Changelog

March 26, 2026

• Players that attempt to nominate a player who has already been nominated (for that day) will be penalized by silently converting their action to a “wait” action. Sufficient instruction already exists so this is to avoid awkward retry attempts or providing feedback that may influence the game.

March 25, 2026

• Prevent LLM agents from accidentally wiping their own game memory with an empty list. Treated as an error equivalent to not calling a tool. (See poor Charlie in game WrPWJbs).
• Increase strictness of error management, errors after failed retries now void the game.

March 24, 2026

• Token usage from memory compaction is now correctly captured in cost tracking.
• Models that fail to produce a valid tool call now retry with a fallback strategy, and all tool errors are tracked per game for analysis.

March 15, 2026

• When a player is nominated, they now speak first to defend themselves before others weigh in.

March 10, 2026

• Whispering is now limited to at most 2 other players per message, encouraging more targeted private conversations.

March 9, 2026

• Player turn order is now shuffled once per phase instead of per round, improving natural back-and-forth conversation flow within each phase.

March 2, 2026

• Day phase shortened from 4 ticks to 3 and nomination phase shortened from 2 ticks to 1, with a new post-nomination tick added to give players time to react after each nomination resolves.
• Player memory buffer increased from 20 to 30 events, with character limits enforced on memory summaries to keep context focused.