feat(llm): add tool approval system for mutating operations

[Kureii]

Summary

Add a human-in-the-loop approval system for destructive/mutating LLM tool calls. Tools marked with mutates: true require explicit user approval before server-side execution.

Changes

• Tool registry: toToolSet() omits execute function for mutating tools, forcing the AI SDK to emit tool-call without auto-executing
• Stream protocol: requiresApproval flag added to tool_use chunks
• Server endpoint: POST /api/llm-chat/execute-tool wraps execution in sql.transactional()
• Client: approveToolCall / rejectToolCall callbacks in useLlmChat
• UI: ToolCallCard shows approve/reject buttons with visual state indicators
• Props threaded through ChatMessage → LlmChat
• ExpandableCard gains defaultExpanded prop for pending approvals

Files (13 files, +257/-32)

• apps/server/src/services/llm/tools/tool_registry.ts - Approval mechanism
• apps/server/src/services/llm/stream.ts - requiresApproval flag
• apps/server/src/routes/api/llm_chat.ts - execute-tool endpoint
• apps/server/src/routes/routes.ts - Route registration
• apps/client/src/services/llm_chat.ts - executeToolCall() client function
• apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.ts - Approval callbacks
• apps/client/src/widgets/type_widgets/llm_chat/ToolCallCard.tsx - Approve/reject UI
• apps/client/src/widgets/type_widgets/llm_chat/ToolCallCard.css - Approval styling
• apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.tsx - Props threading
• apps/client/src/widgets/type_widgets/llm_chat/LlmChat.tsx - Callback forwarding
• apps/client/src/widgets/type_widgets/llm_chat/ExpandableCard.tsx - defaultExpanded
• apps/client/src/widgets/type_widgets/llm_chat/llm_chat_types.ts - Type extensions
• apps/client/src/translations/en/translation.json - Translation keys

[gemini-code-assist]

Code Review

This pull request implements a human-in-the-loop approval mechanism for mutating LLM tool calls. It introduces a workflow where tools flagged as mutating require explicit user approval in the chat interface before execution. Key changes include updating the server-side streaming logic to identify mutating tools, adding a new API endpoint for executing tools after approval, and enhancing the client-side UI with approval/rejection controls and state management. I have no feedback to provide as no review comments were submitted.

[eliandoran]

• Approval seems to be implemented only in the full-screen chat, since in the sidebar I get a tool call with the shield icon but no buttons, making it impossible to continue.
• The footer itself looks a bit out of place, try styling it closer to what we already have and with fewer colors.
• Try looking for inspiration on how other chat interfaces ask for permission to make it stand out a little bit more visually.
• We need a way to be able to skip asking for permission for every mutable tool call. Keep it simple, for example Claude Code in VS Code has "Ask before edits" and "Edit automatically" which is a pretty good mechanism.