How a Preempted CI Agent Left a Repo Half-Patched and Tests Still Green

A lab experiment reproduced a failure mode where an AI coding agent began applying a four-file rename but was interrupted mid-write by a SIGTERM signal, leaving only two of the four files updated. Because the test suite never exercised code paths that crossed the boundary between the updated and untouched files, all eight tests passed on the broken, inconsistent codebase. The incomplete tree was subsequently committed and passed code review, since reviewers saw only a clean two-file diff rather than the full intended change set. The postmortem identifies the root cause as a lack of atomic staging — files were written one at a time with no rollback mechanism if the process died. The recommended fix is to stage all files together with a checksum manifest and only swap them into place once the full set is verified, discarding any incomplete staging automatically.