Claude Code 2.1.183

• [changed] Updated the core system prompt to instruct the agent to always write tests when completing coding tasks.
• [changed] Refined plan mode instructions to more clearly prevent edits and non-readonly tool use until the user is ready.
• [changed] Improved focus mode handling so only the final text message is shown to the user, with clearer context management when conversations are summarized.
• [changed] Adjusted interruption and permission denial messages for consistency across tool use rejections.
• [changed] Updated directory resource handling to return direct children (including subdirectories) when a URI represents a directory.

Models & prompts

• Refined several built-in behavioural instructions around context management, ambient context, and when to act versus gather more information.
• Clarified guidance for handling user corrections or preferences across sessions.

Under the hood

• Added CLAUDE_CODE_CONNECT_TIMEOUT_MS, CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS, and CLAUDE_CODE_WEBSEARCH_USE_CCR_PROXY environment variables for tuning connection and background behaviour.
• Added support for the code-execution-2025-08-25 API beta header.
• Expanded internal feature flags related to MCP first-party scope, rate limit model switching, and thinking behaviour.
• Added backend routes for OAuth v2 access and environment work tracking.

Upgrade notes

No special actions required.