Claude Code v2.1.305: Linux Sandbox Namespaces, Provider Retry Fix, and Self-Hosted Runner Patches
Claude Code v2.1.305 is out, and the headline fix is a 2.1.304 regression: every request failed on Vertex, Bedrock, and other providers when the endpoint or gateway rejected tool search. Behind gateways where ToolSearch is disallowed or left out of --tools, requests also broke. Claude Code now retries without tool search when it sees that rejection.
Sandbox: nested user namespaces on Linux
A new setting, sandbox.allowNestedUserNamespaces, lets sandboxed commands on Linux create user namespaces. That covers things like Chromium and rootless podman — and the release notes are direct about the tradeoff: this lets them get around the sandbox's file protections. Enable it knowingly.
Unattended retry mode
CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS is a new env var for unattended retry mode (CLAUDE_CODE_RETRY_WATCHDOG). It makes the watchdog notice a usage limit that was lifted before its reset time, instead of waiting out the full window.
Other fixes worth knowing
/effortand/autocompactsettings now apply when the model name differs from amodelOverridesvalue by letter case or a[1m]tag.- New worktrees in a bare repository no longer sometimes start from local HEAD instead of the default branch; unchanged subagent worktrees are cleaned up.
- Bash commands on a full disk now say the disk is full and that nothing ran, instead of a raw
ENOSPCfile path. - An MCP server passed by a cloud session or the SDK's
setMcpServersis no longer replaced later in the session by a same-named server from your own settings. - Subagent fallback model switches no longer make Claude name the wrong model for its work.
- Workflow resume serves already-completed
pipeline()andparallel()agents from the run's journal instead of re-running them. - Scheduled Cowork cloud sessions no longer wait five minutes per subagent whose WebFetch needs a URL approval nobody can give.
- Startup warnings moved to the debug log; permission rules that restrict nothing now show as warning rows under the banner.
- Screen reader mode: search boxes say "Search:" instead of a symbol VoiceOver read as "telephone recorder", and highlighted prompt text no longer has spaces inserted inside words.
Self-hosted runner
- Fixed a start script sharing only stdout with the runner or orchestrator failing with
Resource temporarily unavailablewhenTERMis set (2.1.300 regression). - Fixed repository access failing since 2.1.296 when a git credential helper or askpass program reads settings through
GIT_CONFIG. - Fixed the orchestrator reporting a spawn-runner hook as timed out when a background runner kept its output open.
Behavior changes
/usagewarnings are now one-line fixes that name their source; pressfto apply. Same fixes appear in VS Code and/doctor.- MCP consent dialogs are less likely to treat an Enter meant for the prompt as Accept.
allowedMcpServersno longer filtersmanaged-mcp.jsonstdio servers that use${VAR}only inenv— an omitted server now loads on upgrade; usedeniedMcpServersto keep it off./artifactsdelete confirmation now takes1/Enter to keep,2to delete, Esc to cancel;yandnno longer answer it.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Claude Code Randomly Becomes Risk-Averse, Demanding Permission on Routine Tasks
A user reports that Claude Code intermittently shifts from autonomous execution to requiring excessive permissions, even on daily, unchanged workflows like rebuilding a monorepo and running tests.
Claude Boosts Riemann Hypothesis Zero Bound from 41.6% to 67.2%
An unreleased research version of Claude improved the lower bound for zeros on the critical line from 41.6% to 67.2%, using a novel combination of prior work.

Claude's policy filter blocks bioinformatics work with pathogen names
A computational virology researcher reports Claude's usage policy filter flags legitimate bioinformatics scripts when pathogens are named, requiring workarounds like describing tasks without organism names or downgrading to Sonnet 4. The issue affects Claude Code, claude.ai, and both Opus 4.6 and Sonnet 4.6 models.

When RLVR Helps Small Fine-Tuned Models: A 12-Dataset Analysis
A controlled experiment tested adding RLVR reinforcement learning on top of 1.7B parameter models fine-tuned with SFT. Results show text generation tasks improved by +2.0 percentage points on average, while structured tasks declined by -0.7pp.