Repository navigation
A run refused by the lock it did not get could leave a lock of its own behind — one naming a
live pid, so the refusal outlived the run that caused it. The suite that should have noticed
was itself leaving scratch directories in the host temp directory: five days of runs had left 616
of them and about 600 MB. Both are fixed, and the second is why the first stayed hidden.
A refused run kept the lock it took
generate locks the staging root and the output directory, in sorted order, before staging is
touched. The acquisitions came from one map(), so when the second lock was refused the first was
still held — and there was no list of releases to run in the finally, because the throw happened
before the assignment. The file left behind named the current process, so reclaimStaleLock read
it as an active run. An MCP server, watch mode, or a setup that retries would be refused for the
rest of that process's life, with no other run in existence.
Each acquisition is now recorded, and a failure releases the locks already taken. The regression
test starts a run that a fresh staging lock must refuse and asserts the output-directory lock is
gone afterwards.
The test suite stops leaving its scratch directories behind
npm test created every fixture in the host temp directory and relied on each test to remove it.
Two tests never did: agent-simulation.test.js made two fixtures per run with no cleanup, and one
repo() argument in generation-truth.test.js was never assigned, so its directory had no handle
to remove. A wrapper now runs the suite with TEMP, TMP and TMPDIR pointed at one scratch root
that is deleted when the run ends, pass or fail, and reports by name anything a test leaves inside
it. Tests create scratch trees with makeTempDir(t, prefix), which registers the removal with the
test, and a hygiene suite fails the build if a test creates a directory it never removes, if the
npm scripts stop using the wrapper, or if the fixture-heavy generation suite leaves anything behind.
Verified
npm run verify — lint, typecheck, 1032 tests (0 failures, 7 skipped), the docs sync check, the
site build, the site audit, the readme lint and the tarball check — on the release commit. The same
gate runs on every pull request and every push to main, and the release runs it on Node 18, 20
and 22; Node 18 caught a portability defect in this release's own new test before it shipped.