Spaces:
Sleeping
Sleeping
Commit History
-
Multi-model support: per-game model picker + live benchmark (#3)
schipsyUnverified signatureProvider abstraction behind call_structured so each game runs on any of five registered LLMs (Claude Opus 4.6/4.7, Claude Sonnet 4.6, GPT-5.4 via Azure OpenAI Responses, Kimi-K2.5 via Azure AI Inference). Chosen model is pinned on the Game and echoed in CLI, API, UI, and debrief header. Live-benchmarked every model against a fixed one-turn competitor_disruption run; two reports in benchmarks/. Default switched to gpt-5.4 — 45% faster than Opus 4.6 at par quality, the only model that meaningfully exercises server-side prompt caching (~63K re-used tokens per run). Added 429 exponential backoff (10s/30s/60s/120s) to the AI Inference provider after Kimi repeatedly hit its 8-concurrent capacity cap. Two rounds of review feedback addressed in-place. Key fixes: Anthropic cache_control moved from top-level kwarg (silently ignored) to the system content block — the benchmark's cache_read=0 across every Claude model was the smoking gun. Orchestrator now wraps run_orchestration in use_game_mode…b76b644