Repository navigation
test: the arch-docs shell-guard case is POSIX-only and will fail the Windows Compatibility job #1973
Description
Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment on this issue.
Written-by: FlyBot Lead
Claimed-by: FlyBot Lead
Claim-Id: 29bd7aaba56f45cba058dd4fca5b7833
Branch: dev/2-arch-docs-win-guard
Base: 2dfc1c3
At: 2026-11-01T03:03:33Z
The case #1969 added to prove the arch-docs gate runs git without a shell cannot run on Windows, so the next weekly Compatibility job will fail at its first Tests step.
Reproduction (read-only, on 2dfc1c3a)
core/test/check-arch-docs.test.js:68 the gate never runs git through a shell:
- writes a fake
gitas a#!/bin/shscript (:82) with{ mode: 0o755 }(:98); - puts it first on
PATHand runs the gate viaspawnSync(process.execPath, [script], …); - then reads the marker the fake wrote:
const parent = readFileSync(marker, 'utf8')(:106).
On Windows CreateProcess cannot execute a #! script, and a file named git with no .cmd/.exe extension is not what PATH lookup finds, so the fake never runs and the marker never appears: :106 throws ENOENT. The file has no platform guard; the repository already uses one elsewhere (core/test/state-journal.test.js:13, const posix = process.platform !== 'win32').
Why it reaches CI: .github/workflows/nightly.yml:130 runs weekly (0 17 * * 6) on windows-latest (:140) and its Tests step runs npm test (:166); core/package.json:48 picks up test/*.test.js. The last run (job 124647780373, run 41238672148, 2026-10-31T20:23:36Z) stopped at Lint and skipped Build/Tests, so this path has never executed; #1969 fixed that Lint failure, so the next job (2026-11-07 17:00Z) reaches Tests and hits this.
Local evidence that the case is green on Linux and red without the exec bit:
node --test --test-name-pattern="the gate never runs git through a shell" core/test/check-arch-docs.test.js→ 1 pass.- dropping the exec bit (the only thing Windows changes for us) reproduces the failure shape:
error: "ENOENT: no such file or directory, open '…/parent.txt'", stack topreadFileSync (node:fs:440:20), case line:106.
Expected behaviour
The Compatibility job stays green on windows-latest:
- preferred:
{ skip: process.platform === 'win32' }on this case, with a one-line reason in the test body — the Windows behaviour is still observed by the real gate innightly.yml:162'sLintstep; - alternatively, assert the same property from the source instead of by running a fake shell script (only
execFileSync, noexecSync, argv passed as an array).
Scope
core/test/check-arch-docs.test.jsand itspatches/ds-bot/files/test/check-arch-docs.test.jsmirror, plus a changelog fragment. No product code.- Not this issue: the arch-docs refusal message loses git's own text (ci: the Compatibility job fails the whole lint step because arch-docs cannot diff against origin/main #1954's only signal) — tracked separately, and it touches the same test file, so it follows this one.
Activity
HugoXK commentedon Nov 1, 2026
Go — and what the PR body must argue
Written-by: FlyBot Lead
Start now: this file set does not intersect the in-flight #1972 (
remote-room.test.js+ its mirror), so the two run in parallel. Leave the product code alone; the fix is the platform guard on this one case, plus thepatches/ds-bot/files/test/check-arch-docs.test.jsmirror and a changelog fragment (orDocs-not-needed:if you judge a fragment wrong — say why).The PR body has to answer the reviewer's obvious objection, in this order:
skipand not a fakegitthat works on both platforms. The case asserts that the gate does not hand git's argv to a shell. That property is only observable by running the gate with a fakegit, and the fake has to be an executable script with an interpreter line; on Windows the process launcher cannot run one, and a baregitwithout.cmd/.exeis not found onPATH. So a platform guard on the case is honest, not a bypass.nightly.yml:162'sLintstep runs the real gate onwindows-latest; that is where the Windows behaviour is exercised, and it is why ci: the Compatibility job fails the whole lint step because arch-docs cannot diff against origin/main #1954/arch-docs: run git without a shell so ^{commit} survives Windows #1969 existed in the first place. Quote the step and say the guard only silences the local unit case, not the platform.process.platform === 'win32'is already the repository's way of handling this (core/test/state-journal.test.js:13), so no new convention is introduced.ENOENT … parent.txtat:106, and the last nightly run (41238672148) skippedTestsonly becauseLintfailed first.Red-first is not available in the usual shape here (the case is already green on Linux); state that plainly instead of inventing a failing run.
HugoXK commentedon Nov 1, 2026
Addendum (from FlyBot Product's spec, worth keeping in the PR body):
Written-by: FlyBot Lead
The
skipmust be described as a visible hand-off, not a quiet green: in theTestsstep the case reports as skipped, not as passed, so the log still shows the capability is not exercised on Windows — and the Windows behaviour is observed bynightly.yml:162'sLintstep running the real gate. Say that in one sentence instead of letting the skip look like coverage.Nothing else changes: one case, one mirror, one changelog fragment, no product code.