Skip to content

feat(release, luban): part 3 -- one writer of a release's index entries, cross-repo contract read through the client, a move tried first (J47), #699; 2026.10.23.2 - #701

Merged
Sunrisepeak merged 10 commits into
mainfrom
feat/part3
Oct 23, 2026

Conversation

@Sunrisepeak

Copy link
Copy Markdown
Member

Part 3 of the Luban OS ecosystem usability design (.agents/docs/2026-10-23-luban-os-ecosystem-usability-part3-design.md), converged into one PR and one version, 2026.10.23.2.

What

U3 tools/bump_index.py writes xlings, luban, luban-os and xim:xdesk (every platform block, latest) into an xim-pkgindex checkout from the release's .sha256 sidecars -- byte for byte the shape written by hand (checked against 2026.10.23.1), all or nothing, idempotent. tests/scripts/test_bump_index.py (contract-scripts).
U4 release.yml's bump-index and tools/bump_index.sh removed: they called the index's version-check.py (removed by xim-pkgindex#1027) and waited on a mirror job that times out -- they never wrote anything. AGENTS.md, the contributing skill and docs/92-release (en, zh) name the tool.
U5 (supersedes #700) the rootfs self-update fixture asks the candidate client for xim:xlings's versions (interface package_info), not a regex over the recipe: #1027 put an autoupdate table with its own linux = { above xpm, and the distro shard of main and every PR went red. package_info's package now carries versions (protocol 1.18, feature package_versions; IFACE-PACKAGE-VERSIONS).
J47 / U8 luban upgrade of a pre-2026.10.14 system is tried first: xlings subos __domain-check <home> (plumbing; same code as subos new --domain-if-possible's start; feature domain_check; DOM-CHECK-TRIED). Where the host cannot build the domain nothing is exported or removed -- it stays, said once (LUBAN-RELAYOUT-TRIED). An archive kept by a system that lives in the prefix is named in luban status as freeable.
#699 (mcpp#832) cherry-picked with its author: a script package's alias quotes its path.

Tests

  • local: test_interface_protocol, test_xpkg_extensions, test_i18n_domains, test_domain_producer (new check), test_luban (all 32), contract scripts, lints; the fixture run against the live index (picks 2026.10.22.4 for a 2026.10.23.1 candidate).
  • the tool reproduces the hand-written 2026.10.23.1 entries of all four recipes exactly; a second run changes nothing; a missing sidecar writes nothing.

After merge

release 2026.10.23.2 → GitCode filled locally (gtc) → tools/bump_index.py 2026.10.23.2 into xim-pkgindex#1031 (+ its U6 contract check) → merge #1031 → verification (part 3 §4).

Closes #700. Supersedes #699 (its commit is here, author kept).

Co-authored-by: speak-agent 248744407+speak-agent@users.noreply.github.com

Sunrisepeak and others added 10 commits October 23, 2026 10:16
… release's index pointers, cross-repo contracts tested on both sides, the verification loop, what is left

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
…26.10.23.2); self-review

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
A script package is registered with the alias `xlings script <path>`, and
the shim reads an alias as words: in a home under `/tmp/mc probe` it ran
`xlings script /tmp/mc` and failed with `package file not found`. The path
is now quoted with `platform::shell_quote`; a quoted alias is still plain
words to the shim (`split_plain_words`), so it keeps running without a
shell. tests/e2e/script_type_install_test.sh runs a script package's shim
in a home whose path holds a space.

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
…he bump-index job that never ran is gone (part 3, U3, U4)

xlings, luban, luban-os and xim:xdesk have one writer of their versions,
xlings' release. tools/bump_index.py writes every platform block and latest
of the four recipes from the release's .sha256 sidecars, byte for byte the
shape written by hand; a missing sidecar writes nothing; a second run
changes nothing. Run locally after the CN mirror, then open the index PR.

release.yml's bump-index (tools/bump_index.sh) called the index's
version-check.py, which the index removed, and waited on a mirror job that
times out: it never wrote anything. Both are removed; AGENTS.md, the
contributing skill and docs/92-release name the tool.

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
…the client, not a regex over the recipe; package_info carries versions (protocol 1.18; part 3, U5)

tests/e2e/rootfs_lib.sh took the first `linux = {` of the index's
xlings.lua as its version table; the index put autoupdate (with a
`linux = {` of its own) above xpm, and main and every PR went red in the
distro shard. The fixture now asks the candidate client:
`interface package_info xim:xlings` (linux), whose `package` carries
`versions` -- the platform's concrete versions, newest first -- beside
`version` (what latest resolves to). How the index lays a recipe out is
its own. Supersedes #700.

Protocol 1.18 (additive): features package_versions and domain_check;
the spec gains the 1.17 row it lacked.

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
… host cannot build the domain nothing is carried away (J47); 2026.10.23.2

`luban upgrade` of a system made before 2026.10.14 exported it (gigabytes),
removed it and made it again in /xlings; where this host cannot build the
private domain it was made again in the home's prefix and the archive
stayed. `self doctor --isolation` said this host could: it judges the
tools, not the namespace working.

xlings `subos __domain-check <home>` tries the domain as `subos new
--domain-if-possible` starts (the same code: prepare, `self init` inside,
judged by the home it made): 0, or an error saying why. luban asks it
before the move (feature domain_check): where it fails the system stays,
recorded in `.stays` with the reason and said once, and the upgrade goes
on in place. An archive kept by a system living in the prefix is named in
`luban status` as not needed, `rm <archive>` first.

docs/luban 03 (en, zh); design part 3 records the implementation.

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
…ecipe it cannot write; an upgrade's dry run says the move is tried first (self-review)

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
…home is mounted before the system's home (J50); the domain check's test reads bwrap only

A system in this home's prefix has its home where the host has it -- under
/home on most hosts (a runner's /home/runner/work/_temp too). The guest
mounted that home, then /home over it for the machine's users (J32,
2026.10.23.1; persist home before it): no xlings in the machine.
luban-os-ci L2 on runners: core-machine failed every scenario with exit
127. /home is mounted first now; the home's mount point is where it was
before J32 (the two shares are one host directory).

The dynamic candidate initialises a domain where bwrap makes namespaces
(CI): DomainProducer.ADomainCheck... expects success on the bwrap probe
alone.

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
… a failing namespace backend (XDEV_BWRAP)

LUBAN-RELAYOUT-TRIED skipped where the domain builds (CI), and the report
requires a passing execution of every required ID. Both J47 tests now
make the backend fail explicitly (the domain producer's XDEV_BWRAP), so
the path a host without namespaces takes is tested everywhere; the
building path stays where bwrap works.

Co-authored-by: speak-agent <248744407+speak-agent@users.noreply.github.com>
@Sunrisepeak

Copy link
Copy Markdown
Member Author

Self-review (before merge)

angle finding outcome
architecture one writer per release-owned package: tools/bump_index.py, the dead bump-index job removed; cross-repo contract read through the client (package_info.versions), not recipe text in this PR; the index side (U6) checks the same read in xim-pkgindex#1031
stability the first CI run found J50: a machine with users of its own mounted /home over a prefix system's home (under /home on runners and on most hosts) -- no xlings in core-machine (luban-os-ci L2: 12 failed, also on published 2026.10.23.1). Fixed (mount order); reproduced and fixed locally with --work under /home; L2 on runners now 17 passed, 0 failed in this PR (part 3 §2.5)
stability bump_index.py could move latest backwards for an older version; a recipe without xpm raised a traceback fixed; a test each
tests the J47 test assumed only a static client initialises a domain (wrong on runners); LUBAN-RELAYOUT-TRIED skipped on CI, so the report had no passing evidence both branches now run everywhere (a failing XDEV_BWRAP)
UX luban upgrade --dry-run promised a move this host might not make says it is tried first
compatibility luban asks __domain-check only when xlings names domain_check; protocol 1.18 is additive; the fixture's read needs a client with package_versions -- the candidate under test no change for older pairs
cross-platform __domain-check is Linux-only (exit 1 elsewhere); relayout and the guest init are Linux-only paths --
consistency docs: AGENTS.md, contributing skill, docs/92-release (en, zh), docs/luban/03 (en, zh), interface spec (1.17 row was missing), design part 3 --
seamless upgrade a system already in the prefix is not moved again (.stays); an archive an earlier move left is named as freeable, never deleted for the user --

CI: all checks green (asan shards run on main). luban-os-ci L2 for this branch: 17 passed, 0 failed (L2-06's known is mcpp#831, U10).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant