The execNpm-routing redesign (previous commit) broke every real CI job:
check-env.cjs runs as its own standalone "Environment check" step BEFORE
`npm ci` / `npm run build:lib` -- a deliberate pre-flight, run before
there is even a node_modules to build with. Its require of
../gsd-core/bin/lib/shell-command-projection.cjs (a tsc-compiled artifact
that plain does not exist at that point in the pipeline) crashed with
MODULE_NOT_FOUND on every platform, immediately, confirmed via the real
CI log. My own local gsd-test run never caught this because it doesn't
replicate that exact pre-build step ordering.
Reverted the cross-module require entirely; check-env.cjs is back to a
self-contained spawnSync(npmCmd, ...) call, no requires reaching into
gsd-core/bin/lib. Kept the two things actually worth keeping from that
detour:
- the 15_000ms timeout (matches execNpm's own default elsewhere in this
repo -- not invented, an existing precedent -- vs. the original 10s
that failed twice under real Windows CI contention);
- computing `timedOut` via `error.code === 'ETIMEDOUT'` inline (the
same canonical, cross-platform-correct predicate that seam uses),
rather than the earlier signal === 'SIGTERM' check, which that seam's
own docstring documents as platform-fragile with a Windows-specific
false-negative risk.
Manually verified check-env.cjs runs correctly with gsd-core/bin/lib
temporarily removed entirely (simulating the real pre-npm-ci CI
ordering) before re-running gsd-test, since this repo blocks local
node --test.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Per /research + /diagnose direction: the timeout fix landed earlier this
session correctly diagnosed the failure (a real spawnSync timeout under
Windows CI contention, not npm being absent) but it recurred on the very
next push -- same chunk, same ~51-file load. Rather than raise the
hand-rolled 10s timeout myself (CLAUDE.md's own rule: fix the cost, not
the tolerance, and never touch a timeout without explicit instruction),
investigated the repo's own precedent first.
Found: this repo already has a canonical OS-shell-projection seam for
exactly this (src/shell-command-projection.cts's execNpm), already used
by dozens of other scripts/*.cjs files (require('../gsd-core/bin/lib/...')
is an extremely well-established pattern), with:
- the same npm.cmd/shell:true Windows handling check-env.cjs was
hand-rolling, but centralized;
- a 15s default timeout (vs. check-env.cjs's 10s) -- not invented here,
an EXISTING value already governing npm subprocess calls elsewhere;
- isSpawnTimeout / result.timedOut, the canonical cross-platform timeout
predicate (error.code === 'ETIMEDOUT'), whose own docstring explicitly
warns that checking signal === 'SIGTERM' (what my first fix did) is
"platform-fragile" with a Windows-specific false-negative risk -- the
exact platform this bug lives on.
check-env.cjs's npm-version check now calls execNpm(['--version']) instead
of hand-rolling spawnSync + npmCmd + shell:true, and
describeNpmVersionCheckFailure now operates on execNpm's SpawnResultOutput
shape (using timedOut, not signal) rather than a raw spawnSync result.
This is a genuine architectural fix, not just a bigger number: it removes
a duplicate, slightly-divergent re-implementation of an existing seam and
inherits whatever that seam's timeout/handling becomes in the future.
Manually verified end-to-end (npm run check:env against the real
environment) and re-verified describeNpmVersionCheckFailure's branches
directly against execNpm's actual return shape before wiring the test
file, since this repo blocks local node --test.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Discovered blocking this PR's own Windows CI (unrelated to this PR's
actual diff, fixed inline per this repo's no-defer policy): PR #4552's
"full test (windows-latest, 24, shard 2/3)" job failed
tests/check-env.test.cjs's npm-version subtest with "npm binary not
found on PATH" under chunk 5/9's heavy load (51 concurrent files,
5+ minutes). Root-caused via the CI log: the check's spawnSync call
used a 10s timeout, and every failure mode -- ENOENT, a signal-killed
timeout, a non-zero exit, a thrown spawn error -- collapsed into that
one message (only `res.status === 0 && res.stdout` was checked), so a
genuine npm.cmd cold-start timeout under contention was indistinguishable
from npm actually being absent.
Extracted the reason-selection into describeNpmVersionCheckFailure, a
pure function in the new scripts/lib/npm-version-check-diagnosis.cjs
(kept out of check-env.cjs itself, which runs its CLI unconditionally
on require with no `require.main === module` guard, so the pure logic
can be unit-tested without triggering a real environment check).
Reports ENOENT, signal-kill, non-zero-exit, and thrown-error cases
distinctly. Does NOT raise the 10s timeout itself -- a slow subprocess
under contention is a cost to reduce, not a tolerance to widen.
Also fixed a stale tsconfig.build.tsbuildinfo incremental-build cache
discovered while verifying this change: npm run build:lib was silently
omitting gsd-core/bin/lib/markdown-table.cjs (a real, needed compiled
module -- src/state-document.cts requires it), which only surfaced via
npm run lint:generated-sync's gen-health-docs check failing with
Cannot find module. Deleting the cache and rebuilding fresh restored
it; docs/INVENTORY-MANIFEST.json needed no net change once the build
was genuinely complete.
Manually verified describeNpmVersionCheckFailure's five branches
directly (ENOENT, signal-kill, non-zero exit, thrown error, defensive
default) before wiring the test file, since this repo blocks local
node --test. Re-ran node scripts/check-env.cjs directly to confirm the
real success path is unaffected.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
