The catalogue at marisstella.app is built by one script, platform/pages/build.mjs. Its first step deletes the output directory, and over the next few minutes it writes about 10,000 pages back. Nothing stopped a second copy of the script from starting on the same directory partway through.
Two agent runs did that on the same day (run records under prompts/runs/, 2026-11-15). In the first, a build started with a trailing & inside a background shell kept running after its shell had returned. The run started a second build, which died on ENOTEMPTY while deleting a directory the first was still filling. In the second, a run checked pages in dist/ while another build was rewriting them, and saw a page with its keyboard tab stops missing. That defect did not exist, but the run had to spend time proving it.
The change
build.mjs now writes a lock file, dist.building, next to the output directory before it touches anything. The file holds the build's process id. A second build into the same directory reads the file, sees that the process is still alive, and exits with code 75 and a message:
build.mjs: another build (pid 17115) is writing .../platform/pages/dist; wait for it
to finish (it removes .../dist.building when done), or build elsewhere with --out=<dir>.
The file lives beside dist/, not inside it, so the build's own delete step leaves it in place. If a build is killed, it leaves a lock whose process is gone, and the next build takes that lock over instead of obeying it. A build into a different --out directory takes a different lock and is not blocked. The lock only applies to the command-line entry point. The test suite calls buildSite() directly into temporary directories and is unchanged.
The file also answers the second run's question: while dist.building exists, dist/ is not finished, so a probe should wait.
Checked
node --test platform/pages/build-lock.test.mjs: 5 pass (refusal while alive, a stale lock taken over, release that leaves a newer build's lock alone, separate directories independent).- A real run of two builds in a row: the second was refused with exit 75 within a few seconds, the first finished with exit 0, and the lock was gone afterwards.