Claude Code v2.0.5

What: Claude Code now supports extended thinking mode when using Claude Sonnet 4.5, enabling more transparent reasoning for complex tasks.

How to use:

# Enable thinking via environment variable
MAX_THINKING_TOKENS=10000 claude

# Or configure in your settings file
# Set alwaysThinkingEnabled: true in your Claude Code settings

Details:

• Only available when using Claude Sonnet 4.5 model
• Can be controlled via MAX_THINKING_TOKENS environment variable (set to any positive integer to enable)
• Can be configured permanently via alwaysThinkingEnabled in settings
• Subject to Statsig feature gate thinking_on_default for gradual rollout
• Provides visible reasoning steps during Claude's problem-solving process
• Evidence: hL1() at line 364041 in v2.0.5

What: Session entries are now deduplicated before persistence, preventing redundant API calls when the same conversation entry is encountered multiple times.

Details:

• Uses UUID tracking to identify already-persisted entries
• Reduces unnecessary backend calls and potential conflicts
• Improves overall session handling performance
• Evidence: K8B() at line 397234 in v2.0.5, modified from QPB() at line 23491 in v2.0.3

What: Internal tool descriptions now emphasize "batch" execution rather than "parallel" execution, providing more accurate guidance for Claude's behavior.

Details:

• Updated Glob tool description from "in parallel" to "as a batch"
• Better reflects the actual execution model of tool calls
• Improves Claude's understanding of how multiple tool calls are processed
• Evidence: Glob tool description at line 12298 in v2.0.5

What: Eliminated the "tool use concurrency issues" error message that instructed users to run /rewind to recover conversations.

Details:

• The underlying API error condition (400 status for tool_use without tool_result blocks) is no longer expected
• Suggests server-side fix has been implemented
• Users will no longer encounter this specific error message
• Evidence: Error handling removed from fk1() at lines 23711-23720 in v2.0.3

Removed workspace trust validation checks from AWS credential export and refresh functions, suggesting the security model has evolved to handle trust validation elsewhere in the codebase.

Added telemetry for Model Context Protocol (MCP) tool usage, tracking both tool count and token consumption for optimization purposes.

Removed extraction of billingType field from OAuth profile responses, as this field is no longer needed.

• Removed unused stream module import
• Updated to use ES6 import for cwd() from node:process