one and the hub never learns anything happened (measured live, see the window's
 		// comment).
+		repair.noteHeld(targetID) // R-1302: a tier that holds its grant can lose it from now on
 		return storeGrantHealthyVerdict(targetID, critical, s, repair.recentlyRepaired(targetID, time.Now()))
 	}
 
@@ -679,8 +757,25 @@ func storeGrantStatus(ctx context.Context, px *proxmox.Client, targetID string,
 	// produces exactly one alert pair and the operator learns of it. NOTHING NEW WAS BUILT: no wire
 	// change, no hub change, no new event type. The `Feature` text carries the explanation because
 	// that is the field the hub puts in the operator's e-mail (the Reason does not travel).
+	return storeGrantAfterRepair(targetID, critical, repair, s)
+}
+
+// storeGrantAfterRepair is the verdict once a repair is confirmed (split out for the tests). R-1302: the FIRST grant of a
+// tier that never held one is no loss — on a fresh install the whole-guest tier is armed before its storage exists, so
+// every new box logged „investigate the loss" and a DEGRADED capability at its first start, which teaches the reader to
+// skip the line. That grant logs INFO and reports healthy. A tier seen holding its grant, or one this box cannot tell
+// about, keeps R-190's ERROR and its one degraded cycle. Pinned by TestR1302_*.
+func storeGrantAfterRepair(targetID string, critical bool, repair *storeGrantRepairer, healthy capability.Status) capability.Status {
+	if !repair.everHeld(targetID) {
+		repair.noteHeld(targetID)
+		repairLogger(repair).Info("store-grant: granted — the first grant of this tier, nothing was lost (R-1302)",
+			"target", targetID, "privilege", storeGrantRequiredPriv,
+			"action", "felhom-backup-target-apply grant "+targetID, "confirmed_by", "re-read")
+		healthy.Status, healthy.Reason = capability.StatusOK, ""
+		return healthy
+	}
 	repair.noteRepaired(targetID, time.Now())
-	s = storeGrantRepairedVerdict(targetID, critical)
+	s := storeGrantRepairedVerdict(targetID, critical)
 	repairLogger(repair).Error("store-grant: GRANT WAS MISSING AND HAS BEEN SELF-REPAIRED — investigate the loss (R-190)",
 		"target", targetID, "privilege", storeGrantRequiredPriv,
 		"action", "felhom-backup-target-apply grant "+targetID, "confirmed_by", "re-read")
@@ -853,8 +948,9 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 	// The runner is the DIRECT one for the same reason the sudo prober uses it: the wrapper is
 	// invoked through the privileged path, which prepends sudo itself.
 	grantRepairer := &storeGrantRepairer{
-		run: (&proxmox.ExecRunner{Mode: proxmox.RunnerMode(cfg.Privileged.Mode)}).Run,
-		log: logger,
+		run:      (&proxmox.ExecRunner{Mode: proxmox.RunnerMode(cfg.Privileged.Mode)}).Run,
+		log:      logger,
+		seenPath: filepath.Join(agentStateDir(cfg), "store-grant-seen.json"), // R-1302
 	}
 	// R-973: the local-API backup cadence cap's refusals ride the report as a CRITICAL status, so the hub's
 	// operator-only whole_guest_backup_refused event (hub v0.159.0, R-1150) tells the operator. Late-bound like
@@ -907,6 +1003,13 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 	osLeg := newOSLeg(cfg, client, px, logger)
 	collector.SetSystemReporter(&factsReporter{leg: osLeg, guest: firstGuest(px)}) // R-852: the versions
 	desiredSyncer.AddConsumer(osLeg)
+	// R-1380 (D80, felhom.eu `03` §12a): trust in Proxmox by a copy of the box's own root CA. At this start the root
+	// step makes the copy, a REAL call proves it while the pinned fingerprint is still the trust, and only then ONE
+	// write of the config sets ca_file and clears the fingerprint. A call that fails on trust later names the cause and
+	// runs the same order again (a pinned box whose certificate changed moves right then, with no restart).
+	pveTrust := newPVETrust(cfg, px, osLeg, logger)
+	collector.SetProxmoxTrustSource(func() any { return pveTrust.Status() })
+	loop.SetCollectErrorHook(pveTrust.OnCallError)
 	// S5: consume a host_loss restore_directive into an inspectable restore PLAN (derive + surface,
 	// execute nothing). The recipe is fetched on-demand (rare directive) via a fresh Collect.
 	desiredSyncer.AddConsumer(dr.NewConsumer(func(ctx context.Context) *hub.DRRecipeHostHalf {
@@ -937,6 +1040,10 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 	// network and stays under the hub's 45-minute host_stale (its
 	// alerting.stale_threshold; host_down at 90). On a box without the kernel lane (an older wrapper, a BYO host) the check is refused and logged.
 	go osLeg.KernelAfterBoot(ctx, 0, osupdate.KernelJudge{Wait: osupdate.DefaultKernelJudgeWait})
+	// R-1380: the trust order at this start (the copy, the proof, then the config) — before the first report cycle, so
+	// a pinned box whose certificate changed while the agent was down is not blind for a cycle. Then on a timer.
+	pveTrust.Ensure(ctx, "start")
+	go pveTrust.Run(ctx, time.Hour, 24*time.Hour)
 	// R-1192: every guest may swap as much as its RAM — raised once at start (a live change, no restart). A minute
 	// after start so the guest has settled; a failure is logged and the next start retries.
 	go func() {
@@ -1205,6 +1312,8 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 			}
 			return release, nil
 		}}
+	daemonCtx := ctx
+	profileGate := func() (func(), string, bool) { return heavyOps.TryAcquire("os-profile-refresh") }
 	// Agent v0.143.0 (R-840): the config bundle — the box's root-owned files by a signed job; the wrapper verifies it.
 	bundleExec := osupdate.ConfigUpdateExecutor{Leg: osLeg, URLTemplate: suCfg.URLTemplate, Username: suCfg.Username, Token: suCfg.Token,
 		// The capability probe confirms from the agent's side: `sudo -l` lists every command the new sudoers grants.
@@ -1215,7 +1324,17 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 				names = append(names, d.Name)
 			}
 			logger.Warn("osupdate: capability probe after the config bundle", "ok", ok, "total", total, "degraded", strings.Join(names, ","))
+			// R-1388: a new wrapper may know the profile refresh the old one refused — ask again (its own goroutine:
+			// it may wait for the heavy-op gate, and the signed job must not wait with it).
+			go osLeg.ProfileRefresh(daemonCtx, "config bundle installed", profileGate, 12, 5*time.Minute)
+			// R-1380: a new wrapper may know the CA-copy step the old one refused — run the trust order again.
+			go pveTrust.Ensure(daemonCtx, "config bundle installed")
 		}}
+	// R-1388 / R-1301 (`11` §5.12): after a start, load the security profiles from a fresh compile when the box runs
+	// another kernel than at the last refresh (the wrapper keeps that record and answers "not needed" otherwise).
+	// R-1383 (release 3): the last refresh, and how this process's last ask ended, ride the host report.
+	collector.SetProfileRefreshSource(osLeg.ProfileRefreshStanza)
+	go osLeg.ProfileRefresh(daemonCtx, "agent start", profileGate, 12, 5*time.Minute)
 	// R-1285: the NEW agent runs the bundle policy pass once after its self-update commits — in the normal two-job
 	// delivery the bundle was installed by the OLD process, which has no such pass (measured 2026-10-26).
 	wireBundlePolicyAfterCommit(selfUpdateMgr, bundleExec, logger, version)
@@ -1267,6 +1386,21 @@ func runDaemon(cfg config.Config, logger *slog.Logger, logRing *applog.Ring) int
 	}
 	localSrv := buildLocalAPIServer(cfg, px, backupStore, heavyOps, observer, driveKnown, hostOps, gate, collector, client, intentRec, guestBindStore, formatJobStore, logRing, escrowCeremonyCfg, logger, &localTokens)
 	if localSrv != nil {
+		// R-1383 (D78): the guest-power watchdog's give-up rides the host report (it was one journal line).
+		collector.SetGuestPowerGaveUpSource(localSrv.GuestPowerGaveUpStanza)
+	}
+	if localSrv != nil {
+		// 2026-10-29 (operator brief, Part 1): the operator's actions for THIS agent ride the host report's reply — the
+		// closed list (localapi/opactions.go). The server joins the observer chain; its results ride the next report,
+		// which it requests at once through the same cap-1 trigger the storage watchdog uses.
+		localSrv.SetOperatorActionTrigger(func() {
+			select {
+			case storageTrigger <- struct{}{}:
+			default:
+			}
+		})
+		collector.SetOperatorActionResults(localSrv.OperatorActionResults)
+		loop.SetEnvelopeObserver(hub.MultiObserver(desiredSyncer, jobsRunner, localSrv))
 		backupCapRefusal = localSrv.BackupCapRefusal // R-973
 		internalPlaceExec.SetEnsure(localSrv.EnsureInternalPlace)
 		// R-998 (doc 14): POST /host/poweroff — guest shutdown, then host power-off; guards in hostpoweroff.go.
@@ -3838,14 +3972,39 @@ func (f *selftestFlag) Set(v string) error {
 		f.mode = "os-update"
 	case "os-facts", "live-restore": // agent v0.142.0
 		f.mode = v
+	case "kernel-cleanup-dry-run": // R-1368: what the clean-up of old kernels WOULD do; changes nothing
+		f.mode = v
 	case "wgtunnel": // dispatched since S3 but refused here until 2026-10-04 (TestSelftestFlag_AcceptsEveryDispatchedMode)
 		f.mode = "wgtunnel"
 	default:
-		return fmt.Errorf("invalid --selftest value %q (want read|task|hub|storage|backup|restore-test|restore-test-due|pbs-verify|lanresolver|wgtunnel|bring-up|provision|escrow-create|escrow-consume|identity-consume|controller-swap|os-update|os-facts|live-restore)", v)
+		return fmt.Errorf("invalid --selftest value %q (want read|task|hub|storage|backup|restore-test|restore-test-due|pbs-verify|lanresolver|wgtunnel|kernel-cleanup-dry-run|bring-up|provision|escrow-create|escrow-consume|identity-consume|controller-swap|os-update|os-facts|live-restore)", v)
 	}
 	return nil
 }
 
+// newPVETrust wires internal/pvetrust (R-1380): the root step through the OS wrapper, the proof as a throwaway client
+// with EXACTLY the trust under test, and the running client's SetTLS.
+func newPVETrust(cfg config.Config, px *proxmox.Client, leg *osupdate.Leg, logger *slog.Logger) *pvetrust.Manager {
+	m := &pvetrust.Manager{ConfigPath: cfg.SourcePath, StateDir: agentStateDir(cfg), Logger: logger,
+		EnsureCopy: func(ctx context.Context) (pvetrust.Copy, error) {
+			r, err := leg.PVECACopy(ctx)
+			return pvetrust.Copy{Path: r.Path, Outcome: r.Outcome, NotAfter: r.NotAfter, DaysLeft: r.DaysLeft}, err
+		},
+		Probe: func(ctx context.Context, t proxmox.TLSConfig) error {
+			p, err := proxmox.NewClient(proxmox.Config{Endpoint: cfg.Proxmox.Endpoint, Node: cfg.Proxmox.Node, Token: cfg.Proxmox.Token, TLS: t})
+			if err != nil {
+				return err
+			}
+			_, err = p.Version(ctx)
+			return err
+		},
+		Apply: px.SetTLS,
+	}
+	m.Start(proxmox.TLSConfig{CAFile: cfg.Proxmox.TLS.CAFile, Fingerprint: cfg.Proxmox.TLS.Fingerprint,
+		InsecureSkipVerify: cfg.Proxmox.TLS.InsecureSkipVerify})
+	return m
+}
+
 // newOSLeg builds the OS-update leg (agent v0.140.0). The wrapper runs through sudo (FELHOM_OSAPPLY); the plan and
 // the once-per-night marker live in the agent's own os/ dir.
 func newOSLeg(cfg config.Config, client *hub.Client, px *proxmox.Client, logger *slog.Logger) *osupdate.Leg {
@@ -3962,6 +4121,24 @@ func runSelftestFacts(ctx context.Context, cfg config.Config, logger *slog.Logge
 	return 0
 }
 
+// runSelftestKernelCleanupDryRun asks the root wrapper what the clean-up of old kernels WOULD do on this box (R-1368,
+// `11` §5.11.1): the keep set, the removal set, and the refusal or "not run" reason. It changes nothing.
+//
+//	sudo -u felhom-agent felhom-agent --config … --selftest=kernel-cleanup-dry-run
+func runSelftestKernelCleanupDryRun(ctx context.Context, cfg config.Config, logger *slog.Logger, vmid int) int {
+	px, _ := newProxmoxClient(cfg)
+	leg := newOSLeg(cfg, nil, px, logger)
+	raw, err := leg.KernelCleanupDryRun(ctx, vmid)
+	if err != nil {
+		fmt.Fprintln(os.Stderr, "selftest=kernel-cleanup-dry-run:", err)
+		return 1
+	}
+	var v any
+	_ = json.Unmarshal(raw, &v)
+	printJSON("kernel clean-up — DRY RUN (nothing was changed)", v)
+	return 0
+}
+
 // runSelftestLiveRestore is the ONE-TIME live-restore step (`09` decision 87) as a debug action — the night leg does
 // the same before a ring-0 Docker step. It prints the container ids before and after (they must not change).
 //
@@ -4083,3 +4260,13 @@ func agentStateDir(cfg config.Config) string {
 func wireBundlePolicyAfterCommit(m *selfupdate.Manager, e osupdate.ConfigUpdateExecutor, lg *slog.Logger, ver string) {
 	m.SetOnCommitted(func(ctx context.Context) { e.PolicyPassAfterUpdate(ctx, lg, ver) })
 }
+
+// recoverCutConfigWrite repairs agent.json BEFORE it is read, when the last write of it was cut by a power loss
+// (config.RecoverCutWrite; R-1380). It runs before the logger exists, so main logs what it returns.
+// Pinned by TestMainRecoversACutConfigWrite.
+func recoverCutConfigWrite(cfgPath string) (string, error) {
+	if cfgPath == "" {
+		return "", nil
+	}
+	return config.RecoverCutWrite(cfgPath)
+}

cmd/felhom-agent/opactions_wiring_test.go

@@ -0,0 +1,51 @@
+package main
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"testing"
+)
+
+// 2026-10-29 (operator brief, Part 1): the agent's operator actions are unit-green in localapi; main.go must also put
+// the server in the report loop's observer chain, give the collector its results and give the server the report
+// trigger — an inert seam otherwise (the „seam built but never wired" class). Asserted from the AST.
+// RED-PROOF: seen failing before the wiring existed (all three false).
+func TestMainWiresTheAgentOperatorActions(t *testing.T) {
+	fset := token.NewFileSet()
+	f, err := parser.ParseFile(fset, "main.go", nil, 0)
+	if err != nil {
+		t.Fatalf("parse main.go: %v", err)
+	}
+	var observed, results, trigger bool
+	ast.Inspect(f, func(n ast.Node) bool {
+		call, ok := n.(*ast.CallExpr)
+		if !ok {
+			return true
+		}
+		fn, ok := call.Fun.(*ast.SelectorExpr)
+		if !ok {
+			return true
+		}
+		switch fn.Sel.Name {
+		case "MultiObserver":
+			for _, a := range call.Args {
+				if id, ok := a.(*ast.Ident); ok && id.Name == "localSrv" {
+					observed = true
+				}
+			}
+		case "SetOperatorActionResults":
+			if len(call.Args) == 1 {
+				if sel, ok := call.Args[0].(*ast.SelectorExpr); ok && sel.Sel.Name == "OperatorActionResults" {
+					results = true
+				}
+			}
+		case "SetOperatorActionTrigger":
+			trigger = true
+		}
+		return true
+	})
+	if !observed || !results || !trigger {
+		t.Fatalf("operator actions wiring: observer=%v results=%v trigger=%v — all three must be wired in main.go", observed, results, trigger)
+	}
+}

cmd/felhom-agent/pvetrust_wiring_test.go

@@ -0,0 +1,71 @@
+package main
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"testing"
+)
+
+// R-1380: the trust move is a component with seams (the root step, the proof, the running client, the report stanza,
+// the failed-call hook). A unit-green pvetrust that main.go never starts is the inert-seam defect, so the wiring is
+// asserted from the AST: built once, run at start and after a bundle install, on a timer, fed by a failed collect, and
+// reported. The same for the profile refresh (R-1388), which had no such test.
+func TestMainWiresPVETrust(t *testing.T) {
+	fset := token.NewFileSet()
+	f, err := parser.ParseFile(fset, "main.go", nil, 0)
+	if err != nil {
+		t.Fatalf("parse main.go: %v", err)
+	}
+	seen := map[string]int{}
+	ast.Inspect(f, func(n ast.Node) bool {
+		switch node := n.(type) {
+		case *ast.CallExpr:
+			sel, ok := node.Fun.(*ast.SelectorExpr)
+			if !ok {
+				return true
+			}
+			x, _ := sel.X.(*ast.Ident)
+			switch {
+			case x != nil && x.Name == "pveTrust" && sel.Sel.Name == "Ensure":
+				seen["ensure"]++
+			case sel.Sel.Name == "SetProxmoxTrustSource":
+				seen["stanza"]++
+			case sel.Sel.Name == "SetProfileRefreshSource" && len(node.Args) == 1:
+				if a, ok := node.Args[0].(*ast.SelectorExpr); ok && a.Sel.Name == "ProfileRefreshStanza" {
+					seen["profile refresh → report"]++ // R-1383
+				}
+			case sel.Sel.Name == "SetGuestPowerGaveUpSource" && len(node.Args) == 1:
+				if a, ok := node.Args[0].(*ast.SelectorExpr); ok && a.Sel.Name == "GuestPowerGaveUpStanza" {
+					seen["guest-power give-up → report"]++ // R-1383
+				}
+			case sel.Sel.Name == "SetCollectErrorHook" && len(node.Args) == 1:
+				if a, ok := node.Args[0].(*ast.SelectorExpr); ok && a.Sel.Name == "OnCallError" {
+					seen["hook"]++
+				}
+			case x != nil && x.Name == "osLeg" && sel.Sel.Name == "ProfileRefresh":
+				seen["profile-refresh"]++
+			}
+		case *ast.GoStmt:
+			if sel, ok := node.Call.Fun.(*ast.SelectorExpr); ok && sel.Sel.Name == "Run" {
+				if id, ok := sel.X.(*ast.Ident); ok && id.Name == "pveTrust" {
+					seen["timer"]++
+				}
+			}
+		case *ast.KeyValueExpr:
+			if k, ok := node.Key.(*ast.Ident); ok && k.Name == "Apply" {
+				if v, ok := node.Value.(*ast.SelectorExpr); ok && v.Sel.Name == "SetTLS" {
+					seen["apply=SetTLS"]++
+				}
+			}
+		}
+		return true
+	})
+	want := map[string]int{"ensure": 2, "stanza": 1, "hook": 1, "timer": 1, "apply=SetTLS": 1, "profile-refresh": 2,
+		"guest-power give-up → report": 1, "profile refresh → report": 1}
+	for k, n := range want {
+		if seen[k] < n {
+			t.Errorf("main.go wires %q %d time(s), want at least %d (the two-call ones: at start and after a bundle install)", k, seen[k], n)
+		}
+	}
+}

cmd/felhom-agent/storegrant_test.go

@@ -1,11 +1,14 @@
 package main
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"go/ast"
 	"io"
 	"log/slog"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 	"time"
@@ -415,3 +418,101 @@ func TestGrantRepair_TransitionSurvivesALaterProbe(t *testing.T) {
 			"be missed entirely", storeGrantRepairReportWindow)
 	}
 }
+
+// ── R-1302 — the first grant of a tier is no loss ──────────────────────────────────────
+//
+// A fresh install arms the whole-guest tier before its storage exists, so its first probe found no grant, repaired it and
+// logged „GRANT WAS MISSING … investigate the loss" with a DEGRADED capability — on every new box (VM 9375, 2026-10-26).
+//
+// RED-PROOF (2026-11-02): seen failing before storeGrantAfterRepair existed (build), and with its first-grant branch removed
+// TestR1302_TheFirstGrantIsNoLoss fails on „a first grant must not report a loss".
+func r1302Repairer(t *testing.T, path string) (*storeGrantRepairer, *bytes.Buffer) {
+	t.Helper()
+	var buf bytes.Buffer
+	return &storeGrantRepairer{run: (&fakeRepairRunner{}).Run, seenPath: path,
+		log: slog.New(slog.NewTextHandler(&buf, nil))}, &buf
+}
+
+func TestR1302_TheFirstGrantIsNoLoss(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "store-grant-seen.json")
+	r, buf := r1302Repairer(t, path)
+	healthy := capability.Status{Name: "pve:store-grant:felhom-pbs", Status: capability.StatusDegraded, Reason: "missing", Critical: true}
+	s := storeGrantAfterRepair("felhom-pbs", true, r, healthy)
+	if s.Status != capability.StatusOK || s.Reason != "" {
+		t.Fatalf("a first grant must not report a loss: status %q reason %q", s.Status, s.Reason)
+	}
+	got := buf.String()
+	if strings.Contains(got, "investigate the loss") || strings.Contains(got, "level=ERROR") {
+		t.Fatalf("a first grant must not log the loss ERROR: %s", got)
+	}
+	if !strings.Contains(got, "level=INFO") || !strings.Contains(got, "granted") || !strings.Contains(got, "felhom-pbs") {
+		t.Fatalf("a first grant says so at INFO: %s", got)
+	}
+	if r.recentlyRepaired("felhom-pbs", time.Now()) {
+		t.Fatal("a first grant must not arm the loss report's latch")
+	}
+}
+
+func TestR1302_AGrantThatExistedAndVanishedKeepsTheError(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "store-grant-seen.json")
+	first, _ := r1302Repairer(t, path)
+	first.noteHeld("felhom-pbs") // a healthy probe, in an earlier run of the agent
+
+	r, buf := r1302Repairer(t, path) // the agent restarted: the memory is the file, not the process
+	s := storeGrantAfterRepair("felhom-pbs", true, r, capability.Status{})
+	if s.Status != capability.StatusDegraded || !strings.Contains(s.Feature, "R-190") {
+		t.Fatalf("a grant that existed and vanished must keep its degraded cycle; got %q / %q", s.Status, s.Feature)
+	}
+	if got := buf.String(); !strings.Contains(got, "level=ERROR") || !strings.Contains(got, "investigate the loss") {
+		t.Fatalf("a grant that existed and vanished must keep its ERROR: %s", got)
+	}
+	if !r.recentlyRepaired("felhom-pbs", time.Now()) {
+		t.Fatal("the loss must arm the report's latch, or the hub may never hear it")
+	}
+}
+
+// After a first grant the tier counts as held: the NEXT disappearance is a loss.
+func TestR1302_AfterTheFirstGrantALossIsALoss(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "store-grant-seen.json")
+	r, _ := r1302Repairer(t, path)
+	storeGrantAfterRepair("felhom-pbs", true, r, capability.Status{})
+	s := storeGrantAfterRepair("felhom-pbs", true, r, capability.Status{})
+	if s.Status != capability.StatusDegraded {
+		t.Fatalf("the second disappearance is a loss; got %q", s.Status)
+	}
+}
+
+// The direction of a doubt: a broken record, a repairer with no record, or no repairer at all keep the loud way.
+func TestR1302_ADoubtKeepsTheError(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "store-grant-seen.json")
+	if err := os.WriteFile(path, []byte("{not json"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	r, _ := r1302Repairer(t, path)
+	if s := storeGrantAfterRepair("felhom-pbs", true, r, capability.Status{}); s.Status != capability.StatusDegraded {
+		t.Fatalf("a corrupt record must not quieten a loss; got %q", s.Status)
+	}
+	noMemory, _ := r1302Repairer(t, "")
+	if s := storeGrantAfterRepair("felhom-pbs", true, noMemory, capability.Status{}); s.Status != capability.StatusDegraded {
+		t.Fatalf("no record must not quieten a loss; got %q", s.Status)
+	}
+	if s := storeGrantAfterRepair("felhom-pbs", true, nil, capability.Status{}); s.Status != capability.StatusDegraded {
+		t.Fatalf("no repairer must not quieten a loss; got %q", s.Status)
+	}
+}
+
+// The wiring, read from the source: production names the file, and a healthy probe records the tier.
+func TestR1302_TheMemoryIsWired(t *testing.T) {
+	b, err := os.ReadFile("main.go")
+	if err != nil {
+		t.Fatal(err)
+	}
+	src := string(b)
+	if !strings.Contains(src, `seenPath: filepath.Join(agentStateDir(cfg), "store-grant-seen.json")`) {
+		t.Fatal("main.go must give the repairer its file, or every first grant reads as a loss")
+	}
+	start := strings.Index(src, "func storeGrantStatus(")
+	if start < 0 || !strings.Contains(src[start:start+3000], "repair.noteHeld(targetID)") {
+		t.Fatal("a healthy probe must record the tier, or a real loss would read as a first grant")
+	}
+}

configs/felhom-agent.sudoers

@@ -139,6 +139,8 @@ Cmnd_Alias FELHOM_INTERMEDIARY = \
 #   docker inspect -f *         — container running/health/image (read-only; `*` spans the -f template
 #                                 + container across spaces, spike-confirmed)
 #   systemctl restart <fixed unit> — re-run the golden's bootstrap (the only state change)
+#   cat <the controller's busy file> — D65 (2026-10-29): before an operator reboot of the guest, is the controller
+#                                 running a backup, a restore, an off-site run or an app update? (read-only, one path)
 #   felhom-priv-apply controller-image <vmid> — WRITE the ref (R-861 (a) A1, `09` §3 decision 165): the ref goes on
 #                                 STDIN to the ROOT wrapper, which requires our registry + repository + an x.y.z tag
 #                                 and writes the guest file itself. The agent's own `tee` grant is GONE: before, a
@@ -149,6 +151,7 @@ Cmnd_Alias FELHOM_CONTROLLERSWAP = \
   /usr/sbin/pct ^exec [0-9]+ -- docker image inspect gitea\.dooplex\.hu/admin/felhom-controller\:[0-9]+\.[0-9]+\.[0-9]+$, \
   /usr/sbin/pct ^exec [0-9]+ -- docker inspect -f .+ (felhom-controller|cloudflared)$, \
   /usr/sbin/pct ^exec [0-9]+ -- systemctl restart felhom-controller-bootstrap\.service$, \
+  /usr/sbin/pct ^exec [0-9]+ -- cat /var/lib/docker/volumes/felhom-controller-data/_data/data/ctl/busy$, \
   /usr/local/sbin/felhom-priv-apply ^controller-image [0-9]+$
 
 # Stale-lock recovery (F2-b, v0.49.0). A host reboot DURING a vzdump backup leaves the guest with a

configs/felhom-os-apply

@@ -97,9 +97,12 @@ OOMCHECK_SETTLE = 2
 # so it belongs with them. Measured 2026-10-09 on demo-hp (Secure Boot on): left in the pve lane, its upgrade pulled
 # shim-signed-common, the whole pve step was refused R6, and the night's kernel step was skipped with it
 # (`audits/kernel-night-2026-10-08/`). Pinned by test_ring0_pending_pve_leaves_the_secure_boot_meta_out.
+# proxmox-grub is GRUB's Proxmox meta-package (it depends on grub-efi-amd64 of the same version); the `grub` above is
+# anchored at the start, so it missed it. Measured 2026-11-05..07 on demo-felhom: its upgrade pulled grub-efi-amd64 and
+# the whole pve step was refused R6 every night. Pinned by test_ring0_pending_pve_leaves_proxmox_grub_out.
 HOST_SLOW_RE = re.compile(r"^(linux-(image|headers|kbuild|modules|base)|proxmox-kernel|proxmox-default-kernel|pve-kernel|"
                           r"pve-firmware|firmware-|grub|shim|systemd-boot|intel-microcode|amd64-microcode|efibootmgr|"
-                          r"proxmox-secure-boot-support)")
+                          r"proxmox-secure-boot-support|proxmox-grub)")
 # restart_needed() leaves out processes whose cgroup line matches (grep basic regex). Host: the LXC guests' own
 # processes (`0::/lxc/<vmid>/...`) -- NOT lxc-start itself, whose cgroup is `0::/lxc.monitor/<vmid>` (measured
 # 2026-10-04 on demo-felhom: the old pattern "lxc" hid lxc-start with 20 deleted maps, so "reboot needed" stayed false
@@ -152,9 +155,36 @@ KERNEL_IMAGE_RE = re.compile(r"^proxmox-kernel-([0-9]+\.[0-9]+\.[0-9]+-[0-9]+-pv
 # helper and the firmware the kernel loads. Everything else in HOST_SLOW_RE (grub, shim, microcode, efibootmgr) stays out.
 KERNEL_UPGRADE_RE = re.compile(r"^(proxmox-default-kernel|proxmox-kernel-[0-9]+\.[0-9]+|proxmox-kernel-helper|pve-firmware)$")
 KERNEL_MIN_GAP = 20 * 3600     # never two kernel steps in one night
-KERNEL_MODES = ("kernel-status", "kernel-reboot", "kernel-boot", "kernel-good", "kernel-revert", "kernel-cancel")
+KERNEL_MODES = ("kernel-status", "kernel-reboot", "kernel-boot", "kernel-good", "kernel-revert", "kernel-cancel",
+                "kernel-cleanup")
 # phases: staged → oneshot → judging → good | reverting → self_reverted | revert_failed; oneshot → fell_back; staged → cancelled
 KERNEL_ACTIVE = ("staged", "oneshot", "judging", "reverting")
+# The clean-up of old kernels (R-1368, `09` §3 decision 286, `11` §5.11.1). A step that ENDED in one of these phases
+# and is not "good" keeps every kernel (the box fell back or reverted); any other phase is "not ended" (P1).
+# Trust in Proxmox by a copy of the box's own root CA (R-1380, `09` §3 decision 303, felhom.eu `03` §12a). The agent's
+# user cannot read Proxmox's CA file (0640 root:www-data); this root step keeps a readable copy of that ONE public
+# certificate where the agent's `proxmox.tls.ca_file` points.
+PVE_CA_SOURCE = "/etc/pve/pve-root-ca.pem"
+PVE_CA_COPY = "/etc/felhom-agent/pve-root-ca.pem"      # 0644 root: a CA certificate is public; the agent only reads it
+PVE_CA_MAX_BYTES = 64 * 1024
+PVE_LOCAL_API = "127.0.0.1:8006"                        # pveproxy on this host (the installer's proof uses the same)
+PEM_CERT_RE = re.compile(rb"-----BEGIN CERTIFICATE-----\s[A-Za-z0-9+/=\s]+-----END CERTIFICATE-----")
+# The security-profile refresh (R-1388, R-1301, `11` §5.12; measured on demo-hp 2026-11-14,
+# felhom.eu audits/profile-refresh-spike-2026-11-14/). AppArmor keeps one cache folder per kernel feature set; a
+# profile that changed while ANOTHER kernel ran leaves an old compiled copy in this kernel's folder, with a time that
+# looks newer than its source — and the kernel loads it at the next start on this kernel.
+PROFILE_DIR = "/etc/apparmor.d"
+PROFILE_CACHE = "/var/cache/apparmor"
+PROFILE_STATE = "/var/lib/felhom-os-apply/profile-refresh.json"   # 0644 root: the kernel the last refresh ran on
+PROFILE_REFRESH_CMD = ["apparmor_parser", "--replace", "--skip-read-cache", "--write-cache", "--", PROFILE_DIR]
+PROFILE_PATH_NAME_RE = re.compile(r"^/[A-Za-z0-9._/+-]+$")       # a profile whose name is one plain program path
+UNIT_NAME_RE = re.compile(r"^[A-Za-z0-9@:._-]+\.service$")
+PROFILE_RESTART_MAX = 3
+KERNEL_ENDED_NOT_GOOD = ("self_reverted", "revert_failed", "fell_back", "cancelled")
+KERNEL_LEFTOVER_PATHS = ("/lib/modules/{v}", "/boot/vmlinuz-{v}", "/boot/initrd.img-{v}", "/boot/System.map-{v}",
+                         "/boot/config-{v}")
+PINNED_KVER_RE = re.compile(r"gnulinux-([0-9]+\.[0-9]+\.[0-9]+-[0-9]+-pve)-advanced-")
+ANY_KVER_RE = re.compile(r"[0-9]+\.[0-9]+\.[0-9]+-[0-9]+-pve")
 
 # ---------- the config bundle (R-840, agent v0.143.0, `11` §5.4.2) ----------
 # A signed `agent_config_update` job carries {agent_version, bundle_sha256}; the bundle is ONE JSON file built from this
@@ -492,7 +522,7 @@ class Apply:
     def check_plan(self, plan):
         mode = plan.get("mode", "apply")
         if mode not in ("apply", "inventory", "health", "facts", "live-restore-on", "bundle", "bundle-policy", "agent_update",
-                        "oom-check") \
+                        "oom-check", "profile-refresh", "pve-ca-copy") \
                 + KERNEL_MODES:
             raise Refused("R11", f"unknown mode {mode!r}")
         if mode == "agent_update":
@@ -507,6 +537,25 @@ class Apply:
             if plan.get("layer") != "host":
                 raise Refused("R11", "bundle-policy is a host-layer mode")
             return mode, "host", 0, "bundle-policy"
+        if mode == "pve-ca-copy":
+            if plan.get("layer") != "host":
+                raise Refused("R11", "pve-ca-copy is a host-layer mode")
+            # R-1380 (D80): ONE fixed source, ONE fixed place — the plan may name NOTHING (no path, no file, no content)
+            extra = sorted(k for k in plan if k not in ("mode", "layer", "release_id", "run_id", "trigger", "ring"))
+            if extra:
+                raise Refused("T1", f"a pve-ca-copy plan carries {', '.join(extra)} — it takes nothing from a plan: "
+                                    f"the source is {PVE_CA_SOURCE} and the copy is {PVE_CA_COPY}, both fixed")
+            return mode, "host", 0, "pve-ca-copy"
+        if mode == "profile-refresh":
+            if plan.get("layer") != "host":
+                raise Refused("R11", "profile-refresh is a host-layer mode")
+            # it takes NOTHING from the plan but dry_run: no profile name, no unit name, no package (R-1388)
+            for k in ("packages", "allow_new", "signed", "select", "undo", "snapshot", "profiles", "units", "force"):
+                if plan.get(k):
+                    raise Refused("F1", f"a profile-refresh plan carries {k!r} — it takes nothing from a plan but dry_run")
+            if not isinstance(plan.get("dry_run", False), bool):
+                raise Refused("F1", "dry_run must be true or false")
+            return mode, "host", 0, "profile-refresh"
         layer = plan.get("layer")
         if layer not in ("guest", "host", "docker", "pve", "kernel"):
             raise Refused("R12", f"layer {layer!r} is not guest, host, docker, pve or kernel")
@@ -537,6 +586,17 @@ class Apply:
         rid = plan.get("release_id", "")
         if not isinstance(rid, str) or not re.match(r"^[A-Za-z0-9._:-]{1,80}$", rid):
             raise Refused("R11", f"release_id {rid!r} is not a plain id")
+        if mode == "kernel-cleanup":
+            # `11` §5.11.1, rule 4 of the release: the clean-up NEVER takes a package name (or any choice of what to
+            # remove) from a plan — the wrapper computes the set from the box. Pinned by test_cleanup_plan_names_refused.
+            for k in ("packages", "allow_new", "signed", "expect_kver", "undo", "snapshot"):
+                if plan.get(k):
+                    raise Refused("C1", f"a kernel-cleanup plan carries {k!r} — the clean-up takes nothing from a plan "
+                                        f"but dry_run; the wrapper computes the set from the box")
+            if plan.get("select", "listed") != "listed":
+                raise Refused("C1", "a kernel-cleanup plan carries a select — the wrapper computes the set from the box")
+            if not isinstance(plan.get("dry_run", False), bool):
+                raise Refused("C1", "dry_run must be true or false")
         if plan.get("allow_new"):
             raise Refused("R6", "allow_new is a slow-lane field; the fast lane never adds a package")
         select = plan.get("select", "listed")
@@ -1081,6 +1141,13 @@ class Apply:
             return Bundle(self).from_plan(plan)
         if self.mode == "bundle-policy":
             return Bundle(self).policy_pass()
+        if self.mode == "pve-ca-copy":
+            # NO appliance check, on purpose (like the bundle): the copy is the agent's OWN trust file in the agent's own
+            # config folder, and it changes nothing of the host's — a BYO host's agent talks to Proxmox the same way.
+            return self.pve_ca_copy()
+        if self.mode == "profile-refresh":
+            self.check_appliance()
+            return self.profile_refresh(plan)
         if self.mode == "agent_update":
             return self.agent_update(plan)
         if self.layer in ("host", "pve", "kernel"):
@@ -1146,6 +1213,198 @@ class Apply:
             self.report["oom_check"] = self.oom_check()
         return 0
 
+    # ---------- trust in Proxmox by a copy of its root CA (R-1380, D80) ----------
+    def is_live_ca(self, block):
+        """One PEM certificate block: is it a CA certificate that has not ended? (openssl, the bytes on stdin)."""
+        try:
+            rc, out, _ = self.r.host(["openssl", "x509", "-noout", "-ext", "basicConstraints", "-checkend", "0"], 30,
+                                     stdin=block.decode("ascii") + "\n")
+        except (UnicodeDecodeError, subprocess.SubprocessError, OSError):
+            return False
+        return rc == 0 and re.search(r"\bCA:TRUE\b", out) is not None
+
+    def pve_serves_with(self, cafile):
+        """Does the local Proxmox API serve a certificate that verifies with this CA file ALONE? One handshake to
+        PVE_LOCAL_API; nothing is sent. Any doubt (no answer, a timeout) is "no" — the previous root then stays."""
+        try:
+            rc, out, err = self.r.host(["openssl", "s_client", "-connect", PVE_LOCAL_API, "-CAfile", cafile, "-no-CApath",
+                                        "-verify_return_error"], 20, stdin="")
+        except (subprocess.SubprocessError, OSError):
+            return False
+        return rc == 0 and "Verify return code: 0 (ok)" in (out + err)
+
+    def pve_ca_copy(self):
+        """Copy exactly PVE_CA_SOURCE to exactly PVE_CA_COPY — only if the source is ONE CA certificate that has not
+        ended — written whole or not at all (a root-owned temp file beside the target, fsync, rename). "same" when the
+        copy is exactly the source's bytes. The bytes that are checked are the bytes that are written first; only
+        during a change of the root the copy also holds its previous certificate behind them (see below)."""
+        r = self.r
+        res = {"outcome": None, "path": PVE_CA_COPY}
+        self.report["pve_ca_copy"] = res
+        try:
+            data = r.read_bytes(PVE_CA_SOURCE)
+        except OSError as e:
+            raise Refused("T2", f"{PVE_CA_SOURCE} cannot be read ({e.__class__.__name__}) — is this a Proxmox host?")
+        if not data or len(data) > PVE_CA_MAX_BYTES:
+            raise Refused("T3", f"{PVE_CA_SOURCE} is {len(data)} bytes — not one certificate")
+        certs = PEM_CERT_RE.findall(data)
+        if len(certs) != 1 or b"PRIVATE KEY" in data:
+            raise Refused("T3", f"{PVE_CA_SOURCE} holds {len(certs)} certificate(s)"
+                                + (" and a private key" if b"PRIVATE KEY" in data else "") + " — exactly one CA certificate is expected")
+        try:
+            text = data.decode("ascii")      # PEM is ASCII; the runner hands text to openssl's stdin
+        except UnicodeDecodeError:
+            raise Refused("T3", f"{PVE_CA_SOURCE} is not a PEM text file")
+        rc, out, err = r.host(["openssl", "x509", "-noout", "-ext", "basicConstraints", "-enddate"], 30, stdin=text)
+        if rc != 0:
+            raise Refused("T3", f"{PVE_CA_SOURCE} is not a readable certificate: {(out + err).strip()[-160:]}")
+        if not re.search(r"\bCA:TRUE\b", out):
+            raise Refused("T3", f"{PVE_CA_SOURCE} is not a CA certificate (no CA:TRUE) — a leaf or any other file is never copied")
+        m = re.search(r"notAfter=(.+)", out)
+        try:
+            import calendar
+            not_after = calendar.timegm(time.strptime(m.group(1).strip(), "%b %d %H:%M:%S %Y GMT"))   # UTC, no DST
+        except (AttributeError, ValueError):
+            raise Refused("T3", f"the end date of {PVE_CA_SOURCE} cannot be read")
+        days = int((not_after - r.now()) // 86400)
+        if days < 0:
+            raise Refused("T3", f"{PVE_CA_SOURCE} ended {-days} day(s) ago — an ended CA is never copied")
+        res.update(sha256=sha256_hex(data), not_after=_iso(not_after), days_left=days)
+        try:
+            have = r.read_bytes(PVE_CA_COPY)
+        except OSError:
+            have = None
+        if have == data:
+            res["outcome"] = "same"
+            return 0
+        # The copy the agent trusts by is replaced BEFORE anything can be proven with the new root (Proxmox serves its
+        # old certificate until pveproxy is reloaded). So at a CHANGE the copy's previous first certificate — written
+        # by this same step — is kept behind the new one, and the agent trusts both roots across the change, also
+        # after a restart. That second root is kept ONLY for the changeover: it is dropped as soon as Proxmox serves a
+        # certificate that verifies with the new root alone, and at once when it is no longer a CA that has not ended.
+        # The copy is compared EXACTLY: nothing this step did not check is ever left in it. (Review, 2026-11-14.)
+        # To drop the old root at once by hand (a root that was changed because it leaked): delete the copy.
+        prev = PEM_CERT_RE.findall(have) if have and b"PRIVATE KEY" not in have else []
+        if len(prev) == 2 and have == data + prev[1] + b"\n":
+            over = self.pve_serves_with(PVE_CA_SOURCE)
+            if not over and self.is_live_ca(prev[1]):
+                res["outcome"], res["kept_previous"] = "same", True   # the changeover is still on
+                return 0
+            r.put_file(PVE_CA_COPY, data, 0o644)
+            res["outcome"], res["replaced"], res["kept_previous"], res["dropped_previous"] = "written", True, False, True
+            r.log(f"os-apply: PVE CA COPY the previous root is dropped from {PVE_CA_COPY} "
+                  f"({'Proxmox serves a certificate of the new root' if over else 'the previous root is no longer a live CA'})")
+            return 0
+        want, kept = data, False
+        if prev and prev[0].strip() != certs[0].strip() and self.is_live_ca(prev[0]):
+            want, kept = data + prev[0] + b"\n", True
+        r.put_file(PVE_CA_COPY, want, 0o644)
+        res["outcome"], res["replaced"], res["kept_previous"] = "written", have is not None, kept
+        r.log(f"os-apply: PVE CA COPY written {PVE_CA_COPY} sha256={res['sha256'][:16]} ends={res['not_after']} "
+              f"({'replaced an older copy' if have is not None else 'first copy'})")
+        return 0
+
+    # ---------- the security-profile refresh (R-1388, R-1301, `11` §5.12) ----------
+    def profile_cache(self):
+        """{"<folder>/<file>": sha256} of every cached profile, every kernel's folder."""
+        out = {}
+        for d in self.r.list_dir(PROFILE_CACHE):
+            dp = f"{PROFILE_CACHE}/{d}"
+            if not self.r.isdir(dp):
+                continue
+            for f in self.r.list_dir(dp):
+                if f.startswith("."):
+                    continue
+                try:
+                    out[f"{d}/{f}"] = sha256_hex(self.r.read_bytes(f"{dp}/{f}"))
+                except OSError:
+                    out[f"{d}/{f}"] = "unreadable"
+        return out
+
+    def profile_refresh(self, plan):
+        """Load every AppArmor profile from a FRESH compile (the cache ignored and rewritten), once per kernel the box
+        starts on; then restart ONLY a failed service whose own program's profile had a stale cached copy.
+
+        Changes no package and takes no name from the plan. Measured on demo-hp with both guests running: 0.5 s, no
+        guest or container touched, the same profiles loaded in the same modes.
+
+        WHY NOT a "loaded hash changed" check: for a file that defines several profiles the kernel's hash differs
+        between a cache load and a fresh load of the SAME policy (17 of 119 on demo-hp), and one file (lxc-containers)
+        compiles to a different binary every time. The only reliable sign of a stale copy is the CACHED FILE changing
+        when it is rewritten — that is what the restart rule reads."""
+        r, dry = self.r, plan.get("dry_run") is True
+        res = {"dry_run": dry, "outcome": None, "rewritten": [], "restarted": [], "not_restarted": []}
+        self.report["profile_refresh"] = res
+        rc, out, _ = r.host(["uname", "-r"], 30)
+        kernel = out.strip() if rc == 0 else ""
+        try:
+            rec = json.loads(r.read_file(PROFILE_STATE))
+            rec = rec if isinstance(rec, dict) else {}
+        except (OSError, ValueError):
+            rec = {}
+        res["kernel"], res["last_refresh_kernel"], res["last_refresh_at"] = kernel or "unknown", rec.get("kernel"), rec.get("at")
+        if not kernel:
+            raise Refused("F2", "the running kernel could not be read")
+        if not r.isdir(PROFILE_DIR) or not r.isdir(PROFILE_CACHE):
+            res["outcome"] = "not_needed"
+            res["reason"] = "this box has no AppArmor profile folder or no profile cache"
+            return 0
+        needed = rec.get("kernel") != kernel
+        if not needed:
+            res["outcome"], res["reason"] = "not_needed", f"the profiles were refreshed on this kernel ({kernel}) at {rec.get('at')}"
+            return 0
+        if dry:
+            res["outcome"], res["would"] = "dry_run", "refresh"
+            return 0
+        before = self.profile_cache()
+        t0 = time.time()
+        rc, out, err = r.host(list(PROFILE_REFRESH_CMD), 600)
+        res["seconds"] = round(time.time() - t0, 1)
+        if rc != 0:
+            res["outcome"] = "failed"
+            self.report["failed"] = {"rc": rc, "step": "profile-refresh", "tail": (out + err).strip().splitlines()[-3:]}
+            r.log(f"os-apply: FAILED rc={rc} step=profile-refresh (no record written: the next start tries again)")
+            return 3
+        after = self.profile_cache()
+        res["rewritten"] = sorted(k for k in after if before.get(k) != after[k])
+        # --- the narrow restart rule: a FAILED service, whose ExecStart program IS the name of a profile whose cached
+        #     copy was stale. Never "whatever failed"; never a unit named by the plan. ---
+        programs = {}
+        for key in res["rewritten"]:
+            fname = key.split("/", 1)[1]
+            rc_n, names, _ = r.host(["apparmor_parser", "-N", f"{PROFILE_DIR}/{fname}"], 60)
+            if rc_n != 0:
+                continue
+            for n in names.split():
+                if PROFILE_PATH_NAME_RE.match(n):
+                    programs[n] = fname
+        if programs:
+            rc_u, units, _ = r.host(["systemctl", "list-units", "--type=service", "--state=failed", "--no-legend", "--plain"], 60)
+            for line in (units.splitlines() if rc_u == 0 else []):
+                unit = (line.split() or [""])[0]
+                if not UNIT_NAME_RE.match(unit):
+                    continue
+                _, es, _ = r.host(["systemctl", "show", "-p", "ExecStart", "--value", unit], 30)
+                m = re.search(r"\bpath=(\S+)", es)
+                prog = m.group(1) if m else ""
+                if prog not in programs:
+                    continue
+                if len(res["restarted"]) >= PROFILE_RESTART_MAX:
+                    res["not_restarted"].append({"unit": unit, "program": prog, "reason": f"more than {PROFILE_RESTART_MAX} units match — look at the box"})
+                    continue
+                rc_r, o2, e2 = r.host(["systemctl", "restart", unit], 120)
+                _, act, _ = r.host(["systemctl", "is-active", unit], 30)
+                res["restarted"].append({"unit": unit, "program": prog, "profile_file": programs[prog], "rc": rc_r,
+                                         "active": act.strip() == "active"})
+                r.log(f"os-apply: PROFILE REFRESH restarted {unit} (failed; its program {prog} had a stale cached profile) "
+                      f"rc={rc_r} now={act.strip() or 'unknown'}")
+        r.put_file(PROFILE_STATE, (json.dumps({"kernel": kernel, "at": _iso(r.now()), "rewritten": res["rewritten"],
+                                               "restarted": [x["unit"] for x in res["restarted"]]}, indent=2, sort_keys=True) + "\n").encode(), 0o644)
+        res["outcome"] = "refreshed"
+        r.log(f"os-apply: PROFILE REFRESH DONE kernel={kernel} (last {rec.get('kernel') or 'never'}) seconds={res['seconds']} "
+              f"rewritten={len(res['rewritten'])} ({', '.join(res['rewritten'][:8])}) restarted={[x['unit'] for x in res['restarted']]}")
+        return 0
+
     # ---------- R-528: the memory-kill check ----------
     def oom_check(self):
         """Run one throwaway container over its memory cap in the guest and read what the engine says about it.
@@ -1490,6 +1749,11 @@ class Kernel:
       kernel-good    the one-shot boot was healthy: the new kernel becomes the GRUB default.
       kernel-revert  the one-shot boot was NOT healthy: reboot ONCE into the old kernel (still the default).
       kernel-cancel  drop a staged step: clear the flag (the package stays installed, the default never moved).
+      kernel-cleanup remove old kernel images by the rule of `11` §5.11.1 (R-1368, decision 286). The plan names nothing
+                     (C1); `dry_run: true` computes and reports, and changes nothing. Its own refusals: P1–P3, K1, N1,
+                     S1, S2 (the design's table) and C1 (the plan names something). "Not run" (exit 0, nothing
+                     changed): no ring-0 mark, no kernel step yet, the newest step did not end good, the box does not
+                     run its newest kernel, inside a step's 20 h window.
       kernel-status  read only."""
 
     def __init__(self, apply, plan):
@@ -1643,7 +1907,8 @@ class Kernel:
             self.report["kernel"] = v
             return 0
         fn = {"apply": self.stage, "kernel-reboot": self.reboot_staged, "kernel-boot": self.after_boot,
-              "kernel-good": self.good, "kernel-revert": self.revert, "kernel-cancel": self.cancel}[mode]
+              "kernel-good": self.good, "kernel-revert": self.revert, "kernel-cancel": self.cancel,
+              "kernel-cleanup": self.cleanup}[mode]
         rc = fn()
         self.report["kernel"] = self.view()
         return rc
@@ -1879,6 +2144,221 @@ class Kernel:
         self.log(f"os-apply: KERNEL CANCELLED {st.get('to')} (installed, never the default; flag cleared)")
         return 0
 
+    # ---------- the clean-up of old kernels (R-1368, `09` §3 decision 286, `11` §5.11.1) ----------
+    def kernel_images(self):
+        """kver -> [installed image package names] (dpkg status ii only)."""
+        out = {}
+        for name in sorted(self.a.installed()):
+            m = KERNEL_IMAGE_RE.match(name)
+            if m:
+                out.setdefault(m.group(1), []).append(name)
+        return out
+
+    def kver_newest(self, kvers):
+        best = None
+        for k in kvers:
+            if best is None or self.a.dpkg_cmp(k[:-4], "gt", best[:-4]):
+                best = k
+        return best
+
+    def read_or_none(self, path):
+        try:
+            return self.r.read_file(path)
+        except OSError:
+            return None
+
+    def cleanup_sources(self):
+        """Every place that names a kernel the lane still needs, read from the box ONCE: (kver, why) pairs. This is
+        the keep set's only input, and K1's only input — K1 reads it again by itself (cleanup_protected)."""
+        st = self.state()
+        cfg = self.grub_cfg()
+        dflt = self.default_kver(cfg)
+        m = PINNED_KVER_RE.search(self.read_or_none(KERNEL_DEFAULT_CFG) or "")
+        pinned = m.group(1) if m else None
+        flag = self.flag() or ""
+        m = ANY_KVER_RE.search(flag)
+        m2 = ANY_KVER_RE.search(self.read_or_none(KERNEL_PIN_FILE) or "")
+        src = [(self.running(), "running"),
+               (st.get("from"), "the kernel step's `from` (the last kernel that booted well; a self-revert boots it)"),
+               (st.get("to"), "the kernel step's `to`"),
+               (st.get("default_before"), "the kernel step's `default_before`"),
+               (dflt if KVER_RE.match(dflt) else None, "the default in grub.cfg"),
+               (pinned, "the pinned default file"),
+               (m.group(0) if m else None, "the one-shot flag"),
+               (m2.group(0) if m2 else None, "an operator's boot pin")]
+        return st, dflt, pinned, flag, [(k, why) for k, why in src if isinstance(k, str) and KVER_RE.match(k)]
+
+    def cleanup_keep(self, images, sources):
+        """The keep set: every kernel a source names, and the newest installed. (cleanup() adds the next lower
+        installed kernel when no step record names the last good one.)"""
+        keep = {}
+        for k, why in sources:
+            keep.setdefault(k, []).append(why)
+        newest = self.kver_newest(images)
+        if newest:
+            keep.setdefault(newest, []).append("newest installed")
+        return keep
+
+    def cleanup_protected(self):
+        """K1's own reading (`11` §5.11.1: "K1 is the rule's own check and must stay its own"): the kernels that must
+        never reach the purge line, read from the box AGAIN — not taken from the keep set that built the removal set."""
+        _, _, _, _, sources = self.cleanup_sources()
+        return {k: why for k, why in sources}
+
+    def cleanup_simulate(self, names):
+        """apt's simulation of exactly `names`: (rc, removes, installs, last line)."""
+        rc, out, err = self.a.x(APT_ENV + ["apt-get", "-s", "-q", "purge"] + names)
+        removes = sorted({l.split()[1] for l in out.splitlines() if l.startswith(("Remv ", "Purg ")) and len(l.split()) > 1})
+        installs = sorted({l.split()[1] for l in out.splitlines() if l.startswith(("Inst ", "Conf ")) and len(l.split()) > 1})
+        return rc, removes, installs, ((err or out).strip().splitlines() or ["?"])[-1][:200]
+
+    def cleanup(self):
+        a = self.a
+        dry = self.plan.get("dry_run") is True
+        res = {"dry_run": dry, "outcome": None, "not_run": [], "refusals": [], "removed": [], "leftovers": []}
+        self.report["kernel_cleanup"] = res
+        not_run, refuse = res["not_run"], res["refusals"]
+        # --- may this box run the clean-up at all? (ring 0 only in this release; the mark is ROOT-owned) ---
+        try:
+            ring0 = a.load_trust().get("ring0_slow_lane") is True
+        except Refused as e:
+            ring0 = False
+            not_run.append({"code": "no-ring0-mark", "reason": f"no usable trust record: {e.reason}"})
+        if not ring0 and not not_run:
+            not_run.append({"code": "no-ring0-mark", "reason": "this box has no root-owned ring-0 mark — the clean-up runs on "
+                            "ring 0 only in this release"})
+        if a.apt_lock_held():
+            raise Refused("R9", "another apt/dpkg holds the lock on the host")
+        if not dry and not not_run:
+            a.repair()   # repair first (`11` §5.4.1): a purge a power cut interrupted is made whole, THEN the rule is computed
+        # --- read the box ---
+        images = self.kernel_images()
+        running = self.running()
+        newest = self.kver_newest(images)
+        raw_state = self.read_or_none(KERNEL_STATE)
+        st, dflt, pinned, flag, sources = self.cleanup_sources()
+        phase = st.get("phase", "none") if raw_state is not None else "none"
+        res.update(running=running or "unknown", newest=newest, phase=phase, default=dflt, pinned_default=pinned,
+                   installed={k: images[k] for k in sorted(images, key=lambda v: [int(x) for x in re.findall(r"\d+", v)])})
+        # --- "when it may run" (decision 286): not run, nothing changed ---
+        if raw_state is None or not st:
+            not_run.append({"code": "no-step", "reason": "no kernel step yet (no state file) — nothing proves which kernel booted well"})
+        elif phase in KERNEL_ENDED_NOT_GOOD:
+            not_run.append({"code": "step-not-good", "reason": f"the newest kernel step ended {phase!r}, not good — every kernel stays"})
+        elif phase != "good":
+            refuse.append({"code": "P1", "reason": f"a kernel step is in phase {phase!r} — not ended (or not a phase this wrapper knows)"})
+        if running and newest and running in images and running != newest:
+            not_run.append({"code": "not-on-newest", "reason": f"the box runs {running}, not its newest installed kernel {newest} — every kernel stays"})
+        times = [x for x in (_parse_iso(st.get(k)) for k in ("staged_at", "rebooted_at", "result_at")) if x is not None]
+        if times and self.r.now() - max(times) < KERNEL_MIN_GAP:
+            not_run.append({"code": "window", "reason": "a kernel step staged, booted or ended within the last 20 hours — never in its night"})
+        # --- the refusal table ---
+        if flag:
+            refuse.append({"code": "P2", "reason": f"the one-shot flag names {flag!r}"})
+        if pinned is None or pinned != dflt or pinned not in images:
+            refuse.append({"code": "P3", "reason": f"the pinned default {pinned!r} and grub.cfg's default {dflt!r} do not name the same installed kernel"})
+        if not running or running not in images:
+            refuse.append({"code": "K1", "reason": f"the running kernel {running or 'unknown'!r} is not an installed kernel image package"})
+        # --- the keep set and the removal set (exact package names) ---
+        keep = self.cleanup_keep(images, sources)
+        if running and not (st.get("to") == running and st.get("from") in images):
+            lower = [k for k in images if a.dpkg_cmp(k[:-4], "lt", running[:-4])] if running in images else []
+            lg = self.kver_newest(lower)
+            if lg:
+                keep.setdefault(lg, []).append("the next lower installed kernel (no step record names the last good one)")
+        removal = sorted(n for k in images if k not in keep for n in images[k])
+        res["kept"] = {k: keep[k] for k in sorted(keep) if k in images}
+        res["removal"] = removal
+        # K1 — its own reading of the box, against the removal set
+        by_name = {n: k for k, names in images.items() for n in names}
+        protected = self.cleanup_protected()
+        for n in removal:
+            if n not in by_name:
+                refuse.append({"code": "N1", "reason": f"{n} is not an installed kernel image package"})
+            elif by_name[n] in protected or by_name[n] == newest:
+                refuse.append({"code": "K1", "reason": f"{n} is a keep-kernel ({protected.get(by_name[n], 'newest installed')})"})
+        # S1 / S2 — apt's simulation of exactly that set
+        if removal and not [x for x in refuse if x["code"] in ("K1", "N1")]:
+            rc, removes, installs, tail = self.cleanup_simulate(removal)
+            res["simulation"] = {"rc": rc, "removes": removes, "installs": installs}
+            if rc != 0:
+                refuse.append({"code": "S1", "reason": "the simulation failed: " + tail})
+            if installs:
+                refuse.append({"code": "S1", "reason": f"the simulation would install or configure {', '.join(installs[:6])}"})
+            extra = [p for p in removes if p not in removal]
+            if extra:
+                refuse.append({"code": "S2", "reason": f"the simulation would also remove {', '.join(extra[:8])} — not in the removal set"})
+            missing = [p for p in removal if p not in removes]
+            if missing and rc == 0:
+                refuse.append({"code": "S1", "reason": f"the simulation does not remove {', '.join(missing[:6])}"})
+        self.log(f"os-apply: KERNEL CLEANUP {'DRY RUN ' if dry else ''}running={running or 'unknown'} newest={newest} phase={phase} "
+                 f"installed={len(images)} keep={len(res['kept'])} removal={len(removal)} not_run={[x['code'] for x in not_run]} "
+                 f"refusals={[x['code'] for x in refuse]}")
+        # --- the verdict ---
+        if dry:
+            res["outcome"] = "dry_run"
+            res["would"] = "not_run" if not_run else "refuse" if refuse else "remove" if removal else "nothing"
+            return 0
+        if not_run:
+            res["outcome"] = "not_run"
+            return 0
+        if refuse:
+            raise Refused(refuse[0]["code"], "; ".join(f"{x['code']} {x['reason']}" for x in refuse)[:600])
+        if not removal:
+            res["outcome"] = "nothing"
+            self.log("os-apply: KERNEL CLEANUP nothing to remove")
+            return 0
+        # --- the purge: ONE apt-get purge of the exact names; then prove the boot setup did not move ---
+        gens_before = {p: (self.r.lexists(p), self.read_or_none(p)) for p in (ONESHOT_SNIPPET, ONESHOT_ENTRIES)}
+        pin_before = self.read_or_none(KERNEL_DEFAULT_CFG)
+        free_before = a.free_bytes()
+        t0 = time.time()
+        rc, out, err = a.x(APT_ENV + ["apt-get", "-y", "-q", "purge"] + removal)
+        res["seconds"] = round(time.time() - t0, 1)
+        if rc != 0:
+            _, aud, _ = a.x(["dpkg", "--audit"])
+            res["outcome"] = "failed"
+            self.report["failed"] = {"rc": rc, "step": "purge", "dpkg_audit": (aud.strip().splitlines() or ["clean"])[0],
+                                     "tail": (out + err).strip().splitlines()[-3:]}
+            self.log(f"os-apply: FAILED rc={rc} step=purge (the next run repairs dpkg first, then computes the rule again)")
+            return 3
+        after = self.kernel_images()
+        after_names = {n for names in after.values() for n in names}
+        res["removed"] = [n for n in removal if n not in after_names]
+        free_after = a.free_bytes()
+        res["bytes_gained"] = free_after - free_before if free_before >= 0 and free_after >= 0 else None
+        problems = []
+        if len(res["removed"]) != len(removal):
+            problems.append(f"still installed after the purge: {', '.join(n for n in removal if n in after_names)}")
+        gone = [k for k in res["kept"] if k not in after]
+        if gone:
+            problems.append(f"a KEPT kernel is no longer installed: {', '.join(gone)}")
+        got = self.default_kver(self.grub_cfg())
+        if got != dflt:
+            problems.append(f"the default in grub.cfg moved from {dflt} to {got}")
+        if self.read_or_none(KERNEL_DEFAULT_CFG) != pin_before:
+            problems.append("the pinned default file changed")
+        if self.read_or_none(KERNEL_STATE) != raw_state:
+            problems.append("the kernel step's state file changed")
+        for p, (was, body) in gens_before.items():
+            if self.r.lexists(p) != was or self.read_or_none(p) != body:
+                problems.append(f"{p} changed")
+        for k in sorted(set(by_name[n] for n in res["removed"])):
+            if k in after:
+                continue  # another image package of the same kernel is still installed
+            res["leftovers"] += [p.format(v=k) for p in KERNEL_LEFTOVER_PATHS if self.r.lexists(p.format(v=k))]
+        if problems:
+            # reported LOUDLY, never repaired by hand here (`11` §5.11.1): the operator looks at the box
+            res["outcome"] = "failed"
+            self.report["failed"] = {"rc": 3, "step": "verify", "reason": "; ".join(problems)[:600]}
+            self.log("os-apply: FAILED rc=3 step=verify after the kernel purge — " + "; ".join(problems))
+            return 3
+        res["outcome"] = "removed"
+        self.log(f"os-apply: KERNEL CLEANUP DONE removed={len(res['removed'])} ({', '.join(res['removed'])}) "
+                 f"kept={', '.join(res['kept'])} gained={res['bytes_gained']} B seconds={res['seconds']} "
+                 f"leftovers={len(res['leftovers'])}; default {dflt} (proved from grub.cfg)")
+        return 0
+
 
 class Bundle:
     """The config bundle (R-840, `11` §5.4.2): every root-owned file the installer's step 5 writes, installed as ONE

configs/test_felhom_os_apply.py

@@ -1494,6 +1494,24 @@ class PVELane(unittest.TestCase):
         self.assertEqual(sorted(u["name"] for u in rep["upgraded"]), ["pve-manager"])
         self.assertEqual(f.installed["proxmox-secure-boot-support"], "9.0.1")
 
+    # 2026-11-05..07, demo-felhom (EFI): proxmox-grub 2.12-9+pmx2 -> pmx4 depends on grub-efi-amd64 of the same version;
+    # HOST_SLOW_RE's `grub` is anchored at the start, so proxmox-grub fell into the pve plan, the simulation pulled
+    # grub-efi-amd64 and the whole pve step was refused R6 every night. It is a boot-chain package: left out.
+    # RED-PROOF: drop proxmox-grub from HOST_SLOW_RE -> it is in the plan -> this test fails.
+    def test_ring0_pending_pve_leaves_proxmox_grub_out(self):
+        f = pve_fake(ring0=True)
+        f.plan["select"], f.plan["packages"] = "pending-pve", []
+        f.installed.update({"proxmox-grub": "2.12-9+pmx2"})
+        f.pending_sim = [
+            "Inst pve-manager [9.2.2] (9.2.21 Proxmox Debian Repository:stable [amd64])",
+            "Inst proxmox-grub [2.12-9+pmx2] (2.12-9+pmx4 Proxmox Debian Repository:stable [amd64])",
+            "Inst grub-efi-amd64 [2.12-9+pmx2] (2.12-9+pmx4 Proxmox Debian Repository:stable [amd64])",
+        ]
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(sorted(u["name"] for u in rep["upgraded"]), ["pve-manager"])
+        self.assertEqual(f.installed["proxmox-grub"], "2.12-9+pmx2")
+
     def test_select_pending_pve_needs_the_pve_layer(self):
         f = Fake()
         f.plan["layer"], f.plan["select"], f.plan["packages"] = "host", "pending-pve", []
@@ -2013,5 +2031,847 @@ class KernelLane(unittest.TestCase):
         self.assertIn("one-shot", h["kernel_next_boot_source"])
 
 
+
+# ---------- the clean-up of old kernels (R-1368, `09` §3 decision 286, `11` §5.11.1) ----------
+K = ["7.0.2-6-pve", "7.0.14-20-pve", "7.0.14-22-pve", "7.0.14-26-pve", "7.0.14-28-pve"]   # five installed kernels
+C_FROM, C_RUN = "7.0.14-26-pve", "7.0.14-28-pve"
+GB = 1024 ** 3
+
+
+def img(k):
+    return f"proxmox-kernel-{k}-signed"
+
+
+class CFake(KFake):
+    """A ring-0 appliance after a good kernel step 26 -> 28: five kernels installed, the box runs its newest, the
+    default (grub.cfg and the pinned file) is the running kernel, the flag is empty. `apt-get purge` is emulated like
+    the real one: the image leaves dpkg and /boot, and the kernel's postrm runs update-grub."""
+
+    def __init__(self):
+        super().__init__()
+        self.running = C_RUN
+        self.boot = set(K)
+        self.modules = set(K)            # /lib/modules/<kver>
+        for n in [x for x in self.installed if osapply.KERNEL_IMAGE_RE.match(x)]:
+            del self.installed[n]
+        for k in K:
+            self.installed[img(k)] = k[:-4]
+        self.installed["proxmox-kernel-7.0"] = C_RUN[:-4]
+        self.env = {"felhom_next": ""}
+        self.tree[osapply.KERNEL_DEFAULT_CFG] = (f'GRUB_DEFAULT="gnulinux-advanced-{U}>gnulinux-{C_RUN}-advanced-{U}"\n')
+        self.state = {"phase": "good", "step_id": "ring0-x", "from": C_FROM, "to": C_RUN, "default_before": C_FROM,
+                      "vmid": 9201, "staged_at": "2026-10-01T02:50:00Z", "rebooted_at": "2026-10-01T02:51:00Z",
+                      "result_at": "2026-10-01T03:10:00Z", "self_revert_used": False}
+        self.write_state()
+        self.cfg = self.render()
+        self.purges, self.purge_rc, self.purge_moves_default, self.purge_extra_sim = [], 0, False, []
+        self.plan = {"release_id": "kernel", "layer": "kernel", "lane": "slow", "vmid": 9201, "mode": "kernel-cleanup",
+                     "packages": []}
+
+    def write_state(self):
+        if self.state is None:
+            self.tree.pop(osapply.KERNEL_STATE, None)
+        else:
+            self.tree[osapply.KERNEL_STATE] = json.dumps(self.state)
+
+    def lexists(self, p):
+        m = re.match(r"^/lib/modules/(.+)$", p)
+        if m:
+            return m.group(1) in self.modules
+        return super().lexists(p)
+
+    def changing(self):
+        """Every command that changes the box (the purge, a repair, update-grub, a flag write, a file write)."""
+        out = []
+        for c in self.calls:
+            a = [x for x in c[-1] if not re.match(r"^[A-Z_]+=", x) and x != "env"]
+            if (a[0] == "apt-get" and "-s" not in a and "--print-uris" not in a) or a[0] in ("update-grub", "dpkg-reconfigure") \
+                    or (a[0] == "dpkg" and a[1] == "--configure") or (a[0] == "grub-editenv" and a[2] != "list") \
+                    or (a[0] == "systemctl" and a[1] == "reboot"):
+                out.append(a)
+        return out
+
+    def emulate(self, argv):
+        a = [x for x in argv if not re.match(r"^[A-Z_]+=", x) and x != "env"]
+        if a[0] == "apt-get" and "purge" in a and "-s" not in a:
+            names = [x for x in a[a.index("purge") + 1:] if not x.startswith("-")]
+            self.purges.append(names)
+            if self.purge_rc:
+                self.dpkg_journal = ["0001"]     # a purge cut in the middle leaves dpkg's update journal
+                return self.purge_rc, "", "E: Sub-process /usr/bin/dpkg returned an error code (1)"
+            for n in names:
+                self.installed.pop(n, None)
+                m = osapply.KERNEL_IMAGE_RE.match(n)
+                if m:
+                    self.boot.discard(m.group(1))
+                    self.free += GB
+            if self.purge_moves_default:
+                self.tree.pop(osapply.KERNEL_DEFAULT_CFG, None)
+            self.cfg = self.render()     # the kernel's postrm runs update-grub (zz-update-grub)
+            return 0, "Removing proxmox-kernel ...\n", ""
+        return super().emulate(argv)
+
+    def sim(self, a):
+        if "purge" in a:
+            names = [x for x in a[a.index("purge") + 1:] if not x.startswith("-")]
+            if getattr(self, "purge_sim_rc", 0):
+                return self.purge_sim_rc, "", "E: Unable to locate package"
+            out = "".join(f"Purg {n} [{self.installed[n]}]\n" for n in names if n in self.installed)
+            return 0, out + "".join(l + "\n" for l in self.purge_extra_sim), ""
+        return super().sim(a)
+
+
+def cfake(**kw):
+    f = CFake()
+    for k, v in kw.items():
+        setattr(f, k, v)
+    return f
+
+
+def cstate(f, **kw):
+    f.state.update(kw)
+    f.write_state()
+    return f
+
+
+class KernelCleanup(unittest.TestCase):
+    """§5 of the release-2a brief: one test per acceptance scenario, each on the stored effect (the exact apt line or
+    the absence of any changing command, the report, the exit code). Red-proofs (the four mutations the brief names):
+    felhom.eu audits/release-2a-2026-11-14/A/redproof.txt."""
+
+    REMOVE = [img(k) for k in ("7.0.14-20-pve", "7.0.14-22-pve", "7.0.2-6-pve")]   # sorted by name
+
+    def nothing_changed(self, f):
+        self.assertEqual(f.purges, [], "no purge")
+        self.assertEqual(f.changing(), [], "no command that changes the box")
+        self.assertEqual(sorted(f.boot), sorted(K), "every kernel is still there")
+
+    def refused(self, f, code):
+        before = (f.tree.get(osapply.KERNEL_STATE), f.tree.get(osapply.KERNEL_DEFAULT_CFG), f.cfg, dict(f.env or {}))
+        rc, rep = run(f)
+        self.assertEqual(rc, 2, rep)
+        self.assertEqual(rep["refused"]["code"], code, rep)
+        self.nothing_changed(f)
+        self.assertEqual((f.tree.get(osapply.KERNEL_STATE), f.tree.get(osapply.KERNEL_DEFAULT_CFG), f.cfg, dict(f.env or {})), before)
+        return rep
+
+    def not_run(self, f, code):
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        c = rep["kernel_cleanup"]
+        self.assertEqual(c["outcome"], "not_run", rep)
+        self.assertIn(code, [x["code"] for x in c["not_run"]], rep)
+        self.assertIsNone(rep.get("refused"))
+        self.nothing_changed(f)
+        return rep
+
+    # --- scenario 1: the normal night ---
+    def test_s1_the_normal_night_purges_exactly_the_old_kernels_in_one_command(self):
+        f = cfake()
+        state0, pin0 = f.tree[osapply.KERNEL_STATE], f.tree[osapply.KERNEL_DEFAULT_CFG]
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(f.purges, [self.REMOVE], "ONE apt-get purge of the exact names")
+        purge = [c[-1] for c in f.calls if "purge" in c[-1] and "-s" not in c[-1]]
+        self.assertEqual(len(purge), 1)
+        self.assertEqual(purge[0][purge[0].index("apt-get"):], ["apt-get", "-y", "-q", "purge"] + self.REMOVE)
+        self.assertNotIn("autoremove", " ".join(" ".join(c[-1]) for c in f.calls), "never apt's own clean-up")
+        c = rep["kernel_cleanup"]
+        self.assertEqual(c["outcome"], "removed")
+        self.assertEqual(c["removed"], self.REMOVE)
+        self.assertEqual(sorted(c["kept"]), sorted([C_FROM, C_RUN]))
+        self.assertEqual(c["bytes_gained"], 3 * GB)
+        self.assertEqual(sorted(f.boot), sorted([C_FROM, C_RUN]))
+        # the default in grub.cfg, the pinned default file and the step's state file are unchanged
+        self.assertEqual(osapply.Kernel.default_kver(f.cfg), C_RUN)
+        self.assertEqual(f.tree[osapply.KERNEL_DEFAULT_CFG], pin0)
+        self.assertEqual(f.tree[osapply.KERNEL_STATE], state0)
+        self.assertEqual(f.env, {"felhom_next": ""})
+        self.assertEqual(f.reboots, [], "the clean-up never reboots")
+        self.assertEqual(f.update_grubs, 0, "the wrapper itself runs no update-grub (the kernel's postrm did)")
+        # leftovers are reported, never deleted
+        self.assertEqual(c["leftovers"], [f"/lib/modules/{k}" for k in ("7.0.14-20-pve", "7.0.14-22-pve", "7.0.2-6-pve")])
+        self.assertEqual(f.modules, set(K), "a left-over folder is reported, not deleted")
+
+    # --- scenario 2: the second night ---
+    def test_s2_the_second_night_removes_nothing(self):
+        f = cfake()
+        self.assertEqual(run(f)[0], 0)
+        f.calls.clear()
+        f.purges.clear()
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(rep["kernel_cleanup"]["outcome"], "nothing")
+        self.assertEqual(rep["kernel_cleanup"]["removal"], [])
+        self.assertEqual(f.purges, [])
+        self.assertEqual(f.changing(), [], "no apt call that changes anything")
+        self.assertFalse([c for c in f.calls if "purge" in c[-1]], "not even a simulation: there is nothing to simulate")
+
+    # --- scenario 3: a step is not over (P1) ---
+    def test_s3_a_step_that_is_not_over_refuses(self):
+        for phase in ("staged", "oneshot", "judging", "reverting", "some-new-phase"):
+            f = cstate(cfake(), phase=phase)
+            rep = self.refused(f, "P1")
+            self.assertIn(phase, rep["refused"]["reason"])
+
+    # --- scenario 4: the flag (P2); the pinned default against grub.cfg (P3) ---
+    def test_s4_the_flag_names_a_kernel_refuses(self):
+        f = cfake(env={"felhom_next": "7.0.14-26-pve"})
+        self.refused(f, "P2")
+
+    def test_s4_pinned_default_and_grub_cfg_disagree_refuses(self):
+        f = cfake()
+        f.tree[osapply.KERNEL_DEFAULT_CFG] = f'GRUB_DEFAULT="gnulinux-advanced-{U}>gnulinux-{C_FROM}-advanced-{U}"\n'
+        self.refused(f, "P3")       # grub.cfg still says 28, the file says 26
+        g = cfake()
+        del g.tree[osapply.KERNEL_DEFAULT_CFG]
+        self.refused(g, "P3")       # no pinned file at all
+        h = cfake()                  # both name a kernel that is not installed
+        h.tree[osapply.KERNEL_DEFAULT_CFG] = f'GRUB_DEFAULT="gnulinux-advanced-{U}>gnulinux-7.0.14-99-pve-advanced-{U}"\n'
+        h.cfg = h.cfg.replace(f"gnulinux-{C_RUN}-advanced-{U}\"", f"gnulinux-7.0.14-99-pve-advanced-{U}\"", 1)
+        self.refused(h, "P3")
+
+    # --- scenario 5: the box fell back / does not run its newest kernel ---
+    def test_s5_a_box_that_fell_back_keeps_every_kernel(self):
+        for phase in ("fell_back", "self_reverted", "revert_failed", "cancelled"):
+            self.not_run(cstate(cfake(running=C_FROM), phase=phase), "step-not-good")
+
+    def test_s5_a_box_that_does_not_run_its_newest_kernel_keeps_every_kernel(self):
+        f = cfake(running=C_FROM)    # the step says good, but the box runs the older kernel
+        self.not_run(f, "not-on-newest")
+
+    # --- scenario 6: the wrong case (K1) — one case per kept-kernel kind ---
+    def k1(self, kind, prepare):
+        """The keep computation is BROKEN on purpose (it forgets `kind`); K1 reads the box by itself and must refuse
+        before any apt call."""
+        f = cfake()
+        victim = prepare(f)
+        real = osapply.Kernel.cleanup_keep
+
+        def broken(self_, images, sources):
+            keep = real(self_, images, [(k, why) for k, why in sources if k != victim])
+            keep.pop(victim, None)       # … also when it is the newest installed
+            return keep
+        osapply.Kernel.cleanup_keep = broken
+        try:
+            rc, rep = run(f)
+        finally:
+            osapply.Kernel.cleanup_keep = real
+        self.assertEqual(rc, 2, f"{kind}: {rep}")
+        self.assertEqual(rep["refused"]["code"], "K1", f"{kind}: {rep}")
+        self.assertIn(img(victim), rep["refused"]["reason"])
+        self.assertFalse([c for c in f.calls if "purge" in c[-1]], f"{kind}: refused BEFORE any apt call, also the simulation")
+        self.nothing_changed(f)
+
+    def test_s6_k1_the_running_kernel(self):
+        self.k1("running", lambda f: C_RUN)
+
+    def test_s6_k1_the_steps_from(self):
+        self.k1("from", lambda f: C_FROM)
+
+    def test_s6_k1_the_steps_to(self):
+        def prep(f):       # a step whose `to` is not the running kernel any more
+            cstate(f, to="7.0.14-22-pve")
+            return "7.0.14-22-pve"
+        self.k1("to", prep)
+
+    def test_s6_k1_default_before(self):
+        def prep(f):
+            cstate(f, default_before="7.0.14-20-pve")
+            return "7.0.14-20-pve"
+        self.k1("default_before", prep)
+
+    def test_s6_k1_the_grub_cfg_default_and_the_pinned_file(self):
+        # the default (both places) names 26; the box runs 28 after a hand boot. The keep computation forgets 26.
+        def prep(f):
+            f.tree[osapply.KERNEL_DEFAULT_CFG] = f'GRUB_DEFAULT="gnulinux-advanced-{U}>gnulinux-7.0.14-22-pve-advanced-{U}"\n'
+            f.cfg = f.render()
+            return "7.0.14-22-pve"
+        self.k1("default", prep)
+
+    def test_s6_k1_the_flag(self):
+        # a flag is a refusal by itself (P2, first); with the keep computation broken, K1 is in the SAME refusal
+        f = cfake(env={"felhom_next": "7.0.14-20-pve"})
+        real = osapply.Kernel.cleanup_keep
+        osapply.Kernel.cleanup_keep = lambda s_, images, sources: real(s_, images, [(k, w) for k, w in sources if k != "7.0.14-20-pve"])
+        try:
+            rc, rep = run(f)
+        finally:
+            osapply.Kernel.cleanup_keep = real
+        self.assertEqual((rc, rep["refused"]["code"]), (2, "P2"), rep)
+        self.assertIn("K1 " + img("7.0.14-20-pve") + " is a keep-kernel (the one-shot flag)", rep["refused"]["reason"])
+        self.assertFalse([c for c in f.calls if "purge" in c[-1]])
+        self.nothing_changed(f)
+
+    def test_s6_k1_an_operators_boot_pin(self):
+        def prep(f):
+            f.tree[osapply.KERNEL_PIN_FILE] = "7.0.14-20-pve\n"
+            return "7.0.14-20-pve"
+        self.k1("boot pin", prep)
+
+    def test_s6_every_kept_kind_is_in_the_keep_set_and_never_in_the_removal_set(self):
+        f = cfake()
+        cstate(f, default_before="7.0.14-20-pve")
+        f.tree[osapply.KERNEL_PIN_FILE] = "7.0.14-22-pve\n"
+        f.plan["dry_run"] = True
+        rc, rep = run(f)
+        c = rep["kernel_cleanup"]
+        self.assertEqual(sorted(c["kept"]), sorted(["7.0.14-20-pve", "7.0.14-22-pve", C_FROM, C_RUN]))
+        self.assertEqual(c["removal"], [img("7.0.2-6-pve")])
+
+    def test_s6_the_running_kernel_is_not_an_installed_package(self):
+        f = cfake()
+        del f.installed[img(C_RUN)]      # the box runs a kernel dpkg does not know (an unsigned hand install, say)
+        f.installed["proxmox-kernel-7.0.14-28-pve-other"] = "7.0.14-28"
+        self.refused(f, "P3")            # P3 is first (the default names a kernel that is not an installed image) …
+        rc, rep = run(f)
+        self.assertIn("K1", rep["refused"]["reason"], "… and K1 is in the same refusal")
+
+    def test_no_step_record_keeps_the_next_lower_kernel(self):
+        f = cstate(cfake(), **{"from": None, "to": None, "default_before": None})
+        f.plan["dry_run"] = True
+        rc, rep = run(f)
+        self.assertIn("7.0.14-26-pve", rep["kernel_cleanup"]["kept"])
+        self.assertIn("next lower", " ".join(rep["kernel_cleanup"]["kept"]["7.0.14-26-pve"]))
+
+    # --- scenario 7: apt would do more (S1, S2) ---
+    def test_s7_the_simulation_would_remove_something_else(self):
+        f = cfake(purge_extra_sim=["Purg proxmox-kernel-7.0 [7.0.14-28]", "Remv proxmox-ve [9.2.0]"])
+        rep = self.refused(f, "S2")
+        self.assertIn("proxmox-ve", rep["refused"]["reason"])
+
+    def test_s7_the_simulation_would_install_something(self):
+        f = cfake(purge_extra_sim=["Inst proxmox-kernel-7.0.14-28-pve (7.0.14-28 Proxmox Debian Repository:stable [amd64])"])
+        rep = self.refused(f, "S1")
+        self.assertIn("proxmox-kernel-7.0.14-28-pve", rep["refused"]["reason"])
+
+    def test_s7_the_simulation_fails(self):
+        self.refused(cfake(purge_sim_rc=100), "S1")
+
+    # --- scenario 8: not our box ---
+    def test_s8_a_byo_host_is_refused_by_the_appliance_rule(self):
+        f = cfake()
+        f.files[osapply.INSTALL_STATE] = json.dumps({"mode": "byo"})
+        self.refused(f, "R12")
+
+    def test_s8_a_box_without_the_ring0_mark_does_not_run(self):
+        f = cfake()
+        f.files[osapply.TRUST_FILE] = json.dumps({"host_id": "tester-1-d70be4", "ring0_slow_lane": False})
+        rep = self.not_run(f, "no-ring0-mark")
+        self.assertIn("ring 0 only in this release", rep["kernel_cleanup"]["not_run"][0]["reason"])
+        g = cfake()      # an agent-writable trust file proves nothing
+        g.stats[osapply.TRUST_FILE] = St(mode=statmod.S_IFREG | 0o666, uid=0)
+        self.not_run(g, "no-ring0-mark")
+        h = cfake()      # the agent's plan cannot claim the ring
+        h.files[osapply.TRUST_FILE] = json.dumps({"host_id": "x", "ring0_slow_lane": False})
+        h.plan["ring"] = 0
+        self.not_run(h, "no-ring0-mark")
+
+    # --- scenario 9: a plan that names packages ---
+    def test_s9_cleanup_plan_names_refused(self):
+        for extra in ({"packages": [{"name": img("7.0.14-28-pve"), "version": "7.0.14-28", "origin": "Proxmox Debian Repository"}]},
+                      {"select": "pending-kernel"}, {"allow_new": ["x"]}, {"signed": {"blob_b64": "eA==", "sig": "x"}},
+                      {"expect_kver": "7.0.14-20-pve"}, {"dry_run": "yes"}):
+            f = cfake()
+            f.plan.update(extra)
+            self.refused(f, "C1")
+
+    # --- scenario 10: no kernel step ever ran ---
+    def test_s10_no_state_file_does_not_run(self):
+        f = cfake()
+        f.state = None
+        f.write_state()
+        rep = self.not_run(f, "no-step")
+        self.assertIn("no kernel step yet", rep["kernel_cleanup"]["not_run"][0]["reason"])
+
+    # --- scenario 11: a power cut in the purge ---
+    def test_s11_a_cut_purge_is_repaired_first_and_the_rule_is_computed_again(self):
+        f = cfake(purge_rc=100)
+        rc, rep = run(f)
+        self.assertEqual(rc, 3, rep)
+        self.assertEqual(rep["failed"]["step"], "purge")
+        self.assertEqual(rep["kernel_cleanup"]["outcome"], "failed")
+        # the cut removed ONE of the three before it stopped
+        f.installed.pop(self.REMOVE[0])
+        f.boot.discard("7.0.14-20-pve")
+        f.purge_rc, f.purges = 0, []
+        f.calls.clear()
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(rep["repair"]["journal_before"], 1, "repair first: dpkg's journal was replayed")
+        i_repair = [i for i, c in enumerate(f.calls) if c[-1][-3:-1] == ["--configure", "-a"] or "--configure" in c[-1]][0]
+        i_purge = [i for i, c in enumerate(f.calls) if "purge" in c[-1]][0]
+        self.assertLess(i_repair, i_purge, "the repair runs BEFORE the rule is computed and before any purge")
+        self.assertEqual(f.purges, [self.REMOVE[1:]], "the rule was computed again from the box: the two that are left")
+        self.assertEqual(rep["kernel_cleanup"]["outcome"], "removed")
+
+    # --- the 20-hour window, the dry run, after-the-purge proof ---
+    def test_inside_a_steps_20_hour_window_it_does_not_run(self):
+        f = cfake()
+        cstate(f, result_at=osapply._iso(f.clock - 3600))      # the step ended good one hour ago
+        self.not_run(f, "window")
+
+    def test_the_dry_run_reports_and_changes_nothing(self):
+        f = cfake()
+        f.plan["dry_run"] = True
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        c = rep["kernel_cleanup"]
+        self.assertEqual((c["outcome"], c["would"], c["removal"]), ("dry_run", "remove", self.REMOVE))
+        self.assertEqual(sorted(c["kept"]), sorted([C_FROM, C_RUN]))
+        self.nothing_changed(f)
+        self.assertNotIn("repair", rep, "a dry run does not even repair")
+        # a dry run of a refusal says which one, with exit 0
+        g = cstate(cfake(), phase="judging")
+        g.plan["dry_run"] = True
+        rc, rep = run(g)
+        self.assertEqual((rc, rep["kernel_cleanup"]["would"]), (0, "refuse"), rep)
+        self.assertEqual([x["code"] for x in rep["kernel_cleanup"]["refusals"]], ["P1"])
+        self.nothing_changed(g)
+        # … and of a box that would not run
+        h = cfake()
+        h.files[osapply.TRUST_FILE] = json.dumps({"host_id": "x", "ring0_slow_lane": False})
+        h.plan["dry_run"] = True
+        rc, rep = run(h)
+        self.assertEqual((rc, rep["kernel_cleanup"]["would"], rep["kernel_cleanup"]["removal"]), (0, "not_run", self.REMOVE), rep)
+        self.nothing_changed(h)
+
+    def test_a_purge_that_moves_the_default_is_reported_failed_loudly(self):
+        f = cfake(purge_moves_default=True)      # the pinned file vanished in the purge: 10_linux makes the newest the default
+        rc, rep = run(f)
+        self.assertEqual(rc, 3, rep)
+        self.assertEqual(rep["failed"]["step"], "verify")
+        self.assertIn("pinned default file changed", rep["failed"]["reason"])
+        self.assertEqual(f.update_grubs, 0, "never repaired by hand in code")
+        self.assertTrue(any("FAILED" in l and "verify" in l for l in f.logs))
+
+    def test_the_apt_lock_is_respected(self):
+        self.refused(cfake(lock_held=True), "R9")
+
+    def test_both_image_packages_of_one_old_kernel_go_together(self):
+        f = cfake()
+        f.installed["proxmox-kernel-7.0.14-20-pve"] = "7.0.14-20"     # the unsigned image beside the signed one
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertIn("proxmox-kernel-7.0.14-20-pve", f.purges[0])
+        self.assertIn(img("7.0.14-20-pve"), f.purges[0])
+
+    def test_cleanup_needs_no_new_sudo_line(self):
+        """The wrapper is reached by `--plan` only: the new mode rides the one sudo line the agent already has."""
+        sudo = (HERE / "felhom-agent.sudoers").read_text()
+        live = [l.strip().rstrip("\\,").strip() for l in sudo.splitlines() if l.strip() and not l.strip().startswith("#")]
+        self.assertEqual([l for l in live if "felhom-os-apply" in l and "Cmnd_Alias" not in l and "NOPASSWD" not in l],
+                         [osapply.SELF_CHECK_SUDO_LINE], "exactly the one --plan line reaches the wrapper")
+        self.assertFalse([l for l in live if "kernel-cleanup" in l or
+                          ("apt" in l and ("purge" in l or "remove" in l))], "no sudo line removes a package")
+
+
+
+# ---------- the security-profile refresh (R-1388, R-1301, `11` §5.12) ----------
+class PFake(Fake):
+    """A host with two AppArmor cache folders. `stale`: profile files whose cached copy in the running kernel's folder
+    is an old compile; the refresh command rewrites the cache from `fresh`. lxc-containers compiles differently every
+    time (measured on demo-hp). `failed`: unit -> ExecStart program."""
+
+    def __init__(self, stale=("usr.sbin.chronyd",), failed=None, kernel="7.0.14-28-pve", record=None):
+        super().__init__()
+        self.kernel = kernel
+        self.fresh = {"usr.sbin.chronyd": b"chrony-4.8", "usr.bin.man": b"man", "lxc-containers": b"lxc-0", "nvidia_modprobe": b"nv"}
+        self.names = {"usr.sbin.chronyd": "/usr/sbin/chronyd", "usr.bin.man": "/usr/bin/man\nman_filter\nman_groff",
+                      "lxc-containers": "lxc-container-default\nlxc-container-default-cgns", "nvidia_modprobe": "nvidia_modprobe\nnvidia_modprobe//kmod"}
+        self.cache = {"ae1a4050.0": {k: b"old-" + k.encode() for k in self.fresh}, "f6da0254.0": dict(self.fresh)}
+        for k in stale:
+            self.cache["f6da0254.0"][k] = b"OLD-COPY-" + k.encode()
+        self.failed = dict({"chrony.service": "/usr/sbin/chronyd"} if failed is None else failed)
+        self.restarts, self.parser_runs, self.parser_rc, self.tree = [], 0, 0, {}
+        if record is not None:
+            self.tree[osapply.PROFILE_STATE] = json.dumps(record)
+        self.plan = {"release_id": "profile-refresh", "layer": "host", "mode": "profile-refresh"}
+
+    def isdir(self, p):
+        return p in (osapply.PROFILE_DIR, osapply.PROFILE_CACHE) or p.startswith(osapply.PROFILE_CACHE + "/") and p.count("/") == 4
+
+    def list_dir(self, p):
+        if p == osapply.PROFILE_CACHE:
+            return sorted(self.cache)
+        d = p.rsplit("/", 1)[1]
+        return sorted(list(self.cache.get(d, {})) + [".features"])
+
+    def read_bytes(self, p):
+        d, f = p.split("/")[-2:]
+        return self.cache[d][f]
+
+    def put_file(self, path, data, mode):
+        self.tree[path] = data.decode()
+
+    def read_file(self, p):
+        if p in self.tree:
+            return self.tree[p]
+        return super().read_file(p)
+
+    def host(self, argv, timeout=600, stdin=None):
+        if argv[0] == "uname":
+            self.calls.append(("host", argv))
+            return 0, self.kernel + "\n", ""
+        if argv == osapply.PROFILE_REFRESH_CMD:
+            self.calls.append(("host", argv))
+            self.parser_runs += 1
+            if self.parser_rc:
+                return self.parser_rc, "", "AppArmor parser error for /etc/apparmor.d/x"
+            self.fresh["lxc-containers"] = b"lxc-%d" % self.parser_runs     # never the same binary twice
+            self.cache["f6da0254.0"] = dict(self.fresh)
+            return 0, "", ""
+        if argv[:2] == ["apparmor_parser", "-N"]:
+            self.calls.append(("host", argv))
+            return 0, self.names[argv[2].rsplit("/", 1)[1]] + "\n", ""
+        if argv[:2] == ["systemctl", "list-units"]:
+            self.calls.append(("host", argv))
+            return 0, "".join(f"{u} loaded failed failed x\n" for u in self.failed), ""
+        if argv[:2] == ["systemctl", "show"]:
+            self.calls.append(("host", argv))
+            return 0, f"{{ path={self.failed[argv[-1]]} ; argv[]={self.failed[argv[-1]]} -F 1 ; ignore_errors=no }}\n", ""
+        if argv[:2] == ["systemctl", "restart"]:
+            self.calls.append(("host", argv))
+            self.restarts.append(argv[2])
+            return 0, "", ""
+        return super().host(argv, timeout, stdin)
+
+
+class ProfileRefresh(unittest.TestCase):
+    """Red-proofs: felhom.eu audits/release-2a-2026-11-14/B/redproof.txt."""
+
+    def test_a_stale_profile_is_refreshed_and_only_its_failed_unit_restarts(self):
+        f = PFake(failed={"chrony.service": "/usr/sbin/chronyd", "postfix.service": "/usr/sbin/postfix", "other.service": "/usr/bin/man-not"})
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        p = rep["profile_refresh"]
+        self.assertEqual(p["outcome"], "refreshed")
+        self.assertEqual(f.parser_runs, 1)
+        self.assertIn(("host", ["apparmor_parser", "--replace", "--skip-read-cache", "--write-cache", "--", "/etc/apparmor.d"]), f.calls)
+        self.assertEqual(f.cache["f6da0254.0"]["usr.sbin.chronyd"], b"chrony-4.8", "the cache holds the fresh compile")
+        self.assertEqual(p["rewritten"], ["f6da0254.0/lxc-containers", "f6da0254.0/usr.sbin.chronyd"])
+        self.assertEqual(f.restarts, ["chrony.service"], "ONLY the failed unit whose own program's profile was stale")
+        self.assertEqual([x["unit"] for x in p["restarted"]], ["chrony.service"])
+        rec = json.loads(f.tree[osapply.PROFILE_STATE])
+        self.assertEqual(rec["kernel"], "7.0.14-28-pve")
+
+    def test_a_fresh_cache_changes_nothing_and_restarts_nothing(self):
+        f = PFake(stale=())                      # chrony is failed for another reason; its profile is NOT stale
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(rep["profile_refresh"]["rewritten"], ["f6da0254.0/lxc-containers"], "only the file that never compiles the same")
+        self.assertEqual(f.restarts, [], "a failed unit whose profile was not stale is NOT restarted")
+        self.assertEqual(f.cache["f6da0254.0"]["usr.sbin.chronyd"], b"chrony-4.8")
+
+    def test_a_stale_profile_of_a_running_program_restarts_nothing(self):
+        f = PFake(failed={})                     # the unit is active: the kernel swaps the profile of a running task itself
+        rc, rep = run(f)
+        self.assertEqual((rc, f.restarts, rep["profile_refresh"]["outcome"]), (0, [], "refreshed"), rep)
+
+    def test_a_failed_unit_of_another_program_is_not_restarted(self):
+        f = PFake(stale=("usr.bin.man",), failed={"chrony.service": "/usr/sbin/chronyd"})
+        rc, rep = run(f)
+        self.assertEqual((rc, f.restarts), (0, []), rep)
+        g = PFake(stale=("nvidia_modprobe",), failed={"nvidia.service": "nvidia_modprobe"})   # a profile name that is not a path
+        rc, rep = run(g)
+        self.assertEqual((rc, g.restarts), (0, []), rep)
+
+    def test_once_per_kernel(self):
+        f = PFake()
+        self.assertEqual(run(f)[0], 0)
+        f.calls.clear()
+        f.cache["f6da0254.0"]["usr.sbin.chronyd"] = b"OLD-COPY"     # even with a stale file: this kernel was done
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["profile_refresh"]["outcome"], f.parser_runs), (0, "not_needed", 1), rep)
+        self.assertFalse([c for c in f.calls if c[1][0] in ("apparmor_parser", "systemctl")], "no parser, no systemctl")
+        f.kernel = "7.0.14-29-pve"               # the box starts on another kernel: again
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["profile_refresh"]["outcome"], f.parser_runs), (0, "refreshed", 2), rep)
+        self.assertEqual(json.loads(f.tree[osapply.PROFILE_STATE])["kernel"], "7.0.14-29-pve")
+
+    def test_a_failed_parser_writes_no_record_so_the_next_start_tries_again(self):
+        f = PFake()
+        f.parser_rc = 1
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["failed"]["step"], rep["profile_refresh"]["outcome"]), (3, "profile-refresh", "failed"), rep)
+        self.assertNotIn(osapply.PROFILE_STATE, f.tree)
+        self.assertEqual(f.restarts, [])
+        f.parser_rc = 0
+        self.assertEqual(run(f)[1]["profile_refresh"]["outcome"], "refreshed")
+
+    def test_the_plan_names_nothing(self):
+        for extra in ({"profiles": ["usr.sbin.chronyd"]}, {"units": ["ssh.service"]}, {"packages": [{"name": "chrony", "version": "4.8", "origin": "Debian"}]},
+                      {"force": True}, {"dry_run": 1}, {"layer": "guest"}):
+            f = PFake()
+            f.plan.update(extra)
+            rc, rep = run(f)
+            self.assertEqual(rc, 2, rep)
+            self.assertIn(rep["refused"]["code"], ("F1", "R11"), rep)
+            self.assertEqual((f.parser_runs, f.restarts), (0, []))
+
+    def test_a_byo_host_is_refused(self):
+        f = PFake()
+        f.files[osapply.INSTALL_STATE] = json.dumps({"mode": "byo"})
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["refused"]["code"], f.parser_runs), (2, "R12", 0), rep)
+
+    def test_the_dry_run_changes_nothing(self):
+        f = PFake()
+        f.plan["dry_run"] = True
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["profile_refresh"]["outcome"], rep["profile_refresh"]["would"]), (0, "dry_run", "refresh"), rep)
+        self.assertEqual((f.parser_runs, f.restarts), (0, []))
+        self.assertNotIn(osapply.PROFILE_STATE, f.tree)
+
+    def test_no_more_than_three_restarts(self):
+        f = PFake(failed={f"chrony{i}.service": "/usr/sbin/chronyd" for i in range(5)})
+        rc, rep = run(f)
+        self.assertEqual((rc, len(f.restarts), len(rep["profile_refresh"]["not_restarted"])), (0, 3, 2), rep)
+
+    def test_profile_refresh_changes_no_package(self):
+        f = PFake()
+        run(f)
+        self.assertFalse([c for c in f.calls if c[1][0] in ("apt-get", "dpkg", "apt")], "no package command at all")
+
+
+# ---------- trust in Proxmox by a copy of its root CA (R-1380, D80, felhom.eu `03` §12a) ----------
+CA_PEM = b"-----BEGIN CERTIFICATE-----\nQUJDREVGR0hJSktMTU5PUFFSU1RVVldYWVo=\n-----END CERTIFICATE-----\n"
+CA_PEM_2 = b"-----BEGIN CERTIFICATE-----\nWllYV1ZVVFNSUVBPTk1MS0pJSEdGRURDQkE=\n-----END CERTIFICATE-----\n"
+
+
+class TFake(Fake):
+    """A Proxmox host: the root CA file, the copy (or none), and openssl's answer about the bytes it is given."""
+
+    def __init__(self, source=CA_PEM, copy=None, ca=True, ends="Aug 17 15:44:42 2036 GMT"):
+        super().__init__()
+        self.source, self.tree, self.ca, self.ends, self.openssl_rc = source, {}, ca, ends, 0
+        if copy is not None:
+            self.tree[osapply.PVE_CA_COPY] = copy
+        self.openssl_stdin, self.puts = [], []
+        self.serves_new = False        # does pveproxy serve a certificate that verifies with the NEW root alone?
+        self.dead_ca = set()           # certificate texts openssl says are not a live CA (ended)
+        self.handshakes = []
+        self.plan = {"release_id": "pve-ca", "layer": "host", "mode": "pve-ca-copy"}
+
+    def read_bytes(self, p):
+        if p == osapply.PVE_CA_SOURCE:
+            if self.source is None:
+                raise FileNotFoundError(p)
+            return self.source
+        if p in self.tree:
+            return self.tree[p]
+        raise FileNotFoundError(p)
+
+    def put_file(self, path, data, mode):
+        self.puts.append((path, data, mode))
+        self.tree[path] = data
+
+    def host(self, argv, timeout=600, stdin=None):
+        if argv[0] == "openssl" and argv[1] == "s_client":
+            self.handshakes.append(argv)
+            return (0, "Verify return code: 0 (ok)\n", "") if self.serves_new else (1, "", "Verify return code: 19 (self-signed certificate in certificate chain)\n")
+        if argv[0] == "openssl" and "-checkend" in argv:
+            return (1, "X509v3 Basic Constraints: critical\n    CA:TRUE\nCertificate will expire\n", "") if stdin.strip() in self.dead_ca else \
+                (0, "X509v3 Basic Constraints: critical\n    CA:TRUE\nCertificate will not expire\n", "")
+        if argv[0] == "openssl":
+            self.calls.append(("host", argv))
+            self.openssl_stdin.append(stdin)
+            if self.openssl_rc:
+                return self.openssl_rc, "", "Could not read certificate from <stdin>"
+            return 0, ("X509v3 Basic Constraints: critical\n    CA:%s\nnotAfter=%s\n" % ("TRUE" if self.ca else "FALSE", self.ends)), ""
+        return super().host(argv, timeout, stdin)
+
+
+class PVECACopy(unittest.TestCase):
+    """Red-proofs: felhom.eu audits/night-build-2026-11-14/A/redproof-wrapper.txt."""
+
+    def refused(self, f, code):
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["refused"]["code"]), (2, code), rep)
+        self.assertEqual(f.puts, [], "nothing is written on a refusal")
+        return rep
+
+    def test_a_new_box_gets_the_copy(self):
+        f = TFake()
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(f.puts, [(osapply.PVE_CA_COPY, CA_PEM, 0o644)], "ONE file, the fixed place, the source's bytes, 0644")
+        self.assertEqual(f.openssl_stdin, [CA_PEM.decode()], "the bytes that are checked are the bytes that are written")
+        c = rep["pve_ca_copy"]
+        self.assertEqual((c["outcome"], c["path"], c["replaced"]), ("written", "/etc/felhom-agent/pve-root-ca.pem", False))
+        self.assertEqual(c["not_after"], "2036-08-17T15:44:42Z")
+        self.assertGreater(c["days_left"], 3000)
+
+    def test_the_same_copy_is_not_written_again(self):
+        f = TFake(copy=CA_PEM)
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["pve_ca_copy"]["outcome"], f.puts), (0, "same", []), rep)
+
+    def test_a_renewed_ca_refreshes_the_copy(self):
+        f = TFake(source=CA_PEM_2, copy=CA_PEM)
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["pve_ca_copy"]["outcome"], rep["pve_ca_copy"]["replaced"]), (0, "written", True), rep)
+        # the new certificate FIRST, and the copy's previous one kept behind it: Proxmox serves its old certificate until
+        # pveproxy is reloaded, and an agent that restarts in that window must still have a root that verifies it
+        bundle = f.tree[osapply.PVE_CA_COPY]
+        self.assertEqual(bundle, CA_PEM_2 + osapply.PEM_CERT_RE.findall(CA_PEM)[0] + b"\n")
+        self.assertTrue(rep["pve_ca_copy"]["kept_previous"])
+        self.assertEqual(f.openssl_stdin, [CA_PEM_2.decode()], "the SOURCE is checked in full; the kept one by its own check")
+
+    def test_the_previous_root_stays_only_for_the_changeover(self):
+        old = osapply.PEM_CERT_RE.findall(CA_PEM)[0]
+        bundle = CA_PEM_2 + old + b"\n"
+        # Proxmox still serves the OLD certificate: nothing is written, the old root stays
+        f = TFake(source=CA_PEM_2, copy=bundle)
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["pve_ca_copy"]["outcome"], f.puts), (0, "same", []), rep)
+        self.assertTrue(rep["pve_ca_copy"]["kept_previous"])
+        self.assertEqual(len(f.handshakes), 1)
+        self.assertIn(osapply.PVE_CA_SOURCE, f.handshakes[0], "the handshake is verified with the NEW root alone (the source file)")
+        self.assertIn("-no-CApath", f.handshakes[0])
+        self.assertIn("-verify_return_error", f.handshakes[0])
+        # pveproxy was reloaded and serves a certificate of the NEW root: the old root is dropped
+        f = TFake(source=CA_PEM_2, copy=bundle)
+        f.serves_new = True
+        rc, rep = run(f)
+        self.assertEqual(rc, 0, rep)
+        self.assertEqual(f.tree[osapply.PVE_CA_COPY], CA_PEM_2, "after the changeover the copy is the new root alone")
+        self.assertEqual((rep["pve_ca_copy"]["outcome"], rep["pve_ca_copy"]["dropped_previous"], rep["pve_ca_copy"]["kept_previous"]), ("written", True, False))
+        # the old root has ENDED while Proxmox still serves the old certificate: it verifies nothing, so it is dropped
+        f = TFake(source=CA_PEM_2, copy=bundle)
+        f.dead_ca = {old.decode()}
+        rc, rep = run(f)
+        self.assertEqual(f.tree[osapply.PVE_CA_COPY], CA_PEM_2)
+        # an ended previous root is not kept at the change itself either
+        f = TFake(source=CA_PEM_2, copy=CA_PEM)
+        f.dead_ca = {old.decode()}
+        rc, rep = run(f)
+        self.assertEqual(f.tree[osapply.PVE_CA_COPY], CA_PEM_2)
+        self.assertFalse(rep["pve_ca_copy"]["kept_previous"])
+
+    def test_a_copy_with_anything_unchecked_in_it_is_rewritten(self):
+        old = osapply.PEM_CERT_RE.findall(CA_PEM)[0]
+        for name, copy in (("text after the source", CA_PEM_2 + b"anything at all\n"),
+                           ("a third certificate", CA_PEM_2 + old + b"\n" + old + b"\n"),
+                           ("a key in the file", CA_PEM_2 + b"-----BEGIN PRIVATE KEY-----\nAAAA\n-----END PRIVATE KEY-----\n"),
+                           ("the two in the wrong order", CA_PEM + osapply.PEM_CERT_RE.findall(CA_PEM_2)[0] + b"\n")):
+            f = TFake(source=CA_PEM_2, copy=copy)
+            f.serves_new = True
+            rc, rep = run(f)
+            self.assertEqual(rc, 0, (name, rep))
+            got = f.tree[osapply.PVE_CA_COPY]
+            self.assertTrue(got.startswith(CA_PEM_2), name)
+            self.assertNotIn(b"anything", got, name)
+            self.assertNotIn(b"PRIVATE KEY", got, name)
+            self.assertLessEqual(len(osapply.PEM_CERT_RE.findall(got)), 2, name)
+        # a first copy keeps nothing behind it, and makes no handshake
+        f4 = TFake()
+        rc, rep = run(f4)
+        self.assertFalse(rep["pve_ca_copy"]["kept_previous"])
+        self.assertEqual((f4.tree[osapply.PVE_CA_COPY], f4.handshakes), (CA_PEM, []))
+
+    def test_a_source_that_is_not_a_ca_certificate_is_refused(self):
+        self.refused(TFake(ca=False), "T3")                                  # a leaf certificate
+        self.refused(TFake(source=b"hello\n"), "T3")                         # not a certificate at all
+        self.refused(TFake(source=CA_PEM + CA_PEM_2), "T3")                  # two certificates (a chain, a bundle)
+        self.refused(TFake(source=CA_PEM + b"-----BEGIN PRIVATE KEY-----\nAAAA\n-----END PRIVATE KEY-----\n"), "T3")
+        self.refused(TFake(source=b""), "T3")
+        self.refused(TFake(source=CA_PEM + b"x" * (osapply.PVE_CA_MAX_BYTES + 1)), "T3")
+        f = TFake()
+        f.openssl_rc = 1                                                     # openssl cannot parse it
+        self.refused(f, "T3")
+        self.refused(TFake(ends="Aug 17 15:44:42 2020 GMT"), "T3")           # an ended CA
+        self.refused(TFake(ends="garbage"), "T3")
+
+    def test_no_source_is_refused(self):
+        self.refused(TFake(source=None), "T2")
+
+    def test_with_the_real_openssl_and_the_real_runner(self):
+        """NOT a fake: a CA and a leaf made by openssl in a temp folder, judged through Runner.host as on a box (the
+        fake cannot show a wrong argument or a bytes/text mix-up). Skipped where openssl is missing."""
+        import shutil, tempfile
+        if not shutil.which("openssl"):
+            self.skipTest("no openssl here")
+        d = tempfile.mkdtemp()
+        try:
+            def sh(*a):
+                subprocess.run(a, check=True, capture_output=True, cwd=d)
+            sh("openssl", "req", "-x509", "-newkey", "rsa:2048", "-nodes", "-keyout", "ca.key", "-out", "ca.pem", "-days", "30",
+               "-subj", "/CN=test CA", "-addext", "basicConstraints=critical,CA:TRUE")
+            sh("openssl", "req", "-x509", "-newkey", "rsa:2048", "-nodes", "-keyout", "leaf.key", "-out", "leaf.pem", "-days", "30",
+               "-subj", "/CN=leaf", "-addext", "basicConstraints=CA:FALSE")
+            real = osapply.Runner()
+            for name, want in (("ca.pem", "written"), ("leaf.pem", "T3")):
+                data = open(os.path.join(d, name), "rb").read()
+                f = TFake(source=data)
+                f.host = lambda argv, timeout=600, stdin=None: real.host(argv, timeout, stdin) if argv[0] == "openssl" else Fake.host(f, argv, timeout, stdin)
+                f.clock = __import__("time").time()
+                rc, rep = run(f)
+                got = rep["pve_ca_copy"]["outcome"] if rc == 0 else rep["refused"]["code"]
+                self.assertEqual(got, want, f"{name}: {rep}")
+                if want == "written":
+                    self.assertIn(rep["pve_ca_copy"]["days_left"], (29, 30))
+            # the two checks the changeover uses, with the REAL openssl and a REAL TLS server
+            ended = True
+            try:                                        # an ENDED CA needs openssl 3.4+ (-not_after); older: that one check is not made
+                sh("openssl", "req", "-x509", "-newkey", "rsa:2048", "-nodes", "-keyout", "old.key", "-out", "old.pem",
+                   "-not_before", "20190101000000Z", "-not_after", "20200101000000Z",
+                   "-subj", "/CN=ended CA", "-addext", "basicConstraints=critical,CA:TRUE")
+            except subprocess.CalledProcessError:
+                ended = False
+            sh("openssl", "req", "-newkey", "rsa:2048", "-nodes", "-keyout", "srv.key", "-out", "srv.csr", "-subj", "/CN=pve")
+            sh("openssl", "x509", "-req", "-in", "srv.csr", "-CA", "ca.pem", "-CAkey", "ca.key", "-CAcreateserial", "-out", "srv.pem", "-days", "5")
+            app = osapply.Apply.__new__(osapply.Apply)
+            app.r = real
+            block = lambda n: osapply.PEM_CERT_RE.findall(open(os.path.join(d, n), "rb").read())[0]
+            self.assertTrue(app.is_live_ca(block("ca.pem")), "a CA that has not ended")
+            self.assertFalse(app.is_live_ca(block("leaf.pem")), "a leaf is not a CA")
+            if ended:
+                self.assertFalse(app.is_live_ca(block("old.pem")), "an ended CA is not a live one")
+            import socket
+            s0 = socket.socket(); s0.bind(("127.0.0.1", 0)); port = s0.getsockname()[1]; s0.close()
+            srv = subprocess.Popen(["openssl", "s_server", "-accept", f"127.0.0.1:{port}", "-cert", "srv.pem", "-key", "srv.key", "-quiet"],
+                                   cwd=d, stdin=subprocess.PIPE, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+            try:
+                saved, osapply.PVE_LOCAL_API = osapply.PVE_LOCAL_API, f"127.0.0.1:{port}"
+                ok = False
+                for _ in range(40):                     # the server needs a moment to listen
+                    if app.pve_serves_with(os.path.join(d, "ca.pem")):
+                        ok = True
+                        break
+                    __import__("time").sleep(0.1)
+                self.assertTrue(ok, "the server's certificate is signed by ca.pem: the handshake must verify with it alone")
+                self.assertFalse(app.pve_serves_with(os.path.join(d, "leaf.pem")), "another certificate as the CA file must NOT verify")
+                osapply.PVE_LOCAL_API = "127.0.0.1:1"
+                self.assertFalse(app.pve_serves_with(os.path.join(d, "ca.pem")), "no answer is a no")
+            finally:
+                osapply.PVE_LOCAL_API = saved
+                srv.kill(); srv.wait()
+        finally:
+            shutil.rmtree(d, ignore_errors=True)
+
+    def test_a_plan_that_names_a_path_or_a_file_is_refused(self):
+        for extra in ({"source": "/etc/shadow"}, {"path": "/etc/sudoers.d/x"}, {"dest": "/root/.ssh/authorized_keys"},
+                      {"content_b64": "QQ=="}, {"packages": []}, {"file": "x"}, {"vmid": 9201}, {"dry_run": True}):
+            f = TFake()
+            f.plan.update(extra)
+            rep = self.refused(f, "T1")
+            self.assertEqual(f.openssl_stdin, [], "refused before the source is even read")
+        g = TFake()
+        g.plan["layer"] = "kernel"
+        self.refused(g, "R11")
+
+    def test_a_refused_source_keeps_an_older_copy(self):
+        f = TFake(source=b"not a certificate", copy=CA_PEM)
+        self.refused(f, "T3")
+        self.assertEqual(f.tree[osapply.PVE_CA_COPY], CA_PEM, "written whole or not at all")
+
+    def test_it_runs_on_a_byo_host_too_and_changes_nothing_else(self):
+        f = TFake()
+        f.files[osapply.INSTALL_STATE] = json.dumps({"mode": "byo"})
+        rc, rep = run(f)
+        self.assertEqual((rc, rep["pve_ca_copy"]["outcome"]), (0, "written"), rep)
+        self.assertEqual([c[1][0] for c in f.calls], ["openssl"], "one openssl call; no package, no service, no other file")
+
 if __name__ == "__main__":
     unittest.main()

go.mod

@@ -1,6 +1,6 @@
 module gitea.dooplex.hu/admin/felhom-agent
 
-go 1.27.3
+go 1.27.4
 
 require (
 	golang.org/x/crypto v0.58.0

internal/capability/manifest.go

@@ -151,6 +151,7 @@ var manifest = []Capability{
 	{"controllerswap-read", "controller-swap / managed auto-update", "/usr/sbin/pct", []string{"exec", "9201", "--", "cat", "/etc/felhom-controller-image"}, true, ""},
 	{"controllerswap-image-inspect", "controller-swap / managed auto-update", "/usr/sbin/pct", []string{"exec", "9201", "--", "docker", "image", "inspect", "gitea.dooplex.hu/admin/felhom-controller:0.0.0"}, true, ""},
 	{"controllerswap-inspect", "controller-swap / managed auto-update", "/usr/sbin/pct", []string{"exec", "9201", "--", "docker", "inspect", "-f", "{{.State.Running}}", "felhom-controller"}, true, ""},
+	{"controller-busy-read", "operator guest reboot (D65)", "/usr/sbin/pct", []string{"exec", "9201", "--", "cat", "/var/lib/docker/volumes/felhom-controller-data/_data/data/ctl/busy"}, false, ""},
 	{"controllerswap-restart", "controller-swap / managed auto-update", "/usr/sbin/pct", []string{"exec", "9201", "--", "systemctl", "restart", "felhom-controller-bootstrap.service"}, true, ""},
 	// R-861 (a) A1 (decision 165): the write goes through the ROOT verb that checks the ref; the agent has no `tee` grant.
 	{"controllerswap-write", "controller-swap / managed auto-update", "/usr/local/sbin/felhom-priv-apply", []string{"controller-image", "9201"}, true, ""},

internal/config/cutwrite.go

@@ -0,0 +1,188 @@
+package config
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// A cut write of agent.json must never leave a blind box (R-1380, 2026-11-15).
+//
+// agent.json is the agent's own file inside a ROOT-owned folder, so the agent cannot write a new file and rename it
+// over the old one: every writer truncates the file and writes it again. A power cut between the truncate and the
+// sync leaves an empty or half file, the agent does not start, and nothing can repair a box whose agent does not
+// start from the hub. So the three writers (the trust move, the PBS seed, the backup-target move) go through
+// ReplaceInPlace: the OLD bytes are first made durable in a journal file in the agent's own state folder (a new file
+// and a rename, which the agent CAN do there), then the config is written, synced and read back, then the journal is
+// removed. RecoverCutWrite runs at every start, before the config is read: a journal beside a config that does not
+// parse means the write was cut, and the old bytes go back. A journal beside a config that parses means the cut fell
+// before the truncate or after the sync — the file is whole, old or new, and only the journal is removed.
+//
+// Pinned by TestReplaceInPlace_* and TestRecoverCutWrite_* (this package) and by TestMainRecoversACutConfigWrite
+// (cmd/felhom-agent).
+
+// CutWriteDir is where the journal lives. It is NOT read from the config: the recovery runs before the config can be
+// read. /var/lib/felhom-agent is the unit's StateDirectory. A test points it at a temp folder.
+var CutWriteDir = "/var/lib/felhom-agent"
+
+// cutWriteJournal is the journal's file name: the bytes agent.json held before the write that is in flight.
+const cutWriteJournal = "agent.json.write-in-flight"
+
+// cutPoint lets a test stop a write where a power cut would (the names are the steps of ReplaceInPlace).
+var cutPoint = func(step string) error { return nil }
+
+// ReplaceInPlace writes next over the config at path, whose present content is prev. The caller holds FileMu.
+// With no journal folder there is no write: an unprotected write is the defect this exists to remove.
+func ReplaceInPlace(path string, prev, next []byte) error {
+	if !json.Valid(prev) || !json.Valid(next) {
+		return fmt.Errorf("config: refusing to write %s: the old or the new content is not JSON", path)
+	}
+	st, err := os.Stat(path)
+	if err != nil {
+		return err
+	}
+	if err := os.MkdirAll(CutWriteDir, 0o700); err != nil {
+		return fmt.Errorf("config: the journal folder: %w", err)
+	}
+	journal := filepath.Join(CutWriteDir, cutWriteJournal)
+	if err := writeDurable(journal, prev); err != nil {
+		return fmt.Errorf("config: the journal of the old content: %w", err)
+	}
+	if err := cutPoint("journal-written"); err != nil {
+		return err
+	}
+	f, err := os.OpenFile(path, os.O_WRONLY|os.O_TRUNC, st.Mode().Perm())
+	if err != nil {
+		_ = removeDurable(journal)
+		return err
+	}
+	if err := cutPoint("truncated"); err != nil {
+		f.Close()
+		return err
+	}
+	half := len(next) / 2
+	_, werr := f.Write(next[:half])
+	if werr == nil {
+		if err := cutPoint("half-written"); err != nil {
+			f.Close()
+			return err
+		}
+		_, werr = f.Write(next[half:])
+	}
+	if werr == nil {
+		werr = f.Sync()
+	}
+	if cerr := f.Close(); werr == nil {
+		werr = cerr
+	}
+	if werr == nil {
+		var back []byte
+		if back, werr = os.ReadFile(path); werr == nil && !bytes.Equal(back, next) {
+			werr = fmt.Errorf("the file did not read back as written")
+		}
+	}
+	if werr != nil {
+		// The write failed with the process alive: put the old bytes back now. If that fails too, the journal stays
+		// and the next start repairs the file.
+		if rerr := restoreInPlace(path, prev); rerr != nil {
+			return fmt.Errorf("config: writing %s failed (%v) and the old content could not be put back (%v) — the next start repairs it from the journal", path, werr, rerr)
+		}
+		_ = removeDurable(journal)
+		return fmt.Errorf("config: writing %s failed, the old content was put back: %w", path, werr)
+	}
+	if err := cutPoint("synced"); err != nil {
+		return err
+	}
+	return removeDurable(journal)
+}
+
+// RecoverCutWrite repairs a config whose last write was cut. It returns what it did, for the start's log line:
+// "" (no journal), "journal removed — the config is whole", or "the old config was put back".
+func RecoverCutWrite(path string) (string, error) {
+	journal := filepath.Join(CutWriteDir, cutWriteJournal)
+	prev, err := os.ReadFile(journal)
+	if os.IsNotExist(err) {
+		return "", nil
+	}
+	if err != nil {
+		return "", fmt.Errorf("config: reading the write journal: %w", err)
+	}
+	cur, cerr := os.ReadFile(path)
+	if cerr == nil && isJSONObject(cur) {
+		return "journal removed — the config is whole", removeDurable(journal)
+	}
+	if !isJSONObject(prev) {
+		// A journal that is itself broken cannot have been followed by a truncate (it is made durable first), so
+		// this is not a cut write. Leave both for a person.
+		return "", fmt.Errorf("config: %s does not parse and the write journal does not parse either", path)
+	}
+	if err := restoreInPlace(path, prev); err != nil {
+		return "", fmt.Errorf("config: putting the old content back into %s: %w", path, err)
+	}
+	return "the old config was put back", removeDurable(journal)
+}
+
+func isJSONObject(b []byte) bool {
+	var doc map[string]json.RawMessage
+	return json.Unmarshal(b, &doc) == nil && doc != nil
+}
+
+// restoreInPlace writes content into the EXISTING file (no create: the agent cannot make a file in the config's
+// folder), so the file keeps its owner and mode.
+func restoreInPlace(path string, content []byte) error {
+	f, err := os.OpenFile(path, os.O_WRONLY|os.O_TRUNC, 0)
+	if err != nil {
+		return err
+	}
+	if _, err := f.Write(content); err != nil {
+		f.Close()
+		return err
+	}
+	if err := f.Sync(); err != nil {
+		f.Close()
+		return err
+	}
+	return f.Close()
+}
+
+// writeDurable makes path hold content, whole, across a power cut: a new file, synced, renamed, the folder synced.
+func writeDurable(path string, content []byte) error {
+	tmp := path + ".tmp"
+	f, err := os.OpenFile(tmp, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o600)
+	if err != nil {
+		return err
+	}
+	if _, err := f.Write(content); err != nil {
+		f.Close()
+		return err
+	}
+	if err := f.Sync(); err != nil {
+		f.Close()
+		return err
+	}
+	if err := f.Close(); err != nil {
+		return err
+	}
+	if err := os.Rename(tmp, path); err != nil {
+		return err
+	}
+	return syncDir(filepath.Dir(path))
+}
+
+func removeDurable(path string) error {
+	if err := os.Remove(path); err != nil && !os.IsNotExist(err) {
+		return err
+	}
+	return syncDir(filepath.Dir(path))
+}
+
+func syncDir(dir string) error {
+	d, err := os.Open(dir)
+	if err != nil {
+		return err
+	}
+	defer d.Close()
+	return d.Sync()
+}

internal/config/cutwrite_test.go

@@ -0,0 +1,140 @@
+package config
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+const (
+	oldCfg = `{"proxmox":{"tls":{"fingerprint":"AA:BB"}},"keep":"me"}`
+	newCfg = `{"proxmox":{"tls":{"ca_file":"/etc/felhom-agent/pve-root-ca.pem","fingerprint":""}},"keep":"me"}`
+)
+
+// cutBox is a config file and a journal folder in a temp dir, with the write cut at one named step.
+func cutBox(t *testing.T) (cfg string) {
+	t.Helper()
+	dir := t.TempDir()
+	cfg = filepath.Join(dir, "etc", "agent.json")
+	if err := os.MkdirAll(filepath.Dir(cfg), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(cfg, []byte(oldCfg), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	prevDir, prevCut := CutWriteDir, cutPoint
+	CutWriteDir = filepath.Join(dir, "state")
+	t.Cleanup(func() { CutWriteDir, cutPoint = prevDir, prevCut })
+	return cfg
+}
+
+func cutAt(step string) {
+	cutPoint = func(s string) error {
+		if s == step {
+			return errors.New("power cut at " + s)
+		}
+		return nil
+	}
+}
+
+func read(t *testing.T, p string) string {
+	t.Helper()
+	b, err := os.ReadFile(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return string(b)
+}
+
+// The consequence that matters: after a power cut at ANY step of the write and the repair at the next start, the
+// config is the whole old file or the whole new file — never a broken one — and Load reads it.
+func TestRecoverCutWrite_AfterACutAtEveryStepTheConfigIsWholeAndLoads(t *testing.T) {
+	want := map[string]string{
+		"journal-written": oldCfg, // nothing was touched yet
+		"truncated":       oldCfg, // the file is EMPTY at the cut
+		"half-written":    oldCfg, // the file is HALF at the cut
+		"synced":          newCfg, // the new file is whole; only the journal is left
+	}
+	for step, wantAfter := range want {
+		t.Run(step, func(t *testing.T) {
+			cfg := cutBox(t)
+			cutAt(step)
+			if err := ReplaceInPlace(cfg, []byte(oldCfg), []byte(newCfg)); err == nil {
+				t.Fatal("the staged cut did not stop the write")
+			}
+			atCut := read(t, cfg)
+			if (step == "truncated" || step == "half-written") && isJSONObject([]byte(atCut)) {
+				t.Fatalf("the staging is wrong: at %s the file must be broken, it is %q", step, atCut)
+			}
+			cutPoint = func(string) error { return nil } // the next start: no cut
+			note, err := RecoverCutWrite(cfg)
+			if err != nil {
+				t.Fatalf("repair: %v", err)
+			}
+			if note == "" {
+				t.Fatal("the repair saw no journal — a cut write would stay broken")
+			}
+			if got := read(t, cfg); got != wantAfter {
+				t.Fatalf("after a cut at %s and the repair the config is %q, want %q", step, got, wantAfter)
+			}
+			if _, err := Load(cfg); err != nil {
+				t.Fatalf("the agent does not start with the repaired config: %v", err)
+			}
+			if _, err := os.Stat(filepath.Join(CutWriteDir, cutWriteJournal)); !os.IsNotExist(err) {
+				t.Fatalf("the journal stayed after the repair (%v) — an old journal would undo a later change", err)
+			}
+		})
+	}
+}
+
+// A write that is not cut leaves the new file and NO journal: a journal that stayed would put an old config back
+// at some later start.
+func TestReplaceInPlace_AWholeWriteLeavesTheNewFileAndNoJournal(t *testing.T) {
+	cfg := cutBox(t)
+	if err := ReplaceInPlace(cfg, []byte(oldCfg), []byte(newCfg)); err != nil {
+		t.Fatal(err)
+	}
+	if got := read(t, cfg); got != newCfg {
+		t.Fatalf("config = %q", got)
+	}
+	if _, err := os.Stat(filepath.Join(CutWriteDir, cutWriteJournal)); !os.IsNotExist(err) {
+		t.Fatalf("the journal stayed after a whole write: %v", err)
+	}
+	if note, err := RecoverCutWrite(cfg); err != nil || note != "" {
+		t.Fatalf("a start after a whole write must do nothing: note=%q err=%v", note, err)
+	}
+	if got := read(t, cfg); got != newCfg {
+		t.Fatalf("the start changed a whole config to %q", got)
+	}
+}
+
+// No journal folder, no write: the old file stays as it was.
+func TestReplaceInPlace_WithoutAJournalNothingIsWritten(t *testing.T) {
+	cfg := cutBox(t)
+	// the journal cannot be made: a folder sits where its new file must be written
+	if err := os.MkdirAll(filepath.Join(CutWriteDir, cutWriteJournal+".tmp"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	if err := ReplaceInPlace(cfg, []byte(oldCfg), []byte(newCfg)); err == nil {
+		t.Fatal("the write went on with no journal")
+	}
+	if got := read(t, cfg); got != oldCfg {
+		t.Fatalf("the config changed with no journal: %q", got)
+	}
+}
+
+// A config that is broken with NO journal is not a cut write: the repair does not touch it, and says nothing.
+func TestRecoverCutWrite_ABrokenConfigWithNoJournalIsLeftAlone(t *testing.T) {
+	cfg := cutBox(t)
+	if err := os.WriteFile(cfg, []byte(`{"half":`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	note, err := RecoverCutWrite(cfg)
+	if err != nil || note != "" {
+		t.Fatalf("note=%q err=%v", note, err)
+	}
+	if got := read(t, cfg); got != `{"half":` {
+		t.Fatalf("the repair wrote a config it had no journal for: %q", got)
+	}
+}

internal/config/filelock.go

@@ -0,0 +1,12 @@
+package config
+
+import "sync"
+
+// FileMu is held by EVERY read-modify-write of agent.json inside this process: the trust move (pvetrust), the PBS
+// seed (pbsdr) and the backup-target move (localapi). Each of them reads the whole file, changes one section and
+// writes the whole file back in place — without one lock two of them could interleave, and the second writer would
+// put back the document it read BEFORE the first one's change (a silently reverted `escrow.pbs_storage_id`,
+// `backup.local_backup_target` or trust section). Found by review on 2026-11-14, the night the third writer was
+// added. Hold it for the whole read … write, never across a call to Proxmox or the root wrapper.
+// Pinned by TestAgentJSONWritersTakeTheSharedLock (cmd/felhom-agent).
+var FileMu sync.Mutex

internal/hub/client.go

@@ -10,6 +10,7 @@ import (
 	"io"
 	"log/slog"
 	"net/http"
+	"net/url"
 	"os"
 	"strings"
 	"time"
@@ -86,6 +87,7 @@ func (e *HTTPError) Error() string {
 // IS the heartbeat (locked decision 1). Errors are typed (transport vs HTTP) and
 // never include the bearer token.
 func (c *Client) Report(ctx context.Context, r *HostReport) (*ControlEnvelope, error) {
+	r.SentAt = time.Now().UTC().Format(time.RFC3339Nano) // the box clock at SENDING (R-1384); see HostReport.SentAt
 	body, err := json.Marshal(r)
 	if err != nil {
 		return nil, fmt.Errorf("hub: marshaling report: %w", err)
@@ -285,10 +287,37 @@ func (c *Client) Jobs(ctx context.Context) ([]JobWire, error) {
 // /hosts/{id}/jobs/{job_id}, self-scoped. Called after a job is executed OR permanently rejected
 // (the nonce is already durably burned on a passing verify, so re-processing is replay-safe).
 func (c *Client) CompleteJob(ctx context.Context, jobID string) error {
+	return c.CompleteJobWithOutcome(ctx, jobID, JobOutcome{})
+}
+
+// JobOutcome says HOW a signed job ended when the agent clears it (R-1383, D78): Outcome is done | rejected | failed
+// | malformed; Reason is one code of signedjobs.ReasonCodes (empty for done); Op is the job's op class. It rides the
+// DELETE's query, so a hub that does not know it ignores it. Until this, „ran” and „refused” looked the same to the
+// hub, and a refused agent update left no trace outside the box's journal. No secret: no signature, no nonce, no
+// parameter of the job is in it.
+type JobOutcome struct {
+	Outcome string
+	Reason  string
+	Op      string
+}
+
+// CompleteJobWithOutcome clears the job and tells the hub how it ended. The zero JobOutcome sends no query.
+func (c *Client) CompleteJobWithOutcome(ctx context.Context, jobID string, o JobOutcome) error {
 	if c.hostID == "" || jobID == "" {
 		return fmt.Errorf("hub: CompleteJob requires host_id + job_id")
 	}
-	req, err := http.NewRequestWithContext(ctx, http.MethodDelete, c.baseURL+"/api/v1/hosts/"+c.hostID+"/jobs/"+jobID, nil)
+	u := c.baseURL + "/api/v1/hosts/" + c.hostID + "/jobs/" + jobID
+	if o.Outcome != "" {
+		q := url.Values{"outcome": {o.Outcome}}
+		if o.Reason != "" {
+			q.Set("reason", o.Reason)
+		}
+		if o.Op != "" {
+			q.Set("op", o.Op)
+		}
+		u += "?" + q.Encode()
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodDelete, u, nil)
 	if err != nil {
 		return err
 	}

internal/hub/client_sentat_test.go

@@ -0,0 +1,76 @@
+package hub
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"net/http"
+	"testing"
+	"time"
+)
+
+// R-1384 (D79): the hub tells a wrong box clock by the time stamped at the moment of SENDING. A report whose collect
+// started long ago (an old reported_at) must still leave with a fresh sent_at, and a second send of the SAME report
+// struct must carry a new one — or a report that waited on the box reads as a clock that is behind.
+// COMPANION RED-PROOFS (observed, felhom.eu audits/night-build-2026-11-14/B/redproof-agent-sent-at.txt).
+func TestReport_StampsSentAtAtSending(t *testing.T) {
+	var bodies []map[string]any
+	c := testClient(func(r *http.Request) (*http.Response, error) {
+		raw, _ := io.ReadAll(r.Body)
+		var m map[string]any
+		if err := json.Unmarshal(raw, &m); err != nil {
+			t.Fatalf("body: %v", err)
+		}
+		bodies = append(bodies, m)
+		return httpResp(200, `{"status":"ok"}`), nil
+	})
+	old := time.Now().Add(-40 * time.Minute).UTC()
+	rep := &HostReport{HostID: "h", ReportedAt: old.Format(time.RFC3339), SentAt: old.Format(time.RFC3339Nano)}
+	for i := 0; i < 2; i++ {
+		if _, err := c.Report(context.Background(), rep); err != nil {
+			t.Fatalf("Report: %v", err)
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	var sent []time.Time
+	for i, m := range bodies {
+		s, _ := m["sent_at"].(string)
+		at, err := time.Parse(time.RFC3339Nano, s)
+		if err != nil {
+			t.Fatalf("send %d: sent_at = %q, want an RFC3339 time", i, s)
+		}
+		if d := time.Since(at); d < 0 || d > 5*time.Second {
+			t.Errorf("send %d: sent_at is %s old, want the moment of sending (reported_at was 40 minutes old)", i, d)
+		}
+		if m["reported_at"] != old.Format(time.RFC3339) {
+			t.Errorf("send %d: reported_at changed to %v — the collect time is not the send time's to touch", i, m["reported_at"])
+		}
+		sent = append(sent, at)
+	}
+	if len(sent) == 2 && !sent[1].After(sent[0]) {
+		t.Errorf("the second send carries %s, not after the first %s — the stamp is not taken per send", sent[1], sent[0])
+	}
+}
+
+// R-1383 (D78): the clearing call says how the job ended, in the DELETE's query; a plain CompleteJob sends none (an
+// older hub, and the „nothing to say” case, see the same request as before).
+func TestCompleteJob_OutcomeRidesTheQuery(t *testing.T) {
+	var urls []string
+	c := testClient(func(r *http.Request) (*http.Response, error) {
+		if r.Method != http.MethodDelete {
+			t.Errorf("method = %s", r.Method)
+		}
+		urls = append(urls, r.URL.Path+"?"+r.URL.RawQuery)
+		return httpResp(200, `{"status":"ok"}`), nil
+	})
+	if err := c.CompleteJob(context.Background(), "j1"); err != nil {
+		t.Fatal(err)
+	}
+	if err := c.CompleteJobWithOutcome(context.Background(), "j2", JobOutcome{Outcome: "rejected", Reason: "expired", Op: "agent_update"}); err != nil {
+		t.Fatal(err)
+	}
+	want := []string{"/api/v1/hosts/demo-host-01/jobs/j1?", "/api/v1/hosts/demo-host-01/jobs/j2?op=agent_update&outcome=rejected&reason=expired"}
+	if len(urls) != 2 || urls[0] != want[0] || urls[1] != want[1] {
+		t.Fatalf("requests = %q, want %q", urls, want)
+	}
+}

internal/hub/collect.go

@@ -124,10 +124,14 @@ type Collector struct {
 	temp             TempReader                                    // slice 9: host CPU/chassis temp (nil-safe → nil temp)
 	capProbe         func(ctx context.Context) []capability.Status // v0.44.0: privileged-capability self-check (nil → empty)
 	leafFP           string                                        // v0.48.0: served local-API leaf fp (static per process; "" when local API disabled)
+	pveTrust         func() any                                    // R-1380: the proxmox_trust stanza (nil = omitted)
+	guestPowerGaveUp func() any                                    // R-1383: the guest_power_gave_up stanza (nil = omitted)
+	profileRefresh   func() any                                    // R-1383: the profile_refresh stanza (nil = omitted)
 	addrEnum         AddressEnumerator                             // v0.119.0: host interface enumeration; nil => the REAL one (see collectAddresses)
 	wg               WireguardReporter                             // S3: offsite-tunnel status (nil → stanza omitted)
 	pbsdr            PBSDRReporter                                 // slice 2: PBS DR tier bridge state (nil → stanza omitted)
 	ctrlSup          ControllerSupervisorReporter                  // R-523: in-guest controller supervisor (nil → stanza omitted)
+	opResults        func() []OperatorActionResult                 // 2026-10-29: the operator actions' answers (nil → none)
 	guestNet         GuestNetReporter                              // R-54: per-guest network watchdog (nil → stanza omitted)
 	diskTrim         GuestDiskTrimReporter                         // R-444: weekly guest disk trim (nil → stanza omitted)
 	lineTest         LineTestReporter                              // R-1247: weekly upload line test (nil → stanza omitted)
@@ -208,6 +212,24 @@ func (c *Collector) SetCapabilityProber(probe func(ctx context.Context) []capabi
 	return c
 }
 
+// SetProxmoxTrustSource wires the `proxmox_trust` stanza (R-1380). Optional; nil = the stanza is omitted.
+func (c *Collector) SetProxmoxTrustSource(f func() any) *Collector {
+	c.pveTrust = f
+	return c
+}
+
+// SetProfileRefreshSource wires the `profile_refresh` stanza (R-1383). Optional; nil, or a nil result, omits it.
+func (c *Collector) SetProfileRefreshSource(f func() any) *Collector {
+	c.profileRefresh = f
+	return c
+}
+
+// SetGuestPowerGaveUpSource wires the `guest_power_gave_up` stanza (R-1383). Optional; nil, or a nil result, omits it.
+func (c *Collector) SetGuestPowerGaveUpSource(f func() any) *Collector {
+	c.guestPowerGaveUp = f
+	return c
+}
+
 // SetLeafFingerprint records the served local-API leaf fp (v0.48.0) to ride every host report (the hub
 // watches it for a re-key). Static per process — set once at startup. "" when the local API is
 // disabled. Returns the collector for chaining.
@@ -235,6 +257,12 @@ type ControllerSupervisorReporter interface {
 	ControllerSupervisorStatus(ctx context.Context) *ControllerSupervisorStatus
 }
 
+// SetOperatorActionResults wires the source of the operator actions' answers (2026-10-29; nil → none sent).
+func (c *Collector) SetOperatorActionResults(f func() []OperatorActionResult) *Collector {
+	c.opResults = f
+	return c
+}
+
 // SetControllerSupervisorReporter wires the R-523 controller supervisor as a report source (nil-safe).
 func (c *Collector) SetControllerSupervisorReporter(r ControllerSupervisorReporter) *Collector {
 	c.ctrlSup = r
@@ -384,6 +412,27 @@ func (c *Collector) SetOOBReporter(o OOBReporter) *Collector {
 // error (no useful report — the cycle skips the POST); a failed per-guest
 // GuestConfig degrades that guest to status="unknown" without spec but still sends;
 // a cloudflared probe failure yields status="unknown" and is never fatal.
+func (c *Collector) profileRefreshStanza() any {
+	if c.profileRefresh == nil {
+		return nil
+	}
+	return c.profileRefresh()
+}
+
+func (c *Collector) guestPowerGaveUpStanza() any {
+	if c.guestPowerGaveUp == nil {
+		return nil
+	}
+	return c.guestPowerGaveUp()
+}
+
+func (c *Collector) proxmoxTrust() any {
+	if c.pveTrust == nil {
+		return nil
+	}
+	return c.pveTrust()
+}
+
 func (c *Collector) Collect(ctx context.Context) (*HostReport, error) {
 	ns, err := c.px.NodeStatus(ctx)
 	if err != nil {
@@ -409,12 +458,15 @@ func (c *Collector) Collect(ctx context.Context) (*HostReport, error) {
 		RestoreTests:   c.collectRestoreTests(ctx),
 		PBSSnapshots:   c.collectPBSSnapshots(ctx),
 
-		AuditTail:       []AuditEntry{},
-		Cloudflared:     c.cloudflared(ctx),
-		Capabilities:    c.capabilities(ctx),
-		LeafFingerprint: c.leafFP,
-		Addresses:       c.collectAddresses(),
-		System:          c.systemInfo(ctx, ns),
+		AuditTail:        []AuditEntry{},
+		Cloudflared:      c.cloudflared(ctx),
+		Capabilities:     c.capabilities(ctx),
+		LeafFingerprint:  c.leafFP,
+		ProxmoxTrust:     c.proxmoxTrust(),
+		GuestPowerGaveUp: c.guestPowerGaveUpStanza(),
+		ProfileRefresh:   c.profileRefreshStanza(),
+		Addresses:        c.collectAddresses(),
+		System:           c.systemInfo(ctx, ns),
 	}
 	// DR recipe host-half — derived from the just-collected guest/storage/PBS facts (no new reads).
 	// Secret-free by construction (identifiers/intents/sizes/coordinates only).
@@ -432,6 +484,9 @@ func (c *Collector) Collect(ctx context.Context) (*HostReport, error) {
 	if c.ctrlSup != nil {
 		report.ControllerSupervisor = c.ctrlSup.ControllerSupervisorStatus(ctx)
 	}
+	if c.opResults != nil {
+		report.OperatorActionResults = c.opResults()
+	}
 	// R-54: guest-network watchdog state (nil reporter = feature not wired → stanza omitted).
 	if c.guestNet != nil {
 		report.GuestNet = c.guestNet.GuestNetStatus(ctx)

internal/hub/loop.go

@@ -70,8 +70,14 @@ type Loop struct {
 	// reported (R-998) is called after every report the hub ACCEPTED, with that report. The power-off
 	// path waits on it so the hub has the deliberate-off record before the box goes down.
 	reported func(r *HostReport)
+	// collectErr (R-1380) sees every failed collect.
+	collectErr func(ctx context.Context, err error)
 }
 
+// SetCollectErrorHook wires an observer of a failed collect (R-1380: a call to Proxmox that fails on TRUST starts the
+// move to the CA copy, or a reload of a refreshed copy). Optional; it must not block for long.
+func (l *Loop) SetCollectErrorHook(f func(ctx context.Context, err error)) { l.collectErr = f }
+
 // SetReportedHook wires the R-998 after-report observer. Optional.
 func (l *Loop) SetReportedHook(f func(r *HostReport)) { l.reported = f }
 
@@ -144,6 +150,9 @@ func (l *Loop) cycle(ctx context.Context, current time.Duration) time.Duration {
 	report, err := l.collector.Collect(ctx)
 	if err != nil {
 		l.logger.Warn("hub: collect failed; skipping this cycle's report", "err", err)
+		if l.collectErr != nil {
+			l.collectErr(ctx, err)
+		}
 		return current
 	}
 	// Fulfill a pending log-pull: attach the ring tail to THIS report and clear the

internal/hub/report.go

@@ -15,8 +15,13 @@ import (
 // server-side received_at is the hub's dead-man's-switch liveness signal. There is
 // no separate heartbeat endpoint.
 type HostReport struct {
-	HostID       string `json:"host_id"`     // echoes config.Hub.HostID
-	ReportedAt   string `json:"reported_at"` // RFC3339, agent clock
+	HostID     string `json:"host_id"`     // echoes config.Hub.HostID
+	ReportedAt string `json:"reported_at"` // RFC3339, agent clock
+	// SentAt (R-1384, D79) is this box's clock at the moment the report is SENT — stamped by Client.Report, never by
+	// the collector. The hub compares it with its own clock to tell a wrong box clock. ReportedAt cannot serve: it is
+	// stamped when the collect STARTS, and a collect that waits on Proxmox makes it old on a box whose clock is right.
+	// Pinned by TestReport_StampsSentAtAtSending.
+	SentAt       string `json:"sent_at,omitempty"`
 	AgentVersion string `json:"agent_version"`
 	// AgentSHA256 is the sha256 of the binary this process is RUNNING (read through /proc/self/exe,
 	// once per process), R-349. The version string cannot tell a hand-built proof binary from the
@@ -53,6 +58,11 @@ type HostReport struct {
 	// alert. Not a secret (the fp is public; the token is never reported).
 	LeafFingerprint string `json:"leaf_fingerprint"`
 
+	// ProxmoxTrust (R-1380, felhom.eu `03` §12a): how the agent trusts Proxmox's certificate (`mode`: ca | fingerprint |
+	// …), and the CAUSE when something about it is wrong (`problem`, `detail`): the proof with the CA copy failed, the
+	// copy ends soon, the root step gave no copy. Omitted when not wired. No secret: a mode, a cause, a date.
+	ProxmoxTrust any `json:"proxmox_trust,omitempty"`
+
 	// Addresses are the host's routable addresses, one entry per (interface, address) — the LAN
 	// bridge, the WG tunnel, a tailnet. Added v0.119.0 because the hub could not show a managed
 	// box's IP anywhere: nothing in this report carried one. Non-nil so it marshals as [];
@@ -172,6 +182,32 @@ type HostReport struct {
 	// byte-stable. The hub parser is pinned by hub/internal/monitor/controller_supervisor_test.go
 	// against the JSON TestControllerSupervisorStanza_WireShape pins here.
 	ControllerSupervisor *ControllerSupervisorStatus `json:"controller_supervisor,omitempty"`
+	// ProfileRefresh (R-1383, release 3): the box's last security-profile refresh and how the agent's last ask ended
+	// (osupdate.ProfileRefreshStatus, field-exact; hub sysfacts.ProfileRefresh). `any` + omitempty: absent when there
+	// is nothing to say or the source is not wired.
+	ProfileRefresh any `json:"profile_refresh,omitempty"`
+	// GuestPowerGaveUp (R-1383, D78): the guests the guest-power watchdog stopped trying to start (localapi
+	// GuestPowerGiveUp, field-exact: vmid, since, attempts). The hub raises one operator alarm per guest and clears it
+	// when the guest runs. `any` + omitempty: absent when there is none or the source is not wired.
+	GuestPowerGaveUp any `json:"guest_power_gave_up,omitempty"`
+	// OperatorActionResults (2026-10-29, the operator brief's Part 1): the answers to the operator actions the hub's
+	// replies listed — the controller's shape (`04` §6.1). Re-sent until the reply stops listing the id. omitempty: absent
+	// when there is nothing to answer, so the golden stays byte-stable.
+	OperatorActionResults []OperatorActionResult `json:"operator_action_results,omitempty"`
+}
+
+// OperatorAction is ONE entry of the reply's operator_actions (hub store.OperatorActionDirective, field-exact).
+type OperatorAction struct {
+	ID     int64  `json:"id"`
+	Action string `json:"action"`
+	Arg    string `json:"arg,omitempty"`
+}
+
+// OperatorActionResult is one answer (hub api.hostOperatorActionResult, field-exact): outcome done | refused | failed.
+type OperatorActionResult struct {
+	ID      int64  `json:"id"`
+	Outcome string `json:"outcome"`
+	Message string `json:"message"`
 }
 
 // ControllerSupervisorStatus is the R-523 stanza. Carries no secret.
@@ -195,6 +231,17 @@ type ControllerSupervisorGuest struct {
 	Restarts24h        int    `json:"restarts_24h"`
 	SlowCrashloop      bool   `json:"slow_crashloop"`
 	SlowCrashloopSince string `json:"slow_crashloop_since,omitempty"` // RFC3339
+	// R-1344 (2026-10-29) — the controller RUNS but Docker's health check has read „unhealthy" on several consecutive
+	// sweeps. ALARM ONLY: the agent does not restart it for this (the measurement must first show a busy night never
+	// looks like it). UnresponsiveSince is the episode's start; the hub keys on it MOVING (controller_unresponsive).
+	Unresponsive      bool   `json:"unresponsive"`
+	UnresponsiveSince string `json:"unresponsive_since,omitempty"` // RFC3339
+	// R-1344 restart (decided 2026-10-30 from the night's measurement): UnresponsiveRestartAfterS is how long after the
+	// alarm the agent restarts a controller that still does not answer (0 / absent = an older agent: alarm only — the
+	// hub's message says which). UnresponsiveRestartAt is the last such restart; the hub keys on it MOVING
+	// (controller_unresponsive_restart) and it equals LastRestartAt for that act.
+	UnresponsiveRestartAfterS int    `json:"unresponsive_restart_after_s,omitempty"`
+	UnresponsiveRestartAt     string `json:"unresponsive_restart_at,omitempty"` // RFC3339
 }
 
 // PBSDRStatus is the per-heartbeat PBS-DR-tier bridge state (slice 2). States:
@@ -631,6 +678,9 @@ type ControlEnvelope struct {
 	// NEXT heartbeat carries it in log_tail (the report-channel log_tail_requests mirror).
 	// Absent/false on an old hub → nothing happens.
 	LogTailRequested bool `json:"log_tail_requested"`
+	// OperatorActions (2026-10-29): the operator's pending actions for THIS host's agent — a CLOSED list on both sides
+	// (`04` §6.1). Absent on an older hub.
+	OperatorActions []OperatorAction `json:"operator_actions,omitempty"`
 }
 
 // LogTail is the heartbeat's on-demand agent log tail (see HostReport.LogTail).

internal/localapi/backup_target.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
 	"net/http"
 	"os"
 	"path/filepath"
@@ -145,9 +146,11 @@ func (s *Server) validateBackupTargetMount(ctx context.Context, where string) er
 // begin with. A typed round-trip would silently drop any key this build does not know about.
 //
 // Written IN PLACE (O_TRUNC), not tmp+rename: /etc/felhom-agent is root-owned while agent.json is
-// agent-owned 0600, so the non-root agent cannot rename into that directory. A recovery copy is
-// parked first, so a torn write is recoverable.
+// agent-owned 0600, so the non-root agent cannot rename into that directory. config.ReplaceInPlace
+// journals the old bytes first; the next start repairs a cut write (TestRecoverCutWrite_*).
 func (s *Server) setConfiguredBackupTarget(id string) error {
+	config.FileMu.Lock() // one writer of agent.json at a time in this process (config.FileMu)
+	defer config.FileMu.Unlock()
 	path := s.configPath
 	if path == "" {
 		return fmt.Errorf("no config path known to this agent (env-only config)")
@@ -179,24 +182,14 @@ func (s *Server) setConfiguredBackupTarget(id string) error {
 	if err != nil {
 		return err
 	}
-	st, err := os.Stat(path)
-	if err != nil {
-		return err
-	}
 	if s.stateDir != "" {
 		if err := os.MkdirAll(s.stateDir, 0o700); err == nil {
 			_ = os.WriteFile(filepath.Join(s.stateDir, "agent.json.pre-backup-target"), raw, 0o600)
 		}
 	}
-	f, err := os.OpenFile(path, os.O_WRONLY|os.O_TRUNC, st.Mode().Perm())
-	if err != nil {
-		return err
-	}
-	if _, err := f.Write(out); err != nil {
-		f.Close()
-		return err
-	}
-	return f.Close()
+	// config.ReplaceInPlace: the old bytes go into a journal first, so a power cut in the middle of the write is
+	// repaired at the next start (R-1380).
+	return config.ReplaceInPlace(path, raw, out)
 }
 
 // wrapperReason surfaces the wrapper's own REFUSED line when it produced one — it explains WHY in

internal/localapi/controllersupervisor.go

@@ -3,6 +3,7 @@ package localapi
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"os"
 	"path/filepath"
 	"sort"
@@ -83,6 +84,20 @@ const (
 	// controllerSlowCounterFile holds the 24-hour restart times and the last raise, per guest, beside
 	// the parked marker.
 	controllerSlowCounterFile = "controller-restarts-24h.json"
+
+	// R-1344 (operator brief 2026-10-29, Part 2) — a controller that RUNS but does not answer. Docker's own HEALTHCHECK
+	// (`curl -f localhost:8080/api/health`, every 30 s, timeout 5 s) reads „unhealthy" only after 3 failed probes (~90 s);
+	// controllerUnresponsiveConfirm more sweeps (3 minutes) make it an episode: logged, carried by the stanza, mailed by
+	// the hub (controller_unresponsive). Not counted during a swap, a parked guest, a locked guest or a vzdump.
+	controllerUnresponsiveConfirm = 6
+	// controllerUnresponsiveRestart (decided 2026-10-30 from the measurement): at this many consecutive unhealthy sweeps
+	// (5 minutes after Docker's „unhealthy", ~6.5 minutes after the first failed probe; 2 minutes after the alarm) the
+	// agent restarts the controller — the not-running restart's guards and brake, plus the controller's busy file. The
+	// numbers: Tester 1 and demo-hp, 2026-10-29 11:12 → 2026-10-30 07:00 UTC, 4 733 probes through the night chain (a
+	// PBS and a local vzdump, the DB dumps, tier-2, the off-site push, a restore test, two controller updates): no failed
+	// probe, slowest 303 ms against the 5 s timeout. A busy night never came near even ONE failed probe; ten sweeps is
+	// the margin. Pinned by TestSupervisor_UnresponsiveIsAlarmedThenRestarted.
+	controllerUnresponsiveRestart = 10
 )
 
 // controllerSupState is one guest's supervisor record. In-memory on purpose (the guest-power
@@ -103,6 +118,18 @@ type controllerSupState struct {
 	// on the other hand, would hide exactly the box it exists for (one whose agent restarts too).
 	restarts24h        []time.Time
 	slowCrashloopSince time.Time // the last raise; kept after it ages out, the hub keys on it MOVING
+
+	// opBrakeUntil (2026-10-29): an operator restart refused by the brake refuses every press until then (30 min).
+	opBrakeUntil time.Time
+
+	// R-1344: consecutive running-but-unhealthy sweeps, and the current episode (since kept after it ends).
+	unhealthySeen     int
+	unresponsive      bool
+	unresponsiveSince time.Time
+	// unresponsiveRestartAt: the last restart for not answering (kept; the hub keys on it MOVING). unrespRefusal: the
+	// last reason the restart was refused, so a refusal is logged once per change, not every sweep.
+	unresponsiveRestartAt time.Time
+	unrespRefusal         string
 }
 
 type controllerSupervisor struct {
@@ -238,6 +265,10 @@ func (s *Server) ControllerSupervisorTick(ctx context.Context) {
 		if !s.provisionedGuest(g.VMID) {
 			continue
 		}
+		if s.rebooting(g.VMID) { // D65: the operator's reboot owns the guest until it is up again
+			s.resetNotRunning(g.VMID)
+			continue
+		}
 		evaluated++
 		if !s.superviseOneController(ctx, g.VMID, g.Status) {
 			down++
@@ -293,6 +324,7 @@ func (s *Server) superviseOneController(ctx context.Context, vmid int, guestStat
 	if running {
 		st.notRunningSeen = 0
 		s.ctrlSup.mu.Unlock()
+		s.observeControllerHealth(ctx, vmid, parked)
 		return true
 	}
 	st.notRunningSeen++
@@ -333,9 +365,27 @@ func (s *Server) superviseOneController(ctx context.Context, vmid int, guestStat
 		return false
 	}
 
-	// Backoff.
+	if !s.supervisorBrakeAllows(vmid, now) {
+		return false
+	}
+
+	reason := "controller container " + status + " on " + strconv.Itoa(controllerSupervisorConfirm) + " consecutive sweeps"
+	s.logger.Warn("controller-supervisor: controller is NOT running — restarting the bootstrap unit",
+		"vmid", vmid, "status", status, "unit", bootstrapUnit)
+	if _, err := s.guestExec.GuestExec(ctx, vmid, "systemctl", "restart", bootstrapUnit); err != nil {
+		s.logger.Error("controller-supervisor: bootstrap restart failed", "vmid", vmid, "err", err)
+		reason += "; restart FAILED: " + err.Error()
+	}
+	s.recordSupervisorRestart(vmid, now, reason, nil)
+	return false
+}
+
+// supervisorBrakeAllows is the brake every supervisor restart passes (the not-running one and R-1344's): a crash-loop
+// pause in force refuses; controllerCrashloopMax restarts inside controllerCrashloopWindow start one (logged, carried by
+// the stanza, mailed by the hub as controller_crashloop) and refuse.
+func (s *Server) supervisorBrakeAllows(vmid int, now time.Time) bool {
 	s.ctrlSup.mu.Lock()
-	st = s.supState(vmid)
+	st := s.supState(vmid)
 	if !st.crashloopSince.IsZero() {
 		if now.Sub(st.crashloopSince) < controllerCrashloopPause {
 			s.ctrlSup.mu.Unlock()
@@ -358,22 +408,23 @@ func (s *Server) superviseOneController(ctx context.Context, vmid int, guestStat
 		return false
 	}
 	s.ctrlSup.mu.Unlock()
+	return true
+}
 
-	reason := "controller container " + status + " on " + strconv.Itoa(controllerSupervisorConfirm) + " consecutive sweeps"
-	s.logger.Warn("controller-supervisor: controller is NOT running — restarting the bootstrap unit",
-		"vmid", vmid, "status", status, "unit", bootstrapUnit)
-	if _, err := s.guestExec.GuestExec(ctx, vmid, "systemctl", "restart", bootstrapUnit); err != nil {
-		s.logger.Error("controller-supervisor: bootstrap restart failed", "vmid", vmid, "err", err)
-		reason += "; restart FAILED: " + err.Error()
-	}
+// recordSupervisorRestart puts a restart the supervisor performed on the record: the brake's window, the totals, the
+// reason, and R-539's slow counter (persisted; raised at most once per 24 hours — the hub mails on the raise MOVING).
+// mark, when set, runs under the same lock (R-1344 stamps its own record there).
+func (s *Server) recordSupervisorRestart(vmid int, now time.Time, reason string, mark func(st *controllerSupState)) {
 	s.ctrlSup.mu.Lock()
-	st = s.supState(vmid)
+	st := s.supState(vmid)
 	st.restarts = append(st.restarts, now)
 	st.restartsTotal++
 	st.lastRestartAt = now
 	st.lastReason = reason
 	st.notRunningSeen = 0
-	// R-539: the slow counter. Raise at most once per 24 hours — the hub mails on the raise MOVING.
+	if mark != nil {
+		mark(st)
+	}
 	st.restarts24h = append(pruneBefore(st.restarts24h, now.Add(-controllerSlowCrashloopWindow)), now)
 	n24 := len(st.restarts24h)
 	raised := false
@@ -389,7 +440,101 @@ func (s *Server) superviseOneController(ctx context.Context, vmid int, guestStat
 		s.logger.Warn("controller-supervisor: SLOW CRASH-LOOP — the controller keeps dying; still restarting it, raising controller_slow_crashloop",
 			"vmid", vmid, "restarts_24h", n24, "window", controllerSlowCrashloopWindow.String(), "threshold", controllerSlowCrashloopMax)
 	}
-	return false
+}
+
+// observeControllerHealth is R-1344's count for a RUNNING controller: Docker's health status, one read per sweep over
+// the swap's `docker inspect -f` grant. Anything but „unhealthy" — healthy, starting, no health check, a read that
+// fails — resets the count; so do a swap, a parked guest, a locked guest and a vzdump (each is someone else's moment).
+func (s *Server) observeControllerHealth(ctx context.Context, vmid int, parked bool) {
+	s.swapMu.Lock()
+	swapping := s.swapInFlight[vmid]
+	s.swapMu.Unlock()
+	unhealthy := false
+	if !parked && !swapping {
+		if out, err := s.guestExec.GuestExec(ctx, vmid, "docker", "inspect", "-f", opHealthTemplate, controllerContainer); err == nil {
+			f := strings.SplitN(strings.TrimSpace(out), "|", 2)
+			unhealthy = len(f) == 2 && f[0] == "true" && f[1] == "unhealthy"
+		}
+	}
+	if unhealthy {
+		if lock, _, err := s.staleLock.Lock(ctx, vmid); err != nil || lock != "" {
+			unhealthy = false
+		} else if busy, berr := s.staleLock.BackupRunning(ctx, vmid); berr != nil || busy {
+			unhealthy = false
+		}
+	}
+	now := s.clock()
+	s.ctrlSup.mu.Lock()
+	st := s.supState(vmid)
+	if !unhealthy {
+		was := st.unresponsive
+		st.unhealthySeen, st.unresponsive = 0, false
+		s.ctrlSup.mu.Unlock()
+		if was {
+			s.logger.Info("controller-supervisor: the controller answers its health check again", "vmid", vmid)
+		}
+		return
+	}
+	st.unhealthySeen++
+	seen := st.unhealthySeen
+	start := seen >= controllerUnresponsiveConfirm && !st.unresponsive
+	if start {
+		st.unresponsive, st.unresponsiveSince = true, now
+	}
+	since := st.unresponsiveSince
+	s.ctrlSup.mu.Unlock()
+	if start {
+		s.logger.Error("controller-supervisor: the controller RUNS but does not answer its health check — raising controller_unresponsive; restarting it if it stays so",
+			"vmid", vmid, "unhealthy_sweeps", seen, "interval", controllerSupervisorInterval.String(),
+			"restart_after_sweeps", controllerUnresponsiveRestart)
+	}
+	if seen >= controllerUnresponsiveRestart {
+		s.restartUnresponsive(ctx, vmid, seen, since, now)
+	}
+}
+
+// restartUnresponsive is R-1344's restart (decided 2026-10-30): the controller's busy file (fresh = a backup, a
+// restore, an update or an export is running; unreadable = cannot be ruled out) refuses, then the brake every
+// supervisor restart passes. The swap, the parked marker, the lock and the vzdump never reach here — they reset the
+// count. A refusal keeps the alarm and is logged once per reason; the next sweep asks again.
+func (s *Server) restartUnresponsive(ctx context.Context, vmid, seen int, since, now time.Time) {
+	refuse := func(why string) {
+		s.ctrlSup.mu.Lock()
+		st := s.supState(vmid)
+		changed := st.unrespRefusal != why
+		st.unrespRefusal = why
+		s.ctrlSup.mu.Unlock()
+		if changed {
+			s.logger.Warn("controller-supervisor: the controller does not answer, but the restart is refused — the alarm stays",
+				"vmid", vmid, "why", why, "unhealthy_sweeps", seen)
+		}
+	}
+	busy, what, err := s.controllerBusy(ctx, vmid)
+	if err != nil {
+		refuse("the controller's busy file cannot be read: " + err.Error())
+		return
+	}
+	if busy {
+		refuse("the controller is busy: " + what)
+		return
+	}
+	if !s.supervisorBrakeAllows(vmid, now) {
+		refuse("the restart brake")
+		return
+	}
+	reason := fmt.Sprintf("the controller ran but did not answer its health check for %d sweeps (since %s)",
+		seen, since.UTC().Format(time.RFC3339))
+	s.logger.Warn("controller-supervisor: the controller does not answer — restarting the bootstrap unit",
+		"vmid", vmid, "unhealthy_sweeps", seen, "since", since.UTC().Format(time.RFC3339), "unit", bootstrapUnit)
+	if _, err := s.guestExec.GuestExec(ctx, vmid, "systemctl", "restart", bootstrapUnit); err != nil {
+		s.logger.Error("controller-supervisor: bootstrap restart failed", "vmid", vmid, "err", err)
+		reason += "; restart FAILED: " + err.Error()
+	}
+	s.recordSupervisorRestart(vmid, now, reason, func(st *controllerSupState) {
+		st.unresponsiveRestartAt = now
+		st.unhealthySeen = 0
+		st.unrespRefusal = ""
+	})
 }
 
 func (s *Server) resetNotRunning(vmid int) {
@@ -446,6 +591,14 @@ func (s *Server) ControllerSupervisorStatus(_ context.Context) *hub.ControllerSu
 			g.SlowCrashloopSince = st.slowCrashloopSince.UTC().Format(time.RFC3339)
 			g.SlowCrashloop = now.Sub(st.slowCrashloopSince) < controllerSlowCrashloopWindow
 		}
+		g.Unresponsive = st.unresponsive
+		if !st.unresponsiveSince.IsZero() {
+			g.UnresponsiveSince = st.unresponsiveSince.UTC().Format(time.RFC3339)
+		}
+		g.UnresponsiveRestartAfterS = (controllerUnresponsiveRestart - controllerUnresponsiveConfirm) * int(controllerSupervisorInterval/time.Second)
+		if !st.unresponsiveRestartAt.IsZero() {
+			g.UnresponsiveRestartAt = st.unresponsiveRestartAt.UTC().Format(time.RFC3339)
+		}
 		out.Guests = append(out.Guests, g)
 	}
 	sort.Slice(out.Guests, func(i, j int) bool { return out.Guests[i].VMID < out.Guests[j].VMID })

internal/localapi/cutwrite_main_test.go

@@ -0,0 +1,21 @@
+package localapi
+
+import (
+	"os"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
+)
+
+// TestMain points the config write journal (config.CutWriteDir, R-1380) at a temp folder: the real one is the
+// agent's state folder on a box, and a test must not need it.
+func TestMain(m *testing.M) {
+	dir, err := os.MkdirTemp("", "felhom-cutwrite-")
+	if err != nil {
+		panic(err)
+	}
+	config.CutWriteDir = dir
+	code := m.Run()
+	_ = os.RemoveAll(dir)
+	os.Exit(code)
+}

internal/localapi/disks.go

@@ -87,6 +87,11 @@ type GuestAttacher interface {
 	// GuestBootID returns a token that changes on every guest boot (host or guest) but is stable across a
 	// controller-only restart — the deterministic guest-reboot signal the controller recreates apps on.
 	GuestBootID(ctx context.Context, vmid int) string
+	// GuestBootKey is the same signal built from the kernel's boot_id, which a step of the host's clock cannot
+	// move (R-1395). "" = not readable; the controller then reads GuestBootID.
+	GuestBootKey(ctx context.Context, vmid int) string
+	// GuestBootTokens returns (GuestBootID, GuestBootKey) from one read — what /disks sends.
+	GuestBootTokens(ctx context.Context, vmid int) (id, key string)
 	// AttachBind is the LEGACY per-drive `pct set -mpN` bind (pre-intermediary). Retained for the
 	// transition; new attaches use AttachDrive.
 	AttachBind(ctx context.Context, vmid int, mountKey, where string) error
@@ -431,11 +436,12 @@ func (s *Server) handleDisks(w http.ResponseWriter, r *http.Request, vmid int) {
 
 	// Guest boot-id (intermediary model): changes on every guest boot, stable across controller restarts.
 	// The controller persists it and deterministically recreates drive-backed apps when it changes.
-	bootID := ""
+	// R-1395: guest_boot_key is the clock-proof form (kernel boot_id); guest_boot_id stays for older controllers.
+	bootID, bootKey := "", ""
 	if s.guestAttach != nil {
-		bootID = s.guestAttach.GuestBootID(r.Context(), vmid)
+		bootID, bootKey = s.guestAttach.GuestBootTokens(r.Context(), vmid)
 	}
-	writeOK(w, map[string]any{"vmid": vmid, "disks": out, "guest_boot_id": bootID})
+	writeOK(w, map[string]any{"vmid": vmid, "disks": out, "guest_boot_id": bootID, "guest_boot_key": bootKey})
 }
 
 // handleDiskCandidates (Impl-2a) lists the host disks FREE for Felhom to enroll — the raw-device

internal/localapi/disks_test.go

@@ -466,6 +466,10 @@ func (f *fakeGuestAttacher) AttachDrive(_ context.Context, _ int, where string)
 }
 func (f *fakeGuestAttacher) GuestSeesMount(_ context.Context, _ int, _ string) bool { return true }
 func (f *fakeGuestAttacher) GuestBootID(_ context.Context, _ int) string            { return "boot-1" }
+func (f *fakeGuestAttacher) GuestBootKey(_ context.Context, _ int) string           { return "key-1" }
+func (f *fakeGuestAttacher) GuestBootTokens(_ context.Context, _ int) (string, string) {
+	return "boot-1", "key-1"
+}
 func (f *fakeGuestAttacher) attachDriveCount() int {
 	f.mu.Lock()
 	defer f.mu.Unlock()

internal/localapi/guest_boot_key_test.go

@@ -0,0 +1,116 @@
+package localapi
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"log/slog"
+	"strings"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+type initPIDRunner struct{ pid string }
+
+func (r initPIDRunner) Run(_ context.Context, name string, _ ...string) ([]byte, []byte, error) {
+	if name == "lxc-info" {
+		return []byte(r.pid + "\n"), nil, nil
+	}
+	return nil, nil, nil
+}
+
+func (r initPIDRunner) RunStdin(ctx context.Context, _ io.Reader, name string, args ...string) ([]byte, []byte, error) {
+	return r.Run(ctx, name, args...)
+}
+
+// R-1395: on 2026-11-14 the N100's time service came back and stepped the host's clock by +32.7 s. The kernel
+// computes the boot TIME from the wall clock, so guest_boot_id moved and the controller recreated every drive-backed
+// app. The consequence asserted: the boot KEY does not move when the boot time moves; it does move on a host reboot
+// and on a guest reboot; and it is "" (so the controller falls back to the old token) when the boot_id is unreadable.
+// COMPANION RED-PROOFS (observed, felhom.eu audits/night-build-2026-11-14/C/redproof-agent-boot-key.txt).
+func TestGuestBootKey_AClockStepDoesNotMoveIt(t *testing.T) {
+	oldB, oldK, oldS := hostBtimeFn, hostBootIDFn, procStarttimeFn
+	t.Cleanup(func() { hostBtimeFn, hostBootIDFn, procStarttimeFn = oldB, oldK, oldS })
+	btime, bootID, start := "1794112518", "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11", "4242"
+	hostBtimeFn = func() string { return btime }
+	hostBootIDFn = func() string { return bootID }
+	procStarttimeFn = func(pid string) string {
+		if pid != "31337" {
+			t.Errorf("start time read for pid %q, want the guest's init", pid)
+		}
+		return start
+	}
+	b := NewGuestBinder(initPIDRunner{"31337"}, slog.New(slog.NewTextHandler(io.Discard, nil)))
+	ctx := context.Background()
+
+	id0, key0 := b.GuestBootID(ctx, 9201), b.GuestBootKey(ctx, 9201)
+	if id0 != "1794112518-4242" || key0 != "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11-4242" {
+		t.Fatalf("tokens = %q / %q", id0, key0)
+	}
+	btime = "1794112550" // THE STEP: the wall clock jumped +32 s, so the kernel's btime did too
+	id1, key1 := b.GuestBootID(ctx, 9201), b.GuestBootKey(ctx, 9201)
+	if id1 == id0 {
+		t.Fatalf("control: the old token did not move with the boot time (%q) — the test does not stage the defect", id1)
+	}
+	if key1 != key0 {
+		t.Fatalf("the boot key moved with the clock: %q → %q — a clock step reads as a new boot", key0, key1)
+	}
+	if strings.Contains(key1, "1794") {
+		t.Errorf("the boot key carries the boot time: %q", key1)
+	}
+	bootID = "c0ffee00-1111-2222-3333-444455556666" // a host reboot
+	if k := b.GuestBootKey(ctx, 9201); k == key0 {
+		t.Errorf("a host reboot (new boot_id) did not change the key: %q", k)
+	}
+	bootID, start = "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11", "991234" // a guest reboot
+	if k := b.GuestBootKey(ctx, 9201); k == key0 || !strings.HasSuffix(k, "-991234") {
+		t.Errorf("a guest reboot (new init start) did not change the key: %q", k)
+	}
+	for _, bad := range []string{"", "not-a-boot-id", "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11\nx"} {
+		bootID = bad
+		if k := b.GuestBootKey(ctx, 9201); k != "" {
+			t.Errorf("boot_id %q gave the key %q, want \"\" (the controller then reads the old token)", bad, k)
+		}
+		if id := b.GuestBootID(ctx, 9201); id == "" {
+			t.Errorf("the old token must still be sent when the boot_id is unreadable")
+		}
+	}
+	// what /disks sends: both tokens from ONE read of the guest's init — the init is asked for once, not twice
+	bootID, start = "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11", "4242"
+	reads := 0
+	procStarttimeFn = func(string) string { reads++; return start }
+	if id, key := b.GuestBootTokens(ctx, 9201); id != "1794112550-4242" || key != "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11-4242" || reads != 1 {
+		t.Errorf("GuestBootTokens = %q, %q after %d read(s) of the init — want both tokens from ONE read", id, key, reads)
+	}
+	bootID, start = "5b1f6c1e-8f0a-4c56-9d1c-0a8c3b7d2e11", "" // the guest does not run
+	if k := b.GuestBootKey(ctx, 9201); k != "" {
+		t.Errorf("a guest that does not run gave the key %q", k)
+	}
+}
+
+// The key is no use if /disks does not carry it: both tokens are on the wire, the old one unchanged for controllers
+// that do not know the key.
+func TestDisks_CarriesBothBootTokens(t *testing.T) {
+	srv := targetRowServer(t, "", []hub.StorageTarget{})
+	srv.guestAttach = &fakeGuestAttacher{}
+	body := do(t, srv.Handler(), "GET", "/disks", "A", "").Body.Bytes()
+	var w struct {
+		Data struct {
+			ID  string `json:"guest_boot_id"`
+			Key string `json:"guest_boot_key"`
+		} `json:"data"`
+		ID  string `json:"guest_boot_id"`
+		Key string `json:"guest_boot_key"`
+	}
+	if err := json.Unmarshal(body, &w); err != nil {
+		t.Fatalf("decode: %v (%s)", err, body)
+	}
+	if w.Data.ID+w.ID != "boot-1" || w.Data.Key+w.Key != "key-1" {
+		t.Fatalf("/disks boot tokens: id=%q key=%q, want boot-1 and key-1 (%s)", w.Data.ID+w.ID, w.Data.Key+w.Key, body)
+	}
+}
+
+// The real binder is what main.go hands the server: it must have the method (the interface makes this a build
+// error if not — stated here so the reason is on record).
+var _ GuestAttacher = (*GuestBinder)(nil)

internal/localapi/guestpower.go

@@ -3,6 +3,7 @@ package localapi
 import (
 	"context"
 	"log/slog"
+	"sort"
 	"sync"
 	"time"
 
@@ -82,6 +83,40 @@ type guestPowerState struct {
 	attempts int
 	nextAt   time.Time
 	raised   bool // the give-up fault has already been raised for this run
+	raisedAt time.Time
+}
+
+// GuestPowerGiveUp is one guest the watchdog stopped trying to start (R-1383, D78). It rides the host report
+// (`guest_power_gave_up`) so the operator hears it: until this, the give-up was one ERROR line in the host's journal
+// while the household's apps were off. Memory only — an agent restart forgets it (and tries three times again).
+type GuestPowerGiveUp struct {
+	VMID     int    `json:"vmid"`
+	Since    string `json:"since"` // RFC3339 UTC, when the watchdog gave up
+	Attempts int    `json:"attempts"`
+}
+
+// GuestPowerGaveUpStanza is GuestPowerGaveUp shaped for the host report: an UNTYPED nil when there is none, so the
+// report omits the field (a typed nil slice in an `any` would be sent as `null`).
+func (s *Server) GuestPowerGaveUpStanza() any {
+	if g := s.GuestPowerGaveUp(); len(g) > 0 {
+		return g
+	}
+	return nil
+}
+
+// GuestPowerGaveUp returns the guests the watchdog has given up on, by vmid. nil when there is none — the report
+// stanza is then omitted. Pinned by TestGuestPower_GiveUpIsReportedUntilTheGuestRuns.
+func (s *Server) GuestPowerGaveUp() []GuestPowerGiveUp {
+	guestPowerMu.Lock()
+	defer guestPowerMu.Unlock()
+	var out []GuestPowerGiveUp
+	for vmid, st := range s.guestPower {
+		if st.raised {
+			out = append(out, GuestPowerGiveUp{VMID: vmid, Since: st.raisedAt.UTC().Format(time.RFC3339), Attempts: st.attempts})
+		}
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].VMID < out[j].VMID })
+	return out
 }
 
 // WatchGuestPower runs the guest-power sweep every guestPowerInterval until ctx is done. No-op when
@@ -117,6 +152,19 @@ func (s *Server) GuestPowerTick(ctx context.Context) {
 		s.logger.Warn("guest-power: guest list unavailable — skipping sweep (ownership unproven)", "err", err)
 		return
 	}
+	// A guest that left the list (destroyed, or re-made under another number) takes its attempt record with it — a
+	// give-up for a guest that no longer exists would ride every host report until the agent restarts.
+	present := make(map[int]bool, len(guests))
+	for _, g := range guests {
+		present[g.VMID] = true
+	}
+	guestPowerMu.Lock()
+	for vmid := range s.guestPower {
+		if !present[vmid] {
+			delete(s.guestPower, vmid)
+		}
+	}
+	guestPowerMu.Unlock()
 	var stopped int
 	for _, g := range guests {
 		if ctx.Err() != nil {
@@ -152,6 +200,11 @@ func (s *Server) recoverOneStoppedGuest(ctx context.Context, g proxmox.Guest) {
 		s.forgetGuestPower(g.VMID) // healthy again: clear any attempt history
 		return
 	}
+	// D65: a guest the operator's reboot has down for a moment is pct reboot's to start, not ours (a racing start
+	// would fail the reboot's own start half).
+	if s.rebooting(g.VMID) {
+		return
+	}
 
 	lock, onboot, err := s.staleLock.Lock(ctx, g.VMID)
 	if err != nil {
@@ -162,6 +215,7 @@ func (s *Server) recoverOneStoppedGuest(ctx context.Context, g proxmox.Guest) {
 	// SCENARIO B — a deliberately stopped guest is left alone, forever. onboot:0 means the operator
 	// (or the golden-image provisioning) does not want this guest running.
 	if !onboot {
+		s.forgetGuestPower(g.VMID) // not ours to start any more: a give-up for it must not be reported for ever
 		return
 	}
 
@@ -171,6 +225,7 @@ func (s *Server) recoverOneStoppedGuest(ctx context.Context, g proxmox.Guest) {
 	if lock != "" {
 		s.logger.Info("guest-power: guest is stopped but LOCKED — leaving it to the stale-lock path",
 			"vmid", g.VMID, "lock", lock)
+		s.forgetGuestPower(g.VMID) // another operation owns it now; if it ends stopped, the three tries start again
 		return
 	}
 
@@ -255,6 +310,7 @@ func (s *Server) markGuestPowerRaised(vmid int) {
 	defer guestPowerMu.Unlock()
 	st := s.guestPower[vmid]
 	st.raised = true
+	st.raisedAt = s.now()
 	s.guestPower[vmid] = st
 }
 

internal/localapi/guestpower_test.go

@@ -2,6 +2,7 @@ package localapi
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"log/slog"
 	"testing"
@@ -237,3 +238,81 @@ func TestGuestPower_RunningGuestClearsAttemptHistory(t *testing.T) {
 		t.Error("attempt history survived the guest coming back healthy")
 	}
 }
+
+// R-1383 (D78): the give-up was ONE journal line while the household's apps were off. It now rides the host report:
+// nothing before the give-up (and an untyped nil, so the report omits the field), the guest with the time and the
+// attempt count after it, and nothing again once the guest runs.
+// COMPANION RED-PROOFS (observed, felhom.eu audits/night-build-2026-11-14/B/redproof-agent-guest-power.txt).
+func TestGuestPower_GiveUpIsReportedUntilTheGuestRuns(t *testing.T) {
+	ctl := &fakeGuestPowerCtl{
+		guests:   []proxmox.Guest{{VMID: 9201, Status: "stopped"}},
+		locks:    map[int]string{9201: ""},
+		onboot:   map[int]bool{9201: true},
+		startErr: errors.New("cannot start: storage offline"),
+	}
+	now := time.Date(2026, 11, 14, 2, 0, 0, 0, time.UTC)
+	s := gpServer(t, ctl, func() time.Time { return now })
+	if st := s.GuestPowerGaveUpStanza(); st != nil {
+		t.Fatalf("before any attempt the stanza is %#v, want an untyped nil (the report omits the field)", st)
+	}
+	for i := 0; i < guestPowerMaxAttempts; i++ { // the three failed starts
+		s.GuestPowerTick(context.Background())
+		if g := s.GuestPowerGaveUp(); len(g) != 0 {
+			t.Fatalf("after %d failed start(s) the watchdog still tries, but it reports a give-up: %+v", i+1, g)
+		}
+		now = now.Add(10 * time.Minute)
+	}
+	gaveUpAt := now
+	s.GuestPowerTick(context.Background()) // the tick that gives up
+	now = now.Add(time.Hour)
+	s.GuestPowerTick(context.Background())
+	g := s.GuestPowerGaveUp()
+	if len(g) != 1 || g[0].VMID != 9201 || g[0].Attempts != guestPowerMaxAttempts || g[0].Since != gaveUpAt.Format(time.RFC3339) {
+		t.Fatalf("give-up list = %+v, want guest 9201, %d attempts, since %s (the give-up's time, not a later tick's)", g, guestPowerMaxAttempts, gaveUpAt.Format(time.RFC3339))
+	}
+	b, _ := json.Marshal(s.GuestPowerGaveUpStanza())
+	if string(b) != `[{"vmid":9201,"since":"2026-11-14T02:30:00Z","attempts":3}]` {
+		t.Fatalf("the stanza on the wire = %s", b)
+	}
+	// the operator starts the guest: the next sweep sees it running and the list is empty again
+	ctl.guests = []proxmox.Guest{{VMID: 9201, Status: "running"}}
+	s.GuestPowerTick(context.Background())
+	if st := s.GuestPowerGaveUpStanza(); st != nil {
+		t.Fatalf("the guest runs again and the stanza is still %#v", st)
+	}
+}
+
+// Found by review, 2026-11-14: the give-up was forgotten only when the guest was seen RUNNING. A guest the operator
+// then switches to onboot:0, a guest another operation locks, and a guest that is destroyed would be reported as
+// „given up” for ever (the hub's alarm could never clear). Each of the three ends the report.
+func TestGuestPower_GiveUpEndsWhenTheGuestIsNoLongerOurs(t *testing.T) {
+	for _, c := range []struct {
+		name   string
+		change func(ctl *fakeGuestPowerCtl)
+	}{
+		{"onboot switched off", func(ctl *fakeGuestPowerCtl) { ctl.onboot = map[int]bool{9201: false} }},
+		{"locked by another operation", func(ctl *fakeGuestPowerCtl) { ctl.locks = map[int]string{9201: "backup"} }},
+		{"destroyed", func(ctl *fakeGuestPowerCtl) { ctl.guests = nil }},
+	} {
+		ctl := &fakeGuestPowerCtl{
+			guests:   []proxmox.Guest{{VMID: 9201, Status: "stopped"}},
+			locks:    map[int]string{9201: ""},
+			onboot:   map[int]bool{9201: true},
+			startErr: errors.New("cannot start"),
+		}
+		now := time.Date(2026, 11, 14, 2, 0, 0, 0, time.UTC)
+		s := gpServer(t, ctl, func() time.Time { return now })
+		for i := 0; i < guestPowerMaxAttempts+1; i++ {
+			s.GuestPowerTick(context.Background())
+			now = now.Add(10 * time.Minute)
+		}
+		if len(s.GuestPowerGaveUp()) != 1 {
+			t.Fatalf("%s: setup — no give-up recorded", c.name)
+		}
+		c.change(ctl)
+		s.GuestPowerTick(context.Background())
+		if g := s.GuestPowerGaveUp(); len(g) != 0 {
+			t.Errorf("%s: the give-up is still reported: %+v", c.name, g)
+		}
+	}
+}

internal/localapi/intermediary.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"fmt"
 	"os"
+	"regexp"
 	"strconv"
 	"strings"
 )
@@ -288,21 +289,82 @@ func (b *GuestBinder) guestInitPID(ctx context.Context, vmid int) string {
 // /proc/<pid>/stat — ticks since host boot, unique per process launch) changes on a guest reboot. ""
 // on any read error (the controller then keeps its last-seen → no spurious recreate).
 func (b *GuestBinder) GuestBootID(ctx context.Context, vmid int) string {
-	pid := b.guestInitPID(ctx, vmid)
-	if pid == "" {
-		return ""
-	}
-	start := procStarttime(pid)
+	start := b.guestInitStart(ctx, vmid)
 	if start == "" {
 		return ""
 	}
-	bt := hostBtime()
+	bt := hostBtimeFn()
 	if bt == "" {
 		bt = "0"
 	}
 	return bt + "-" + start
 }
 
+// guestInitStart is the guest init's start time in ticks since the host's boot ("" = the guest does not run, or a
+// read error). Both boot tokens end with it.
+func (b *GuestBinder) guestInitStart(ctx context.Context, vmid int) string {
+	pid := b.guestInitPID(ctx, vmid)
+	if pid == "" {
+		return ""
+	}
+	return procStarttimeFn(pid)
+}
+
+// GuestBootKey (R-1395) is GuestBootID with the host's boot TIME replaced by the kernel's boot_id:
+// "<boot_id>-<guest-init-starttime>". The kernel computes the boot time (`btime`) from the wall clock, so a STEP of
+// the host's clock moves GuestBootID — on the N100 on 2026-11-14 the time service came back, stepped +32.7 s, and the
+// controller recreated every drive-backed app as if the guest had booted. The kernel's boot_id is a random value made
+// once per boot; no clock can move it. Same meaning as before: it changes on a host reboot (new boot_id) and on a
+// guest reboot (new init start), and is stable across an agent or controller restart.
+//
+// "" when the guest does not run or the boot_id cannot be read — the controller then decides by GuestBootID, which
+// is still sent (controllers that do not know the key read only that). Pinned by TestGuestBootKey_*.
+func (b *GuestBinder) GuestBootKey(ctx context.Context, vmid int) string {
+	start := b.guestInitStart(ctx, vmid)
+	return guestBootKey(hostBootIDFn(), start)
+}
+
+// GuestBootTokens returns both tokens from ONE read of the guest's init: the old token and the key always describe
+// the same moment. Read separately, one could succeed and the other fail (the guest stops between the two) — the
+// controller would then get an old token with no key, read „an older agent”, and recreate every drive-backed app.
+func (b *GuestBinder) GuestBootTokens(ctx context.Context, vmid int) (id, key string) {
+	start := b.guestInitStart(ctx, vmid)
+	if start == "" {
+		return "", ""
+	}
+	bt := hostBtimeFn()
+	if bt == "" {
+		bt = "0"
+	}
+	return bt + "-" + start, guestBootKey(hostBootIDFn(), start)
+}
+
+// guestBootKey joins the two halves; "" when either is missing or the boot_id is not the kernel's shape.
+func guestBootKey(bootID, start string) string {
+	if start == "" || !kernelBootIDRe.MatchString(bootID) {
+		return ""
+	}
+	return bootID + "-" + start
+}
+
+var kernelBootIDRe = regexp.MustCompile(`^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$`)
+
+// The three reads of /proc behind the boot tokens, as seams (the tests move the boot time and keep the boot_id).
+var (
+	hostBtimeFn     = hostBtime
+	hostBootIDFn    = hostBootID
+	procStarttimeFn = procStarttime
+)
+
+// hostBootID returns the kernel's boot_id (/proc/sys/kernel/random/boot_id): random, made once per boot. "" on error.
+func hostBootID() string {
+	data, err := os.ReadFile("/proc/sys/kernel/random/boot_id")
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(data))
+}
+
 // procStarttime returns field 22 (starttime) of /proc/<pid>/stat. The comm field (2) can contain spaces
 // and parentheses, so we split AFTER the last ')': field 22 is index 19 of the post-comm fields
 // (field 3 = state is index 0). "" on any error.

internal/localapi/opactions.go

@@ -0,0 +1,258 @@
+package localapi
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// The operator's actions for THIS agent (operator brief 2026-10-29, Part 1; felhom.eu `04` §6.1).
+//
+// THE DOOR IS THE HOST REPORT'S REPLY, the controller's shape: the operator presses a button on the hub's host page, the
+// hub stores a row and pokes the box (the poke's out-of-cycle report gets the reply in seconds; without the tunnel the
+// 15-minute cycle carries it), the reply lists `operator_actions: [{id, action, arg}]`, this agent acts ONCE per id, and
+// the next report answers `operator_action_results: [{id, outcome, message}]` — the agent asks for that report at once.
+// The result is re-sent until the reply stops listing the id.
+//
+// THE LIST IS CLOSED: two verbs, restart_controller <vmid> and (D65 = A, 2026-10-29) reboot_guest <vmid> (opreboot.go). Anything else — a verb, a vmid that does not parse, a guest
+// this agent did not provision — is refused and NOTHING is called. Pinned by TestOpActions_AgentClosedList.
+//
+// THE ACT is the supervisor's own (controllersupervisor.go): `systemctl restart felhom-controller-bootstrap.service` in
+// the guest over the existing grants, with the supervisor's guards — not during a controller swap, not parked, not on a
+// stopped, locked or vzdump-busy guest, not when the guest list or the vzdump state cannot be read — and the thrash brake
+// counts these restarts too: 3 restarts (the supervisor's or the operator's) inside 15 minutes refuse the press, and
+// presses stay refused for 30 minutes. It works with the controller stopped or stuck: the agent does it, not the
+// controller. Every decision is one `audit:` line (the customer-visible audit form of reconcile.SlogAudit).
+
+const (
+	// OpRestartController is the agent's first operator verb.
+	OpRestartController = "restart_controller"
+
+	opOutcomeDone    = "done"
+	opOutcomeRefused = "refused"
+	opOutcomeFailed  = "failed"
+
+	// opBrakeRefusal is how long presses stay refused after the brake refused one.
+	opBrakeRefusal = 30 * time.Minute
+	// opForgetAfter drops a started id the reply no longer lists.
+	opForgetAfter = 24 * time.Hour
+
+	defaultOpHealthPoll = 2 * time.Second
+	// defaultOpHealthWait covers the bootstrap's pull-less rm/run (seconds), Docker's start period (10 s) and its first
+	// probe (30 s interval) with room: the controller's own HEALTHCHECK is the answer.
+	defaultOpHealthWait = 3 * time.Minute
+	// opHealthTemplate reads running + health over the swap's `docker inspect -f` grant.
+	opHealthTemplate = "{{.State.Running}}|{{if .State.Health}}{{.State.Health.Status}}{{else}}none{{end}}"
+)
+
+// operatorActionNames is THE agent's closed list. Adding an entry fails TestOpActions_AgentClosedList by design.
+var operatorActionNames = []string{OpRestartController, OpRebootGuest}
+
+// OperatorActionNames returns the closed list (sorted copy).
+func OperatorActionNames() []string {
+	out := append([]string(nil), operatorActionNames...)
+	sort.Strings(out)
+	return out
+}
+
+type opActionState struct {
+	mu      sync.Mutex
+	started map[int64]time.Time
+	results map[int64]hub.OperatorActionResult
+}
+
+// SetOperatorActionTrigger wires the report loop's out-of-band trigger (the result goes at once, not in 15 minutes).
+func (s *Server) SetOperatorActionTrigger(f func()) { s.opTrigger = f }
+
+// OnEnvelope is the report loop's hook (hub.EnvelopeObserver): it starts each newly listed action once and forgets
+// the ids the reply no longer lists (the hub has their results).
+func (s *Server) OnEnvelope(_ context.Context, env *hub.ControlEnvelope) {
+	if env == nil {
+		return
+	}
+	listed := map[int64]bool{}
+	var start []hub.OperatorAction
+	now := time.Now()
+	s.opActs.mu.Lock()
+	if s.opActs.started == nil {
+		s.opActs.started = map[int64]time.Time{}
+		s.opActs.results = map[int64]hub.OperatorActionResult{}
+	}
+	for _, a := range env.OperatorActions {
+		listed[a.ID] = true
+		if _, ok := s.opActs.started[a.ID]; ok {
+			continue
+		}
+		s.opActs.started[a.ID] = now
+		start = append(start, a)
+	}
+	for id, at := range s.opActs.started {
+		if listed[id] {
+			continue
+		}
+		if _, answered := s.opActs.results[id]; answered || now.Sub(at) > opForgetAfter {
+			delete(s.opActs.started, id)
+			delete(s.opActs.results, id)
+		}
+	}
+	s.opActs.mu.Unlock()
+	for _, a := range start {
+		go s.runOperatorAction(a) // a restart waits for the health answer — never block the report loop
+	}
+}
+
+// OperatorActionResults is the report's answer list, oldest id first.
+func (s *Server) OperatorActionResults() []hub.OperatorActionResult {
+	s.opActs.mu.Lock()
+	defer s.opActs.mu.Unlock()
+	out := make([]hub.OperatorActionResult, 0, len(s.opActs.results))
+	for _, r := range s.opActs.results {
+		out = append(out, r)
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].ID < out[j].ID })
+	return out
+}
+
+func (s *Server) runOperatorAction(a hub.OperatorAction) {
+	limit := s.opWait() + time.Minute
+	if a.Action == OpRebootGuest {
+		limit += opRebootTimeout
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), limit)
+	defer cancel()
+	outcome, msg := opOutcomeRefused, fmt.Sprintf("unknown action %q — the agent's list holds restart_controller and reboot_guest only", a.Action)
+	switch a.Action {
+	case OpRestartController:
+		outcome, msg = s.operatorRestartController(ctx, a.ID, a.Arg)
+	case OpRebootGuest:
+		outcome, msg = s.operatorRebootGuest(ctx, a.ID, a.Arg)
+	}
+	s.logger.Warn("audit: operator action", "id", a.ID, "action", a.Action, "arg", a.Arg, "by", "the operator via the hub",
+		"outcome", outcome, "message", msg)
+	s.opActs.mu.Lock()
+	if _, still := s.opActs.started[a.ID]; still {
+		s.opActs.results[a.ID] = hub.OperatorActionResult{ID: a.ID, Outcome: outcome, Message: msg}
+	}
+	s.opActs.mu.Unlock()
+	if s.opTrigger != nil {
+		s.opTrigger()
+	}
+}
+
+func (s *Server) opPoll() time.Duration {
+	if s.opHealthPoll > 0 {
+		return s.opHealthPoll
+	}
+	return defaultOpHealthPoll
+}
+
+func (s *Server) opWait() time.Duration {
+	if s.opHealthWait > 0 {
+		return s.opHealthWait
+	}
+	return defaultOpHealthWait
+}
+
+// operatorRestartController is the press: the supervisor's guards, the brake, the restart, then the wait for the
+// controller's healthy answer. Returns the outcome and the sentence the host page shows.
+func (s *Server) operatorRestartController(ctx context.Context, id int64, arg string) (string, string) {
+	vmid, err := strconv.Atoi(arg)
+	if err != nil || strconv.Itoa(vmid) != arg {
+		return opOutcomeRefused, fmt.Sprintf("refused: %q is not a guest id", arg)
+	}
+	if s.staleLock == nil || s.guestExec == nil {
+		return opOutcomeRefused, "refused: the controller supervisor is not wired on this agent"
+	}
+	if !s.provisionedGuest(vmid) {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d is not a guest this agent provisioned", vmid)
+	}
+	guests, err := s.staleLock.Guests(ctx)
+	if err != nil {
+		return opOutcomeRefused, "refused: the guest list cannot be read — ownership unproven, nothing touched"
+	}
+	status := ""
+	for _, g := range guests {
+		if g.VMID == vmid {
+			status = g.Status
+		}
+	}
+	if status != "running" {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d is not running (%s) — the guest-power watchdog owns it", vmid, orDash(status))
+	}
+	if s.controllerParked(vmid) {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d's controller is parked (remove the controller-parked marker first)", vmid)
+	}
+	s.swapMu.Lock()
+	swapping := s.swapInFlight[vmid]
+	s.swapMu.Unlock()
+	if swapping {
+		return opOutcomeRefused, "refused: a controller swap (update) is in flight — it owns the controller"
+	}
+	if lock, _, err := s.staleLock.Lock(ctx, vmid); err != nil {
+		return opOutcomeRefused, "refused: the guest lock cannot be read — nothing touched"
+	} else if lock != "" {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d is locked (%s) — another operation owns it", vmid, lock)
+	}
+	if busy, berr := s.staleLock.BackupRunning(ctx, vmid); berr != nil || busy {
+		return opOutcomeRefused, "refused: a whole-guest backup (vzdump) may be in flight for the guest"
+	}
+
+	// The brake: every restart counts — the supervisor's and the operator's.
+	now := s.clock()
+	s.ctrlSup.mu.Lock()
+	st := s.supState(vmid)
+	if now.Before(st.opBrakeUntil) {
+		until := st.opBrakeUntil
+		s.ctrlSup.mu.Unlock()
+		return opOutcomeRefused, fmt.Sprintf("refused: 3 restarts in 15 minutes — presses are refused until %s UTC", until.UTC().Format("15:04"))
+	}
+	st.restarts = pruneBefore(st.restarts, now.Add(-controllerCrashloopWindow))
+	if len(st.restarts) >= controllerCrashloopMax {
+		st.opBrakeUntil = now.Add(opBrakeRefusal)
+		until := st.opBrakeUntil
+		s.ctrlSup.mu.Unlock()
+		return opOutcomeRefused, fmt.Sprintf("refused: 3 restarts in 15 minutes — presses are refused until %s UTC", until.UTC().Format("15:04"))
+	}
+	reason := fmt.Sprintf("the operator via the hub (action #%d)", id)
+	st.restarts = append(st.restarts, now)
+	st.restartsTotal++
+	st.lastRestartAt = now
+	st.lastReason = reason
+	st.notRunningSeen = 0
+	s.ctrlSup.mu.Unlock()
+
+	s.logger.Warn("controller-supervisor: RESTARTING the controller on the operator's press", "vmid", vmid, "action_id", id, "unit", bootstrapUnit)
+	began := time.Now()
+	if _, err := s.guestExec.GuestExec(ctx, vmid, "systemctl", "restart", bootstrapUnit); err != nil {
+		return opOutcomeFailed, fmt.Sprintf("the restart of guest %d's controller failed: %v", vmid, err)
+	}
+	for time.Since(began) < s.opWait() {
+		out, err := s.guestExec.GuestExec(ctx, vmid, "docker", "inspect", "-f", opHealthTemplate, controllerContainer)
+		if err == nil {
+			f := strings.SplitN(strings.TrimSpace(out), "|", 2)
+			if len(f) == 2 && f[0] == "true" && (f[1] == "healthy" || f[1] == "none") {
+				secs := int(time.Since(began).Round(time.Second) / time.Second)
+				return opOutcomeDone, fmt.Sprintf("restarted; the controller answered after %d s (Docker's health check: healthy)", secs)
+			}
+		}
+		select {
+		case <-ctx.Done():
+			return opOutcomeFailed, "restarted, but the wait for its answer was cut"
+		case <-time.After(s.opPoll()):
+		}
+	}
+	return opOutcomeFailed, fmt.Sprintf("restarted, but no healthy answer within %s — look at the box", s.opWait().Round(time.Second))
+}
+
+func orDash(s string) string {
+	if s == "" {
+		return "not listed"
+	}
+	return s
+}

internal/localapi/opactions_test.go

@@ -0,0 +1,222 @@
+package localapi
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"reflect"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// Operator brief 2026-10-29, Part 1: „Restart controller" from the hub. The agent's closed list holds one verb; the act
+// is the supervisor's own restart with the supervisor's guards, and the brake counts operator restarts too.
+
+// opExec answers the supervisor's fake plus the health read the operator restart waits on.
+type opExec struct {
+	*supExec
+	health string // what {{.State.Health.Status}} reads after the restart
+}
+
+func (f *opExec) GuestExec(ctx context.Context, vmid int, args ...string) (string, error) {
+	if len(args) >= 5 && args[0] == "docker" && args[1] == "inspect" && args[3] == opHealthTemplate {
+		f.mu.Lock()
+		st := f.status[vmid]
+		f.mu.Unlock()
+		if st == "" {
+			return "", errors.New("pct exec: exit status 1: Error: No such object: felhom-controller")
+		}
+		return map[bool]string{true: "true", false: "false"}[st == "running"] + "|" + f.health + "\n", nil
+	}
+	return f.supExec.GuestExec(ctx, vmid, args...)
+}
+
+func opServer(t *testing.T, status string) (*Server, *opExec, *fakeGuestPowerCtl, *supClock, string, *atomic.Int32) {
+	t.Helper()
+	ex := &opExec{supExec: &supExec{status: map[int]string{9201: status}, onRestart: "running"}, health: "healthy"}
+	ctl := runningGuest(9201)
+	s, clk, dir := supServer(t, nil, ctl, 9201)
+	s.guestExec = ex
+	s.opHealthPoll = time.Millisecond
+	s.opHealthWait = 2 * time.Second
+	var triggered atomic.Int32
+	s.SetOperatorActionTrigger(func() { triggered.Add(1) })
+	return s, ex, ctl, clk, dir, &triggered
+}
+
+func env(acts ...hub.OperatorAction) *hub.ControlEnvelope {
+	return &hub.ControlEnvelope{OperatorActions: acts}
+}
+
+// waitResult waits for the action's result (the act runs in the background).
+func waitResult(t *testing.T, s *Server, id int64) hub.OperatorActionResult {
+	t.Helper()
+	deadline := time.Now().Add(3 * time.Second)
+	for time.Now().Before(deadline) {
+		for _, r := range s.OperatorActionResults() {
+			if r.ID == id {
+				return r
+			}
+		}
+		time.Sleep(2 * time.Millisecond)
+	}
+	t.Fatalf("no result for #%d", id)
+	return hub.OperatorActionResult{}
+}
+
+// RED-PROOF: seen failing before the list existed (build failure, then the names).
+func TestOpActions_AgentClosedList(t *testing.T) {
+	if got, want := OperatorActionNames(), []string{"reboot_guest", "restart_controller"}; !reflect.DeepEqual(got, want) {
+		t.Fatalf("the agent's operator actions = %v, want exactly %v — a new verb needs the operator's word", got, want)
+	}
+}
+
+// Anything outside the list — and a bad argument — is refused and NOTHING is called.
+func TestOpActions_UnknownOrBadArgIsRefused(t *testing.T) {
+	s, ex, _, _, _, _ := opServer(t, "running")
+	s.OnEnvelope(context.Background(), env(
+		hub.OperatorAction{ID: 1, Action: "reboot_host", Arg: "9201"}, // never on the list: only the guest is rebooted
+		hub.OperatorAction{ID: 2, Action: "restart_controller", Arg: "abc"},
+		hub.OperatorAction{ID: 3, Action: "restart_controller", Arg: "9202"}, // not provisioned here
+	))
+	for _, id := range []int64{1, 2, 3} {
+		if r := waitResult(t, s, id); r.Outcome != "refused" {
+			t.Errorf("#%d: %+v, want refused", id, r)
+		}
+	}
+	if n := ex.count(9201); n != 0 {
+		t.Fatalf("a refused action restarted the controller (%d)", n)
+	}
+}
+
+// The press restarts ONCE, says how long the controller took to answer, asks for a report at once, and re-sends the
+// result until the hub stops listing the id. RED-PROOF: drop the GuestExec restart → restarts 0 → FAILS.
+func TestOpActions_RestartReportsTheAnswerTime(t *testing.T) {
+	s, ex, _, _, _, trig := opServer(t, "running")
+	a := hub.OperatorAction{ID: 7, Action: "restart_controller", Arg: "9201"}
+	s.OnEnvelope(context.Background(), env(a))
+	r := waitResult(t, s, 7)
+	if r.Outcome != "done" || !strings.Contains(r.Message, "answered after") {
+		t.Fatalf("result %+v", r)
+	}
+	if n := ex.count(9201); n != 1 {
+		t.Fatalf("restarts = %d, want 1", n)
+	}
+	if trig.Load() == 0 {
+		t.Fatal("the agent did not ask for a report — the result would wait 15 minutes")
+	}
+	s.OnEnvelope(context.Background(), env(a)) // the hub still lists it: re-send, never act again
+	time.Sleep(100 * time.Millisecond)         // a wrong second act runs in the background — give it time to show
+	if n := ex.count(9201); n != 1 {
+		t.Fatalf("a re-listed id acted again (%d)", n)
+	}
+	if len(s.OperatorActionResults()) != 1 {
+		t.Fatal("the result must be re-sent while listed")
+	}
+	s.OnEnvelope(context.Background(), env()) // the hub recorded it
+	if left := s.OperatorActionResults(); len(left) != 0 {
+		t.Fatalf("a recorded result is still sent: %+v", left)
+	}
+}
+
+// The point of the agent doing it: a STOPPED controller is restarted too (the controller cannot restart itself).
+func TestOpActions_WorksWithTheControllerStopped(t *testing.T) {
+	s, ex, _, _, _, _ := opServer(t, "exited")
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: 8, Action: "restart_controller", Arg: "9201"}))
+	if r := waitResult(t, s, 8); r.Outcome != "done" || ex.count(9201) != 1 {
+		t.Fatalf("result %+v, restarts %d", r, ex.count(9201))
+	}
+}
+
+// Each supervisor guard refuses the press, with the reason. RED-PROOF: drop the lock check → the locked case
+// restarts → FAILS (each guard seen failing by its own mutation).
+func TestOpActions_GuardsRefuse(t *testing.T) {
+	cases := []struct {
+		name, want string
+		setup      func(s *Server, ctl *fakeGuestPowerCtl, dir string)
+	}{
+		{"parked", "parked", func(_ *Server, _ *fakeGuestPowerCtl, dir string) {
+			if err := os.WriteFile(filepath.Join(dir, "9201", ControllerParkedMarker), nil, 0o600); err != nil {
+				panic(err)
+			}
+		}},
+		{"swap", "swap", func(s *Server, _ *fakeGuestPowerCtl, _ string) { s.swapInFlight[9201] = true }},
+		{"locked", "locked", func(_ *Server, ctl *fakeGuestPowerCtl, _ string) { ctl.locks[9201] = "backup" }},
+		{"vzdump", "backup", func(_ *Server, ctl *fakeGuestPowerCtl, _ string) { ctl.backupRun = map[int]bool{9201: true} }},
+		{"vzdump unknown", "backup", func(_ *Server, ctl *fakeGuestPowerCtl, _ string) { ctl.backupErr = errors.New("tasks unreadable") }},
+		{"guest stopped", "not running", func(_ *Server, ctl *fakeGuestPowerCtl, _ string) { ctl.guests[0].Status = "stopped" }},
+		{"guest list", "guest list", func(_ *Server, ctl *fakeGuestPowerCtl, _ string) { ctl.guestsErr = errors.New("api down") }},
+	}
+	for i, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			s, ex, ctl, _, dir, _ := opServer(t, "running")
+			tc.setup(s, ctl, dir)
+			id := int64(100 + i)
+			s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: id, Action: "restart_controller", Arg: "9201"}))
+			r := waitResult(t, s, id)
+			if r.Outcome != "refused" || !strings.Contains(strings.ToLower(r.Message), tc.want) {
+				t.Fatalf("result %+v, want refused naming %q", r, tc.want)
+			}
+			if n := ex.count(9201); n != 0 {
+				t.Fatalf("guard %q did not hold: restarted %d", tc.name, n)
+			}
+		})
+	}
+}
+
+// The thrash brake counts operator restarts AND the supervisor's: 3 in 15 minutes → the next press is refused, and
+// presses stay refused for 30 minutes. RED-PROOF: drop the brake → the 4th press restarts → FAILS.
+func TestOpActions_TheBrakeCountsEveryRestart(t *testing.T) {
+	s, ex, _, clk, _, _ := opServer(t, "running")
+	// one restart by the supervisor itself
+	ex.mu.Lock()
+	ex.status[9201] = "exited"
+	ex.mu.Unlock()
+	s.ControllerSupervisorTick(context.Background())
+	s.ControllerSupervisorTick(context.Background())
+	if ex.count(9201) != 1 {
+		t.Fatalf("setup: the supervisor did not restart (%d)", ex.count(9201))
+	}
+	for id := int64(1); id <= 2; id++ {
+		clk.t = clk.t.Add(time.Minute)
+		s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: id, Action: "restart_controller", Arg: "9201"}))
+		if r := waitResult(t, s, id); r.Outcome != "done" {
+			t.Fatalf("press %d: %+v", id, r)
+		}
+	}
+	clk.t = clk.t.Add(time.Minute)
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: 3, Action: "restart_controller", Arg: "9201"}))
+	r := waitResult(t, s, 3)
+	if r.Outcome != "refused" || !strings.Contains(r.Message, "3 restarts in 15 minutes") {
+		t.Fatalf("the 4th restart in 15 minutes: %+v", r)
+	}
+	if n := ex.count(9201); n != 3 {
+		t.Fatalf("restarts = %d, want 3", n)
+	}
+	clk.t = clk.t.Add(20 * time.Minute) // the window has passed, but the 30-minute refusal has not
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: 4, Action: "restart_controller", Arg: "9201"}))
+	if r := waitResult(t, s, 4); r.Outcome != "refused" {
+		t.Fatalf("inside the 30-minute refusal: %+v", r)
+	}
+	clk.t = clk.t.Add(11 * time.Minute)
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: 5, Action: "restart_controller", Arg: "9201"}))
+	if r := waitResult(t, s, 5); r.Outcome != "done" {
+		t.Fatalf("after the refusal: %+v", r)
+	}
+}
+
+// A controller that does not come back healthy is a FAILED press, said plainly.
+func TestOpActions_NoHealthyAnswerIsFailed(t *testing.T) {
+	s, ex, _, _, _, _ := opServer(t, "running")
+	ex.health = "unhealthy"
+	s.opHealthWait = 30 * time.Millisecond
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: 9, Action: "restart_controller", Arg: "9201"}))
+	if r := waitResult(t, s, 9); r.Outcome != "failed" || !strings.Contains(r.Message, "no healthy answer") {
+		t.Fatalf("result %+v", r)
+	}
+}

internal/localapi/opreboot.go

@@ -0,0 +1,329 @@
+package localapi
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// reboot_guest — the operator reboots a guest from the hub (D65 = A, operator ruling 2026-10-29 15:50; felhom.eu `04`
+// §6.1, `03`). No signature: a reboot destroys nothing (`03`, „healing … non-destructive by construction").
+//
+// WHY: a stuck Docker or a stuck network inside the box is out of reach of „Restart controller" — that restarts one
+// container. The reboot is the existing guest-reboot path (`pct reboot <vmid>`, the household's „A Felhom
+// újraindítása", the same exact sudo vector) — of THAT guest, never the host.
+//
+// THE GUARDS (each refuses with the reason in the result; nothing is called):
+//   - the guest is not one this agent provisioned, is not running, is locked, or its controller is parked;
+//   - a controller swap (update) is in flight;
+//   - a heavy operation holds the host-wide gate (a backup of any tier the agent runs, a restore test, an OS step),
+//     or jobBlocker finds one (an agent backup job, a drive or NAS job, an escrow ceremony, a vzdump of the guest);
+//   - the CONTROLLER says it is busy: a backup, a restore, an off-site run or an app update (its `ctl/busy` file,
+//     read over one exact `pct exec … cat` grant; controllerBusy below). Fails closed on any read error but „no file";
+//   - at most ONE reboot per hour per guest (the record below, persisted — it survives an agent restart);
+//   - the operator restart's brake: a reboot counts as a restart, and the brake's refusal refuses a reboot too.
+// The heavy-operation gate is HELD from the checks until the guest is up again, so no backup, restore test or OS
+// step of this agent can start in between. The guest-power watchdog and the controller supervisor leave the guest
+// alone while the reboot runs (rebootInFlight).
+//
+// AFTER: the result says when the guest was up and when the controller answered (both from the press). The record
+// `<guests>/<vmid>/operator-reboot.json` is written BEFORE the reboot, so the controller that starts after it finds
+// it: GET /guest/operator-reboot hands it to the household's dashboard („Felhom support restarted your box at …",
+// one day). Pinned by opreboot_test.go.
+
+const (
+	// OpRebootGuest is the agent's second operator verb (D65 = A).
+	OpRebootGuest = "reboot_guest"
+
+	// opRebootEvery: at most one operator reboot per guest per hour.
+	opRebootEvery = time.Hour
+	// opRebootNoticeFor: how long the household's dashboard shows the line.
+	opRebootNoticeFor = 24 * time.Hour
+	// opRebootTimeout bounds `pct reboot` (it blocks until the guest is back).
+	opRebootTimeout = 5 * time.Minute
+
+	operatorRebootFile = "operator-reboot.json"
+
+	// ControllerBusyMarkerGuestPath is the controller's busy file as the GUEST sees it: the controller data volume
+	// (`felhom-controller-data`, mounted at /opt/docker/felhom-controller in the container) + data/ctl/busy. One exact
+	// sudo vector reads it (configs/felhom-agent.sudoers, FELHOM_CONTROLLERSWAP). It reads without Docker, so a stuck
+	// Docker does not hide it.
+	ControllerBusyMarkerGuestPath = "/var/lib/docker/volumes/felhom-controller-data/_data/data/ctl/busy"
+	// controllerBusyFresh: the controller rewrites the file every 15 s while busy; an older file is a controller that
+	// stopped writing (stuck or killed) — not a busy one.
+	controllerBusyFresh = 90 * time.Second
+)
+
+// operatorRebootRecord is the persisted record of the last operator reboot of a guest.
+type operatorRebootRecord struct {
+	At               time.Time `json:"at"`
+	ActionID         int64     `json:"action_id"`
+	GuestUpAfterS    int       `json:"guest_up_after_s,omitempty"`
+	ControllerAfterS int       `json:"controller_after_s,omitempty"`
+}
+
+func (s *Server) operatorRebootPath(vmid int) string {
+	return filepath.Join(s.guestsStateDir(), strconv.Itoa(vmid), operatorRebootFile)
+}
+
+// lastOperatorReboot reads the record; ok=false when there is none. A corrupt record is an error (the hourly limit
+// must not silently vanish).
+func (s *Server) lastOperatorReboot(vmid int) (operatorRebootRecord, bool, error) {
+	raw, err := os.ReadFile(s.operatorRebootPath(vmid))
+	if errors.Is(err, os.ErrNotExist) {
+		return operatorRebootRecord{}, false, nil
+	}
+	if err != nil {
+		return operatorRebootRecord{}, false, err
+	}
+	var rec operatorRebootRecord
+	if err := json.Unmarshal(raw, &rec); err != nil {
+		return operatorRebootRecord{}, false, fmt.Errorf("corrupt %s: %w", operatorRebootFile, err)
+	}
+	return rec, !rec.At.IsZero(), nil
+}
+
+func (s *Server) saveOperatorReboot(vmid int, rec operatorRebootRecord) error {
+	raw, err := json.Marshal(rec)
+	if err != nil {
+		return err
+	}
+	p := s.operatorRebootPath(vmid)
+	tmp := p + ".tmp"
+	if err := os.WriteFile(tmp, raw, 0o644); err != nil {
+		return err
+	}
+	return os.Rename(tmp, p)
+}
+
+// rebooting reports whether an operator reboot of the guest is in flight (the watchdogs leave it alone).
+func (s *Server) rebooting(vmid int) bool {
+	s.rebootMu.Lock()
+	defer s.rebootMu.Unlock()
+	return s.rebootInFlight[vmid]
+}
+
+func (s *Server) setRebooting(vmid int, on bool) {
+	s.rebootMu.Lock()
+	defer s.rebootMu.Unlock()
+	if s.rebootInFlight == nil {
+		s.rebootInFlight = map[int]bool{}
+	}
+	if on {
+		s.rebootInFlight[vmid] = true
+	} else {
+		delete(s.rebootInFlight, vmid)
+	}
+}
+
+func (s *Server) rebooter() interface {
+	RebootGuest(ctx context.Context, vmid int) error
+} {
+	if s.opRebooter != nil {
+		return s.opRebooter
+	}
+	if s.guestAttach != nil {
+		return s.guestAttach
+	}
+	return nil
+}
+
+// controllerBusy reads the controller's busy file. busy=true with the controller's reason when the file is fresh;
+// a missing file is „not busy" (an idle controller, or one older than the file); any other read failure is an error
+// (the caller refuses — it cannot prove no backup runs). Two callers: the operator's reboot (D65) and R-1344's restart
+// of a controller that does not answer (controllersupervisor.go restartUnresponsive).
+func (s *Server) controllerBusy(ctx context.Context, vmid int) (bool, string, error) {
+	out, err := s.guestExec.GuestExec(ctx, vmid, "cat", ControllerBusyMarkerGuestPath)
+	if err != nil {
+		if strings.Contains(strings.ToLower(err.Error()+" "+out), "no such file") {
+			return false, "", nil
+		}
+		return false, "", err
+	}
+	lines := strings.SplitN(strings.TrimSpace(out), "\n", 2)
+	at, perr := time.Parse(time.RFC3339, strings.TrimSpace(lines[0]))
+	if perr != nil {
+		return false, "", fmt.Errorf("the controller's busy file does not start with a time: %q", lines[0])
+	}
+	if s.clock().Sub(at) > controllerBusyFresh {
+		s.logger.Info("controller-busy: the controller's busy file is stale — the controller stopped writing it; not busy",
+			"vmid", vmid, "written", at.UTC().Format(time.RFC3339))
+		return false, "", nil
+	}
+	why := "work in progress"
+	if len(lines) == 2 && strings.TrimSpace(lines[1]) != "" {
+		why = strings.TrimSpace(lines[1])
+	}
+	return true, why, nil
+}
+
+// operatorRebootGuest is the press: every guard, then the reboot, then the wait for the guest and the controller.
+func (s *Server) operatorRebootGuest(ctx context.Context, id int64, arg string) (string, string) {
+	vmid, err := strconv.Atoi(arg)
+	if err != nil || strconv.Itoa(vmid) != arg {
+		return opOutcomeRefused, fmt.Sprintf("refused: %q is not a guest id", arg)
+	}
+	rb := s.rebooter()
+	if s.staleLock == nil || s.guestExec == nil || rb == nil {
+		return opOutcomeRefused, "refused: the guest reboot is not wired on this agent"
+	}
+	if !s.provisionedGuest(vmid) {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d is not a guest this agent provisioned", vmid)
+	}
+	if s.rebooting(vmid) {
+		return opOutcomeRefused, fmt.Sprintf("refused: a reboot of guest %d is already running", vmid)
+	}
+	guests, err := s.staleLock.Guests(ctx)
+	if err != nil {
+		return opOutcomeRefused, "refused: the guest list cannot be read — ownership unproven, nothing touched"
+	}
+	status := ""
+	for _, g := range guests {
+		if g.VMID == vmid {
+			status = g.Status
+		}
+	}
+	if status != "running" {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d is not running (%s) — the guest-power watchdog owns it", vmid, orDash(status))
+	}
+	if s.controllerParked(vmid) {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d's controller is parked (remove the controller-parked marker first)", vmid)
+	}
+	// A controller swap (update) is refused by jobBlocker below, with the agent's other jobs.
+	if lock, _, err := s.staleLock.Lock(ctx, vmid); err != nil {
+		return opOutcomeRefused, "refused: the guest lock cannot be read — nothing touched"
+	} else if lock != "" {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d is locked (%s) — another operation owns it", vmid, lock)
+	}
+
+	// At most one per hour.
+	now := s.clock()
+	last, had, err := s.lastOperatorReboot(vmid)
+	if err != nil {
+		return opOutcomeRefused, "refused: the last reboot's record cannot be read — " + err.Error()
+	}
+	if had && now.Sub(last.At) < opRebootEvery {
+		return opOutcomeRefused, fmt.Sprintf("refused: guest %d was rebooted from the hub at %s UTC — at most one reboot per hour (next from %s UTC)",
+			vmid, last.At.UTC().Format("15:04"), last.At.Add(opRebootEvery).UTC().Format("15:04"))
+	}
+
+	// The brake: a refusal in force refuses a reboot too, and 3 restarts in 15 minutes refuse it.
+	s.ctrlSup.mu.Lock()
+	st := s.supState(vmid)
+	if now.Before(st.opBrakeUntil) {
+		until := st.opBrakeUntil
+		s.ctrlSup.mu.Unlock()
+		return opOutcomeRefused, fmt.Sprintf("refused: 3 restarts in 15 minutes — presses are refused until %s UTC", until.UTC().Format("15:04"))
+	}
+	st.restarts = pruneBefore(st.restarts, now.Add(-controllerCrashloopWindow))
+	if len(st.restarts) >= controllerCrashloopMax {
+		st.opBrakeUntil = now.Add(opBrakeRefusal)
+		until := st.opBrakeUntil
+		s.ctrlSup.mu.Unlock()
+		return opOutcomeRefused, fmt.Sprintf("refused: 3 restarts in 15 minutes — presses are refused until %s UTC", until.UTC().Format("15:04"))
+	}
+	s.ctrlSup.mu.Unlock()
+
+	// The heavy-operation gate, held until the guest is up again.
+	release, busy, ok := s.inFlight.TryAcquire("guest-reboot")
+	if !ok {
+		return opOutcomeRefused, "refused: a backup or another heavy operation runs on the host (" + busy + ")"
+	}
+	defer release()
+	if code, detail := s.jobBlocker(ctx, vmid); code != "" {
+		return opOutcomeRefused, "refused: " + rebootBlockerSentence(code) + " (" + detail + ")"
+	}
+	if cbusy, why, err := s.controllerBusy(ctx, vmid); err != nil {
+		return opOutcomeRefused, "refused: the controller's busy file cannot be read, so a running backup cannot be ruled out — " + err.Error()
+	} else if cbusy {
+		return opOutcomeRefused, "refused: the controller is busy — " + why
+	}
+
+	// The record FIRST: the hourly limit and the household's line both read it.
+	rec := operatorRebootRecord{At: now, ActionID: id}
+	if err := s.saveOperatorReboot(vmid, rec); err != nil {
+		return opOutcomeRefused, "refused: the reboot cannot be recorded — " + err.Error()
+	}
+	s.ctrlSup.mu.Lock()
+	st = s.supState(vmid)
+	st.restarts = append(st.restarts, now)
+	st.restartsTotal++
+	st.lastRestartAt = now
+	st.lastReason = fmt.Sprintf("guest reboot by the operator via the hub (action #%d)", id)
+	st.notRunningSeen = 0
+	s.ctrlSup.mu.Unlock()
+	s.setRebooting(vmid, true)
+	defer s.setRebooting(vmid, false)
+
+	s.logger.Warn("operator reboot: REBOOTING the guest on the operator's press", "vmid", vmid, "action_id", id)
+	began := time.Now()
+	rctx, cancel := context.WithTimeout(ctx, opRebootTimeout)
+	err = rb.RebootGuest(rctx, vmid)
+	cancel()
+	if err != nil {
+		return opOutcomeFailed, fmt.Sprintf("the reboot of guest %d failed: %v — look at the box (the guest-power watchdog starts a stopped guest)", vmid, err)
+	}
+	upS := int(time.Since(began).Round(time.Second) / time.Second)
+	rec.GuestUpAfterS = upS
+	// The controller's own HEALTHCHECK is the answer, as for a restart; the wait starts when the guest is up.
+	for deadline := time.Now().Add(s.opWait()); time.Now().Before(deadline); {
+		out, err := s.guestExec.GuestExec(ctx, vmid, "docker", "inspect", "-f", opHealthTemplate, controllerContainer)
+		if err == nil {
+			f := strings.SplitN(strings.TrimSpace(out), "|", 2)
+			if len(f) == 2 && f[0] == "true" && (f[1] == "healthy" || f[1] == "none") {
+				rec.ControllerAfterS = int(time.Since(began).Round(time.Second) / time.Second)
+				if serr := s.saveOperatorReboot(vmid, rec); serr != nil {
+					s.logger.Warn("operator reboot: the times could not be added to the record", "vmid", vmid, "err", serr)
+				}
+				return opOutcomeDone, fmt.Sprintf("rebooted; the guest was up after %d s, the controller answered after %d s", upS, rec.ControllerAfterS)
+			}
+		}
+		select {
+		case <-ctx.Done():
+			return opOutcomeFailed, fmt.Sprintf("rebooted; the guest was up after %d s, but the wait for the controller was cut", upS)
+		case <-time.After(s.opPoll()):
+		}
+	}
+	_ = s.saveOperatorReboot(vmid, rec)
+	return opOutcomeFailed, fmt.Sprintf("rebooted; the guest was up after %d s, but the controller gave no healthy answer within %s — look at the box", upS, s.opWait().Round(time.Second))
+}
+
+// rebootBlockerSentence names jobBlocker's code in the result's words.
+func rebootBlockerSentence(code string) string {
+	switch code {
+	case PowerOffBackup:
+		return "a backup is running"
+	case PowerOffCtrlUpdate:
+		return "a controller update is running"
+	case PowerOffDrive:
+		return "a drive or network-storage job is running"
+	case PowerOffEscrow:
+		return "an escrow ceremony is running"
+	case PowerOffUnknown:
+		return "the running jobs cannot be read"
+	default:
+		return "the host is busy"
+	}
+}
+
+// handleOperatorRebootNotice — GET /guest/operator-reboot (self-scoped): the last operator reboot of THIS guest, when it
+// is less than a day old, for the household's dashboard line. {"at": ""} when there is none.
+func (s *Server) handleOperatorRebootNotice(w http.ResponseWriter, _ *http.Request, vmid int) {
+	rec, ok, err := s.lastOperatorReboot(vmid)
+	if err != nil {
+		writeErr(w, http.StatusInternalServerError, "the reboot record cannot be read")
+		return
+	}
+	if !ok || s.clock().Sub(rec.At) >= opRebootNoticeFor {
+		writeOK(w, map[string]any{"vmid": vmid, "at": ""})
+		return
+	}
+	writeOK(w, map[string]any{"vmid": vmid, "at": rec.At.UTC().Format(time.RFC3339), "by": "operator"})
+}

internal/localapi/opreboot_test.go

@@ -0,0 +1,306 @@
+package localapi
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/backup"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// D65 = A (operator 2026-10-29 15:50): „Reboot guest" from the hub. The agent's reboot_guest — the existing
+// `pct reboot` of THAT guest — with its guards, the hourly limit, the brake, and the record the household's line reads.
+
+// rbExec adds the controller's busy file to the operator-restart fake.
+type rbExec struct {
+	*opExec
+	busy    string // the file's content; "" = no file
+	busyErr error  // pct exec itself failed
+}
+
+func (f *rbExec) GuestExec(ctx context.Context, vmid int, args ...string) (string, error) {
+	if len(args) == 2 && args[0] == "cat" && args[1] == ControllerBusyMarkerGuestPath {
+		if f.busyErr != nil {
+			return "", f.busyErr
+		}
+		if f.busy == "" {
+			return "", errors.New("pct exec 9201 [cat ...]: exit status 1: cat: " + ControllerBusyMarkerGuestPath + ": No such file or directory")
+		}
+		return f.busy, nil
+	}
+	return f.opExec.GuestExec(ctx, vmid, args...)
+}
+
+type fakeRebooter struct {
+	mu    sync.Mutex
+	n     int
+	err   error
+	after func()
+}
+
+func (r *fakeRebooter) RebootGuest(context.Context, int) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.n++
+	if r.after != nil {
+		r.after()
+	}
+	return r.err
+}
+
+func (r *fakeRebooter) count() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.n
+}
+
+func rbServer(t *testing.T) (*Server, *rbExec, *fakeRebooter, *fakeGuestPowerCtl, *supClock, string) {
+	t.Helper()
+	s, ox, ctl, clk, dir, _ := opServer(t, "running")
+	ex := &rbExec{opExec: ox}
+	s.guestExec = ex
+	rb := &fakeRebooter{}
+	s.opRebooter = rb
+	s.inFlight = &backup.InFlight{}
+	return s, ex, rb, ctl, clk, dir
+}
+
+func pressReboot(t *testing.T, s *Server, id int64) hub.OperatorActionResult {
+	t.Helper()
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: id, Action: "reboot_guest", Arg: "9201"}))
+	return waitResult(t, s, id)
+}
+
+// The press reboots ONCE and says when the guest was up and when the controller answered; the record is written for
+// the household's line. RED-PROOF: drop the RebootGuest call → reboots 0 → FAILS.
+func TestOpReboot_RebootsAndReportsTheTimes(t *testing.T) {
+	s, _, rb, _, clk, _ := rbServer(t)
+	r := pressReboot(t, s, 1)
+	if r.Outcome != "done" || !strings.Contains(r.Message, "the guest was up after") || !strings.Contains(r.Message, "the controller answered after") {
+		t.Fatalf("result %+v", r)
+	}
+	if rb.count() != 1 {
+		t.Fatalf("reboots = %d, want 1", rb.count())
+	}
+	rec, ok, err := s.lastOperatorReboot(9201)
+	if err != nil || !ok || !rec.At.Equal(clk.t) || rec.ActionID != 1 {
+		t.Fatalf("record %+v ok=%v err=%v", rec, ok, err)
+	}
+	if s.inFlight.Busy() != "" {
+		t.Fatalf("the heavy-operation gate is still held after the reboot: %q", s.inFlight.Busy())
+	}
+	if s.rebooting(9201) {
+		t.Fatal("the reboot-in-flight flag outlived the reboot")
+	}
+}
+
+// Each guard refuses with its reason and NOTHING is rebooted. RED-PROOF: each guard seen failing by deleting it.
+func TestOpReboot_GuardsRefuse(t *testing.T) {
+	fresh := func(clk *supClock) string {
+		return clk.t.Add(-20*time.Second).Format(time.RFC3339) + "\nan off-site backup is running\n"
+	}
+	cases := []struct {
+		name, want string
+		setup      func(s *Server, ex *rbExec, ctl *fakeGuestPowerCtl, clk *supClock, dir string)
+	}{
+		{"parked", "parked", func(_ *Server, _ *rbExec, _ *fakeGuestPowerCtl, _ *supClock, dir string) {
+			if err := os.WriteFile(filepath.Join(dir, "9201", ControllerParkedMarker), nil, 0o600); err != nil {
+				panic(err)
+			}
+		}},
+		{"swap", "swap", func(s *Server, _ *rbExec, _ *fakeGuestPowerCtl, _ *supClock, _ string) { s.swapInFlight[9201] = true }},
+		{"locked", "locked", func(_ *Server, _ *rbExec, ctl *fakeGuestPowerCtl, _ *supClock, _ string) { ctl.locks[9201] = "backup" }},
+		{"stopped", "not running", func(_ *Server, _ *rbExec, ctl *fakeGuestPowerCtl, _ *supClock, _ string) {
+			ctl.guests[0].Status = "stopped"
+		}},
+		{"vzdump", "backup", func(_ *Server, _ *rbExec, ctl *fakeGuestPowerCtl, _ *supClock, _ string) {
+			ctl.backupRun = map[int]bool{9201: true}
+		}},
+		{"vzdump unknown", "cannot be read", func(_ *Server, _ *rbExec, ctl *fakeGuestPowerCtl, _ *supClock, _ string) {
+			ctl.backupErr = errors.New("tasks unreadable")
+		}},
+		{"heavy gate", "backup:felhom-pbs", func(s *Server, _ *rbExec, _ *fakeGuestPowerCtl, _ *supClock, _ string) {
+			s.inFlight.TryAcquire("backup:felhom-pbs")
+		}},
+		{"agent backup job", "a backup is running", func(s *Server, _ *rbExec, _ *fakeGuestPowerCtl, _ *supClock, _ string) {
+			s.jobs = map[backupJobKey]*backupJob{{vmid: 9201, target: "local"}: {JobID: "j1", Phase: PhaseRunning}}
+		}},
+		{"controller busy", "off-site backup", func(_ *Server, ex *rbExec, _ *fakeGuestPowerCtl, clk *supClock, _ string) {
+			ex.busy = fresh(clk)
+		}},
+		{"busy file unreadable", "cannot be ruled out", func(_ *Server, ex *rbExec, _ *fakeGuestPowerCtl, _ *supClock, _ string) {
+			ex.busyErr = errors.New("pct exec 9201: exit status 255: lxc-attach failed")
+		}},
+		{"brake in force", "3 restarts in 15 minutes", func(s *Server, _ *rbExec, _ *fakeGuestPowerCtl, clk *supClock, _ string) {
+			s.ctrlSup.mu.Lock()
+			s.supState(9201).opBrakeUntil = clk.t.Add(10 * time.Minute)
+			s.ctrlSup.mu.Unlock()
+		}},
+		{"not provisioned", "not a guest this agent provisioned", func(_ *Server, _ *rbExec, _ *fakeGuestPowerCtl, _ *supClock, dir string) {
+			if err := os.RemoveAll(filepath.Join(dir, "9201", "bootstrap")); err != nil {
+				panic(err)
+			}
+		}},
+	}
+	for i, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			s, ex, rb, ctl, clk, dir := rbServer(t)
+			tc.setup(s, ex, ctl, clk, dir)
+			r := pressReboot(t, s, int64(100+i))
+			if r.Outcome != "refused" || !strings.Contains(r.Message, tc.want) {
+				t.Fatalf("result %+v, want refused naming %q", r, tc.want)
+			}
+			if rb.count() != 0 {
+				t.Fatalf("guard %q did not hold: rebooted", tc.name)
+			}
+			if _, ok, _ := s.lastOperatorReboot(9201); ok {
+				t.Fatal("a refused press wrote the reboot record (the household would read a reboot that never happened)")
+			}
+		})
+	}
+}
+
+// A busy file the controller stopped writing (stuck, killed) is not a busy controller — the reboot exists for exactly
+// that box. RED-PROOF: drop the freshness check → refused → FAILS.
+func TestOpReboot_StaleBusyFileDoesNotBlock(t *testing.T) {
+	s, ex, rb, _, clk, _ := rbServer(t)
+	ex.busy = clk.t.Add(-10*time.Minute).Format(time.RFC3339) + "\na backup is running\n"
+	if r := pressReboot(t, s, 1); r.Outcome != "done" || rb.count() != 1 {
+		t.Fatalf("result %+v, reboots %d", r, rb.count())
+	}
+}
+
+// At most one reboot per hour per guest, and the limit survives an agent restart (the record is on disk).
+// RED-PROOF: drop the hourly check → the second press reboots → FAILS.
+func TestOpReboot_AtMostOncePerHour(t *testing.T) {
+	s, _, rb, ctl, clk, dir := rbServer(t)
+	if r := pressReboot(t, s, 1); r.Outcome != "done" {
+		t.Fatalf("first press: %+v", r)
+	}
+	clk.t = clk.t.Add(59 * time.Minute)
+	if r := pressReboot(t, s, 2); r.Outcome != "refused" || !strings.Contains(r.Message, "at most one reboot per hour") {
+		t.Fatalf("second press inside the hour: %+v", r)
+	}
+	// a restarted agent (a fresh in-memory state over the same guests dir) still refuses
+	s2, ox2, _, _, _, _ := opServer(t, "running")
+	s2.guestsDir, s2.staleLock, s2.now = dir, ctl, clk.now
+	s2.guestExec = &rbExec{opExec: ox2}
+	rb2 := &fakeRebooter{}
+	s2.opRebooter, s2.inFlight = rb2, &backup.InFlight{}
+	if r := pressReboot(t, s2, 3); r.Outcome != "refused" || rb2.count() != 0 {
+		t.Fatalf("after an agent restart, inside the hour: %+v", r)
+	}
+	clk.t = clk.t.Add(2 * time.Minute)
+	if r := pressReboot(t, s, 4); r.Outcome != "done" || rb.count() != 2 {
+		t.Fatalf("after the hour: %+v (reboots %d)", r, rb.count())
+	}
+}
+
+// The brake counts a reboot as a restart: two operator restarts and one reboot inside 15 minutes refuse the next
+// restart press. RED-PROOF: drop the restarts append in operatorRebootGuest → the 4th press restarts → FAILS.
+func TestOpReboot_TheBrakeCountsAReboot(t *testing.T) {
+	s, ex, _, _, clk, _ := rbServer(t)
+	for id := int64(1); id <= 2; id++ {
+		s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: id, Action: "restart_controller", Arg: "9201"}))
+		if r := waitResult(t, s, id); r.Outcome != "done" {
+			t.Fatalf("restart %d: %+v", id, r)
+		}
+		clk.t = clk.t.Add(time.Minute)
+	}
+	if r := pressReboot(t, s, 3); r.Outcome != "done" {
+		t.Fatalf("reboot: %+v", r)
+	}
+	clk.t = clk.t.Add(time.Minute)
+	s.OnEnvelope(context.Background(), env(hub.OperatorAction{ID: 4, Action: "restart_controller", Arg: "9201"}))
+	if r := waitResult(t, s, 4); r.Outcome != "refused" || !strings.Contains(r.Message, "3 restarts in 15 minutes") {
+		t.Fatalf("the 4th restart (2 restarts + 1 reboot) in 15 minutes: %+v", r)
+	}
+	if n := ex.count(9201); n != 2 {
+		t.Fatalf("restarts = %d, want 2", n)
+	}
+}
+
+// A failed `pct reboot` is a FAILED press, said plainly; the gate and the flag are given back.
+func TestOpReboot_FailedRebootIsFailed(t *testing.T) {
+	s, _, rb, _, _, _ := rbServer(t)
+	rb.err = errors.New("pct reboot 9201: exit status 255")
+	if r := pressReboot(t, s, 1); r.Outcome != "failed" || !strings.Contains(r.Message, "failed") {
+		t.Fatalf("result %+v", r)
+	}
+	if s.inFlight.Busy() != "" || s.rebooting(9201) {
+		t.Fatal("the gate or the flag outlived a failed reboot")
+	}
+}
+
+// While the reboot runs, the guest-power watchdog does not start the (momentarily stopped) guest and the supervisor
+// does not restart its controller. RED-PROOF: drop either rebooting() check → a start or a restart → FAILS.
+func TestOpReboot_TheWatchdogsLeaveARebootingGuestAlone(t *testing.T) {
+	s, ex, rb, ctl, _, _ := rbServer(t)
+	var started, restarted int
+	rb.after = func() {
+		ctl.guests[0].Status = "stopped"
+		s.GuestPowerTick(context.Background())
+		ctl.guests[0].Status = "running"
+		ex.mu.Lock()
+		ex.status[9201] = "exited"
+		ex.mu.Unlock()
+		s.ControllerSupervisorTick(context.Background())
+		s.ControllerSupervisorTick(context.Background())
+		started, restarted = len(ctl.started), ex.count(9201)
+		ex.mu.Lock()
+		ex.status[9201] = "running"
+		ex.mu.Unlock()
+	}
+	if r := pressReboot(t, s, 1); r.Outcome != "done" {
+		t.Fatalf("result %+v", r)
+	}
+	if started != 0 || restarted != 0 {
+		t.Fatalf("during the reboot: guest-power started %d, the supervisor restarted %d — both must leave it to pct reboot", started, restarted)
+	}
+}
+
+// The household's line: GET /guest/operator-reboot returns the time for a day, then nothing.
+// RED-PROOF: drop the 24-hour check → the 25-hour read still returns the time → FAILS.
+func TestOpReboot_TheNoticeExpiresAfterADay(t *testing.T) {
+	s, _, _, _, clk, _ := rbServer(t)
+	read := func() string {
+		w := httptest.NewRecorder()
+		s.handleOperatorRebootNotice(w, httptest.NewRequest("GET", "/guest/operator-reboot", nil), 9201)
+		var body struct {
+			Data struct {
+				At string `json:"at"`
+			} `json:"data"`
+			At string `json:"at"`
+		}
+		if err := json.Unmarshal(w.Body.Bytes(), &body); err != nil {
+			t.Fatalf("body %q: %v", w.Body.String(), err)
+		}
+		if body.Data.At != "" {
+			return body.Data.At
+		}
+		return body.At
+	}
+	if got := read(); got != "" {
+		t.Fatalf("no reboot yet, got %q", got)
+	}
+	at := clk.t
+	if r := pressReboot(t, s, 1); r.Outcome != "done" {
+		t.Fatalf("result %+v", r)
+	}
+	clk.t = at.Add(23 * time.Hour)
+	if got := read(); got != at.UTC().Format(time.RFC3339) {
+		t.Fatalf("after 23 h: %q, want %s", got, at.UTC().Format(time.RFC3339))
+	}
+	clk.t = at.Add(25 * time.Hour)
+	if got := read(); got != "" {
+		t.Fatalf("after 25 h the line must be gone, got %q", got)
+	}
+}

internal/localapi/server.go

@@ -411,6 +411,20 @@ type Server struct {
 	guestsDir string
 	ctrlSup   controllerSupervisor
 
+	// 2026-10-29: the operator's actions from the hub's reply (opactions.go). opHealthPoll/opHealthWait bound the wait
+	// for a restarted controller's healthy answer (tests shorten them; zero → the defaults).
+	opActs       opActionState
+	opTrigger    func()
+	opHealthPoll time.Duration
+	opHealthWait time.Duration
+	// D65 = A (2026-10-29): the operator's guest reboot (opreboot.go). opRebooter is a test seam (nil → guestAttach);
+	// rebootInFlight keeps the guest-power watchdog and the controller supervisor off a guest being rebooted.
+	opRebooter interface {
+		RebootGuest(ctx context.Context, vmid int) error
+	}
+	rebootMu       sync.Mutex
+	rebootInFlight map[int]bool
+
 	// Network-storage verify job (SPIKE-nas-verify): the IN-MEMORY single slot + the seams the
 	// detached pipeline runs through (tests inject; production defaults set in NewServer).
 	netVerifyMu  sync.Mutex
@@ -596,6 +610,8 @@ func (s *Server) Handler() http.Handler {
 	mux.HandleFunc("POST /disks/guest-attach", s.withGuest(s.handleDiskGuestAttach))
 	// Guest reboot (slice 10 P2 activation): user-triggered restart to activate pending drive binds.
 	mux.HandleFunc("POST /guest/reboot", s.withGuest(s.handleGuestReboot))
+	// D65 = A: the last operator reboot of THIS guest (< 1 day), for the household's dashboard line (opreboot.go).
+	mux.HandleFunc("GET /guest/operator-reboot", s.withGuest(s.handleOperatorRebootNotice))
 	// Host power-off (R-998, doc 14): the household's „A gép kikapcsolása". Guarded in hostpoweroff.go.
 	mux.HandleFunc("POST /host/poweroff", s.withGuest(s.handleHostPowerOff))
 	// Guest RAM resize (v0.90.0, R-24): read current allocation + bounds, and apply a bounded resize

internal/localapi/stalelock_inflight_test.go

@@ -256,8 +256,10 @@ func TestBackupRunning_AllCallersUseTheOneHelper(t *testing.T) {
 			}
 		}
 	}
-	if callers != 5 {
-		t.Errorf("expected the five known callers (stale-lock, guest-power, power-off, supervisor, the R-1006 cleanup), found %d", callers)
+	// 2026-10-29: +2 deliberate callers — the operator's restart_controller (opactions.go) and R-1344's unresponsive count
+	// (controllersupervisor.go observeControllerHealth). Both through s.staleLock, the one helper.
+	if callers != 7 {
+		t.Errorf("expected the seven known callers (stale-lock, guest-power, power-off, supervisor, the R-1006 cleanup, the operator restart, the unresponsive count), found %d", callers)
 	}
 	b, _ := os.ReadFile("stalelock.go")
 	if !strings.Contains(string(b), "vzdumpInFlight(tasks, vmid, proc)") {

internal/localapi/supervisor_unresponsive_test.go

@@ -0,0 +1,156 @@
+package localapi
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+	"time"
+)
+
+// R-1344 (operator brief 2026-10-29, Part 2; decided 2026-10-30 from the night's measurement): a controller that RUNS
+// but whose Docker health check reads „unhealthy" on controllerUnresponsiveConfirm consecutive sweeps is reported
+// unresponsive (the stanza; the hub raises controller_unresponsive). If it still reads „unhealthy" at
+// controllerUnresponsiveRestart sweeps, the agent restarts it — with the not-running restart's guards and brake, plus
+// the controller's own busy file. RED-PROOF: the restart tests seen failing before the restart existed.
+
+func unrespServer(t *testing.T, health string) (*Server, *rbExec, *fakeGuestPowerCtl, *supClock) {
+	t.Helper()
+	s, ex, ctl, clk, _, _ := opServer(t, "running")
+	ex.health = health
+	rb := &rbExec{opExec: ex}
+	s.guestExec = rb
+	return s, rb, ctl, clk
+}
+
+func sweep(s *Server, n int) {
+	for i := 0; i < n; i++ {
+		s.ControllerSupervisorTick(context.Background())
+	}
+}
+
+func TestSupervisor_UnresponsiveIsAlarmedThenRestarted(t *testing.T) {
+	s, ex, _, _ := unrespServer(t, "unhealthy")
+	sweep(s, controllerUnresponsiveConfirm-1)
+	if g := s.ControllerSupervisorStatus(context.Background()).Guests; len(g) != 1 || g[0].Unresponsive {
+		t.Fatalf("unresponsive before %d sweeps: %+v", controllerUnresponsiveConfirm, g)
+	}
+	sweep(s, 1)
+	g := s.ControllerSupervisorStatus(context.Background()).Guests
+	if len(g) != 1 || !g[0].Unresponsive || g[0].UnresponsiveSince == "" {
+		t.Fatalf("after %d unhealthy sweeps: %+v", controllerUnresponsiveConfirm, g)
+	}
+	if g[0].UnresponsiveRestartAfterS != (controllerUnresponsiveRestart-controllerUnresponsiveConfirm)*int(controllerSupervisorInterval/time.Second) {
+		t.Fatalf("the stanza must tell the hub when the restart comes: %+v", g[0])
+	}
+	since := g[0].UnresponsiveSince
+	sweep(s, controllerUnresponsiveRestart-controllerUnresponsiveConfirm-1)
+	if n := ex.count(9201); n != 0 {
+		t.Fatalf("restarted after %d unhealthy sweeps — the restart waits for %d", controllerUnresponsiveRestart-1, controllerUnresponsiveRestart)
+	}
+	if g := s.ControllerSupervisorStatus(context.Background()).Guests; g[0].UnresponsiveSince != since {
+		t.Fatal("unresponsive_since moved while it stayed unresponsive — the hub would mail again")
+	}
+	sweep(s, 1)
+	if n := ex.count(9201); n != 1 {
+		t.Fatalf("after %d unhealthy sweeps: %d restarts, want 1", controllerUnresponsiveRestart, n)
+	}
+	g = s.ControllerSupervisorStatus(context.Background()).Guests
+	if g[0].UnresponsiveRestartAt == "" || g[0].UnresponsiveRestartAt != g[0].LastRestartAt || g[0].Restarts24h != 1 ||
+		!strings.Contains(g[0].LastReason, "health check") {
+		t.Fatalf("the restart is not on the record the hub reads: %+v", g[0])
+	}
+	// A restart that did not help (still „unhealthy" on the next sweep) is not repeated at once: the count starts again.
+	sweep(s, 1)
+	if n := ex.count(9201); n != 1 {
+		t.Fatalf("restarted again on the sweep after a restart: %d restarts", n)
+	}
+	// The restarted controller starts, then answers: the episode ends, its record stays.
+	ex.health = "starting"
+	sweep(s, 1)
+	if g := s.ControllerSupervisorStatus(context.Background()).Guests; g[0].Unresponsive || g[0].UnresponsiveSince != since {
+		t.Fatalf("after the restart: %+v (unresponsive must clear; since stays as the record)", g[0])
+	}
+	// A second episode needs the whole count again — never a restart on the first unhealthy read after one.
+	ex.health = "unhealthy"
+	sweep(s, controllerUnresponsiveRestart-1)
+	if n := ex.count(9201); n != 1 {
+		t.Fatalf("a second episode restarted early: %d restarts", n)
+	}
+}
+
+// Slow but answering (healthy / starting), a swap, a vzdump, a locked or a parked guest never counts — and so is never
+// restarted.
+func TestSupervisor_UnresponsiveGuards(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		setup func(s *Server, ex *rbExec, ctl *fakeGuestPowerCtl)
+	}{
+		{"healthy", func(_ *Server, ex *rbExec, _ *fakeGuestPowerCtl) { ex.health = "healthy" }},
+		{"starting", func(_ *Server, ex *rbExec, _ *fakeGuestPowerCtl) { ex.health = "starting" }},
+		{"swap", func(s *Server, _ *rbExec, _ *fakeGuestPowerCtl) { s.swapInFlight[9201] = true }},
+		{"vzdump", func(_ *Server, _ *rbExec, ctl *fakeGuestPowerCtl) { ctl.backupRun = map[int]bool{9201: true} }},
+		{"locked", func(_ *Server, _ *rbExec, ctl *fakeGuestPowerCtl) { ctl.locks[9201] = "backup" }},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			s, ex, ctl, _ := unrespServer(t, "unhealthy")
+			tc.setup(s, ex, ctl)
+			sweep(s, controllerUnresponsiveRestart+2)
+			if g := s.ControllerSupervisorStatus(context.Background()).Guests; len(g) == 1 && g[0].Unresponsive {
+				t.Fatalf("%s counted as unresponsive: %+v", tc.name, g[0])
+			}
+			if n := ex.count(9201); n != 0 {
+				t.Fatalf("%s: restarted %d times", tc.name, n)
+			}
+		})
+	}
+}
+
+// At the restart moment: the controller's busy file (fresh, or unreadable) and the brake refuse it; the alarm stays.
+// RED-PROOF: each seen failing by deleting its guard.
+func TestSupervisor_UnresponsiveRestartRefusals(t *testing.T) {
+	for _, tc := range []struct {
+		name      string
+		setup     func(s *Server, ex *rbExec, clk *supClock)
+		crashloop bool
+	}{
+		{"busy file fresh", func(_ *Server, ex *rbExec, clk *supClock) {
+			ex.busy = clk.t.Add(-20*time.Second).Format(time.RFC3339) + "\na backup, a restore or an off-site run is running\n"
+		}, false},
+		{"busy file unreadable", func(_ *Server, ex *rbExec, _ *supClock) { ex.busyErr = errors.New("pct exec: exit status 255") }, false},
+		{"brake: 3 restarts in 15 minutes", func(s *Server, _ *rbExec, clk *supClock) {
+			st := s.supState(9201)
+			st.restarts = []time.Time{clk.t.Add(-10 * time.Minute), clk.t.Add(-5 * time.Minute), clk.t.Add(-time.Minute)}
+		}, true},
+		{"crash-loop pause in force", func(s *Server, _ *rbExec, clk *supClock) {
+			s.supState(9201).crashloopSince = clk.t.Add(-5 * time.Minute)
+		}, true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			s, ex, _, clk := unrespServer(t, "unhealthy")
+			tc.setup(s, ex, clk)
+			sweep(s, controllerUnresponsiveRestart+3)
+			if n := ex.count(9201); n != 0 {
+				t.Fatalf("%s: restarted %d times", tc.name, n)
+			}
+			g := s.ControllerSupervisorStatus(context.Background()).Guests
+			if len(g) != 1 || !g[0].Unresponsive || g[0].UnresponsiveRestartAt != "" {
+				t.Fatalf("%s: the alarm must stay and no restart be recorded: %+v", tc.name, g)
+			}
+			if tc.crashloop && !g[0].Crashloop {
+				t.Fatalf("%s: the brake did not raise the crash-loop: %+v", tc.name, g[0])
+			}
+		})
+	}
+}
+
+// A busy file the controller stopped writing (older than controllerBusyFresh) is not „busy" — a hung controller must
+// not shield itself with its last word.
+func TestSupervisor_UnresponsiveStaleBusyFileDoesNotShield(t *testing.T) {
+	s, ex, _, clk := unrespServer(t, "unhealthy")
+	ex.busy = clk.t.Add(-10*time.Minute).Format(time.RFC3339) + "\nan app update is running\n"
+	sweep(s, controllerUnresponsiveRestart)
+	if n := ex.count(9201); n != 1 {
+		t.Fatalf("a stale busy file blocked the restart: %d restarts", n)
+	}
+}

internal/osupdate/kernel.go

@@ -21,6 +21,8 @@ import (
 	"fmt"
 	"log/slog"
 	"regexp"
+	"sort"
+	"strings"
 	"time"
 
 	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
@@ -444,3 +446,147 @@ func (l *Leg) StageKernelSigned(ctx context.Context, vmid int, p KernelStepParam
 	}
 	return l.finish(ctx, lg, rep)
 }
+
+// ---- the clean-up of old kernels (R-1368, `09` §3 decision 286, `11` §5.11.1) ----
+//
+// The rule, its refusals and "may this box do it" live in the root wrapper (mode kernel-cleanup); the plan names NO
+// package. This file decides only WHEN: in the night leg, after a healthy host step, never in a night with a kernel
+// step. The wrapper adds its own 20-hour window after a step.
+
+// KernelCleanupResult is the wrapper's kernel_cleanup object.
+type KernelCleanupResult struct {
+	Outcome string `json:"outcome"` // removed | nothing | not_run | dry_run | failed
+	DryRun  bool   `json:"dry_run"`
+	Would   string `json:"would"` // dry run only: remove | nothing | refuse | not_run
+	NotRun  []struct {
+		Code   string `json:"code"`
+		Reason string `json:"reason"`
+	} `json:"not_run"`
+	Refusals []struct {
+		Code   string `json:"code"`
+		Reason string `json:"reason"`
+	} `json:"refusals"`
+	Kept        map[string][]string `json:"kept"`
+	Removal     []string            `json:"removal"`
+	Removed     []string            `json:"removed"`
+	BytesGained *int64              `json:"bytes_gained"`
+	Leftovers   []string            `json:"leftovers"`
+}
+
+// kernelCleanupSkip says why tonight's pass must not call the clean-up ("" = call it).
+func kernelCleanupSkip(appliance bool, trigger string, host, pve, kernel Report, blk hub.WireOSUpdate, okStep func(Report) bool) string {
+	switch {
+	case trigger != "night":
+		return "it runs only in the night leg (never a debug pass)"
+	case !appliance:
+		return "not an appliance install (a BYO host belongs to its owner)"
+	case !blk.Enabled:
+		return "OS updates are switched off for this box"
+	case host.Layer == "" || !okStep(host):
+		return "no healthy host step this pass"
+	case pve.Layer != "" && !okStep(pve):
+		return "the Proxmox step did not end healthy"
+	case kernel.Layer != "":
+		return "a kernel step ran in this night (" + kernel.Outcome + ") — never in the same night"
+	}
+	if due, _ := kernelDue(blk, trigger); due {
+		return "a kernel step is due tonight — never in the same night"
+	}
+	return ""
+}
+
+// runKernelCleanup calls the wrapper once. The hub hears about it only when there is something to tell: kernels were
+// removed (outcome "cleaned"), the wrapper refused, or the purge failed. "Nothing to remove" and "not run" are log
+// lines — one report per box per night would bury the kernel step's own line on the hub's System page.
+func (l *Leg) runKernelCleanup(ctx context.Context, runID string, vmid int, trigger string, blk hub.WireOSUpdate) Report {
+	lg := l.log().With("run", runID, "layer", LayerKernel, "mode", "kernel-cleanup", "vmid", vmid, "trigger", trigger)
+	start := l.now()
+	wr, err := l.call(ctx, runID, kernelPlan("kernel-cleanup", vmid, map[string]any{"release_id": "kernel-cleanup",
+		"run_id": runID, "trigger": trigger, "ring": blk.Ring}))
+	// NO kernel view and no kernel version in this report: the hub reads a kernel-layer report's `kernel.to` /
+	// `release_id` as "the kernel this STEP was for" (kernelOf), and a clean-up is not a step for any kernel.
+	rep := Report{RunID: runID, Layer: LayerKernel, Trigger: trigger, Ring: blk.Ring, VMID: vmid, Mode: "kernel-cleanup",
+		ReleaseID: "kernel-cleanup", PassSeconds: wr.PassSeconds}
+	var res KernelCleanupResult
+	_ = json.Unmarshal(wr.KernelCleanup, &res)
+	took := l.now().Sub(start).Round(time.Second).String()
+	switch {
+	case err != nil:
+		lg.Warn("osupdate: kernel clean-up — the wrapper gave no answer; not retried tonight", "err", err, "took", took)
+		rep.Outcome, rep.HealthReason = "failed", "kernel clean-up: "+err.Error()
+		return l.finish(ctx, lg, rep)
+	case wr.refused():
+		lg.Warn("osupdate: kernel clean-up REFUSED — nothing was removed; not retried tonight", "refused", string(wr.Refused), "took", took)
+		rep.Outcome, rep.Refused, rep.HealthReason = "refused", wr.Refused, "the clean-up of old kernels was refused; nothing was removed"
+		return l.finish(ctx, lg, rep)
+	case wr.failed():
+		lg.Error("osupdate: kernel clean-up FAILED — look at the box; not retried tonight", "failed", string(wr.Failed), "took", took)
+		rep.Outcome, rep.Refused, rep.HealthReason = "failed", wr.Failed, "the clean-up of old kernels failed"
+		return l.finish(ctx, lg, rep)
+	}
+	switch res.Outcome {
+	case "removed":
+		gained := "unknown"
+		if res.BytesGained != nil {
+			gained = fmt.Sprintf("%.1f GiB", float64(*res.BytesGained)/(1<<30))
+		}
+		rep.Outcome, rep.Healthy = "cleaned", true
+		rep.KernelCleanup = &CleanupDetail{Removed: capNames(res.Removed, 40), Kept: capNames(sortedKeys(res.Kept), 40),
+			BytesGained: res.BytesGained, Leftovers: len(res.Leftovers)}
+		rep.HealthReason = truncate(fmt.Sprintf("removed %d old kernel package(s): %s; kept %s; %s gained; %d left-over path(s)",
+			len(res.Removed), strings.Join(res.Removed, ", "), strings.Join(sortedKeys(res.Kept), ", "), gained, len(res.Leftovers)), 600)
+		lg.Info("osupdate: kernel clean-up — old kernels removed", "removed", len(res.Removed), "names", strings.Join(res.Removed, ","),
+			"kept", strings.Join(sortedKeys(res.Kept), ","), "bytes_gained", res.BytesGained, "leftovers", len(res.Leftovers), "took", took)
+		return l.finish(ctx, lg, rep)
+	case "nothing":
+		lg.Info("osupdate: kernel clean-up — nothing to remove", "kept", strings.Join(sortedKeys(res.Kept), ","), "took", took)
+	case "not_run":
+		code, why := "", ""
+		if len(res.NotRun) > 0 {
+			code, why = res.NotRun[0].Code, res.NotRun[0].Reason
+		}
+		lg.Info("osupdate: kernel clean-up — not run on this box tonight", "code", code, "reason", why, "took", took)
+	default:
+		lg.Warn("osupdate: kernel clean-up — the wrapper answered with no outcome this agent knows (an older wrapper?)", "outcome", res.Outcome, "took", took)
+	}
+	return Report{}
+}
+
+// KernelCleanupDryRun asks the wrapper what the clean-up WOULD do, and changes nothing (the debug tool: `--selftest
+// kernel-cleanup-dry-run`). The raw kernel_cleanup object, or the refusal.
+func (l *Leg) KernelCleanupDryRun(ctx context.Context, vmid int) (json.RawMessage, error) {
+	wr, err := l.call(ctx, "kcdry"+l.now().UTC().Format("150405"), kernelPlan("kernel-cleanup", vmid,
+		map[string]any{"release_id": "kernel-cleanup", "dry_run": true}))
+	if err != nil {
+		return nil, err
+	}
+	if wr.refused() {
+		return nil, fmt.Errorf("kernel-cleanup dry run refused: %s", wr.Refused)
+	}
+	return wr.KernelCleanup, nil
+}
+
+// capNames bounds a list that goes to the hub: at most n names, each at most 120 bytes; never nil (the hub reads
+// "removed: []" as "none", not as "unknown").
+func capNames(in []string, n int) []string {
+	out := make([]string, 0, len(in))
+	for _, x := range in {
+		if len(out) == n {
+			break
+		}
+		if len(x) > 120 {
+			x = x[:120]
+		}
+		out = append(out, x)
+	}
+	return out
+}
+
+func sortedKeys(m map[string][]string) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}

internal/osupdate/kernel_test.go

@@ -28,13 +28,23 @@ func tonight(ring int) *hub.WireOSUpdate {
 func kernelCalls(w *fakeWrapper) []string {
 	var m []string
 	for _, p := range w.plans {
-		if p["layer"] == LayerKernel {
+		if p["layer"] == LayerKernel && p["mode"] != "kernel-cleanup" { // the kernel STEP's calls (the clean-up: cleanupCalls)
 			m = append(m, p["mode"].(string))
 		}
 	}
 	return m
 }
 
+func cleanupCalls(w *fakeWrapper) []map[string]any {
+	var m []map[string]any
+	for _, p := range w.plans {
+		if p["layer"] == LayerKernel && p["mode"] == "kernel-cleanup" {
+			m = append(m, p)
+		}
+	}
+	return m
+}
+
 // Ring 0, a told night: after the healthy host step the leg stages the pending kernel (select pending-kernel, the
 // kernel the household was told about), tells the hub "staged", THEN reboots — the kernel step ends the night.
 func TestKernel_Ring0ToldNightStagesThenReboots(t *testing.T) {
@@ -326,3 +336,196 @@ func TestKernelSet(t *testing.T) {
 		t.Fatal("a non-kernel string must give no set")
 	}
 }
+
+// ---- the clean-up of old kernels (R-1368, `09` §3 decision 286, `11` §5.11.1) ----
+
+func cleanupRep(js string) map[string][]WrapperReport {
+	return map[string][]WrapperReport{"kernel-cleanup": {{KernelCleanup: json.RawMessage(js)}}}
+}
+
+// The night: the clean-up is the LAST call, after the host step, exactly once, and its plan names NO package.
+// COMPANION RED-PROOF (observed): make kernelCleanupSkip return "" before its checks → the BYO / debug / unhealthy
+// sub-cases of TestKernelCleanup_WhenItMayNotRun fail; put the call before runFast → this test's order check fails.
+func TestKernelCleanup_NightOrderOnceAndNoNames(t *testing.T) {
+	w := &fakeWrapper{t: t, kernelRep: cleanupRep(`{"outcome":"removed","removed":["proxmox-kernel-7.0.14-20-pve-signed"],` +
+		`"kept":{"7.0.14-22-pve":["running"],"7.0.2-6-pve":["from"]},"bytes_gained":1073741824,"leftovers":[]}`)}
+	l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 0, Enabled: true})
+	p := l.Run(context.Background(), 9201, "night")
+	cc := cleanupCalls(w)
+	if len(cc) != 1 {
+		t.Fatalf("clean-up calls = %d, want exactly one a night", len(cc))
+	}
+	if last := w.plans[len(w.plans)-1]; last["mode"] != "kernel-cleanup" {
+		t.Fatalf("the clean-up must be the night's last call (after the host step), last = %v", last)
+	}
+	hostAt, cleanAt := -1, -1
+	for i, x := range w.plans {
+		if x["layer"] == LayerHost {
+			hostAt = i
+		}
+		if x["mode"] == "kernel-cleanup" {
+			cleanAt = i
+		}
+	}
+	if hostAt < 0 || cleanAt < hostAt {
+		t.Fatalf("the clean-up ran before the host step: host %d clean-up %d", hostAt, cleanAt)
+	}
+	plan := cc[0]
+	if pk, _ := json.Marshal(plan["packages"]); string(pk) != "[]" {
+		t.Fatalf("the clean-up plan names packages: %s", pk)
+	}
+	for _, k := range []string{"select", "signed", "allow_new", "expect_kver", "dry_run"} {
+		if _, ok := plan[k]; ok {
+			t.Fatalf("the clean-up plan carries %q: %v", k, plan)
+		}
+	}
+	if plan["lane"] != "slow" || plan["layer"] != LayerKernel {
+		t.Fatalf("plan = %v", plan)
+	}
+	// the hub hears "cleaned" with what was removed, kept and gained — and NO kernel view / version (the hub would read
+	// it as a kernel step for that kernel)
+	rep := p.KernelCleanup
+	if rep.Outcome != "cleaned" || !rep.Healthy || rep.Mode != "kernel-cleanup" || rep.Layer != LayerKernel {
+		t.Fatalf("report = %+v", rep)
+	}
+	for _, want := range []string{"removed 1 old kernel package(s): proxmox-kernel-7.0.14-20-pve-signed", "kept 7.0.14-22-pve, 7.0.2-6-pve", "1.0 GiB gained"} {
+		if !strings.Contains(rep.HealthReason, want) {
+			t.Fatalf("report reason %q lacks %q", rep.HealthReason, want)
+		}
+	}
+	lastRep := h.reports[len(h.reports)-1]
+	if lastRep.Mode != "kernel-cleanup" || lastRep.Outcome != "cleaned" {
+		t.Fatalf("the hub did not hear the clean-up: %+v", lastRep)
+	}
+	if len(lastRep.Kernel) != 0 || kverRE.MatchString(lastRep.ReleaseID) {
+		t.Fatalf("the clean-up's report names a kernel (the hub reads that as a kernel step): %+v", lastRep)
+	}
+	// R-1383 (release 3): the detail the hub's System page shows — removed, kept, gained, left-overs — rides the same
+	// report as its own object. On the wire it is `kernel_cleanup`, and the report STILL has no `kernel` key.
+	d := lastRep.KernelCleanup
+	if d == nil || len(d.Removed) != 1 || d.Removed[0] != "proxmox-kernel-7.0.14-20-pve-signed" ||
+		strings.Join(d.Kept, ",") != "7.0.14-22-pve,7.0.2-6-pve" || d.BytesGained == nil || *d.BytesGained != 1073741824 || d.Leftovers != 0 {
+		t.Fatalf("the clean-up's detail did not reach the hub: %+v", d)
+	}
+	wire, _ := json.Marshal(lastRep)
+	var keys map[string]json.RawMessage
+	_ = json.Unmarshal(wire, &keys)
+	if _, has := keys["kernel"]; has {
+		t.Fatalf("the clean-up's report carries a `kernel` view on the wire: %s", wire)
+	}
+	if string(keys["kernel_cleanup"]) != `{"removed":["proxmox-kernel-7.0.14-20-pve-signed"],"kept":["7.0.14-22-pve","7.0.2-6-pve"],"bytes_gained":1073741824,"leftovers":0}` {
+		t.Fatalf("kernel_cleanup on the wire = %s", keys["kernel_cleanup"])
+	}
+	if string(keys["release_id"]) != `"kernel-cleanup"` {
+		t.Fatalf("release_id on the wire = %s (a kernel version there is read as a step for it)", keys["release_id"])
+	}
+	// the same night again: the pass is skipped as a whole — never a second clean-up
+	l.Run(context.Background(), 9201, "night")
+	if len(cleanupCalls(w)) != 1 {
+		t.Fatalf("a second clean-up in the same night: %d calls", len(cleanupCalls(w)))
+	}
+}
+
+// "Nothing to remove" and "not run" are log lines, never a report (one a night would bury the kernel step's own line).
+func TestKernelCleanup_NothingAndNotRunSendNoReport(t *testing.T) {
+	for name, js := range map[string]string{
+		"nothing": `{"outcome":"nothing","removal":[],"kept":{"7.0.14-22-pve":["running"]}}`,
+		"not_run": `{"outcome":"not_run","not_run":[{"code":"no-ring0-mark","reason":"ring 0 only in this release"}]}`,
+	} {
+		w := &fakeWrapper{t: t, kernelRep: cleanupRep(js)}
+		l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 1, Enabled: true})
+		p := l.Run(context.Background(), 9201, "night")
+		if len(cleanupCalls(w)) != 1 {
+			t.Fatalf("%s: the wrapper was not asked", name)
+		}
+		if p.KernelCleanup.Layer != "" {
+			t.Fatalf("%s: a report was made: %+v", name, p.KernelCleanup)
+		}
+		for _, r := range h.reports {
+			if r.Mode == "kernel-cleanup" {
+				t.Fatalf("%s: the hub got a clean-up report: %+v", name, r)
+			}
+		}
+	}
+}
+
+// A refusal and a failure reach the hub (the operator), once, and are not retried.
+func TestKernelCleanup_RefusedAndFailedReachTheHub(t *testing.T) {
+	for name, wr := range map[string]WrapperReport{
+		"refused": {Refused: json.RawMessage(`{"code":"P3","reason":"the pinned default and grub.cfg disagree"}`)},
+		"failed":  {Failed: json.RawMessage(`{"rc":3,"step":"verify","reason":"the default in grub.cfg moved"}`), KernelCleanup: json.RawMessage(`{"outcome":"failed"}`)},
+	} {
+		w := &fakeWrapper{t: t, kernelRep: map[string][]WrapperReport{"kernel-cleanup": {wr}}}
+		l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 0, Enabled: true})
+		p := l.Run(context.Background(), 9201, "night")
+		if p.KernelCleanup.Outcome != name || p.KernelCleanup.Healthy || len(p.KernelCleanup.Refused) == 0 {
+			t.Fatalf("%s: report = %+v", name, p.KernelCleanup)
+		}
+		if last := h.reports[len(h.reports)-1]; last.Mode != "kernel-cleanup" || last.Outcome != name {
+			t.Fatalf("%s: the hub did not hear it: %+v", name, last)
+		}
+		if len(cleanupCalls(w)) != 1 {
+			t.Fatalf("%s: retried the same night (%d calls)", name, len(cleanupCalls(w)))
+		}
+	}
+}
+
+// When it may NOT run: never with a kernel step in the night (due, staged, or reported), not in a debug pass, not on a
+// BYO host, not with OS updates off, not after an unhealthy host step.
+func TestKernelCleanup_WhenItMayNotRun(t *testing.T) {
+	type tc struct {
+		blk     *hub.WireOSUpdate
+		trigger string
+		prep    func(*fakeWrapper, *Leg)
+	}
+	bad := &Health{DockerOK: false}
+	cases := map[string]tc{
+		"a kernel step is due tonight (ring 0)": {tonight(0), "night", func(w *fakeWrapper, _ *Leg) {
+			w.kernelRep = map[string][]WrapperReport{"kernel-status": {{Kernel: kview("none")}},
+				"apply":         {{Upgraded: []Package{{Name: "proxmox-kernel-7.0", Version: "7.0.14-22"}}, Authority: "ring0", Kernel: kview("staged")}},
+				"kernel-reboot": {{Kernel: kview("oneshot")}}}
+		}},
+		"a kernel step is due tonight (ring 1, nothing staged)": {tonight(1), "night", func(w *fakeWrapper, _ *Leg) {
+			w.kernelRep = map[string][]WrapperReport{"kernel-status": {{Kernel: kview("none")}}}
+		}},
+		"a debug pass": {&hub.WireOSUpdate{Ring: 0, Enabled: true}, "debug", nil},
+		"a BYO host":   {&hub.WireOSUpdate{Ring: 0, Enabled: true}, "night", func(_ *fakeWrapper, l *Leg) { l.Appliance = false }},
+		"updates off":  {&hub.WireOSUpdate{Ring: 0, Enabled: false}, "night", nil},
+		"host unhealthy": {&hub.WireOSUpdate{Ring: 0, Enabled: true, HostRelease: &hub.WireOSRelease{ID: "os-host-1", Packages: []hub.WireOSPackage{{Name: "openssl", Version: "u3", Origin: "Debian"}}}}, "night", func(w *fakeWrapper, _ *Leg) {
+			w.applyRep[LayerHost] = WrapperReport{Upgraded: []Package{{Name: "openssl", Version: "u3"}}, HealthAfter: &Health{HostServices: map[string]string{"pveproxy": "failed"}, GuestRunning: yes(), Guest: bad}}
+			w.healthSeq[LayerHost] = []*Health{{HostServices: map[string]string{"pveproxy": "failed"}, GuestRunning: yes(), Guest: bad}}
+		}},
+	}
+	for name, c := range cases {
+		w := &fakeWrapper{t: t}
+		l, _ := newLeg(t, w, c.blk)
+		if c.prep != nil {
+			c.prep(w, l)
+		}
+		l.HealthWait = time.Second
+		l.Run(context.Background(), 9201, c.trigger)
+		if n := len(cleanupCalls(w)); n != 0 {
+			t.Fatalf("%s: the clean-up ran (%d call(s)); calls = %s", name, n, calls(w))
+		}
+	}
+}
+
+// The dry run is the same mode with dry_run: true — and it sends nothing to the hub.
+func TestKernelCleanup_DryRun(t *testing.T) {
+	w := &fakeWrapper{t: t, kernelRep: cleanupRep(`{"outcome":"dry_run","would":"remove","removal":["proxmox-kernel-7.0.14-20-pve-signed"]}`)}
+	l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 0, Enabled: true})
+	raw, err := l.KernelCleanupDryRun(context.Background(), 9201)
+	if err != nil || !strings.Contains(string(raw), `"would":"remove"`) {
+		t.Fatalf("dry run = %s, %v", raw, err)
+	}
+	cc := cleanupCalls(w)
+	if len(cc) != 1 || cc[0]["dry_run"] != true {
+		t.Fatalf("the dry run's plan = %v", cc)
+	}
+	if pk, _ := json.Marshal(cc[0]["packages"]); string(pk) != "[]" {
+		t.Fatalf("the dry run's plan names packages: %s", pk)
+	}
+	if len(h.reports) != 0 {
+		t.Fatalf("a dry run reported to the hub: %+v", h.reports)
+	}
+}

internal/osupdate/leg.go

@@ -134,6 +134,13 @@ type WrapperReport struct {
 	// found after a boot; OutcomeHint "nothing" when no kernel was pending.
 	Kernel      json.RawMessage `json:"kernel"`
 	KernelEvent string          `json:"kernel_event"`
+	// KernelCleanup (R-1368, `11` §5.11.1): mode kernel-cleanup's result {outcome, dry_run, kept, removal, removed,
+	// bytes_gained, not_run, refusals, leftovers}.
+	KernelCleanup json.RawMessage `json:"kernel_cleanup"`
+	// ProfileRefresh (R-1388, `11` §5.12): mode profile-refresh's result {outcome, kernel, rewritten, restarted, seconds}.
+	ProfileRefresh json.RawMessage `json:"profile_refresh"`
+	// PVECACopy (R-1380, felhom.eu `03` §12a): mode pve-ca-copy's result {outcome, path, sha256, not_after, days_left}.
+	PVECACopy   json.RawMessage `json:"pve_ca_copy"`
 	OutcomeHint string          `json:"outcome_hint"`
 	// R-868 (v0.144.0): the agent's ids, echoed from the plan, so a report kept on disk can be sent without the
 	// agent process that started the pass. ReleaseID / VMID were always in the report.
@@ -181,10 +188,23 @@ type Report struct {
 	// outcomes of this layer: staged | applied (the new kernel is the default) | fell_back | health_failed (self-revert
 	// started) | self_reverted | revert_failed | judging | nothing | refused | failed.
 	Kernel json.RawMessage `json:"kernel,omitempty"`
+	// KernelCleanup (R-1383, release 3): ONLY on the mode kernel-cleanup report with outcome "cleaned" — what the
+	// clean-up removed, kept and gained, for the hub's System page. It is not the `kernel` view: this report still
+	// carries no kernel view and no kernel as its release_id, or the hub would read it as a kernel STEP for that kernel
+	// (the hub's kernelOf reads only those two). Pinned by TestKernelCleanup_NightOrderOnceAndNoNames.
+	KernelCleanup *CleanupDetail `json:"kernel_cleanup,omitempty"`
 
 	unsent string // R-868: the wrapper's kept copy of this pass's report — deleted once the hub has it
 }
 
+// CleanupDetail is the clean-up's result as the hub shows it (hub osupdates.CleanupDetail, field-exact).
+type CleanupDetail struct {
+	Removed     []string `json:"removed"`                // package names
+	Kept        []string `json:"kept"`                   // kernel versions that stay
+	BytesGained *int64   `json:"bytes_gained,omitempty"` // nil = the wrapper could not measure it
+	Leftovers   int      `json:"leftovers"`              // paths of a removed kernel that are still there
+}
+
 // Reporter posts a report to the hub (*hub.Client).
 type Reporter interface {
 	PostOSReport(ctx context.Context, body []byte) error
@@ -204,9 +224,15 @@ type Leg struct {
 	MinGap     time.Duration // between night runs (default 20 h)
 	Now        func() time.Time
 	Sleep      func(context.Context, time.Duration)
+	// ProfileRecordPath is the root wrapper's record of the last security-profile refresh (0644, read-only for the
+	// agent); "" = DefaultProfileRecordPath. A seam for the tests.
+	ProfileRecordPath string
 
-	mu    sync.Mutex
-	block *hub.WireOSUpdate
+	mu           sync.Mutex
+	block        *hub.WireOSUpdate
+	profileAsk   string // how the last ask of the profile refresh ended in THIS process ("" = never asked)
+	profileAskAt time.Time
+	profileLast  ProfileRefreshStatus // the last refresh as the wrapper answered it to this process
 }
 
 // planFile / reportFile: the plan the agent writes and the copy of the report the wrapper keeps beside it (R-868).
@@ -531,6 +557,8 @@ func (l *Leg) call(ctx context.Context, runID string, plan map[string]any) (Wrap
 // Pass is one leg's reports; an empty Layer means the step did not run.
 type Pass struct {
 	Guest, Host, Docker, PVE, Kernel Report
+	// KernelCleanup (R-1368): the clean-up of old kernels. Layer "" when it did not run or had nothing to tell the hub.
+	KernelCleanup Report
 }
 
 // Run is one pass: the guest layer, then (on an appliance, after a good guest step) the host layer, then (ring 0
@@ -582,6 +610,15 @@ func (l *Leg) Run(ctx context.Context, vmid int, trigger string) Pass {
 	default:
 		p.Kernel = l.runKernel(ctx, g.RunID, vmid, trigger, blk)
 	}
+	// R-1368 (`09` §3 decision 286, `11` §5.11.1): the clean-up of old kernels — its own step, at most once a night (this
+	// pass is the night's only one), after the host step ended healthy, NEVER in a night with a kernel step. Whether
+	// this box may do it at all is the root wrapper's answer (its root-owned ring-0 mark), never the agent's.
+	// Pinned by TestKernelCleanup_*.
+	if why := kernelCleanupSkip(l.Appliance, trigger, h, p.PVE, p.Kernel, blk, okStep); why != "" {
+		lg.Info("osupdate: kernel clean-up skipped — " + why)
+	} else {
+		p.KernelCleanup = l.runKernelCleanup(ctx, g.RunID, vmid, trigger, blk)
+	}
 	return p
 }
 

internal/osupdate/leg_test.go

@@ -26,6 +26,7 @@ type fakeWrapper struct {
 	keep        bool // R-868: like the real wrapper, keep an apply report beside the plan
 	pveGateHeld bool
 	kernelRep   map[string][]WrapperReport // R-836: per kernel-layer mode, successive answers (the last one repeats)
+	profileRep  *WrapperReport             // R-1388: the answer to mode profile-refresh
 }
 
 func yes() *bool { b := true; return &b }
@@ -71,6 +72,10 @@ func (f *fakeWrapper) Run(_ context.Context, name string, args ...string) ([]byt
 	}
 	var rep WrapperReport
 	switch plan["mode"] {
+	case "profile-refresh":
+		if f.profileRep != nil {
+			rep = *f.profileRep
+		}
 	case "inventory":
 		rep = WrapperReport{Mode: "inventory", Pending: f.pending, HealthBefore: ok, HealthAfter: ok,
 			Installed: []Package{{Name: "libc6", Version: "u3", Origin: "Debian"}}}
@@ -176,8 +181,9 @@ func TestRing0_OneCallPerLayer(t *testing.T) {
 	if g.Outcome != "applied" || !g.Healthy || ho.Outcome != "applied" || !ho.Healthy {
 		t.Fatalf("guest %+v\nhost %+v", g, ho)
 	}
-	if calls(w) != "guest:apply,host:apply,guest:live-restore-on,docker:apply,pve:apply" {
-		t.Fatalf("calls = %s, want one apply per layer, guest first, then live-restore, the ring-0 docker step and the pve step", calls(w))
+	// R-1368: the clean-up of old kernels is the night's last call when no kernel step is due
+	if calls(w) != "guest:apply,host:apply,guest:live-restore-on,docker:apply,pve:apply,kernel:kernel-cleanup" {
+		t.Fatalf("calls = %s, want one apply per layer, guest first, then live-restore, the ring-0 docker step, the pve step, the kernel clean-up", calls(w))
 	}
 	for _, p := range w.plans[:2] {
 		if p["select"] != "pending-fast" || p["snapshot"] != "" || len(p["packages"].([]any)) != 0 {
@@ -221,7 +227,8 @@ func TestRing1_NoReleaseIsInventory(t *testing.T) {
 	w := &fakeWrapper{t: t, pending: pend}
 	l, _ := newLeg(t, w, &hub.WireOSUpdate{Ring: 1, Enabled: true})
 	g, ho := run2(l, "night")
-	if g.Outcome != "nothing" || ho.Outcome != "nothing" || calls(w) != "guest:inventory,host:inventory" {
+	// R-1368: a ring-1 box asks the wrapper too — whether it may clean up is the wrapper's answer (its root-owned mark)
+	if g.Outcome != "nothing" || ho.Outcome != "nothing" || calls(w) != "guest:inventory,host:inventory,kernel:kernel-cleanup" {
 		t.Fatalf("g=%+v h=%+v calls=%s", g, ho, calls(w))
 	}
 }

internal/osupdate/profilerefresh.go

@@ -0,0 +1,217 @@
+package osupdate
+
+// The security-profile refresh (R-1388, R-1301, `11` §5.12). After a kernel went back and forth a box can load an OLD
+// compiled AppArmor profile for a program (on the N100 this stopped the time service). The root wrapper's mode
+// profile-refresh loads every profile from a fresh compile, once per kernel the box starts on, and restarts only a
+// failed service whose own program's profile was stale. This file decides WHEN to ask: at every agent start and after
+// a config bundle was installed (a new wrapper may know the mode the old one refused) — the wrapper itself answers
+// "not needed" when this kernel was already done. No OS report is sent for it (an unknown layer reads as a GUEST step on
+// the hub); since release 3 the result rides the host report as the `profile_refresh` stanza (ProfileRefreshStanza).
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"regexp"
+	"strings"
+	"time"
+)
+
+// ProfileRefreshResult is the wrapper's profile_refresh object.
+type ProfileRefreshResult struct {
+	Outcome   string   `json:"outcome"` // refreshed | not_needed | dry_run | failed
+	Reason    string   `json:"reason"`
+	Kernel    string   `json:"kernel"`
+	Rewritten []string `json:"rewritten"`
+	Restarted []struct {
+		Unit    string `json:"unit"`
+		Program string `json:"program"`
+		RC      int    `json:"rc"`
+		Active  bool   `json:"active"`
+	} `json:"restarted"`
+	Seconds float64 `json:"seconds"`
+	// The wrapper's own record of the refresh BEFORE this ask (it answers them at every ask, also "not needed").
+	LastRefreshKernel string `json:"last_refresh_kernel"`
+	LastRefreshAt     string `json:"last_refresh_at"`
+}
+
+// DefaultProfileRecordPath is where the root wrapper keeps the record of its last refresh (PROFILE_STATE, 0644).
+const DefaultProfileRecordPath = "/var/lib/felhom-os-apply/profile-refresh.json"
+
+// ProfileRefreshStatus is the host report's `profile_refresh` stanza (R-1383, release 3; hub sysfacts.ProfileRefresh,
+// field-exact): the last refresh the box did — from the root wrapper's own record, so it outlives an agent restart —
+// and how the last ask of THIS agent process ended. The refresh restarted a failed service on the N100 and nobody
+// outside the box's journal could see it.
+type ProfileRefreshStatus struct {
+	Kernel    string   `json:"kernel,omitempty"`    // the kernel the last refresh ran on
+	At        string   `json:"at,omitempty"`        // RFC3339, when
+	Rewritten *int     `json:"rewritten,omitempty"` // how many compiled profiles changed; nil = not known to this process
+	Restarted []string `json:"restarted,omitempty"` // units it restarted (failed, with a stale profile)
+	LastAsk   string   `json:"last_ask,omitempty"`  // refreshed | not_needed | failed | refused | no_answer
+	LastAskAt string   `json:"last_ask_at,omitempty"`
+}
+
+var profileNameRe = regexp.MustCompile(`^[A-Za-z0-9@._:+-]{1,120}$`)
+
+// ProfileRefreshStanza builds the stanza, or an UNTYPED nil when there is nothing to say (never asked and no record: a
+// BYO host, or an agent that did not ask yet) — the report then omits the field.
+//
+// Two sources, the first that knows wins per field. (1) What the wrapper ANSWERED to this process's last ask: a
+// refresh gives everything; "not needed" gives the kernel and time of the earlier refresh (the wrapper answers its own
+// record at every ask) — so the last refresh is still told after an agent restart, without the detail. (2) The
+// wrapper's record file, when this process may read it: on boxes where its folder was first made by a signed job it
+// is root-only (0700), so the file is a bonus, never the only source (found by review, 2026-11-14; not read on a box).
+// It never calls the wrapper. Names that are not plain unit / kernel names are dropped, the list is capped.
+func (l *Leg) ProfileRefreshStanza() any {
+	l.mu.Lock()
+	st, ask, at := l.profileLast, l.profileAsk, l.profileAskAt
+	l.mu.Unlock()
+	have := ask != ""
+	if have {
+		st.LastAsk, st.LastAskAt = ask, at.UTC().Format(time.RFC3339)
+	}
+	path := l.ProfileRecordPath
+	if path == "" {
+		path = DefaultProfileRecordPath
+	}
+	if raw, err := os.ReadFile(path); err == nil && len(raw) < 64<<10 {
+		var rec struct {
+			Kernel    string   `json:"kernel"`
+			At        string   `json:"at"`
+			Rewritten []string `json:"rewritten"`
+			Restarted []string `json:"restarted"`
+		}
+		if json.Unmarshal(raw, &rec) == nil && (rec.Kernel != "" || rec.At != "") {
+			have = true
+			fromRec := profileStatusFrom(rec.Kernel, rec.At, rec.Rewritten, rec.Restarted, true)
+			if st.At == "" || fromRec.At == st.At || st.Rewritten == nil {
+				fromRec.LastAsk, fromRec.LastAskAt = st.LastAsk, st.LastAskAt
+				st = fromRec
+			}
+		}
+	}
+	if !have {
+		return nil
+	}
+	return st
+}
+
+// profileStatusFrom checks and bounds what goes to the hub. detail=false: the count and the units are not known.
+func profileStatusFrom(kernel, at string, rewritten, restarted []string, detail bool) ProfileRefreshStatus {
+	var st ProfileRefreshStatus
+	if profileNameRe.MatchString(kernel) {
+		st.Kernel = kernel
+	}
+	if _, err := time.Parse(time.RFC3339, at); err == nil {
+		st.At = at
+	}
+	if detail {
+		n := len(rewritten)
+		st.Rewritten = &n
+		for _, u := range restarted {
+			if len(st.Restarted) < 16 && profileNameRe.MatchString(u) {
+				st.Restarted = append(st.Restarted, u)
+			}
+		}
+	}
+	return st
+}
+
+// noteProfileAsk remembers how an ask ended and, when the wrapper answered, what it said about the last refresh.
+func (l *Leg) noteProfileAsk(outcome string, res *ProfileRefreshResult) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.profileAsk, l.profileAskAt = outcome, l.now()
+	if res == nil {
+		return
+	}
+	switch outcome {
+	case "refreshed":
+		units := make([]string, 0, len(res.Restarted))
+		for _, u := range res.Restarted {
+			units = append(units, u.Unit)
+		}
+		l.profileLast = profileStatusFrom(res.Kernel, l.now().UTC().Format(time.RFC3339), res.Rewritten, units, true)
+	case "not_needed":
+		if res.LastRefreshAt != "" || res.LastRefreshKernel != "" {
+			l.profileLast = profileStatusFrom(res.LastRefreshKernel, res.LastRefreshAt, nil, nil, false)
+		}
+	}
+}
+
+// ProfileGate takes the heavy-op gate (a backup, a restore-test or an OS step must not run beside the refresh).
+type ProfileGate func() (release func(), busy string, ok bool)
+
+// ProfileRefresh asks the wrapper once. It returns the outcome ("" when the wrapper could not be asked or refused).
+// It waits for the gate: up to `tries` attempts, `wait` apart. why is the log's short reason for this call.
+func (l *Leg) ProfileRefresh(ctx context.Context, why string, gate ProfileGate, tries int, wait time.Duration) string {
+	lg := l.log().With("mode", "profile-refresh", "why", why)
+	if !l.Appliance {
+		lg.Info("osupdate: profile refresh skipped — not an appliance install (a BYO host belongs to its owner)")
+		return ""
+	}
+	var release func()
+	for i := 1; ; i++ {
+		if gate == nil {
+			break
+		}
+		r, busy, ok := gate()
+		if ok {
+			release = r
+			break
+		}
+		if i >= tries || ctx.Err() != nil {
+			lg.Warn("osupdate: profile refresh NOT run — the box stayed busy; the next agent start asks again", "busy", busy, "tries", i)
+			return ""
+		}
+		lg.Info("osupdate: profile refresh waits — a heavy operation runs", "busy", busy, "try", i)
+		l.sleep(ctx, wait)
+	}
+	if release != nil {
+		defer release()
+	}
+	unlock := l.lockPass(true) // never beside an OS-update pass
+	defer unlock()
+	start := l.now()
+	wr, err := l.call(ctx, "prefresh"+l.now().UTC().Format("150405"), map[string]any{"release_id": "profile-refresh",
+		"layer": LayerHost, "mode": "profile-refresh"})
+	took := l.now().Sub(start).Round(100 * time.Millisecond).String()
+	switch {
+	case err != nil:
+		lg.Warn("osupdate: profile refresh — the wrapper gave no answer; the next agent start asks again", "err", err, "took", took)
+		l.noteProfileAsk("no_answer", nil)
+		return ""
+	case wr.refused():
+		lg.Info("osupdate: profile refresh refused (an older wrapper that does not know the mode, or a BYO host)", "refused", string(wr.Refused))
+		l.noteProfileAsk("refused", nil)
+		return ""
+	case wr.failed():
+		lg.Error("osupdate: profile refresh FAILED — the profiles are as before; the next agent start asks again", "failed", string(wr.Failed), "took", took)
+		l.noteProfileAsk("failed", nil)
+		return "failed"
+	}
+	var res ProfileRefreshResult
+	_ = json.Unmarshal(wr.ProfileRefresh, &res)
+	switch res.Outcome {
+	case "refreshed":
+		units := make([]string, 0, len(res.Restarted))
+		for _, u := range res.Restarted {
+			units = append(units, u.Unit)
+		}
+		if len(units) > 0 {
+			lg.Warn("osupdate: profile refresh — security profiles reloaded from a fresh compile; a FAILED service whose profile was stale was restarted",
+				"kernel", res.Kernel, "restarted", strings.Join(units, ","), "rewritten", len(res.Rewritten), "wrapper_seconds", res.Seconds, "took", took)
+		} else {
+			lg.Info("osupdate: profile refresh — security profiles reloaded from a fresh compile", "kernel", res.Kernel,
+				"rewritten", len(res.Rewritten), "wrapper_seconds", res.Seconds, "took", took)
+		}
+	case "not_needed":
+		lg.Info("osupdate: profile refresh not needed", "kernel", res.Kernel, "reason", res.Reason)
+	default:
+		lg.Warn("osupdate: profile refresh — the wrapper answered with no outcome this agent knows", "outcome", res.Outcome)
+	}
+	if res.Outcome == "refreshed" || res.Outcome == "not_needed" {
+		l.noteProfileAsk(res.Outcome, &res)
+	}
+	return res.Outcome
+}

internal/osupdate/profilerefresh_test.go

@@ -0,0 +1,169 @@
+package osupdate
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+)
+
+// ---- the security-profile refresh (R-1388, R-1301, `11` §5.12) ----
+
+func profileCalls(w *fakeWrapper) []map[string]any {
+	var m []map[string]any
+	for _, p := range w.plans {
+		if p["mode"] == "profile-refresh" {
+			m = append(m, p)
+		}
+	}
+	return m
+}
+
+func refreshed() *WrapperReport {
+	return &WrapperReport{ProfileRefresh: json.RawMessage(`{"outcome":"refreshed","kernel":"7.0.14-28-pve","rewritten":["f6da0254.0/usr.sbin.chronyd"],` +
+		`"restarted":[{"unit":"chrony.service","program":"/usr/sbin/chronyd","rc":0,"active":true}],"seconds":0.5}`)}
+}
+
+// One call, a host-layer plan that names NOTHING, and nothing to the hub.
+// COMPANION RED-PROOF (observed): add "units" to the plan → this test fails; drop the Appliance check → the BYO test fails.
+func TestProfileRefresh_OneCallNoNamesNoHubReport(t *testing.T) {
+	w := &fakeWrapper{t: t, profileRep: refreshed()}
+	l, h := newLeg(t, w, &hub.WireOSUpdate{Ring: 1, Enabled: true})
+	if got := l.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "refreshed" {
+		t.Fatalf("outcome = %q", got)
+	}
+	pc := profileCalls(w)
+	if len(pc) != 1 {
+		t.Fatalf("calls = %d", len(pc))
+	}
+	if len(pc[0]) != 3 || pc[0]["layer"] != LayerHost || pc[0]["release_id"] != "profile-refresh" {
+		t.Fatalf("the plan carries more than release_id, layer and mode: %v", pc[0])
+	}
+	if len(h.reports) != 0 {
+		t.Fatalf("the hub got a report: %+v", h.reports)
+	}
+}
+
+func TestProfileRefresh_ByoHostIsNeverAsked(t *testing.T) {
+	w := &fakeWrapper{t: t, profileRep: refreshed()}
+	l, _ := newLeg(t, w, nil)
+	l.Appliance = false
+	if got := l.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "" || len(profileCalls(w)) != 0 {
+		t.Fatalf("a BYO host was asked: %q %v", got, profileCalls(w))
+	}
+}
+
+// The heavy-op gate: busy → wait and ask again; still busy after the tries → not run (the next start asks again).
+// COMPANION RED-PROOF (observed): ignore `ok` from the gate → the "always busy" half fails.
+func TestProfileRefresh_WaitsForTheHeavyOpGate(t *testing.T) {
+	w := &fakeWrapper{t: t, profileRep: refreshed()}
+	l, _ := newLeg(t, w, nil)
+	n, released := 0, 0
+	gate := func() (func(), string, bool) {
+		n++
+		if n < 3 {
+			return nil, "primary-backup", false
+		}
+		return func() { released++ }, "", true
+	}
+	if got := l.ProfileRefresh(context.Background(), "agent start", gate, 5, time.Minute); got != "refreshed" {
+		t.Fatalf("outcome = %q after %d gate tries", got, n)
+	}
+	if n != 3 || released != 1 || len(profileCalls(w)) != 1 {
+		t.Fatalf("gate tries %d, released %d, calls %d", n, released, len(profileCalls(w)))
+	}
+	w2 := &fakeWrapper{t: t, profileRep: refreshed()}
+	l2, _ := newLeg(t, w2, nil)
+	busy := func() (func(), string, bool) { return nil, "restore-test", false }
+	if got := l2.ProfileRefresh(context.Background(), "agent start", busy, 3, time.Minute); got != "" || len(profileCalls(w2)) != 0 {
+		t.Fatalf("ran beside a heavy operation: %q %d call(s)", got, len(profileCalls(w2)))
+	}
+}
+
+// An older wrapper refuses the mode: logged, nothing else. A failure is said as "failed".
+func TestProfileRefresh_RefusedAndFailed(t *testing.T) {
+	w := &fakeWrapper{t: t, profileRep: &WrapperReport{Refused: json.RawMessage(`{"code":"R11","reason":"unknown mode 'profile-refresh'"}`)}}
+	l, _ := newLeg(t, w, nil)
+	if got := l.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "" {
+		t.Fatalf("refused → %q", got)
+	}
+	w2 := &fakeWrapper{t: t, profileRep: &WrapperReport{Failed: json.RawMessage(`{"rc":1,"step":"profile-refresh"}`)}}
+	l2, _ := newLeg(t, w2, nil)
+	if got := l2.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "failed" {
+		t.Fatalf("failed → %q", got)
+	}
+}
+
+// R-1383 (release 3): the refresh restarted a failed service on the N100 and only the box's journal knew. The stanza
+// that rides the host report: nothing before any ask (an UNTYPED nil — the report omits the field); after a refresh,
+// everything; after an agent RESTART, the kernel and time of that refresh again — from the wrapper's answer to the new
+// process's ask, because the wrapper's record FILE may be root-only on a box (its folder is 0700 where a signed job
+// made it first). Where the file can be read, it adds the detail back.
+// COMPANION RED-PROOFS (observed, felhom.eu audits/night-build-2026-11-14/B/redproof-agent-b3.txt).
+func TestProfileRefreshStanza_RecordAndLastAsk(t *testing.T) {
+	w := &fakeWrapper{t: t, profileRep: refreshed()}
+	l, _ := newLeg(t, w, &hub.WireOSUpdate{Ring: 1, Enabled: true})
+	unreadable := filepath.Join(t.TempDir(), "root-only", "profile-refresh.json") // does not exist: as a 0700 folder reads
+	l.ProfileRecordPath = unreadable
+	now := time.Date(2026, 11, 14, 11, 30, 11, 0, time.UTC)
+	l.Now = func() time.Time { return now }
+	if st := l.ProfileRefreshStanza(); st != nil {
+		t.Fatalf("before any ask the stanza is %#v, want an untyped nil", st)
+	}
+	if got := l.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "refreshed" {
+		t.Fatalf("outcome = %q", got)
+	}
+	b, _ := json.Marshal(l.ProfileRefreshStanza())
+	if string(b) != `{"kernel":"7.0.14-28-pve","at":"2026-11-14T11:30:11Z","rewritten":1,"restarted":["chrony.service"],"last_ask":"refreshed","last_ask_at":"2026-11-14T11:30:11Z"}` {
+		t.Fatalf("stanza after a refresh, record file unreadable = %s", b)
+	}
+	// an agent restart: a new process asks, the wrapper answers "not needed" with ITS record of the last refresh
+	l2, _ := newLeg(t, &fakeWrapper{t: t, profileRep: &WrapperReport{ProfileRefresh: json.RawMessage(
+		`{"outcome":"not_needed","reason":"this kernel was done","kernel":"7.0.14-28-pve","last_refresh_kernel":"7.0.14-28-pve","last_refresh_at":"2026-11-14T11:30:11Z"}`)}},
+		&hub.WireOSUpdate{Ring: 1, Enabled: true})
+	l2.ProfileRecordPath = unreadable
+	l2.Now = func() time.Time { return now.Add(3 * time.Hour) }
+	if st := l2.ProfileRefreshStanza(); st != nil {
+		t.Fatalf("a new process that has not asked and cannot read the record: %#v, want nil", st)
+	}
+	if got := l2.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "not_needed" {
+		t.Fatalf("outcome = %q", got)
+	}
+	b, _ = json.Marshal(l2.ProfileRefreshStanza())
+	if string(b) != `{"kernel":"7.0.14-28-pve","at":"2026-11-14T11:30:11Z","last_ask":"not_needed","last_ask_at":"2026-11-14T14:30:11Z"}` {
+		t.Fatalf("stanza after an agent restart (the wrapper's answer, no detail) = %s", b)
+	}
+	// where the record file CAN be read, the detail is back — and a name that is not a unit name is dropped
+	rec := filepath.Join(t.TempDir(), "profile-refresh.json")
+	if err := os.WriteFile(rec, []byte(`{"at":"2026-11-14T11:30:11Z","kernel":"7.0.14-28-pve","restarted":["chrony.service","bad name; rm -rf"],`+
+		`"rewritten":["a","b"]}`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	l2.ProfileRecordPath = rec
+	b, _ = json.Marshal(l2.ProfileRefreshStanza())
+	if string(b) != `{"kernel":"7.0.14-28-pve","at":"2026-11-14T11:30:11Z","rewritten":2,"restarted":["chrony.service"],"last_ask":"not_needed","last_ask_at":"2026-11-14T14:30:11Z"}` {
+		t.Fatalf("stanza with a readable record = %s", b)
+	}
+	// a failed ask is told, beside the older record
+	l2.Runner = (&fakeWrapper{t: t, profileRep: &WrapperReport{Failed: json.RawMessage(`{"code":"F1","reason":"apparmor_parser failed"}`)}})
+	l2.Now = func() time.Time { return now.Add(4 * time.Hour) }
+	if got := l2.ProfileRefresh(context.Background(), "agent start", nil, 1, time.Second); got != "failed" {
+		t.Fatalf("outcome = %q", got)
+	}
+	b, _ = json.Marshal(l2.ProfileRefreshStanza())
+	if !strings.Contains(string(b), `"last_ask":"failed","last_ask_at":"2026-11-14T15:30:11Z"`) || !strings.Contains(string(b), `"kernel":"7.0.14-28-pve"`) {
+		t.Fatalf("stanza after a failed ask = %s", b)
+	}
+	// a record that is not JSON says nothing and breaks nothing
+	_ = os.WriteFile(rec, []byte("not json"), 0o644)
+	l3, _ := newLeg(t, &fakeWrapper{t: t}, &hub.WireOSUpdate{Ring: 1, Enabled: true})
+	l3.ProfileRecordPath = rec
+	if st := l3.ProfileRefreshStanza(); st != nil {
+		t.Fatalf("an unreadable record gave the stanza %#v", st)
+	}
+}

internal/osupdate/pvecacopy.go

@@ -0,0 +1,45 @@
+package osupdate
+
+// The root step that keeps a readable copy of Proxmox's root CA for the agent (R-1380, D80, felhom.eu `03` §12a).
+// The rule lives in the wrapper (mode pve-ca-copy: one fixed source, one fixed place, a CA certificate only); the plan
+// names nothing. internal/pvetrust decides when to ask and what to do with the answer.
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+)
+
+// PVECACopyPath is where the wrapper puts the copy (felhom-os-apply PVE_CA_COPY) — the agent's `proxmox.tls.ca_file`.
+const PVECACopyPath = "/etc/felhom-agent/pve-root-ca.pem"
+
+// PVECACopyResult is the wrapper's pve_ca_copy object.
+type PVECACopyResult struct {
+	Outcome  string `json:"outcome"` // written | same
+	Path     string `json:"path"`
+	SHA256   string `json:"sha256"`
+	NotAfter string `json:"not_after"`
+	DaysLeft int    `json:"days_left"`
+	Replaced bool   `json:"replaced"`
+}
+
+// PVECACopy asks the wrapper to make or refresh the copy. An error names a refusal (T1–T3, or R11 from a wrapper that
+// does not know the mode yet) or a wrapper that gave no answer.
+func (l *Leg) PVECACopy(ctx context.Context) (PVECACopyResult, error) {
+	wr, err := l.call(ctx, "pveca"+l.now().UTC().Format("150405.000"), map[string]any{"release_id": "pve-ca-copy",
+		"layer": LayerHost, "mode": "pve-ca-copy"})
+	if err != nil {
+		return PVECACopyResult{}, err
+	}
+	if wr.refused() {
+		return PVECACopyResult{}, fmt.Errorf("pve-ca-copy refused: %s", wr.Refused)
+	}
+	if wr.failed() {
+		return PVECACopyResult{}, fmt.Errorf("pve-ca-copy failed: %s", wr.Failed)
+	}
+	var res PVECACopyResult
+	if jerr := json.Unmarshal(wr.PVECACopy, &res); jerr != nil || res.Outcome == "" || res.Path == "" {
+		return PVECACopyResult{}, fmt.Errorf("pve-ca-copy: the wrapper gave no result (%v)", jerr)
+	}
+	return res, nil
+}

internal/pbsdr/cutwrite_main_test.go

@@ -0,0 +1,21 @@
+package pbsdr
+
+import (
+	"os"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
+)
+
+// TestMain points the config write journal (config.CutWriteDir, R-1380) at a temp folder: the real one is the
+// agent's state folder on a box, and a test must not need it.
+func TestMain(m *testing.M) {
+	dir, err := os.MkdirTemp("", "felhom-cutwrite-")
+	if err != nil {
+		panic(err)
+	}
+	config.CutWriteDir = dir
+	code := m.Run()
+	_ = os.RemoveAll(dir)
+	os.Exit(code)
+}

internal/pbsdr/manager.go

@@ -27,6 +27,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
 	"log/slog"
 	"net/netip"
 	"os"
@@ -483,8 +484,10 @@ func (m *Manager) consumedFail(h, msg string, block *hub.WirePBSDR) {
 // seedEscrowStorageID sets escrow.pbs_storage_id in agent.json when empty/absent — the bare
 // `--selftest=escrow-create` one-liner must find its storage with no flags. A different existing
 // value is NEVER clobbered (warn-and-keep). Unknown config keys are preserved verbatim
-// (map[string]RawMessage read-modify-write, atomic rename).
+// (map[string]RawMessage read-modify-write, in place behind a journal — config.ReplaceInPlace).
 func (m *Manager) seedEscrowStorageID(storageID string) error {
+	config.FileMu.Lock() // one writer of agent.json at a time in this process (config.FileMu)
+	defer config.FileMu.Unlock()
 	if m.configPath == "" {
 		return nil
 	}
@@ -527,30 +530,18 @@ func (m *Manager) seedEscrowStorageID(storageID string) error {
 	if err != nil {
 		return err
 	}
-	st, err := os.Stat(m.configPath)
-	if err != nil {
-		return err
-	}
 	// The config DIRECTORY (/etc/felhom-agent) is root-owned while the FILE is agent-owned —
 	// tmp+rename is impossible for the non-root agent (rename mutates the directory; live
-	// finding on the demo). So: park a recovery copy in the agent state dir, then rewrite the
-	// file IN PLACE (O_TRUNC). The agent is the file's only writer and the content is small;
-	// a torn write is recoverable from the parked copy.
+	// finding on the demo). So: park a copy for a person in the agent state dir, then rewrite the
+	// file IN PLACE through config.ReplaceInPlace, which journals the old bytes first — the next
+	// start repairs a cut write (R-1380, TestRecoverCutWrite_*).
 	if err := os.MkdirAll(m.stateDir, 0o700); err != nil {
 		return err
 	}
 	if err := os.WriteFile(filepath.Join(m.stateDir, "agent.json.pre-seed"), raw, 0o600); err != nil {
 		return fmt.Errorf("parking the pre-seed copy: %w", err)
 	}
-	f, err := os.OpenFile(m.configPath, os.O_WRONLY|os.O_TRUNC, st.Mode().Perm())
-	if err != nil {
-		return err
-	}
-	if _, err := f.Write(out); err != nil {
-		f.Close()
-		return err
-	}
-	if err := f.Close(); err != nil {
+	if err := config.ReplaceInPlace(m.configPath, raw, out); err != nil {
 		return err
 	}
 	m.logger.Info("pbsdr: seeded escrow.pbs_storage_id (the ceremony one-liner needs no flags)", "storage_id", storageID)

internal/proxmox/client.go

@@ -10,6 +10,7 @@ import (
 	"net/http"
 	"net/url"
 	"strings"
+	"sync"
 	"time"
 
 	"gitea.dooplex.hu/admin/felhom-agent/internal/httpx"
@@ -45,6 +46,39 @@ type Client struct {
 	node  string
 	token string
 	http  doer
+
+	// tlsMu guards http: SetTLS swaps the transport while calls run (R-1380 — the move from a pinned certificate to the
+	// box's own root CA, and a refreshed CA copy, take effect with no agent restart).
+	tlsMu   sync.RWMutex
+	timeout time.Duration
+}
+
+func (c *Client) httpDoer() doer {
+	c.tlsMu.RLock()
+	defer c.tlsMu.RUnlock()
+	return c.http
+}
+
+// SetTLS replaces how this client trusts the Proxmox certificate, for every call that starts after it returns. On an
+// error (an unreadable CA file, a bad fingerprint) nothing changes.
+func (c *Client) SetTLS(t TLSConfig) error {
+	tlsCfg, err := t.build()
+	if err != nil {
+		return err
+	}
+	timeout := c.timeout
+	if timeout == 0 {
+		timeout = 30 * time.Second
+	}
+	hc := &http.Client{Timeout: timeout, Transport: httpx.NewTransport(tlsCfg, 0)}
+	c.tlsMu.Lock()
+	old := c.http
+	c.http = hc
+	c.tlsMu.Unlock()
+	if o, ok := old.(*http.Client); ok && o != nil {
+		o.CloseIdleConnections() // a kept-alive connection made under the old trust must not serve the next call
+	}
+	return nil
 }
 
 // NewClient builds an API client. It validates required config and constructs the
@@ -74,10 +108,11 @@ func NewClient(cfg Config) (*Client, error) {
 		Transport: httpx.NewTransport(tlsCfg, 0),
 	}
 	return &Client{
-		base:  strings.TrimRight(cfg.Endpoint, "/") + "/api2/json",
-		node:  cfg.Node,
-		token: cfg.Token,
-		http:  hc,
+		base:    strings.TrimRight(cfg.Endpoint, "/") + "/api2/json",
+		node:    cfg.Node,
+		token:   cfg.Token,
+		http:    hc,
+		timeout: timeout,
 	}, nil
 }
 
@@ -125,7 +160,7 @@ func (c *Client) doBody(ctx context.Context, method, path string, body io.Reader
 		req.Header.Set("Content-Type", contentType)
 	}
 
-	resp, err := c.http.Do(req)
+	resp, err := c.httpDoer().Do(req)
 	if err != nil {
 		return fmt.Errorf("proxmox: %s %s: %w", method, path, err)
 	}

internal/proxmox/trust.go

@@ -0,0 +1,43 @@
+package proxmox
+
+import (
+	"crypto/x509"
+	"errors"
+	"strings"
+)
+
+// The causes a call to Proxmox can fail on TRUST (R-1380, felhom.eu `03` §12a). A symptom alarm („the agent channel is
+// down") exists; these name the cause.
+const (
+	TrustPinnedCertChanged = "pinned_cert_changed" // the pinned certificate is not the one Proxmox serves any more
+	TrustCAMismatch        = "ca_mismatch"         // the CA file did not sign the served certificate
+	TrustCertExpired       = "cert_expired"        // the served certificate (or the CA) has ended
+	TrustCAFileUnreadable  = "ca_file_unreadable"  // the CA file is missing or holds no certificate
+	TrustNameMismatch      = "name_mismatch"       // the certificate does not name the address the agent calls
+)
+
+// TrustProblem names why err is a TRUST failure, or "" when it is none (a refused connection, a 401, a timeout).
+func TrustProblem(err error) string {
+	if err == nil {
+		return ""
+	}
+	var ua x509.UnknownAuthorityError
+	var inv x509.CertificateInvalidError
+	var hn x509.HostnameError
+	switch {
+	case strings.Contains(err.Error(), "TLS pin mismatch"):
+		return TrustPinnedCertChanged
+	case errors.As(err, &ua):
+		return TrustCAMismatch
+	case errors.As(err, &inv):
+		if inv.Reason == x509.Expired {
+			return TrustCertExpired
+		}
+		return TrustCAMismatch
+	case errors.As(err, &hn):
+		return TrustNameMismatch
+	case strings.Contains(err.Error(), "reading TLS CA file"), strings.Contains(err.Error(), "contained no usable certificates"):
+		return TrustCAFileUnreadable
+	}
+	return ""
+}

internal/pvetrust/cutwrite_main_test.go

@@ -0,0 +1,21 @@
+package pvetrust
+
+import (
+	"os"
+	"testing"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
+)
+
+// TestMain points the config write journal (config.CutWriteDir, R-1380) at a temp folder: the real one is the
+// agent's state folder on a box, and a test must not need it.
+func TestMain(m *testing.M) {
+	dir, err := os.MkdirTemp("", "felhom-cutwrite-")
+	if err != nil {
+		panic(err)
+	}
+	config.CutWriteDir = dir
+	code := m.Run()
+	_ = os.RemoveAll(dir)
+	os.Exit(code)
+}

internal/pvetrust/pvetrust.go

@@ -0,0 +1,406 @@
+// Package pvetrust moves the agent's trust in Proxmox from ONE pinned certificate to the box's own root CA, and names
+// the cause when a call fails on trust (R-1380, D80, felhom.eu `09` §3 decision 303; design `03-host-agent.md` §12a).
+//
+// WHY. Proxmox makes a new node certificate by itself, at most every two years. An agent that pins the old one is then
+// blind: every call fails, no host report reaches the hub, and the only alarm names a symptom. A copy of the box's own
+// root CA verifies every certificate that CA signs.
+//
+// THE ORDER, and it is the whole point (a box must never be left with neither way working):
+//
+//  1. the root step makes or refreshes the copy (it refuses anything that is not one CA certificate);
+//  2. a REAL call to Proxmox is made with the copy ALONE — while the fingerprint still is the configured trust;
+//  3. only after that call worked, ONE write of the agent's config sets `ca_file` and clears the fingerprint;
+//  4. the running client takes the new trust (no restart).
+//
+// If 1 or 2 fails nothing is written and the fingerprint stays. If 3 fails the fingerprint stays, in the file and in
+// the running client. `proxmox.tls` keeps its order in code — a set fingerprint wins over a CA file — so the
+// fingerprint is cleared LAST, in the same write that sets `ca_file`.
+//
+// WHO WRITES THE CONFIG. The agent itself: `agent.json` is the agent's own file (0600) in a root-owned folder, so it is
+// rewritten in place after a recovery copy was parked in the agent's state folder — the same way the backup-target move
+// and the PBS seed write it. No root step touches it.
+package pvetrust
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/config"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"sync"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// EndsSoonDays: the copy's certificate ends within this many days → the problem „ca_copy_ends_soon". Proxmox renews its
+// root CA when less than two years are left, so a healthy box never comes near it.
+const EndsSoonDays = 30
+
+// Problems this package names beyond proxmox.Trust*.
+const (
+	ProblemCopyUnavailable = "ca_copy_unavailable"     // the root step refused or did not answer
+	ProblemCopyNoProof     = "ca_copy_does_not_verify" // a call with the copy alone failed — the fingerprint stays
+	ProblemCopyEndsSoon    = "ca_copy_ends_soon"
+	ProblemConfigNotSaved  = "config_not_written" // the proof worked but the config could not be written — the fingerprint stays
+)
+
+// Copy is what the root step answered.
+type Copy struct {
+	Path     string
+	Outcome  string // written | same
+	NotAfter string
+	DaysLeft int
+}
+
+// Status is the stanza that rides the host report (`proxmox_trust`). No secret: a mode, a cause, a date.
+type Status struct {
+	Mode       string `json:"mode"`              // ca | fingerprint | ca_file_own | insecure | system
+	Problem    string `json:"problem,omitempty"` // "" = none
+	Detail     string `json:"detail,omitempty"`
+	CANotAfter string `json:"ca_not_after,omitempty"`
+	CADaysLeft *int   `json:"ca_days_left,omitempty"`
+	MovedAt    string `json:"moved_at,omitempty"` // when this box moved from the pinned certificate to the CA copy
+	CheckedAt  string `json:"checked_at,omitempty"`
+}
+
+// Manager holds the trust state of one agent process.
+type Manager struct {
+	// EnsureCopy runs the root step (osupdate.Leg.PVECACopy).
+	EnsureCopy func(ctx context.Context) (Copy, error)
+	// Probe makes one real call to Proxmox with EXACTLY this trust (a throwaway client).
+	Probe func(ctx context.Context, t proxmox.TLSConfig) error
+	// Apply gives the running client the new trust (proxmox.Client.SetTLS).
+	Apply func(t proxmox.TLSConfig) error
+	// ConfigPath is agent.json ("" = the config came from the environment: nothing can be saved, so nothing is moved).
+	ConfigPath string
+	StateDir   string
+	Logger     *slog.Logger
+	Now        func() time.Time
+	// MinGap between two runs that a failed call asked for (default 10 minutes).
+	MinGap time.Duration
+
+	runMu   sync.Mutex // serialises Ensure
+	mu      sync.Mutex
+	cur     proxmox.TLSConfig
+	status  Status
+	lastRun time.Time
+}
+
+// copyStepTimeout bounds the root step that makes the CA copy.
+const copyStepTimeout = 90 * time.Second
+
+func (m *Manager) now() time.Time {
+	if m.Now != nil {
+		return m.Now()
+	}
+	return time.Now()
+}
+
+func (m *Manager) log() *slog.Logger {
+	if m.Logger != nil {
+		return m.Logger
+	}
+	return slog.Default()
+}
+
+// Start records the trust the process started with. Call once, before Ensure.
+func (m *Manager) Start(t proxmox.TLSConfig) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	m.cur = t
+	m.status = Status{Mode: modeOf(t)}
+}
+
+func modeOf(t proxmox.TLSConfig) string {
+	switch {
+	case t.InsecureSkipVerify:
+		return "insecure"
+	case t.Fingerprint != "":
+		return "fingerprint"
+	case t.CAFile != "":
+		return "ca"
+	}
+	return "system"
+}
+
+// Status is the current stanza (nil-safe).
+func (m *Manager) Status() *Status {
+	if m == nil {
+		return nil
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	s := m.status
+	return &s
+}
+
+// Moved reports whether the box trusts Proxmox by the CA copy (nothing is left to move).
+func (m *Manager) Moved() bool {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.cur.Fingerprint == "" && m.cur.CAFile != ""
+}
+
+// Run asks again on a timer until ctx ends: every `pending` while the box is not moved yet (a wrapper that did not
+// know the mode, a Proxmox that did not answer), every `settled` after (a renewed CA refreshes the copy).
+func (m *Manager) Run(ctx context.Context, pending, settled time.Duration) {
+	for {
+		d := settled
+		if !m.Moved() {
+			d = pending
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(d):
+			m.Ensure(ctx, "timer")
+		}
+	}
+}
+
+// Current is the trust the running client uses.
+func (m *Manager) Current() proxmox.TLSConfig {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.cur
+}
+
+// OnCallError is the hook for a failed Proxmox call: when the failure is one of TRUST, it names the cause and runs
+// Ensure (at most once per MinGap) — a pinned box whose certificate changed moves to the CA copy right then, and a box
+// whose CA copy was refreshed reloads it. A failure of another kind does nothing.
+func (m *Manager) OnCallError(ctx context.Context, err error) {
+	cause := proxmox.TrustProblem(err)
+	if cause == "" {
+		return
+	}
+	m.mu.Lock()
+	gap := m.MinGap
+	if gap <= 0 {
+		gap = 10 * time.Minute
+	}
+	tooSoon := !m.lastRun.IsZero() && m.now().Sub(m.lastRun) < gap
+	if m.status.Problem == "" {
+		m.status.Problem, m.status.Detail = cause, causeText(cause)
+	}
+	m.mu.Unlock()
+	m.log().Error("pvetrust: a call to Proxmox failed on TRUST — "+causeText(cause), "cause", cause, "err", err)
+	if tooSoon {
+		return
+	}
+	m.Ensure(ctx, cause)
+}
+
+func causeText(cause string) string {
+	switch cause {
+	case proxmox.TrustPinnedCertChanged:
+		return "the pinned certificate is not the one Proxmox serves any more (Proxmox made a new certificate)"
+	case proxmox.TrustCAMismatch:
+		return "the CA copy did not sign the certificate Proxmox serves (the box's root CA changed, or the copy is not this box's)"
+	case proxmox.TrustCertExpired:
+		return "the certificate Proxmox serves, or its CA, has ended"
+	case proxmox.TrustCAFileUnreadable:
+		return "the CA copy is missing or holds no certificate"
+	case proxmox.TrustNameMismatch:
+		return "the certificate Proxmox serves does not name the address the agent calls"
+	}
+	return cause
+}
+
+// Ensure runs the order of the package comment once. why is a log word ("start", "daily", a trust cause). It returns
+// true when the running trust changed.
+func (m *Manager) Ensure(ctx context.Context, why string) bool {
+	// One run at a time. Ensure is called from the start, the timer, a failed collect and after a bundle install; two
+	// overlapping runs would both read the config, and the second would park the ALREADY-MOVED file as the "before"
+	// copy (losing the fingerprint from the recovery copy). Found by review, 2026-11-14.
+	m.runMu.Lock()
+	defer m.runMu.Unlock()
+	m.mu.Lock()
+	cur := m.cur
+	m.lastRun = m.now()
+	m.mu.Unlock()
+	lg := m.log().With("why", why, "mode", modeOf(cur))
+	set := func(mode, problem, detail string, c *Copy) {
+		m.mu.Lock()
+		defer m.mu.Unlock()
+		moved := m.status.MovedAt
+		m.status = Status{Mode: mode, Problem: problem, Detail: detail, MovedAt: moved, CheckedAt: m.now().UTC().Format(time.RFC3339)}
+		if c != nil {
+			d := c.DaysLeft
+			m.status.CANotAfter, m.status.CADaysLeft = c.NotAfter, &d
+		}
+	}
+	if cur.InsecureSkipVerify {
+		lg.Warn("pvetrust: certificate checks are switched off in the config (insecure_skip_verify) — nothing is moved")
+		set("insecure", "", "", nil)
+		return false
+	}
+	if m.EnsureCopy == nil || m.Probe == nil {
+		return false
+	}
+	// 1. the copy. The root step gets a deadline of its own: Ensure runs on the daemon's start path and inside the
+	// report loop, and a wrapper that hangs (sudo, a stuck read of /etc/pve) must not stop either.
+	cctx, ccancel := context.WithTimeout(ctx, copyStepTimeout)
+	c, err := m.EnsureCopy(cctx)
+	ccancel()
+	if err != nil {
+		lg.Warn("pvetrust: the root step gave no CA copy — the trust stays as it is", "err", err)
+		set(modeOf(cur), ProblemCopyUnavailable, truncate(err.Error(), 240), nil)
+		return false
+	}
+	ca := proxmox.TLSConfig{CAFile: c.Path}
+	onCopy := cur.Fingerprint == "" && cur.CAFile == c.Path
+	if cur.Fingerprint == "" && cur.CAFile != "" && cur.CAFile != c.Path {
+		// an operator's own CA file: never replaced
+		lg.Info("pvetrust: the config names its own CA file — left as it is", "ca_file", cur.CAFile)
+		set("ca_file_own", "", "", &c)
+		return false
+	}
+	if onCopy && c.Outcome != "written" && why != proxmox.TrustCAMismatch && why != proxmox.TrustCAFileUnreadable && why != proxmox.TrustCertExpired {
+		// the box already trusts by the copy and the copy did not change: nothing to prove again
+		problem, detail := "", ""
+		if c.DaysLeft < EndsSoonDays {
+			problem, detail = ProblemCopyEndsSoon, fmt.Sprintf("the CA copy ends in %d day(s), on %s", c.DaysLeft, c.NotAfter)
+			lg.Error("pvetrust: "+detail, "days_left", c.DaysLeft)
+		}
+		set("ca", problem, detail, &c)
+		return false
+	}
+	// 2. the proof: a real call with the copy ALONE — the fingerprint is still the configured trust
+	pctx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	perr := m.Probe(pctx, ca)
+	cancel()
+	if perr != nil {
+		cause := proxmox.TrustProblem(perr)
+		detail := "a call to Proxmox with the CA copy alone failed"
+		if cause != "" {
+			detail += ": " + causeText(cause)
+		} else {
+			detail += " (not a trust failure: " + truncate(perr.Error(), 160) + ")"
+		}
+		if onCopy {
+			lg.Error("pvetrust: the box trusts Proxmox by the CA copy and the copy does not verify the served certificate", "err", perr)
+			set("ca", firstNonEmpty(cause, ProblemCopyNoProof), detail, &c)
+			return false
+		}
+		if cause == "" {
+			// Proxmox did not answer at all (it restarts, the box just booted): nothing is known about the copy, so no
+			// cause is named; the next run asks again.
+			lg.Warn("pvetrust: the proof could not be made now (not a trust failure) — the trust stays as it is; asked again later", "err", perr)
+			set(modeOf(cur), "", "", &c)
+			return false
+		}
+		lg.Error("pvetrust: the proof with the CA copy FAILED — the fingerprint STAYS, nothing was written", "err", perr)
+		set(modeOf(cur), ProblemCopyNoProof, detail+" — the fingerprint stays", &c)
+		return false
+	}
+	if onCopy {
+		// the copy was refreshed (or a trust failure asked): the running client reloads it
+		if m.Apply != nil {
+			if aerr := m.Apply(ca); aerr != nil {
+				set("ca", proxmox.TrustCAFileUnreadable, truncate(aerr.Error(), 200), &c)
+				return false
+			}
+		}
+		lg.Warn("pvetrust: the CA copy was refreshed and proven — the running client uses it now", "outcome", c.Outcome)
+		set("ca", "", "", &c)
+		return true
+	}
+	// 3. the config: ONE write sets ca_file and clears the fingerprint — only now, after the proof
+	if m.ConfigPath == "" {
+		lg.Warn("pvetrust: the proof worked, but the config did not come from a file — nothing can be saved, so the trust stays as it is")
+		set(modeOf(cur), ProblemConfigNotSaved, "the config is not a file; the move was not made", &c)
+		return false
+	}
+	if werr := m.writeConfig(c.Path); werr != nil {
+		lg.Error("pvetrust: the proof worked but the config could NOT be written — the fingerprint stays", "err", werr)
+		set(modeOf(cur), ProblemConfigNotSaved, truncate(werr.Error(), 200), &c)
+		return false
+	}
+	// 4. the running client
+	if m.Apply != nil {
+		if aerr := m.Apply(ca); aerr != nil {
+			lg.Error("pvetrust: the config is written but the running client could not take the CA copy — it keeps the old trust until the next start", "err", aerr)
+		}
+	}
+	m.mu.Lock()
+	m.cur = ca
+	m.mu.Unlock()
+	lg.Warn("pvetrust: MOVED — the agent trusts Proxmox by the box's own root CA now; the pinned fingerprint was cleared AFTER a working call with the copy",
+		"ca_file", c.Path, "ca_not_after", c.NotAfter)
+	set("ca", "", "", &c)
+	m.mu.Lock()
+	m.status.MovedAt = m.now().UTC().Format(time.RFC3339)
+	m.mu.Unlock()
+	return true
+}
+
+func firstNonEmpty(a, b string) string {
+	if a != "" {
+		return a
+	}
+	return b
+}
+
+func truncate(s string, n int) string {
+	if len(s) <= n {
+		return s
+	}
+	return s[:n]
+}
+
+// writeConfig sets proxmox.tls.ca_file and clears proxmox.tls.fingerprint in ONE write of agent.json. Every other key is
+// kept byte for byte. A recovery copy is parked in the state folder first; the result is read back and checked.
+func (m *Manager) writeConfig(caPath string) error {
+	config.FileMu.Lock() // one writer of agent.json at a time in this process (config.FileMu)
+	defer config.FileMu.Unlock()
+	raw, err := os.ReadFile(m.ConfigPath)
+	if err != nil {
+		return err
+	}
+	var doc map[string]json.RawMessage
+	if err := json.Unmarshal(raw, &doc); err != nil {
+		return fmt.Errorf("parse %s: %w", m.ConfigPath, err)
+	}
+	px := map[string]json.RawMessage{}
+	if cur, ok := doc["proxmox"]; ok {
+		if err := json.Unmarshal(cur, &px); err != nil {
+			return fmt.Errorf("parse the proxmox section: %w", err)
+		}
+	}
+	tlsSec := map[string]json.RawMessage{}
+	if cur, ok := px["tls"]; ok {
+		if err := json.Unmarshal(cur, &tlsSec); err != nil {
+			return fmt.Errorf("parse the proxmox.tls section: %w", err)
+		}
+	}
+	caJSON, _ := json.Marshal(caPath)
+	tlsSec["ca_file"] = caJSON
+	tlsSec["fingerprint"] = json.RawMessage(`""`)
+	b, _ := json.Marshal(tlsSec)
+	px["tls"] = b
+	b, _ = json.Marshal(px)
+	doc["proxmox"] = b
+	out, err := json.MarshalIndent(doc, "", "  ")
+	if err != nil {
+		return err
+	}
+	out = append(out, '\n')
+	if m.StateDir != "" {
+		if err := os.MkdirAll(m.StateDir, 0o700); err != nil {
+			return err
+		}
+		if err := os.WriteFile(filepath.Join(m.StateDir, "agent.json.pre-pve-trust"), raw, 0o600); err != nil {
+			return fmt.Errorf("parking the recovery copy: %w", err)
+		}
+	}
+	// In place, not tmp+rename: the folder is root-owned and the file is the agent's. config.ReplaceInPlace keeps the
+	// old bytes in a journal first, so a power cut in the middle is repaired at the next start (R-1380); it also reads
+	// the file back and puts the old bytes back when the write fails with the agent alive.
+	if err := config.ReplaceInPlace(m.ConfigPath, raw, out); err != nil {
+		return err
+	}
+	return nil
+}

internal/pvetrust/pvetrust_test.go

@@ -0,0 +1,489 @@
+package pvetrust
+
+import (
+	"context"
+	"crypto/ecdsa"
+	"crypto/elliptic"
+	"crypto/rand"
+	"crypto/sha256"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/hex"
+	"encoding/json"
+	"encoding/pem"
+	"errors"
+	"fmt"
+	"math/big"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/proxmox"
+)
+
+// ---- a small Proxmox: a root CA that signs node certificates, and a server that can change its certificate ----
+
+type pki struct {
+	t      *testing.T
+	caKey  *ecdsa.PrivateKey
+	caCert *x509.Certificate
+	caPEM  []byte
+}
+
+func newPKI(t *testing.T, cn string) *pki {
+	t.Helper()
+	k, _ := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	return (&pki{t: t, caKey: k}).renewCA(cn)
+}
+
+// renewCA makes a new CA certificate with the SAME key and name (what Proxmox does when its root CA is renewed).
+func (p *pki) renewCA(cn string) *pki {
+	tpl := &x509.Certificate{SerialNumber: big.NewInt(time.Now().UnixNano()), Subject: pkix.Name{CommonName: cn},
+		NotBefore: time.Now().Add(-time.Hour), NotAfter: time.Now().Add(24 * time.Hour), IsCA: true, BasicConstraintsValid: true,
+		KeyUsage: x509.KeyUsageCertSign | x509.KeyUsageCRLSign}
+	der, err := x509.CreateCertificate(rand.Reader, tpl, tpl, &p.caKey.PublicKey, p.caKey)
+	if err != nil {
+		p.t.Fatal(err)
+	}
+	p.caCert, _ = x509.ParseCertificate(der)
+	p.caPEM = pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: der})
+	return p
+}
+
+func (p *pki) leaf() tls.Certificate {
+	k, _ := ecdsa.GenerateKey(elliptic.P256(), rand.Reader)
+	tpl := &x509.Certificate{SerialNumber: big.NewInt(time.Now().UnixNano()), Subject: pkix.Name{CommonName: "node"},
+		NotBefore: time.Now().Add(-time.Hour), NotAfter: time.Now().Add(12 * time.Hour), IPAddresses: []net.IP{net.ParseIP("127.0.0.1")},
+		KeyUsage: x509.KeyUsageDigitalSignature, ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth}}
+	der, err := x509.CreateCertificate(rand.Reader, tpl, p.caCert, &k.PublicKey, p.caKey)
+	if err != nil {
+		p.t.Fatal(err)
+	}
+	return tls.Certificate{Certificate: [][]byte{der}, PrivateKey: k}
+}
+
+func fp(c tls.Certificate) string {
+	s := sha256.Sum256(c.Certificate[0])
+	return hex.EncodeToString(s[:])
+}
+
+type pve struct {
+	srv  *httptest.Server
+	mu   sync.Mutex
+	cert tls.Certificate
+}
+
+func newPVE(t *testing.T, c tls.Certificate) *pve {
+	p := &pve{cert: c}
+	p.srv = httptest.NewUnstartedServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		fmt.Fprint(w, `{"data":{"version":"9.2.2","release":"9.2","repoid":"x"}}`)
+	}))
+	// GetConfigForClient, not GetCertificate: httptest adds its own certificate, and for a call by IP address (no
+	// server name) that one would win over GetCertificate.
+	p.srv.TLS = &tls.Config{GetConfigForClient: func(*tls.ClientHelloInfo) (*tls.Config, error) {
+		p.mu.Lock()
+		defer p.mu.Unlock()
+		return &tls.Config{Certificates: []tls.Certificate{p.cert}}, nil
+	}}
+	p.srv.StartTLS()
+	t.Cleanup(p.srv.Close)
+	return p
+}
+
+// newCert is `pveupdate` making a new node certificate: served from the next connection on.
+func (p *pve) newCert(c tls.Certificate) {
+	p.mu.Lock()
+	p.cert = c
+	p.mu.Unlock()
+	p.srv.CloseClientConnections()
+}
+
+// ---- the agent's side ----
+
+type box struct {
+	t        *testing.T
+	m        *Manager
+	px       *proxmox.Client
+	cfgPath  string
+	copyPath string
+	source   []byte // what the root step would copy (nil = the root step refuses)
+	events   []string
+	applyErr error
+}
+
+func cfgJSON(fingerprint, caFile string) string {
+	return fmt.Sprintf(`{
+  "hub": {"url": "https://hub.example", "host_id": "box-1", "api_key": "KEEP-ME"},
+  "proxmox": {"endpoint": "X", "node": "n", "token": "u@pve!t=s", "tls": {"fingerprint": %q, "ca_file": %q, "insecure_skip_verify": false}, "custom": [1, 2, 3]},
+  "unknown_key": {"stays": true}
+}`, fingerprint, caFile)
+}
+
+func newBox(t *testing.T, srv *pve, start proxmox.TLSConfig, source []byte) *box {
+	t.Helper()
+	dir := t.TempDir()
+	b := &box{t: t, cfgPath: filepath.Join(dir, "agent.json"), copyPath: filepath.Join(dir, "pve-root-ca.pem"), source: source}
+	if start.CAFile == "COPY" {
+		start.CAFile = b.copyPath
+		if err := os.WriteFile(b.copyPath, source, 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := os.WriteFile(b.cfgPath, []byte(cfgJSON(start.Fingerprint, start.CAFile)), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	pcfg := proxmox.Config{Endpoint: srv.srv.URL, Node: "n", Token: "u@pve!t=s", TLS: start, HTTPTimeout: 5 * time.Second}
+	px, err := proxmox.NewClient(pcfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+	b.px = px
+	b.m = &Manager{ConfigPath: b.cfgPath, StateDir: filepath.Join(dir, "state"),
+		EnsureCopy: func(context.Context) (Copy, error) {
+			b.events = append(b.events, "copy")
+			if b.source == nil {
+				return Copy{}, errors.New(`pve-ca-copy refused: {"code": "T3", "reason": "not a CA certificate"}`)
+			}
+			outcome := "same"
+			if cur, _ := os.ReadFile(b.copyPath); string(cur) != string(b.source) {
+				outcome = "written"
+				if err := os.WriteFile(b.copyPath, b.source, 0o644); err != nil {
+					t.Fatal(err)
+				}
+			}
+			return Copy{Path: b.copyPath, Outcome: outcome, NotAfter: "2036-08-17T15:44:42Z", DaysLeft: 3560}, nil
+		},
+		Probe: func(ctx context.Context, tl proxmox.TLSConfig) error {
+			b.events = append(b.events, fmt.Sprintf("probe(fp=%q,ca=%v) while the file says fp=%q", tl.Fingerprint, tl.CAFile != "", b.fileTLS().Fingerprint))
+			c := pcfg
+			c.TLS = tl
+			p, err := proxmox.NewClient(c)
+			if err != nil {
+				return err
+			}
+			_, err = p.Version(ctx)
+			return err
+		},
+		Apply: func(tl proxmox.TLSConfig) error {
+			b.events = append(b.events, fmt.Sprintf("apply(fp=%q,ca=%v)", tl.Fingerprint, tl.CAFile != ""))
+			if b.applyErr != nil {
+				return b.applyErr
+			}
+			return b.px.SetTLS(tl)
+		},
+	}
+	b.m.Start(start)
+	return b
+}
+
+type fileTLS struct {
+	Fingerprint string `json:"fingerprint"`
+	CAFile      string `json:"ca_file"`
+}
+
+func (b *box) fileTLS() fileTLS {
+	var d struct {
+		Proxmox struct {
+			TLS fileTLS `json:"tls"`
+		} `json:"proxmox"`
+	}
+	raw, _ := os.ReadFile(b.cfgPath)
+	if err := json.Unmarshal(raw, &d); err != nil {
+		b.t.Fatalf("agent.json is not JSON any more: %v\n%s", err, raw)
+	}
+	return d.Proxmox.TLS
+}
+
+func (b *box) call() error {
+	_, err := b.px.Version(context.Background())
+	return err
+}
+
+func (b *box) has(prefix string) bool {
+	for _, e := range b.events {
+		if strings.HasPrefix(e, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
+// Scenario: a NEW box — the installer wrote the copy and ca_file, no fingerprint. Nothing is proven again, nothing written.
+func TestNewBox_CAOnly_NothingToMove(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	srv := newPVE(t, p.leaf())
+	b := newBox(t, srv, proxmox.TLSConfig{CAFile: "COPY"}, p.caPEM)
+	before, _ := os.ReadFile(b.cfgPath)
+	if b.m.Ensure(context.Background(), "start") {
+		t.Fatal("a box that already trusts by the copy changed its trust")
+	}
+	after, _ := os.ReadFile(b.cfgPath)
+	if string(before) != string(after) || b.has("probe") || b.has("apply") {
+		t.Fatalf("the config was touched or a proof ran: %v", b.events)
+	}
+	if st := b.m.Status(); st.Mode != "ca" || st.Problem != "" || st.CADaysLeft == nil || *st.CADaysLeft != 3560 {
+		t.Fatalf("status = %+v", st)
+	}
+	if err := b.call(); err != nil {
+		t.Fatalf("the call fails: %v", err)
+	}
+}
+
+// Scenario: an OLD box moves. The proof runs with the CA copy ALONE while the file still holds the fingerprint; only
+// then ONE write sets ca_file and clears the fingerprint; every other key stays; a recovery copy is parked.
+// COMPANION RED-PROOFS (observed, felhom.eu audits/night-build-2026-11-14/A/redproof-agent.txt): write the config
+// BEFORE the probe → the order check fails; skip the probe → TestProofFails_TheFingerprintStays fails.
+func TestOldBox_MovesAndTheFingerprintIsClearedOnlyAfterTheProof(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, p.caPEM)
+	old, _ := os.ReadFile(b.cfgPath)
+	if !b.m.Ensure(context.Background(), "start") {
+		t.Fatalf("the box did not move: %+v %v", b.m.Status(), b.events)
+	}
+	want := []string{"copy", fmt.Sprintf("probe(fp=\"\",ca=true) while the file says fp=%q", fp(leaf)), "apply(fp=\"\",ca=true)"}
+	if strings.Join(b.events, " | ") != strings.Join(want, " | ") {
+		t.Fatalf("order = %v\nwant    %v", b.events, want)
+	}
+	if got := b.fileTLS(); got.Fingerprint != "" || got.CAFile != b.copyPath {
+		t.Fatalf("the config's tls = %+v", got)
+	}
+	raw, _ := os.ReadFile(b.cfgPath)
+	for _, keep := range []string{`"KEEP-ME"`, `"unknown_key"`, `"stays"`, `"custom"`, `"u@pve!t=s"`, `"insecure_skip_verify"`} {
+		if !strings.Contains(string(raw), keep) {
+			t.Fatalf("the write lost %s:\n%s", keep, raw)
+		}
+	}
+	if parked, _ := os.ReadFile(filepath.Join(b.m.StateDir, "agent.json.pre-pve-trust")); string(parked) != string(old) {
+		t.Fatal("no recovery copy of the old config was parked")
+	}
+	if st := b.m.Status(); st.Mode != "ca" || st.MovedAt == "" || st.Problem != "" {
+		t.Fatalf("status = %+v", st)
+	}
+	// and now Proxmox makes a new certificate: the agent goes on, with no restart and no new client
+	srv.newCert(p.leaf())
+	if err := b.call(); err != nil {
+		t.Fatalf("after a new Proxmox certificate the moved box fails: %v", err)
+	}
+}
+
+// Scenario: the proof fails (the copy is another box's CA). The fingerprint stays — in the file, byte for byte, and in
+// the running client — and the cause is named.
+func TestProofFails_TheFingerprintStays(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	other := newPKI(t, "another box")
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, other.caPEM)
+	before, _ := os.ReadFile(b.cfgPath)
+	if b.m.Ensure(context.Background(), "start") {
+		t.Fatal("the trust changed although the proof failed")
+	}
+	after, _ := os.ReadFile(b.cfgPath)
+	if string(before) != string(after) {
+		t.Fatal("the config was written although the proof failed")
+	}
+	if b.has("apply") {
+		t.Fatalf("the running client was changed: %v", b.events)
+	}
+	if st := b.m.Status(); st.Mode != "fingerprint" || st.Problem != ProblemCopyNoProof || !strings.Contains(st.Detail, "did not sign") {
+		t.Fatalf("status = %+v", st)
+	}
+	if err := b.call(); err != nil {
+		t.Fatalf("the fingerprint no longer works: %v", err)
+	}
+}
+
+// Scenario: Proxmox makes a new certificate on a box that still PINS the old one. The failed call names the cause, the
+// box moves to the CA copy right then, and the next call works — no restart.
+func TestPinnedBox_NewCertificate_HealsWithNoRestart(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, p.caPEM)
+	srv.newCert(p.leaf())
+	err := b.call()
+	if proxmox.TrustProblem(err) != proxmox.TrustPinnedCertChanged {
+		t.Fatalf("fixture: want a pin mismatch, got %v", err)
+	}
+	b.m.OnCallError(context.Background(), err)
+	if err := b.call(); err != nil {
+		t.Fatalf("the box is still blind after the move: %v (%v)", err, b.events)
+	}
+	if got := b.fileTLS(); got.Fingerprint != "" || got.CAFile != b.copyPath {
+		t.Fatalf("the config's tls = %+v", got)
+	}
+	// a failure that is NOT one of trust does nothing
+	n := len(b.events)
+	b.m.OnCallError(context.Background(), errors.New("proxmox: GET /version: dial tcp 127.0.0.1:8006: connect: connection refused"))
+	if len(b.events) != n {
+		t.Fatalf("a refused connection started a trust run: %v", b.events[n:])
+	}
+}
+
+// Scenario: the root CA itself is renewed with the same key. The copy is refreshed, proven, and the running client
+// reloads it; the config is not written again.
+func TestCARenewedWithTheSameKey_TheCopyIsRefreshedAndReloaded(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	srv := newPVE(t, p.leaf())
+	b := newBox(t, srv, proxmox.TLSConfig{CAFile: "COPY"}, p.caPEM)
+	oldCA := string(p.caPEM)
+	p.renewCA("PVE CA")
+	srv.newCert(p.leaf())
+	b.source = p.caPEM
+	before, _ := os.ReadFile(b.cfgPath)
+	if !b.m.Ensure(context.Background(), "daily") {
+		t.Fatalf("a refreshed copy was not taken: %+v %v", b.m.Status(), b.events)
+	}
+	if cur, _ := os.ReadFile(b.copyPath); string(cur) == oldCA || string(cur) != string(p.caPEM) {
+		t.Fatal("the copy was not refreshed")
+	}
+	if !b.has("probe") || !b.has("apply") {
+		t.Fatalf("a refreshed copy must be proven and reloaded: %v", b.events)
+	}
+	if after, _ := os.ReadFile(b.cfgPath); string(after) != string(before) {
+		t.Fatal("the config was written again")
+	}
+	if err := b.call(); err != nil {
+		t.Fatalf("the call fails after the CA's renewal: %v", err)
+	}
+}
+
+// The root step refuses (the source is not a CA certificate): nothing changes, the cause is named.
+func TestRootStepRefuses_NothingChanges(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, nil)
+	before, _ := os.ReadFile(b.cfgPath)
+	if b.m.Ensure(context.Background(), "start") || b.has("probe") || b.has("apply") {
+		t.Fatalf("something ran without a copy: %v", b.events)
+	}
+	if after, _ := os.ReadFile(b.cfgPath); string(after) != string(before) {
+		t.Fatal("the config was written")
+	}
+	if st := b.m.Status(); st.Problem != ProblemCopyUnavailable || !strings.Contains(st.Detail, "T3") {
+		t.Fatalf("status = %+v", st)
+	}
+}
+
+// The proof works but the config cannot be written: the fingerprint stays, in the file and in the running client.
+func TestConfigCannotBeWritten_TheFingerprintStays(t *testing.T) {
+	if os.Geteuid() == 0 {
+		t.Skip("root writes a read-only file")
+	}
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, p.caPEM)
+	if err := os.Chmod(b.cfgPath, 0o400); err != nil {
+		t.Fatal(err)
+	}
+	if b.m.Ensure(context.Background(), "start") || b.has("apply") {
+		t.Fatalf("the trust changed without a saved config: %v", b.events)
+	}
+	if got := b.fileTLS(); got.Fingerprint != fp(leaf) {
+		t.Fatalf("the fingerprint is gone from the file: %+v", got)
+	}
+	if st := b.m.Status(); st.Mode != "fingerprint" || st.Problem != ProblemConfigNotSaved {
+		t.Fatalf("status = %+v", st)
+	}
+}
+
+// What is never moved: an operator's own CA file, a box with checks switched off, a config that is not a file.
+func TestNeverMoved(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	own := filepath.Join(t.TempDir(), "own-ca.pem")
+	os.WriteFile(own, p.caPEM, 0o644)
+	for name, start := range map[string]proxmox.TLSConfig{"own CA file": {CAFile: own}, "insecure": {InsecureSkipVerify: true}} {
+		b := newBox(t, srv, start, p.caPEM)
+		before, _ := os.ReadFile(b.cfgPath)
+		if b.m.Ensure(context.Background(), "start") || b.has("apply") {
+			t.Fatalf("%s: moved (%v)", name, b.events)
+		}
+		if after, _ := os.ReadFile(b.cfgPath); string(after) != string(before) {
+			t.Fatalf("%s: the config was written", name)
+		}
+	}
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, p.caPEM)
+	b.m.ConfigPath = ""
+	if b.m.Ensure(context.Background(), "start") || b.has("apply") || b.m.Status().Problem != ProblemConfigNotSaved {
+		t.Fatalf("a config that is not a file was 'moved': %v %+v", b.events, b.m.Status())
+	}
+}
+
+// The copy ends soon: the cause is named (the box still works).
+func TestCopyEndsSoon(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	srv := newPVE(t, p.leaf())
+	b := newBox(t, srv, proxmox.TLSConfig{CAFile: "COPY"}, p.caPEM)
+	inner := b.m.EnsureCopy
+	b.m.EnsureCopy = func(ctx context.Context) (Copy, error) {
+		c, err := inner(ctx)
+		c.DaysLeft, c.NotAfter = 12, "2026-11-26T00:00:00Z"
+		return c, err
+	}
+	b.m.Ensure(context.Background(), "daily")
+	if st := b.m.Status(); st.Problem != ProblemCopyEndsSoon || !strings.Contains(st.Detail, "12 day") {
+		t.Fatalf("status = %+v", st)
+	}
+}
+
+// A trust failure asks at most once per MinGap.
+func TestOnCallError_Throttled(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	other := newPKI(t, "another box")
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, other.caPEM)
+	srv.newCert(p.leaf())
+	err := b.call()
+	now := time.Date(2026, 11, 15, 3, 0, 0, 0, time.UTC)
+	b.m.Now = func() time.Time { return now }
+	b.m.OnCallError(context.Background(), err)
+	n := len(b.events)
+	b.m.OnCallError(context.Background(), err)
+	if len(b.events) != n {
+		t.Fatalf("a second failure inside the gap ran again: %v", b.events[n:])
+	}
+	now = now.Add(11 * time.Minute)
+	b.m.OnCallError(context.Background(), err)
+	if len(b.events) == n {
+		t.Fatal("no new run after the gap")
+	}
+	if st := b.m.Status(); st.Problem == "" {
+		t.Fatalf("a blind box names no cause: %+v", st)
+	}
+}
+
+// Proxmox does not answer at all (it restarts): nothing is known about the copy — no cause is named, nothing is written,
+// and the box counts as not moved, so the timer asks again soon.
+func TestProofCannotBeMadeNow_NoCauseNamed(t *testing.T) {
+	p := newPKI(t, "PVE CA")
+	leaf := p.leaf()
+	srv := newPVE(t, leaf)
+	b := newBox(t, srv, proxmox.TLSConfig{Fingerprint: fp(leaf)}, p.caPEM)
+	b.m.Probe = func(context.Context, proxmox.TLSConfig) error {
+		return errors.New("proxmox: GET /version: dial tcp 127.0.0.1:8006: connect: connection refused")
+	}
+	before, _ := os.ReadFile(b.cfgPath)
+	if b.m.Ensure(context.Background(), "start") {
+		t.Fatal("moved without a proof")
+	}
+	if after, _ := os.ReadFile(b.cfgPath); string(after) != string(before) {
+		t.Fatal("the config was written")
+	}
+	if st := b.m.Status(); st.Problem != "" || st.Mode != "fingerprint" || b.m.Moved() {
+		t.Fatalf("status = %+v moved=%v", st, b.m.Moved())
+	}
+}

internal/selfupdate/commit.go

@@ -128,4 +128,11 @@ func (m *Manager) MaybeCommit(ctx context.Context) {
 	if m.onCommitted != nil {
 		m.onCommitted(ctx)
 	}
+	// R-1368 part 2: the update ended healthy — the folder keeps the running version's binary and no other
+	// (prune.go). A refusal or a failed removal is logged and changes nothing about the commit.
+	if removed, err := pruneStaged(m.stateDir, m.runningVersion, m.logger); err != nil {
+		m.logger.Warn("selfupdate: old staged binaries NOT removed after the commit", "err", err)
+	} else if len(removed) > 0 {
+		m.logger.Info("selfupdate: old staged binaries removed after the commit", "count", len(removed), "kept", "felhom-agent-"+m.runningVersion)
+	}
 }

internal/selfupdate/executor.go

@@ -103,6 +103,9 @@ func (e *Executor) Execute(ctx context.Context, op string, params json.RawMessag
 	}
 	e.logger.Debug("agent_update: op invariants passed (semver + sha format)", "version", p.Version)
 
+	// The staged file is this call's from the download to the hand-off: a prune (prune.go) waits for it.
+	stagingMu.Lock()
+	defer stagingMu.Unlock()
 	dir := stagingDir(e.stateDir)
 	if err := os.MkdirAll(dir, 0o750); err != nil {
 		return fmt.Errorf("agent_update: staging dir: %w", err)

internal/selfupdate/prune.go

@@ -0,0 +1,96 @@
+package selfupdate
+
+import (
+	"fmt"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"sync"
+)
+
+// The self-update folder keeps what it needs and no more (R-1368 part 2; space list 2026-11-10, place H12).
+//
+// WHAT WAS WRONG. Execute downloads each new agent to <StateDir>/selfupdate/felhom-agent-<version> and removes that
+// file only when the download or its checksum fails. A good update left it there for ever: measured 2026-11-10, 49 /
+// 48 / 33 binaries = 740 / 724 / 509 MB on demo-hp, demo-felhom and Tester 1 — one for each version the box ever got.
+//
+// WHAT THE FOLDER IS NEEDED FOR. The root wrapper (felhom-os-apply, mode agent_update) reads the staged file ONCE,
+// copies it to its own root-owned folder and flips from that copy. After that hand-off nothing reads the staged file:
+//   - the rollback (felhom-selfupdate-guarded `rollback`) needs two files and neither is a staged binary:
+//     `pending.json` in this folder, and `/usr/local/bin/felhom-agent.prev` outside it;
+//   - the commit needs `pending.json` only.
+//
+// So a staged binary is dead weight as soon as its update ended. Pinned by TestPrune_RollbackReadsNoStagedBinary (the
+// wrapper's text) and TestPrune_RealRollbackStillWorksAfterNUpdates (the wrapper's rollback verb, run for real).
+//
+// THE RULE. After an update ended HEALTHY (the wrapper's `commit` returned success), remove every staged binary but
+// the one of the running version. Nothing is removed:
+//   - while a pending marker exists or cannot be read — an update is in flight, and that is the only time a rollback
+//     can fire;
+//   - after a rollback, or at a plain start — only the commit path calls this, so the binary of a version that failed
+//     stays for the operator to look at until the next good update;
+//   - that is not a regular file with exactly the name Execute gives a staged binary. `pending.json`, a folder, a
+//     symlink, and any name this code does not know are never touched.
+//
+// NOT REACHED BY THIS: `/usr/local/bin/felhom-agent.bak-*` (space list H13). Those files are root-owned, in a folder
+// the agent cannot write, and no code of the agent makes them (a person's copy before a hand deploy). R-1368 part 4.
+
+// stagedNameRe is the exact name Execute gives a staged binary: "felhom-agent-" + a version that versionRe accepts.
+// Kept in step with versionRe by TestPrune_NameShapeFollowsTheVersionRule.
+var stagedNameRe = regexp.MustCompile(`^felhom-agent-([0-9]+\.[0-9]+\.[0-9]+(-[0-9A-Za-z.]+)?)$`)
+
+// stagingMu makes a download and a prune take turns. Without it a prune that runs while Execute downloads the NEXT
+// version would remove that half-written file (its name is a staged binary's and it is not the running version).
+var stagingMu sync.Mutex
+
+// pruneStaged removes the staged binaries the folder no longer needs and returns the names it removed, sorted. It
+// refuses (removes nothing, returns an error) while an update is in flight. A file it cannot remove is logged and
+// skipped: a failed clean-up must never fail the commit that came before it.
+func pruneStaged(stateDir, runningVersion string, logger *slog.Logger) ([]string, error) {
+	logger = orDefaultLogger(logger)
+	stagingMu.Lock()
+	defer stagingMu.Unlock()
+
+	if !versionRe.MatchString(runningVersion) {
+		return nil, fmt.Errorf("selfupdate prune: refusing — running version %q is not a version this code knows; it could not name the file to keep", runningVersion)
+	}
+	pm, err := readPending(stateDir)
+	if err != nil {
+		return nil, fmt.Errorf("selfupdate prune: refusing — the pending marker cannot be read: %w", err)
+	}
+	if pm != nil {
+		return nil, fmt.Errorf("selfupdate prune: refusing — an update to %s is pending (a rollback may still fire)", pm.NewVersion)
+	}
+	dir := stagingDir(stateDir)
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil
+		}
+		return nil, fmt.Errorf("selfupdate prune: read %s: %w", dir, err)
+	}
+	keep := "felhom-agent-" + runningVersion
+	var removed []string
+	for _, ent := range entries {
+		name := ent.Name()
+		if name == keep || !stagedNameRe.MatchString(name) {
+			continue // the running version's binary, pending.json, or a name this code does not know
+		}
+		if !ent.Type().IsRegular() {
+			logger.Warn("selfupdate prune: a staged-binary name that is not a regular file — left alone", "name", name, "type", ent.Type().String())
+			continue
+		}
+		// name came from ReadDir and matched stagedNameRe, so it holds no path separator: this is one exact file.
+		if err := os.Remove(filepath.Join(dir, name)); err != nil {
+			logger.Warn("selfupdate prune: cannot remove an old staged binary — left for the next good update", "name", name, "err", err)
+			continue
+		}
+		logger.Debug("selfupdate prune: removed an old staged binary", "name", name)
+		removed = append(removed, name)
+	}
+	sort.Strings(removed)
+	logger.Info("selfupdate prune: the self-update folder holds what it needs", "kept", keep, "removed", len(removed))
+	return removed, nil
+}

internal/selfupdate/prune_test.go

@@ -0,0 +1,373 @@
+package selfupdate
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strings"
+	"testing"
+)
+
+// R-1368 part 2 — the self-update folder keeps what it needs and no more (prune.go).
+//
+// RED-PROOFS (each observed 2026-11-10, then restored):
+//   - drop the pruneStaged call from MaybeCommit → TestPrune_AfterNUpdatesTheFolderHoldsOneBinary fails with
+//     "after update 1 (0.170.0) the folder holds [… felhom-agent-0.169.0 felhom-agent-0.170.0 …]", and
+//     TestPrune_RealRollbackStillWorksAfterNUpdates fails too;
+//   - drop the pending-marker refusal from pruneStaged → TestPrune_RefusesWhileAnUpdateIsPending fails with
+//     "removed [...] while an update was pending";
+//   - drop the IsRegular check → TestPrune_AfterNUpdatesTheFolderHoldsOneBinary fails with "a symlink with a
+//     staged-binary name was removed".
+
+// committingWrapper is the guarded wrapper's `commit` as far as this package can see it: it clears the pending marker.
+type committingWrapper struct {
+	stateDir string
+	calls    int
+}
+
+func (w *committingWrapper) Run(_ context.Context, _ string, args ...string) ([]byte, []byte, error) {
+	if len(args) == 1 && args[0] == "commit" {
+		w.calls++
+		if err := os.Remove(pendingPath(w.stateDir)); err != nil {
+			return nil, []byte(err.Error()), err
+		}
+	}
+	return []byte("ok"), nil, nil
+}
+
+func stage(t *testing.T, stateDir, version string) {
+	t.Helper()
+	if err := os.MkdirAll(stagingDir(stateDir), 0o750); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(stagingDir(stateDir), "felhom-agent-"+version), []byte("binary "+version), 0o755); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func folder(t *testing.T, stateDir string) []string {
+	t.Helper()
+	ents, err := os.ReadDir(stagingDir(stateDir))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var out []string
+	for _, e := range ents {
+		out = append(out, e.Name())
+	}
+	sort.Strings(out)
+	return out
+}
+
+// The consequence: after N good updates the folder holds ONE staged binary — the running version's — and every file
+// that is not a staged binary is still there.
+func TestPrune_AfterNUpdatesTheFolderHoldsOneBinary(t *testing.T) {
+	stateDir := t.TempDir()
+	dir := stagingDir(stateDir)
+	if err := os.MkdirAll(dir, 0o750); err != nil {
+		t.Fatal(err)
+	}
+	// Bystanders: names this code does not know, and two things that carry a staged binary's NAME but are not files.
+	outside := filepath.Join(t.TempDir(), "precious")
+	if err := os.WriteFile(outside, []byte("not ours"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	for _, n := range []string{"notes.txt", "felhom-agent-0.1.0.partial", "felhom-agent.bak-0.139.0"} {
+		if err := os.WriteFile(filepath.Join(dir, n), []byte("x"), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := os.Mkdir(filepath.Join(dir, "felhom-agent-9.9.9"), 0o750); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(outside, filepath.Join(dir, "felhom-agent-8.8.8")); err != nil {
+		t.Fatal(err)
+	}
+	bystanders := []string{"felhom-agent-0.1.0.partial", "felhom-agent-8.8.8", "felhom-agent-9.9.9", "felhom-agent.bak-0.139.0", "notes.txt"}
+
+	versions := []string{"0.170.0", "0.171.0", "0.172.0", "0.172.1", "0.173.0-rc.1", "0.173.0"}
+	prev := "0.169.0"
+	stage(t, stateDir, prev) // what an old agent left behind
+	for i, v := range versions {
+		stage(t, stateDir, v)                                                                                   // Execute downloaded it
+		writePending(t, stateDir, PendingMarker{OldVersion: prev, NewVersion: v, SHA256: "ab", AppliedAt: "t"}) // the wrapper flipped
+		wrap := &committingWrapper{stateDir: stateDir}
+		m := newMgr(t, stateDir, v, wrap, nil)
+		m.MaybeCommit(context.Background())
+		if wrap.calls != 1 {
+			t.Fatalf("update %d (%s): commit calls = %d, want 1", i+1, v, wrap.calls)
+		}
+		want := append([]string{"felhom-agent-" + v}, bystanders...)
+		sort.Strings(want)
+		got := folder(t, stateDir)
+		if strings.Join(got, " ") != strings.Join(want, " ") {
+			if _, err := os.Lstat(filepath.Join(dir, "felhom-agent-8.8.8")); err != nil {
+				t.Fatalf("a symlink with a staged-binary name was removed: %v", err)
+			}
+			t.Fatalf("after update %d (%s) the folder holds %v, want %v", i+1, v, got, want)
+		}
+		prev = v
+	}
+	if b, err := os.ReadFile(outside); err != nil || string(b) != "not ours" {
+		t.Fatalf("the file behind the symlink changed: %q %v", b, err)
+	}
+}
+
+// While an update is pending — the only time a rollback can fire — nothing is removed.
+func TestPrune_RefusesWhileAnUpdateIsPending(t *testing.T) {
+	stateDir := t.TempDir()
+	for _, v := range []string{"0.170.0", "0.171.0", "0.172.0"} {
+		stage(t, stateDir, v)
+	}
+	writePending(t, stateDir, PendingMarker{OldVersion: "0.171.0", NewVersion: "0.172.0", SHA256: "ab", AppliedAt: "t"})
+	before := folder(t, stateDir)
+
+	removed, err := pruneStaged(stateDir, "0.172.0", nil)
+	if err == nil || len(removed) != 0 {
+		t.Fatalf("removed %v while an update was pending (err=%v) — want a refusal", removed, err)
+	}
+	if got := folder(t, stateDir); strings.Join(got, " ") != strings.Join(before, " ") {
+		t.Fatalf("removed %v while an update was pending; folder now %v", before, got)
+	}
+
+	// The same through the commit path: a `commit` that reports success but leaves the marker (fakeWrapper does) must
+	// not lead to a prune.
+	m := newMgr(t, stateDir, "0.172.0", &fakeWrapper{}, nil)
+	m.MaybeCommit(context.Background())
+	if got := folder(t, stateDir); strings.Join(got, " ") != strings.Join(before, " ") {
+		t.Fatalf("the commit path pruned with the marker still there; folder now %v", got)
+	}
+
+	// A marker nobody can read is a refusal too.
+	if err := os.WriteFile(pendingPath(stateDir), []byte("{not json"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if removed, err := pruneStaged(stateDir, "0.172.0", nil); err == nil || len(removed) != 0 {
+		t.Fatalf("an unreadable marker must refuse; removed %v err=%v", removed, err)
+	}
+}
+
+// Only a commit prunes: a plain start, a version mismatch and a failed commit remove nothing — so the binary of a
+// version that was rolled back stays for the operator.
+func TestPrune_OnlyAHealthyCommitPrunes(t *testing.T) {
+	cases := []struct {
+		name    string
+		pending *PendingMarker
+		running string
+		wrap    WrapperRunner
+	}{
+		{"plain start, no marker (also the state after a rollback)", nil, "0.171.0", &fakeWrapper{}},
+		{"marker names another version", &PendingMarker{OldVersion: "0.171.0", NewVersion: "0.172.0"}, "0.171.0", &fakeWrapper{}},
+		{"commit fails", &PendingMarker{OldVersion: "0.171.0", NewVersion: "0.172.0"}, "0.172.0", &fakeWrapper{err: fmt.Errorf("sudo: no")}},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			stateDir := t.TempDir()
+			for _, v := range []string{"0.170.0", "0.171.0", "0.172.0"} {
+				stage(t, stateDir, v)
+			}
+			if c.pending != nil {
+				writePending(t, stateDir, *c.pending)
+			}
+			before := folder(t, stateDir)
+			newMgr(t, stateDir, c.running, c.wrap, nil).MaybeCommit(context.Background())
+			if got := folder(t, stateDir); strings.Join(got, " ") != strings.Join(before, " ") {
+				t.Fatalf("folder changed without a healthy commit: %v → %v", before, got)
+			}
+		})
+	}
+	// A running version this code cannot name a file for: refuse, do not guess.
+	stateDir := t.TempDir()
+	stage(t, stateDir, "0.170.0")
+	if removed, err := pruneStaged(stateDir, "dev", nil); err == nil || len(removed) != 0 {
+		t.Fatalf("running version \"dev\" must refuse; removed %v err=%v", removed, err)
+	}
+}
+
+// The name rule and the version rule say the same thing: every name Execute can write is a name the prune knows, and
+// nothing else is.
+func TestPrune_NameShapeFollowsTheVersionRule(t *testing.T) {
+	for _, v := range []string{"0.70.1", "0.172.1", "10.0.0", "0.70.2-crash", "0.173.0-rc.1"} {
+		if !versionRe.MatchString(v) {
+			t.Fatalf("test input %q is not a version versionRe accepts", v)
+		}
+		if !stagedNameRe.MatchString("felhom-agent-" + v) {
+			t.Errorf("Execute can stage felhom-agent-%s but the prune does not know that name", v)
+		}
+	}
+	for _, n := range []string{"pending.json", "felhom-agent", "felhom-agent-", "felhom-agent-1.2", "felhom-agent-0.1.0.partial",
+		"felhom-agent.bak-0.139.0", "felhom-agent.prev", "felhom-agent-0.1.0/x", "../felhom-agent-0.1.0", "felhom-agent-0.1.0 ", "xfelhom-agent-0.1.0"} {
+		if stagedNameRe.MatchString(n) {
+			t.Errorf("the prune would treat %q as a staged binary", n)
+		}
+	}
+	m := stagedNameRe.FindStringSubmatch("felhom-agent-0.173.0-rc.1")
+	if m == nil || !versionRe.MatchString(m[1]) {
+		t.Fatalf("the version part of a staged name must satisfy versionRe; got %v", m)
+	}
+}
+
+func wrapperText(t *testing.T) string {
+	t.Helper()
+	b, err := os.ReadFile(filepath.Join("..", "..", "configs", "felhom-selfupdate-guarded"))
+	if err != nil {
+		t.Fatalf("the guarded wrapper is the other half of this proof and must be readable: %v", err)
+	}
+	return string(b)
+}
+
+// The stated dependency, pinned: the wrapper's `rollback` and `commit` read the pending marker and `.prev`, and no
+// staged binary. If either verb starts to read the staging folder, this fails and prune.go's rule must be looked at.
+func TestPrune_RollbackReadsNoStagedBinary(t *testing.T) {
+	s := wrapperText(t)
+	for _, want := range []string{"\nBIN=/usr/local/bin/felhom-agent\n", "\nPREV=$BIN.prev\n", "\nSTAGING=/var/lib/felhom-agent/selfupdate\n", "\nPENDING=$STAGING/pending.json\n"} {
+		if !strings.Contains(s, want) {
+			t.Fatalf("the wrapper no longer holds the line %q — the folder's rule (prune.go) was written against it", strings.TrimSpace(want))
+		}
+	}
+	for _, verb := range []string{"commit", "rollback"} {
+		i := strings.Index(s, "\n"+verb+")\n")
+		if i < 0 {
+			t.Fatalf("no %s) block in the wrapper", verb)
+		}
+		block := s[i:]
+		j := strings.Index(block, "\n    ;;\n")
+		if j < 0 {
+			t.Fatalf("no end of the %s) block", verb)
+		}
+		block = block[:j]
+		if !strings.Contains(block, "$PENDING") {
+			t.Fatalf("%s) does not read $PENDING any more", verb)
+		}
+		code := regexp.MustCompile(`(?m)^\s*#.*$`).ReplaceAllString(block, "")
+		for _, bad := range []string{"STAGING", "staged", "felhom-agent-"} {
+			if strings.Contains(code, bad) {
+				t.Fatalf("%s) now mentions %q — it may read a staged binary, and the prune removes those", verb, bad)
+			}
+		}
+	}
+	if stagedNameRe.MatchString("pending.json") || stagedNameRe.MatchString("felhom-agent.prev") {
+		t.Fatal("the prune would remove a file the rollback needs")
+	}
+}
+
+// The consequence, with the wrapper's own code: after N good updates (each committed by the REAL `commit` verb and
+// pruned), update N+1 goes bad and the REAL `rollback` verb puts the previous binary back. The wrapper runs from a copy
+// whose four fixed paths point into a temp folder; systemctl, logger and chown are stubs.
+func TestPrune_RealRollbackStillWorksAfterNUpdates(t *testing.T) {
+	sh, err := exec.LookPath("sh")
+	if err != nil {
+		t.Skip("no sh on this machine")
+	}
+	root := t.TempDir()
+	stateDir := filepath.Join(root, "state")
+	binDir := filepath.Join(root, "bin")
+	stubs := filepath.Join(root, "stubs")
+	for _, d := range []string{stagingDir(stateDir), binDir, stubs, filepath.Join(root, "rootstaging")} {
+		if err := os.MkdirAll(d, 0o750); err != nil {
+			t.Fatal(err)
+		}
+	}
+	live := filepath.Join(binDir, "felhom-agent")
+	s := wrapperText(t)
+	for old, repl := range map[string]string{
+		"\nBIN=/usr/local/bin/felhom-agent\n":                    "\nBIN=" + live + "\n",
+		"\nSTAGING=/var/lib/felhom-agent/selfupdate\n":           "\nSTAGING=" + stagingDir(stateDir) + "\n",
+		"\nROOT_STAGING=/var/lib/felhom-os-apply/agent-update\n": "\nROOT_STAGING=" + filepath.Join(root, "rootstaging") + "\n",
+	} {
+		if strings.Count(s, old) != 1 {
+			t.Fatalf("the wrapper's path line %q is not there exactly once — this test cannot sandbox it", strings.TrimSpace(old))
+		}
+		s = strings.Replace(s, old, repl, 1)
+	}
+	// Safety: no assignment in the copy may still point at a real system path (PREV and PENDING are built from BIN and
+	// STAGING, so they follow).
+	for _, line := range strings.Split(s, "\n") {
+		if regexp.MustCompile(`^[A-Z_]+=/`).MatchString(line) && !strings.Contains(line, root) {
+			t.Fatalf("the sandboxed wrapper still holds a real path: %q", line)
+		}
+	}
+	script := filepath.Join(root, "wrapper.sh")
+	if err := os.WriteFile(script, []byte(s), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	stubLog := filepath.Join(root, "stub.log")
+	for _, n := range []string{"systemctl", "logger", "chown", "systemd-run"} {
+		body := "#!/bin/sh\necho \"" + n + " $*\" >> '" + stubLog + "'\nexit 0\n"
+		if err := os.WriteFile(filepath.Join(stubs, n), []byte(body), 0o755); err != nil {
+			t.Fatal(err)
+		}
+	}
+	run := func(verb string) (string, error) {
+		cmd := exec.Command(sh, script, verb)
+		cmd.Env = []string{"PATH=" + stubs + ":/usr/bin:/bin"}
+		out, err := cmd.CombinedOutput()
+		return string(out), err
+	}
+	// applied is what the wrapper's `apply` leaves behind (it needs root, so the test writes the same four things).
+	applied := func(oldV, newV string) {
+		cur, err := os.ReadFile(live)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(live+".prev", cur, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(live, []byte("binary "+newV), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writePending(t, stateDir, PendingMarker{OldVersion: oldV, NewVersion: newV, SHA256: "ab", AppliedAt: "t"})
+	}
+	if err := os.WriteFile(live, []byte("binary 0.169.0"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+
+	prev := "0.169.0"
+	for _, v := range []string{"0.170.0", "0.171.0", "0.172.0", "0.172.1"} {
+		stage(t, stateDir, v)
+		applied(prev, v)
+		m := newMgr(t, stateDir, v, runnerFunc(func(args ...string) ([]byte, []byte, error) {
+			out, err := run(args[0])
+			return []byte(out), nil, err
+		}), nil)
+		m.MaybeCommit(context.Background()) // the REAL commit verb, then the prune
+		if got := folder(t, stateDir); len(got) != 1 || got[0] != "felhom-agent-"+v {
+			t.Fatalf("after the good update to %s the folder holds %v", v, got)
+		}
+		prev = v
+	}
+
+	// Update N+1 is applied and goes bad: systemd's OnFailure runs the wrapper's rollback.
+	stage(t, stateDir, "0.173.0")
+	applied("0.172.1", "0.173.0")
+	out, err := run("rollback")
+	if err != nil {
+		t.Fatalf("the real rollback verb failed after the prunes: %v\n%s", err, out)
+	}
+	if b, _ := os.ReadFile(live); string(b) != "binary 0.172.1" {
+		t.Fatalf("after the rollback the live binary is %q, want the previous version's", b)
+	}
+	if p, _ := readPending(stateDir); p != nil {
+		t.Fatal("the rollback left the pending marker")
+	}
+	if b, _ := os.ReadFile(stubLog); !bytes.Contains(b, []byte("systemctl restart felhom-agent.service")) {
+		t.Fatalf("the rollback did not restart the agent; stub log:\n%s", b)
+	}
+	// The rolled-back agent starts: no marker, so no commit and no prune — the bad version's binary stays to be read.
+	newMgr(t, stateDir, "0.172.1", &fakeWrapper{}, nil).MaybeCommit(context.Background())
+	if got := folder(t, stateDir); strings.Join(got, " ") != "felhom-agent-0.172.1 felhom-agent-0.173.0" {
+		t.Fatalf("after the rollback the folder holds %v, want the running version and the one that failed", got)
+	}
+}
+
+// runnerFunc adapts a function to WrapperRunner; it gets the verb arguments only.
+type runnerFunc func(args ...string) ([]byte, []byte, error)
+
+func (f runnerFunc) Run(_ context.Context, _ string, args ...string) ([]byte, []byte, error) {
+	return f(args...)
+}

internal/signedjobs/outcome_test.go

@@ -0,0 +1,101 @@
+package signedjobs
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.dooplex.hu/admin/felhom-agent/internal/hub"
+	"gitea.dooplex.hu/admin/felhom-agent/internal/reconcile"
+)
+
+// outcomeQueue is a jobsQueue that also has the optional half (OutcomeSource), as *hub.Client does.
+type outcomeQueue struct {
+	jobsQueue
+	omu      sync.Mutex
+	outcomes map[string]hub.JobOutcome
+}
+
+func (q *outcomeQueue) CompleteJobWithOutcome(ctx context.Context, jobID string, o hub.JobOutcome) error {
+	q.omu.Lock()
+	if q.outcomes == nil {
+		q.outcomes = map[string]hub.JobOutcome{}
+	}
+	q.outcomes[jobID] = o
+	q.omu.Unlock()
+	return q.jobsQueue.CompleteJob(ctx, jobID)
+}
+
+// R-1383 (D78): a signed job that is refused or fails was cleared from the hub's queue exactly like one that ran —
+// „an agent update never arrives” left one journal line on the box. The runner now says how each job ended, through
+// the REAL gate and verifier: the stored effect is the outcome the hub receives with the clearing call.
+// The two clock reasons are the ones the ruling is about: a box with a wrong clock refuses every job.
+// COMPANION RED-PROOFS (observed, felhom.eu audits/night-build-2026-11-14/B/redproof-agent-job-outcome.txt).
+func TestRunner_TellsTheHubHowEachJobEnded(t *testing.T) {
+	s := newTestSigner(t)
+	_, _, exec := newRealGateRunner(t, s)
+	base, _, _ := newRealGateRunner(t, s)
+	src := &outcomeQueue{}
+	r := NewRunner(src, base.gate, exec, testHost, quiet())
+	now := time.Now().UTC()
+
+	src.add(mintJob(t, s, "ok", testHost, "", "ops-1", wipeParamsJSON, now, now.Add(time.Hour)))
+	src.add(mintJob(t, s, "expired", testHost, "", "ops-1", wipeParamsJSON, now.Add(-3*time.Hour), now.Add(-2*time.Hour)))
+	src.add(mintJob(t, s, "early", testHost, "", "ops-1", wipeParamsJSON, now.Add(2*time.Hour), now.Add(3*time.Hour)))
+	src.add(mintJob(t, s, "other-box", "some-other-host", "", "ops-1", wipeParamsJSON, now, now.Add(time.Hour)))
+	src.add(corruptSig(t, mintJob(t, s, "forged", testHost, "", "ops-1", wipeParamsJSON, now, now.Add(time.Hour))))
+	src.add(hub.JobWire{JobID: "garbage", BlobB64: "!!!not-base64!!!"})
+	if _, err := r.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	// and a job that is accepted and whose run fails
+	exec.err = errors.New("the wrapper refused")
+	src.add(mintJob(t, s, "run-fails", testHost, "", "ops-1", wipeParamsJSON, now, now.Add(time.Hour)))
+	if _, err := r.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	want := map[string]hub.JobOutcome{
+		"ok":        {Outcome: "done", Op: "storage_wipe"},
+		"expired":   {Outcome: "rejected", Reason: "expired", Op: "storage_wipe"},
+		"early":     {Outcome: "rejected", Reason: "not_yet_valid", Op: "storage_wipe"},
+		"other-box": {Outcome: "rejected", Reason: "target", Op: "storage_wipe"},
+		"forged":    {Outcome: "rejected", Reason: "malformed", Op: "storage_wipe"}, // corruptSig breaks the armour itself
+		"garbage":   {Outcome: "malformed", Reason: "malformed"},
+		"run-fails": {Outcome: "failed", Reason: "execution_failed", Op: "storage_wipe"},
+	}
+	for id, w := range want {
+		if got, ok := src.outcomes[id]; !ok || got != w {
+			t.Errorf("job %q: the hub was told %+v (told=%v), want %+v", id, got, ok, w)
+		}
+	}
+	if len(src.outcomes) != len(want) {
+		t.Errorf("outcomes for %d jobs, want %d: %+v", len(src.outcomes), len(want), src.outcomes)
+	}
+	if exec.count() != 2 {
+		t.Errorf("executor ran %d times, want 2 (ok and run-fails only)", exec.count())
+	}
+}
+
+// Every code reasonCode can return is on the closed list the hub holds a sentence for.
+func TestReasonCode_ClosedList(t *testing.T) {
+	known := map[string]bool{}
+	for _, c := range ReasonCodes {
+		known[c] = true
+	}
+	for _, d := range []reconcile.Decision{
+		{}, {Reason: reconcile.ReasonRejected, Err: errors.New("something new")},
+		{Reason: reconcile.ReasonRoleDenied}, {Reason: reconcile.ReasonBindingMismatch}, {Reason: reconcile.ReasonPendingSignature},
+	} {
+		if c := reasonCode(d); !known[c] {
+			t.Errorf("reasonCode(%+v) = %q, not on ReasonCodes", d, c)
+		}
+	}
+	if reasonCode(reconcile.Decision{Reason: reconcile.ReasonRejected, Err: errors.New("x")}) != "other" {
+		t.Error("an unknown refusal must be told as „other”, never as the error's own text")
+	}
+	if !known["execution_failed"] || !known["malformed"] {
+		t.Error("the two codes the runner sets itself are not on the list")
+	}
+}

internal/signedjobs/runner.go

@@ -35,6 +35,42 @@ type JobSource interface {
 	CompleteJob(ctx context.Context, jobID string) error
 }
 
+// OutcomeSource is the optional half of a JobSource: it clears a job AND tells the hub how it ended (R-1383, D78).
+// *hub.Client has it. A source without it is cleared as before.
+type OutcomeSource interface {
+	CompleteJobWithOutcome(ctx context.Context, jobID string, o hub.JobOutcome) error
+}
+
+// ReasonCodes is the CLOSED list of reasons the agent sends with a job that did not run to a good end. The hub holds
+// one operator sentence per code (felhom.eu hub/internal/api, signedJobReasons); a code the hub does not know is
+// told as „other”. Pinned by TestReasonCode_ClosedList.
+var ReasonCodes = []string{"expired", "not_yet_valid", "window_too_long", "replay", "bad_signature", "unknown_signer",
+	"target", "namespace", "malformed", "role_denied", "binding_mismatch", "execution_failed", "other"}
+
+// reasonCode names why the gate refused: the authz error first (the two clock reasons matter most — a box with a
+// wrong clock refuses EVERY job), then the gate's own reason.
+func reasonCode(dec reconcile.Decision) string {
+	for _, m := range []struct {
+		err  error
+		code string
+	}{
+		{authz.ErrExpired, "expired"}, {authz.ErrNotYetValid, "not_yet_valid"}, {authz.ErrWindowTooLong, "window_too_long"},
+		{authz.ErrReplay, "replay"}, {authz.ErrBadSignature, "bad_signature"}, {authz.ErrUnknownSigner, "unknown_signer"},
+		{authz.ErrTarget, "target"}, {authz.ErrNamespace, "namespace"}, {authz.ErrMalformed, "malformed"},
+	} {
+		if errorsIs(dec.Err, m.err) {
+			return m.code
+		}
+	}
+	switch dec.Reason {
+	case reconcile.ReasonRoleDenied:
+		return "role_denied"
+	case reconcile.ReasonBindingMismatch:
+		return "binding_mismatch"
+	}
+	return "other"
+}
+
 // Authorizer is the reversibility gate. Satisfied by *reconcile.Gate.
 type Authorizer interface {
 	Authorize(intent reconcile.Intent, signed *reconcile.SignedOp) reconcile.Decision
@@ -141,7 +177,7 @@ func (r *Runner) processJob(ctx context.Context, j hub.JobWire) bool {
 		// A malformed/forged envelope is permanently bad — clear it loudly (a queued garbage job
 		// must not wedge the queue). No nonce was consumed (we never reached Verify).
 		r.logger.Error("signedjobs: malformed job envelope — clearing", "job", j.JobID)
-		r.complete(ctx, j.JobID)
+		r.complete(ctx, j.JobID, hub.JobOutcome{Outcome: "malformed", Reason: "malformed"})
 		return true
 	}
 
@@ -166,7 +202,7 @@ func (r *Runner) processJob(ctx context.Context, j hub.JobWire) bool {
 		// it as a security event (a compromised hub queuing a forged op lands here).
 		r.logger.Warn("signedjobs: REJECTED signed op — executor not called",
 			"job", j.JobID, "op", ob.Op, "reason", dec.Reason, "err", errStr(dec.Err))
-		r.complete(ctx, j.JobID)
+		r.complete(ctx, j.JobID, hub.JobOutcome{Outcome: "rejected", Reason: reasonCode(dec), Op: ob.Op})
 		return true
 	}
 
@@ -177,7 +213,7 @@ func (r *Runner) processJob(ctx context.Context, j hub.JobWire) bool {
 	switch {
 	case err == nil:
 		r.logger.Warn("signedjobs: signed op COMPLETED", "job", j.JobID, "op", ob.Op)
-		r.complete(ctx, j.JobID)
+		r.complete(ctx, j.JobID, hub.JobOutcome{Outcome: "done", Op: ob.Op})
 		return true
 	case errorsIs(err, ErrNoExecutor):
 		// No executor for this op class in this build — leave it queued for the owning slice.
@@ -189,7 +225,7 @@ func (r *Runner) processJob(ctx context.Context, j hub.JobWire) bool {
 		// nonce to retry) and log the failure loudly.
 		r.logger.Error("signedjobs: signed op execution FAILED (nonce spent — clearing)",
 			"job", j.JobID, "op", ob.Op, "err", err)
-		r.complete(ctx, j.JobID)
+		r.complete(ctx, j.JobID, hub.JobOutcome{Outcome: "failed", Reason: "execution_failed", Op: ob.Op})
 		return true
 	}
 }
@@ -215,8 +251,14 @@ func (r *Runner) decode(j hub.JobWire) (*reconcile.SignedOp, authz.OpBlob, bool)
 	return &reconcile.SignedOp{Blob: opBlob, Sig: []byte(env.SigArmored)}, ob, true
 }
 
-func (r *Runner) complete(ctx context.Context, jobID string) {
-	if err := r.source.CompleteJob(ctx, jobID); err != nil {
+func (r *Runner) complete(ctx context.Context, jobID string, o hub.JobOutcome) {
+	var err error
+	if os, ok := r.source.(OutcomeSource); ok {
+		err = os.CompleteJobWithOutcome(ctx, jobID, o) // R-1383: the hub hears HOW the job ended
+	} else {
+		err = r.source.CompleteJob(ctx, jobID)
+	}
+	if err != nil {
 		r.logger.Warn("signedjobs: clearing job failed (will retry next pass)", "job", jobID, "err", err)
 	}
 }

scripts/release-agent.sh

@@ -33,6 +33,9 @@
 # Env: GITEA_USER/GITEA_TOKEN (package write) — same credentials publish-agent.sh already takes.
 #      GITEA_BASE / GITEA_OWNER override the defaults.
 #      RELEASE_ALLOW_DIRTY=1 skips the clean-tree gate (for a rehearsal; never for a real release).
+#      GITEA_GIT_TOKEN — the tag push (`git push origin <tag>`) needs it since 2026-11-11: the address of a CC clone
+#      on DooPlex holds no credential, and git takes the token from the secret store through the helper
+#      felhom.eu/scripts/git-credential-sec.sh. Run the release as `sec run GITEA_GIT_TOKEN,... -- scripts/release-agent.sh <version>`.
 set -euo pipefail
 
 GITEA_BASE="${GITEA_BASE:-https://gitea.dooplex.hu}"

scripts/retention-policy.json

@@ -24,10 +24,10 @@
   "readers": [
     "scripts/check-published-versions.py — asserts the newest N versions plus every pinned version"
   ],
-  "recorded": "2026-10-27",
-  "recorded_by": "CC 2026-10-27: day0_agent 0.167.0 (operator vouch 2026-10-27, read from the hub: version + agent and bundle sha256 match the release); earlier 0.166.0 (vouch 2026-10-26 10:04:55 CET); D50 = A (operator 2026-10-25); earlier history: R-287 (2026-08-09), R-291 (2026-10-05)",
+  "recorded": "2026-11-14",
+  "recorded_by": "CC 2026-11-14: day0_agent 0.174.0 (operator vouch 2026-11-14 15:21:45 CET, read from the hub log and from the hub's System page: agent=0.174.0, bundle 7d87d058… = the release; the save took 428 ms); earlier 0.173.0 (CC 2026-11-10) (operator vouch 2026-11-10 16:29:16 CET, read from the hub log: agent=0.173.0, bundle 51ebc1f4… = the release; the save took 232 ms); earlier 0.172.1 (CC 2026-11-07) (operator vouch 2026-11-07 ~23:00 CET, read from the hub DB with its WAL: artifact_agent_version=0.172.1, bundle 3e7acaee… = the release); earlier 0.172.0 (CC 2026-11-03) (operator vouch 2026-11-03 10:18:10 CET, read from the hub log and DB: agent=0.172.0, bundle cc34b0bc… = the release); earlier 0.171.0 (CC 2026-10-30) (operator vouch 2026-10-30 10:47:36 CET, read from the hub log: agent=0.171.0, bundle 95f977f6… = the release; the save took 132 ms with hub 0.178.0, R-1346); earlier 0.170.0 (operator vouch 2026-10-29 18:57:48 CET, read from the hub log: agent=0.170.0, bundle 03f2ff64… = the release; the save took 38.3 s, R-1346); earlier 0.169.0 (operator vouch 2026-10-29 14:06:09 CET, read from the hub log: agent=0.169.0, bundle 3ca33f9c… = the release; his browser got a 504 but the save went through); earlier 0.168.0 (vouch 10:21:56 CET); 0.167.0 (vouch 2026-10-27); 0.166.0 (vouch 2026-10-26 10:04:55 CET); D50 = A (operator 2026-10-25); earlier history: R-287 (2026-08-09), R-291 (2026-10-05)",
   "pinned": {
     "hub_min_agent": "0.158.1",
-    "day0_agent": "0.167.0"
+    "day0_agent": "0.174.0"
   }
 }