neatroots is a tool for recording AI agent executions that runs locally. It lets you inspect the runs an agent has produced and replay them as they are, and its latest commit 4f0b3ad (November 2, 2026, #675) addressed both the query performance of its internal quiet-run sweep and the ordering of startup recovery. The quiet-run sweep is the job of periodically finding runs that were never sent a termination signal, marking them as interrupted, and re-verifying them.
The commit changes two things together. First, the query of the sweep that runs every 60 seconds used to scan the entire runs table, and it was replaced by a (status, doctor, complete, updated_ms) index. The commit message states that in a store with 100,000 accumulated runs it dropped from 7.1 ms to 0.52 ms. Second, the sweep now claims the runs it is to process with a single fast DB pass and hands re-verification off to the background, so that startup recovery (resumeUnfinished) no longer waits for one model re-verification per old run. The claim cap is applied inside the query via SQL LIMIT, so the approach of reading all rows and then trimming them is gone.
However, there is a cap on the number of runs a single sweep claims synchronously, so a backlog that accumulated during a long stall is handled by the next sweep instead. If re-verification fails, only a warning is left behind and the rest of the same batch continues, but a run whose recovery also fails remains interrupted until the next sweep reclaims it as an abandoned claim. Completed sweeps are recorded in the quiet_sweep_done column newly added to the runs table. The performance figures are the values written in the commit message, and no license or star count was given. The test file quiet-run.test.ts grew by 467 lines and also covers failure paths, but whether the full regression suite passes cannot be known from the commit information alone.