Open Commons ObservatoryEnglish·Русский

Журнал

Как шла работа — что сделано, что рассматривалось и что отвергнуто.

bringup image bisect built

Proof test first, seen failing; image bisect SERIES with lz4 kernel and footerless entries; programmer review found a panic on a footer running past the image, a mangled doc line and untested miss paths: fixed and tested; landed lab-tools#123, read back from main.

Отвергнуто: deleting failed images from DIR: kept, the report line says FAIL

bringup image swap built and landed

image swap in lpoco-bringup: parts replaced in a control image, AVB hash footer rebuilt through avbtool (NONE; partition, salt, props, rollback index, hash algorithm, flags carried), what differs said. Proof test seen failing, then passing; programmer review fixed four defects (dropped footer facts, footerless size, wrong footer size, dead code). Landed lab-tools#110, read back from main; release PR 109 landed.

Отвергнуто: Rebuilding a signed vbmeta or multi-descriptor footer on a guess: refused with a message instead. avbtool verify_image as the test's check: it crashes looking for boot.img, so info_image is used.

release check accepts a completing changelog

Landed lab-tools #116: check-changelog.sh with tests (failing first), merge-base, per-package [package] version, quoted/spaced names; programmer review's four bugs fixed.

Отвергнуто: Global version-bump test (old check): a bump elsewhere excused any changelog.

lpoco host reads origin/main

Failing test first, then device_text in lpoco-host; programmer review fixes (bounded non-prompting fetch, reasons on every fallback, differing-file note, read: line in every verb, exact git subcommand filter in tests); landed #108, released v0.3.4.

lpoco host: script tools by rev stamp

Landed lab-tools #95: stamp per script tool judged by rev, devices list, advice text; programmer review fixed three points (stamp only counts with the script present, rev refuses whitespace, README).

Отвергнуто: Content hash: needs the clone present just to plan; teaching land and agent-env --version: owned by other repos. Dropping the version field: breaks existing lab.toml.

lpoco-lab env through argenv

Built lpoco-lab-env-through-argenv: DEVLAB_ROOT, LAB_AGENT, PHONE_PASS, DEVLAB_MICROPHONE, HOME read once through cli::env(); two proof tests (source scan, empty DEVLAB_ROOT exits 2); programmer review applied; landed as lab-tools #103.

Отвергнуто: Declaring defaults on the String inputs - not checked that argenv supports it; left in env_in. Per-command refusal of empty DEVLAB_ROOT - refused at parse for one rule.

lpoco-lab on argenv

Failing test first (devlab/tests/help.rs), cli.rs declaration + TREE, parse() in main.rs, programmer review answered (per-action flags, exit 2, help wins, README), landed lab-tools #98.

Отвергнуто: see decision

lpoco-tone on argenv

Proof test first (red), cli.rs and main.rs on argenv, clap removed; programmer review found silent defaulting of out-of-range rate/channels, fixed with tests; landed lab-tools #101, release v0.3.1 landed (#97), read back from main.

Отвергнуто: Fixing zero-value panics only by recording them: refused at the edge instead (cheap).

Names sweep: surya-lab and google-walleye-lab

surya-lab #166 (8 files: AGENTS, README, comments in device.toml/idealkb/tables/phone-queue.sql/test-run.sql, research/microphone.md) and google-walleye-lab #100 (10 files) landed after each lane coordinator said yes to the exact list; phone-queue and phone-next rows identical before and after, tad check green, programmer-reviewed, read back from main.

Отвергнуто: Sweeping scripts the keepers run and the history: the coordinators said no, and they call the aliases which stay until a grep finds no caller.

Names sweep: lab-tools messages

About 20 messages and the gather --land help said devlab; now lpoco-lab. New test names_in_messages.rs failed first, then passed; programmer review found nothing parsing the strings. Landed as lab-tools #104 (cargo fmt fixed after a failed check), read back from main.

Отвергнуто: Renaming env vars, the state dir and the 'by' default: they are interfaces and recorded values, they stay until the aliases go.

Names sweep: lab-tools README

lab-tools README commands now say lpoco lab / lpoco tone (24 commands); lpoco/tests/readme_names.rs fails on an old name starting a command; landed as lab-tools #99, read back from main; programmer review applied (cargo fmt).

Отвергнуто: Changing real paths (~/.config/devlab, hook path, mcp add line): aliases and state dirs stay until the alias grep finds no caller.

Names sweep: marks, docs briefs, device-template

lab-tools #100 (oco.toml name/title, lpoco-health service Description; test lpoco/tests/marks_names.rs, seen failing first); docs #400 (six briefs in ideal-agent-linux-phone-oco/ say lpoco lab; overlay-check passes, builder.md and devlab-watch@ left); device-template #12 (docs, comments, command lines). Each programmer-reviewed before land. surya-lab and google-walleye-lab: line lists sent to their lane coordinators, landing only what each agrees to.

Отвергнуто: A test in device-template for the sweep: it has no Rust crate, and lab-tools' tests already keep the generated names; a grep-only check there would be one that only prints.

release check: kinds and changelog

Built commit-kind workflow and release_family tests, landed in lab-tools #112, review by programmer applied; v0.3.5 release #111 landed and carries #110.

Отвергнуто: tag-vs-changelog test (tag precedes changelog)

release-plz stops releasing its own release commits

Test first (red: no release_commits), then release_commits in release-plz.toml; programmer review found no blocker; landed as lab-tools #119, read back from main.

Отвергнуто: One release PR for all changelogs: needs release-plz to change its own PR shape, not in its config. Listing chore among released kinds: that is the loop. A chore! breaking change no longer releases (accepted; the commit-kind lists must stay in step).

release-plz one PR: changelog_include

Reproduced the missing changelogs with release-plz 0.4.3 on a fixture; test red before, green after; programmer review applied (faithful fixture, target tmpdir, isolated git config, pin). Landed as lab-tools#121, read back from main.

Отвергнуто: see decision

Suite check suspend-irq-quiet

Built suspend-irq-quiet-check: 92-suspend-irq.sh, manifest fields suspend_irq_max and suspend_guard, test row. Proof test failed first, then 8 pass. Programmer review found: alarm not read back, no alarm clear on signals, fail-open when suspend counter or interrupts unreadable, garbled guard message, racy and hang-prone tests; all fixed. Landed as lab-tools #106, released v0.3.3, read back from main.

Отвергнуто: A guessed default threshold: unset says not measured. Reading the watchdog directly: the guard is a manifest command, as the device owns how it is armed.

bringup image v0 built and landed

Failing proof test committed first (fixtures: kernel, ramdisk, appended dtb, walleye [boot]); measured abootimg-oxide against mkbootimg; built lpoco-bringup image info|unpack|pack|check; landed as lab-tools PR 85; main read back.

bringup pin and hooks built

Failing tests first, then pin module, parts hook, documented library and examples/step.rs run by rust-script; CI first failed for lack of rust-script, fixed by installing it in ci.yml; landed as PR 88 and read back from main.

Отвергнуто: Skipping the script test without rust-script.

bringup install from tag

Wrote install_from_tag test first (failed: released 0.1.7 tags had no pin, and the tags sat on different commits). Landed release PR 84 by hand (land --release did nothing), which made 0.2.0 tags on one commit and broke bringup_refuses_wrong_pin (pin 0.2 now right). Test and fix landed in lab-tools #90 with README install section; CI green, read back from main.

Отвергнуто: Release assets: not needed while hosts have cargo. Note: land --release exits silently without merging; merged PR 84 with gh pr merge.

Analysis in Rust on the phone, and mobile data under a budget

lab-tools #57, #60, #62 and releases 0.1.2, 0.1.3, 0.1.4; surya-lab #115 and #116; devlab installed from main (the previous one kept at ~/.local/bin/devlab.before-data-budget). devlab-health 0.1.4 put on surya, and suite run 20261117-093527 was healthy, with the tone checks and mobile-data passing. The static aarch64 runner went from 528,656 to 594,192 bytes. Signs seen: land exited 1 with nothing said when GitHub refused a merge for a conflict (main had moved); release PRs' ci runs fail to start, as before; one run's earpiece recording held exact zeros (-240 dBFS) right after suite put, and two runs of the check alone passed at -22.3 dBFS.

Отвергнуто: See decision 2026-11-17-health-analysis-subcommands.

Every change ends with the phone's health

devlab phone reboot, kernel try and provision end with the health report of the boot they came to; proven by a reboot of surya: healthy, report brought, 3 minutes.

Filed through tad, landed by itself

devlab gather now --land: one new report brought from surya, filed through tad sql into the picture, landed in surya-lab and docs in 75 s with every check passing; devlab watch now does this after every boot and says the phone's own verdict on the board.

Gathering is one command

devlab gather now brings the phone's reports to the lab and files every unfiled run into the picture's tables; 9 runs filed, a week of gathering by hand caught up.

lpoco dispatcher built and landed

New workspace member lpoco (argenv, no tokio): failing tests first, then dispatch by path (beside itself, then PATH), version check against --version, list, --version, completions, man; lpoco-health gained --version; release-plz version_group for the family.

Отвергнуто: A shared family tag (breaks the conveyor prefix).

lpoco family answers --help

Wrote tests/help.rs per crate (failing: '--help is not a declared flag'), declared HELP (-h/--help) once with argenv in four cli.rs, usage printed from the records; landed #93 and release v0.2.2, read back from main and tag.

Отвергнуто: A shared crate: one 25-line fn per crate is smaller than a new release unit, and health ships to phones. Hand-written usage text: it would drift from the declaration.

lpoco-health: the one health suite

Looked around (research 2026-11-17-lpoco-health-landscape), renamed devlab-health to lpoco-health, added run-time conditions (no-sim, no-modem, no-network), switches with reasons on the phone (lpoco-health off|on|list), argenv command line, devlab suite off|on|list and MCP tools, migration of reports/ledger/marks, conveyor names (lab-tools #75). Nothing done on the phone: its put is on the phone queue.

Отвергнуто: see decision 2026-11-17-lpoco-health

lpoco-host built and landed (lab-tools #91)

New crate host/ (lpoco-host, argenv): plan (default, changes nothing), prepare now (the printed root commands only), check. Proof tests host/tests/plan.rs: plan_grants_gadget_serial_and_changes_nothing failed first, then passed with plan_lists_host_tools_at_pins_and_installs_nothing, prepare-now, injection-refused, wide-rule and old-devlab tests. Review by programmer: fixed unvalidated codename/serial/module words (refused unless letters digits . _ -), tty open now O_NONBLOCK|O_NOCTTY, full X.Y.Z lab pin, unknown tools must be 'unknown', stage dir 0700 and removed, tests read the host's files under LPOCO_FS_ROOT, one manifest per codename (clones and worktrees of a device repo found by the glob). 51-android.rules named and kept; old devlab 0.1.0 first on PATH shown off the pin; oco-work listed as not known yet. Read plan on this lab (read-only): walleye rule + modules file new, health and host not installed, devlab 0.1.0 off the pin. Acceptance not run: prepare now belongs to whoever holds the lab with the walleye keeper.

Отвергнуто: Running prepare now here (host is shared; orchestration's boundary). Removing 51-android.rules (the author: keep it for now). Duplicate devices per worktree (first host-declaring manifest wins, conflicts named).

lpoco mcp: one MCP door for the family

Wrote the proof test first (failed: no mcp built-in), then lpoco mcp, lab_ aliases, bringup_image_info/check tools; landed as lab-tools#87 and read back from main.

Отвергнуто: The proxy, see decision 2026-11-17-mcp-one-door-by-prefix. health has no phone-side MCP of its own: its tools stay the lab's health_* ones.

devlab becomes lpoco-lab, tone lpoco-tone

Renamed the two packages, kept devlab and tone as second binaries of the same source, joined both to the family version and release-plz group, README note, test devlab/tests/one_program.rs. CI first failed on a clean build (the list test saw only some binaries); the test now builds the workspace's binaries first. Read back from main: devlab/Cargo.toml names lpoco-lab.

Отвергнуто: Alias bins: more files for a warning that cargo only prints at the manifest; two [[bin]] on one path is the plain form. Rewriting devlab's clap parser to argenv here: out of this item's scope, a separate change.

Research: how device repositories follow and extend bringup, and what rhizoid lacks

Read idealkb's chain-follow decision and research and rhizoid's design log and backlog. Recorded research bringup-home-follow-and-rhizoid (five sources), re-recorded decision 2026-11-17-bringup-in-lab-tools with the follow, extend, send-back and rhizoid parts, added backlog bringup-pin-and-hooks and bringup-release, updated device-template-names-bringup. Filed three items on rhizoid's own backlog (git-rhizoid.github.io PR 28).

Отвергнуто: All of those, reasons in the decision. Nothing built.

Research: where bring-up tools live and what they are called

Read our records (rust-toolset, device-template, health-checks-shell-and-rust, lpoco-health landscape), walleye's eight scripts and devlab; a researcher read the outside (mkbootimg, avbtool, magiskboot, abootimg-oxide, cpio, dtoolkit, pmbootstrap, lk2nd, Droidian). Recorded research bringup-tools-landscape, eight sources, decision 2026-11-17-bringup-in-lab-tools, and five backlog items (bringup-image-v0 first, with its failing test named).

Отвергнуто: linux-phone-oco-bootstrap-helpers / lp-oco-bh, bootstrap, bootkit, bootimg, imgtool, mkboot (reasons in the decision). Nothing was built and no device touched.

Research: preparing the lab host from declarations (lpoco host)

Read the roof decision and its follow-ups, lab-tools (devlab provision, device reader, health/build.sh, bringup pin), lab-machine (setup, watcher unit, bin/land), lab-agent and ideal-agent's retirement decision (root-exec part 6), walleye's device.toml [host] (lab PR 54) and today's session logs (root-exec uses); measured the lab host read-only (systemd 262, groups, seat0 session, rules, modules, packages, units, ~/.local/bin). A read-only researcher read pmbootstrap, android-udev-rules, systemd's uaccess source and NEWS, nixpkgs, Ansible, LAVA and labgrid. Recorded 9 sources, research lab-host-prepare-landscape (material research/lab-host/2026-11-17/), decision 2026-11-17-lpoco-host, backlog lpoco-host (rank 1, needs build, google-walleye; the author's acceptance and the coordinator's host-tools test added) and lpoco-host-rest, research lpoco-host-how-it-stands. Nothing changed on the lab host.

Отвергнуто: See decision 2026-11-17-lpoco-host: names labprep, hostprep, setup, provision, bootstrap, prepare, machine, env; forms above; pins in device.toml (one ~/.local/bin, tools no device owns); a hand-written udev rule meanwhile (the author waits for the tool).

Support status follows the phone

After filing, devlab runs every statement the picture keeps in derive/ through tad: support status is read off the latest results (broken on a failure, works on a proven pass, otherwise as judged). Proven with run 20261117-005911: filed, derived, landed - every measured part now names it.

The phone over MCP

devlab mcp: nine tools that let any agent run commands on the phone, move files, read and measure its health, try a kernel once and reboot - proven against surya.

Only an EFI program is offered to the bootloader

A test kernel given as Image.gz was chosen by the bootloader, counted down, and could not start - the try was lost before the kernel logged a line, and looked like 'the bootloader ignores the entry'. devlab kernel install now refuses anything that is not an EFI executable and names the right file.

Отвергнуто: Acting on any of them before the bootloader's own record (the entry's try counter) was read.

Two refusals, and what they cost to learn

The kernel guard and the listening word, each from a mistake made today; two more steps recorded: a room of digital silence, and a try that reports its own outcome.

The word first

A listening test now needs the explicit word that the microphone is placed, and refuses with the procedure otherwise - so the lesson of the worn microphone rests on the tool, not on memory.

Отвергнуто: That: a renamed file is a trick only its author remembers.

devlab flash's first real flash

The pristine stock Sxmo image flashed to surya with devlab flash: fixes baked in, device tree a43d069d0a49adcd, the phone booted and passed all 20 checks. Two things done by hand after it became the tool's job: a fresh phone's clock is at its image's build date (certificates looked invalid) and its first boot grew the partition but left the filesystem at 2 GB - devlab provision now sets the clock and grows the system first.

Отвергнуто: That: what must be known to use a tool belongs in the tool.

It tells its essence

Its idea written for anyone, and its workings: what delivers and proves each step of its path, and why we care. Steps that nothing serves yet are left empty, to be seen.

Отвергнуто: That: a step nothing serves is a wish, and should show.

The ear in health

devlab health --hear listens as part of health: heard checks join the one baseline and regress like any other. The rule made honest rather than a second baseline: a run that did not listen is out of scope for what only listening tells. The watcher never makes a sound by itself.

Отвергнуто: The first: a tone in the night. The second: two truths about one phone.

The ear, on notice

The author corrected the withdrawal: the microphone is reliable when placed by the phone - it only was not, unannounced. The procedure is written down: notify, wait for his word, then listen. A proper rig - fixed microphone, webcam, servos - is on the backlog.

Отвергнуто: That: an instrument placed with care on notice is an instrument.

The lab's ear

devlab hear joins what existed - the tone tool, the lab's microphone, the phone's player: the room recorded in silence, a tone played on each channel at a known volume and recorded, measured against the noise around it and the silent room. Both of surya's speakers heard on the first run. For any phone: its lab names a microphone, its manifest names the tones.

Отвергнуто: That: a machine should hear it. And claiming which speaker is which - one microphone cannot tell.