Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS
7.6 KiB
REPORT — the Space list (pilot of the gap review), 2026-11-10
Task: every place that holds data on a box, eight questions each. Read-only. Session 07:15–10:00 UTC.
Everything is in documentation/audits/space-list-2026-11-10/ (README.md first).
1. Claims in the prompt or the plan that were wrong
- The helper images. The prompt names
alpine:3.20,curlandhello-worldas the product's helper images. The product pullsalpinewith no tag (seven call sites).alpine:3.20andcurlcame from test tools.hello-worldcomes from the template bake. - Not in the prompt, but found wrong on the way: closed row R-521 says the guest root disk has a fill watcher. It has none since the two guest volumes became one (measured on three boxes). Capability-map row C13 says "per filesystem"; this one filesystem is the exception. Filed in R-1371. The map's status is not changed.
- My own number: an early draft of the table said
sudowrote "92 % session lines". R-1367 says 61 %. Both are right:sudowrites three lines for each call; R-1367 counts two of them. The table and the row say so now. - The register count in the prompt (225) was right. The section headings in the register were not: they added up to 237. I set all eleven to the real numbers and moved two rows that sat in the wrong section (R-1268, R-1312 → Backup & restore). No row was changed, closed or re-ranked by that.
2. Five lines
- Places: 45 customer places (18 host, 17 guest, 10 data drive) · 7 test-only · 6 off-box, from documents only.
- No document named: 8 of the 45. 16 more are named only in passing.
- Marks, 360 answers: A 294 · D 0 · K 19 · N 47. The 47 N are 5 classes = 5 new rows.
- Worst N: nothing removes old kernels from the host. A ring-0 box holds 8 (8.15 GB) and gets about one a week.
- The operator decides four things: D67 (remove old kernels automatically), D68 (who gets the host-disk critical mail), D69 (write a "disk space" section), D70 (run the next list the same way).
3. The method's verdict
It worked. It found real gaps nobody had asked about, in about 1.5 hours of work plus a 1.5-hour wait for the
second reading. New and measured: kernels and old agent binaries are never removed; the guest root disk has no
watcher; a household gets the host-disk mail; off-site restore copies have no age limit.
The weak parts: question 7 ("is there a runbook") gave the same answer 21 times — one gap. Some places are too
small to matter. No D answer appeared. The off-box places are candidates only: a helper filled them and the main
session did not read each pointer. The Prometheus rules in homelab-manifests were not read.
For "Endings": ask the runbook question once for each disk; drop tiny places in the first hour; a helper's answer
enters the table only after the main session read its pointer.
4. Machines, and what was changed on them
Read: demo-hp host, guests 9201 and 9202 · Tester 1 host and guest 9201 · demo-felhom (N100) host and guest 9201. Two readings: 07:38–07:39 and 09:44–09:45 UTC. Before each one I checked that no backup, restore or update ran. Nothing was changed on any machine. Provisioned nothing; nothing to tear down (machine, host, hub). Not read, by the brief: ep0, DooPlex's disks, the off-site box, the hub database. Evidence was copied to the audit folder at the end of each reading. No secret is in it (scanned); the N100 evidence holds folder sizes only, no file names.
5. Commits, CI, register
- Commit and CI: see the last section of this file.
- Register: 225 before · 230 after · 5 opened · 0 closed. New: R-1368 (host places that only grow), R-1369 (no
written way out of a full disk), R-1370 (no space budget for the host disk), R-1371 (no space alarm; the host-disk
mail), R-1372 (small guest places). Every N answer belongs to exactly one of them;
MATRIX.mdlists the members. - A dated line, no other change, on R-548, R-1366, R-1367, R-995, R-1267, R-1268, R-1087, R-1294, R-540. No rank change is proposed for R-1367: the journal stops at its own cap.
STATUS.md: new "Updated" paragraph (the old one is indocumentation/backlog/STATUS-HISTORY.md); D67–D70.documentation/lived-in/n100.md: one dated line with its rows.- Facts corrected in place (facts only; no rule, fence or ruling):
documentation/architecture/05-hub-architecture.md:615— before: "the volume grew from 1 GiB to 2 GiB"; after: adds "4 GiB since 2026-10-07". Evidence:manifests/hub.yaml:58(storage: 4Gi).documentation/architecture/11-os-updates.md:107— before: "json-file log caps. Nolive-restore.", citedbuild-golden.sh:138-144; after: "log caps (10 MB × 3 files),live-restore: true", cited:193-203. Evidence:felhom-agent/configs/build-golden.sh:195-201; the same document, line 429.documentation/backlog/OPEN-ITEMS.md— eleven section headings set to the real row counts (see 1.4).
- Capability map, poster, CHANGELOG: no change. The poster facts file: no fact in it changed.
python3 scripts/unproven.py --summary: 23 of 55 not walked. The number did not move.
6. Architecture documents read
07-backup-architecture.md · 09-update-architecture.md · 11-os-updates.md · 03-host-agent.md ·
08-alarm-ladder.md · 16-storage-physical-identity.md · 05-hub-architecture.md ·
02-controller-module-map.md · _recovery-inventory-2026-07-28.md · 00-capability-map.md (rows A19, C13, C15) ·
felhom-system-poster.facts.md. No architecture document covers disk space as a whole — the absence the task
tests is real: 8 customer places are named nowhere, and host_disk_* appears in no architecture document.
7. Observations outside the scope
- Tester 1 holds 3 kernels, the ring-0 boxes 8. NOT-A-FINDING: Tester 1 is ring 1 and gets fewer kernels.
- Tester 1's drive has four empty folders with no app. NOT-A-FINDING: 0 bytes; the guards refuse a write there.
- Tester 1's host swap use rose from 186 MB to 513 MB between the two readings. NOT-A-FINDING: swap is a fixed volume and its use moves with memory load; two readings cannot say more.
runbooks/ep0-datastore-copy.md:85says the removal job is "NOT INSTALLED"; closed row R-901 says it was installed on 2026-10-09. NOT-A-FINDING for this run (a DooPlex procedure; not edited); listed in the README under "for a later session".- The comment at
felhom-controller/controller/cmd/controller/main.go:2694-2707is stale. FILED: R-1371. - Register section headings were stale; two rows sat in the wrong section. Fixed in this run without a row (small).
8. Commit and CI
- Work commit:
47e372a5f45f60fce08a0ce781111bb0b30dc4d2onmain(basea6705ceb04). Pre-push gates: OK. - CI: workflow
gates, run id 2604, run number #1410, for exactly this commit hash: Success (running at 09:55:54 UTC, success at 09:56:14 UTC). - How I read it: the public page
https://gitea.dooplex.hu/admin/felhom.eu/actions(the row that links tocommit/47e372a5…; its status attribute and tooltip). Not thejobsendpoint. Tried: thejobsendpoint with no credential → HTTP 401;sec list→ no Gitea name with Actions rights (GITEA_PKG_TOKENhas none). The admin credential is in a credentials file that thesecrule forbids me to read by hand. - Not read: the attempt number. The public page does not show it. The run finished two minutes after the push and nobody else was at the repo, so a re-run is unlikely — but that is a guess, not a reading.
- This file's own second commit changes only this section; its CI result is in the session's last chat message.