I've tried to reproduce it with Roo Code and Qwen3.6-27B-UD-Q5_K_XL as well as Qwen3.6-35B-A3B-UD-Q5_K_XL, 80k context, -ctv q8_0. Single-shot, not automated testing via browser. Got semi-working results.

27B had a bunch of missing function parameters and exports. The A3B run just had a single missing parameter, likely because I added an instruction to the final prompt to check the whole generated code vs. the original plan once done, and it fixed a few things then.

27B version was looking good and playing nice, but you got pushed through a forest once you entered it, and enemies apparently respawned into a stronger version once killed.

A3B lacked the nice item overlay and attack animation. You couldn't walk through forests, but on water.

The token stats are interesting. These are just generated tokens during the agentic run. Any different in input tokens / file re-reads or web-search used? In my experience Qwen 3.6 is rather verbose, while Opus 4.7 was tuned to be more concise. Yet still Opus used more tokens. (extra)high reasoning maybe?

Any specific quant used for Qwen? And what context length was available (and used)?

Also: Single-line "build a cozy roguelike" prompt, or maybe a more sophisticated, half a page description? The models infer a lot that's not given via prompt, just because the genre is familiar. For example when I asked Qwen to make a web-based multiplayer pong game, it automatically made it so that the ball speeds up over time, without me ever mentioning it, as that's part of such kind of game.