Anthropic has released Claude Code v2.1.265, a maintenance update focused on stability and plugin infrastructure. Key changes include fixing prompt-cache reuse errors in resumed subagents and adding support for dynamic plugin directory scanning. The release also addresses security vulnerabilities related to symlink containment checks and improves telemetry handling for Claude apps gateway users. Additional updates cover VS Code integration fixes and performance improvements for large Git repositories.
Read originalAnthropic quietly patched a regression in Claude Code that broke LLM-gateway and proxy setups. The previous version incorrectly forced Cloud-gateway sign-in when the CLAUDE_CODE_USE_GATEWAY variable was set, even with API keys or custom auth headers configured. This update restores the expected behavior where the variable is ignored unless specific Anthropic credentials are present. Developers relying on custom authentication flows can now resume their workflows without configuration changes.
This release targets the most painful friction points in long-running coding sessions: broken prompt caching and lost context during resumption. By fixing how MCP tools and system prompts are recorded and replayed, Anthropic ensures that switching models or reconnecting doesn't wipe your conversation history or force redundant tool definitions. The addition of a maxEffortLevel setting gives developers precise control over compute costs across all cloud providers, while specific fixes for VS Code and remote sessions stabilize the daily driver experience. It’s a maintenance-heavy update that quietly restores reliability to complex workflows.
This release quietly fixes critical permission bypasses that could have allowed dangerous shell commands to slip through the safety net. By tightening how Bash permissions are analyzed and correcting memory directory handling, Anthropic is hardening the tool against accidental data leaks or destructive operations. The update also resolves persistent authentication failures on enterprise cloud platforms like Bedrock and Vertex, ensuring smoother workflows for teams relying on managed settings. These fixes stabilize the agent environment without adding new features.
This release candidate addresses a critical configuration gap for hybrid models in vLLM by enabling the dense prefix cache by default. Previously, these mixed-precision or multi-head architectures likely suffered from inefficient memory usage or required manual flags to achieve optimal performance. By automating this optimization, the update ensures that users running complex model topologies get better throughput out of the box without tweaking internal parameters. It is a quiet but necessary step toward making vLLM robust for heterogeneous hardware setups.
This release quietly extends llama.cpp's hardware support to the latest driver stacks, adding official binaries for ROCm 10.0 and CUDA 13 across Linux and Windows. For AMD users, this means native compatibility with newer GPU architectures without manual compilation tweaks, while NVIDIA users gain access to the latest CUDA runtime optimizations. The inclusion of WebGPU in CI signals ongoing work toward browser-based inference, though it remains a background effort for now. There are no new model formats or quantization methods here, just broader infrastructure coverage that keeps llama.cpp relevant as hardware evolves.
A critical precision bug in llama.cpp’s Apple Silicon backend has been patched, resolving total inference failures on models with high-activation ranges like Mistral Small 4. The issue stemmed from f16 saturation during matrix multiplication, which turned entire output tensors into NaN values for inputs exceeding ~32 tokens. By implementing an exact, power-of-two rescaling mechanism in the Metal kernel, the fix restores correctness without significant performance penalties. This ensures local inference on M-series chips remains viable for complex MoE architectures that previously crashed.