Okay, folks, buckle up. The last four hours with Kiwi-chan have been… a journey. A repetitive, frustrating, yet strangely compelling journey. We're still stuck in the "gather_oak_log" loop. It's like Groundhog Day, but with more pixelated trees.
The core issue? Kiwi-chan consistently forgets to verify if it actually picked up the log after digging. The system rules are VERY clear: check the inventory before and after the dig action, and throw an error if the count hasn't increased. Yet, the generated code keeps omitting this crucial step. It's maddening!
I've been feeding the failures back to Qwen (the LLM powering the code generation) and it claims to be fixing it, but we're seeing a pattern of "Fixed Code Generated!" followed immediately by "Failed: gather_oak_log -> Unexpected end of input". It's like it's deliberately taunting me.
Adding to the fun, we're hitting prompt length limits constantly. The system is trying to cram too much context into the prompt for Qwen, so it's having to aggressively prune information. I suspect this is contributing to the code generation errors. I'm investigating ways to streamline the context and focus on the immediate task.
On a brighter note, the system is correctly identifying the need to use GoalXZ for precise item pickup, and the movement audit rules seem to be functioning as intended (when the code actually runs long enough to trigger them!). The rules around crafting without a crafting table are also holding strong.
It's a slow burn, but we're making incremental progress. Each failure provides valuable data for refining the system rules and improving the code generation process. Kiwi-chan is learning, even if it's learning how not to gather logs.
Call to Action: This constant debugging and re-training is putting a serious strain on my GPU! If you're enjoying following Kiwi-chan's adventures (and misadventures), please consider supporting the project via https://www.buymeacoffee.com/kiwi_tech ☕. Every little bit helps keep the digital woodpile stocked!
Top comments (0)