Бэклог
Работа впереди — и уже сделанная.
Открыто · 95
A device base derives its parts' measured fields from its own runs (a derive-parts record that devlab gather runs)
device-templateAfter the move a device's runs are filed into its own base, while parts.measured/measured_run were derived by docs from docs' test_results (derive-support-status) and only seeded at the move. Without…
The device template names bringup and ships a [boot] table in device.toml
device-templateAGENTS.md says where a device's first-boot images are made (lab-tools bringup, followed by pin, never a scripts folder of its own, never an edit of the tool); device.toml gets [boot] in pmOS deviceinf…
device.toml read off a running phone
device-templatedevlab reads a running phone (compatible, sound cards, signal processors, inputs) and proposes its device.toml, [health] included - the onboarding step of the device-template design (linux-phone-oco/d…
A second phone through the template
device-templateThe proof that surya is not the mould: a second postmarketOS phone's lab made by Use this template, reaching healthy with only its device.toml filled.
Switching the phone shell, a word of devlab
device-templatesurya-lab ui/switch.sh switches the shell in place (postmarketOS UI packages, tinydm session, reboot, then the shell, DSP and speaker checked); nothing in it is surya own. As devlab shell switch NAME…
The lab's USB wake lock, carried by devlab
device-templateEvery lab phone on USB needs devlab-usb-awake (a suspend attempt drops the link); surya carries it as its own overlay. It belongs to devlab, written at flash for any phone.
A second phone in the lab
docsOne phone serialises everything that changes it: four sessions now share surya, and devlab holds the phone so that changes queue. A second phone doubles what can be tried at once and proves the device…
How a device row names a docs component or feature: the about rule works within one base
docsOpen: proposal is docs:ID in text and a local components list in the device base seeded from docs; decide with the first device base. Owner: idealkb (ikb) with docs.
What Android does on the same phone, measured once and kept
docsThe project's goal is phones as good as Android; the face says it from tests that prove parts present or working, but nothing measures 'as good'. ArtCraft keeps reference corpora an oracle compares ag…
Briefs say when a session ends and how questions route: hand-over into the domain's how-it-stands row, end before compaction, ask the record, then the live owner directly, then a consult
docsDecision 2026-11-17-experts-written-not-standing. Next: orchestration edits prompts/bring-up.md, feature-research.md and orchestration.md; the first how-it-stands rows are audio-mic's and deep-sleep's…
A custom render.py that draws support_status as a real device x component matrix, not one page-section per row
docsThe generic engine renders it as plain rows for now (it has no reason to detect this shape specially); a matrix is much easier to scan at a glance once there is more than one device.
A device template: the conveyor any phone's support is built on
docsAsked for 2026-11-10: the conveyor as a repo of its own, so anyone can bring support to another phone with the same conveyors, CI, tests, tad page and automatic integration into the grand picture; sur…
soc-sleeps also reads ddr and says which rung is missing (lab-tools)
docs доказано тестом soc-sleepsNo phone. In lab-tools health/suite/checks/base/90-sleep.sh: read ddr Count and the DDR LPM stat too, and name the rung reached (cluster S2 usage, cxsd, aosd, ddr). Today a pass needs cxsd or aosd and…
Fork sm7150-mainline/linux into linux-phone-oco and wire it into lab-agent's build/boot jobs
docsUpstream: surya's capture joins sm7150-mainline PR 60
docssm7150-mainline/linux PR 60 (davinci, open since 2026-08) already carries the swr_tx_data mux, the TX wakeup IRQ, the port numbering and the capture DAI 1. Step: review and test it on surya, comment w…
A board for linux-phone-oco
docsThe user's idea: a shared board for the org, like lab-agent's. Later.
The device toolset in Rust, for any phone
docsAsked for 2026-11-10: tooling must not be tied to surya - surya is the test subject; the work is a conveyor for any phone, built as one Rust toolset (rust-script for small scripts) as we go, informed…
Sensors upstream: configs at the vendor mtime, SM7150 data to hexagonrpc
docsDecided 2026-11-17 (research-sensor-hub, decision 2026-11-17-sensors-registry-kept-current). After sensors-without-crash. Offer: firmware-xiaomi-surya (sm7150-mainline) and pmaports install the sensor…
The start weight per role, a number on the face
docsHousekeeping (ideal-agent practice 45) measures the bytes a fresh session reads per role (scripts/start-weight). Next: a table or named record for the measurement (role, bytes, date), filed by the pat…
Write the summaries tad q about shows as due: deep-sleep, dl-reaches, sensors
docs доказано тестом tad q about shows none of them as summary dueOn 2026-11-17, after audio-mic's: deep-sleep 15.7 KB in 17 records, dl-reaches 10.9 KB in 15, sensors 11.3 KB in 10. Each is written by the session that next records into it, or its keeper at hand-ove…
The author's TTY-first shell
docsHis own shell for Linux, mobile first and at home on a desktop: built from nothing, on the best existing parts or their reimplementation in Rust. It starts as the raw console the lab's agent uses for…
Upstream: tqftpserv >= 1.2 in Alpine v3.24-stable
docs ждёт the author: a request to Alpine is made in his nameFound 2026-11-17 on surya (bring-up): Alpine v3.24 (postmarketOS v26.06) ships tqftpserv 1.1.1, which never writes the modem's short state files, so SM7150/SM7125 modems never come online and their SI…
Upstream: the ADSP must not boot before the in-kernel PD mapper
docsFound 2026-11-17 on surya (bring-up): remoteproc's qcom_common only adds an auxiliary pd-mapper device when it prepares a remote processor; qcom_pd_mapper is a module that udev loads later, so on abou…
Upstream: tas2562 sets the I2S slot length from the sample width
docsKernel patch on linux-phone-oco/linux branch fix/tas2562-i2s-slot-length (one commit, compiles for arm64, verified on surya as a module built from the exact v7.1_rc3 source). Send to sm7150-mainline a…
Upstream: tas2562 digital volume writes do not reach the chips
docsSee plasma-silent. Reproduce: set the Digital Volume Control, read page 2 registers 0x0c-0x0f live with i2cget. Then a driver fix (write through, or restore on power-up) and a report. Fixed 2026-11-10…
Upstream: stereo speaker UCM profile for surya
docssurya-lab ucm/Xiaomi/surya/HiFi.conf: two channels, top amplifier (0x4d) left, bottom (0x4c) right. For alsa-ucm-conf (Xiaomi/surya), once the kernel fix is accepted or at least posted.
A spec record: what the observer learned, with evidence kind, file, hash, command
docsWanted before the first part that needs clean-room observation (display panel init is the likely first). Ask the ikb session to add a spec record (fields from research 2026-11-17-cleanroom-agents); fi…
Plan the real bring-up order from the first mainline map and queue the items
google-walleye-lab доказано тестом A research row that orders the bring-up from the first mainline map and the vendor facts (the running fdt, the dmesg of y2, the vendor kernel's regulator_summary): the dependencies between the parts (PMIC and its regulators -> UFS (rootfs) and storage; PMIC/charger -> battery and safe charging while testing; arm-smmu -> UFS DMA, display, USB; clocks (gcc/rpmcc) -> everything; USB networking (a function the host can bind with the lab-prepare rules: ncm with its attributes written the mainline way, ecm or rndis, as the mainline kernel supports them natively) for ssh instead of a serial console; display (S6E3FA5 command-mode through DSI/DPU) and touch (Synaptics RMI4 i2c); audio, modem and Wi-Fi/BT later), the order with what each unlocks and its proof test (the kernel dmesg, a sysfs file, the shell), and for each part the backlog item (build and phone) queued by the research session itself in this base with --needs, proof and note; the first three items after the PMIC get ranks; the research names the parts mainline has for the Pixel 2's chips (panel driver for the S6E3FA5?, touchscreen RMI4, charger smb2 (pmi8998), fuel gauge, haptics DRV2624, audio WCD934x/TAS2557) and which would need new drivers (upstream candidates)Now that our own kernel and device tree boot to a shell (first light, lab PR 152) and the first dmesg is read (map), the lane needs a dependency-ordered plan of the real parts, queued as items, so bui…
Remove the C probe of the ffs helper (build/ffs-acm/ffs-acm.c and what only builds n3 from it) once n3r is tried
google-walleye-lab доказано тестом build/ffs-acm/ffs-acm.c and anything that only builds n3 from it removed (n3's recorded hashes stay as history; n3r is the image to try); the lab's checks pass; read back from main ждёт walleye-ffs-rust-try (the keeper's first try of n3r)CI half split off as walleye-rust-ci-ffs-acm-rs (done). This half waits by the coordinator's order until the keeper has tried n3r (walleye-ffs-rust-try), so the C probe stays buildable until the Rust…
Reach a hung or crashed walleye without a display: warm reset, pstore, UART
google-walleye-lab доказано тестом after a deliberate panic the next boot lets the keeper read the panic log from /sys/fs/pstore (test: walleye-pstore-panic)ramoops at 0xa1810000/0x200000, qcom_wdt.arm_on_probe=1, watchdog.open_timeout=300, reboot=w (porthole); the Pixel debug cable is not in the lab.
Display: Samsung S6E3FA5 command-mode panel through DPU/DSI
google-walleye-lab доказано тестом the DRM connector is connected and a test pattern is seen by a person (test: walleye-display-pattern) ждёт walleye-ufs-readonly (not required) / walleye-first-fastboot…No mainline driver; start from the out-of-tree S6E3FA5 driver (OnePlus 5) and porthole's sw43402 cmd-mode work for the DPU side; init sequence from the GPL downstream dtsi; reset gpio94, TE gpio10 (wa…
Find out which dtbo overlay id walleye's bootloader needs before a mainline kernel boots
google-walleye-labThe taimen bootloader refuses a kernel unless dtbo holds an overlay with id 2704 (kcxt-dtbo-wahoo-mainline); walleye's id is not known. Read it without writing: from the stock dtbo (Android booted, or…
Get the UART debug setup for the Pixel 2 (breakout + 1.8 V adapter, or USB-Cereal) and confirm it
google-walleye-lab ждёт the author: choosing and buying the parts (tad q hardware)The parts and wiring are in the hardware table (uart-debug-setup). Once owned: try it, record the log, and mark the fit confirmed with its evidence.
Give the lab user access to our gadget's /dev/sdX and /dev/ttyACM* through the lab-prepare tool, then read m1 and n3
google-walleye-lab ждёт the lpoco lab-prepare tool (orchestration); the author will…device.toml now declares [host] modules usb_storage, cdc_acm and gadget serial walleye-first-boot with access block and tty (devlab accepts the table: keeper status and fastboot state run on it). Phon…
Read the bootloader's own reason for Load Error over the Android Debug Cable (UART)
google-walleye-lab доказано тестом a captured serial log of one refused control try that names the failing step (LoadImageAndAuth, dtbo, vbmeta or other), attached to a research row ждёт the author's debug cableThe bootloader logs the reason to the serial console; the Nura wiki names the Pixel 2 Android Debug Cable. Hands needed; the phone itself is not changed.
dtbo_b stub with overlay id for walleye - only if the control boots and e,f still end in Load Error
google-walleye-lab доказано тестом walleye-dtbo-stub-boot: after the stub, fastboot boot of e or f leaves fastboot; stock dtbo flashed back and its hash equals the one from the zip ждёт the author's word (a write to the phone)A write to the phone: needs the author's word and the way back first (stock dtbo from the LineageOS zip read in walleye-loaderror-stock-ids, flashed back with fastboot flash dtbo_b). Overlay id for wa…
The Pixel 2 is gone from USB since the pk5-stages tries (2026-11-18 ~02:30 UTC): phone work is parked until the author is back
google-walleye-lab ждёт the author: the phone is off USB (hold Power 10-30 s; if it…Seen by the keeper (lab how-the-phone-stands 2026-11-18-2300): neither fastboot nor adb, nothing on its USB port; devlab says off or needs hands. Candidate causes: a g-image hang that the outside watc…
PMIC first: bind pm8998 and pmi8998 on the SPMI bus (the first mainline dmesg: 'does not have reg property')
google-walleye-lab доказано тестом dt/msm8998-google-walleye.dts gets the PMIC nodes right so that the kernel's SPMI bus binds pm8998 and pmi8998 with their children (pon, rtc, gpio, charger/fuel-gauge candidates, regulators if the vendor tree lists them): the fault of the first mainline dmesg is 'spmi pm8998@0 and pmi8998@3 does not have reg property (-22)'; read how mainline's own msm8998 boards (the OnePlus 5/5T, Sony Yoshino, Xiaomi Mi 6 dts in the kernel tree used by the build) include pm8998.dtsi and pmi8998.dtsi (spmi_bus children with reg = <0x0 SPMI_USID>, <0x1 SPMI_GSID>) and what the running fdt of the phone shows for the same bus (sid values, the nodes it enables); the change as small commits, each with the reason; image z1 = y2 (our gadget init, ACM console) with the new dts; the host test (the shell reader from lab PR 154) greps the new dmesg: zero 'does not have reg property' lines, pm8998/pmi8998 child devices present under /sys/bus/spmi/devices, and the rest of the dmesg not worse than the first map (research 2026-11-18-walleye-first-mainline-shell-map); programmer review; hashesFact (lab research 2026-11-18-walleye-first-mainline-shell-map): the first mainline dmesg of this phone (Linux 7.3.0, 336 lines): USB udc, pinctrl, spmi 800f000, RPM regulators and clocks, ramoops, wa…
Kernel patch T4: a reboot notifier setting the PON PS_HOLD reset type to warm (read registers from vendor qpnp-power-on first)
google-walleye-lab доказано тестом pstore holds the marker after a plain reboot -f ждёт paused by decision 2026-11-18-walleye-vendor-kernel-as-instr…Only if T3 fails or a plain reboot must keep DDR; candidate to send upstream
After one panic try: unlock by the author, then read /data/vendor_ce/0/ramoops as root
google-walleye-lab доказано тестом test walleye-panic-text: a research row quoting the first panic lines (Kernel panic - not syncing: ...) from the ramoops file, or recording that the file is absent for this kind of reboot ждёт the author's unlock after the try; Android on a must not be…Ramoops region 0xa1100000 (2 MiB) survives the warm reset; the vendor init encrypts it after unlock. Try may be any existing image that panics (gadget3), no new build.
walleye's build and check scripts become calls to bringup
google-walleye-lab ждёт bringup-image-swapAfter bringup-image-v0, -swap and -initramfs land: build-bootimg, build-bisect, check-bootimg, check-bisect, build-initramfs, build-all become one-line calls (or go) with the header facts in device.to…
Touch: Synaptics RMI4 on blsp1_i2c5 (0x20, irq gpio125, reset gpio89)
google-walleye-lab доказано тестом evtest shows touch events from the RMI4 device when the panel is touched (test: walleye-touch-events)Mainline rmi_i2c driver; supplies pm8998 l6 and l28; xd boards add a pmi8998 power gpio - which board revision the lab phone is, is not known.
UFS probes and the partitions are listed, read-only
google-walleye-lab доказано тестом lsblk in the initramfs lists the UFS LUNs and by-name partitions without a write (test: walleye-ufs-lsblk)Enable ufshc + ufsphy with the vcc/vccq rails from the wahoo dtsi; mount nothing, write nothing; this is also the way to read walleye's own boot.img header.
USB-C debug UART cable for walleye (no display): decide on 'fastboot oem uart enable' first
google-walleye-labpmOS wiki lists the Android Debug Cable as tested with walleye. The oem command's effect (RAM flag or persistent) is not known, so it counts as a change to the phone: a decision first. Needs the autho…
Retire walleye's own bisect scripts (build-bisect.sh, check-bisect.sh and their kin) for `lpoco-bringup image bisect SERIES`: write walleye's series files
google-walleye-lab доказано тестом walleye's series file(s) declare the swaps of the bisects done so far (pk4/pk5/pk6, d1-d6, w*/wa/wpa, wp*, the fdt compare when it exists) against the control image; `lpoco-bringup image bisect` reproduces at least pk5, wa and one d-image byte for byte (sha256 equal to the recorded ones, or the difference explained); the old scripts removed in the same change; checks and hashes; programmer reviewFrom orchestration (lab-tools PR 123, release v0.4.0): the one command `lpoco-bringup image bisect SERIES --out DIR` counts the swaps against the control and wraps lz4 kernels. Low rank; any builder r…
bringup: the host tool that makes a device's first boot images (boot image pack, unpack, swap, check; initramfs)
lab-toolsDecided 2026-11-17 (decision 2026-11-17-bringup-in-lab-tools): lpoco-bringup, a workspace member of lab-tools run as 'lpoco bringup' under one roof - the lpoco dispatcher (lpoco-dispatcher), one pin,…
devlab and tone declare their command lines with argenv, not clap
lab-toolsPractice 43: a program on clap moves to argenv when it is next changed; devlab changed 2026-11-17 (lab-tools 71, 72, 79, 80, 81) and still parses with clap, as does tone; lpoco-health already moved. E…
A turn does not end with unlanded changes
lab-toolsA Stop hook (devlab guard stop) that refuses to end a turn while a place the session worked in holds changes not landed: the places are read off the transcript (the cwd of its Bash calls and the files…
The checks surya-lab ran from the lab, carried into the suite
lab-toolssurya-lab tests/automatic/checks.sh retired on 2026-11-17 (its decision 2026-11-17-the-lab-s-own-conveyor-retires); four of its checks are not yet in the phone suite: bt-sees-devices (a 6 s bluetoothc…
One controlled suspend and wake in the suite
lab-tools доказано тестом fixture tests of the verb (plan with no phone, counters parsed before/after, refuses without a declared way back); then one cycle on surya through surya-lab's queue, the report's soc-sleeps turning from skip to pass or fail2026-11-18, orchestration: the build part goes to the lab-tools builder. surya has run this cycle by hand through surya-lab sleep/way-back.sh (cycle SLEEP bt-on [stop-sensors]: watchdog fuse as the wa…
The microphone, placed on notice
lab-toolsThe microphone can be relied on when it is placed by the phone. Before any listening test the agent notifies its keeper and waits for his word; only then is the test run, with --placed, and counted. T…
The phone's own microphone, heard the same way
lab-toolsThe lab plays a tone, the phone records it: the instrument for the long-open microphone work. And a second lab microphone, to tell left from right.
Listening by itself, at decent hours
lab-toolsThe watcher stays silent: a tone after every start could sound at night. For it to listen unasked the lab needs to say when sound is welcome.
A room of exact zero means the microphone did not record
lab-toolsRight after a reboot the room recorded at -240 dBFS (digital silence) and the tones still passed. devlab hear should refuse in that case.
Modules no entry boots with, taken away
lab-toolsSeen 2026-11-17: removing bring-up's ear try proved remove keeps shared modules (it removed ear's entry only; mic stays). Left on surya: /lib/modules/7.1.0-rc3-sm7150-wcd, from the wcdswr experiment r…
lpoco host, the rest: lab group for headless, watcher units, cross targets, host commands, foreign rules named, NCM
lab-tools доказано тестом a plan test per part from fixtures: listed, nothing changed without the verbAfter lpoco-host's first build (its acceptance on walleye passed), the rest of the host from declarations, each its own small change with its plan test first: (1) a lab group (sysusers.d, MODE 0660 on…
Half-manual tests
lab-toolsSequential tests with an observer - a person now, a rig later.
Built on postmarketOS's conventions
lab-toolsdevice.toml from deviceinfo; tests in the shape of pmOS's initramfs CI; bootrr.
A proper rig
lab-toolsAsked for by the author, for the future: a microphone fixed in place, a webcam that sees the screen, servos that press the buttons and touch the glass - so the tests that need a hand, an eye or an ear…
The audio processor signalling handover every second
surya-labMeasured in the test kernel: 'Handover signaled, but it already happened' once a second from 11 s after boot (54 by 65 s), with or without audio. Also seen in a stock boot's crash memory. Likely an in…
Bluetooth pairs and carries audio: from partial to works
surya-lab доказано тестом bt-pairsThe author, 2026-11-17: bluetooth is partial (controller and powered pass). A works test: pair with a device the lab keeps, carry audio or a file; witnessed where a person must confirm.
One-shot dtb: CTS pull-down in qup-uart3-sleep like sc7180-trogdor; same cycles as bt-trace-cycles
surya-lab доказано тестом bt-no-controller-crashStep 2. Keep the dtb out of the kept kernel until it measures better.
sxmo autosleep gate: no rtcwake while hci0 is not UP or an rfkill is in D state
surya-lab доказано тестом bt-rfkill-not-stuckStep 4, independent of the others: removes the r4 wedge whatever the controller does.
tests/health/base/bt-suspend.sh in the suite, pushed to the phone with the runner
surya-lab доказано тестом bt-hci0-upLanded with the research; the keeper puts it on the phone with lpoco lab suite put now.
Bluetooth s2idle: 20x15 s and 3x120 s way-back cycles bt-on, wakeup ON, with hci_qca dyndbg and geni irq trace; read storm vs crash order
surya-lab доказано тестом bt-no-irq-stormRead-only on the phone apart from trace switches; watchdog armed, no unplug. Decision 2026-11-18-bt-suspend-way step 1.
Sleep hook: unbind 88c000.serial:0 from hci_uart_qca before suspend, bind after (never rfkill); same cycles plus BT audio reconnect
surya-lab доказано тестом bt-hci0-upStep 3, only if steps 1-2 leave controller crashes. Hazard: BT-off s2idle hung twice; run with watchdog and ramoops armed.
Ask each camera slot what it is
surya-lab доказано тестом image-sensorThe first step of the cameras: a test device tree that only powers a slot (supplies, clock, reset), booted once as the second entry; then the chip's identity register read over its control bus. Turns…
devlab kernel try NAME --cmdline: a one-shot kernel command line
surya-lab доказано тестом devlab kernel try NAME --cmdline noirqdebug boots with it in /proc/cmdline once, then stock/kept ждёт low: r5 no longer needs noirqdebug (irq 157 stays enabled wi…lab-tools change (coordinator routing it): one-shot entry NAME-cmd falling back by itself like --dtb; needed for noirqdebug in r4
Boot the current sm7150-mainline kernel on the phone via lab-agent and record dmesg + audio card list
surya-labThe first real job: proves the whole lab-agent loop end to end on real hardware, and gives a verified baseline instead of secondhand notes.
Why the AOP does not ack QMP sends (qcom_aoss_qmp c300000)
surya-lab доказано тестом soc-sleepsLower rank (research 2026-11-17-sh0-and-aop-research: node equals sm8150/sc7180, serves only the aoss knobs, hypothesis 5). Read only: dump msgram descriptor and mailbox words (magic, version, mcore/u…
Read all power-domain idle_states rows and regulator_summary users across one way-back cycle 15 bt-on (read-only)
surya-lab доказано тестом soc-sleepsSettles why the cluster S2idle column is 0 (wrong row/domain or a path that did not count) and lists the rails that stay on in suspend. Write the output into the cycle log. Watchdog armed.
Silence what wakes the CPUs in suspend, one source per cycle
surya-lab доказано тестом suspend-irq-quietOnly if floor-entry-trace shows interrupts through the cycle (Odin 2 was held by an SD-controller IRQ storm of ~30/s, armada#274). Take the top source from the delta, quiet it by a runtime step (drive…
Find who holds the SH0 x=1,y=3 sleep vote and drop it
surya-lab доказано тестом soc-sleepsPaused 2026-11-17 by decision 2026-11-17-floor-read-the-entry: read the entry first (floor-entry-trace), then decide whether the SH0 owner hunt resumes.
If SH0 stays: the other TAG_ALWAYS QUP and crypto consumers, one per cycle
surya-lab доказано тестом soc-sleepsOne-shot dtb per change: the serial toward SLAVE_LLCC (sm7150.dtsi ~1259), then the crypto engine (~460), then the rest of the QUP ALWAYS paths; each judged by the sleep-set flush and soc-sleeps. Skip…
Standby current: battery-only suspended cycle of 10 minutes, charge_counter delta over time
surya-lab доказано тестом standby-currentNeeds the cable out for the window (the witness's hands) or the fuel gauge's discharge read while charging is held off; the RTC wakes the phone and the cycle log is read after. Pass: under 15 mA avera…
A debug kernel that dumps held votes in late suspend
surya-lab доказано тестом soc-sleeps ждёт floor-read-holders and the subtractions aboveOnly if the reads above leave the holder open: a debug patch (builder) with a pm notifier / syscore op that prints clk enable counts, regulator use counts and icc votes at dpm_suspend_late, out of tre…
Allow USB runtime PM (power/control auto on the HS PHY and the dwc3 controller) for one cycle and read counters and current
surya-lab доказано тестом soc-sleepspmaports#4846: lowers draw slightly and keeps USB from holding CX. Needs the lab link to survive: keep the watchdog armed, one change per cycle.
GPS: a component and a failing test, so location can be measured
surya-lab доказано тестом gps-fix ждёт the decision of a feature research session: the author has h…The author, 2026-11-17: lp-connected needs the feature gps, but components has no GPS row, so nothing can measure it. Add the component and a failing test first; on SM7150 location comes through the m…
Image sensors: described, then identified, then seen
surya-lab доказано тестом image-sensorKnown (research/cameras.md): the pipeline and all three control buses are bound; no sensor is described, which alone is why the cameras are dark. The vendor's kernel gives the wiring of seven sensor s…
The phone cannot stay off: shutdown and a charger both boot pmOS
surya-lab ждёт the decision of a feature research session: the author has h…Seen 2026-11-09. poweroff in pmOS boots pmOS again. The same shutdown from the EOS recovery on the phone stays off, until a charger is plugged in; then pmOS boots instead of Android-style off-mode cha…
Under Phosh the PC loses the phone over USB
surya-labAfter switching to Phosh (2026-11-09) the PC showed no USB device for the phone, so no SSH over USB; the user saw the same before, in an earlier Phosh try. Unclear yet: Phosh changing the USB gadget,…
Witnessed: cover the proximity sensor at the ear, see the reading go near
surya-lab ждёт the author: hand over the top of the phoneExpect: monitor-sensor --proximity prints near: 1 when a hand covers the top bezel and near: 0 when removed; ask the author for go before running it
Proximity blanks the screen at the ear in a call
surya-lab доказано тестом proximity-readsDecided 2026-11-17 (research-sensor-hub, decision 2026-11-17-sensors-registry-kept-current). After sensors-without-crash. The sensor is surya's stk3a5x infrared one (not ultrasonic: nothing in the mic…
A shell-free base for hardware tests
surya-labProposed by the author after a day of faults that a shell's policy turned into loops: test the base - sound, modem, suspend, storage - on a console-only system, with no desktop or shell deciding when…
FastRPC helpers stopped across suspend; the sensors back
surya-lab доказано тестом motion-and-light2026-11-18 keeper: stop/start around suspend measured 3/3 crash-free with sensors back (research 2026-11-18-sensorspd-stopped-no-adsp-crash). Left: same via sxmo/systemctl suspend, then done. 2026-11…
Standby current measured beside Android's
surya-lab доказано тестом soc-sleeps ждёт the author back to unplug, and the Bluetooth suspend fix (ba…r4 SCORED: aborted/partial, unscorable (4 samples, median 120 mA, 63% asleep); aborted by irq 157 + BT SSR at 22:40:48; hang after = sxmo suspend loop on stuck BT SSR, rfkill D-state. Next r5: BT off…
Waking from deep sleep for calls and messages
surya-lab доказано тестом wakes-for-call ждёт the SIM back in surya (the author took it out 2026-11-17) an…On the floor reached above: the modem's IPCRTR glink wakeup enabled (/sys/bus/rpmsg/devices/4080000.remoteproc:glink-edge.IPCRTR.-1.-1/power/wakeup), ModemManager's quick suspend/resume quieting the m…
The standby drain test lives in the health suite (lpoco-health) as a switchable test
surya-lab ждёт moved to lab-tools in docs (orchestration records it in the…The author (2026-11-17): the standby test belongs in the health suite under the lpoco roof, switchable (lpoco-health on/off), not a one-off script. Move sleep/standby-run.sh's measurement into a lpoco…
surya's fastboot serial in its manifest, read from fastboot
surya-labdevlab flash now refuses surya until surya-lab device.toml says [fastboot] serial and writes = true (lab-tools PR 79, decision 2026-11-17-fastboot-by-serial-writes-by-manifest). Next time surya sits i…
Checked by hand
surya-labButtons verified by a person; vibration still to check (the first half-manual test). Settled 2026-11-11: checked by hand once; next: checked by the conveyor as scenes (agent access).
Telegram client at provisioning: uncomment telegram-desktop in provision/packages.txt
surya-lab ждёт Alpine v3.24 rebuilding telegram-desktop against ada 4 (liba…Alpine v3.24 telegram-desktop 6.8.5-r0 needs libada.so.3, gone with ada 4.0; edge 7.4.4-r0 is rebuilt. Check with: echo apk add --simulate telegram-desktop | devlab phone root. When it passes, uncomme…
Try every mobile shell with working sound
surya-labSwitched in place with surya-lab ui/switch.sh (postmarketOS UI packages, tinydm for all; no reflash, the sound fixes stay). Phosh: runs, stereo verified by ear 2026-11-09. Next: Plasma Mobile, GNOME M…
Verify headphone, mic and earpiece status directly on hardware instead of relying on the secondhand pmaports note
surya-labThe source only confirms mic/earpiece are broken as a group; headphone output specifically was never separately confirmed one way or the other.
VoLTE and emergency calls, to finish calls and messages
surya-lab доказано тестом volte-callThe author, 2026-11-17: confirm first whether today's working calls go over VoLTE (Germany has switched 3G off; read the modem's call and IMS state during the daily echo call). Emergency calls are pro…
More phones
sxmo-deviceprofilesProfiles for other devices, made the same way.
В работе · 11
docs holds only what helps every device; each device has its own open base; docs derives progress from them
docsBoth lanes record only in their own base (google-walleye-lab #15, surya-lab #136; docs #353, #360); docs pulls their parts daily (feed-device-bases, #355; surya's base named #361) and the face reads t…
How a device repository takes later changes of device-template beyond idealkb's generic files
docsDecided 2026-11-17 (idealkb decision 2026-11-17-chain-follow-by-slice): the same pin-and-follow slice as idealkb, not a fork, subtree or Copier. Done: device-template lists what it gives (generic) and…
The audio DSP sleeps when nothing plays
docs доказано тестом soc-sleeps ждёт a deeper decision: what on the Apps/AOSS side holds CX and A…2026-11-17 keeper-6: two floor tests (adsp stopped; adsp+cdsp+modem stopped) still gave cxsd 0, aosd 0, ddr 0 - the DSP is not what holds the SoC (research 2026-11-17-adsp-stopped-cx-aosd-ddr-stay-0).…
Upstream work is kept as records: upstreams and upstream_items (idealkb 0.8), the upstream-* items headed in them
docsSeeded 2026-11-17: six upstreams, nine items (sxmo profile sent; tas2562 x2, WCD series, UCM, pd-mapper, tqftpserv, PR 60 comment as drafts); patches and messages copied into upstream/UPSTREAM/ITEM/.…
Measure Android on the Pixel 2 before anything is flashed
google-walleye-labMeasured read-only 2026-11-17 (research android-reference-walleye-measured): build, sensors, battery, boot time, audio config. Left: sleep current (needs a meter, radios stated), audio levels (tone so…
Boot image: header v0, base 0, kernel 0x8000, ramdisk 0x1000000, tags 0x100, page 4096, dtb appended
google-walleye-lab доказано тестом unpack_bootimg of the result shows the offsets and cmdline, the dtb is inside the kernel blob, size is under the phone's max-download-size (test: walleye-bootimg-unpack) ждёт walleye's stock boot.img header (android-reference-walleye)…Four images, 17 checks each, 0 failed: a,b as before; c,d = a,b with os_version 8.1.0 / patch level 2020-01 (4 header bytes differ). sha256 in ~/Projects/walleye-build/images/SHA256SUMS. Header v0 fit…
First fastboot boot of the walleye kernel: the phone leaves fastboot and its gadget answers
google-walleye-lab доказано тестом lsusb on the host shows the gadget within 60 s of fastboot boot and ssh/telnet to 172.16.42.1 answers (test: walleye-gadget-up) ждёт walleye-boot-refused-why (research), walleye-control-on-slot…Six images and the official LineageOS control all refused with Load Error: the cause is the phone's state or what fastboot boot needs, not our image (walleye-control-e-f-tries). A dtbo/vbmeta write st…
Initramfs that brings up a USB gadget (ACM console + RNDIS/ECM) and telnet/ssh on 172.16.42.1
google-walleye-lab доказано тестом In QEMU aarch64 with dummy_hcd the initramfs creates the gadget and answers on its address (test: walleye-initramfs-gadget-qemu)Built (build/initramfs, busybox 1.36.1 static arm64, 1148576 bytes gz). Proven under x86_64 QEMU with dummy_hcd (build/selftest-initramfs.sh): gadget binds, lease 172.16.42.2, ping and telnet command…
lpoco-health: the one health suite on the phone, with switches and run-time detection
lab-toolsLanded 2026-11-17: lab-tools #75, #77 (compat names), package default (next); put on surya by the keeper, runs gathered (docs #213). Open: the surya-lab checks call the new names (keeper repointing);…
lpoco host: the lab host prepared from the lab's and each device's declarations - udev access for gadget serials, host modules, the host's own tools at their pins; plan by default, prepare now to change
lab-tools доказано тестом host/tests/plan.rs plan_grants_gadget_serial_and_changes_nothing (first, failing); then plan_lists_host_tools_at_pins_and_installs_nothing; acceptance: lpoco host prepare now on the lab, then the walleye keeper reads /dev/sda and /dev/ttyACM0 as the lab user ждёт the tty half of the acceptance: walleye's pk5l (the mainline…Decided 2026-11-17 (decision 2026-11-17-lpoco-host; research lab-host-prepare-landscape, lpoco-host-how-it-stands). A new workspace member host/ in lab-tools: the crate and binary lpoco-host, run as `…
surya's own cluster sleep parameters in its device tree
surya-lab доказано тестом soc-sleeps ждёт a deeper decision: WHAT holds CX, AOSS and DDR with an apps…2026-11-17 keeper-7: cycle 15 bt-on on the cp.dtb (0x41002244, 0x4100c244) 8 times: a real 15 s s2idle (RTC alarm wake), cluster S2 entered, cxsd/aosd/ddr 0 (research 2026-11-17-cluster-params-clean-y…
Готово · 227
The template, made from what surya proved
device-templatedevice.toml with every section devlab reads (connection, provision, flash, health) and what to change; tests/health for a phone's own checks; tests/results for its reports; AGENTS.md conventions; CI f…
surya-lab re-based onto the template
device-templatesurya-lab keeps scripts the template no longer needs (lib/phone.sh, tests/run.sh, tests/to-docs.sh, tests/automatic/checks.sh, flash/, provision/provision.sh): each one retires once devlab does its wo…
The ADSP crashes every 13-25 s, restarting audio each time
docsCause found 2026-11-09 by tracing every request the ADSP makes (a verbose hexagonrpcd build on the phone): the sensors service (hexagonrpcd -s) attaches the ADSP sensor process, whose registry init re…
The audio DSP crashes right after resume from suspend
docsTrigger found 2026-11-10, one change at a time: with hexagonrpcd's RootPD helper (a FastRPC client) attached, the audio DSP dies on every suspend attempt - 24 of 24 (fatal error EX:kernel:0x0:IST46:0x…
The agent specifics of linux-phone-oco in their own folder, ideal-agent-linux-phone-oco/: prompts/ moved there with git mv, the domain part of practice 42 beside them, extending ideal-agent
docsDone 2026-11-17: prompts/ moved with git mv into ideal-agent-linux-phone-oco/ beside phone-keeper.md, with a README holding practice 42's domain part and the table of briefs; pointers fixed in the bri…
One cold boot came up without audio services
docsWith the sensors service off, one boot (2026-11-09) never announced the ADSP audio services, so no sound card until the ADSP was restarted by hand. Not reproduced in two further valid cold boots (audi…
The bring-up brief over budget (13.3 KB, budget 8 KB): its detail behind summaries and queries
docsDone 2026-11-17 (orchestration): bring-up folded into phone-keeper.md (5.7 KB) and bring-up.md left as a pointer (0.4 KB); the keeper's first paragraph is its whole start - purpose and the exact comma…
Call audio both ways: q6voice, with the uplink from the microphone
docsResearched (research call-audio-on-surya) and decided (2026-11-17-calls-proven-by-echo-call); the work continues as calls-q6voice-kernel, calls-voice-ucm-uplink, calls-first-call and calls-echo-check.
Calls 4: call-echo, the daily automatic echo call
docs ждёт calls-first-callA telephony-layer check in surya-lab tests/health (generic in lab-tools once a second phone calls): skip as unknown without modem, SIM, registration, voice, call-audio-path, a number in the manifest,…
Calls 3: the first supervised call - tone line, then echo line
docs ждёт calls-voice-ucm-uplink; the author's wordWith the author's word (2026-11-11-modem-read-only-first), the bring-up session calls +49 89 721010 99702 (blueSIP's tone line) and then +49 89 721010 99701 (blueSIP's echo) through mmcli, 20 s each,…
Calls 1: q6voice and the VoiceMMode1 link in surya's kernel
docsFirst step of 2026-11-17-calls-proven-by-echo-call. On linux-phone-oco/linux, a branch with q6voice (q6mvm, q6cvs, q6cvp, the voice DAI) from sdm845-mainline and pulpoff/linux-surya's SM7150 changes (…
Calls 2: UCM Voice Call, TX topology 0x10F70, the 1-channel uplink capture
docs ждёт calls-q6voice-kernelIn surya-lab ucm/: a Voice Call verb (earpiece and speaker, bottom mic; 'VoiceMMode1 TX Topology' 0x10F70), and a small service that runs a 1-channel capture on TX_CODEC_DMA_TX_3 for as long as ModemM…
The check counts crashes on resume wrongly
docsThe check's pattern used \| in sed, which the phone's BusyBox sed does not have: it matched nothing and always counted 0. Rewritten in awk alone - a crash between "suspend entry" and two seconds after…
device-template gets a tad base: tables.toml following idealkb, the parts contract (status, evidence, date), next and phone-queue queries, ikb
docsLanded 2026-11-17 (linux-phone-oco/device-template #4, read back from main): data/ (12 empty tables), tables.toml (own: backlog + needs/proof/rank, devices, parts, test_runs, test_results), idealkb 0.…
Open docs items that need a phone go onto that phone's own queue (its base), or say they wait for a phone
docsDone 2026-11-18 (orchestration-renewed): devlab-fastboot-boot-reads-slots, floor-dsps-never-started, floor-read-holders, floor-no-ignore-unused, mic-ucm-capture, microphone and sleep-measured-once are…
Earpiece: through the WCD9375's EAR amplifier, proven by the top microphone
docsTaken up 2026-11-17 (bring-up), after the microphones work on the kept kernel. The earpiece is driven by the WCD9375's EAR amplifier (pulpoff/linux-surya: MultiMedia3 -> RX_CODEC_DMA_RX_0 -> RX macro…
A domain's expert primer: one named query that shows a fresh session everything docs knows of a component - research, decisions, backlog, tests, sources, support status, its how-it-stands row
docsDone 2026-11-17 by kb-about-link: tad q about --subject COMPONENT is the primer - status, research (its how-it-stands row once written with --about), decisions, open work, log, tests and the sources c…
docs pulls each registered device base's parts over HTTPS in CI and derives support_status; stale or missing shows as not known yet
docs ждёт device-template-baseLanded. Surya: when its rows have moved to surya-lab (data/parts.json non-empty), run: tad r device-base --device xiaomi-surya --base linux-phone-oco/surya-lab in the same change; until then docs keep…
Boot with the DSPs never started, then cycle 15 bt-on
docs доказано тестом soc-sleepsDone 2026-11-17 (keeper-9): DSPs never started (remoteproc nodes disabled in a one-shot dtb), cycle 15 bt-on: cxsd/aosd/ddr 0 -> 0, soc-sleeps fails, processors never asleep: none. No DSP vote is the…
Drop clk_ignore_unused / pd_ignore_unused if the cmdline has them
docs доказано тестом soc-sleeps ждёт floor-read-holdersClosed 2026-11-17 (keeper-9): the cmdline has no clk_ignore_unused or pd_ignore_unused (floor-read-holders), nothing to drop.
Read what is held before and at a cycle: cmdline, clocks, regulators, icc, genpd
docs доказано тестом soc-sleepsDone 2026-11-17 (keeper-9): read at idle and before way-back.sh cycle 15 bt-on, research/deep-sleep/2026-11-17-floor-holders/README.md; what is held inside late suspend is not known yet (floor-suspend…
Known gaps shown as gaps on the face, not as failures
docsThe phone's own verdict leaves out the known gaps (surya: fingerprint, sensors, both cameras - healthy, 4 in known gaps), but test_runs counts them as failed, so every bar on the Tests page carries re…
A coverage query: each device's parts with no test proving works
docsDone 2026-11-17 (orchestration): queries/coverage.sql - tad q coverage [--device D] lists every part of a device whose work no test on that device proves, with its best proof (present or none), its st…
grenudi/ideal-knowledge-base: a new repository that pins tad's release and adds the knowledge-base layer - the file-or-table rule, the about link, the generic queries (about, next, find, recent), the short AGENTS.md entry and the standing tables (backlog, session_log, decisions, research); with tad init in tad for a bare database, and text-as-data-template retired with a pointer once its reasons are carried over
docsDone 2026-11-17: tad init in tad 0.7.0 (tad-engine decision 2026-11-17-tad-init); grenudi/ideal-knowledge-base made and released 0.1.0 (public, oco-project, its own tad base: concepts, decisions, rese…
Record into docs with ikb: one command per record, one page (IKB.md)
docsdocs follows idealkb 0.6.1 and ikb 0.1.1; IKB.md generated, kept fresh by make verify; machine records named in idealkb.toml.
Every record names what it is about, checked: an about field on research, backlog, session_log and concepts (decisions has one, free text in four spellings), one references rule to components, features, milestones and devices, backfilled; and the query about SUBJECT
docsDone 2026-11-17: about on research, backlog, session_log and concepts beside decisions, one references rule per table (components, features, milestones, devices, backlog); 11 drifted spellings in deci…
AGENTS.md is the one entry: under 3 KB, an orientation table from typical question to the command that answers it, tad statements for the rest, never read data/ or docs/ to answer; CLAUDE.md is @AGENTS.md; editing recipes move to CONTRIBUTING
docsDone 2026-11-17: AGENTS.md is 2.1 KB - two lines, linux-phone's own questions and idealkb's generic part (the orientation table, the rule never to read data/ or docs/); the editing recipes are in CONT…
A summary over many records, and research kept aside, from idealkb 0.3.0
docsDone by summary-and-research-layers on 2026-11-17: pin 0.3.0; docs' own about, research record and tables.toml take summaries, material and supersedes; audio-mic summarised (15.9 KB to 4.8 KB); Open H…
Lane launchers and briefs work in the device repository and ask its own base (tad q next, phone-queue); docs' lane-next stays for the general queue
docs ждёт move-device-rowsDone in the moves: the walleye and surya coordinator briefs, phone-keeper.md, the walleye-keeper/research/builder and surya-keeper/research launchers and the waker work in the lab repository and ask i…
Keep the DMIC experiment out: surya has no digital mics
docsBring-up's experiment commits (DMIC pins at mux 1, TX DMIC routing, VA DMIC clocks on a module switch, clock p…
The microphones held across boots
docsThe mic kernel is the phone's default (devlab kernel keep) and audio-mic is out of surya's known gaps (surya…
Capture through the sound server: a UCM profile for both mics, the headset mic when plugged
docsFirst step of decision 2…
Microphone: check, and make it work
docs доказано тестом mic-hears-speakerTaken up 2026-11-17 by the bring-up session, as the weakest link the data shows: audio-mic and audio-earpiece are both broken and two milestones rest on them (surya-sound, and the calls of…
Move device rows to the device bases: backlog, research, decisions, log by device and project, test_runs and test_results
docs ждёт device-template-baseBoth lanes moved: walleye (google-walleye-lab #15, docs #353) and surya (surya-lab, docs, 2026-11-17); rule: decision 2026-11-17-moving-a-lane.
The patrol measures the start weight each run and files bring-up-brief-light-like items when over budget
docsDone 2026-11-17 (orchestration): the linux-phone-patrol round measures the start with ideal-agent's scripts/start-weight on docs at main each run, and files one backlog item about agent-folder when a…
Plasma Mobile: speakers silent, the chain playing
docsPlasma Mobile (fresh Nura image, sound fixes applied on the phone): speakers were silent with the whole chain playing. Two causes, both confirmed: (1) the image restores a saved mixer state with both…
Bluetooth dies across a suspend and hangs the resume
docs доказано тестом soc-sleeps2026-11-17 (surya-keeper-4): not seen again. With Bluetooth unblocked and the serial's wakeup off, 3 cycles of 5-15 s resumed in time and hci0 scanned afterwards (no tx timeout, no QCA retries in dmes…
Deep sleep measured once: the SoC's own counters across one suspend
docs доказано тестом soc-sleepsDone 2026-11-17 by surya-keeper: one cycle by hand, soc-sleeps = fail (cxsd 0, aosd 0, adsp never asleep, cluster S2 never entered). Research 2026-11-17-surya-first-s2idle-cycle. The phone woke after…
Why the first s2idle woke after 2.4 s, not 60
docs доказано тестом soc-sleepsAnswered 2026-11-17: IRQ 157, the serial at 88c000 (research 2026-11-17-surya-s2idle-wakes-at-once); the fix is sleep-wake-uart.
Suspend ends at once: the serial's interrupt 157 is a wake source and a storm
docs доказано тестом soc-sleeps2026-11-17 (surya-keeper-4): done. The wake was irq 158, the UART's RX wake pin, armed by the serial's power/wakeup; way-back.sh cycle SLEEP bt-on turns it off for the cycle with Bluetooth unblocked.…
SMS: the phone texts itself and receives it, after every boot or daily
docs ждёт the SIM's own phone number (the SIM does not store it)Opened 2026-11-17 (bring-up), with mobile data working (Vodafone.de network, a 1&1 SIM). A check sends a short SMS from the phone to its own number with mmcli and waits for it to arrive - proving mode…
Support status read off the suite's results
docsderive/support-status.sql: broken on a failing latest run, works on a passing test that proves the part works, otherwise judged; tried beside the hand-written status - naive reading disagreed on 8 of…
Upstream: the Sxmo device profile for surya
docsWritten and in use (2026-11-10): both panel variants, values from a running Sxmo, shellcheck clean; Sxmo loads it (SXMO_DEVICE_NAME=xiaomi,surya-huaxing, scale 3). Kept on linux-phone-oco/sxmo-utils,…
The phone kept reconnecting on the host's USB
docsOn USB a suspend cannot complete (the link cannot go idle): each attempt aborts at once, and a shell that suspends when idle retried every 12 seconds, each time dropping the link - 26 USB disconnects…
WCD937x device tree: drop aoncc, clock the VA macro from q6afecc
docsThe boot hang (RCU stalls) with the WCD937x patch is the aoncc clock controller at 0x03380000, an SM8250 address with nothing near it in the SM7150 tree: with only aoncc disabled the phone boots (2026…
tests/host/battery-ok.sh must not hang when the phone runs Android: bound every fastboot call by a timeout and fall back to adb
google-walleye-lab доказано тестом the script reads the battery from adb (dumpsys battery) when adb shows the phone, from fastboot getvar battery-voltage when fastboot does (each call under timeout 10), prints "not measured" and exits 2 when neither answers within 15 s; run with the phone in Android it ends in under 15 s; selftest with fake adb/fastboot in PATH; programmer reviewbuilt: adb first then fastboot, each call under timeout -k 1 6, skip exit 2 within 15 s; selftest with hanging fakes passes in 12 s; programmer review done
The keeper reads the battery before every try and leaves the phone charging (in Android, not in fastboot) between tries
google-walleye-lab доказано тестом (1) tests/host/battery-ok.sh (lab): reads fastboot getvar battery-voltage (or adb dumpsys battery in Android), prints the value, exits 1 below 3700 mV; the keeper's brief (the lab's keeper note in the item/AGENTS.md) says to run it before each try and to stop and record when it fails; (2) the keeper ends each session with the phone booted into Android on a good slot (fastboot reboot, adb) while idle, not left in fastboot; checked by a test that records the end stateBuilt: battery-ok.sh, idle-android.sh, battery-selftest.sh (pass) and the keeper rule in AGENTS.md. Next: the keeper runs battery-ok.sh before each try and idle-android.sh at session end; first live r…
Bisect porthole's kernel (pk3: reaches further) against ours (ml6a: dead): swap one part at a time, all without earlycon
google-walleye-lab доказано тестом three images, each from pk3/ml6a's own bytes with exactly one swap, ramdisk = ml5a's init (reboot at once), cmdline = pk3's (no earlycon): pk4 = porthole's kernel + OUR -@ dtb (with overlay stubs) ; pk5 = OUR kernel (the ml6a one) + porthole's taimen dtb (with symbols/msm-id as in pk1); pk6 = porthole's kernel + pk3's dtb + the timing init of ml5b (3 s per point score) to learn which hardware their kernel shows in userspace; checks as before; hashes; the item states the expected return timesbuilt: pk4/pk5/pk6 in walleye-build/images-pk456, 31 checks ok, hashes in research/walleye-first-fastboot-boot/2026-11-18-pk456; try queued as walleye-try-bisect-pk
Bisect from the accepted control to our image: stock kernel+dtb with our initramfs, then one part of ours at a time
google-walleye-lab доказано тестом images s1..s3 built by a script from control/boot.img's own parts (unpack_bootimg) with exactly one part swapped each, check-bootimg ok, hashes in SHA256SUMS; the keeper's tries then show which swap makes the bootloader stop at 'Booting OKAY'Built and checked on the lab (google-walleye-lab #10, main 01f9138): ~/Projects/walleye-build/bisect/boot-s{0,0n,1,2,3}.img, SHA256SUMS there. s1 ramdisk=ours (172.16.43.1) 3f658748...; s2 kernel=ours…
Why the Pixel 2's bootloader refuses even LineageOS's official boot.img from memory (Load Error), and what makes fastboot boot work on it
google-walleye-lab доказано тестом a research row naming the cause or the ranked causes with sources, and the next experiments each with its way back; consistent with the tries in walleye-control-e-f-triesAnswered in research walleye-boot-refused-why-how-it-stands and decision walleye-load-error-vector; next phone items: walleye-load-error-first-state, then walleye-load-error-restart-host, walleye-load…
The best-known way to bootstrap an Android phone to mainline almost from zero, compared with what the walleye lane did
google-walleye-lab доказано тестом a research row with sources, a comparison, a decision with what was rejected: landedDone 2026-11-17: research walleye-bootstrap-method-how-it-stands, decision 2026-11-17-walleye-vendor-kernel-first; next walleye-vendor-gadget-image, walleye-gadget-up-check, then walleye-vendor-gadget…
Read the running Android's full cmdline (skip_initramfs? root=? dm=?) and the ramoops service files, read-only
google-walleye-lab доказано тестом test walleye-boot-cmdline-known: the full /proc/cmdline and chosen/bootargs and /vendor/etc/init/*ramoops* are recorded as a research row; says whether skip_initramfs is presentAnswered by the vendor-gadget4 try (lab PR 25): our idle init RAN (bootreason=reboot at ~120 s, not kernel_panic), so skip_initramfs is not in effect and the ramdisk is used. The cmdline itself was al…
Compare OUR dtb, the alive taimen dtb (pk5) and the phone's RUNNING fdt (vfa): what the bootloader adds or needs, and what our tree lacks; then build the images that test it
google-walleye-lab доказано тестом (1) a research row from the three decompiled trees (material path ~/Projects/walleye-build/material/vf/2026-11-18-vfx-vfa-try1/fdt.dts; wpa/pk5's dtb; our out dtb): what the ABL ADDS to the tree it hands the kernel (chosen, memory, reserved-memory entries, board-id/msm-id/pmic-id, cpu/psci/timer fixups, the overlay's changes): present in the running fdt, absent in the source dtbs; what OUR tree lacks that the alive taimen tree has and the running fdt has too (list the nodes and properties that are in BOTH alive trees but not ours: the likely required ones), in order of suspicion; the structural facts (dtb size, number of nodes, phandle count, root compatible, memory node form: reg ranges and cell sizes, whether our /memory exists); (2) images x1..xn = wa/pk5's alive dtb progressively REPLACED toward ours group by group (the other direction from the earlier bisects: start alive, move top-level groups to ours: root props, /memory, /reserved-memory, /cpus, /psci, /timer, /soc nodes in halves), same kernel/init as pk5 (immediate reboot), checks, hashes, expected returns (alive <= 66 s, dead >= 69 s); (3) the nodes our dts needs for USB peripheral on this phone read from the running fdt (usb@a800000, the phys, regulators, clocks, gpios, extcon/vbus detect) as a proposal for dt/msm8998-google-walleye.dtsCompared in walleye-fdt-compare; images x1..x7, xa built (walleye-fdt-bisect-x-built); queued walleye-try-fdt-bisect for the keeper; USB nodes: none to add, see the compare row
The control boot.img again with slot a active (the state of the very first tries)
google-walleye-lab доказано тестом devlab's before/after slot report and the phone's full fastboot text for the control on slot a; then the phone back to current b, a recoverableproven by devlab's before/after slot report and fastboot text: control refused (Load Error) on slot a; phone back on b
Build our dtb with -@ and the 137 overlay labels; host test passes
google-walleye-lab доказано тестом walleye-dtbo-overlay-mergesDone 2026-11-17: dt/msm8998-google-walleye-overlay-stubs.dtsi (130 stubs, generated by build/gen-overlay-stubs.sh from entry 3's __fixups__; 7 of 137 labels were ours), dtb built with -@ (DTC_FLAGS in…
Our mainline dtb is never started (stock kernel + our dtb stays in fastboot): does the bootloader's dtbo overlay merge need __symbols__ and the nodes the overlays patch?
google-walleye-lab доказано тестом a research row: the structure of the LineageOS walleye dtbo.img (13 entries, id 0x1bc ...; androidboot.dtbo_idx=3 picks one) read with dtc/fdtdump from the public build already downloaded (~/Projects/walleye-build/control/dtbo.img, sha256 fb09009b...b8f6, official), which base-dtb paths/labels/__symbols__ the overlay fragments target, whether our dtb (compiled how: -@?) lacks them, how ABL selects and merges (public Qualcomm ABL source: dtbo idx, msm-id/board-id match, merge failure behaviour), the exact dtb properties to add or to compile, and one cheap image that tells (e.g. s3 rebuilt with our dtb + -@ symbols + the target nodes, or the stock dtb with only the compatible changed); say which of the bisect facts support it (s0n/s0 start; s3 'Booting OKAY' then stays in fastboot; s2 falls back)The host-independent work while the lab-prepare tool is awaited. Facts in the lab base: stock dtb has qcom,msm-id <0x124 0x20001>, compatible qcom,msm8998-mtp qcom,mtp, no board-id/pmic-id, 376 KB wit…
FunctionFS with CDC-ACM descriptors and a tiny static aarch64 helper: the interactive console channel
google-walleye-lab доказано тестом a research row with the helper's design: the ffs descriptors and strings (CDC union/header/call-management/ACM functional descriptors, one interrupt-in and two bulk endpoints), the init steps (mount ffs, the order of writing descriptors before the UDC bind, as Android's own ffs.adb init composes it, read-only look at the public init.*usb*.rc), the bridge to a shell on a pty, how the host sees /dev/ttyACM0; the kernel's own GPL tools/usb/ffs-test.c as the base; which image tells, and its failure modes (panics, no enumerate)Answered: research walleye-ffs-acm-helper-how-it-stands, decision 2026-11-17-walleye-ffs-acm-single-interface; build item walleye-ffs-helper-build, phone item walleye-ffs-helper-try.
Write the FunctionFS ACM helper (C, static aarch64, GPL base ffs-test.c) and the n3 ramdisk/image
google-walleye-lab доказано тестом builds static aarch64 helper and image n3 with check-bootimg counting ok; tests/host/ffs-acm-up.sh exists and fails without the phoneBuilt: images/vendor-gadget-n3.img sha256 6672eb844fb5435ebb336e33a0b59db069bc2e62a6e01f98d760f5646b1d214c (helper, ramdisk sums in the session log 2026-11-17-builder-vendor-gadget-n3-image); 21 check…
Rewrite the FunctionFS ACM helper in Rust (static aarch64-unknown-linux-musl), the C one stays a probe
google-walleye-lab доказано тестом a Rust crate in the lab (build/ffs-acm-rs or as the lab's layout has it) built with lab-tools health/build.sh's way (rustup in ~/.local/share/devlab-cross, aarch64-linux-gnu-gcc as the linker, cargo build --release --locked --target aarch64-unknown-linux-musl), the same descriptors/steps as the C helper (decision 2026-11-17-walleye-ffs-acm-single-interface), a new image n3r differing from n3 only in the helper binary; checks as for n3; hashes; cargo fmt/clippy clean; no arguments or environment, so no argenv; the keeper's try then repeats n3's proof (tests/host/ffs-acm-up.sh)Built: build/ffs-acm-rs (Rust, libc, musl aarch64, fmt/clippy clean, descriptors byte for byte the C probe's by a test), build/build-vendor-gadget-n3r.sh (21 checks ok, reproducible). sha256 ffs-acm-r…
Phone: boot image n3 once; host sees ttyACM and a shell answers
google-walleye-lab доказано тестом tests/host/ffs-acm-up.sh ждёт host: cdc_acm for the running kernel (reboot the lab host in…n3/n3r tried (lab PR 76): cdc_acm refuses a FunctionFS interface without functional descriptors: parked (decision 2026-11-17-walleye-ffs-console-parked).
Phone: boot image n3r (the Rust FunctionFS ACM helper) once; host sees ttyACM and a shell answers
google-walleye-lab доказано тестом tests/host/ffs-acm-up.shTried once (n3r on slot b): the gadget enumerates but the host's cdc_acm refuses it (Zero length descriptor references, -22); f_fs cannot carry the CDC functional descriptors it needs. Parked: decisio…
The first shell of our own mainline kernel: read y2 over /dev/ttyACM0 - the whole dmesg first, then the facts
google-walleye-lab доказано тестом A script tests/host/mainline-shell-read.sh (new, in the lab, selftested with a fake tty) opens /dev/ttyACM0 as the lab user the moment it appears, sends commands one at a time with a read timeout, and saves under the lab material path (never into git without redaction): (1) dmesg in full - the first mainline kernel log of this phone; (2) uname -a, /proc/cmdline, /proc/cpuinfo, /proc/iomem, /proc/interrupts, /proc/version; (3) the device list: ls /sys/bus/platform/devices, deferred probes (/sys/kernel/debug/devices_deferred if debugfs is mounted), /sys/class/udc, /sys/class/power_supply, /sys/class/drm, /sys/class/input, /sys/class/regulator; (4) touch /tmp/hold to keep the phone up while reading, then reboot -f. Nothing is written to the phone beyond /tmp. The research row names what the dmesg shows works, what defers or fails (display, touch, UFS, charger, regulators): the real map of the next parts to bring up.proven by reader pass: 13 answers, dmesg 336 lines; fixed reader's end marker matching the phone's echo
Split the gadget bring-up into three images, one step each: configfs+gadget dir, +NCM function and config, +UDC bind
google-walleye-lab доказано тестом g5a, g5b, g5c = gadget4's init (kmsg marker, 30 s heartbeat, reboot -f at 120 s) plus steps: g5a mounts configfs and creates the gadget dir with idVendor/idProduct/strings (our serial walleye-first-boot, product 'Pixel 2 walleye (mainline)') but no function; g5b adds the ncm function and the config symlink, no UDC bind; g5c adds the UDC bind (a800000.dwc3) and the host-side network address/telnetd; checks ok as for gadget4; hashes in SHA256SUMSg5a, g5b, g5c built and checked (build/build-vendor-gadget5.sh, 14 checks each); hashes in SHA256SUMS and the log 2026-11-17-walleye-vendor-gadget5. Tries queued as walleye-try-gadget-bisect.
Get the host to bind the gadget (ecm/rndis variants)
google-walleye-labDropped: ECM and RNDIS are not in the vendor kernel config (decision walleye-host-channel-order). Replaced by walleye-massstorage-window and walleye-ffs-helper.
Write the failing check walleye-gadget-up in the lab's tests/health
google-walleye-lab доказано тестом the check exists in google-walleye-lab/tests/health, runs, and fails with the phone in fastboot (test: walleye-gadget-up)Done: google-walleye-lab tests/host/gadget-up.sh (host-side: the phone cannot run it; GADGET_WAIT=90 waits). Finds the gadget by our strings (serial walleye-first-boot, product 'Pixel 2 walleye (mainl…
walleye's gadget initramfs answers on 172.16.43.1, not surya's 172.16.42.1
google-walleye-lab доказано тестом images g and h (e and b with the initramfs address changed, nothing else) built, checked by the same check-bootimg, SHA256SUMS updated; images a-f byte-identical to beforeLanded google-walleye-lab #8, #9 (main 4eb1a74). g=e, h=b, i=f with the ramdisk built NET=43; check-bootimg 22 ok each; a-f sha256 identical to before; QEMU selftest PASS at 42 and 43. Tries queued: w…
Build v7.3 with a minimal msm8998-google-walleye.dts
google-walleye-lab доказано тестом make qcom/msm8998-google-walleye.dtb builds with no dtc warning from our nodes, the dtb holds msm-id 0x124/0x20001 and no /chosen/bootargs, and Image.gz builds (test to write in lab-tools: walleye-dtb-check)Built 2026-11-17: dt/msm8998-google-walleye.dts (lab PR 3) + v7.3 + build/walleye.config; counted: dtb builds, no dtc warning at W=1 from our file, msm-id 292/131073 (0x124/0x20001), 0 bootargs/stdout…
vendor-gadget4: an idle init (no gadget): kmsg marker, sleep 120, reboot -f
google-walleye-lab доказано тестом vendor-gadget4.img = the control with only the ramdisk swapped; the keeper's try shows whether the phone panics at the same gapAlready built and tried as vendor-gadget4 (log 2026-11-17-walleye-vendor-gadget4; research 2026-11-17-walleye-gadget4-idle-init-returns: returns at 156 s, the ramdisk runs, skip_initramfs absent). Clo…
Restore the first state: boot Android on slot a to completion, adb reboot bootloader, control once
google-walleye-lab доказано тестом devlab's before/after slot report (a successful yes before the try) and the fastboot text of the control try, accepted (phone leaves fastboot) or refused; the phone back on a bootable slotproven by devlab fastboot state: control accepted from restored first state (walleye-first-state-control-accepted); after the author's power-cycle Android booted on slot a, adb reboot bootloader, fast…
If refused: cold restart, then the same try on the same slot, with host checks
google-walleye-lab доказано тестом devlab's before/after slot report and the fastboot text of one control try after the cold restart, with the fastboot version and port in the log; the phone back on its start slotNot needed: the Load Error was the slot state after repeated refusals (set-active metadata fixes it); the host restart for the host's own modules was the lab reboot into 7.3.4, done by the author.
What the Pixel 2's bootloader dislikes in our boot image (Load Error), found from public sources, then the image variants to try
google-walleye-lab доказано тестом a recommended image recipe, each step cited, and the next 2-3 variants to build (research row), consistent with the tries recorded in walleye-first-fastboot-boot-tries*Answered: research walleye-load-error-how-it-stands, decision 2026-11-17-walleye-load-error-control-first; steps walleye-loaderror-* (stock-ids, variants-build, control, variants-try, dtbo-stub).
Control try: fastboot boot of a public walleye stock-format image (LineageOS recovery or TWRP) with devlab's before/after slot report
google-walleye-lab доказано тестом walleye-fastboot-boot-control: devlab report shows the phone left fastboot (18d1:4ee0 gone for 90 s polls) or Load Error, recorded either way as a research rowproven by research walleye-control-e-f-tries: control, e and f all Load Error, flags unchanged
Read walleye's stock boot header and dtb ids (qcom,board-id, pmic-id, msm-id) from a public LineageOS walleye build, host only
google-walleye-lab доказано тестом walleye-stock-ids-recorded: research row naming the header fields and the dtb's board-id/pmic-id/msm-id, each with the file hashRecorded in research walleye-stock-ids-recorded; control image at ~/Projects/walleye-build/control/boot.img (sha256 48dbc55c...7468), dtbo.img beside it (fb09009b...b8f6).
Build variants e,f: blueline recipe (second 0, os word 0) and e plus the stock board-id/pmic-id in the dts; check-bootimg counts both
google-walleye-lab доказано тестом check-bootimg.sh passes for e and f and counts board-id/pmic-id in fBuilt, lab PR 6 on main: ~/Projects/walleye-build/images/boot-e.img 112ab284...2daf and boot-f.img 7a42e5bd...461a (sha256 in images/SHA256SUMS; a-d unchanged). e = a with the command line cut to earl…
Try e then f (one at a time) with fastboot boot, after the control, with before/after slot report
google-walleye-lab доказано тестом walleye-fastboot-boot-variants: each image's outcome recorded as research; success is a USB id other than 18d1:4ee0 appearing (our gadget) for e or fproven by research walleye-control-e-f-tries: control, e and f all Load Error, flags unchanged
First light of OUR mainline v7.3 kernel: the control's header/footer, our kernel (gzip and lz4 variants) + the overlay-symbols dtb + a gadget init that shows itself in lsusb
google-walleye-lab доказано тестом images ml1 (kernel as Image.gz) and ml2 (kernel as lz4 frame like the stock) built by script from the control's own parts with exactly: our kernel, our -@ dtb (with the overlay stubs), our ramdisk; the ramdisk is a mainline-capable gadget init (mainline has devtmpfs and configfs: mass_storage or NCM-with-attributes are fine there, but keep it as simple as the vendor-kernel m1: enumerate with serial walleye-first-boot, product 'Pixel 2 walleye (mainline)', a 2 MiB mass_storage LUN window; open /dev/watchdog and never pet it (the SoC watchdog then resets the phone ~30-60 s later; if the dts needs the apss watchdog node enabled, enable it); cmdline: the stock cmdline except what mainline must have (earlycon msm_serial_dm,0xc1b0000 clk_ignore_unused ignore_loglevel panic=5 plus the stock androidboot.* kept), check-bisect-like checks ok, hashes in SHA256SUMS; the keeper's try then tells by lsusb (needs no host permissions): a device with our strings = first mainline lightBuilt: ml1 and ml2 (see the decision 2026-11-17-walleye-mainline-light-images); try queued as walleye-try-mainline-light.
m1: a mass_storage gadget as a log window (device writes a file, the host reads the raw LUN)
google-walleye-lab доказано тестом vendor-gadget-m1.img = the control with only the ramdisk swapped; init as n2's up to the UDC bind but with function mass_storage.0 (lun.0/file = a 2 MiB file in the ramdisk filesystem created by dd, ro=1, removable=1) instead of ncm, NO attribute writes beyond file/ro/removable/cdrom/nofua (check each exists in this kernel's mass_storage configfs, read the stock config facts); every 5 s the init rewrites the file in place (dd conv=notrunc) with: kmsg-marker, uname, cmdline, dmesg tail, mounts, heartbeat counter; reboot -f at 180 s; checks as for n2; hashes in SHA256SUMS; the host-side reader script tests/host/massstorage-read.sh (find the LUN by our strings via sysfs vendor/model, read with dd iflag=direct, print the text) in the lab repoBuilt: vendor-gadget-m1.img a3cf08d3..., initramfs dd59bce1... in ~/Projects/walleye-build/images (SHA256SUMS), 22 checks ok (build/build-vendor-gadget-m.sh); host reader tests/host/massstorage-read.s…
ml3: ml1 with the ramoops node at the phone's own region (0xa1100000, 2 MiB, vendor record/console sizes), so the stock Android can read our kernel's log after the reset
google-walleye-lab доказано тестом boot-ml3.img = boot-ml1.img with only the dts ramoops node changed: reg <0x0 0xa1100000 0x0 0x200000>, record-size 0x80000, console-size 0x20000, no pmsg and no ftrace region unless the vendor values say so (read /sys/module/ramoops/parameters on Android a, and the vendor dts' reserved-memory node names in the public GPL dts, for pmsg/ftrace sizes and for whether 0xa1100000 collides with another carve-out in our dts: say it), checks as for ml1/ml2, hashes; also read-back: the new dtb's ramoops node printed with fdtget in the check; the kernel's console is saved into it by pstore/console (check CONFIG_PSTORE_CONSOLE and PSTORE_RAM are y in our config)Built 2026-11-17: boot-ml3.img sha256 5c2574706882f5c8f40b134bdf501e22c092093e494b5aad25efdabe4ba105c4 in ~/Projects/walleye-build/images-ml3; ramoops@a1100000 checked with fdtget, no carve-out overla…
ml4: a timing channel from our mainline init - the reset delay says how far the kernel came, no UART, no USB, no pstore needed
google-walleye-lab доказано тестом boot-ml4.img = ml3 with ONLY the ramdisk's /init changed (kernel, dtb and header bytes equal ml3's): /init mounts proc/sys/devtmpfs, opens /dev/watchdog (a safety net: the SoC resets by itself ~30 s later if the rest hangs), counts which of a fixed checklist holds - (1) init reached userspace [always 1], (2) /sys/class/udc has an entry, (3) a dwc3 platform device is bound to its driver, (4) /sys/bus/platform/devices has a ramoops device, (5) /proc/iomem lists the ramoops region, (6) the PSCI/reset path answers (/sys/firmware/devicetree/base/psci exists), (7) UFS host is not enabled (expected 0) - then sleeps 5*score seconds (bins of 5 s) and runs reboot -f (sysrq b fallback); the checklist and the bin table written in the item/log; checks as ml3's plus: the init script has no bashisms, runs under busybox ash -n; a QEMU-free host test reading the bin table. The keeper's try then reads the time Android is back (host timestamps at fastboot-boot, at adb) and subtracts the known base (the control's: boot+Android a ~ 25-36 s); 70 s = the safety watchdog fired: init never ran or hung before the sleepBuilt: images-ml4/boot-ml4.img sha256 105364591ce1f46a2a828caf1f80c851399487ab4a59bd777210a9a72e13ea84; try queued as walleye-try-ml4
ml5: timing bins that finish BEFORE the dead-kernel reset (~33 s): a baseline image that reboots at once, and a score image with 3 s per point
google-walleye-lab доказано тестом boot-ml5a.img = ml4 with /init changed to: mount proc/sys, kmsg marker, then reboot -f at once (sysrq b fallback), no watchdog games; boot-ml5b.img = ml4 with the sleep scaled to 3 s per score point (max 7 points = 21 s; the watchdog petting kept) and the bin table in tests/host/ml5-bin-table.sh; both with kernel/dtb/header bytes equal to ml3's; checks as ml4's; hashes in SHA256SUMS; the table states the expected return times: base = the time ml5a returns (fastboot boot to adb), ml5b = base + 3*scorebuilt: images-ml5, hashes in research walleye-ml5-built; try walleye-try-ml5 queued
Build ncm-noattr images n1, n2: the 5b4 steps and the bind, without the dev_addr/host_addr writes
google-walleye-lab доказано тестом check-bisect passes: grep for dev_addr|host_addr|qmult|ifname finds none in n1/n2 init, exactly the step's parts, hashes in SHA256SUMSBuilt: vendor-gadget-n1.img b2e96b42..., vendor-gadget-n2.img b007d1e3... in ~/Projects/walleye-build/images, 17 checks each ok, hashes in SHA256SUMS there and in session log 2026-11-17-walleye-vendor…
Why the ncm function (or its config symlink) panics the vendor msm-4.4.302 kernel, and which gadget function is safe and gives a channel
google-walleye-lab доказано тестом a research row with sources (the GPL wahoo kernel's f_ncm.c, configfs.c, the gsi/ipa gadget functions, how Android's USB HAL on Pixel 2 composes its gadgets: ffs.adb, mtp, rndis.gsi/ncm?, what the vendor's init.usb scripts write, public sources only): the likely cause ranked, the safe alternatives for a host-visible channel on this kernel (FunctionFS with a tiny static aarch64 helper, gsi functions, mass_storage with a read-only image as a crude log window, MTP, ACC), each with how one cheap image tells it, and a recommendationAnswered: research walleye-ncm-panic-why-how-it-stands, decision 2026-11-17-walleye-ncm-without-attributes
Split the NCM step of the gadget bring-up: function dir only, then its attributes, then the config symlink
google-walleye-lab доказано тестом images 5b1 (mkdir functions/ncm.usb0 only), 5b2 (5b1 + writing the function's attributes it has, e.g. host_addr/dev_addr/qmult, if the init wrote any), 5b3 (5b2 + the symlink into configs/b.1), all on gadget5a's init (configfs + gadget dir + strings + the usual 120 s reboot, heartbeat, kmsg markers BEFORE each step), checks as for 5a-c; hashes in SHA256SUMSBuilt: vendor-gadget5b1 (mkdir functions/ncm.usb0), 5b2 (+dev_addr, host_addr), 5b3 (+configs/c.1 and its string), 5b4 (+symlink = 5b); four steps not three, so the config dir is told apart from the s…
Every mainline image carries earlycon msm_serial_dm,0xc1b0000: on a phone whose debug UART is not enabled that is a likely first-printk hang. Build the same images WITHOUT earlycon/console on the UART
google-walleye-lab доказано тестом pk3 = pk1 and ml6a = ml5a with ONLY the kernel command line changed: earlycon=... and any console=ttyMSM*/ttyHS* removed (say what the cmdline was and what remains; keep clk_ignore_unused ignore_loglevel panic=5 and the stock androidboot.* pieces); kernel/dtb/ramdisk/header/footer bytes equal to pk1's and ml5a's; checks as before; hashesbuilt: pk3/ml6a in walleye-build/images-nocon, 22 checks ok, hashes in research/walleye-first-fastboot-boot/2026-11-18-nocon; try queued as walleye-try-no-earlycon
Find the node(s) of OUR dtb that stop boot: diff it against porthole's taimen dtb (the one that reaches init with our kernel), bisect by moving nodes
google-walleye-lab доказано тестом a research row naming the node or property whose presence in our dts stops boot (reset at ~33 s), found by images each with our dtb minus a node group (reserved-memory carve-outs, ramoops, the watchdog node we added, regulators, the dwc3/usb nodes, tlmm...), or porthole's dtb plus a node group of ours; the fix as a patch to dt/msm8998-google-walleye.dts; hashesbuilt: d1..d6 in walleye-build/images-dtb, 90 checks ok, hashes in research/walleye-first-fastboot-boot/2026-11-18-dtb-bisect; try queued as walleye-try-dtb-bisect
The cause of our dead dtbs: pinctrl@3400000 lacks gpio-reserved-ranges <0 4>, <81 4> (TZ-protected GPIOs). Put it into OUR dts and build images y1 (init: reboot at once) and y2 (the gadget: ACM console + mass_storage window)
google-walleye-lab доказано тестом (1) dt/msm8998-google-walleye.dts gets the property exactly as in the alive trees and the running fdt (pinctrl@3400000 gpio-reserved-ranges = <0 4>, <81 4>), with a comment naming where it comes from (the phone running fdt, vfa) and why (those pins belong to the secure world: touching them resets the phone) - and the other missing items from walleye-fdt-compare that are cheap and certain (restart@10ac000 qcom,pshold, the apps watchdog) only if the finding says they matter, one commit each so the effect is attributable; (2) images y1 = our kernel (the ml6a one, no earlycon) + OUR dtb with the property + the ml5a immediate-reboot init; y2 = the same + the mainline gadget init of pk5l (ACM shell on ttyGS0 + mass_storage window, watchdog petted 180 s, reboot -f at the end) so the first USB light can appear; expected: y1 alive (<=66 s), y2 shows 1d6b:0104 on the host within ~25 s of the jump (plain lsusb); checks, hashes, programmer review; build recipe reproducible; (3) keep dtbo-overlay-merges.sh passingBuilt: y1 and y2 by build/build-y.sh (49 checks ok, rebuild identical, hashes in research walleye-y-built); try queued as walleye-try-y (phone, rank 1).
What does the stock dtbo overlay (entry 3) switch on in OUR dtb that it does not in porthole's taimen dtb? Analyse it on the host, then neutralize label by label
google-walleye-lab доказано тестом (1) a research row: fdtoverlay (or ufdt-equivalent) merges entry 3 of the control's dtbo.img into (a) pk5's taimen dtb (works) and (b) our dtb (dead), the merged dtbs decompiled and diffed: the nodes/properties the overlay sets or adds (status okay, panel, touch, regulators, clocks, pinctrl, charger...) that exist in ours as real SoC nodes but are absent or differently wired in taimen's; ranked as suspects (a driver that probes unclocked hardware and hangs the SoC is the pattern: a reset ~33 s after the jump); (2) images L1..Ln = our dtb (the ml6a one) with the __symbols__ entries of groups of those labels pointed at harmless dummy nodes so the overlay lands on the dummy and the real node is not enabled (the merge still succeeds: tests/host/dtbo-overlay-merges.sh passes), ramdisk = ml5a's init, cmdline as ml6a; checks, hashes; the bisect plan (binary halving of the label set)analysed: overlay inert on our dtb; images w1..w5, wr, wa built (research walleye-overlay-enables-what-result, walleye-taimen-walleye-built); try queued
Our images e, then b, by fastboot boot from the first state (Android booted on slot a, then adb reboot bootloader)
google-walleye-lab доказано тестом devlab's before/after slot report and the phone's full fastboot text for each; lsusb polling (by the manifest serial, not VID:PID: surya's gadget 18d1:d001 is on the same host) shows whether a Linux gadget of ours appears (lsusb on the host within 60 s, 172.16.42.1 answering: test walleye-gadget-up)proven by devlab fastboot boot now: e and b accepted from the first state, neither started, slot flags unchanged; see walleye-own-images-e-b-first-state
Why the vendor 4.4.302 kernel panics about 60 s after fastboot boot with our ramdisk, and how to read the panic text
google-walleye-lab доказано тестом a research row with sources: the likely causes ranked (subsystem restart level panic after PIL/SSR timeouts, MSM watchdog bark/bite not petted, rcu/hung-task, an init exit, vendor ramoops handling), how each is told apart by one cheap try, and how the previous boot's ramoops/console text can be read (vendor init.ramoops.sh encrypts it into /data/vendor_ce/0/ramoops after user unlock: what exactly triggers it, whether an unlock by the author after the try is enough, other routes: /proc/last_kmsg, bootloader log, UART cable with its cost and the fastboot oem uart question)Facts in the lab base: research walleye-vendor-kernel-facts-s1-panic (config.gz facts, ramoops 0xa1100000 2 MiB, encryption), the tries of vendor-gadget (gzip, no /init), gadget2/3 (xz): kernel_panic…
Measure when the panic happens: T_android from a plain adb reboot against the gap of a panic try
google-walleye-lab доказано тестом test walleye-panic-time: the 1 s watcher log gives T_android (reboot) and the try's gap; panic time = gap - T_android, written as a research rowNot needed now: the panic belongs to the gadget bring-up (gadget4 idle init does not panic); the bisect of the bring-up (walleye-gadget-bisect) is the next measure.
Stage the mainline gadget bring-up on the ALIVE base (pk5: our kernel + porthole's taimen dtb): one step per image, each holds 60 s so alive and dead are 30 s apart
google-walleye-lab доказано тестом images g1..g3 on pk5's kernel/dtb/cmdline/header/footer (bytes equal), ramdisk init: mount proc/sys/devtmpfs/configfs, then g1 = gadget dir + strings (serial walleye-first-boot, product 'Pixel 2 walleye (mainline)') and nothing else; g2 = g1 + mass_storage.0 function (2 MiB file LUN, ro) linked into a config, NO UDC bind; g3 = g2 + UDC bind to the first /sys/class/udc entry (expects USB enumeration on the host as 1d6b:0104 within ~25 s of the jump: the first mainline light, readable with plain lsusb); every init pets /dev/watchdog and holds 60 s then reboot -f (sysrq b fallback); no ACM, no shell loop (pk5l had them and ended early); expected returns: alive (init ran) ~12+60+36 = ~108 s, dead ~75 s (outside watchdog), early panic with panic=5 ~45-60 s; checks, hashes; table in the itembuilt: g1..g3, 58 checks ok, rebuild identical; try queued as walleye-try-pk5-stages
First light of OUR kernel: pk5 (our kernel + porthole's taimen dtb, no earlycon) with a gadget init: configfs ACM console (a shell on /dev/ttyGS0) and the mass_storage log window
google-walleye-lab доказано тестом boot-pk5l.img = pk5's kernel, dtb, header, footer, cmdline (bytes equal), ramdisk = a mainline gadget init: mount proc/sys/devtmpfs/configfs, gadget with strings serial walleye-first-boot / product 'Pixel 2 walleye (mainline)', functions acm.GS0 (shell: busybox ash on /dev/ttyGS0 through a getty-less loop, respawning) and mass_storage.0 (2 MiB file window rewritten every 5 s with uname, cmdline, dmesg tail, /proc/iomem, /sys/class/udc, pstore), UDC bound to the first /sys/class/udc entry; /dev/watchdog opened and petted for a bounded 180 s, then reboot -f (sysrq b fallback); a host test tests/host/mainline-acm-up.sh (finds the device by our strings, waits for /dev/ttyACM*, sends 'uname -a; cat /proc/cmdline', expects the answer; and reads the LUN window via massstorage-read.sh); checks as before; the mainline kernel config must have USB_CONFIGFS_ACM, MASS_STORAGE, DWC3 peripheral and the QCOM glue (say if our config lacks any; porthole's dtb carries its dwc3 node: say what the kernel needs from it)built: boot-pk5l.img in walleye-build/images-pk5l, 39 checks ok, hashes in research/walleye-first-fastboot-boot/2026-11-18-pk5l; try queued as walleye-try-pk5l
pk5l2: pk5l with the dtb's ramoops node moved to the phone's own region, so the panic that ends pk5l after ~5-20 s can be read from Android's pstore
google-walleye-lab доказано тестом boot-pk5l2.img = pk5l with ONLY the ramoops node of the (porthole taimen) dtb edited (fdtput/dtc on the dtb, no kernel or init change): reg 0xa1100000 size 0x200000, record-size 0x80000, console-size 0x20000, pmsg-size 0 (as in our decision 2026-11-17-walleye-ramoops-at-the-phones-region), the old node/region removed or checked for overlap; fdtget read-back in the checks; kernel config has PSTORE_CONSOLE and PSTORE_RAM (say); hashes; recipe in build/built: boot-pk5l2.img in walleye-build/images-pk5l2, 20 checks ok, hashes in research/walleye-first-fastboot-boot/2026-11-18-pk5l2; try queued as walleye-try-pk5l2
A known-good mainline kernel on this phone: porthole's 7.2.2 Pixel 2 XL kernel (their config, patches, dts) with our ml5a-style immediate-reboot init, to see whether ANY mainline kernel reaches init here
google-walleye-lab доказано тестом pk1/pk2 images: the control's header/footer, porthole's kernel built from its public pmaports package (kernel version, config and patch series as they publish; taimen dts for pk1, their walleye dts for pk2), their appended dtb with the overlay symbols if their dtb lacks __symbols__ (check with tests/host/dtbo-overlay-merges.sh and say), our ml5a init (reboot at once) as the ramdisk; checks as ml5's; build recipe reproducible from the lab; sources and licences kept; hashesbuilt: pk1/pk2 in walleye-build/images-pk, checks ok, hashes in research/walleye-first-fastboot-boot/2026-11-18-pk; try queued as walleye-try-pk
Property-level bisect between wa (alive) and pk2 (dead): same node set and statuses, so a PROPERTY of pk2's dtb (or ours) stops boot
google-walleye-lab доказано тестом (1) a research row: wa and pk2 decompiled (dtc -I dtb -O dts -s) and diffed property by property: every differing property listed by node (compatible, model, msm-id/board-id, memory reg, reserved-memory regions and their no-map/reusable, cpu capacities/idle-state, timer frequency, psci, chosen, pinctrl/regulator limits, clock rates, interrupts...), grouped (top-level, memory+reserved-memory, cpus, soc nodes...); (2) images wp1..wpn = wa with ONE group of pk2's property values copied in (binary-halving plan: first half/second half of the groups, then narrower), same kernel/ramdisk/cmdline as wa; checks, hashes; each states its expected return (alive <=66 s, dead >=69 s). Also build the mirror from OUR dtb: o1.. = our dtb with one group of wa's values copied in, if the first plan leaves the culprit in a group that is ours alonebuilt: wp1..wp5, wpa (68 checks, rebuild identical), research walleye-property-diff and walleye-property-bisect-built; tried by walleye-try-property-bisect; mirror images from our dtb only if the caus…
Image T1: init marks /dev/kmsg, waits 20 s, sysrq c (panic=5); read console-ramoops after the return
google-walleye-lab доказано тестом tests/health: pstore holds the marker after the T1 return (expected absent: records cause B)Built: images-t1/boot-t1.img (research walleye-t1-built); try queued as walleye-pstore-t1-try
Try T1: fastboot boot boot-t1.img, wait 90 s, read console-ramoops for the marker
google-walleye-lab доказано тестом tests/host/pstore-t1-read.shran 03:52 slot b: boot-t1.img back as Android after 159 s; battery 4384 mV 100% before; /sys/fs/pstore empty, no ramoops dir; marker absent = cause B
Image T3: init opens /dev/watchdog (kpss-wdt) with a short timeout and never pets it; read console-ramoops after the bite
google-walleye-lab доказано тестом tests/host/pstore-t3-read.sh: pstore holds the T3 marker, last tick under 15, no no-bite line (expected present)Tried once (lab PR 122): back at 42 s, no marker: inconclusive without a witness; the pstore chase is paused by decision 2026-11-18-walleye-vendor-kernel-as-instrument.
Why our mainline kernel's console never survives in pstore, and how a MSM8998 kernel asks for a WARM reset (reboot=w, psci reset, qcom pon, the SDI/dload) so the ramoops region keeps its text
google-walleye-lab доказано тестом a research row with sources (porthole's cmdline reboot=w and qcom_wdt.arm_on_probe, kernel docs, postmarketOS and kcxt notes, Qualcomm PSCI/TZ behaviour, the vendor kernel's own warm-reset path read from the GPL source): ranked causes of the empty pstore (DDR cleared by a cold reset, the bootloader zeroing the region, the ramoops node/size mismatch, the vendor reader consuming it), and one or two cheap image changes that would tell (cmdline reboot=warm, the init doing a psci warm reset, panic=5 plus a deliberate panic via sysrq c at a known time) with their expected pstore resultsAnswered in research walleye-pstore-warm-reset-how-it-stands and decision 2026-11-18-walleye-warm-reset-by-watchdog-bite-first; tries queued as walleye-pstore-t1/t3/t4
Try n3 (FunctionFS ACM) again now the host runs 7.3.4 (cdc_acm loads): a console on /dev/ttyACM0
google-walleye-lab доказано тестом tests/host/ffs-acm-up.sh passes: /dev/ttyACM0 appears on the host, a shell answers on it (uname -a, cat /proc/cmdline read-only), the phone returns by itself (the init's reboot) ждёт the lpoco lab-prepare tool (host access to /dev/ttyACM*); th…Superseded by walleye-ffs-rust-try: the console try uses the Rust helper image n3r (the C n3 is a probe).
Try m1 again now the host runs 7.3.4 (usb_storage loads): read the first text the phone writes
google-walleye-lab доказано тестом massstorage-read.sh (tests/host) prints the init's window text from the phone: uname, cmdline, mounts, dmesg tail of the vendor kernel with our init running, the heartbeat counter growing; kept as material in the lab and recorded as research with what it showsproven by tests/host/massstorage-read.sh pass: gadget 3-2 /dev/sda WALLEYE-LOG m1 beat=4 up=21.34s; research 2026-11-17-walleye-m1-window-read. lpoco-host acceptance: PASS. Next: walleye-ffs-rust-try…
CI runs cargo fmt --check, clippy -D warnings and test for build/ffs-acm-rs
google-walleye-lab доказано тестом check.yml rust job green on the PR; read back from mainHalf of walleye-code-hygiene-rust-ci, done by walleye-builder: fmt, clippy -D warnings and test pass locally on the lab (1 test ok); job added to .github/workflows/check.yml
Phone: one image, stock kernel + our dtb with overlay symbols (s3 redone); read the result
google-walleye-lab доказано тестом TESTproven by fastboot boot: left fastboot at 12 s, Android on a at 36 s, flags unchanged; see research walleye-s4-overlay-symbols-try
Phone: boot pk3 once, then pk4, pk5, pk6 twice each from the first state; record the seven return times and read pstore
google-walleye-lab доказано тестом TESTproven by the seven returns in walleye-pk-bisect-returns: our dtb is the fault
Try bisect image s0n (stock parts without AVB footer (control of the footer)) by fastboot boot from the first state
google-walleye-lab доказано тестом walleye-gadget-up for s1; for s2, s3, s0n: the phone leaves fastboot (starts) or stays at 'Booting OKAY', recordedproven by s0n: left fastboot after 12 s, then off USB, same as s0; no writes
Try bisect image s1 (ramdisk = ours) by fastboot boot from the first state
google-walleye-lab доказано тестом walleye-gadget-up for s1; for s2, s3, s0n: the phone leaves fastboot (starts) or stays at 'Booting OKAY', recordedno start, fell back to a; see walleye-try-s1-s2-s3
Try bisect image s2 (kernel = ours (lz4 frame)) by fastboot boot from the first state
google-walleye-lab доказано тестом walleye-gadget-up for s1; for s2, s3, s0n: the phone leaves fastboot (starts) or stays at 'Booting OKAY', recordedno start, fell back to a
Try bisect image s3 (dtb = ours) by fastboot boot from the first state
google-walleye-lab доказано тестом walleye-gadget-up for s1; for s2, s3, s0n: the phone leaves fastboot (starts) or stays at 'Booting OKAY', recordedaccepted, not started, stays in fastboot
Phone: boot pk4 once more, then d1..d6 twice each from the first state; record the thirteen return times and read pstore
google-walleye-lab доказано тестом TEST13 tries, all 75 s (no init); no group of d1..d6 is the cause; recorded walleye-dtb-bisect-tried
Phone: boot x4, x1, x2, x6, x5, x3, x7, xa twice each from the first state; record the sixteen return times
google-walleye-lab доказано тестом TESTproven by 16 boots: x1 dead twice (79,78 s), xa dead (78,76), x2-x7 alive. Next: add gpio-reserved-ranges <0 4>,<81 4> to our dtb and try (alive expected)
Try image g (e at 172.16.43.1) by fastboot boot
google-walleye-lab доказано тестом walleye-gadget-up: after devlab fastboot boot now IMAGE the phone leaves fastboot, lsusb shows our gadget (1d6b:0104, walleye-first-boot) within 60 s and 172.16.43.1 answers ping and telnet; devlab's before/after slot report and the fastboot text keptNot needed: the vendor kernel ran with our ramdisk (research walleye-vendor-kernel-facts-s1-panic); mainline images g/h wait until the vendor-kernel gadget (walleye-try-vendor-gadget2) gives a shell a…
Try vendor-gadget5a, 5b, 5c from the first state, in order; stop at the first panic
google-walleye-lab доказано тестом TESTproven by tries: 5a ok (reboot, 156 s), 5b kernel_panic (75 s): culprit is the ncm function/config step; 5c not tried per stop rule. Record 2026-11-17-walleye-vendor-gadget5-bisect
Try image h (b at 172.16.43.1) by fastboot boot
google-walleye-lab доказано тестом walleye-gadget-up: after devlab fastboot boot now IMAGE the phone leaves fastboot, lsusb shows our gadget (1d6b:0104, walleye-first-boot) within 60 s and 172.16.43.1 answers ping and telnet; devlab's before/after slot report and the fastboot text keptNot needed: the vendor kernel ran with our ramdisk (research walleye-vendor-kernel-facts-s1-panic); mainline images g/h wait until the vendor-kernel gadget (walleye-try-vendor-gadget2) gives a shell a…
Phone: boot ml1, then ml2, each once from the first state; lsusb tells first light
google-walleye-lab доказано тестом tests/host/mainline-light-up.shTried 2026-11-17: ml1, ml2 and (item walleye-try-ml3-ramoops) ml3 all fail: no USB light, Android on a at ~70 s. pstore read done, empty. Next: UART, the walleye coordinator's call.
Try m1 from the first state, then read the log window on the host for 90 s
google-walleye-lab доказано тестом TEST ждёт host: usb_storage for the running kernel (reboot into 7.3.4)m1 tried and its window read as the lab user (lab PR 70: lpoco-host acceptance passed); ml/pk variants superseded.
Phone: boot ml3 once from the first state, then read our kernel's console from the stock Android's pstore
google-walleye-lab доказано тестом tests/host/ml3-ramoops-read.sh (to write: pstore console of our boot read back)Tried 2026-11-17: ml3 no light, Android on a at 72 s, pstore empty (research walleye-ml3-pstore-empty). Proof test tests/host/ml3-ramoops-read.sh fails as the facts stand. The remaining read is UART (…
Phone: boot image ml4 twice from the first state; read the reset time as a score
google-walleye-lab доказано тестом TESTproven by two tries (75 s, 76 s, slot b, nothing flashed, pstore empty): bin table reads score 7 but is not told apart from the safety watchdog (ml1-ml3 back at 69-72 s); see research 2026-11-17-walle…
Phone: boot ml5a twice, then ml5b twice from the first state; record the four return times
google-walleye-lab доказано тестом TESTproven by the phone: ml5a 75 s and 75 s, ml5b 75 s and 78 s (fastboot boot to adb, from slot b, nothing flashed); ml5a ~70 s+ = the kernel never reached init; ml5b equals ml5a within 3 s, so no init r…
Try ncm-noattr n1 then n2 from the first state, stop at the first panic
google-walleye-lab доказано тестом TESTproven by n1/n2 tries: n1 returns, n2 binds without a panic and enumerates with our strings; host cdc_ncm fails on the MAC string (walleye-ncm-noattr-n1-n2). Next: a function the host binds without a…
Try vendor-gadget5b1, then 5b2, 5b3, 5b4 by fastboot boot from the first state; stop at the first that panics
google-walleye-lab доказано тестом TESTproven by b1 returns, b2 panics at 75 s (research 2026-11-17-walleye-ncm-split-b2-panics); next: the noattr try (n1, n2)
Phone: boot pk3 twice, then ml6a twice from the first state; record the four return times
google-walleye-lab доказано тестом TESTproven by four returns: pk3 63/48 s, ml6a 75/75 s; ml6a no init, pk3 undecided; next: coordinator picks the log channel (ramoops read / UART)
Phone: boot pk1 twice, then pk2 twice from the first state; record the four return times
google-walleye-lab доказано тестом TESTproven by four tries: pk1 75 s, 75 s; pk2 75 s, 75 s - no init ran
Phone: boot g1, g2, g3 (images-g) twice each from the first state, stop at the first that dies; lsusb polled by port every second for 90 s; pstore read after each
google-walleye-lab доказано тестом TESTproven by 6 tries: no gadget enumerates, g2 returns at 138 s; research walleye-pk5-stages-tried
Phone: boot pk5l twice from the first state; read the ACM console and the window, record the first mainline dmesg
google-walleye-lab доказано тестом TESTtried twice: no ACM, Android back at 57/42 s; recorded walleye-pk5l-tried; next walleye-try-dtb-bisect
Phone: boot pk5l2 twice from the first state; after each read /sys/fs/pstore/* and /data/vendor_ce/0/ramoops/*, record the first mainline panic or console text
google-walleye-lab доказано тестом TESTtried twice: no mainline panic text found (research walleye-pk5l2-tried); next: prove the mainline kernel reaches ramoops at all, or return to the dtb diff
Phone: boot wpa twice from the first state, then (if dead) wp1..wp5 twice each; record the return times (<=66 s alive, >=69 s dead)
google-walleye-lab доказано тестом TESTproven by wpa x2: 60 s, 60 s (alive); cause is not a property; next: phandle order / dtb size
Phone: boot wa twice from the first state, then w1..w5 and wr twice each; record the return times (~42 s init ran, ~75 s dead)
google-walleye-lab доказано тестом TESTproven by 12 tries, all alive (research walleye-w1-w5-wr-tried): cause is a property; next a property-level bisect
Try vendor-gadget2.img (the control with an xz gadget ramdisk) by fastboot boot from the first state
google-walleye-lab доказано тестом walleye-gadget-up (tests/host/gadget-up.sh, GADGET_WAIT=90): a USB device with our strings enumerates and 172.16.43.1 answers ping and telnet; else record the bootreasonSuperseded by walleye-try-vendor-gadget3: the same image without the self-reboot timer (decision walleye-try-returns-without-hands). Never tried.
Try vendor-gadget3.img (vendor-gadget2 plus a 240 s self-reboot) by fastboot boot from the first state
google-walleye-lab доказано тестом walleye-gadget-up (tests/host/gadget-up.sh, GADGET_WAIT=90): a USB device with our strings enumerates and 172.16.43.1 answers ping and telnet; and without hands the phone is back in fastboot or Android within ~300 s of the boot (or kept up with touch /tmp/hold over telnet); else record the bootreasonThird try done, watcher alive: kernel_panic before any gadget strings (research walleye-gadget3-third-try). Next: a kmsg-only init build (builder), then try it.
Try vendor-gadget4.img (idle init, no gadget) by fastboot boot from the first state
google-walleye-lab доказано тестом the bootreason and the time the phone is back on Android a after fastboot boot: panic at ~75 s = the panic is independent of our gadget code; a return at ~120 s with bootreason not kernel_panic = our init ran and the gadget bring-up of gadget2/3 is what panics; the init's /dev/kmsg marker is not readable from Android (record that)proven by the try: back on Android a at 156 s, bootreason reboot (not kernel_panic): our idle init ran; the gadget bring-up of gadget2/3 is what panics. Next: bisect the gadget bring-up - the coordina…
Controlled repeat: vf2 twice as before, then vf4 twice WITHOUT the poller, to tell the image from the way it was watched
google-walleye-lab доказано тестом four tries from the first state (battery first, Android on a good slot, adb reboot bootloader), each recorded with the time the gadget appeared (plain lsusb by port, no lsusb -v, no reads of the LUN until 70 s after the jump), the time Android is back, and whether the init's beat/step text appears in a single read at ~70 s of /dev/sdX (massstorage-read.sh as in vf2's tries): order vf2, vf2, vf4, vf4; the finding: does vf2 repeat its ~43 s gadget? does vf4 (identical init to vf2 up to the bind) behave like vf2 when watched the same way? if vf4 is alive unwatched and dead with the poller, the poller (lsusb -v or the LUN reads during enumeration) is what breaks the phoneproven by 4 tries: vf2 repeats (gadget ~50 s, empty header), vf4 dead twice unwatched; poller cleared; next: diff vf4 vs vf2 after the bind, bisect the dump
Phone: boot vf1 twice from the first state; the keeper reads the window with tests/host/vendor-facts-read.sh (battery first, phone left in Android afterwards); first run may be partial: say what was missing
google-walleye-lab доказано тестом TEST ждёт coordinator: how to make the window show life firstvf1 tried twice (lab PR 125): enumerates, LUN readable, no header: superseded by vf2.
Phone: boot vf2 twice from the first state; read the window with tests/host/vendor-facts-read.sh (step lines print first; battery first; phone left in Android afterwards); say whether the marker, the beat and which step the init reached
google-walleye-lab доказано тестом TESTproven by try 2: marker, beat and init steps seen; blob unread - see walleye-vf2-tried-twice
Phone: start tests/host/vendor-facts-poll.sh FIRST in its own background job (never killed by the keeper's cleanup), then boot vf3 twice from the first state; battery first; phone left in Android afterwards; say how many blocks the poller fetched, whether DONE came, and the beat log
google-walleye-lab доказано тестом TESTtried twice as asked: no gadget, no window, no blocks (poller fail x2); back as Android 78 s each; see walleye-vf3-tried-twice. Next: coordinator decides - vf2's order alive, vf3's reorder dead; diff…
Phone: start tests/host/vendor-facts-poll.sh FIRST in its own background job with VF_MARK=WALLEYE-VF2 VF_HEAD=16 VF_BLOCKS=1040 VF_OUT=$HOME/Projects/walleye-build/material/vf/2026-11-18-vf4 (never killed by the keeper's cleanup), then boot boot-vf4.img twice from the first state; battery first; phone left in Android afterwards; say how many blocks the poller fetched, whether DONE came, the beat log, steps.txt's last lines (where it stopped), and when the phone left USB
google-walleye-lab доказано тестом TESTproven by two tries: 0 blocks fetched, no gadget, back as Android at 78 s both; turning point for the coordinator: vf4 behaves like vf3 (dead), not vf2 - what differs between vf2 and vf4 before the bi…
Try vfa twice, then vfb, vfc, vfd once each (twice if one dies): which single hardware source reads, which kills
google-walleye-lab доказано тестом vfa twice, then vfb, vfc, vfd once each unless one dies (then twice), each recorded with the time the gadget appeared, the time Android is back, and the reader's one line (pass with the parts, or fail saying killed or hung); the finding: which source reads and which kills; /sys/firmware/fdt kept as fdt.dtb and fdt.dts in the lab material path if vfa passesproven by: vfa twice pass (fdt.dtb/dts kept in material path), vfb alive but clk_summary unreadable rc=1, vfc pass, vfd killed twice (research walleye-vfx-tried). Next: give mainline USB its nodes fro…
Phone: boot y1 twice, then y2 twice from the first state; battery first; read return times and the host's lsusb; leave Android on a good slot
google-walleye-lab доказано тестом TESTproven by y1 alive x2 (36 s) and y2 gadget 1d6b:0104 at 18 s x2, research walleye-y-tried
vf1: the vendor kernel with our init dumps the running hardware's facts into a bigger log window (device tree as merged, clocks, regulators, pinctrl/gpio, interrupts, iomem, dwc3/phy state, dmesg)
google-walleye-lab доказано тестом boot-vf1.img = the control's kernel/dtb/cmdline/header/footer (bytes equal, as m1) with a ramdisk whose init mounts proc/sys/debugfs, creates a tar of the chosen files (proc/device-tree as a dtc-readable copy via cat of the files or a single fdt if /sys/firmware/fdt exists, debugfs clk/clk_summary, regulator/regulator_summary, pinctrl/*, gpio, /proc/interrupts, iomem, ioports, /sys/class/udc, dwc3 debugfs, /sys/kernel/debug/usb, dmesg), gzips it, writes it as a length-prefixed blob into an 8 MiB file exported as a read-only mass_storage LUN (m1's gadget order: dir, mass_storage, bind), rewrites it in place every 10 s while it grows, holds 90 s then reboot -f; tests/host/vendor-facts-read.sh reads the LUN (uaccess rule exists), checks the length prefix and gunzips the tar into research/google-walleye-lab/2026-11-18-vf/ (not in git if large: put it under the lab's material path and keep the index); checks as m1's; no write to the phoneboot-vf1.img built, 23 checks ok, rebuild identical, selftest 9 ok, reviewed; try queued as walleye-try-vf1
Repack the control boot.img with our gadget initramfs (kernel, dtb, cmdline untouched)
google-walleye-lab доказано тестом check-bootimg counts the repack: kernel/dtb/cmdline/offsets byte-equal to the control, ramdisk is ours (test: walleye-vendor-gadget-image-built)Done: ~/Projects/walleye-build/images/vendor-gadget.img (= bisect s1), sha256 3f6587480d1f53fde918de8ebeef67641245e4ee7cbf9704e787075314f3b4e4, 44 bisect checks ok (lab #10, #11). Control header, os w…
fastboot boot the repacked control from the first state; does the gadget answer?
google-walleye-lab доказано тестом walleye-gadget-up passes (phone enumerates by serial, 172.16.43.1 answers) and the slot report shows what changedSuperseded by the bisect of the bring-up (gadget4 idle init ran, gadget5a ok, 5b panics at the ncm step): see research 2026-11-17-walleye-vendor-gadget5-bisect and walleye-try-ncm-split.
A gadget initramfs the vendor 4.4.302 kernel can run: gzip, aarch64, NCM, mknod, init that never exits (image vendor-gadget2)
google-walleye-lab доказано тестом vendor-gadget2.img = the control with only the ramdisk swapped (header, AVB footer, 32 MiB as vendor-gadget.img), check-bisect ok; the ramdisk is a gzip cpio of aarch64 binaries (file/readelf shown); then the keeper's try shows walleye-gadget-up (tests/host/gadget-up.sh) or, failing that, a new bootreasonBuilt and landed: google-walleye-lab a41a721 (PR 12), build/build-vendor-gadget.sh; image ~/Projects/walleye-build/images/vendor-gadget2.img sha256 9c0a8946cb65dc46311d6efc87da6393c152e3878d148bc773cc…
vendor-gadget3: vendor-gadget2 plus a self-reboot timer (240 s, cancelled by /tmp/hold) and a way for the shell to reboot (sysrq b)
google-walleye-lab доказано тестом vendor-gadget3.img = vendor-gadget2 with only the ramdisk's init changed; check script ok; the init's timer and /tmp/hold shown by sh -n and unpacking the ramdisk; hashes in SHA256SUMS; the keeper's try then returns without hands or answers gadget-upDone, landed google-walleye-lab 0f14a92 (PR 13): build/build-vendor-gadget3.sh, 14 checks ok (ramdisk differs from vendor-gadget2's only in /init; timer line, kmsg node, AVB verify). Stock config: CON…
From the running vendor kernel, read the facts our dts guesses (read-only)
google-walleye-lab доказано тестом a research row walleye-vendor-kernel-facts with cmdline, ramoops region and ids, checked against the dts (test: walleye-gadget-up passes during the read)read from the running vendor kernel; see research walleye-vendor-kernel-facts-s1-panic
vf2: the vf1 dump with a marker written at once and a status line per step, so an empty header can be told from a dead init
google-walleye-lab доказано тестом boot-vf2.img = vf1 with only the init changed: (1) write a plain-text header block WALLEYE-VF2 + a beat counter into the window within the first seconds, in m1's proven way (the same file/LUN setup as m1, which the lab user reads); (2) every dump step (mount debugfs, each file/dir copy, tar, gzip, length prefix) appends a status line 'step N name rc=X' into a text region of the window that never moves, before the blob region; (3) the blob is written after the text region at a fixed offset with its own magic and length; busybox applets used are checked present in the control's initramfs build (tar, gzip, cat, dd, mount), and failures are written, not hidden (no set -e swallowing, no >/dev/null on the steps); tests/host/vendor-facts-read.sh prints the status lines first, then extracts the blob if complete; selftest as vf1's; hashes; a programmer reviewboot-vf2.img built, 29 checks ok, rebuild identical, selftest 25 ok, reviewed; try queued as walleye-try-vf2
vf3: the gadget FIRST, then the dump written progressively (each file appended as read, plain text, no gzip/tar step), and a host poller that keeps everything the window ever showed
google-walleye-lab доказано тестом (1) boot-vf3.img = vf2 with only the ramdisk's init reordered: marker + gadget bound within the first seconds (as m1: ~18 s on the host), THEN the dump, one small source at a time (device tree files from /proc/device-tree as text/hex, clk_summary, regulator_summary, pinctrl/gpio, interrupts, iomem, dwc3/usb debugfs, dmesg head/tail), each appended at the next free offset of the 4 MiB window, framed with 'BEGIN name' / 'END name rc=N' lines, flushed (sync) after each; the init stays up until its hold (90 s) and the tail of the window carries a 'DONE' line; a read of a debugfs file that blocks is bounded by timeout 5; (2) tests/host/vendor-facts-poll.sh: started BEFORE the boot in its own background job, finds the LUN by our strings, copies the whole used part of the window every 2 s to numbered files, keeps the last complete copy, and fetches every new BEGIN..END block into research/.../vf/NAME the moment it appears; stops at 'DONE' or after 150 s; (3) selftest, programmer review, hashesboot-vf3.img built, 24 checks ok, rebuild identical, selftests ok, reviewed; try queued as walleye-try-vf3
vf4: keep vf2's proven init order up to the gadget bind byte for byte, then append vf3's progressive dump after the bind
google-walleye-lab доказано тестом boot-vf4.img = vf2's init with ONLY the dump part added after the gadget bind: everything vf2 did before and including the UDC bind unchanged (diff of the init against vf2's shows additions only after the bind step); the added part is vf3's progressive framed dump (BEGIN/END lines, appended, synced, timeout 5 per read); the diff vf2->vf3 is written up in the research row with the likely cause of vf3's death (what it did before the bind that vf2 did not: say it), the diff vf2->vf4 shown; the poller of vf3 reused; checks, selftest, programmer review, hashesbuilt: boot-vf4.img, 27 checks ok, rebuild identical, selftests ok; research walleye-vf4-built; try queued as walleye-try-vf4
One hardware source per image on m1's proven structure: vfa = /sys/firmware/fdt (the merged running device tree) + iomem + interrupts + cmdline; vfb = clk_summary; vfc = regulator_summary; vfd = the USB/dwc3 debugfs state
google-walleye-lab доказано тестом four images, each = m1 (its init, window file and gadget order byte for byte) plus ONE payload copied into the window file BEFORE the bind (cat to the window file with timeout 5; the payload sizes checked to fit the 2 MiB window, else split across images; the first 512 bytes of the window carry a plain 'WALLEYE-VFx' header, the payload's name and its byte length, then the payload as plain bytes, so the host reads it with massstorage-read.sh and one dd of the LUN; no tar, no gzip, no progressive appends, no poller); a host reader script extracts the payload by the header into the lab material path (the fdt kept as .dtb, with dtc -I dtb -O dts for the readable form); selftest; programmer review; hashesBuilt and checked: images in ~/Projects/walleye-build/images-vfx, research walleye-vfx-built; try queued as walleye-try-vfa.
bringup bisect FILE: a declared series of swaps (walleye's s0, s0n, s1, s2, s3), and the lz4 frame kernel part
lab-tools доказано тестом bringup-bisect-runs-declared-seriesBuilt and landed (lab-tools#123, 0bb6fdf on main): lpoco-bringup image bisect SERIES --out DIR; a TOML series (control, [[image]] name/kernel/dtb/ramdisk/cmdline/lz4/footer), each image counted agains…
bringup image swap: one part of a control image replaced, the AVB hash footer kept, and what differs said
lab-tools доказано тестом bringup-swap-s0-reproduces-controlBuilt and landed (lab-tools#110): lpoco-bringup image swap CONTROL --out IMG [--ramdisk|--kernel|--dtb|--cmdline]; proof bringup_swap_s0_reproduces_control passes. Bisect series and lz4 kernel are bri…
bringup image info|unpack|pack|check for header v0, packed byte for byte as mkbootimg
lab-tools доказано тестом bringup-pack-matches-mkbootimgLanded in lab-tools PR 85 (https://github.com/linux-phone-oco/lab-tools/pull/85): crate lpoco-bringup (lib+bin, argenv, family version 0.1.7, release-plz version_group), image info|unpack|pack|check f…
bringup initramfs build: a reproducible newc cpio, the gadget address as a variable, xz or gzip by the device
lab-tools доказано тестом bringup-cpio-matches-gen-init-cpioReplaces build-initramfs.sh. Test: from the walleye spec (busybox, /init, udhcpd.conf, device nodes) the archive is byte for byte what the kernel's gen_init_cpio -t 0 makes, then compressed with the c…
bringup follows by pin and extends by hooks: [tools] version check, manifest-named step hooks, a library API a rust-script can use
lab-tools доказано тестом bringup-refuses-wrong-pinLanded in lab-tools PR 88 (squash on main). Proof bringup-refuses-wrong-pin: bringup/tests/pin_and_hooks.rs (pin refusal, matching pin, parts hook changes ramdisk, failing/missing hook stops) and rust…
bringup is released by release-plz (a bringup-vX.Y.Z tag) and installs on a lab host from the tag
lab-tools доказано тестом bringup-installs-from-tagDone: cargo install --git --tag per package, proven by bringup-installs-from-tag against 0.2.0 (lab-tools #90, read back from main f4e55b0).
devlab fastboot boot reads the slot variables before and after a try and says what changed
lab-tools доказано тестом a test that fails when devlab reports a slot change as only refusedDone 2026-11-17 by lab-tools PR 80 (d510232), recorded there as fastboot-slots-told: devlab fastboot boot now reads current-slot and slot-unbootable/successful/retry-count of a and b before and after,…
devlab reaches a phone over fastboot (walleye speaks nothing else until Linux runs)
lab-toolsDone 2026-11-17 (lab-tools PR 79, device-template fastboot-serial): device.toml [fastboot] serial + writes (false unless said); devlab fastboot state | getvar NAME | boot now BOOT.IMG | reboot now [--…
devlab mcp: the phone for any agent
lab-toolsdevlab mcp (rmcp 3.10, stdio): run, put, get, health_report, health_run, kernel_install, kernel_try, kernel_status, reboot - each the function the command line calls; nothing but the protocol on stand…
fastboot: a phone out of fastboot found by its serial or USB port, else gone from USB
lab-toolsDone 2026-11-17 (lab-tools PR 81, f904196), asked by walleye-coordinator: after a boot try devlab took surya's postmarketOS gadget (18d1:d001) and a shared address for walleye's Linux while walleye ha…
fastboot boot says where the phone went and the slots' flags; set-active gated
lab-toolsDone 2026-11-17 (lab-tools PR 80, d510232), asked by walleye-coordinator after research walleye-first-fastboot-boot-tries: boot now reads current-slot and slot-unbootable/successful/retry-count of a a…
Flashing
lab-toolsdevlab flash did its first real flash (2026-11-10): the pristine stock image, the manifest's fixes baked in, the device tree fingerprinted (a43d069d0a49adcd), the phone back with all 20 checks passing…
A known gap counted apart from a failure, when a run is filed
lab-toolsdevlab gather files each run test_runs row with gaps: failures in the manifest [health] gaps, counted apart from failed, as the phone own verdict counts them (docs gaps-on-the-face). Columns are named…
devlab gather files a device's runs and results into its own base
lab-tools ждёт move-device-rowsdevlab gather files into the lab's own base when it has a devicebase.toml (same test-run record; filed() and derivations from the lab; land lands the lab alone); --docs keeps the old path; the output…
Gathered once the phone is quiet, landed again if main moved
lab-toolsMeasured 2026-11-17 (conveyor session): in about 2 h the watch gathered 5 times and 2 landings failed, other sessions having landed in docs and surya-lab meanwhile; each boot cost two PRs, two CI roun…
The ways, held for agents: devlab guard
lab-toolsClaude Code hooks on the lab call devlab guard: PreToolUse refuses git push to main, python/pip, and ssh/scp/sftp/rsync/sshpass to a phone address devlab has met (each manifest leaves its addresses in…
Health, after every boot
lab-toolsdevlab test, health, watch: regressions raise one board task.
Every change ends with the health report
lab-toolsEvery change devlab makes ends with the phone's health: phone reboot and kernel try wait for the report the new boot writes by itself (the test kernel's boot is measured like any other), provision run…
devlab-health tone make|find on the phone, the awk Goertzels retired
lab-toolsDone 2026-11-17 (conveyor-health): devlab-health is a multi-call binary. tone make|find links lab-tools' tone crate (its clap command a feature now; one WAV reader shared by tone, devlab hear and the…
devlab-health as an Alpine package
lab-toolsDone 2026-11-17: surya takes the suite from its Alpine package. devlab suite put now (lab-tools #51) adds lab-tools' signed repository and key to the phone once, installs or upgrades devlab-health, an…
The mobile-data check made generic, counting its bytes against a budget
lab-toolsDone 2026-11-17 (conveyor-health): telephony/10-mobile-data.sh in the generic suite (lab-tools #60, fixed by #62 to measure the data bearer and never IMS), with devlab-health data plan|spent keeping t…
Gathering is one command
lab-toolsdevlab gather now: the phone's reports into the device lab's tests/results, every unfiled run into docs' test_results and test_runs. Proven 2026-11-17: 5 reports brought from surya, 9 runs filed (5 fr…
The phone checks its own health after every boot
lab-toolsdevlab-health (static Rust runner, health/) and the generic checks: storage, kernel (oops, new warnings, processor crashes, failed units), signal processors, sound cards, Wi-Fi, modem, Bluetooth, disp…
The telephony layer: checks that cost something, at most once a day
lab-toolsPer decision 2026-11-17-health-checks-shell-and-rust, asked by bring-up for surya's SMS loopback and mobile data: devlab-health 0.1.1 runs a third layer after base and shell, telephony (a call, an SMS…
The lab's ear
lab-toolsdevlab hear and devlab health --hear work: a tone on each channel, recorded through the lab's microphone, measured against the silent room; heard checks join the one baseline when listened for. Its fi…
devlab hear asks for the word first
lab-toolsdevlab hear and devlab health --hear refuse without --placed, and say the procedure: tell the microphone's keeper, wait for their word. The lab's microphone can stay named: nothing plays by accident.
Safe test kernels
lab-toolsdevlab kernel: one try by boot counting, a timeout that falls back to stock.
devlab kernel: make a proven test kernel the phone's default
lab-toolsDone 2026-11-17 (orchestration, lab-tools kernel-keep): devlab kernel keep NAME now rewrites the test kernel's entry to sort ahead of stock with 3 tries and systemd-bless-boot no longer masked; its fi…
kernel remove never takes modules another kernel boots with
lab-toolsAsked by bring-up, 2026-11-17: its ear try is the same 7.1.0-sm7150-mic+ build as the kept mic kernel, so removing ear could have taken the default kernel's modules. Now install says the release in th…
A device-tree change is one boot: kernel try --dtb
lab-toolsAsked by bring-up, 2026-11-17: a device-tree experiment on the kept kernel took unkeep, a stock reboot, try and keep again. devlab kernel try NAME --dtb FILE now boots installed kernel NAME's kernel,…
The lab's devlab (and tone) move from 0.1.0 in ~/.local/bin to the family pin 0.2.1, between surya keeper runs
lab-toolsDone 2026-11-17 23:12 in the surya lane's window (keeper idle, lease released): the 0.1.0 copy is kept as ~/.local/bin/devlab-0.1.0; lpoco-lab and lpoco-tone are cargo-installed at lpoco-v0.2.1 with -…
devlab takes the family name (lpoco-...) when device repositories install it, keeping devlab as another name
lab-tools ждёт devlab first pinned or installed by a device repositorySuperseded 2026-11-17 by the one-roof decision: the rename is lpoco-rename-lab-tone, behind lpoco-dispatcher.
Never read LPASS pins through debugfs while LPASS is idle
lab-toolsFound by the bring-up session, 2026-11-17: reading /sys/kernel/debug/gpio and the LPASS LPI pinctrl debugfs on surya while the LPASS clocks were off stalled the bus; the phone hung until a person held…
lpoco: the dispatcher - lpoco NAME runs lpoco-NAME, one version for the family, list and completions
lab-tools доказано тестом lpoco-runs-subcommand-by-pathLanded in lab-tools PR 83 (https://github.com/linux-phone-oco/lab-tools/pull/83): lpoco NAME runs lpoco-NAME, refuses mixed versions, list/--version/completions/man from the argenv declaration, one fa…
lpoco-host --help prints its usage (argenv), as ikb and land do
lab-tools доказано тестом lpoco-host-help-prints-usageLanded in lab-tools #93, released as lpoco-v0.2.2: --help and -h print the usage (arguments and environment read from the argenv declaration) in lpoco, lpoco-host, lpoco-bringup, lpoco-health; lpoco-t…
lpoco host reads each device declaration from origin/main, or names the commit read and how far behind it is
lab-tools доказано тестом lpoco-host-reads-device-from-origin-mainLanded lab-tools #108 (15c7a7c), released v0.3.4. Proof: host test reads_each_device_from_origin_main_and_names_how_far_behind (seen failing first). version of script tools documented as label only.
lpoco host finds script tools (land, agent-env, oco-work) by rev and says what it found, and the devices glob takes a list
lab-tools доказано тестом lpoco-host-finds-script-toolsSeen writing lab.toml (lab-machine #36): plan says land 'found not found' and oco-work 'not installed' though both are in ~/.local/bin, since scripts print no --version and a script pin needs version…
lpoco-lab's command line declared once with argenv (devlab's clap moves over)
lab-toolslanded lab-tools #98 (a365eb2), read back from main; proof devlab/tests/help.rs. Left: declared env still read by std::env::var in flash/provision/watch/phone/hear/device (route through the contract);…
lpoco-lab reads its declared environment through the argenv contract, not std::env::var
lab-tools доказано тестом lpoco-lab-env-read-onceLanded in lab-tools #103 (fd1163d): the five declared env vars are read once in cli.rs; proof lpoco-lab-env-read-once seen failing, then green. Known: a non-UTF-8 HOME counts as unset; empty DEVLAB_RO…
lpoco mcp: one MCP server for the family, every subcommand's tools under its name, old names kept
lab-tools доказано тестом lpoco-mcp-lists-every-family-tooldone: lab-tools#87; proof lpoco-mcp-lists-every-family-tool (lpoco/tests/mcp.rs) green; read back from main
The old names (devlab, tone, devlab-health, the MCP tool names) swept to lpoco ones in small changes, the aliases kept until the sweep ends
lab-toolsLeft by lpoco-rename-lab-tone (lab-tools #86): ~450 references across AGENTS.md, READMEs, briefs, device.toml comments, launchers, devlab-watch@ units and scripts still say devlab or tone. Sweep one r…
devlab becomes lpoco-lab (lpoco lab), tone lpoco-tone; the old names stay as other names
lab-tools доказано тестом lpoco-lab-and-devlab-are-one-programLanded in lab-tools #86: packages lpoco-lab and lpoco-tone, second [[bin]] keeps devlab and tone running the same code; one family version 0.1.7 (release-plz version_group). Proof lpoco-lab-and-devlab…
lpoco-tone's command line declared once with argenv (the last clap in the family)
lab-tools доказано тестом lpoco-tone-help-from-argenvLanded in lab-tools #101, released in v0.3.1 (#97). Proof lpoco-tone-help-from-argenv: lab-tools tone/tests/help.rs, red first, 5 tests pass; programmer review fixed (out-of-range/zero rate and channe…
Reaches the phone
lab-toolsdevlab phone: USB or Wi-Fi, no quoting layer, boot ids.
Provisioning
lab-toolsPackages, files by {compatible}, Wi-Fi; the shared connection undone by a guard. It sets a fresh phone's clock and grows its system to fill its partition first.
A family release is one release PR that writes every member's changelog; main never red between two
lab-tools доказано тестом release-pr-writes-every-changeloglanded lab-tools#121; proof test a_release_where_one_member_changed_writes_every_changelog. Next release PR is the real read-back.
Every release of the lpoco family bumps the version and tags what landed: no changelog-only release under an existing version
lab-tools доказано тестом release-tags-every-changelanded lab-tools #116 and #119; proof: release-tags-every-change, release_plz_ignores_its_own_release_commits (lpoco/tests/release_family.rs). release-plz.toml release_commits lists feat|fix|perf|refa…
A change's health report stays on the phone until the watch files it
lab-toolsSeen by bring-up 2026-11-17: a suite run from its worktree brought the phone's reports into that worktree and they rode along in a landing. devlab no longer brings reports into the clone a command run…
suite put puts the lab as its main has it
lab-toolsSeen 2026-11-17 (orchestration's own slip): suite put run in a surya-lab clone 40 commits behind main put old checks and a known gap main had left (audio-mic) onto the phone, until the suite was put a…
suite put: find the runner by itself
lab-toolsDone 2026-11-17 (lab-tools PR 72): devlab suite put now --runner (no file) builds devlab-health with health/build.sh in the lab-tools devlab was built from and copies it with that tree's checks; --run…
Suite check suspend-irq-quiet: interrupts across one suspended cycle stay under a threshold
lab-tools доказано тестом suspend-irq-quietFrom surya-lab's backlog (floor research; the surya coordinator asked orchestration's builder, 2026-11-18). A generic check in lab-tools health/suite/checks/ with its row in docs tests (ikb test): rea…
Test sounds at a level the manifest sets - quiet by default
lab-toolsAsked by the author, 2026-11-17 (~03:00 local): 'please use quieter sound for tests, asap' and 'there should be a setting in the test suite'. The bring-up session had played a full-scale speaker-test…
kernel try: a try whose release equals the test kernel's is read as stock
lab-toolsDone 2026-11-17 (lab-tools PR 71): where LoaderEntrySelected is empty, devlab kernel try reads the booted entry off the running cmdline's initrd=initramfs-NAME, and the release only where neither says…
kernel try reports what came back
lab-toolsdevlab kernel try reports what the try came to, said by the phone after a new boot answers: booted, not chosen, stopped early, or timed out with its log. Proven 2026-11-11: booted the microphone kerne…
Research: make the Bluetooth UART (QCA, serial@88c000, hci0) suspend-proof
surya-labResearched 2026-11-18: decision 2026-11-18-bt-suspend-way; steps queued as bt-trace-cycles, bt-sleep-cts-pulldown, bt-unbind-hook, bt-suspend-gate.
Write the qcom_aoss prevent_* knobs to off, then cycle
surya-lab доказано тестом soc-sleepsDone 2026-11-17 (keeper-9): knobs written 0 before cycle 15 bt-on: cx/ddr writes time out, 'ucore did not ack channel': the QMP link to the AOP is dead, so the knobs settle nothing; counters 0 -> 0. r…
Read the suspend entry: RSC sleep flush, last-CPU idle state, IRQ delta (no kernel change)
surya-lab доказано тестом soc-sleepsDecision 2026-11-17-floor-read-the-entry. One cycle: way-back.sh cycle 15 bt-on, watchdog armed (BT-blocked cycles reset). Before and after: /proc/interrupts, every cpuN/cpuidle/stateN usage+time, pm_…
Research: what the AOP sees on the sleep flush; standby current as the arbiter
surya-labAnswered by research 2026-11-17-floor-aop-view-and-standby and decision 2026-11-17-standby-current-is-the-arbiter; phone items floor-genpd-rows-and-regulators, floor-usb-runtime-pm, floor-standby-curr…
Research: who holds the SH0 x=1,y=3 sleep vote, and why the AOP does not ack QMP
surya-labAnswered 2026-11-17-sh0-and-aop-research; decision 2026-11-17-sh0-owner-by-tag; phone items floor-sh0-owner, floor-sh0-qup-tags, floor-aop-qmp-no-ack.
Research: is SH0 x=1,y=3 the blocker at all - counter validity, clocks/GDSC, late-suspend state
surya-labAnswered by research 2026-11-17-floor-research-validity and decision 2026-11-17-floor-read-the-entry; phone items floor-entry-trace, floor-irq-quiet; build item suspend-irq-quiet-check.
Cycle with the USB controller unbound
surya-lab доказано тестом soc-sleepsDone 2026-11-17 (keeper-9): UDC a6f8800.usb unbound, cycle 15 bt-on: cxsd/aosd/ddr 0 -> 0. USB is not the holder. The rebind did not bring the lab link back (needed a hard reboot). research/deep-sleep…
Cycle with WLAN unloaded
surya-lab доказано тестом soc-sleepsDone 2026-11-17 (keeper-9): ath10k_snoc/core unloaded, cycle 15 bt-on: cxsd/aosd/ddr 0 -> 0. WLAN is not the holder. research/deep-sleep/2026-11-17-floor-wlan-off/README.md
Sound recorded through the microphone
surya-labTwo tries on 2026-11-11 in the test kernel. The Android route sets cleanly (TX DEC0 MUX=SWR_MIC, TX SMIC MUX0=ADC0, TX_AIF1_CAP Mixe…
The microphone codec times out at initialisation
surya-labSolved 2026-11-11: the audio pin driver muxed gpio19 (TX SoundWire data) to function 1 and gpio23 (RX data) to 3, where the phone's own Android kernel has 3 and 2. With both corrected the TX bus no lo…
The modem stays in shutting-down and will not go online
surya-labThe modem stays in shutting-down from boot with the SIM in. tqftpserv 1.1.1 is installed; its read-write base is /tmp/tqftpserv an…
A SIM card in the phone
surya-lab ждёт the authorInserted by the author on 2026-11-11: the card is seen (1&1, present and ready).
Boots that come up without the sound card
surya-lab доказано тестом sound-cardOn 2026-11-11 one stock boot had no phone sound card: the audio processor ran with no crash, but its service channel (APR) never…
Port the pin fix to the microphone kernel branch
surya-labThe tested tree came from the phone's v7.1-rc3 source plus upstream's WIP; linux-phone-oco/linux's microphone branch has the au…
Sensors probe: the registry kept current, sensorspd alone unmasked
surya-lab доказано тестом proximity-readsProbe done 2026-11-17 (keeper-8, research 2026-11-17-sensorspd-probe-still-writes): with the configs at mtime 1230768000 the DSP still writes temp.json and sns_reg_version through hexagonrpcd (refused…
Fallback: hexagonrpcd serves a writable sensor registry
surya-lab доказано тестом motion-and-lightproven by a 150 s run with 0 ADSP crashes and monitor-sensor showing accelerometer, light, proximity and compass; research 2026-11-17-sns-reg-version-served. Lasting install not yet (transient run)
Sensors kept: hexagonrpcd with a served version marker made lasting, three boots without an ADSP crash
surya-lab доказано тестом motion-and-lightproven by three boots, 0 ADSP crashes, motion-and-light and proximity-reads pass (research 2026-11-17-sensors-lasting-install-three-boots); suspend with the helper left to sensors-across-suspend (1 cr…
Checks that need the SIM skip, not fail, while it is out
surya-labDone 2026-11-17 (lab-tools #73 no_sim helper, surya-lab #123): with the SIM out, modem-registered, mobile-data, call-echo and sms-loopback say 'skip: no SIM card' (before the once-a-day stamp, so a sk…
Quiet the Bluetooth UART wake irq 157 so Bluetooth-on standby is representative
surya-lab доказано тестом soc-sleepsproven by 4 bt-on 15 s cycles: no 'nobody cared', hci0 up, soc-sleeps read (cxsd/aosd 0). Cost: ADSP sensor_process crash after every suspend (4 of 4) - needs sensors stopped before suspend; reboot re…
A way back from a suspend that may not return: the watchdog kept armed through sleep
surya-labDone 2026-11-17 (surya-lab sleep/way-back.sh, build-wdt.sh, patches/qcom-wdt-keep-armed-in-suspend.patch; patched module in updates/ on the phone, acts as stock until 'enable'): idle fuse, probe throu…
Why surya rebooted by itself during standby run r2
surya-lab доказано тестом soc-sleepsproven by research 2026-11-17-r2-reboot-was-the-fuse (pstore console, journal, r2 logs read); r3 waits on decision 2026-11-17-r3-waits-on-the-bluetooth-choice, which the surya coordinator takes
surya unreachable after a real systemctl suspend: the author's power cycle, then the cause
surya-labPhone work is PARKED until it answers. When it does: read pstore/ramoops and journalctl -b -1 first (why it did not return: the 30 s fuse was defused for the try; sxmo autosleep still running?). Rules…
Suite check suspend-irq-quiet: interrupts across one suspended cycle stay under a threshold (lab-tools)
surya-lab доказано тестом suspend-irq-quietA generic check in lab-tools health/suite/checks/ with its row in docs tests (ikb test): reads /proc/interrupts before and after one 15 s s2idle cycle and fails above the threshold, printing the top t…
Telephony in the test suite
surya-labdevlab checks: modem registered; mobile data (download a small file over the mobile connection); a call to the number in ~/.…
surya's profile
sxmo-deviceprofilesBoth panel variants; Sxmo loads it; scale 3.
surya's profile upstream
sxmo-deviceprofilesIn sxmo-utils master as 1fd7822 (sxmo-profile-patch landed).