Repository navigation
feat(registry): spike fixture package builds and registry constraints - #761
Conversation
Quest-Claim: ecf7c9e0a12360fc
|
Maintainer decision (2026-11-16, via /quest-run):
(Written by Claude Opus 5.5) |
|
Applied the maintainer's decision in 9677e99: accept the report's no-go, with release templates and shared builders as the full version. The quest records the decision and rejected alternatives; the PR body records it under Decisions. Added the release templates follow-up for moq's Python and Swift releases. The remaining reusable-workflow validation is now running through the branch's manual Check invocation. No merge has been selected. (Written by GPT-6) |
kixelated
left a comment
There was a problem hiding this comment.
Automated review by review (OpenAI)
Head: 31f4aa2c31cc07f7de4639687717b3f7d3b78d12
No blocker for the fixture spike. The CPython tag matches the generator’s stable-ABI setting, and the clean consumers compare installed artifacts against native Rust rather than importing from the build tree. The Swift artifact establishes the host slice, not a portable release matrix. Shared builders plus per-project release templates are the right boundary here: the fixture proves packaging, not registry ownership or publishing policy.
The Node launcher now registers signal handlers before spawn, and the regression exercises a real OS signal during the startup window. I verified the successful consumer jobs at 13710f7b, not this head. Since this head changes the workflow timeouts and packaging checks, rerun Check here before treating that evidence as current. Static review; I did not build or publish anything.
There was a problem hiding this comment.
ℹ️ No critical issues — one pinning nit inline.
Reviewed changes
The full registry-packaging spike at 31f4aa2: the reusable workflow and its callers, the package builders and consumers, the crabx launcher signal fix, the Swift fixture rename and the quest handoff.
- Reusable
registry.yml: builds the api fixture's wheel and XCFramework on Linux arm64 and macOS arm64, then checks the installed artifacts in fresh consumer jobs that have no checkout. Runs in-repo use the triggering commit; external callers use a fixed revision. - Callers: the manual
CheckandNightlyworkflows invoke it. Nightly'sreportjob now counts it, so a spike regression opens the nightly-failure issue. package.py/build.sh: the wheel gets a complete hashedRECORDand native-dependency allowlists (RUNPATH stripped and glibc ≤ 2.39 on Linux;@rpathID and system-onlyotool -Lon macOS). The Swift package's link settings come from rustc'snative-static-libs. On macOS the build uses Xcode's SDK and Apple's linker.- Consumers:
consume.pycompares the Python transcript byte for byte, prints a unified diff on mismatch, then runs the extra cases.consume-swift.shbuilds a fresh SwiftPM executable against the unpacked package. - crabx launcher: supervisor-signal handlers are now installed before
spawn. Before this, aSIGTERMthat arrived during startup hit the default disposition and killed the launcher. The new regression test sends a real signal from inside a mockedspawn. - Swift fixture: renames a local that shadowed the
debug()helper under Swift 6. - Quests: deletes
quest/m2/registry.mdand adds therelease-templatesfollow-up. No stale links to the deleted file remain.
claude-opus-5-5 | 𝕏
kixelated
left a comment
There was a problem hiding this comment.
Automated review by review (OpenAI)
Head: 8d942c3477176729d4861297cc041c8e3b4ebb19
No new blocker in the downloader update. The new pinned SHA already matches crabx.yml and nightly.yml, and the fixture still downloads an exact artifact name from its own run. The earlier successful build/install transcript predates this pin change, and ordinary PR checks don't run the dispatch-only registry fixture. Capture a fresh fixture run before treating the updated workflow as validated for reuse. Static review; I did not dispatch anything.
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
The delta since the prior Pullfrog review at 31f4aa2.
- Aligned the
download-artifactpin:registry.ymlnow pinsactions/download-artifactat v8.0.2 (9000827c…), matchingnightly.ymlandcrabx.yml, which resolves the earlier inline finding. - Recorded the merge pick:
quest/m2/release-templates.mdnotes the maintainer's decision to keep the follow-up unchanged.
claude-sonnet-5-5 | 𝕏

Problem
Completes the registry-packaging spike: can one workflow replace hand-maintained binding releases across registries?
Approach
Add a no-input reusable workflow for the fixed api fixture, native CPython wheels, a host macOS XCFramework plus Swift source package, and clean consumer jobs comparing each installed package with native Rust. The macOS builder selects the installed Xcode SDK and exports the fixture suite's current WASI target. The Swift fixture avoids a helper name shadow in Swift 6. Also close a base-branch Node launcher signal race exposed by local CI, with a deterministic regression test. The maintainer accepted the no-go and chose release templates with shared builders. Both artifact consumers passed in GitHub Actions; the completed spike quest is deleted.
The report surveys Maven, npm, NuGet, pub.dev and Go. Accepted direction: release templates with shared builders. Cargo metadata cannot establish registry namespace ownership.
Impact
registry.ymlworkflow call for this fixture only. No caller-crate packaging or publication is implemented.Alternatives
Use maturin for Python alone; reuse BoltFFI's packaging but accept its different generation model; infer names from Cargo. None solves existing registry identities without platform setup.
Decisions
Maintainer merge pick, 2026-11-16: ✅
/quest-merge; keep the release-templates follow-up unchanged (rejected: drop it). The coordinator will queue the resulting head.Maintainer decision, 2026-11-16: ✅ accept the report's no-go; the full version is release templates with shared builders (rejected: keep investigating a configless publisher; stop the spike with no full version). Cargo metadata cannot establish registry identities and hosting policy.
Choices
Validation
Pullfrog pin fix at
8d942c3477176729d4861297cc041c8e3b4ebb19: applied the exact v8.0.2 download-artifact pin already used in nightly.yml and crabx.yml; checked the pinned action'snameandpathinputs.quest gate nix develop --command just ciandgit diff --checkpassed. Claude round 3 reviewed this head at 9/10 with no findings. The finding thread is answered and resolved. No additional artifact-consumer run was needed for this pin alignment.Repository-pinned
nix develop --command quest check: 43 documents pass.actionlint, shellcheck and shfmt for the new scripts pass.Final
quest gate nix develop --command just ci: exit 0 after integration fixes, includingapi_swiftand all 92 TS differential fixtures.git diff --checkalso passed.quest gate nix develop --command bench/registry/build.sh: Linux x86_64 wheel built.Installed the wheel via uv into a fresh CPython 3.14.10 venv with no index/dependencies; ran from a clean directory with
-I: byte-for-byte native transcript plus futures, ownership, values and calls passed.Independently verified wheel RECORD coverage, sizes and SHA-256 hashes. After review, stripped Linux RUNPATH and bounded native dependencies and GLIBC versions; installed consumer passed again.
Initial export failed because imported Cargo build directories lacked CACHEDIR.TAG; repaired only local generated cache markers. No harness or cache tool was disabled.
macOS validation on kixtop, 2026-11-16: macOS 26.6.2 arm64, Xcode 26.6 (17F113), Apple Swift 6.3.3. Both external-call pin 901f7baa and the merged source at 8b7af48d passed. Main was merged only to resolve the Swift fixture conflict, retaining its lossless
RustDurationimplementation.Ran
nix develop --command quest gate bench/registry/build.shvia gated SSH: exit 0;xcodebuild -create-xcframeworkproducedApi.zipwith the C static library and Swift source package.Copied only
Api.zip,native.txt,consumer.swift, andconsume-swift.sh(asverify-swift.sh) into a fresh artifact directory, then ranunzip -q Api.zip. Ran the verifier vianix develop --command quest gate env -i HOME=/Users/kixelated PATH=/usr/bin:/bin:/usr/sbin:/sbin /bin/bash <artifact-directory>/verify-swift.sh <artifact-directory>: exit 0. Apple's SwiftPM built the release consumer, its transcript matched native Rust byte for byte, then it printedfutures okandowned ok.Independent checks:
plutil -p Api/ApiFFI.xcframework/Info.plistandlipo -info Api/ApiFFI.xcframework/macos-arm64/libapi.aconfirmed one macOS arm64 slice.shasum -a 256 native.txt consumer.*/actual.txtreturned729ffddcec8928c564d29fd1ee3c4fc34ccaecf18900d34be63c4e6fd0a50714for both merged-source transcripts.Reproduced and fixed the Mac failures: Nix's incomplete SDK/developer-directory selection, a Swift 6 helper name shadow, and the spike's stale wasm target. The older external pin also needed an explicit
@Sendablehelper type; the merged base no longer captures that helper. Apple Swift still warns about the fixture's separate shared-call stress test's pointer capture; this consumer does not execute that stress test.The first attempt to launch the merged consumer referenced a garbage-collected Quest store path. Using the current pinned CLI to launch the clean Apple environment above succeeded.
CI exposed a timeout in crabx's Node signal test. A deterministic regression sent real SIGTERM during spawn and reproduced premature launcher death (exit -15). Installing signal handlers before spawning the child fixes the startup window; the regression passed after the fix and checks signal forwarding and the child's exit code. The final full CI run passed all 28 release tests and all remaining suites.
Ran
gh workflow run check.yml --repo kixel-ai/rust2 --ref quest/m2/registry. Direct dispatch of the newregistry.ymlreturned 404 because it is not on the default branch; the existing registered Check workflow now invokes it locally. First run 43968322384 built both packages and passed the Linux consumer, but the clean Mac wheel import exposed a Nix-store libiconv dependency. Fixed the builder to use Apple's linker; wheel packaging now normalizes the dylib ID and rejects non-system dependencies. Kixtop was unreachable for this repeat, so validation used hosted macOS.Run 43971814703 invoked the reusable workflow at
13710f7b7f1dcc649fa5cdabff348b497bc835ff. Both builders passed and uploadedregistry-ubuntu-24.04-armandregistry-macos-26. Separate consumers had no checkout, Nix, Rust or exporter: they created fresh venvs, installed viapip install --no-index --no-deps artifacts/*.whl, and ran the verifier with Python-Ifrom a clean directory. Linux arm64 and macOS arm64 each printedinstalled wheel: transcript matches Rust; futures, ownership, values and calls pass.The same macOS 26.7.2 consumer unpacked
Api.zipand ranbash verify-swift.sh .: release SwiftPM build passed, transcript matched Rust byte for byte, thenfutures okandowned ok. Both consumer workspaces were cleaned. The builder's independentotool -Lcheck showed only the relative dylib ID, system libiconv and libSystem. The external fixture pin now uses this validated revision.Downloaded both run artifacts with
gh run download: independently checked complete wheel RECORD coverage, sizes and SHA-256 hashes, and the XCFramework plist's single macOS arm64 slice. Both packaged native transcripts have SHA-256729ffddcec8928c564d29fd1ee3c4fc34ccaecf18900d34be63c4e6fd0a50714.All five jobs in the dispatched Check run finished successfully, including its existing full
just gatejob.Review hardening: final
quest gate nix develop --command just cipassed with the job timeouts, Linux libm allowlist, Swift-only revision parsing and mismatch diagnostics. A forced transcript mismatch printed a byte-level unified diff and exited 1; importing the wheel packager while refusing reads of ffi/justfile passed. Two consecutive pinned-toolchaincargo rustc -- --print=native-static-libsruns replayed the same note, including the cached 0.00 s run. Real Node signal probes proved that its event-loop turn is required; the signal regression passed in the full suite.No moq migration, release publication, iOS or universal Apple packaging is claimed.
Caused by
debug()helper in Swift 6 (found by check).Follow-ups
Models
Size: M
Implementer: Codex, gpt-6.1-sol, high effort
Implementer: Codex, gpt-6, unknown effort (iterate, XS)
Review
Rewritten by
quest reviewafter each round.Head: 8d942c3
Base: origin/main at 2c10cde
Risk: high,
it changes .github/workflows/check.yml, which is always high, as are 2 more of its paths: workflows run with the repository's secrets and decide what merges and shipsClassified: high by the always-high paths
Reviewer: Claude Code, claude-opus-5-5, medium effort
Session:
6d852b42-26ff-4ca5-88a2-1a1ddf458cb1, round 3Score: 9/10,
The new commit only changes the download-artifact pin to the v8.0.2 SHA the other workflows use and adds a plan note, so no findings remain, but the registry CI wiring is still high risk.Rounds: 3; first round 9 (medium 3, low 6); findings 9 (medium 3, low 6): fixed 4, refuted 5, open 0
Findings: none
Summary by CodeRabbit
(Written by GPT-6)
Summary by CodeRabbit