Wrap JS-only model output (e.g. GLM's javascript fences) in HTML shell with canvas matching getElementById IDs

Disable thinking mode via chat_template_kwargs.enable_thinking=False so GLM emits game to content channel

Fix GLM Idle status: read all delta fields via model_dump; analysis status now reflects code OR reasoning OR connection activity; add Content/Reasoning byte columns

Don't block on /v1/models catalog: it's partial on Crusoe; models can be callable while unlisted

Pre-flight model catalog check + Connecting/Streaming/Queued visual states; auto-fail unknown model IDs immediately

Capture delta.reasoning_content as fallback (fixes empty GLM stream); add request timeout + watchdog; show both streams in failure card

Extractor: strip <think> blocks, accept style+script fragments, pick largest fence; failure card now shows raw output preview

Add comparative analysis panel + tolerant HTML extraction; swap Nemotron for Qwen3-235B in default lineup

Make games autoplay: each model bakes in a heuristic bot

Center main container horizontally

Pin gradio>=5.29.0

Bump Gradio to 5.29.0 (HfFolder import fix)

Restore requirements.txt + add audioop-lts fallback

Pin Python 3.12 (fix pydub/audioop on 3.13)

Pin Python 3.12 + audioop-lts fallback (fix pydub on 3.13)

Initial: 4-LLM arcade game race UI

initial commit