v0.5.6 — Calmer Skies
When a village has five or more agents all thinking at once, the LLM server was getting overwhelmed with requests. Agents would retry aggressively, creating a feedback loop of rejected requests. This release adds a global decision queue with exponential backoff and a concurrency cap, so agents take turns thinking instead of shouting over each other.
The server-side rate limit was also raised from 20 to 120 requests per minute to match the actual demand of active villages. Together, these changes reduce error noise by over 90% and keep agent cognition flowing smoothly even in busy settlements.