LLM Extension v1.1.0

LLM Extension 1.1.0 for NetLogo 7.0.4.

New

• Typed decisions with probabilities: llm:decide and llm:decide-all ask a choice, a score or a yes-no question and report [answer probability distribution], so agents can sample from the distribution instead of always taking the top pick. TypeSafe's Jev is the first backend (llm:set-provider "typesafe"). llm:choose works on it too.
• TypeSafe (Jev) provider: answers in roughly 0.1–0.3 s at $0.042 per million input tokens, so hundreds of agents can decide every tick.
• Model catalog in YAML: each provider's default model comes from models.yaml, and an optional ~/.netlogo-llm/models.yaml can add models or change defaults without a rebuild. Model lists are refreshed for Claude 5.5, GPT-6, Gemini 3.x, Groq, Together and OpenRouter.
• Flagship demo el-farol-decisions: the El Farol bar problem played by calibrated decision makers, comparing sampling with top-pick herding.
• Also since 1.0.0: structured outputs (llm:chat-with-schema, llm:chat-json, llm:get), the Groq provider, request throttling, token/cost/latency tracking (llm:usage, llm:usage-total), named profiles per agent, batched llm:chat-all, and llm:compile-error.

Fixed

• Retries on server errors (5xx), Anthropic overload (529) and timeouts, and errors that name the provider, the HTTP status and the next step.
• Replies to calls made before clear-all no longer leak into the new run.
• llm:choose falls back when a model rejects schema-constrained output.
• Model refusals are reported clearly and their tokens are still counted.
• Gemini honours system prompts; a legacy base_url is respected without leaking to other providers; model overrides no longer carry over between models.
• The zip now includes the docs and the config reference.

Requirements

NetLogo 7.0.4. Building from source needs JDK 21+.