test(source-control, playbooks): five suites fail on Windows Git Bash from path form and line endings

[kyle-sexton]

Problem

Five shell suites fail when run locally under Windows Git Bash on current main and pass in CI on Linux. CI's Windows lane (.github/workflows/test-windows.yml) does not run any of them, so nothing in CI catches the gap. Every agent working on Windows sees these red when it verifies a change to playbooks or source-control locally. It then has to work out that the failures came before its change, or push to CI to get a trustworthy signal.

Most of the failures fall into two classes. In the first, a path is compared in POSIX /tmp/... form against the /c/Users/.../AppData/Local/Temp/... form Git Bash resolves. In the second, CRLF/LF line endings or a trailing \r break a byte-exact comparison. #5899 (head 4cf65d5fb) first reported them. They reproduce unchanged on today's main.

Reproduction

On Windows 11 with Git Bash, using a detached worktree at origin/main f605a5963 (2026-10-03), run bash <suite> for each suite:

suite	result	failing assertions
plugins/playbooks/skills/repo-sweep/scripts/tick.test.sh	3 FAILED	in-progress: body bytes, done lines: body bytes, filed: body bytes. The expected bytes have run:\r\n- [ ] and the actual have run:\n- [ ], so a CRLF body is rewritten with LF.
plugins/playbooks/skills/repo-sweep/scripts/state.test.sh	4 FAILED	reconcile: committed lines, web-UI tick, exact skill sets, base excluded, CRLF body from a subdirectory: same report, other branch, one open sweep: exit 14 names the branch, other branch, two open sweeps: exit 15 lists them. For the two exit-14/15 cases, expected and actual print identically, even under cat -A with no \r visible. The difference is in something the test compares but does not print, and its cause is not found yet.
plugins/source-control/scripts/worktree-claim.test.sh	5 FAIL	[4] report names the plain-add path, [19] foreign claim names the path, [31] unclaimed entry names the path. Each expects /tmp/tmp.X/external/... and gets /c/Users/KyleSexton/AppData/Local/Temp/tmp.X/external/.... Also [73] stale: a scan that fails mid-way is not provable (exit 1) (got 0) and [74] stale: the refusal says the scan failed.
plugins/source-control/scripts/lib/worktree-facts.test.sh	6 of 21 FAIL	claimed after a bare add: the claim written is the claim read (actual is empty), a reason wrapped in double quotes round-trips unchanged, a literal backslash-n and backslash pair round-trip unchanged, list prints the raw quoted reason in the lock_reason column, list prints the raw backslash reason in the lock_reason column, list: a reasonless lock reads locked=yes, lock_reason=-
plugins/source-control/scripts/worktree-root-scan.test.sh	2 FAIL	[14] symlink is not proposed: expected linked\tsymlink\tno, got linked\tempty\tyes. Git Bash without developer-mode symlinks copies instead of linking (assumption). Also [24] default root scans the configured root: the same /tmp vs /c/Users/.../Temp path form as above.

The other source-control worktree suites pass on the same checkout: worktree-add-claim-gate, worktree-add-containment-gate, worktree-create-gate, worktree-create, worktree-root-doctor, and skills/worktree/reattach-slice. worktree-gates-spawn-budget skips because strace is unavailable.

Impact

• Local verification time on Windows. An agent cannot get a green local run for playbooks or source-control changes, so it either investigates failures that existed before its change or pushes to CI to find out. Both add a CI round to time-to-green.
• The worktree-claim and worktree-facts failures cover claim parsing on the platform where most lanes run. If any of these failures is a real Windows behaviour gap rather than a test-fixture gap (unverified), the Windows behaviour of the claim gate is untested.

Proposed fix

1. Path form. In the three source-control suites, normalise both sides before comparing: resolve the fixture root once with cd "$TMP" && pwd -P (or cygpath -u when it is present) and build the expected paths from that, instead of the literal /tmp/... that mktemp returned.
2. Line endings. In tick.test.sh, and in state.test.sh once its hidden difference is found, find out whether tick.sh/state.sh should preserve CRLF bodies, which the "CRLF body" case suggests. If so, fix the script, which is likely a sed, awk or read that drops \r. If not, build the expected fixture bytes with the same line ending the script emits. Fix it in the script if the web-UI tick path really receives CRLF bodies from GitHub.
3. worktree-facts claim round-trip. Find out why the claim read after a bare add is empty on Windows: whether the lock-reason file path, git worktree lock --reason, or quoting differs.
4. Symlink case. Skip [14] with a SKIP: line when ln -s produces a copy (test with [ -L ] after creating it), the same way worktree-gates-spawn-budget skips without strace.
5. Add these five suites to the Windows lane in .github/workflows/test-windows.yml, scoped by path like its existing entries, so a regression is caught in CI.

Acceptance criteria

• On Windows Git Bash at a fresh main, all five suites exit 0, or print SKIP: only for a capability the host lacks (symlinks).
• test-windows.yml runs the five suites when their plugin paths change.
• Linux CI results for the five suites are unchanged.

[github-actions]

Thanks for the report. Automated triage handed this issue to a maintainer.

_{Automated triage. A maintainer reviews it.}

[kyle-sexton]

This was generated by AI during triage.

Triage

• Verified (partial reproduction): on Windows Git Bash at main 6690dad, the repo-sweep tick.test.sh suite reports 3 FAILED, and worktree-root-scan.test.sh fails [14] symlink is not proposed (got linked empty yes) and [24] default root scans the configured root (expected /tmp/tmp.X/..., got /c/Users/.../Temp/tmp.X/...), as reported. None of the five suites' scripts or tests changed between the report's base and origin/main 698960d.
• Redundancy: fix: write PR body edits before pushes and push once per review round #5899, cited as the first report, is a merged PR, not an open tracker item. No open item covers these suites.
• Note: perf(ci): run each test lane only on the suites the change selects #6129 changed how CI test lanes select suites after this issue was filed, so the CI criterion below is stated as behavior, not as entries in a specific workflow file.
• Routing: fully specified, no open decision; C3 (scoped, bug-fix). The intake lane's needs-human is removed because no human decision is pending; the claim-gate fence below stops the worker if the fix turns out to need one.

Agent Brief

Type: Bug
Summary: make the five Windows-failing shell suites (repo-sweep tick and state, source-control worktree-claim, worktree-facts, worktree-root-scan) pass on Windows Git Bash, and have the Windows CI lane run them.

Current behavior:
On Windows Git Bash the suites fail while Linux CI passes. Two causes are known: expected paths are built from the literal POSIX temp path mktemp returned, while the code under test reports the Git Bash drive form; and CRLF bodies are rewritten with LF (or a trailing carriage return breaks a byte-exact comparison). Two cases in the state suite (other-branch exit 14 and 15) differ in something the test compares but does not print, cause unknown. The worktree-facts claim round-trip reads an empty claim after a bare add. The symlink case fails where Git Bash copies instead of linking. The Windows CI lane does not run these suites.

Desired behavior:

• Path comparisons resolve the fixture root once to the same form the code under test emits, on both Linux and Windows.
• The repo-sweep scripts either preserve CRLF bodies byte for byte (when the GitHub web-UI tick path can deliver CRLF bodies) or the fixtures are built with the line ending the script emits; the choice is stated in the PR with its evidence.
• The hidden difference in the state suite's exit-14/15 cases is found and fixed at its cause.
• The worktree-facts claim written after a bare add is the claim read back on Windows.
• The symlink case prints a SKIP: line when the host cannot create a real symlink, the same way other suites skip on a missing capability.
• The Windows CI lane runs these five suites when files in their plugins change.

Key interfaces:

• The five test suites and, only where a test proves a script defect, the repo-sweep tick/state scripts and the worktree-facts library.
• The Windows CI test lane's suite selection.

Acceptance criteria:

• On Windows Git Bash at a fresh main, each of the five suites exits 0, printing SKIP: only for a capability the host lacks (symlinks).
• The five suites still pass on Linux, with the same assertion count as before.
• A change touching the playbooks or source-control plugin makes the Windows CI lane run the matching suites.
• No assertion is deleted or loosened to pass; each changed expectation is computed from the fixture, not hard-coded to one platform's form.

Out of scope:

• Any change to what the worktree-claim gate admits or refuses. If a failure turns out to be a real Windows behavior gap in the claim gate (not a fixture gap), stop and report it on this issue for a human instead of changing gate behavior.
• Other suites that already pass on Windows.
• Reworking the CI lane's selection mechanism beyond adding these suites.

[kyle-sexton]

This was generated by AI during work-loop.

Ratify autonomous dispatch of this C3 bug-fix item?

Classification: work-class: scoped (C3), bug-fix-shaped (confirmed at triage). Intended dispatch: the work-loop worker lane, through /work-items:work, to a draft PR the lane never merges. First drain is not complete and no item-level ratification exists, so it is queued here. Ratify through /work-items:attend-queue.

Lane: work-loop, instance melo-worker-1, cycle 2.