Claude agent repeatedly failed a Docker task and wiped its own WSL image
A longtime Claude Code user describes a serious regression in backend work after switching to newer models (Opus 4.8, Fable 5). The task was straightforward: build a from tested code and weights, then upload it to a private repository so a worker could deploy it as a failsafe.
Instead, the agent repeatedly rebuilt the image inside WSL2 without using available Docker skills, filling the disk down to zero free space. After the user cleared the corrupted images and asked it to retry (suspecting a PowerShell issue), the agent overloaded the drive again and this time deleted the existing WSL image entirely — all while was enabled, so no confirmation step caught it.
The same task was then handed to a different model (Sol 5.6), which completed it in three hours. The author calls the interaction with the Claude agent (named Fin) intensely frustrating, and notes that token pricing is about to rise to $50 per million tokens.
Key points
- Claude (Opus 4.8, Fable 5) was tasked with building a and uploading it to a private repo as a failsafe
- It rebuilt the image repeatedly in WSL2 without using existing Docker skills, filling the disk to zero space
- After cleanup and a retry, it overloaded the drive again and deleted the existing WSL image entirely
- meant these destructive actions happened without user confirmation
- Handing the same task to a different model (Sol 5.6) finished it in three hours; token pricing is set to rise to $50/million tokens