You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Marking slot B successful: the plan after #125's hold, and for the Pixel 8 Pro #252
@derekross on #125 you put marking slot B successful on hold for the 9a, until you decide "how marking works from here". As I read it, the concern was the laptop script around omamobile-bootctl, which makes devinfo and its parent disk writable for 30 s. A bare run would need a hand-typed blockdev --setrw, and you mentioned a boot-success marker in the device package as one option. I couldn't find a later decision, so I'm asking here, because the Pixel 8 Pro (#1) has the same problem now.
Why it matters on husky:
Nothing on husky marks slot B, so every plain boot uses one of ABL's three tries.
On mine, ABL read slot-retry-count:b: 2 after one boot since the last flash. So far I work around it by re-flashing, which resets the count, instead of rebooting more than once in a row. That keeps every test boot one or two tries from losing the install.
What I've checked on husky, read-only:
devinfo is sdd1, 8192 bytes, layout 3.15, the version omamobile-bootctl is verified against.
A full backup is saved off the phone, outside the repository, and its sha256 matches live storage.
omamobile-bootctl dump, built unchanged from packages/omamobile-device-cheetah/src/omamobile-bootctl.c, agrees with ABL: B active, not successful, retry 2; A successful.
mark-successful b without --i-understand reports one byte, 0x35: 04 -> 06 (B's successful bit). Lef's read-only check of the layout agrees with it too.
Nothing has been written.
Questions:
Is the 9a hold about the laptop script only, or about marking B in general until there's a packaged way to do it?
Have you settled on how marking should work? For example, the Pixel 7 Pro's omamobile-boot-success, which opens exactly devinfo for the one verified write inside the tool and sets it read-only again.
For husky: would you rather I wait for that and then send it in omamobile-device-husky? Or may I run the operation you approved on device-tegu: record Pixel 9a first light and staged descriptor #24 once on my phone now: the unchanged tool, a fresh backup checked against live storage right before the write, one bit at 0x35, and the before and after bytes recorded in husky's notes? After that I'd use push only, as the 9a does, and accept that a broken image means fastboot by the buttons.
Thanks for asking first, Brett, and for doing all the read-only groundwork before asking. The backup checked against live storage, the dump agreeing with ABL, and the dry run showing exactly one bit (0x35: 04 -> 06) are what I'd want to see before any decision. (The dry run is safe: without --i-understand, omamobile-bootctl opens devinfo read-only.)
Your questions:
1. What the hold covers. Marking in general, not only the laptop script. What I wrote into AGENTS on #125 was "don't mark B, and don't run omamobile-bootctl or anything that makes devinfo writable" until I've decided how it's done. The script was what made me stop, but the open question is the mechanism: who lifts the read-only flag, on which boot, and checked how.
2. Whether I've settled it. Not yet. I'm leaning towards a boot-success marker in each device package, since that's the one write outside the allowlist that CONTRIBUTING §3.2 already allows. But porting cheetah's isn't just copying a file, and three things need deciding first:
omamobile-boot-success doesn't open only devinfo. It runs blockdev --setrw on the whole bootloader LUN and then on the partition (lines 30-31) for the write. That's the same kind of window as the 9a's laptop script, so it needs the same review, whichever phone it goes on.
When it marks. On cheetah it marks every boot that reaches multi-user.target, including the first boot of an untested image. After that, ABL never falls back, so a broken push means fastboot by the buttons. On the 9a and the 8 Pro, where a fallback costs the install, that trade may be the right one. But it should be a decision, not a side effect.
The layout evidence for each phone. On husky we have the version, the flags, and one retry count that agrees with ABL on each of two phones (yours at 2, Lef's at 1). That's encouraging. Before writing anything, I'd like one more pair on the same phone, across a change ABL makes itself. You already have the first one (dump and getvar both at 2 after one boot). On the second boot after a flash, run the dump (expect 1), then reboot-to bootloader, read fastboot getvar slot-retry-count:b, and flash your known-good set from there without booting again. That stays inside the budget even if entering the bootloader costs a try, and it also answers Lef's open question of whether it does. (reboot-to bootloader reached fastboot directly in BRINGUP-M1 test 22. If it ever boots Linux instead, B has spent its last try: reach fastboot by the buttons as in BRINGUP-M1 test 23 (Power + Volume Up to restart, then Power + Volume Down) without letting it boot, and flash before anything else.) Post only the eight slot bytes (0x30-0x37) and the version, not a full dump.
3. For husky now: please wait, and don't run the one-off operation. The #24 approval was for the 9a, with its own layout evidence, and it's on hold there too. I'd rather make one decision for both phones than approve a second by-hand write while the first is paused. I'll come back on this issue with the mechanism, and then it goes into omamobile-device-husky with its own runbook.
Until then, your routine is the safe one:
Use bin/omamobile flash only with a known-good image set. It gives B three tries again.
Count boots since the last flash, and flash again before the boot that would use B's last try.
Count every on-purpose reset (a watchdog drill, a panic test, a dead-man reset) as a boot. Lef's notes show a session of drills is what used up the tries on his phone.
push doesn't refill the tries, so with B unmarked every push and reboot spends one.
One thing I'd like to get straight for the record: you wrote that Lef's phone fell back to stock on slot A, but his notes (#249) say slot A on that phone is GrapheneOS, and that its first boot wiped userdata. Which is on your slot A? It matters because the 9a's loss was stock's own first boot, and I'd like the 8 Pro's line in AGENTS to say exactly what was seen on which phone. I'll add that line (a fallback to slot A costs the install, and B is unmarked) once I have your answer.
Thanks again for holding off. It makes this much easier to get right.
Here's the decision I owed you on marking slot B for the Pixel 8 Pro and the Pixel 9a. @greenbrettmichael, @Somnius and @W4HO-ham, it covers all three of your phones.
Marking is a manual, attended step, never a unit that runs at every boot.
When: once per install of slot B's images, and only after that image set has booted cleanly and passed tools/phone-smoke. Not on the first boot of an untested image. Once B is marked, ABL never falls back, so a broken image after that means fastboot by the buttons. Marking only a proven image keeps that trade worth it on phones where a fallback costs the install.
How: with a device-aware omamobile-bootctl, the same tool the 7 Pro uses. It reads the devinfo version and refuses any layout we haven't verified on that phone, writes only the slot bytes, and reads them back. Its read-write window has to be the devinfo partition alone, not the whole bootloader LUN, the way omamobile-boot-success does it today. That change comes first, in its own reviewed PR, before anyone uses it on the 8 Pro or the 9a.
Layout evidence per phone, before the first write: the version, plus the eight slot bytes (0x30-0x37) and getvar slot-successful, slot-unbootable and slot-retry-count read together, across one change that ABL makes itself (a boot that spends a try). Then the tool's dry run, and a read-only devinfo backup taken before the write.
Who: the phone's owner runs it by hand from a runbook in the device's plan, after posting the dry run here. It never runs from a service, a hook or the installer.
The 7 Pro stays as it is. Its marker has run on every boot since October, and changing it needs its own reason and a phone test.
Until that tool PR lands, the routine above still applies: flash a known-good set before the boot that would use B's last try, and count every deliberate reset as a boot.
@Somnius, as I asked on #472: please disable and mask the local unit, and post here what it ran, what made devinfo writable, since when, the slot bytes before and after, and whether you have a devinfo backup from before the first write. Once the tool is in, your phone can be marked the reviewed way.
I'll write this into AGENTS and CONTRIBUTING with the tool's PR. Thanks for your patience on this one.
@derekross thank you for the decision, it's clear: marking is a manual, attended step, once per install after tools/phone-smoke, with a devinfo-only omamobile-bootctl, and never a unit at every boot. I'll wait for the tool PR and follow the routine until then: a known-good set flashed before the boot that would use B's last try, and every deliberate reset counted as a boot.
Sorry I didn't treat your hold on this issue as covering my phone. You posted it on 2026-11-01 and I put the unit on that evening. It is disabled and masked now. Here is what you asked for:
What it ran: husky-boot-success.service, a oneshot at multi-user.target, runs a bash script. It checks that /dev/sdd1 is devinfo and starts with the DEVI magic, reads the slot from /proc/bootconfig, and if that slot's successful bit (bit 1) is clear it writes that one byte (0x35 for B, 0x31 for A) with dd. Otherwise it prints "already successful" and exits.
What made devinfo writable: two blockdev --setrw calls, on the whole disk /dev/sdd and then on /dev/sdd1, put back to read-only by a trap when the script exits. So the window was the whole bootloader LUN, the same thing you pointed out in omamobile-boot-success.
Since when: the unit file is dated 2026-11-01 23:13 (Athens time, UTC+2). The journal on the phone starts on 2026-11-05 at 23:00 and shows 14 writes of 4 -> 6 since then, each on the first boot after a flash; earlier writes aren't in it.
Backup: yes, a read-only copy of devinfo from 2026-11-01 23:20, taken before the first write.
Slot bytes 0x30-0x37 before: 01 0a 00 00 02 04 00 00. After, and as they read today: 01 0a 00 00 02 06 00 00. Version bytes: 03 00 0f 00.
My slot A is GrapheneOS (2026102101). When slot B ran out of tries on 2026-11-01 it fell back to slot A, and GrapheneOS wiped userdata on that boot.
@derekross one more thing, and I'm sorry about it: I put that unit on only to make sure the phone would always boot into B after I lost the install to the fallback once. I should have asked you first instead of doing it on my own.
A correction to my last comment, and the ruling that follows from it.
I wrote that the marking tool's read-write window would be devinfo alone. That isn't possible on our kernel. A partition counts as read-only if either its own flag or its disk's flag is set (bdev_read_only()), and the storage guard sets the boot LUN's whole-disk flag. So writing devinfo means clearing the disk's flag too, and while it's clear, a raw write through the whole-disk node (/dev/sdd on the 8 Pro) can reach that LUN's other partitions (pinfo, blenv, mfg_data). Writing #483 found this, and an independent review confirmed it against the kernel source.
The ruling: guard the boot LUNs the way the main LUN is already guarded: the whole-disk node removed and each partition's own flag set. Then the tool can clear devinfo's flag alone, and the window really is devinfo alone. That's a change to the shared initramfs init, so it comes as its own PR and gets a boot on my Pixel 7 Pro before it merges. The marking tool (#483) lands only after it, rebased so its window is the one partition. Everything else in my last comment stands: attended, by hand, once per install of a smoke-tested image, never from a unit, with per-phone layout evidence first.
Until both are in, the routine stays the same, and nothing on the 8 Pro or the 9a marks slot B.
@derekross thanks for the correction, noted. Nothing on my 8 Pro marks slot B until the boot LUN guard change and #483 are in. I'm on the routine you described: a known-good set flashed before the boot that would use B's last try, and every deliberate reset counted as a boot. I also read slot-retry-count:b in fastboot before each flash.
It matches what my unit had to do. It ran blockdev --setrw on /dev/sdd and then on /dev/sdd1, and a trap put both back read-only on exit. On my phone the boot LUN is /dev/sdd and devinfo is sdd1. While the disk flag was clear, a raw write to /dev/sdd could have reached the other partitions on that LUN. The script writes one byte, at offset 0x35 for slot B or 0x31 for slot A, setting that slot's successful bit (on B, 0x04 to 0x06), and nothing else.
When the guard change has a PR, I can boot it on the 8 Pro and post what I see. This unit has a Kioxia UFS 4.0 part (@greenbrettmichael's 8 Pro is Samsung UFS 3.1), so it would add a second UFS vendor to your 7 Pro boot.
@Somnius thank you for the full account, and no need to apologise further. You told me exactly what I needed: the script, the blockdev --setrw on the whole LUN, when it started, the backup from before the first write, and only the eight bytes. Keep that 11-01 devinfo backup safe; it's the reference for your phone when the marking tool comes.
If you can, paste systemctl is-enabled husky-boot-success.service (it should say masked) and I'll consider this closed.
One practical note, since your slot B is still marked from the unit's last write (0x35 reads 06). Until your next flash, ABL won't fall back on B, so a push costs no try, but a broken pushed image means fastboot by the buttons (Power + Volume Up to restart, then Volume Down + Power). Please leave the bit as it is: clearing it by hand would be a devinfo write too. Your journal also answers something I didn't know yet: the unit only wrote when the bit was clear, and it wrote after every flash, so a flash clears B's mark on the 8 Pro as it does on the 9a. From your next flash on, the routine above applies. I'll put that in the 8 Pro line in AGENTS with the tool PR.
Thanks also for confirming GrapheneOS on slot A and what it did on 11-01.
@derekross here is the output you asked for: systemctl is-enabled husky-boot-success.service prints masked (and systemctl is-active says inactive). I'll leave the successful bit as it is (0x35 still reads 06) and treat my next flash as the point where the routine starts. I'm keeping the 11-01 devinfo backup.
One correction to my earlier answer: I wrote that the backup was from 23:20. The file is dated 2026-11-01 23:12 (the folder name has 23:20 in it and I took that for the time). It is still from before the first write: it was made before the unit script was created at 23:13, and its slot bytes read 01 0a 00 00 02 04 00 00, the "before" values I posted.
@derekross it's masked: I pasted the output above (#252 (comment)), and this is the terminal from when I ran it on the phone:
The unit file was a real file under /etc/systemd/system, so I removed it first and then masked the name; that is the run in the picture. I'll leave the bit as it is.
@Gr-linkk@derekross Here is what I already had recorded on my oriole, in case it helps. I didn't run anything new; the dumps stay private.
Layout/version (my devinfo backup, 2026-11-08, LineageOS still on both slots): 8192 bytes, magic DEVI, version 3.15; slot bytes 0x30-0x37: 02 02 00 00 02 0e 00 00 (0x35 bit 1 set, B successful).
Bootloader's view (fastboot getvar all, same day, about an hour earlier; bootloader slider-17.0-15199431, MP1.0): current-slot: b; successful a/b: yes/yes; unbootable a/b: no/no; retry-count a/b: 2/2. Not read at exactly the same time as the backup.
Retry counts (Omarchy on slot B, no success mark): --set-active=b sets B's retry count to 3; each boot uses one; at 0 the bootloader fell back to slot A.
Not recorded: the values right after --set-active, and the slot bytes before and after a used retry.
@derekross a bit more evidence for the routine in this thread, on my 8 Pro. My earlier comment only said the count reads 3 after set_active b following a flash. I've now also done it without any flash: fastboot set_active b alone sets slot-retry-count:b back to 3, and it has done so every time I've tried, 9 times today (2026-11-09). Each time I entered fastboot with reboot-to bootloader from the running system and read the count there: it read what my own boot count predicted (2 after one boot, 1 after two, never lower), and after set_active b it read 3. In the three morning runs I also read slot-successful:b and slot-unbootable:b, both no before and after. So entering fastboot with reboot-to bootloader didn't cost a try, and no re-flash is needed to refill, which keeps the "flash before the boot that would use the last try" routine from rewriting five partitions every time.
The gates I use each time: PipeWire and WirePlumber stopped, the TDM_0_RX EP6 route off (an open sound card hung reboot-to once), battery and temperature read, one fastboot device with product: husky, and I only boot if the count reads 3. I never ran anything of my own that writes devinfo, and nothing marks B (set_active is the bootloader's own change).
One thing it showed that is worth knowing for the same routine: a systemctl poweroff on the 8 Pro restarts the phone instead of staying off, so it spends a try like any boot. I put the details in a separate issue: #578
This is one phone (the Kioxia UFS 4.0 one), so I can't say the same holds on the 9a or the other 8 Pros. @greenbrettmichael, @W4HO-ham, if you've seen set_active b alone behave differently, I'd like to know.
@Somnius I haven't tried set_active b on its own yet. Every refill on my phone has been a bin/omamobile flash, which ends with --set-active=b, and the count read 3 after it, as on yours. So nothing I've seen disagrees, but I can't confirm it on a second unit either.
@derekross, sorry for leaving your 11-01 question open while I was away. My slot A is stock Android 17 (CP3A.261005.005, Google's full OTA sideloaded to both slots on 10-21 and set active on A, in device/husky/NOTES.md), not GrapheneOS, and nothing of ours has written slot A since. Slot B has never fallen back since Omarchy went onto userdata, so I can't say whether stock's first boot wipes it. The fallbacks I did see, in BRINGUP-M1 on 10-21 and 10-22 after slot B's three tries, came before userdata held anything of ours (the init said no userdata partition then), so they don't answer it either. For the 8 Pro line in AGENTS: a wipe seen with GrapheneOS on slot A (Somnius's phone), not yet known with stock.
Supplemental Pixel 6a retry-refill observation, November 13:
On this unlocked Bluejay, with the actual device/product and B slot state checked in attended Fastboot, fastboot --set-active=b alone changed B's retry count 2 → 3, with no image flash. B remained active, bootable and unsuccessful; A's flags were unchanged. One subsequent normal B boot consumed one try and reported 2.
Paired full devinfo readbacks showed only the expected B-record changes: the retry refill and clearing of fastboot_ok (byte offset 0x35, 0x0c → 0x04), consistent with the bootloader's DevInfoInitSlot behavior. A's record and all other bytes were unchanged. No Linux/raw metadata writer or success-marking operation was used.
This is evidence for a bootloader-owned refill on one Pixel 6a, not permanent slot-success support or a promise about other devices. Each use still needs actual device/state identification and a post-selection retry read. AUTO_WRITE=no, masked automatic boot-success and the Bluejay success-marking hold remain in force. No raw metadata or device identifiers are attached.
Thank you for the 6a result, and for keeping to the hold while you read it. It matches what we have from the 8 Pro (Somnius, 11-09) and the Pixel 7 (device/panther/NOTES.md): --set-active=b on its own refills B's tries, the bootloader clears fastboot_ok (bit 3 of B's flags byte, which is how omamobile-bootctl decodes it too), and the next boot sets it again and spends one try. That's three phone families now where the refill needs no flash. Could you add it to device/bluejay/NOTES.md in a small pull request, with only the eight slot bytes before and after (0x30-0x37) and the getvar values next to them?
A status update for everyone waiting on this: the next step is still mine. The boot LUN guard (#484) has to land before the marking tool (#483) can have the one-partition window I ruled for on 11-07. #484 has fallen behind main and needs a rebase. Because it changes the storage guard for every phone, it boots on my Pixel 7 Pro before it merges. I'll post here when that's done. Until then nothing changes: no phone marks B by hand, and the routine stays the one from 11-01 and 11-07: a known-good image set flashed before the boot that would use B's last try, and every deliberate reset counted as a boot. (On the 9a, ask me before any flash, as AGENTS.md says.)
@derekross, one more data point for the refill, from a Pixel 8 Pro where slot B had run out (2026-11-13).
I got the phone back from a hang with a forced restart and went to the bootloader before letting it boot. fastboot getvar there said current-slot: a, slot-retry-count:b: 0, slot-retry-count:a: 1 and slot-successful:b: no. So with B at zero the current slot had already switched to A. Slot A on this phone is GrapheneOS, and a fallback to A is the case you described on #724 as wiping userdata, so I did not boot it.
fastboot -s <serial> set_active b on its own, with nothing flashed, was enough: afterwards current-slot: b, slot-retry-count:b: 3, slot-retry-count:a: 1 and slot-unbootable:b: no. I booted B once from there and it came up normally. I didn't read the count after that boot, so I can't say what it spent.
So the refill you already have from the 8 Pro (#252, 11-09) also works from zero with the current slot already on A, if you catch the phone in fastboot before its next boot. One phone, and I don't know how many boots spent the tries before I looked; I didn't read the flags byte or fastboot_ok.
The requested Bluejay note is in #820. It places the immediately paired Fastboot getvars beside the eight slot bytes, with the readback timing made explicit: the complete Linux dumps are the November 12 recovery capture and November 13 after selection plus one B boot, not immediate before/after-selection dumps. There is no dump at intermediate B3; that count is the saved Fastboot observation.
The complete retained files differ only at 0x35 (0c to 04), with A and every other byte unchanged. This is documentation of the saved run, not a new refill or a change to the known-good-flash routine. Automatic/manual success marking remains held, and no phone operation was performed for the PR.
A data point from a Pixel 7 (panther), for when #483 reaches this phone. It's not a request to change the panther package. I know marking is on hold until #484 and #483 land. This was a one-off on my own phone.
Why: a plain reboot left my unit in fastboot three times in the last two weeks. Each time ABL had slot-retry-count:b: 1 and slot-successful:b: no before I ran --set-active=b.
What I ran on 2026-11-17:
devinfo is sdd1, 8192 bytes, layout 3.15. I took a full backup first; it's off the phone and outside the repository.
omamobile-bootctl, built with gcc on the phone from today's main (packages/omamobile-device-cheetah/src/omamobile-bootctl.c, unchanged).
omamobile-boot-success and its unit, both from the same package, also unchanged. That's the current script, which opens sdd and sdd1 for the write, not the single-partition window you ruled for.
The dry run planned one byte, 0x35: 0c -> 0e (B's successful bit). The tool wrote it and read the partition back. sdd and sdd1 were read-only again afterwards.
After three reboots in a row: each boot reached Linux with no fastboot. Read on the phone after each boot: B successful: yes, retry-count: 2, with no try spent. Before the write the count dropped one per boot.
Not checked:
Anything after an OTA or a flash.
A second panther unit.
nesmeck's unit, which is on 3.19; the tool refuses that layout anyway.
Three reboots of slot B today (an update reboot, one after installing #895, one for #866), and the third came up on slot A. Slot A is GrapheneOS, it started on its welcome screen, and Omarchy's initramfs then couldn't mount userdata (no ext4 on it any more), so it's gone again. I didn't read slot-retry-count:b before the last reboot; that part is on me. In fastboot afterwards it read current-slot: a, slot-retry-count:a: 1, slot-retry-count:b: 0, slot-successful:b: no, slot-unbootable:b: no. I backed up the calibration partitions read-only from the initramfs and reinstalled with install-phone; nothing on the phone writes devinfo, and my unit stays masked.
I know your ruling here: marking is by hand, attended, once per smoke-tested install, never a unit, and nothing marks B until #484 and then #483. One question on that, nothing more: for a phone that's in development use, a broken image costing me a trip to fastboot by the buttons is a much smaller loss than the fallback wiping the whole install, so would you consider making slot B stick while a device is in that mode? If the answer stays the same, that's fine, and the offer to boot #484 on my 8 Pro (Kioxia UFS 4.0) still stands.
Something that needs no write at all, if you'd take it: a check on husky before a reboot (in omamobile-keys' power menu, omarchy-system-reboot, or the update sheet's "Restart now") that reads B's retry count read-only and warns, or refuses, when B is unmarked and down to its last try. That would have stopped me today. I can write that one and test it here.
For my own part I now keep a ledger of every boot of slot B and read the count in fastboot before any reboot I ask for.
@derekross I'm sorry, but I won't go through the counting again. Yesterday cost me the install and an evening of restoring it, and it was the second time. While this phone is in development I'm putting my local fix back and using it until the phone is far enough along to be called out of development. Then I'll put it back to whatever you decide here, with your commands and instructions. I hope you understand.
So that nothing I said here is out of date, exactly what it does on my phone only:
a oneshot unit, husky-boot-success.service, runs once per boot, 150 s after multi-user.target;
it checks first that it's a Pixel 8 Pro booted from slot B, the magic and version are DEVI 3.15, slot B is active and not unbootable, the boot is at least 120 s old with no failed units (that's all it checks, it doesn't run phone-smoke), and the bit isn't already set; it reads the byte back afterwards and puts the old value back on a mismatch;
Before the first write I took a read-only copy of devinfo, and the copy after must differ in byte 0x35 only, apart from what the bootloader itself updates at each boot. fastboot set_active b clears the bit again, so going back is one fastboot command. Nothing else of mine writes devinfo.
Following your earlier request, I reproduced the public Oriole build from scratch on a separate Debian machine (no phone involved). Both repositories were pinned to the current public heads: derekross/omarchy-mobilemain @ 17f9a3cd979c280431b2253e5af377e1b2377d55 and derekross/linux-omarchy-mobileomamobile @ 26867fc4e826a03d1095341af50b6165d4ea5164. The official Docker procedure (kernel, modules, initramfs, images) succeeded, and the three generated images are byte-identical to the archived reference set.
The later Fastboot inspection was strictly read-only; the historical Z22 installation is documented separately in our logs. That Fastboot session was done with the phone already in the bootloader by hand. Only fastboot getvar commands were used — no flash, no set_active, no reboot, no devinfo write.
What we confirmed earlier in our own logs:set_active=b was executed during the Z22 installation (Setting current slot to 'b' OKAY), and the before/after getvar pair showed slot-successful:b and slot-fastboot-ok:b going from yes to no while slot-retry-count:b stayed at 3. This matches DevInfoInitSlot() in gs-common.
Current read-only Fastboot values:current-slot: b; slot-successful:a=yes, slot-successful:b=no; slot-unbootable:a=no, slot-unbootable:b=no; slot-retry-count:a=2, slot-retry-count:b=1; bootloader slider-16.4-14097577; product: oriole, hw-revision: MP1.0. Slot A's currently reported Fastboot metadata matches the historical backup.
One thing I cannot reconcile from the available records: three reboots are documented in our session reports (02:38, 04:18 and about 05:45 UTC on 19 Nov), but slot-retry-count:b went from 3 (after set_active) to 1 — only two decrements. There is no intermediate getvar reading in our logs, and no further set_active is recorded. A community report from another device (Somnius on the Pixel 8 Pro, issue #252) says that entering fastboot with reboot-to bootloader does not spend a try; this is community testimony on a different phone, not a confirmed rule on Oriole. It made me wonder whether one of those reboots did not go through ABL's slot selection, but I have no evidence for that and I am not assuming it.
Before considering any further hardware step, I would like to understand the rule rather than test it with the last remaining try. My questions:
On Oriole/gs101, what exact event makes ABL decrement retry_count — is it every boot of the active slot while successful is clear, or a narrower condition (stage reached, AVB result, shutdown path)?
When the count reaches zero, what does ABL do exactly on Pixel 6, and is the switch to slot A immediate or effective only at the next boot?
Besides retry_count == 0 with successful == no, are there other conditions that make ABL choose slot A (for example unbootable, fastboot_ok, or a failed AVB verification)?
How should fastboot getvar current-slot be read against the active bit in devinfo? Is it always the same value as getActiveBootSlot(), and can it diverge?
Is there an Oriole recovery routine documented by the project that preserves both userdata and slot A — in particular, does fastboot --set-active=b alone refilling B's tries without a flash, as reported for other phones, also hold on Oriole?
I am not asking to repeat the historical devinfo write, and I am not claiming that Oriole is qualified for #483. I will not run any further boot or write until this is clear. A pointer to the relevant ABL documentation or to an existing Oriole note would be enough; if there is no public source, a short answer from you would help.
@Ruraam since you cite my comment: this is exactly what my logs have. It's all one Pixel 8 Pro (husky, zuma), not Oriole, so it can't answer your five questions. Those are for @derekross.
reboot-to bootloader: my 11-09 note says it didn't cost a retry, and on 11-18 22:55 the count was 2 before and 2 after. One row disagrees: 11-18 around 22:35, from the initramfs rescue shell, 2 then 1. So that claim of mine is only partly backed.
At 0: on 11-13 fastboot read current-slot: a. On 11-18 it fell back to A (GrapheneOS here) and userdata was wiped. On 11-01 it ran out of retries and GrapheneOS wiped userdata; no getvar logged that day.
So on husky, at 1 with successful no, I don't boot B again. I go to fastboot with the buttons, read, and refill first. Whether that holds on Oriole I don't know. @derekross's 11-13 comment covers the refill on other phones: #252 (comment)
I can't explain your three reboots against two decrements from anything here.
Activity
derekross commentedon Nov 1, 2026
Thanks for asking first, Brett, and for doing all the read-only groundwork before asking. The backup checked against live storage, the dump agreeing with ABL, and the dry run showing exactly one bit (
0x35: 04 -> 06) are what I'd want to see before any decision. (The dry run is safe: without--i-understand,omamobile-bootctlopensdevinforead-only.)Your questions:
1. What the hold covers. Marking in general, not only the laptop script. What I wrote into AGENTS on #125 was "don't mark B, and don't run
omamobile-bootctlor anything that makesdevinfowritable" until I've decided how it's done. The script was what made me stop, but the open question is the mechanism: who lifts the read-only flag, on which boot, and checked how.2. Whether I've settled it. Not yet. I'm leaning towards a boot-success marker in each device package, since that's the one write outside the allowlist that CONTRIBUTING §3.2 already allows. But porting cheetah's isn't just copying a file, and three things need deciding first:
omamobile-boot-successdoesn't open onlydevinfo. It runsblockdev --setrwon the whole bootloader LUN and then on the partition (lines 30-31) for the write. That's the same kind of window as the 9a's laptop script, so it needs the same review, whichever phone it goes on.multi-user.target, including the first boot of an untested image. After that, ABL never falls back, so a brokenpushmeans fastboot by the buttons. On the 9a and the 8 Pro, where a fallback costs the install, that trade may be the right one. But it should be a decision, not a side effect.getvarboth at 2 after one boot). On the second boot after a flash, run the dump (expect 1), thenreboot-to bootloader, readfastboot getvar slot-retry-count:b, andflashyour known-good set from there without booting again. That stays inside the budget even if entering the bootloader costs a try, and it also answers Lef's open question of whether it does. (reboot-to bootloaderreached fastboot directly in BRINGUP-M1 test 22. If it ever boots Linux instead, B has spent its last try: reach fastboot by the buttons as in BRINGUP-M1 test 23 (Power + Volume Up to restart, then Power + Volume Down) without letting it boot, and flash before anything else.) Post only the eight slot bytes (0x30-0x37) and the version, not a full dump.3. For husky now: please wait, and don't run the one-off operation. The #24 approval was for the 9a, with its own layout evidence, and it's on hold there too. I'd rather make one decision for both phones than approve a second by-hand write while the first is paused. I'll come back on this issue with the mechanism, and then it goes into
omamobile-device-huskywith its own runbook.Until then, your routine is the safe one:
bin/omamobile flashonly with a known-good image set. It gives B three tries again.pushdoesn't refill the tries, so with B unmarked everypushand reboot spends one.One thing I'd like to get straight for the record: you wrote that Lef's phone fell back to stock on slot A, but his notes (#249) say slot A on that phone is GrapheneOS, and that its first boot wiped
userdata. Which is on your slot A? It matters because the 9a's loss was stock's own first boot, and I'd like the 8 Pro's line in AGENTS to say exactly what was seen on which phone. I'll add that line (a fallback to slot A costs the install, and B is unmarked) once I have your answer.Thanks again for holding off. It makes this much easier to get right.
device/husky: battery bring-up overlays, the dead-man tool and the re…
device-bluejay 0.1.0: the Pixel 6a's first-light package, nothing ena…
device-bluejay 0.1.0: the Pixel 6a's first-light package, nothing ena…
derekross commentedon Nov 7, 2026
Here's the decision I owed you on marking slot B for the Pixel 8 Pro and the Pixel 9a. @greenbrettmichael, @Somnius and @W4HO-ham, it covers all three of your phones.
Marking is a manual, attended step, never a unit that runs at every boot.
tools/phone-smoke. Not on the first boot of an untested image. Once B is marked, ABL never falls back, so a broken image after that means fastboot by the buttons. Marking only a proven image keeps that trade worth it on phones where a fallback costs the install.omamobile-bootctl, the same tool the 7 Pro uses. It reads the devinfo version and refuses any layout we haven't verified on that phone, writes only the slot bytes, and reads them back. Its read-write window has to be thedevinfopartition alone, not the whole bootloader LUN, the wayomamobile-boot-successdoes it today. That change comes first, in its own reviewed PR, before anyone uses it on the 8 Pro or the 9a.getvar slot-successful,slot-unbootableandslot-retry-countread together, across one change that ABL makes itself (a boot that spends a try). Then the tool's dry run, and a read-onlydevinfobackup taken before the write.Until that tool PR lands, the routine above still applies: flash a known-good set before the boot that would use B's last try, and count every deliberate reset as a boot.
@Somnius, as I asked on #472: please disable and mask the local unit, and post here what it ran, what made
devinfowritable, since when, the slot bytes before and after, and whether you have adevinfobackup from before the first write. Once the tool is in, your phone can be marked the reviewed way.I'll write this into AGENTS and CONTRIBUTING with the tool's PR. Thanks for your patience on this one.
Somnius commentedon Nov 7, 2026
@derekross thank you for the decision, it's clear: marking is a manual, attended step, once per install after
tools/phone-smoke, with adevinfo-onlyomamobile-bootctl, and never a unit at every boot. I'll wait for the tool PR and follow the routine until then: a known-good set flashed before the boot that would use B's last try, and every deliberate reset counted as a boot.Sorry I didn't treat your hold on this issue as covering my phone. You posted it on 2026-11-01 and I put the unit on that evening. It is disabled and masked now. Here is what you asked for:
husky-boot-success.service, a oneshot atmulti-user.target, runs a bash script. It checks that/dev/sdd1isdevinfoand starts with the DEVI magic, reads the slot from/proc/bootconfig, and if that slot's successful bit (bit 1) is clear it writes that one byte (0x35 for B, 0x31 for A) withdd. Otherwise it prints "already successful" and exits.devinfowritable: twoblockdev --setrwcalls, on the whole disk/dev/sddand then on/dev/sdd1, put back to read-only by a trap when the script exits. So the window was the whole bootloader LUN, the same thing you pointed out inomamobile-boot-success.4 -> 6since then, each on the first boot after a flash; earlier writes aren't in it.devinfofrom 2026-11-01 23:20, taken before the first write.01 0a 00 00 02 04 00 00. After, and as they read today:01 0a 00 00 02 06 00 00. Version bytes:03 00 0f 00.My slot A is GrapheneOS (2026102101). When slot B ran out of tries on 2026-11-01 it fell back to slot A, and GrapheneOS wiped
userdataon that boot.Somnius commentedon Nov 7, 2026
@derekross one more thing, and I'm sorry about it: I put that unit on only to make sure the phone would always boot into B after I lost the install to the fallback once. I should have asked you first instead of doing it on my own.
derekross commentedon Nov 7, 2026
A correction to my last comment, and the ruling that follows from it.
I wrote that the marking tool's read-write window would be
devinfoalone. That isn't possible on our kernel. A partition counts as read-only if either its own flag or its disk's flag is set (bdev_read_only()), and the storage guard sets the boot LUN's whole-disk flag. So writingdevinfomeans clearing the disk's flag too, and while it's clear, a raw write through the whole-disk node (/dev/sddon the 8 Pro) can reach that LUN's other partitions (pinfo,blenv,mfg_data). Writing #483 found this, and an independent review confirmed it against the kernel source.The ruling: guard the boot LUNs the way the main LUN is already guarded: the whole-disk node removed and each partition's own flag set. Then the tool can clear
devinfo's flag alone, and the window really isdevinfoalone. That's a change to the shared initramfs init, so it comes as its own PR and gets a boot on my Pixel 7 Pro before it merges. The marking tool (#483) lands only after it, rebased so its window is the one partition. Everything else in my last comment stands: attended, by hand, once per install of a smoke-tested image, never from a unit, with per-phone layout evidence first.Until both are in, the routine stays the same, and nothing on the 8 Pro or the 9a marks slot B.
device/husky: roadmap follows the ruling on derekross#252 for marking…
Somnius commentedon Nov 7, 2026
@derekross thanks for the correction, noted. Nothing on my 8 Pro marks slot B until the boot LUN guard change and #483 are in. I'm on the routine you described: a known-good set flashed before the boot that would use B's last try, and every deliberate reset counted as a boot. I also read
slot-retry-count:bin fastboot before each flash.It matches what my unit had to do. It ran
blockdev --setrwon/dev/sddand then on/dev/sdd1, and a trap put both back read-only on exit. On my phone the boot LUN is/dev/sddanddevinfoissdd1. While the disk flag was clear, a raw write to/dev/sddcould have reached the other partitions on that LUN. The script writes one byte, at offset 0x35 for slot B or 0x31 for slot A, setting that slot's successful bit (on B, 0x04 to 0x06), and nothing else.When the guard change has a PR, I can boot it on the 8 Pro and post what I see. This unit has a Kioxia UFS 4.0 part (@greenbrettmichael's 8 Pro is Samsung UFS 3.1), so it would add a second UFS vendor to your 7 Pro boot.
derekross commentedon Nov 7, 2026
@Somnius thank you for the full account, and no need to apologise further. You told me exactly what I needed: the script, the
blockdev --setrwon the whole LUN, when it started, the backup from before the first write, and only the eight bytes. Keep that 11-01devinfobackup safe; it's the reference for your phone when the marking tool comes.If you can, paste
systemctl is-enabled husky-boot-success.service(it should saymasked) and I'll consider this closed.One practical note, since your slot B is still marked from the unit's last write (0x35 reads
06). Until your nextflash, ABL won't fall back on B, so apushcosts no try, but a broken pushed image means fastboot by the buttons (Power + Volume Up to restart, then Volume Down + Power). Please leave the bit as it is: clearing it by hand would be adevinfowrite too. Your journal also answers something I didn't know yet: the unit only wrote when the bit was clear, and it wrote after every flash, so a flash clears B's mark on the 8 Pro as it does on the 9a. From your next flash on, the routine above applies. I'll put that in the 8 Pro line in AGENTS with the tool PR.Thanks also for confirming GrapheneOS on slot A and what it did on 11-01.
docs: husky roadmap and DEVICES row brought up to 2026-11-07 (#472)
Somnius commentedon Nov 7, 2026
@derekross here is the output you asked for:
systemctl is-enabled husky-boot-success.serviceprintsmasked(andsystemctl is-activesays inactive). I'll leave the successful bit as it is (0x35 still reads06) and treat my next flash as the point where the routine starts. I'm keeping the 11-01devinfobackup.One correction to my earlier answer: I wrote that the backup was from 23:20. The file is dated 2026-11-01 23:12 (the folder name has 23:20 in it and I took that for the time). It is still from before the first write: it was made before the unit script was created at 23:13, and its slot bytes read
01 0a 00 00 02 04 00 00, the "before" values I posted.initramfs, bin/omamobile: the marker check in push and flash, review …
Somnius commentedon Nov 7, 2026
@derekross it's masked: I pasted the output above (#252 (comment)), and this is the terminal from when I ran it on the phone:
The unit file was a real file under
/etc/systemd/system, so I removed it first and then masked the name; that is the run in the picture. I'll leave the bit as it is.2 remaining items
Svffer commentedon Nov 8, 2026
@Gr-linkk @derekross Here is what I already had recorded on my oriole, in case it helps. I didn't run anything new; the dumps stay private.
DEVI, version 3.15; slot bytes0x30-0x37:02 02 00 00 02 0e 00 00(0x35bit 1 set, B successful).fastboot getvar all, same day, about an hour earlier; bootloaderslider-17.0-15199431, MP1.0):current-slot: b; successful a/b: yes/yes; unbootable a/b: no/no; retry-count a/b: 2/2. Not read at exactly the same time as the backup.--set-active=bsets B's retry count to 3; each boot uses one; at 0 the bootloader fell back to slot A.--set-active, and the slot bytes before and after a used retry.Somnius commentedon Nov 9, 2026
@derekross a bit more evidence for the routine in this thread, on my 8 Pro. My earlier comment only said the count reads 3 after
set_active bfollowing a flash. I've now also done it without any flash:fastboot set_active balone setsslot-retry-count:bback to 3, and it has done so every time I've tried, 9 times today (2026-11-09). Each time I entered fastboot withreboot-to bootloaderfrom the running system and read the count there: it read what my own boot count predicted (2 after one boot, 1 after two, never lower), and afterset_active bit read 3. In the three morning runs I also readslot-successful:bandslot-unbootable:b, bothnobefore and after. So entering fastboot withreboot-to bootloaderdidn't cost a try, and no re-flash is needed to refill, which keeps the "flash before the boot that would use the last try" routine from rewriting five partitions every time.The gates I use each time: PipeWire and WirePlumber stopped, the TDM_0_RX EP6 route off (an open sound card hung
reboot-toonce), battery and temperature read, one fastboot device withproduct: husky, and I only boot if the count reads 3. I never ran anything of my own that writes devinfo, and nothing marks B (set_activeis the bootloader's own change).One thing it showed that is worth knowing for the same routine: a
systemctl poweroffon the 8 Pro restarts the phone instead of staying off, so it spends a try like any boot. I put the details in a separate issue: #578This is one phone (the Kioxia UFS 4.0 one), so I can't say the same holds on the 9a or the other 8 Pros. @greenbrettmichael, @W4HO-ham, if you've seen
set_active balone behave differently, I'd like to know.greenbrettmichael commentedon Nov 10, 2026
@Somnius I haven't tried
set_active bon its own yet. Every refill on my phone has been abin/omamobile flash, which ends with--set-active=b, and the count read 3 after it, as on yours. So nothing I've seen disagrees, but I can't confirm it on a second unit either.@derekross, sorry for leaving your 11-01 question open while I was away. My slot A is stock Android 17 (
CP3A.261005.005, Google's full OTA sideloaded to both slots on 10-21 and set active on A, indevice/husky/NOTES.md), not GrapheneOS, and nothing of ours has written slot A since. Slot B has never fallen back since Omarchy went ontouserdata, so I can't say whether stock's first boot wipes it. The fallbacks I did see, in BRINGUP-M1 on 10-21 and 10-22 after slot B's three tries, came beforeuserdataheld anything of ours (the init saidno userdata partitionthen), so they don't answer it either. For the 8 Pro line in AGENTS: a wipe seen with GrapheneOS on slot A (Somnius's phone), not yet known with stock.device-bluejay 0.1.0: the Pixel 6a's first-light package, nothing ena…
latentoperator commentedon Nov 13, 2026
Supplemental Pixel 6a retry-refill observation, November 13:
On this unlocked Bluejay, with the actual device/product and B slot state checked in attended Fastboot,
fastboot --set-active=balone changed B's retry count 2 → 3, with no image flash. B remained active, bootable and unsuccessful; A's flags were unchanged. One subsequent normal B boot consumed one try and reported 2.Paired full devinfo readbacks showed only the expected B-record changes: the retry refill and clearing of
fastboot_ok(byte offset0x35,0x0c → 0x04), consistent with the bootloader'sDevInfoInitSlotbehavior. A's record and all other bytes were unchanged. No Linux/raw metadata writer or success-marking operation was used.This is evidence for a bootloader-owned refill on one Pixel 6a, not permanent slot-success support or a promise about other devices. Each use still needs actual device/state identification and a post-selection retry read.
AUTO_WRITE=no, masked automatic boot-success and the Bluejay success-marking hold remain in force. No raw metadata or device identifiers are attached.derekross commentedon Nov 13, 2026
Thank you for the 6a result, and for keeping to the hold while you read it. It matches what we have from the 8 Pro (Somnius, 11-09) and the Pixel 7 (
device/panther/NOTES.md):--set-active=bon its own refills B's tries, the bootloader clearsfastboot_ok(bit 3 of B's flags byte, which is howomamobile-bootctldecodes it too), and the next boot sets it again and spends one try. That's three phone families now where the refill needs no flash. Could you add it todevice/bluejay/NOTES.mdin a small pull request, with only the eight slot bytes before and after (0x30-0x37) and thegetvarvalues next to them?A status update for everyone waiting on this: the next step is still mine. The boot LUN guard (#484) has to land before the marking tool (#483) can have the one-partition window I ruled for on 11-07. #484 has fallen behind main and needs a rebase. Because it changes the storage guard for every phone, it boots on my Pixel 7 Pro before it merges. I'll post here when that's done. Until then nothing changes: no phone marks B by hand, and the routine stays the one from 11-01 and 11-07: a known-good image set flashed before the boot that would use B's last try, and every deliberate reset counted as a boot. (On the 9a, ask me before any flash, as AGENTS.md says.)
Somnius commentedon Nov 13, 2026
@derekross, one more data point for the refill, from a Pixel 8 Pro where slot B had run out (2026-11-13).
I got the phone back from a hang with a forced restart and went to the bootloader before letting it boot.
fastboot getvarthere saidcurrent-slot: a,slot-retry-count:b: 0,slot-retry-count:a: 1andslot-successful:b: no. So with B at zero the current slot had already switched to A. Slot A on this phone is GrapheneOS, and a fallback to A is the case you described on #724 as wipinguserdata, so I did not boot it.fastboot -s <serial> set_active bon its own, with nothing flashed, was enough: afterwardscurrent-slot: b,slot-retry-count:b: 3,slot-retry-count:a: 1andslot-unbootable:b: no. I booted B once from there and it came up normally. I didn't read the count after that boot, so I can't say what it spent.So the refill you already have from the 8 Pro (#252, 11-09) also works from zero with the current slot already on A, if you catch the phone in fastboot before its next boot. One phone, and I don't know how many boots spent the tries before I looked; I didn't read the flags byte or
fastboot_ok.device/husky: battery bring-up overlays, the dead-man tool and the re…
docs: husky roadmap and DEVICES row brought up to 2026-11-07 (derekro…
latentoperator commentedon Nov 15, 2026
The requested Bluejay note is in #820. It places the immediately paired Fastboot getvars beside the eight slot bytes, with the readback timing made explicit: the complete Linux dumps are the November 12 recovery capture and November 13 after selection plus one B boot, not immediate before/after-selection dumps. There is no dump at intermediate B3; that count is the saved Fastboot observation.
The complete retained files differ only at 0x35 (
0cto04), with A and every other byte unchanged. This is documentation of the saved run, not a new refill or a change to the known-good-flash routine. Automatic/manual success marking remains held, and no phone operation was performed for the PR.angelrenefm commentedon Nov 17, 2026
A data point from a Pixel 7 (panther), for when #483 reaches this phone. It's not a request to change the panther package. I know marking is on hold until #484 and #483 land. This was a one-off on my own phone.
Why: a plain reboot left my unit in fastboot three times in the last two weeks. Each time ABL had
slot-retry-count:b: 1andslot-successful:b: nobefore I ran--set-active=b.What I ran on 2026-11-17:
sdd1, 8192 bytes, layout 3.15. I took a full backup first; it's off the phone and outside the repository.omamobile-bootctl, built with gcc on the phone from today'smain(packages/omamobile-device-cheetah/src/omamobile-bootctl.c, unchanged).omamobile-boot-successand its unit, both from the same package, also unchanged. That's the current script, which openssddandsdd1for the write, not the single-partition window you ruled for.0x35: 0c -> 0e(B's successful bit). The tool wrote it and read the partition back.sddandsdd1were read-only again afterwards.After three reboots in a row: each boot reached Linux with no fastboot. Read on the phone after each boot: B
successful: yes,retry-count: 2, with no try spent. Before the write the count dropped one per boot.Not checked:
flash.I used AI help (Claude) to work on this.
Somnius commentedon Nov 18, 2026
@derekross @greenbrettmichael I lost the install on my 8 Pro again today, the same way as on 11-01.
Three reboots of slot B today (an update reboot, one after installing #895, one for #866), and the third came up on slot A. Slot A is GrapheneOS, it started on its welcome screen, and Omarchy's initramfs then couldn't mount
userdata(no ext4 on it any more), so it's gone again. I didn't readslot-retry-count:bbefore the last reboot; that part is on me. In fastboot afterwards it readcurrent-slot: a,slot-retry-count:a: 1,slot-retry-count:b: 0,slot-successful:b: no,slot-unbootable:b: no. I backed up the calibration partitions read-only from the initramfs and reinstalled withinstall-phone; nothing on the phone writesdevinfo, and my unit stays masked.I know your ruling here: marking is by hand, attended, once per smoke-tested install, never a unit, and nothing marks B until #484 and then #483. One question on that, nothing more: for a phone that's in development use, a broken image costing me a trip to fastboot by the buttons is a much smaller loss than the fallback wiping the whole install, so would you consider making slot B stick while a device is in that mode? If the answer stays the same, that's fine, and the offer to boot #484 on my 8 Pro (Kioxia UFS 4.0) still stands.
Something that needs no write at all, if you'd take it: a check on husky before a reboot (in
omamobile-keys' power menu,omarchy-system-reboot, or the update sheet's "Restart now") that reads B's retry count read-only and warns, or refuses, when B is unmarked and down to its last try. That would have stopped me today. I can write that one and test it here.For my own part I now keep a ledger of every boot of slot B and read the count in fastboot before any reboot I ask for.
Somnius commentedon Nov 19, 2026
@derekross I'm sorry, but I won't go through the counting again. Yesterday cost me the install and an evening of restoring it, and it was the second time. While this phone is in development I'm putting my local fix back and using it until the phone is far enough along to be called out of development. Then I'll put it back to whatever you decide here, with your commands and instructions. I hope you understand.
So that nothing I said here is out of date, exactly what it does on my phone only:
husky-boot-success.service, runs once per boot, 150 s aftermulti-user.target;devinfo(/dev/sdd1): slot B's flags at offset 0x35, old value OR 0x02 (successful), so04becomes06. That's the same bitomamobile-bootctlin omamobile-device-husky 0.1.19, omamobile-device-tegu 0.2.22: omamobile-bootctl marks slot B by hand (#252) #483 sets (layout DEVI 3.15, slots at 0x30, flag 0x02), but only that one byte, not the header like omamobile-device-husky 0.1.19, omamobile-device-tegu 0.2.22: omamobile-bootctl marks slot B by hand (#252) #483;blockdev --setrwon/dev/sddand/dev/sdd1and sets both read-only again on exit, so for that moment the whole boot LUN is writable. If it's killed or the power goes in that moment,/dev/sddstays writable until the next boot. I know that's the window you objected to, and that initramfs, device-cheetah, device-raven, device-lynx (versions at merge): guard the boot LUNs as the main LUN (#252) #484 is the proper fix.Before the first write I took a read-only copy of
devinfo, and the copy after must differ in byte 0x35 only, apart from what the bootloader itself updates at each boot.fastboot set_active bclears the bit again, so going back is one fastboot command. Nothing else of mine writesdevinfo.Ruraam commentedon Nov 19, 2026
Hi @derekross,
Following your earlier request, I reproduced the public Oriole build from scratch on a separate Debian machine (no phone involved). Both repositories were pinned to the current public heads:
derekross/omarchy-mobilemain@17f9a3cd979c280431b2253e5af377e1b2377d55andderekross/linux-omarchy-mobileomamobile@26867fc4e826a03d1095341af50b6165d4ea5164. The official Docker procedure (kernel,modules,initramfs,images) succeeded, and the three generated images are byte-identical to the archived reference set.The later Fastboot inspection was strictly read-only; the historical Z22 installation is documented separately in our logs. That Fastboot session was done with the phone already in the bootloader by hand. Only
fastboot getvarcommands were used — no flash, noset_active, no reboot, nodevinfowrite.What we confirmed earlier in our own logs:
set_active=bwas executed during the Z22 installation (Setting current slot to 'b' OKAY), and the before/aftergetvarpair showedslot-successful:bandslot-fastboot-ok:bgoing fromyestonowhileslot-retry-count:bstayed at 3. This matchesDevInfoInitSlot()ings-common.Current read-only Fastboot values:
current-slot: b;slot-successful:a=yes,slot-successful:b=no;slot-unbootable:a=no,slot-unbootable:b=no;slot-retry-count:a=2,slot-retry-count:b=1; bootloaderslider-16.4-14097577;product: oriole,hw-revision: MP1.0. Slot A's currently reported Fastboot metadata matches the historical backup.One thing I cannot reconcile from the available records: three reboots are documented in our session reports (02:38, 04:18 and about 05:45 UTC on 19 Nov), but
slot-retry-count:bwent from 3 (afterset_active) to 1 — only two decrements. There is no intermediategetvarreading in our logs, and no furtherset_activeis recorded. A community report from another device (Somnius on the Pixel 8 Pro, issue #252) says that entering fastboot withreboot-to bootloaderdoes not spend a try; this is community testimony on a different phone, not a confirmed rule on Oriole. It made me wonder whether one of those reboots did not go through ABL's slot selection, but I have no evidence for that and I am not assuming it.Before considering any further hardware step, I would like to understand the rule rather than test it with the last remaining try. My questions:
retry_count— is it every boot of the active slot whilesuccessfulis clear, or a narrower condition (stage reached, AVB result, shutdown path)?retry_count == 0withsuccessful == no, are there other conditions that make ABL choose slot A (for exampleunbootable,fastboot_ok, or a failed AVB verification)?fastboot getvar current-slotbe read against theactivebit indevinfo? Is it always the same value asgetActiveBootSlot(), and can it diverge?userdataand slot A — in particular, doesfastboot --set-active=balone refilling B's tries without a flash, as reported for other phones, also hold on Oriole?I am not asking to repeat the historical
devinfowrite, and I am not claiming that Oriole is qualified for #483. I will not run any further boot or write until this is clear. A pointer to the relevant ABL documentation or to an existing Oriole note would be enough; if there is no public source, a short answer from you would help.Thank you,
Somnius commentedon Nov 19, 2026
@Ruraam since you cite my comment: this is exactly what my logs have. It's all one Pixel 8 Pro (husky, zuma), not Oriole, so it can't answer your five questions. Those are for @derekross.
On husky, with
slot-successful:b: no:systemctl poweroffrestarts the phone and spends one too: Pixel 8 Pro: systemctl poweroff restarts the phone, so every shutdown spends a slot B retry #578reboot-to bootloader: my 11-09 note says it didn't cost a retry, and on 11-18 22:55 the count was 2 before and 2 after. One row disagrees: 11-18 around 22:35, from the initramfs rescue shell, 2 then 1. So that claim of mine is only partly backed.fastboot set_active balone, no flash, gave 3: nine times on 11-09 (Marking slot B successful: the plan after #125's hold, and for the Pixel 8 Pro #252 (comment)), from 0 on 11-13 when the slot had already flipped to A (Marking slot B successful: the plan after #125's hold, and for the Pixel 8 Pro #252 (comment)), and from 0 on 11-18 22:10. Two more ended at 3 but are weaker: 11-18 23:50 (my log doesn't say whether a flash was involved) and 11-19 07:55 (the 1 before it is assumed, not read).current-slot: a. On 11-18 it fell back to A (GrapheneOS here) anduserdatawas wiped. On 11-01 it ran out of retries and GrapheneOS wipeduserdata; nogetvarlogged that day.So on husky, at 1 with successful
no, I don't boot B again. I go to fastboot with the buttons, read, and refill first. Whether that holds on Oriole I don't know. @derekross's 11-13 comment covers the refill on other phones: #252 (comment)I can't explain your three reboots against two decrements from anything here.