Claude Code v2.1.305 fixes Vertex/Bedrock tool search regression and adds Linux sandbox improvements
Version 2.1.305 of Anthropic’s Claude Code CLI addresses a 2.1.304 regression that broke tool search on Vertex, Bedrock, and other providers, and introduces Linux sandbox changes to support nested user namespaces for rootless podman and Chromium.
1 source · cross-referenced
- Fixes a regression in v2.1.304 that caused every request to fail on Vertex, Bedrock, and other providers when endpoints or gateways rejected tool search or disallowed it.
- Adds Linux sandbox option to allow nested user namespaces, enabling rootless podman and Chromium to bypass sandbox file protections.
- Improves model name handling for subagents, worktree behavior in bare repos, and MCP server identity consistency.
- Updates /usage command to show one-line fixes with a single-letter apply command and adds VS Code and /doctor parity.
- Adjusts Cloud sessions, Remote Control, and Claude Tag behaviors including confirmation prompts and admin page fixes.
Anthropic released Claude Code CLI v2.1.305 with fixes for a regression introduced in v2.1.304 that caused every request to fail on Vertex, Bedrock, and other providers when the endpoint or a gateway rejected tool search or disallowed it. The update also adds a Linux sandbox option to allow nested user namespaces, enabling sandboxed commands on Linux to create user namespaces for rootless podman and Chromium and bypass the sandbox’s file protections.
The changelog documents improvements to subagent behavior, including fixing cases where Claude named the wrong model for a subagent after it switched to a fallback model and ensuring cloud sessions and teammates notify Claude when compaction omits the oldest messages to fit summaries. Worktree handling in bare repositories is corrected so new worktrees start from the default branch rather than local HEAD, and unchanged subagent worktrees are cleaned up reliably.
Model configuration issues are addressed: /effort and /autocompact settings now apply even when the model name differs from a modelOverrides value by letter case or a bracketed tag. MCP server identity is stabilized so servers passed by cloud sessions or the SDK’s setMcpServers are not later replaced by same-named servers from user settings.
The /usage command is updated to display each warning as a one-line fix that names its source, and a single-letter command applies the fix; the same fixes are now visible in VS Code and via /doctor. Consent dialogs for MCP servers are improved to reduce accidental acceptance of prompts, and an Image mod’s rendering in iTerm2 3.5+ is corrected for fullscreen mode.
Cloud sessions and Remote Control receive fixes: Cloud environments admin pages now accept existing environments without an 'installs packages' error, and deleting self-hosted environments shows what blocks deletion and what it affects. Remote Control’s spawn-mode messages are clarified, and confirmation prompts for /artifacts delete are changed to use 1/Enter to keep, 2 to delete, and Esc to cancel.
Claude Tag receives multiple fixes and improvements: public Slack channels not joined by Claude are added to the admin interface for place search setup; repository moves are handled so previously added repos can be re-added; routines in Slack threads no longer stop posting after a restart; and admin pages in large organizations now allow edits while the channel list loads.
Self-hosted runner fixes include correcting a regression since v2.1.300 where a start script sharing only stdout with the runner failed with 'Resource temporarily unavailable' when TERM is set, and restoring repository access since v2.1.296 when git credential helpers or askpass programs read settings through GIT_CONFIG.
- Oct 20, 2026 · Hugging Face Blog
NVIDIA releases Cosmos3-H-Dreams, a generative surgical simulator for real-time clinical training
Trust79 - Oct 20, 2026 · TechCrunch — AI
Microsoft launches Guardian Agents to monitor AI agents amid insider-threat warnings
Trust72 - Oct 20, 2026 · Hugging Face
PyTorch 2.15 brings FP8, faster LLM decoding, and memory reductions to Apple MPS
Trust84