Cache-billing leak fixed, and everything from robots to diffusers gets harder to break

The month's most consequential change landed in tau, where Anthropic prompts were being re-billed on every turn of a conversation instead of being cached [1]. "Cache Anthropic prompts instead of re-billing every turn" [2] fixes a billing leak that inflated API costs for subscription users, and it pairs with a separate fix for Anthropic subscription login and OAuth token refresh [3]. If you run tau against Anthropic billing, upgrade and verify your prompts now hit the cache.

Beyond that, the through-line is reliability engineering across the tools. lerobot stopped simulating environments whose episode had already ended, halving eval overhead [4]; the D405 RealSense camera now auto-recovers from connection timeouts at startup instead of hanging [5]; and LeKiwi bus reads retry on transient Feetech errors instead of dying [6]. On the training side, TRL now uses the accelerator device instead of hardcoded CUDA in its FSDP2 vLLM weight sync [7], and it requires `peft>=0.20.0` for reference adapters with `target_parameters` [8]. Each fix is narrow, but together they close the classes of bugs that silently corrupt training runs.

Transformers spent the month hardening its codebase ahead of v5: it split vLLM CI across dedicated runners [9], added a linter that hunts `nn.Buffer` misuse , and shipped v5.16.0 with Qwen4-Exp support , while its Mixture-of-Experts loss was cut by 99.7% . Diffusers likewise tightened its own corners: it now supports group offloading under auto offloading [9], and it fixed LoRA state corruption and pinned revisions once to prevent drift . The pattern is consistent, if not glamorous: each project is auditing its own state management before the next major release.

Finally, moon-ide shipped a reliability overhaul across messaging and sessions, seeding session running status from the backend on load so the UI no longer guesses [10], while Harbor-HF walled off its sandboxes and unified its TypeScript control service . These are the kinds of changes that don't make headlines about new features, but they are why the platform stops flaking in production.