Anthropic released Claude Code v2.1.278, updating the default behavior of auto mode for API, Enterprise, and gateway users. The update shifts the classifier execution to the server side by default, removing the associated overhead costs from user bills. A new row in the /status command indicates whether the server-side classifier is active. Users on Bedrock, Vertex, and Foundry can opt out using the CLAUDE_CODE_AUTO_MODE_SERVER=0 variable.
Read originalThis release stabilizes Claude Code by patching a cascade of crashes and session hangs that plagued recent versions. The most notable functional shift is the fallback to AGENTS.md when CLAUDE.md is absent, aligning with broader industry standards for agent configuration. Gateway improvements allow better proxy handling for egress-bound environments, while numerous fixes address edge cases in file editing, plugin management, and resume functionality. It’s a maintenance-heavy update that restores reliability rather than introducing new capabilities.
This release stabilizes Claude Code by fixing crashes that occurred when resuming sessions with malformed transcripts or memory files. It also improves the developer experience by syncing skills and plugins from your claude.ai account to terminal sessions, ensuring consistency across environments. Security is tightened by preventing install scripts from running on npm-sourced plugins, while usability sees gains in artifact publishing and image handling. The update addresses critical edge cases that previously broke workflows, making the tool more reliable for daily use.
Anthropic quietly patched a critical regression in Claude Code that broke all API requests for users routing traffic through proxies or gateways. The issue stemmed from an invalid input tag 'advisor_20260301' introduced in the previous version, effectively locking out enterprise and custom infrastructure setups. This fix restores basic functionality for developers who cannot connect directly to Anthropic's endpoints, highlighting the fragility of recent CLI updates.
This release quietly closes the hardware gap for local inference by adding default builds for CUDA 13 and ROCm 10.0. NVIDIA users on newer driver stacks can finally run without workarounds, while AMD GPU owners get parity with the latest ROCm version. Apple Silicon support is explicitly disabled in this build, a notable regression for Mac users who need to wait for the next patch. The inclusion of OpenVINO and SYCL builds further cements llama.cpp as the universal runtime for diverse hardware, ensuring no major accelerator is left behind.
This release prioritizes stability over new features, addressing critical correctness issues in vector handling for GET_ROWS operations. By fixing vec4 alignment checks and updating CUDA libraries to versions 12.8 and 13.3, it ensures reliable performance across NVIDIA hardware on both Linux and Windows. The inclusion of ROCm 10.0 builds further solidifies AMD GPU support without requiring complex configuration. While no new model architectures are added, these fixes prevent silent corruption in local inference tasks that could otherwise go unnoticed.
This release finally closes a gap in how llama.cpp handles sliding window attention patterns during model conversion. Previously, loaders silently ignored array-based SWA configurations from models like OLMo2 and Gemma3n, relying on hardcoded defaults that masked potential precision loss. The new Model-Saver now explicitly writes per-layer SWA flags and MLA geometry, ensuring bit-exact roundtrips for a dozen architectures including Plamo3 and Cohere2. This matters because it guarantees that converted GGUF files preserve the exact inference behavior of their original checkpoints, eliminating silent degradation for complex attention mechanisms.