Claude Code v2.1.305 Adds Sandbox Namespace Controls, Fixes Provider Gateway Failures
What changed
Anthropic released Claude Code v2.1.305, introducing configuration options for Linux sandboxing, addressing provider gateway regressions, and updating Model Context Protocol (MCP) server handling.
Key additions and fixes in this release include:
- Linux Sandbox Namespace Controls: Added the
sandbox.allowNestedUserNamespacesconfiguration key. This allows commands running in a Linux sandbox to spawn user namespaces, which is required for tools like Chromium and rootless Podman, though enabling it bypasses the sandbox’s default file isolation protections. - Unattended Watchdog Re-checks: Introduced
CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MSfor theCLAUDE_CODE_RETRY_WATCHDOGunattended mode, allowing the runner to detect when usage limits clear prior to their scheduled reset time. - Gateway and Provider Fallbacks: Fixed a regression introduced in version 2.1.304 where requests failed on AWS Bedrock, Google Cloud Vertex AI, or behind enterprise gateways that rejected tool search requests or omitted
ToolSearchfrom--tools. Claude Code now automatically retries the request without tool search. - Workflow Journal Resumption: Resuming a workflow now correctly loads previously executed
pipeline()andparallel()agents from the execution journal instead of re-running them. - Model Settings and Case Sensitivity: Fixed an issue where
/effortand/autocompactcommands failed to apply when model names differed by letter case or included a[1m]tag relative tomodelOverridesvalues. - MCP Server Management: Modified
allowedMcpServerssomanaged-mcp.jsonstdio servers using${VAR}syntax solely withinenvare no longer filtered out. Enterprise teams wanting to block these must now explicitly list them underdeniedMcpServers. Additionally, path-mounted plugins in Claude Desktop now stripallowed-toolsif organizational policy enforcesallowManagedPermissionRulesOnly. - Self-Hosted Runner Fixes: Resolved a regression from version 2.1.300 where start scripts failed with
Resource temporarily unavailablewhenTERMwas set, and fixed a version 2.1.296 regression affecting git credential programs relying onGIT_CONFIG. - UI and Prompt Updates: Updated the
/artifactsdeletion prompt to accept1orEnterto keep,2to delete, andEscto cancel (yandnare no longer accepted). In/usage, actionable system warnings can now be automatically resolved by pressingf.
Why it matters
For developers building complex workflows on enterprise platforms like Vertex AI and Bedrock, the tool search fix restores stability following gateway filtering changes introduced across recent releases like Claude Code v2.1.295.
Similarly, adding sandbox.allowNestedUserNamespaces addresses a long-standing constraint for developers using browser automation (Chromium) or container tooling inside Linux sandboxes. The change to workflow resumption also ensures multi-agent pipelines do not burn redundant tokens by re-executing steps that were previously logged in the local journal, aligning with recent updates focused on session stability and terminal diagnostic controls.
What to do next
- Update Claude Code: Upgrade to v2.1.305 via your global package installation or CLI runner.
- Configure Linux Sandboxes: If your agent tasks require running rootless Podman containers or headless Chromium browsers inside Linux sandboxes, add
sandbox.allowNestedUserNamespacesto your settings, noting the security trade-offs. - Review MCP Allowlist Configuration: If you maintain
managed-mcp.jsonconfigurations using environment variable expansion, review whether unwanted stdio servers now auto-load upon upgrade, and restrict them usingdeniedMcpServersif necessary. See earlier coverage on MCP protocol updates for underlying permission models. - Update Shortcuts: Note the breaking keybinding change for interactive artifact deletion (
1/Enterto keep,2to delete).
Caveats
- Enabling
sandbox.allowNestedUserNamespacesexplicitly bypasses file protection mechanisms provided by the Linux sandbox environment. Use it only when nested namespaces are required by your tooling. - Traditional
y/nresponses are no longer recognized by the/artifactsdeletion prompt, which may affect manual habits in terminal sessions.
Sources
Share Article
If this article helped you, please share it with others!