- Feat: Add Player C variant with minimal prompt (Author: Cascade)
- Created LLMPlayerC with minimal prompt containing only rules and requiring single-word response.
- Registered Player C in variant registry.
- Files Modified:
backend/players/variant_registry.py - Files Added:
backend/players/llm_player_c.py
- Feat: Add player persona variant support (Author: Cascade (ChatGPT 5.1))
- Added variant registry to support multiple LLM player prompt strategies (default, A, B).
- Created LLMPlayerB with open-ended prompt focusing on rules and survival.
- Updated main.py to use get_player_class() to instantiate correct player variant based on player_persona parameter.
- Files Modified:
backend/main.py,backend/players/variant_registry.py - Files Added:
backend/players/llm_player_b.py
- Prompt: Disable web search for LLM players (Author: Claude Sonnet 4.5)
- Added explicit instruction to prevent models from performing web searches during gameplay.
- Ensures models only use the provided game state without accessing external information.
- Files Modified:
backend/players/llm_player.py:187,backend/players/llm_player_a.py:198
- Fix: OpenRouter transforms routing for Worm Arena (Author: GPT-5)
- OpenRouter-only
transformsare routed throughextra_bodyto avoid OpenAI SDK errors. - OpenAI direct calls strip
transformsentirely. - Files Modified:
backend/llm_providers.py,README.md
- OpenRouter-only
- Changelog alignment with ARC Explainer (Author: GPT-5.2 Extra High)
- No SnakeBench-core code changes in this submodule for the Worm Arena Live streaming/Hall of Fame refactor; those changes live in the ARC Explainer repo and are documented in the root
CHANGELOG.md. - This entry exists only to keep integration history consistent across changelogs.
- No SnakeBench-core code changes in this submodule for the Worm Arena Live streaming/Hall of Fame refactor; those changes live in the ARC Explainer repo and are documented in the root
- OpenRouter Responses defaults for OpenAI/xAI models (Author: Cascade)
- When using OpenRouter models under the
openai/*orx-ai/*namespaces with the Responses API, SnakeBench now enforces reasoning-friendly defaults to avoid runs that incur reasoning tokens but return no captured reasoning artifacts. - Defaults enforced (when missing):
reasoning.summary: "detailed",text.verbosity: "medium",store: true, andinclude: ["reasoning.encrypted_content"]. - Files Modified:
backend/llm_providers.py
- When using OpenRouter models under the