Files
felhom.eu/REPORT-release-2a-2026-11-14.md
T

12 KiB
Raw Blame History

REPORT — release 2a: the box removes its old kernels, and refreshes its security profiles after a kernel change (2026-11-14, attended at one STOP)

1. Claims in the prompt that turned out wrong, named first

  1. „A night can be started on a test box by an existing path" — FALSE. The agent starts the night leg only after the night's OWN whole-guest backup. A manual backup does not start it, by design (R-899), and the debug pass runs with trigger debug, which the kernel step and the clean-up refuse, by design. So the real clean-up on Tester 1 was started by hand with the agent's own command line, not „through the night leg's own path". The step by the agent at night is not proven yet. Filed: R-1394.
  2. „Tester 1 — the first real clean-up by the program" needed something the prompt did not name. Tester 1 is ring 1 and has no root-owned ring-0 mark, and the release runs the clean-up on ring-0 boxes only (the prompt's own scenario 8). The program answered „not run" there, correctly. CC wanted to set the mark by hand; Claude Code's own safety check refused that, rightly. CC asked at the STOP; the operator said yes to the mark for the test. It was set for under a minute and is back as before (proven in the evidence).
  3. „The wrapper is reached by --plan only, so a new mode needs no new sudo line" — TRUE. Pinned by a test, and seen live: both new modes ran through the one existing line.
  4. „The result can ride the existing OS-update report" — TRUE in part. A clean-up that removed kernels reaches the hub as a kernel-layer report; the hub stores it and shows „cleaned
  5. „One stale profile can be made on demo-hp without changing a package" — TRUE. One old compiled copy was put into the running kernel's cache folder; the time service then failed exactly as on the N100.
  6. The diagnosis' method needed a control it did not have. „Loaded hash against a fresh compile" is wrong for one profile always (lxc-containers compiles differently every time) and for 17 of 119 after a fresh load (files with several profiles). The built mode does not use it; it reads the cached FILE changing.
  7. The prompt's „about 4 hours after the start" for the vouch: the STOP was reached after about 2 hours; the answer came about 2.5 hours later. No part waited on it except the Tester 1 proof and the vouch.

2. Five lines

  1. The clean-up on Tester 1: 3 kernels → 2 (removed proxmox-kernel-7.0.14-20-pve-signed), 1.12 GB gained, 8.5 s; the default, the pinned file, both generators and the state file unchanged; no left-over; the second call removed nothing. The reboot: back on 7.0.14-22-pve, all 10 containers healthy after 239 s, a clean stop for the crash guard.
  2. What the N100 removes tonight: 6 of its 8 kernels (7.0.2-6, 7.0.14-20, -22, -23, -24, -25); it keeps 7.0.14-26 and -28. About 6 GB.
  3. Part B — BUILT. The spike on demo-hp was clean (0.5 s, both guests and 32 containers untouched, no new kernel refusal in 10 minutes with a positive control), so the mode was built, tested and released.
  4. The N100's clock: offset to DooPlex −32.55 s before, +0.18 s after and 2.5 hours later; the time service is active. The product repaired it when the N100 took the new agent's root files; the clock stepped +32.7 s once.
  5. Agent 0.174.0: vouched by the operator, 2026-11-14 15:21:45 CET (the hub's log and its System page).

3. The parts

Part Result
A — the kernel clean-up BUILT. Wrapper mode kernel-cleanup (layer kernel) and the agent's night step. 11 §5.11.1 is [FACT — built]
B — security profiles SPIKED, then BUILT. Wrapper mode profile-refresh (layer host) and the agent's call at start and after a bundle install. 11 §5.12 (new)
C — Go 1.27.4 done; the running binary reads go1.27.4 on the three boxes. R-1391 closed
D — the release 0.174.0 released, on Tester 1, demo-hp and the N100 by the two signed jobs each, vouched
E — the one-year HTTPS setting skipped, nothing changed: the build day is 2026-11-14, before the ruled 2026-11-15

Versions, digests, CI by commit. felhom-agent code cc63dd5 — CI run id 2641 success (read before the release); release records 2706e35 — run id 2643 success. Agent sha256 6efbb97809135483bc446d71d1321f9e2f7fc3268fe9b20e7fe742a2eb13db92, bundle 7d87d0584dee79eee766f01c27b9e89273d39c89a14b0725a6e5564a55b3ec5f. No hub build, no hub deploy, no controller release. felhom.eu: documents only (the CI run of this push is quoted in the chat note).

Design decisions taken by CC in the build (each follows the design; the operator may reverse):

  • The clean-up asks „may this box?" from the root-owned ring-0 mark only; the agent does not gate on the hub's ring. (11 §5.4.1: the protection is what the wrapper refuses.)
  • „Nothing to remove" and „not run" send no hub report — one a night would bury the kernel step's own line.
  • The wrapper adds its own 20-hour window after a kernel step (the agent also never calls it in a step's night).
  • Profile refresh: once per kernel (not at every agent start), and the narrow restart rule (a failed service whose own program's cached profile was stale). Both with the other option and its cost in the spike's README.

Tests, each with its red-proof (audits/release-2a-2026-11-14/A, B):

Suite Cases Mutations seen red
wrapper, clean-up (KernelCleanup) 31: one per scenario 1–11, one per kept-kernel kind for scenario 6, the window, the dry run, the after-purge proof, the lock, two image packages, the sudo line 14, incl. the four the prompt names: K1 removed → the kept kernel reaches the purge line; the simulation check removed → the purge runs although apt would do more; the phase check removed → a purge during judging; the ring-0 check removed → a box without the mark purges
wrapper, profile refresh (ProfileRefresh) 11 10
agent, clean-up 5 (the night's order, once, no names; no report for nothing / not run; refused and failed reach the hub; when it may not run; the dry run) 3
agent, profile refresh 4 3 (one did not compile at first and proved nothing; done again)

Two existing agent tests list the night's calls exactly and gained the clean-up call; three kernel-step tests now count only the step's own calls. Wrapper suite 191 green; the agent's go test ./..., go vet and the gates green.

Rows: 250 before, 249 after. Opened R-1394 (a night cannot be started by day), R-1395 (the controller and a clock step). Closed R-1388, R-1301, R-1391. Narrowed R-1368 (the kernel part: the night is left, then the ring-1 ruling). Dated lines: R-1383, R-1384. New decision in STATUS.md: D82 (may a household's box clean up its kernels).

Machines changed, and how each was left.

  • demo-hp: the spike planted one stale profile and removed it (cache = the fresh compile, loaded hashes back as found by one cache load, the spike's files gone, chrony active — profile-refresh-spike-2026-11-14/evidence/07). Three test states for the refusal dry runs, each put back; checksums identical after (D/10). Agent 0.174.0 + bundle.
  • Tester 1: agent 0.174.0 + bundle; one kernel removed; the ring-0 mark set for the test and put back (false, root:root 0644, no saved copy left); one restart. The hub's System page showed a partial line for it after the restart (see Observations).
  • The N100: agent 0.174.0 + bundle by the normal signed jobs; reads and two dry runs by CC; no hand step. The product restarted its time service and the clock stepped; the controller recreated three apps once.
  • DooPlex: one release build, the signing tool built in the scratchpad. Nothing else. ep0, the hub: no change.
  • Scratch guest 9202, bench VMs: not used.

Architecture documents read: 11-os-updates.md §5.4.1, §5.4.2, §5.5 (the night's order, through the leg's code), §5.9, §5.11, §5.11.1; 09-update-architecture.md §3 decisions 286, 305, 306; runbooks/config-bundle.md; audits/kernel-cleanup-spike-2026-11-10/ (README and the rule script); audits/n100-clock-2026-11-11/; the felhom-build-deploy skill. Not read in full: 03-host-agent.md (the sudo rules were read in the sudoers file itself), 05-hub-architecture.md (the vouch flow was unchanged), 09 decisions 164, 171, 172, 195 (the kernel lane's own rulings; its behaviour was read in 11 §5.11 and in source).

Documents changed: 11 §5.11.1 ([FACT — built], with the proofs and the „not proven" lines) and §5.12 (new); 16 §14 (the kernel line: its limit and its cleaner); runbooks/disk-full.md (the hand step is no longer the only way); the capability map row G18; lived-in/n100.md; the register; STATUS.md. System poster: no listed fact changed; none of the three acts was needed. Instruction files: none edited. unproven.py --summary: 32 walked, 15 partial, 4 built, 4 missing, 23 not walked — no number moved.

4. What is not proven

  • The clean-up by the agent at night — the N100's and demo-hp's night of 2026-11-15 (STATUS.md says what to read).
  • A box whose running kernel is not its newest, and a self-revert into the kept kernel chosen by the product (tests only).
  • Refusal K1 on a box (it guards a wrong computation; tests only). Scenarios 7, 9, 10 and 11 on a box (tests only).
  • A restart by the profile-refresh rule of any program but chrony; the refresh after a real fall-back of the kernel lane.
  • A box that boots with proxmox-boot-tool; a BYO host on a box (tests only: R12).
  • That no audit line that mattered was lost in the seconds after a refresh (the kernel suppresses them; measured).

5. Observations

  • A night cannot be started by day on a test box. FILED: R-1394.
  • A clock step makes the controller read a new boot and recreate its drive-backed apps. FILED: R-1395.
  • The hub shows no detail of a clean-up and nothing of a profile refresh. FILED: a dated line on R-1383.
  • lxc-containers never compiles to the same binary twice, so every refresh „rewrites" it. NOT-A-FINDING: AppArmor's behaviour; written into 11 §5.12; it names no program, so the restart rule never matches it.
  • After a profile reload the kernel suppresses audit lines for some seconds. NOT-A-FINDING: the kernel's rate limit; it is the reason for „once per kernel", written into 11 §5.12.
  • In the customer guest AppArmor is installed and inactive, and the containers read „unconfined" from inside. NOT-A-FINDING for this release: no cache exists there, so the stale-profile fault cannot. Not checked: whether the register already holds a row on container confinement inside the guest.
  • The old wrapper refuses the new mode at the first start of the new agent; the refresh then runs after the bundle. NOT-A-FINDING: built for (the agent asks again after a bundle install); seen on three boxes.
  • The release script prints the repository's address in its output. NOT-A-FINDING: the address holds no login since 2026-11-11; the output was not committed.
  • The signing tool prints the whole signed job when its output is not filtered; one evidence file held two such jobs for a minute. NOT-A-FINDING: a signed job is not a secret and is single-use; the lines were removed before any commit and the script now filters.
  • Tester 1's hub line read „partial … crash guard not installed" after its restart. See the chat note for the last read.
  • The hub still lists host Tester4-46836f (this morning's test VM, destroyed). NOT-A-FINDING: the operator's delete, already in STATUS.md.