Claude Code 2.1.248 adds restricted mode and stronger diagnostics

Claude Code’s August 27–28 changelog adds a new restricted mode and several controls for teams running Claude Code in managed or automated environments. Version 2.1.248 adds the mode and related diagnostics; the August 28 entry for 2.1.250 contains bug fixes and reliability improvements.

A tighter tool boundary with restricted mode

Users can enable restricted mode with --restricted or CLAUDE_CODE_RESTRICTED=1. Anthropic says this removes built-in tools that run commands or code and removes WebFetch unless it is explicitly named in --tools. File tools remain limited to the working directory, bypassPermissions is refused, and user, project, and local settings files are ignored.

This is a configuration boundary, not a guarantee that every workflow is safe by default. Teams should test the exact tool allowlist, working-directory assumptions, and provider configuration they need before using the mode for automation.

Agent caching and self-hosted runner controls

Agent frontmatter now accepts experimental cacheTtl values of "5m" or "1h". The per-agent prompt-cache setting applies when no subagent TTL setting is configured, giving teams a more explicit way to balance cache reuse and prompt freshness.

Self-hosted runner users can override the registered client label with claude self-hosted-runner --client-label <label> or SELF_HOSTED_RUNNER_CLIENT_LABEL. The default remains the hostname, while the override can make runner selection clearer in a managed fleet.

Better visibility into managed settings and usage

Claude Code now warns at startup when server-managed settings fail to load. /doctor and /status explain the load failure or why settings were not fetched, including deployments using Bedrock, Vertex, third-party providers, and a custom ANTHROPIC_BASE_URL.

The changelog also adds Enterprise usage-credit support for organizations billed through AWS Marketplace, self-serve Enterprise, and Enterprise trials. Members can request a higher usage limit from their administrator.

Reliability fixes across long-running work

The release fixes a prompt-cache miss and lost extended-thinking context that could occur roughly once an hour in long sessions after an OAuth token refresh. It also fixes a cache miss on the first resumed turn when ScheduleWakeup changed between a session and its --resume.

Other fixes cover authentication lock contention, stale agent-view sessions, hook errors that left background sessions waiting silently, and several Windows and /web-setup edge cases. The August 28 entry itself is labeled bug fixes and reliability improvements, so this article does not infer additional features from that entry.

Sources

• Claude Code Docs — “Changelog” (versions 2.1.248 and 2.1.250): https://code.claude.com/docs/en/changelog
• Claude Code Docs — “What’s new”: https://code.claude.com/docs/en/whats-new
• ClaudeDevs on X — official developer updates: https://x.com/ClaudeDevs