admin/felhom.eu
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

night 2026-09-24: findings doc, decisions 29-30, STATUS/CONTEXT, register 335->344 (7 closed), teardown evidence, floor 0.269.1

Browse Source 
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0159rPz1ZhFKsS53msqPYxtS
This commit is contained in:
authentik Default Admin
2026-09-24 14:55:06 +02:00
parent 4502af6bb1
commit 83e436f609
41 changed files with 5537 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
STATUS.md
+21 -23
View File
@@ -1,30 +1,28 @@
# STATUS — what works, what's broken, what's next
**Updated 2026-09-24 (morning session). The fleet is on controller 0.268.0. The undo works after a restore. A stopped app's page tells the truth. A box that is several updates behind now climbs one tested step per press.**
**Updated 2026-09-24 (night shift, run in daytime from 11:07). The demo boxes are on controller 0.269.1 — except demo-hp's customer box, which is damaged and needs you (first item under "What needs you").**
**Decisions I took on my own:** none. Two small changes to your wording, both below.
**Decisions I took on my own (you may reverse either).**
- **A second controller release tonight (0.269.1).** The first one (0.269.0) had a flaw I found in its own live test: when the catalog re-tested an image, the box wrote the new image into a running app's file before anyone pressed Update, so any restart would have changed the image with no backup and no undo. I fixed it and released again rather than send the flaw to the fleet. Now an installed app keeps its image until an Update moves it.
- **A crash loop is 6 restarts in 10 minutes, not the 10 you wrote.** Measured: Docker slows a steady crash loop to about one restart a minute, so 10 would never catch it. No healthy app in any test record restarted more than once while starting.
**What I did, and it worked.**
- **The demo boxes got 0.267.0**, as you said yes to. Then, after every live test passed, they got **0.268.0**. Both boxes arrived healthy about 15 seconds after each change.
- **The undo after a restore.** A restore used to rebuild an app's storage without the tag the undo looked for, so the undo copied nothing. Now the undo finds the storage by its name. A restore also puts the tag back. Tested on the scratch machine with an app restored by the OLD version, the way customer boxes are today: the failed update was undone, and data written before and after the restore came back.
- **A stopped app's page (your option A).** When no copy on the box can bring an app back with its files, the page now says so, says support is informed, and shows no restore button. You get an urgent alarm. Tested by repeating last night's exact case, in Hungarian and English.
- **Small fixes.** A stopped app no longer raises a second, false alarm. A removed app leaves no old files behind. The test bench now marks a memory test with no real load as "not proven", and it starts every run with an empty drive folder.
- **The update ladder.** One press now does one tested step, never a jump. Tested with RomM on the scratch machine: first press moved only the app, second press moved only the database engine. The data came back after each press. The page says how many steps remain.
**What I tested, and it worked.**
- **The second drive now brings an app with files back whole** (your ruling). Tested four times on the scratch machine, twice under a power cut or a nearly full disk. Every time: the account and the files came back, a file the household had edited later was kept, and nothing was deleted.
- **A stranded app can only be removed keeping its data** (your ruling). The box refuses to delete the data, in both languages.
- **The box stops a crash loop or a memory storm** (your ruling), tells the household and you, and Start gives one more try. A second stop within a day says support is informed.
- **Exact image fingerprints.** The box runs exactly the image the catalog tested, and says "update available" when a newer tested image exists for the same version name.
- **Automatic updates: measured, not built.** A test caller updated three apps (up to three steps each) and set a failing one aside in about 7 minutes. The build plan is written with the numbers.
- **A chaos hour, 12 rounds** (power cuts, killed controller, Docker restarts, a full disk, backups). Updates resumed or undid themselves correctly every time.
**Your wording, changed in two small ways.**
- English: I removed the word "please". The house rule for English copy forbids it.
- Hungarian: I kept your text exactly. It uses the formal "Ön" form; the rest of the screens use the informal form. That mismatch is already on the list.
**What broke.**
- **demo-hp's customer box (not caused by tonight's work).** This morning the host's automatic restore test copied a whole guest onto the same full disk pool. The pool filled, and the customer box's disks became read-only. I freed the pool with an agent restart (the product's own clean-up). I could not repair the box itself.
- **An install interrupted by a restart is lost without a word**, and a remove interrupted the same way is left half-done. Both recorded, not fixed.
- After a failed update is restored, the box can keep the failed version's health check, and a later undo then fails for no reason. Recorded.
- Not done: moving more apps in the catalog. It needs a throwaway test machine, and this session could not delete one afterwards.
**What I found (new, not fixed).**
- **The second drive cannot bring back an app with files, even though it holds everything.** Its restore button refuses such apps, and the file restore brings back only files, not the database. So for Nextcloud, Immich, Paperless and Calibre, only the off-site copy counts. This needs your decision.
- **A long crash loop never raises an alarm** if the app looks "up" for a moment between crashes. Seen on the scratch machine: 385 restarts, zero alarms.
- The stopped-app sentence says "do not remove the app", but the Remove button is still there. This needs your word.
- Smaller: a ladder step has no health check of its own yet; right after a restore, an update may be judged with the restored health check; one restore option ("database only") is described in the code but offered nowhere.
**Rows.** 16 opened, 7 closed. The list went from 335 to 344.
**Rows.** Seven closed, six opened. The list went from 336 to 335.
**What needs you — two questions.**
1. **The second drive and apps with files.** (a) Build a "files plus database" restore for the second drive (my recommendation — then the second drive is a real way back, as customers would expect). (b) Rule that the second drive is files-only for these apps. If you do nothing, the page keeps naming only the off-site copy for these apps, which is true.
2. **The Remove button on a stranded app.** (a) Hide it while support is informed (my recommendation — a removal destroys what support needs). (b) Keep it and soften the sentence. If you do nothing, both stay as they are.
**Not done.** No automatic updates yet (plan part 7). No customer box touched by hand. Nothing on Peti's machine, the off-site box or your own machine was touched, beyond the normal hub update. The scratch machine is back on the real catalog with its standing apps.
**What needs you.**
1. **Repair demo-hp's customer box:** stop it, check its two disks, start it. If you do nothing, it keeps running but cannot save anything: no backups, no logs, no updates, and it stays on the old controller.
2. **Decide how the restore test may use disk space** (for example: it must check free space first and never use the pool of the box it tests). If you do nothing, the next scheduled restore test can fill the pool again, on any box with a small disk.
3. **Optional:** the two decisions above. If you do nothing, they stay as built.