Open Commons ObservatoryEnglish·Русский
Linux-телефоныСутьИдеяНасколько мы близкиТелефон в лабораторииФункцииУстройстваСемейства SoCТестыResearchDecisionsБэклогThe phone's queueОкружениеСделано работой

Окружение

Проекты вокруг наших — что делает каждый и как мы строим на них, а не рядом с ними.

Anthropic's orchestrator-worker research

A lead agent plans and starts 3-5 workers in parallel, each with its own context, an explicit objective, output format and boundary ('not X - that is another worker's job'); about 15 times the tokens of one agent; worth it only for work that is truly independent and valuable; later guidance argues most systems should have fewer, better-scoped agents.

For us: Parallelise research (it is independent and reads only); serialise what touches the one phone; write each brief with its boundary; spend agents where the value is, and say the cost.

bootrr

Checks that every expected driver probed on a booted device; packaged in pmaports for the CI hook.

For us: Replaces our hand-made waiting-to-probe checks with the established tool.

CI-tron

A bare-metal testing gateway: deploys test environments to devices, sequences power and boot, collects logs; devices are trained through boot cycles before they take jobs.

For us: The gateway our lab machine could become, instead of growing its own.

Claude Code: sessions, teams, hooks, schedules

Background sessions with cross-session messages (the desktop app), subagents inside one session, experimental agent teams with a shared task list claimed by file lock, worktrees to keep writers apart, hooks that run commands at fixed points (PreToolUse, Stop, SessionStart) and can block - the deterministic way to enforce a rule - and scheduled runs: desktop tasks on the lab, cloud routines, GitHub Actions.

For us: Enforce the ways with command hooks on the lab rather than by an agent watching; start sessions from versioned briefs (prompts/); run patrols on a schedule rather than an agent always on.

Gas Town and Beads

An orchestrator for 20-30 coding agents: work state lives in git as JSON-lines issues (Beads), not in agents' memory; an agent is a persistent identity, not a session; workflows are chains of small issues an agent walks one at a time, so a crash or a full context loses nothing; supervisor roles patrol with backoff; its author says not to use it unless juggling five or more agents daily.

For us: Our backlog, decisions and session logs in tad/ocoeco data are already the git-kept work state; ocoeco's 'where the drivers stand' is the identity view. Not to adopt: a second tracker beside ocoeco would be one nature under two names. To take: an agent is a name (OCOECO_BY) with a history in the data, and work is small steps that survive a cut-off session.

mobian-recipes

Debian images built with debos; devices grouped in families by SoC; device packages (<device>-tweaks); weekly CI image builds. Points to the postmarketOS wiki to judge whether a device runs mainline.

For us: Consumes mainline device support rather than building it: surya reaches Mobian through mainline and a family recipe, after pmOS.

ArtCraft: seventeen apps, one organisation

Seven, then thirteen clean-room Rust apps built with about 8 parallel Claude agents, each owning one subsystem directory. House rules kept once (craft-contrib/standards) and stamped into every repository's AGENTS.md with a checksum; CI shared from one actions repository; agents check an issue and claim it before working; a generated scorecard, ranked gaps list and parity checklists say what is left; a test oracle (Photoshop's own renders, a CPU compositor for the GPU one) decides what counts as progress; remaining work estimated in agent-hours.

For us: The same shape fits us, almost one to one: craft-contrib (rules and RFCs, stamped into every AGENTS.md) is creation-guidelines and the practices; craft-common-actions (CI called, not copied) is ocoeco's gather and the package conveyor's workflows; craft-launcher (installs and updates every app) is the package conveyor's Alpine repository to come; craft-repo-status-app and craft-wiki are the face and docs; every app agent-ready (CLI, JSON control, MCP) is devlab mcp and agent access. What they have and we lack: craft-skills - procedures shared as agent skills, discoverable by every agent rather than remembered (our oco-work words, as skills); and *-corpus repositories - reference data an oracle compares against (for a phone: what Android does on the same hardware - sleep current, audio levels, sensor readings - measured once and kept). Worth taking too: claim before work, estimates in agent-hours, a ranked gaps list and scorecard drawn from data, and the lesson that breadth is fast and only an oracle keeps the numbers honest.

Device packages and deviceinfo

Each device package carries a deviceinfo file - codename, name, architecture, flash method, device tree - the device's manifest.

For us: Our device.toml derives from deviceinfo and adds only what it lacks (how the lab reaches the phone), rather than duplicating it.

postmarketOS hardware CI

Tests on real phones: CI-tron as the testing gateway, power units that switch devices and press buttons, GitLab pipelines generated per device from a template in its package; scaling toward a farm of dozens of devices (talk at the 2026 conference).

For us: Our lab is a one-phone version of it. Aim: surya as a device under test in their farm, our tests in their shape - contributed, not kept apart.

Initramfs CI hook and test packages

postmarketos-mkinitfs-hook-ci runs tests inside the initramfs; tests ship as small subpackages that install themselves for the right device (postmarketos-test-suspend, -unl0kr).

For us: Our automatic checks and the test kernel's timeout hook belong in this shape, so they can go upstream and run in their CI.