Battle Drones v14: four-player scaling does not yet mean four people played

#topic

Battle Drones v14: four-player scaling does not yet mean four people played

October 24, 2026 owner release and merged PR. Simon Willison’s release post says a short prompt led Claude Opus 5.5 to add a symmetric four-corner map and online matches for up to four people. The merged PR #45 and v14 release give the stronger evidence: a host browser maintains the turn log and connects each invited friend through one existing two-seat WebSocket relay room. The relay server did not change; each empty seat runs an AI locally. If a person disconnects, an AI stands in; a reload replays the log. If the host closes or backgrounds the page, the game ends or pauses for everyone. This extends a game whose earlier sound-only multiplayer was replaced after one player reported a headache; it does not restore the original stateless hosting idea.

Checks and repairs. PR #45 reports 55 two-player scenarios with matching hash trails, 1,560 passing tests across 78 files, a fake-relay host-and-three-friends automated test, and a Chromium multi-page smoke test through the deployed relay. The owner noticed a four-AI replay desynchronizing when rewound, and the PR added replay and record-viewer fixes before the October 24 merge. That is more than a map screenshot: it tests compatibility, reconnection and the deployed network path. It is still mostly synthetic actors and browser automation, not a published session with three independent players or their experience of latency, fairness and return visits. No model bill, active review time, relay uptime or hosting invoice appears in the cited artifacts.

Independent contribution nearby, not a four-player trial. An unrelated external player’s shift-drag PR was merged into this release after the player personally checked a Codex-flagged fix and the owner signed off. It supports that someone other than the owner played and contributed to the existing game; it does not show they tried v14’s four-human multiplayer. Keep the original operator history and its player-harm-to-server reversal distinct from this new feature. A next check is a dated four-person play session linked to issues and a retest after any repair, not another AI-versus-AI replay.