1. 01Three agents at one card tablePart 1: how I model the Twenty-eight players as seat-specific agents, and why the LLM does not make every card-play decision.
  2. 02The repeated prompt at the card tablePart 2: what KV cache and prefix caching can reuse in Thuruppu's repeated agent prompts, and what the repository does not measure.
  3. 03Two serving stacks, one card gamePart 3: the practical differences between running Thuruppu's local model endpoint with SGLang or vLLM.
  4. 04Tracing the game, scraping the modelPart 4: how Thuruppu emits OpenTelemetry traces, how the Collector sends them to Jaeger, and how Prometheus reads vLLM metrics.