* feat: add SkillTool component and integrate into agent tools
- Introduced SkillTool component for rendering skill-related functionality.
- Updated MessageAgentTools to include SkillTool in the tool renderers.
- Enhanced MessageTool to recognize 'Skill' as a valid agent tool type.
- Modified handleUserMessage to conditionally handle text blocks based on skill inclusion.
- Added SkillToolInput and SkillToolOutput types for better type safety.
* feat: implement command tag filtering in message handling
- Added filterCommandTags function to remove command-* tags from text content, ensuring internal command messages do not appear in the user-facing UI.
- Updated handleUserMessage to utilize the new filtering logic, enhancing the handling of text blocks and improving user experience by preventing unwanted command messages from being displayed.
* refactor: rename tool prefix constants for clarity
- Updated variable names for tool prefixes in MessageTool and SkillTool components to enhance code readability.
- Changed `prefix` to `builtinToolsPrefix` and `agentPrefix` to `agentMcpToolsPrefix` for better understanding of their purpose.
* fix(ToolPermissionRequestCard): simplify button rendering by removing suggestion handling
* ✨ feat: add CachedPluginsDataSchema for plugin cache file
- Add Zod schema for .claude/plugins.json cache file format
- Schema includes version, lastUpdated timestamp, and plugins array
- Reuses existing InstalledPluginSchema for type safety
- Cache will store metadata for all installed plugins
* ✨ feat: add cache management methods to PluginService
- Add readCacheFile() to read .claude/plugins.json
- Add writeCacheFile() for atomic cache writes (temp + rename)
- Add rebuildCache() to scan filesystem and rebuild cache
- Add listInstalledFromCache() to load plugins from cache with fallback
- Add updateCache() helper for transactional cache updates
- All methods handle missing/corrupt cache gracefully
- Cache auto-regenerates from filesystem if needed
* ✨ feat: integrate cache loading in AgentService.getAgent()
- Add installed_plugins field to GetAgentResponseSchema
- Load plugins from cache via PluginService.listInstalledFromCache()
- Gracefully handle errors by returning empty array
- Use loggerService for error logging
* 🐛 fix: break circular dependency causing infinite loop in cache methods
- Change cache method signatures from agentId to workdir parameter
- Update listInstalledFromCache(workdir) to accept workdir directly
- Update rebuildCache(workdir) to accept workdir directly
- Update updateCache(workdir, updater) to accept workdir directly
- AgentService.getAgent() now passes accessible_paths[0] to cache methods
- Removes AgentService.getAgent() calls from PluginService methods
- Fixes infinite recursion bug where methods called each other endlessly
Breaking the circular dependency:
BEFORE: AgentService.getAgent() → PluginService.listInstalledFromCache(id)
→ AgentService.getAgent(id) [INFINITE LOOP]
AFTER: AgentService.getAgent() → PluginService.listInstalledFromCache(workdir)
[NO MORE RECURSION]
* 🐛 fix: update listInstalled() to use agent.installed_plugins
- Change from agent.configuration.installed_plugins (old DB location)
- To agent.installed_plugins (new top-level field from cache)
- Simplify validation logic to use existing plugin structure
- Fixes UI not showing installed plugins correctly
This was causing the UI to show empty plugin lists even though plugins
were correctly loaded in the cache by AgentService.getAgent().
* ♻️ refactor: remove unused updateCache helper
* ♻️ refactor: centralize plugin directory helpers
* feat: Implement Plugin Management System
- Added PluginCacheStore for managing plugin metadata and caching.
- Introduced PluginInstaller for handling installation and uninstallation of plugins.
- Created PluginService to manage plugin lifecycle, including installation, uninstallation, and listing of available plugins.
- Enhanced AgentService to integrate with PluginService for loading installed plugins.
- Implemented validation and sanitization for plugin file names and paths to prevent security issues.
- Added support for skills as a new plugin type, including installation and management.
- Introduced caching mechanism for available plugins to improve performance.
* ♻️ refactor: simplify PluginInstaller and PluginService by removing agent dependency and updating plugin handling
* ⬆️ chore: upgrade claude agent sdk to 0.1.15
* ✨ feat: add initial tool permission approval system
- Add promptForToolApproval function for real-time tool approval UI
- Integrate canUseTool callback into ClaudeCodeService
- Create tool-permissions.ts module for permission handling
- Set foundation for enhanced tool permission system #10738
This provides the basic infrastructure for displaying tool approval
prompts and getting user consent before agents execute potentially
dangerous operations.
* restore to main for
restore to main for
* ✨ feat: implement agent tool permission request system
Add comprehensive tool permission management for Claude Code agents with:
- IPC channels for bidirectional permission requests/responses between main and renderer
- Permission request queue with timeout (30s), abort signal handling, and auto-cleanup
- Auto-approval system via CHERRY_AUTO_ALLOW_TOOLS env var and default allow-list (Read, Glob, Grep)
- New ToolPermissionRequestCard UI component for user approval with input preview
- Redux store (toolPermissions) for tracking pending/resolved permission requests
- User input stream architecture allowing dynamic user messages during agent execution
Key changes:
- packages/shared/IpcChannel.ts: Add AgentToolPermission_* channels
- src/main/services/agents/services/claudecode/: Refactor canUseTool with permission prompts
- src/renderer/src/store/toolPermissions.ts: New Redux slice for permission state
- src/renderer/src/pages/home/Messages/Tools/ToolPermissionRequestCard.tsx: Interactive approval UI
* refactor: simplify ToolPermissionRequestCard by removing unused imports and suggestion handling logic
* feat: add i18n
* fix(i18n): Auto update translations for PR #10743
---------
Co-authored-by: dev <verc20.dev@proton.me>
Co-authored-by: GitHub Action <action@github.com>
* Revert "fix: make anthropic model provided by cherryin visible to agent (#10695)"
This reverts commit 7b3b73d390.
* fix: agent supported model filter
- Unify buildClaudeCodeSystemMessage implementation in shared package
- Refactor MessagesService to provide comprehensive message processing API
- Extract streaming logic, error handling, and header preparation into service methods
- Remove duplicate anthropic config from renderer, use shared implementation
- Update ClaudeCodeService to use append mode for custom instructions
- Improve type safety and request validation in message processing
- Replace @anthropic-ai/claude-code with @anthropic-ai/claude-agent-sdk@0.1.1
- Update all import statements across 4 files
- Migrate patch for Electron compatibility (fork vs spawn)
- Handle breaking changes: replace appendSystemPrompt with systemPrompt preset
- Add settingSources configuration for filesystem settings
- Update vendor path in build scripts
- Update package name mapping in CodeToolsService
- Add support for anthropicApiHost configuration in providers
- Improve model filtering for Anthropic-compatible providers
- Add isAnthropicModel function to validate Anthropic models
- Update ClaudeCode service to support compatible providers
- Enhance logging and error handling in API routes
- Fix model transformation and validation logic
- Log each content block when processing user or assistant messages
- Maintain existing switch case logic for text block handling
- Improve debugging visibility for multi-block message processing
- Correct toolName lookup to use tool_use_id instead of blockKey in tool-result chunks
- Add debug logging for stream event handling
- Update contentBlockState key to use event.content_block.id for tool_use events
- Correctly reference toolName from contentBlockState using blockKey instead of block.tool_use_id
- Ensure proper tool result chunk generation when handling assistant messages
- Maintain consistent data structure for tool call processing
- Populate toolName in tool-result chunks from contentBlockState
- Add onSessionCreated callback to SessionModal for post-creation actions
- Return created session from useSessions hook and update SWR cache optimistically
- Rename built_in_tools field to tools for consistency
- Add type field to Tool schema (builtin/mcp/custom)
- Consolidate tool handling in BaseService with listMcpTools method
- Remove unused CreateSessionResponse and related schemas
- Clean up unused imports and dead code in session handlers
- Unify agent and session tool resolution logic
- Add shared Anthropic utilities package with OAuth and API key client creation
- Implement provider-specific message routing alongside existing v1 API
- Enhance authentication middleware with priority handling (API key > Bearer token)
- Add comprehensive auth middleware test suite with timing attack protection
- Update session handling and message transformation for Claude Code integration
- Improve error handling and validation across message processing pipeline
- Standardize import formatting and code structure across affected modules
This establishes the foundation for Claude Code OAuth authentication while maintaining
backward compatibility with existing API key authentication methods.
