forked from longtao.wu/droboticflow
executeProjectDeletion processed every target (runtime, then workspace, then bos, then records) strictly one at a time — for a project with thousands of real runs, that meant thousands of sequential K8s/BOS/Mongo round trips. Observed live on Fanyu: BLF-data's runtime phase (stopping 4,641 runs) took tens of minutes at that pace. Replaced the per-target sequential loop with a bounded worker pool (PROJECT_DELETION_TARGET_CONCURRENCY, default 12) applied within each phase — targets of the same kind now run concurrently, but phases still run strictly in order (runtime, then workspace, then bos, then records), preserving the "stop/verify runtime, then remove files, then DB" ordering the plan is built with. The one-time phase-transition safety checks (wait for in-flight human work, re-verify runtime absence, revalidate references) still run exactly once, at the same point, between the runtime and workspace phases. Every read-modify-write of shared task state (a target's Status/Error, and every persist() call) is guarded by a mutex; the actual I/O (deleteProjectTarget) runs unlocked so workers genuinely overlap. Added TestProjectDeletionExecutesTargetsConcurrentlyAndCompletes (30 runs) to lock in that concurrent execution reaches the same end state — every target done, every workspace directory removed, project record gone — as the old sequential loop. Full suite passes under -race with no warnings; 75 pre-existing unrelated failures unchanged. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>