env_client: retry transient 5xx + typed EnvUnavailableError; trainer survives env outages

action_parser: strict variant strips Qwen3 <think>...</think> before json.loads

publish: auto push adapter (and optional merged) to HF Hub after training

Switch base model to Qwen/Qwen3-8B (Qwen3.5-9B is multimodal Qwen3_5ForConditionalGeneration, unsupported by unsloth)

Fix training pipeline: TRL>=0.25 rollout, real generation, curriculum/W&B callbacks

Add episode replay dashboard and GRPO training pipeline

Update Red/Blue showdown behavior and refresh Qwen benchmark artifacts.

refactor: enhance type safety in inference and evaluation scripts; update pyright config to exclude specific directories

refactor: enhance type safety in inference and evaluation scripts; update pyright config to exclude specific directories

feat: improve blue showdown runtime context

feat: add red reasoning metadata

feat: harden environment integration

feat: add dense red reward scoring

feat: add blue defender curriculum

feat: add raw bash red agent access

feat: port WarGames phase 0 environment