Open Commons ObservatoryEnglish·Русский

Понятия и решения

Понятия проекта и принятые решения — с тем, что было отвергнуто.

A release tag is moved only after the change has landed - by land --tag, which confirms the merge first.

A job moved v1.0.1 after its change had been refused, so the tag claimed a fix it did not hold; a rule kept in mind failed, so land enforces it.

Отвергнуто: Tagging by hand in the same breath as landing.

2026-11-11

Projects call the conveyor's reusable workflows, pinned to a version; nothing is copied into them.

snapall and lab-agent carry about 500 lines of nearly identical workflow each; copies drift and fixes do not travel.

Отвергнуто: A template repository; copy-paste with care; composite actions alone (they cannot hold whole jobs).

2026-11-11

The extraction is proven by moving snapall and lab-agent onto the conveyor with byte-for-byte the same published repositories - same URLs, same keys.

An extraction that changes what users receive cannot be told from a regression.

Отвергнуто: Proving it on a new project first.

2026-11-11

'Nothing changed for users' means the same addresses and key, every signature verifying, and the README's install line installing the new release - not byte-identical packages.

Arch packages built twice differ in their timestamps; the byte promise could not be kept and would test nothing a user meets.

Отвергнуто: Byte-for-byte identical repositories (the earlier wording of this decision).

2026-11-11

Each project keeps its own signing key for now; one key for the whole commons is considered later.

A single key would change the trust of every existing installation at once.

Отвергнуто: One organisation key from the start.

2026-11-11

The conveyor's direction is to build in the project's own Nix environment (the shell.nix the lab machine uses), from Nix's binary cache; the patient retry stays as a fallback meanwhile.

The author asked whether we must rely on Ubuntu's archive; Nix gives one pinned toolchain for the lab and CI, from a more robust source.

Отвергнуто: Only retrying; switching to another single distribution's mirrors.

2026-11-11

A project describes its delivery in one small declaration file; the workflow inputs stay few.

Inputs grow without limit and are scattered across callers; a file is read, reviewed and checked like data.

Отвергнуто: Workflow inputs only.

2026-11-11

The conveyor's scripts live inside its workflows, not as files it fetches.

A called workflow does not know which version of itself was called, so fetched files could drift from the workflow; inline, they move together with the tag.

Отвергнуто: A scripts folder checked out at run time.

2026-11-11

Arch (pacman) and Alpine (apk, for postmarketOS) first, each with a signed repository.

These are what the lab, the desktops and the phone actually run.

Отвергнуто: Debian and RPM repositories first - installation on them stays verified, as today.

2026-11-11

The declaration is read by a typed Rust reader that refuses unknown fields, not by Python inside the workflows.

Shipped code that other projects call: with Python a misspelled field is silently replaced by a default - a bug Rust's types rule out by design. Asked by the author: no Python for new things.

Отвергнуто: Python's tomllib inline (the first version); a shell TOML parser.

2026-11-11

Cutting versions stays with each language's own tool (release-plz for Rust); the conveyor starts from a release tag.

Versioning is part of a project's own decisions and tooling.

Отвергнуто: The conveyor owning versions.

2026-11-11

A release is said once it can be taken. Each program's releaser makes it a draft, with its tag - release-please with draft and force-tag-creation, release-plz with git_release_draft - and the conveyor, having attached every file, says it: one step, once, at the end of the conveyor every program already calls. A release already said stays as it was.

A release was made and its files were built after it: for minutes the newest release of a program had nothing to take, and each who takes one had to know of it - the author's observatory took the newest that had its files.

Отвергнуто: Each who takes a release taking the newest one with its files (every reader carries the gap); the conveyor making the release itself from the tag (versions and notes stay with each program's own releaser); a release made public, then taken back to a draft by the conveyor until its files are there (said, then unsaid: the gap only shortened); a draft without its tag (release-please would not find the release before it, and the PKGBUILD's source is the tag's).

2026-11-16

An Alpine repository is built in one container and signed in another: the project code (its APKBUILD build) runs where no key is; abuild rootpkg and index, which sign, run in a fresh container that alone receives APK_SIGNING_KEY. The key is the project own (an RSA key made like abuild-keygen), its public half committed beside the APKBUILD.

As with the pacman path, the code being packaged never shares a process space with the key; a fresh container leaves nothing running from the build. abuild signs each package at packaging time, so signing cannot wait for a later job without re-packing the files.

Отвергнуто: Building and signing in one step (the project code beside the key); signing in a later job by splitting and re-signing each .apk (re-packing the format by hand, fragile); GitHub JavaScript actions inside an Alpine container (they do not run on arm64 Alpine); one key for the whole commons (decided per project for now: 2026-11-11-keys-per-project).

2026-11-17