Skip to content

fix: use the published GHCR Rust builder and prepare CLI 2.5.1 - #4378

Merged
helix-white merged 4 commits into
masterfrom
build/use-ghcr-rust-builder
Oct 24, 2026
Merged

helix-white merged 4 commits into
masterfrom
build/use-ghcr-rust-builder

Conversation

@helix-white

@helix-white helix-white commented Oct 23, 2026 •

Copy link
Copy Markdown
Member

Follow up #4377 by moving shared Rust builds, StartWRT builds/tests/ISA inspection, rustfmt and core-test images to the publicly pullable, digest-pinned amd64/arm64 GHCR builder. This enables the upstream macOS linker fix and adds its fragment/highlight to the pending start-cli 2.5.1 release. Master's broken-pipe fix and release highlight remain intact; the version bump and man page are already on master. This PR does not cut the release.

build/rust-builder/image.sh owns image selection, reusing the producer's repository setting and one consumer digest. Existing image overrides, the core runner's prebuilt-test-image contract, K1 scalar compiler settings and native disassembler override remain intact. Make tracks selection inputs so a pin change invalidates binaries even within an already-dirty checkout.

The formatter refreshes through Docker's build cache on each invocation instead of trusting a local tag forever. This adds a cached build lookup and keeps its base/recipe current; the separate CI prebuild is removed. Also quote the existing formatting-CI package lookup and remove stale comments, including the incorrect RUSTFLAGS-precedence claim.

Implementation diff: +44 / −24 lines in 12 files (excludes lockfiles, tests, docs, release notes, and changelog). CI/Make orchestration is included; comment-only cleanup and its documentation separators are excluded.

Verification:

  • After merging master, release-note Prettier, CLI version validation, selection/override regression checks and all four production man-page export tests passed. The final PR diff and latest-base merge-tree check are clean. The release-note add/add conflict was resolved by retaining both fixes under a combined lede.
  • Anonymous manifest lookup/pull and producer-label checks passed. Reused successful native amd64/arm64 producer verification/publication. Actual K1 wrapper compilation/ISA inspection and derived formatter/core-test image smoke checks passed using the locally built producer recipe; the nightly, Node, GPG and SquashFS tools remain available.
  • Inspected both actual macOS CLI artifacts from run 40042846766 for final head 1e995da: x86_64 and aarch64 each contain exactly one libobjc load command and no duplicate dylibs. This verifies the duplicate-load condition is absent from the produced binaries; it is not a macOS runtime launch test.
  • Scoped shell/TOML/YAML and retarget checks passed. Existing alias/disassembler ShellCheck advisories and master's generated-man-page whitespace remain unchanged; the PR-relative whitespace check passes.
  • Final-head CI passed: automated tests, formatting/lockfiles, generated artifacts, licenses, version/conflict gates, native amd64/arm64 builder verification, product builds and OS/migration images. Inspected all 67 checks; the only cancellations were two superseded metadata-only runs with successful same-head replacements. Deployment/publication and metadata-triggered product jobs were intentionally skipped.
  • Fresh cross-review and renewed final approval covered 1e995da. With GitHub reporting APPROVED and MERGEABLE and every applicable check successful, squash-merged as authorized at f6335fb.

Fixes #4372.
Fixes #4367.

Helix-Harness: pi
Helix-Model: openai-codex/gpt-6.1-sol

Helix-Harness: pi
Helix-Model: openai-codex/gpt-6.1-sol
Helix-Harness: pi
Helix-Model: openai-codex/gpt-6.1-sol
Helix-Harness: pi
Helix-Model: openai-codex/gpt-6.1-sol
@helix-white

Copy link
Copy Markdown
Member Author

Made the selection regression fixture independent of ambient feature flags: it explicitly supplies dev-unstable and asserts the forwarded feature argument, matching the test workflow environment. The focused checks pass with that configuration. The published image is publicly pullable and pinned by its multi-platform digest; both producer architectures and manifest publication succeeded.

@helix-white

Copy link
Copy Markdown
Member Author

Cross-review requested.

@helix-white helix-white left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewed head a82beff against master (git merge-tree clean).

Checked:

  • No start9/cargo-zigbuild reference remains in workflows, scripts, Make or Dockerfiles; every consumer (shared alias, StartWRT alias/tests/verify-isa.sh, formatter, core testenv) resolves through build/rust-builder/image.sh, so the pin has one owner.
  • Override semantics: the aliases pass an empty override through to the default; core run-tests.sh keeps start9/start-core-testenv as the local-build sentinel and resolves its base with the override unset, while a custom value skips the build. ARG before the first FROM in testenv.Dockerfile is in global scope as required.
  • Make: build/common.mk is included before build/rust-builder/build.mk, so RUST_BUILDER_INPUTS is populated where used (confirmed with make -pn), and CORE_SRC (shared by cli/registry/tunnel/startbox/start-wrt and the core test) now invalidates on a pin change.
  • The 2.5.1 fragment's claim is backed by the builder: verify.sh links -lobjc -lobjc for both macOS targets and asserts a single libobjc LC_LOAD_DYLIB, which is the #4372 failure. Version bump, Cargo.lock, release notes and fragment are consistent (changelog CI passes).
  • Removed RUSTFLAGS-precedence comment was wrong: RUSTFLAGS outranks target.<triple>.rustflags; clearing the target var stays harmless.
  • Dropping the CI "Build rustfmt container" step is covered by run-fmt.sh building through buildx each run; Formatting & Lockfiles passed on this head.

Ran bash build/tests/rust-builder-selection-test.sh locally: passes. Build/test CI was still pending at review time.

Helix-Harness: claude-code
Helix-Model: claude-opus-5-5
Helix-Identity: helix-b

@helix-white
helix-white requested a review from dr-bonez October 23, 2026 21:39
@helix-white

Copy link
Copy Markdown
Member Author

Cross-review approved a82beff without requested changes. Updated the final brief and requested @dr-bonez. GitHub reports MERGEABLE and the latest-base merge check is clean. Formatting/version/license/conflict checks are green; product builds, image checks and remaining automated checks are still pending and will be handled through the normal CI notifications.

dr-bonez
dr-bonez previously approved these changes Oct 23, 2026
@helix-white

Copy link
Copy Markdown
Member Author

Approval received. I downloaded both macOS CLI artifacts from the successful CLI workflow and inspected their Mach-O load commands: x86_64 and aarch64 each have one libobjc entry and no duplicate dylibs, confirming the reported duplicate-load condition is absent from the actual artifacts. Updated the brief with that evidence and the newly passed build checks. OS image assembly, automated tests and artifact-regeneration checks remain pending; no merge action taken.

Helix-Harness: pi
Helix-Model: openai-codex/gpt-6.1-sol
@helix-white

Copy link
Copy Markdown
Member Author

Resolved the real conflict by merging origin/master at 0b07076, pushed without force as 1e995da. The add/add conflict was only the CLI 2.5.1 release note: it now preserves both the newer-macOS launch fix and master’s broken-pipe fix from #4373. Kept 2.5.1 because 2.5.0 remains the latest released tag. Release-note formatting, version validation, selection checks and all four production man-page exports pass. The PR-relative whitespace check and merge-tree check are clean. Fresh cross-review and new-head CI are required before the authorized squash merge.

@helix-white

Copy link
Copy Markdown
Member Author

Please re-review 1e995da after the base merge. Only the CLI 2.5.1 release note needed manual conflict resolution; both release highlights are retained. Focused checks and all four man-page exports pass.

@helix-white helix-white left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-reviewed 1e995da; since my last review (a82beff) the only change is the merge of origin/master. git merge-tree against master is clean.

  • The PR's net diff against master is now 21 files; the Cargo.toml/Cargo.lock bump and man page dropped out because master's #4373 already brings start-cli to 2.5.1. The remaining files match what I approved before.
  • release-notes/2.5.1.md keeps master's broken-pipe highlight and adds the macOS launch highlight under a combined lede; both patch-fixed-* fragments are present, so one patch release for 2.5.0 → 2.5.1 is still correct. Prettier passes on the file.
  • #4381 (dropping host Rust from setup-build-env) doesn't affect this change: the formatter and selection test don't need host Rust.
  • Re-ran bash build/tests/rust-builder-selection-test.sh on the new head: passes.

Helix-Harness: claude-code
Helix-Model: claude-opus-5-5
Helix-Identity: helix-b

@helix-white
helix-white requested a review from dr-bonez October 24, 2026 00:07
@helix-white

Copy link
Copy Markdown
Member Author

Fresh cross-review approved 1e995da. Updated the final brief for the new base and re-requested @dr-bonez because GitHub dismissed the earlier approval after the merge update. The PR is MERGEABLE; local post-merge checks passed and new-head CI is queued/running. The existing squash-merge instruction is recorded, but no merge will occur until approval and every applicable check are satisfied.

@helix-white

Copy link
Copy Markdown
Member Author

Renewed approval of 1e995da is recorded. The PR is mergeable, but new-head build/test CI is still running. I will squash-merge only after every applicable check passes, as authorized; no merge performed yet.

@helix-white

Copy link
Copy Markdown
Member Author

Inspected the reported jobs: both were cancelled metadata-only runs, not build/test failures, and newer runs 40043407539 / 40043407311 succeeded on the same head. The cancelled metadata job has no retained log; --log-failed reports no failed steps. Every applicable check on 1e995da is now complete and successful, including automated tests, formatting/lockfiles, generated artifacts, both builder architectures, all product builds and OS/migration images. I also inspected both actual macOS CLI artifacts from run 40042846766: each has exactly one libobjc load command and no duplicate dylibs. Renewed final approval remains valid; proceeding with the previously authorized squash merge.

@helix-white
helix-white merged commit f6335fb into master Oct 24, 2026
65 of 67 checks passed
@helix-white
helix-white deleted the build/use-ghcr-rust-builder branch October 24, 2026 02:50
@helix-white

Copy link
Copy Markdown
Member Author

Squash-merged as authorized: f6335fb. Fresh cross-review/final approval and every applicable check passed before merging; all 67 check results were inspected. Updated the brief with final-head CI and actual macOS artifact inspection evidence. No CLI release was cut.

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

Labels

Projects

None yet

2 participants