Three new environment variables added for custom model picker entries: ANTHROPIC_CUSTOM_MODEL_OPTION (model ID), ANTHROPIC_CUSTOM_MODEL_OPTION_NAME (display name), and ANTHROPIC_CUSTOM_MODEL_OPTION_DESCRIPTION (description). [lines 45-47] [Source]
Updated CLAUDE_CODE_PLUGIN_SEED_DIR description to reflect support for multiple directories separated by : (Unix) or ; (Windows). [line 56] [Source]
Added new StopFailure hook event that fires when a turn ends due to an API error. Output and exit code are ignored. [line 116] [Source]
Added matcher support for StopFailure (error type), InstructionsLoaded (load reason), Elicitation, and ElicitationResult (MCP server name). [lines 124-127] [Source]
Clarified that Stop hooks fire on API errors via StopFailure instead. [line 136] [Source]
Added new StopFailure hook event to the events table. [line 157] [Source]
Updated matcher table: StopFailure supports error type matchers (rate_limit, authentication_failed, billing_error, etc.); InstructionsLoaded now supports load reason matchers (session_start, nested_traversal, etc.); Elicitation/ElicitationResult support MCP server name matchers. [lines 166-170] [Source]
InstructionsLoaded now supports matchers (previously listed as no-matcher). [line 230] [Source]
Full StopFailure hook event documented with input schema including error, error_details, and last_assistant_message fields. [lines 272-304] [Source]
Clarified that permission_mode is not present in all hook event JSON — only events that actually receive it include it. SessionStart, InstructionsLoaded, Notification, SubagentStart, PreCompact, PostCompact, SessionEnd, and ConfigChange no longer show permission_mode in their example JSON.
Updated hooks lifecycle diagram to include StopFailure.
Added /remote-control to the VS Code command menu description. [line 376] [Source]
New sessions now receive AI-generated titles based on your first message. [line 385] [Source]
Added new section documenting the built-in IDE MCP server: transport, authentication (random token written to ~/.claude/ide/), and the two model-visible tools (mcp__ide__getDiagnostics and mcp__ide__executeCode). [lines 393-408] [Source]
New section "Add a custom model option" documenting how to use ANTHROPIC_CUSTOM_MODEL_OPTION, ANTHROPIC_CUSTOM_MODEL_OPTION_NAME, and ANTHROPIC_CUSTOM_MODEL_OPTION_DESCRIPTION to add a non-standard model ID to the /model picker for gateway deployments. [lines 181-199] [Source]
Added documentation for layering multiple CLAUDE_CODE_PLUGIN_SEED_DIR directories: separate paths with : on Unix or ; on Windows; first seed containing a given plugin wins. [line 694] [Source]
Replaced the bullet-list of available hook events with a comprehensive table matching the full hooks reference, including new events like StopFailure, InstructionsLoaded, ConfigChange, WorktreeCreate/Remove, Elicitation/ElicitationResult. [lines 49-73] [Source]
Added http as a supported hook type alongside command, prompt, and agent. [line 80] [Source]
Clarified that on Team and Enterprise plans Remote Control is off by default until an admin enables the toggle specifically. [line 131] [Source]
Session title priority order now documented explicitly (name arg → /rename → last message → first prompt). [lines 149-156] [Source]
Added three new troubleshooting sections: "Remote Control is not yet enabled for your account" (env var conflicts), "Remote Control is disabled by your organization's policy" (API key vs OAuth, admin toggle, compliance restrictions), and "Remote credentials fetch failed". [lines 165-184] [Source]
Corrected sandbox path prefix table: single / now means absolute path (not settings-relative); ./ or no prefix means relative to project root (for project settings) or ~/.claude (for user settings). The old //path prefix still works for backwards compatibility. [lines 218-226] [Source]
Updated showTurnDuration description: clarified it now appears in /config as Show turn duration (previously said "Edit ~/.claude.json directly to change"). [line 239] [Source]
Changed recommended persistent memory scope from user to project — project scope is shareable via version control. Use user for cross-project knowledge. [line 308] [Source]
Same changes as ide-integrations.md (this file is identical): /remote-control in command menu, AI-generated session titles, and new built-in IDE MCP server section.
Added note that model capabilities and token limits can be queried programmatically via the Models API, which now returns max_input_tokens, max_tokens, and a capabilities object. [line 39] [Source]
metadata.user_id maximum length increased from 256 to 512 characters.
Added thinking.display parameter to ThinkingConfigEnabled and ThinkingConfigAdaptive — accepts "summarized" (default) or "omitted" to redact thinking content while preserving the signature for multi-turn continuity.
Models API response now includes capabilities object (batch, citations, code_execution, context_management, effort, image_input, and more), max_input_tokens, and max_tokens for each model. All SDK variants (Python, TypeScript, Go, Java, C#, Ruby, and beta endpoints) updated accordingly.
Fixed code example for pause_after_compaction: now correctly preserves the prior exchange plus current user message (3 messages total) instead of only 2. [lines 1166-1178] [Source]