The windows-test-parity ratchet greps test source for fs.rmSync-without-
maxRetries (and six other Windows-portability anti-patterns), failing when an
integer offender COUNT exceeds a frozen baseline (rmSync: 95). A count ratchet
is a Goodhart metric: fixing one offender and adding another keeps the count
constant, so a new defect slips through green. Replace it — and every other
count ratchet in the repo — with a layered, masking-proof design.
Behavioral seam test
- tests/helpers-cleanup.test.cjs proves helpers.cleanup() carries the Windows
EBUSY retry budget. cleanup() delegates retries to Node's fs.rmSync via
maxRetries (it owns no loop), so the test asserts the option contract
(recursive/force/maxRetries>0/retryDelay>0) + real-FS removal + the cwd-guard,
rather than a loop that does not exist. The EBUSY risk is now tested ONCE at
the helper, not approximated textually at every call site.
Write-time ESLint rule (AST-accurate, replaces the grep)
- eslint-rules/no-raw-rmsync-in-tests.cjs (error in tests/**/*.test.cjs) bans
raw fs.rmSync, steering to cleanup(). Catches member, computed (fs['rmSync']),
destructured and aliased forms; escape hatch is inline
`// eslint-disable-next-line local/no-raw-rmsync-in-tests -- <reason>` only.
- Migrated 336 raw fs.rmSync teardown calls across ~116 test files to cleanup().
~18 genuinely load-bearing sites (mid-test SUT/fault-injection removals,
error-swallowing or name-colliding local teardown helpers) keep the raw call
with an inline eslint-disable + reason.
Shared anti-ratchet primitive
- scripts/lib/allowlist-ratchet.cjs:
- assertWithinAllowlist: fails on NOVEL ids (new offender introduced) AND on
STALE ids (a known offender was fixed but not pruned) — identity, not count,
and a ratchet DOWN toward zero.
- assertTightCeiling: a size/length budget whose ceiling must stay within a
grace band of the high-water mark, so budgets may only tighten, never creep.
Ratchets converted onto the primitive
- windows-test-parity-guard.test.cjs: rmSync rule deleted (now ESLint-enforced);
the remaining six patterns moved from integer baselines to named-set
allowlists with ratchet-down.
- scripts/lint-test-file-count.{cjs,allowlist.json}: per-module integer counts →
named filename sets (closes the swap-a-file-keep-the-count blind spot); a
module dropping under cap now FAILS to force pruning its allowlist entry.
- enh-2790 skill-count `<= 63` → named skill allowlist (ratchets toward ~58).
Size budgets hardened (tighten-only)
- agent-size / workflow-size / feat-3039 help-tiered: ceilings lowered to the
current high-water mark and an assertTightCeiling anti-creep check added per
tier. Fixed external-contract limits (description ≤100 chars, agent ≤100 KB)
are intentionally left as-is — they are not grandfathered creeping budgets.
No user-facing behavior change (tests + tooling only); no USER_FACING_PREFIXES
touched, so no changeset fragment is required.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* test(installer): add regression tests for #3670 migration lock self-deadlock
- T1: same-process PID re-entry reclamation (primary regression)
- T2: dead-PID stale lock reclamation
- T3: unlinkSync EPERM surfaces (not silently swallowed via force:true)
- T4: counter-test — normal round-trip still works
- T5: counter-test — genuinely-held live lock still errors clearly
- Update existing 'reports lock release failures' test to mock
fs.unlinkSync (not fs.rmSync) matching the fixed release path
Windows wall-clock deadlock repro depends on Docker matrix Windows runners.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(installer): break migration lock self-deadlock on Windows (#3670)
Root cause: `acquireInstallMigrationLock` release closure called
`fs.rmSync(lockPath, { force: true })`. On Windows NTFS, a file
recently closed via `closeSync(fd)` may still return EPERM from
`unlink` until the OS fully releases the handle. The `force: true`
flag silently swallows EPERM, leaving the lock file on disk. The
subsequent `runInstallerMigrations` call in the same install()
invocation hits EEXIST, spins for 30 s, then throws
"installer migration lock is held".
Fix:
1. Release closure uses `fs.unlinkSync` (not rmSync+force) so
EPERM propagates via releaseError instead of being swallowed.
2. `acquireInstallMigrationLock` closes the fd before writing the
payload (path-based write), eliminating the open handle that
caused the deferred EPERM on Windows.
3. Stale-lock reclamation: on EEXIST, parse the on-disk PID and
reclaim immediately if it matches process.pid (same-process
re-entry, the primary #3670 failure mode) or if the PID is
dead (ESRCH). Live alien PIDs still trigger the 30 s timeout.
4. Error message on a genuinely-held lock now includes the holder
PID and acquiredAt timestamp for operator diagnostics.
No public API change. All callers of runInstallerMigrations are
inside installer-migrations.cjs and bin/install.js.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* chore(3670): update changeset to reference PR #3765
* fix(3670): timeout on reclaim-unlink failure to prevent spin-loop regression
When unlinkSync throws (e.g. Windows EPERM on an open handle) in the
same-PID / dead-PID reclamation path, the original code continued
unconditionally — bypassing the timeout check and reintroducing the
exact deadlock the PR is supposed to fix.
Guard the continue behind a `reclaimed` flag: only loop back to
openSync if unlink SUCCEEDED. On failure, fall through to the existing
bounded sleep + timeout, which surfaces "installer migration lock is
held" within lockTimeoutMs instead of spinning indefinitely.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(3670): tighten T5 — assert bounded failure when reclaim unlink fails on live lock
The old T5 accepted BOTH success and throw, which allowed over-reclamation
of a genuinely un-reclaimable lock to pass undetected.
Rewrite T5 to force deterministically unreclaimable conditions:
- Pre-seed lock with process.pid (triggers isSameProcess path)
- Mock fs.unlinkSync via mock.method() to throw EPERM for the lock file
With the production fix: reclaimed=false → falls through to timeout →
throws "installer migration lock is held" within ~200ms.
Without the production fix: unlink throws but continue runs anyway →
process spins and eventually OOMs (confirmed RED: 136s runtime, V8 heap
exhaustion from infinite readLockFile + new Error() allocations).
assert.throws() now makes success a hard failure, closing the
over-reclamation gap.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(3670): clean up orphan lock file when writeFileSync fails after closeSync
If closeSync(fd) succeeds (fd=null) but the subsequent writeFileSync
throws, the empty lock file was left on disk. readLockFile returns null
for an empty/invalid-JSON file, so the stale-lock reclamation path
skips it, causing the next acquire attempt to spin to timeout.
Track ownership with lockCreatedByUs flag; add a second cleanup branch
in the catch block for the fd-already-closed case.
Also fix changeset body to use the bold-prefix format required by all
other fragments in .changeset/.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
