Repository navigation
Build changed packages on pull requests with the release builder's container - #388
Closed
spencerbull wants to merge 1 commit into
Closed
spencerbull wants to merge 1 commit into
spencerbull wants to merge 1 commit into
Conversation
…ntainer A package could reach master with nothing but self-tests having run, and the first time its dependencies were resolved inside the channel's container was the release build. claude-desktop (omacom#348) was reviewed on bare makepkg, which never resolves dependencies on a machine that already has them, and then failed its first rc release build. A plan job diffs pkgbuilds/ against the base and asks package_builds_for_mirror which mirrors each changed package ships to, so the policy stays in one helper. A matrix job then runs bin/build per mirror for x86_64, the same command and container the release builder uses. Package names are validated and passed through env rather than interpolated into a script, because a pull request chooses its own directory names. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
joshuaswarren
added a commit
to joshuaswarren/omarchy-pkgs
that referenced
this pull request
Nov 10, 2026
Wheel and vendored set rebuilt at v0.7.40 (7ec53476). The release carries the KV cache pre-size and the prefill chunk re-choice (omacom#390, omacom#392) and the assistant server stop fix (omacom#388). Honeykrisp stays at 439dc032d7. The driver file is byte-identical to the one in 0.7.39-2. The vendored set resolved annotated-doc 0.0.6 and markdown-it-py 4.3.0 at build time.
ryanrhughes
pushed a commit
that referenced
this pull request
Nov 12, 2026
…tant stop fix (#839) * omarchy-mlx 0.7.35: !strip (sealed ANE files), patch dependency, package install hook * omarchy-mlx 0.7.38: 6.8% faster decode on the M1 Max Pins the v0.7.38 tag tarball, the cp314 aarch64 wheel and the vendor wheel tar. Packaging files are unchanged since 0.7.35, and the Mesa pin is unchanged. * omarchy-mlx 0.7.38-2: Honeykrisp with the G14X tracked CDM barrier Pin Mesa to joshuaswarren/mesa-1 ef4219af57 (honeykrisp-omarchy-v3 plus one commit: tracked CDM barrier by default on G14X gen 14). On a 14-inch M2 Max the packaged driver issued a full CDM flush per dispatch. Measured, Qwen3-4B 4bit decode, ABBA, same token ids: ctx512 16.249 to 14.749 ms/tok, ctx2048 19.351 to 17.258 ms/tok. G13G and G13X code paths are unchanged. * omarchy-mlx 0.7.39: KV cache grows in place, up to 37% faster decode at long context The KV cache was copied in full on every decode step (72 copies per token). It now grows in place (0 copies). Measured on a 14-inch M2 Max, same prompt, same tokens, A/B on separate builds: 36.6% faster at 16384 context, 27.4% at 8192, 8.2% at 512. A 13-inch M1 and a 16-inch M1 Max pass the same token check at 8192 context (34.9% and 29.3% faster). The Vulkan driver pin stays at ef4219af57 (Honeykrisp with the tracked CDM barrier). The wheel and vendor set are built from the v0.7.39 tag in a clean Arch Linux ARM container. * omarchy-mlx 0.7.39-2: Honeykrisp lineage 439dc032d7, no non-ASCII venv alias Mesa moves from ef4219af57 to 439dc032d7 (joshuaswarren/mesa-1 w7b/hk-lineage-hoist-bf16). That lineage adds the SPIR-V image availability fix on top of the G14X tracked CDM barrier, the bf16 cooperative matrices and the G13 compute load hoist. The archive hash is fc26edaef4b6025435416b9ac9d145869b967b5ce564d7a216ab3a79bd6a6ecf. A Python 3.14 venv holds a symlink named U+1D70B plus "thon". pacman -U in a POSIX locale prints success and stops extracting at that name. On 0.7.39-1 it left 13216 of 13282 files missing and a 180 KB venv. build() now deletes any non-ASCII path from the stage and check() fails the build if one remains. With this change, pacman -U in the same locale installs all 13281 files and the venv is 783 MB. * omarchy-mlx 0.7.40: no prefill out-of-memory on the 13-inch M1 Wheel and vendored set rebuilt at v0.7.40 (7ec53476). The release carries the KV cache pre-size and the prefill chunk re-choice (#390, #392) and the assistant server stop fix (#388). Honeykrisp stays at 439dc032d7. The driver file is byte-identical to the one in 0.7.39-2. The vendored set resolved annotated-doc 0.0.6 and markdown-it-py 4.3.0 at build time.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A package can reach master with only the self-tests having run, so the first time its dependencies are resolved inside the channel's container is the release build. claude-desktop (#348) was reviewed on bare
makepkgon a machine that already had every dependency installed, which provespackage()and the artifact and nothing about resolution, and then failed its firstrcrelease build on "Could not resolve all dependencies" while the same commit shipped to edge and stable. Whether that failure was the package or the builder could only be answered by running the real thing, and nothing had.This adds a workflow with two jobs.
plandiffspkgbuilds/against the base and askspackage_builds_for_mirror— the same helper the release builder uses — which mirrors each changed package ships to, so ring and channel policy stays in one place: edge always, rc and stable forrelease_ring: fast, pinned packages to edge only.buildthen runsbin/build --package … --mirror … --arch x86_64per mirror as a matrix, the same command and container the release builder runs. A pull request chooses its own directory names, so names are validated against^[a-z0-9@._+-]+$and handed to the build step throughenv, never interpolated into a script. With nothing changed underpkgbuilds/, the plan is empty and the build job is skipped in seconds, so the check is always present.workflow_dispatchtakes an explicit package list for rebuilding on demand.x86_64 only: an emulated aarch64 build of a half-gigabyte Electron tree is not a check anyone would wait for, and #171 covers native ARM runners. The proof is a run of this workflow on a fork against a touched claude-desktop, where edge, rc and stable each built in six to seven minutes on
ubuntu-latest: https://github.com/spencerbull/omarchy-pkgs/actions/runs/34531142335🤖 Generated by Fable 5 in Claude Code.