3 engineers in packages/shared spent 14 days editing the same files. None of them reviewed each other's work. PR #35918 merged into the same area PR #35985 was still open.

When engineers edit the same files without reviewing each other, merge conflicts are the best-case outcome. The worse case is silent semantic conflicts that pass CI and break production. This is not a tooling problem. It is a visibility problem.

hoxyq, eps1lon, and jackpope all modified packages/shared within the same 14-day window. Zero cross-reviews between them.

42 detection windows flagged G4_CoordinationGap on packages/shared. PR #35918 by hoxyq merged while PR #35985 by eps1lon was still open in the same directory. PR #36026 by jackpope landed 3 days later touching overlapping files.

hoxyq authored 56 PRs that waited for first review. The bottleneck is not the reviewer. It is the absence of one.

56 detection windows flagged G1_ReviewBottleneck on hoxyq as author. Reviews piled up without a first reviewer assigned. The median wait exceeded the team baseline.

react-devtools-shared had 18 review bottleneck windows. A subsystem with its own velocity, decoupled from the main review cadence.

18 detection windows flagged G1_ReviewBottleneck on packages/react-devtools-shared. This package operates on a different review cycle than the core React packages, creating a consistent pattern of stalled reviews.

The merge rate shifted 60 times during the analysis window. Each shift is a signal the team did not act on.

60 detection windows flagged G3_VelocityChange across the repository. The highest-gravity velocity drop reached 0.425, indicating significant throughput disruption that went unaddressed.

This pattern was detectable within 24 hours of PR #35985 opening in the same directory as #35918. A notification at that point would have given the team 13 days to coordinate before PR #36026 landed.

Analysis of 328 situation snapshots from facebook/react. ARC detection types: G1_ReviewBottleneck, G3_VelocityChange, G4_CoordinationGap. Detection window: 14-day rolling lookback.

The patterns above were detectable within 24 hours of the first signal. Your AI agent opened files in the same area. It did not know.

Zamski's MCP server gives every agent on the team coordination context before it writes code. One setup.

Want to see this for your repos?