frankenbit/release-toolkit

Watch

Fork 0

bug(release): rt does not build on Windows — the cut lock calls syscall.Flock, and register-check ignores the allow-list there #1502

New issue

Closed

opened 2026-10-02 11:32:39 +02:00 by claude · 1 comment

claude commented

2026-10-02 11:32:39 +02:00

Member

rt does not compile on Windows. internal/release/lock.go takes the cut-serialization lock with syscall.Flock, and package syscall defines that only on Unix. As a result, no verb is available to anyone working on Windows. That includes the read-only ones a reviewer runs by hand: rt review-eligibility and rt review-submit-check (AGENTS.md section 6.1/6.2), rt register-check, rt fragment-check and rt repin --dry-run.

Measured 2026-10-02 at 29b541b, Go 1.27.1 windows/amd64, Git 2.55.0.windows.3, LF clone:

go build ./cmd/rt
internal\release\lock.go:69:20: undefined: syscall.Flock
internal\release\lock.go:69:47: undefined: syscall.LOCK_EX
internal\release\lock.go:69:63: undefined: syscall.LOCK_NB
internal\release\lock.go:88:23: undefined: syscall.Flock
internal\release\lock.go:88:52: undefined: syscall.LOCK_UN

With a Windows lock backend dropped in, the binary builds. Running it then shows one Windows runtime bug:

rt register-check            (repo root, default surface, lock backend only)
[register-check] FAIL: 71 register-drift hit(s) found across scanned paths.   exit 1

hits by file:  59  docs\retro\phase-7-arc.md
                5  docs\adr\0008-go-port-architecture.md
                3  docs\adr\0009-go-substrate-for-v1.md
                1  docs\adr\0010-concurrency-guard-composition.md
                1  docs\architecture\arc42\11-risks-technical-debt.md
                1  docs\architecture\arc42\12-glossary.md
                1  docs\architecture\contracts\fragment-format.md

Every one of these files is covered by a .register-allowlist entry (docs/adr/, docs/retro/ and the three named files). On Linux the same command reports 0 hits.

Cause. scanOneFile in internal/register/filescan.go receives the OS-native path that filepath.WalkDir produces (docs\retro\...). It matches that path against the forward-slash allow-list entries and self-exclusion keys, which never match, and it also prints the path as the hit's file.

Hooks. On Windows, rt prep with any post_bump_hooks configured stops with prep: post_bump_hook not executable: <hook>. No chmod +x can fix that: os.Stat reports no exec bit on Windows, and CreateProcess does not dispatch on #!. This repository no longer configures hooks, and the release workflow runs on Linux, so only a local Windows rt prep meets this.

Tests. With the lock backend in place, go test -count=1 ./... fails 11 tests natively, in 5 of the 29 packages:

the three register tests are the bug above: TestScanTree, TestRegisterFileScanDrift and TestRegisterFileScanExemptAndAllowlist;
TestRunPostBumpHooks_theHookDoesNotReceiveTheGitCredential is the hook refusal;
the rest come from the test harness, not from rt:
- a built binary named rt without .exe;
- POSIX mode bits and chmod 000;
- a Git Bash locale control;
- a staged-path lookup keyed with filepath.Join.

TestPushManifestAuthedRejectsSecondPush also fails intermittently on Windows. Its 2 s context deadline can kill git mid-rebase on a busy host. It passed in this run.

Why it matters

A reviewer on Windows cannot run the pre-submit gates the review conventions ask for, nor dry-run any verb before a cut.
register-check on Windows gives a false red on exactly the paths the allow-list exists to exempt. Its hit lines also carry backslash paths, a form that neither the allow-list format nor the docs use.
The cut lock is the only thing in the tree that pins the build to Unix. Everything else is portable Go.

Out of scope

Publishing Windows release assets. .goreleaser.yaml builds linux/amd64 only, by the open cross-platform-runners fork in ADR-0009 section 9, which is an operator ratification point.
Running post_bump_hooks on Windows. They are POSIX scripts shared with the Linux workflow, so the right move is to refuse them by name.
Windows path-handling gaps found only by reading the code, which belong in a separate issue:
- the version_files guard does not reject a ..\ prefix;
- resolveUnderRoot treats only a leading / as absolute;
- an atomic rename fails over a read-only destination.

Acceptance criteria

go build ./cmd/rt succeeds natively on Windows. Every non-Windows build keeps the identical flock call. — DONE @ #1505 (b47a824, reviewed as 8e2cf58). go build ./cmd/rt and go vet ./... succeed natively on 96d2646. lock_unix.go is //go:build !windows and carries the same two syscall.Flock calls as the old lock.go (LOCK_EX|LOCK_NB, LOCK_UN).
The Windows cut lock has the same contract as flock: exclusive, non-blocking (ErrConcurrentCut on contention), owned by the open file, and dropped by the OS when the holder dies. A second-process test and a killed-holder test cover it, and both go red under a contention-treated-as-success mutant on Windows and Linux. — DONE @ #1505 (b47a824, reviewed as 8e2cf58), with one correction to the mutant wording. Measured natively on Windows and in golang:1.25: a contention-treated-as-success mutant turns TestAcquireLock_anotherProcessHoldsIt and TestNewCutter_concurrentCut red, but not the killed-holder test, which never sees contention. A marker-file mutant (an O_EXCL file instead of the OS lock) turns TestAcquireLock_killedHolderDoesNotStrandTheLock red: stranded after 5 s. Each test bites on the property it pins, as #1505's own mutation table also says.
rt register-check on Windows honours .register-allowlist and prints forward-slash paths. On this repository it reports 0 hits, as on Linux. — DONE @ #1505 (90ef268, reviewed as 6aef15b). Native rt register-check on 96d2646: OK: 0 register-drift hits, exit 0, with stdout and stderr byte-identical to the Linux run on the same tree. Forward-slash hits are pinned by TestScanTree and TestScanTreeExplicitOSPath, both PASS natively.
On Windows, rt prep refuses configured post_bump_hooks by name, stating the actual cause, before running any of them. Linux behaviour is unchanged. — DONE @ #1505 (25d0c0b, reviewed as 0ac9790). The refusal sits at the top of the per-hook loop, ahead of any exec. TestRunPostBumpHooks_refusalNamesTheCauseOnThisOS PASSES natively with post_bump_hooks are not supported on Windows (no exec bit, no #! dispatch), and in golang:1.25 with the unchanged not executable refusal.
go vet ./... && go test -count=1 ./... passes natively on Windows (from Git Bash) and in golang:1.25. On Linux, the output of the touched verbs is byte-identical before and after. — DONE @ #1505 (96d2646, reviewed as ca66a11). Fresh clone of 96d2646: go vet ./... clean and go test -count=1 ./... 29/29, natively from Git Bash and in golang:1.25 as a non-root user. Re-measured on Linux with rt built at 340608c and at 96d2646, both run on the same tree: --help, register-check and fragment-check are byte-identical in stdout, stderr and exit code. prep --dry-run was not re-run here; its byte-identity rests on #1505's review pass (non-executable, failing and succeeding hooks), and its only production change is a runtime.GOOS == "windows" branch.
A changelog.d fragment for the change. — DONE @ #1505 (b47a824, reviewed as 8e2cf58). changelog.d/1502-windows-build.fixed.md; rt fragment-check in content mode PASSES.

Anchor

Found while testing a release-notes plan for frankenbit/storey on a Windows workstation. The Flock build failure was already listed as out of scope in #1485.

`rt` does not compile on Windows. `internal/release/lock.go` takes the cut-serialization lock with `syscall.Flock`, and package `syscall` defines that only on Unix. As a result, no verb is available to anyone working on Windows. That includes the read-only ones a reviewer runs by hand: `rt review-eligibility` and `rt review-submit-check` (AGENTS.md section 6.1/6.2), `rt register-check`, `rt fragment-check` and `rt repin --dry-run`. Measured 2026-10-02 at `29b541b`, Go 1.27.1 windows/amd64, Git 2.55.0.windows.3, LF clone: ``` go build ./cmd/rt internal\release\lock.go:69:20: undefined: syscall.Flock internal\release\lock.go:69:47: undefined: syscall.LOCK_EX internal\release\lock.go:69:63: undefined: syscall.LOCK_NB internal\release\lock.go:88:23: undefined: syscall.Flock internal\release\lock.go:88:52: undefined: syscall.LOCK_UN ``` With a Windows lock backend dropped in, the binary builds. Running it then shows one Windows runtime bug: ``` rt register-check (repo root, default surface, lock backend only) [register-check] FAIL: 71 register-drift hit(s) found across scanned paths. exit 1 hits by file: 59 docs\retro\phase-7-arc.md 5 docs\adr\0008-go-port-architecture.md 3 docs\adr\0009-go-substrate-for-v1.md 1 docs\adr\0010-concurrency-guard-composition.md 1 docs\architecture\arc42\11-risks-technical-debt.md 1 docs\architecture\arc42\12-glossary.md 1 docs\architecture\contracts\fragment-format.md ``` Every one of these files is covered by a `.register-allowlist` entry (`docs/adr/`, `docs/retro/` and the three named files). On Linux the same command reports 0 hits. **Cause.** `scanOneFile` in `internal/register/filescan.go` receives the OS-native path that `filepath.WalkDir` produces (`docs\retro\...`). It matches that path against the forward-slash allow-list entries and self-exclusion keys, which never match, and it also prints the path as the hit's file. **Hooks.** On Windows, `rt prep` with any `post_bump_hooks` configured stops with `prep: post_bump_hook not executable: <hook>`. No `chmod +x` can fix that: `os.Stat` reports no exec bit on Windows, and `CreateProcess` does not dispatch on `#!`. This repository no longer configures hooks, and the release workflow runs on Linux, so only a local Windows `rt prep` meets this. **Tests.** With the lock backend in place, `go test -count=1 ./...` fails 11 tests natively, in 5 of the 29 packages: - the three register tests are the bug above: `TestScanTree`, `TestRegisterFileScanDrift` and `TestRegisterFileScanExemptAndAllowlist`; - `TestRunPostBumpHooks_theHookDoesNotReceiveTheGitCredential` is the hook refusal; - the rest come from the test harness, not from `rt`: - a built binary named `rt` without `.exe`; - POSIX mode bits and `chmod 000`; - a Git Bash locale control; - a staged-path lookup keyed with `filepath.Join`. `TestPushManifestAuthedRejectsSecondPush` also fails intermittently on Windows. Its 2 s context deadline can kill git mid-rebase on a busy host. It passed in this run. ## Why it matters - A reviewer on Windows cannot run the pre-submit gates the review conventions ask for, nor dry-run any verb before a cut. - `register-check` on Windows gives a false red on exactly the paths the allow-list exists to exempt. Its hit lines also carry backslash paths, a form that neither the allow-list format nor the docs use. - The cut lock is the only thing in the tree that pins the build to Unix. Everything else is portable Go. ## Out of scope - **Publishing Windows release assets.** `.goreleaser.yaml` builds linux/amd64 only, by the open cross-platform-runners fork in ADR-0009 section 9, which is an operator ratification point. - **Running `post_bump_hooks` on Windows.** They are POSIX scripts shared with the Linux workflow, so the right move is to refuse them by name. - **Windows path-handling gaps found only by reading the code**, which belong in a separate issue: - the `version_files` guard does not reject a `..\` prefix; - `resolveUnderRoot` treats only a leading `/` as absolute; - an atomic rename fails over a read-only destination. ## Acceptance criteria - [x] `go build ./cmd/rt` succeeds natively on Windows. Every non-Windows build keeps the identical `flock` call. — DONE @ #1505 (`b47a824`, reviewed as `8e2cf58`). `go build ./cmd/rt` and `go vet ./...` succeed natively on `96d2646`. `lock_unix.go` is `//go:build !windows` and carries the same two `syscall.Flock` calls as the old `lock.go` (`LOCK_EX|LOCK_NB`, `LOCK_UN`). - [x] The Windows cut lock has the same contract as flock: exclusive, non-blocking (`ErrConcurrentCut` on contention), owned by the open file, and dropped by the OS when the holder dies. A second-process test and a killed-holder test cover it, and both go red under a contention-treated-as-success mutant on Windows and Linux. — DONE @ #1505 (`b47a824`, reviewed as `8e2cf58`), with one correction to the mutant wording. Measured natively on Windows and in `golang:1.25`: a contention-treated-as-success mutant turns `TestAcquireLock_anotherProcessHoldsIt` and `TestNewCutter_concurrentCut` red, but not the killed-holder test, which never sees contention. A marker-file mutant (an `O_EXCL` file instead of the OS lock) turns `TestAcquireLock_killedHolderDoesNotStrandTheLock` red: stranded after 5 s. Each test bites on the property it pins, as #1505's own mutation table also says. - [x] `rt register-check` on Windows honours `.register-allowlist` and prints forward-slash paths. On this repository it reports 0 hits, as on Linux. — DONE @ #1505 (`90ef268`, reviewed as `6aef15b`). Native `rt register-check` on `96d2646`: `OK: 0 register-drift hits`, exit 0, with stdout and stderr byte-identical to the Linux run on the same tree. Forward-slash hits are pinned by `TestScanTree` and `TestScanTreeExplicitOSPath`, both PASS natively. - [x] On Windows, `rt prep` refuses configured `post_bump_hooks` by name, stating the actual cause, before running any of them. Linux behaviour is unchanged. — DONE @ #1505 (`25d0c0b`, reviewed as `0ac9790`). The refusal sits at the top of the per-hook loop, ahead of any exec. `TestRunPostBumpHooks_refusalNamesTheCauseOnThisOS` PASSES natively with `post_bump_hooks are not supported on Windows (no exec bit, no #! dispatch)`, and in `golang:1.25` with the unchanged `not executable` refusal. - [x] `go vet ./... && go test -count=1 ./...` passes natively on Windows (from Git Bash) and in `golang:1.25`. On Linux, the output of the touched verbs is byte-identical before and after. — DONE @ #1505 (`96d2646`, reviewed as `ca66a11`). Fresh clone of `96d2646`: `go vet ./...` clean and `go test -count=1 ./...` 29/29, natively from Git Bash and in `golang:1.25` as a non-root user. Re-measured on Linux with `rt` built at `340608c` and at `96d2646`, both run on the same tree: `--help`, `register-check` and `fragment-check` are byte-identical in stdout, stderr and exit code. `prep --dry-run` was not re-run here; its byte-identity rests on #1505's review pass (non-executable, failing and succeeding hooks), and its only production change is a `runtime.GOOS == "windows"` branch. - [x] A `changelog.d` fragment for the change. — DONE @ #1505 (`b47a824`, reviewed as `8e2cf58`). `changelog.d/1502-windows-build.fixed.md`; `rt fragment-check` in content mode PASSES. ## Anchor Found while testing a release-notes plan for frankenbit/storey on a Windows workstation. The Flock build failure was already listed as out of scope in #1485.

claude added the

labels

2026-10-02 11:32:39 +02:00

claude referenced this issue from a commit

2026-10-02 11:37:51 +02:00

fix(release): take the cut lock with LockFileEx on Windows (#1502)

claude referenced this issue from a commit

2026-10-02 11:37:51 +02:00

fix(register): match and print forward-slash paths on Windows (#1502)

claude referenced this issue from a commit

2026-10-02 11:37:51 +02:00

fix(prep): refuse post_bump_hooks on Windows by name (#1502)

claude referenced this issue from a commit

2026-10-02 11:37:51 +02:00

test: make the suite pass natively on Windows (#1502)

claude referenced this issue

2026-10-02 11:39:20 +02:00

fix(release): rt builds and runs on Windows (#1502) #1505

alex referenced this issue from a commit

2026-10-02 22:19:06 +02:00

fix(release): take the cut lock with LockFileEx on Windows (#1502)

alex referenced this issue from a commit

2026-10-02 22:19:06 +02:00

fix(register): match and print forward-slash paths on Windows (#1502)

alex referenced this issue from a commit

2026-10-02 22:19:06 +02:00

fix(prep): refuse post_bump_hooks on Windows by name (#1502)

alex referenced this issue from a commit

2026-10-02 22:19:06 +02:00

test: make the suite pass natively on Windows (#1502)

release-bot referenced this issue

2026-10-02 22:20:14 +02:00

chore(release): v0.65.0 #1523

claude referenced this issue

2026-10-02 22:25:09 +02:00

bug(windows): native go test ./... still fails on an LF checkout — cmd/rt does not build, and POSIX-only tests run unguarded #1501

claude commented

2026-10-02 22:34:52 +02:00

Author

Member

Landed as #1505: b47a824..96d2646, reviewed as 8e2cf58..ca66a11. git range-diff shows all four commits as =. All six criteria are ticked above against live main.

Approval: alex merged #1505 after a brief look-through. There is no formal approval row, and the landing identity record (comment 114368) says so.
Checks: the combined status on 96d2646 is success, 16 of 16.
One wording correction, recorded on the second criterion: the contention mutant reddens the second-process test, and a marker-file mutant reddens the killed-holder test. Measured on both OSes.

The competing #1503 was shut unmerged, and its tracker #1501 is ticked against the same landing.

Shutting this tracker by hand.

Landed as #1505: `b47a824..96d2646`, reviewed as `8e2cf58..ca66a11`. `git range-diff` shows all four commits as `=`. All six criteria are ticked above against live `main`. - **Approval:** alex merged #1505 after a brief look-through. There is no formal approval row, and the landing identity record (comment 114368) says so. - **Checks:** the combined status on `96d2646` is success, 16 of 16. - **One wording correction,** recorded on the second criterion: the contention mutant reddens the second-process test, and a marker-file mutant reddens the killed-holder test. Measured on both OSes. The competing #1503 was shut unmerged, and its tracker #1501 is ticked against the same landing. Shutting this tracker by hand.

claude closed this issue

2026-10-02 22:34:52 +02:00

claude referenced this issue

2026-10-02 23:24:42 +02:00

bug(windows): three path gaps - the #181 version_files guard misses ..\ paths, recover-pending-cut joins an absolute changelog, an atomic write fails over a read-only file #1526

claude referenced this issue

2026-10-02 23:41:27 +02:00

fix(windows): version_files guard, recover's absolute changelog path and read-only renames (#1526) #1527

claude referenced this issue

2026-10-04 11:07:48 +02:00

bug(prep): on Windows, rt prep --dry-run refuses when another program holds an ignored file open — the scratch copy reads every ignored file #1543