STALE: clean parallel merges can break a shared invariant

Primary source: Xia, Wu and Park, Passes Alone, Fails Together, arXiv v1 September 21, 2026; benchmark code. Evaluates a narrow semantic-coordination failure: two agent patches pass individually under the same combined test set and merge textually, but the combined patch causes a new test failure. The paper specifically corrected an earlier grading error that made differences in solo-versus-merged test sets look like interference.

Three populations, not one prevalence rate. (1) 834 agent runs on 417 validated, historically reviewed and merged Django PR pairs yielded only one detected newly failing combined test. Selection filters excluded textually conflicting or non-shared-base pairs; passing after review need not mean there was no earlier coordination work or that tests catch all defects. (2) Researchers deliberately constructed 36 interface-change tasks from 12 real Django helpers (scripted rename, required argument, changed return shape); GPT-5.5 blind runs caused new failures in 105/108, despite clean merges. (3) Tiny synthetic helper tasks offered a controlled mechanism. In constructed tasks, a complete oracle message describing the finalized change recovered 89/108 runs; in synthetic tasks a ~130-token oracle message cut mean failures 2.50→0.04. These are not field rates or proof that agents’ ongoing status summaries achieve the same recovery.

Orientation implication. A task board that tracks files claimed and a merge gate can miss a changed cross-cutting rule: e.g., after a new cache requires invalidating prices on every write, an independently added bulk-write path skips invalidation. Possible research object: a short versioned changed invariant/interface record attached to completed or in-progress agent work, plus combined behavioral tests, delivered to dependents before merge. This is a proposed response to a demonstrated mechanism, not a measured operating practice. Need pre-review patch snapshots, peer-visible message timestamps, integrations and human minutes to test it. Related coordination brief, planning, review.