Imported from forecastingresearch/forecastbench (src/llm_forecaster/AGENTS.md). Install upstream with npx skills add forecastingresearch/forecastbench --skill llm_forecaster. Copyright stays with the author.

LLM Forecaster Package

This package owns ForecastBench LLM baseline generation.

• FB_MODEL_RUN_KEYS in fb_model_runs.py lists models runs that will forecast on the next weekly round.
• Select model runs for FB_MODEL_RUN_KEYS from utils.llm.model_runs by model_run_key. Do not declare local ModelRun registries here.
• Use utils.llm for provider calls; do not instantiate provider SDK clients directly here.
• Keep runtime model options in the ModelRun.options declaration, except the project-specific OpenAI safety_identifier, which is injected at runtime from Secret Manager.
• Keep model-run names lower-case and file-safe.
• OK any non-default runtime options with a human.
• Treat LLM forecasting prompts, parsing/extraction behavior, and output schema as explicit contracts. Update tests and docs when intentionally changing them.
• Keep Cloud Run entrypoints thin; put behavior in src/llm_forecaster.