* fix: resolve subagent persona lookup for 'default' and unify resolution logic
- Add PersonaManager.get_persona_v3_by_id() to centralize v3 persona resolution
- Handle 'default' persona_id mapping to DEFAULT_PERSONALITY in subagent orchestrator
- Fix HandoffTool.default_description using agent_name parameter correctly
- Add tests for default persona in subagent config and tool deduplication
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor: simplify get_default_persona_v3 using get_persona_v3_by_id
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: whatevertogo <whatevertogo@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Soulter <37870767+Soulter@users.noreply.github.com>
Address Sourcery AI review feedback: the image-handling logic was
duplicated for ImageContent and EmbeddedResource cases.
Changes:
- Extract _handle_image_content() helper function
- Consolidate image caching, result appending, and yielding logic
- Reduce code duplication and improve maintainability
Fixes#6140
When a tool returns CallToolResult with multiple content items (e.g.,
both TextContent and ImageContent), the agent was only processing
content[0], ignoring the rest.
Changes:
- Replace direct content[0] access with enumerate(res.content) loop
- Process all content items: TextContent, ImageContent, EmbeddedResource
- Use content_index for image caching to distinguish multiple images
This fixes the issue where tools like Bilibili plugin return both
text descriptions and screenshots, but LLM only received one of them.
Two changes to make the tool schema sent to the LLM deterministic:
1. ToolSet.normalize() — sort tools by name before serialization.
Called at the end of build_main_agent() after all injection passes.
Eliminates ordering drift from plugin load order, MCP reconnection,
and persona tool list differences.
2. Always inject full sandbox tool set — ComputerToolProvider now
returns get_default_sandbox_tools() unconditionally, regardless of
sandbox boot state. Browser tools are always in the schema even if
the sandbox profile lacks browser capability. The executor rejects
calls to unavailable browser tools with a descriptive error instead
of silently omitting them from the schema.
This eliminates the pre-boot/post-boot tool set jump that caused
prefix cache misses on the second request of a conversation.
* feat: add stop functionality for active agent sessions and improve handling of stop requests
* feat: update stop button icon and tooltip in ChatInput component
* fix: correct indentation in tool call handling within ChatRoute class
* feat:为subagent添加后台任务参数
* ruff
* fix: update terminology from 'handoff mission' to 'background task' and refactor related logic
* fix: update terminology from 'background_mission' to 'background_task' in HandoffTool and related logic
* fix(HandoffTool): update background_task description for clarity on usage
---------
Co-authored-by: Soulter <905617992@qq.com>
* feat: implement fallback provider support for chat models and update configuration
* feat: enhance provider selection display with count and chips for selected providers
* feat: update fallback chat providers to use provider settings and add warning for non-list fallback models
* feat: temporary file handling and introduce TempDirCleaner
- Updated various modules to use `get_astrbot_temp_path()` instead of `get_astrbot_data_path()` for temporary file storage.
- Renamed temporary files for better identification and organization.
- Introduced `TempDirCleaner` to manage the size of the temporary directory, ensuring it does not exceed a specified limit by deleting the oldest files.
- Added configuration option for maximum temporary directory size in the dashboard.
- Implemented tests for `TempDirCleaner` to verify cleanup functionality and size management.
* ruff
* Fix TypeError when MCP schema type is a list
Fixes crash in Gemini native tools with VRChat MCP.
* Refactor: avoid modifying schema in place per feedback
* Fix formatting and cleanup comments
- Implemented proactive cron job tools in InternalAgentSubStage for scheduling tasks.
- Created SendMessageToUserTool for sending messages to users based on cron job triggers.
- Added CreateActiveCronTool, DeleteCronJobTool, and ListCronJobsTool for cron job management.
- Introduced CronRoute for handling cron job API requests in the dashboard.
- Developed CronJobPage.vue for managing cron jobs in the dashboard UI.
- Updated SubAgentPage.vue to include persona selection for subagents.
* feat: support anthropic skills
closes: #4687
* chore: ruff
* feat: implement skills management and selection in persona configuration
* feat: enhance skills management with local environment tools and permissions
* feat: context compressor
Co-authored-by: kawayiYokami <289104862@qq.com>
* Add comprehensive tests for ContextManager and ContextTruncator
- Implemented a full test suite for ContextManager covering initialization, message processing, token-based compression, and error handling.
- Added tests for ContextTruncator focusing on message fixing, truncation by turns, dropping oldest turns, and halving.
- Ensured that both test suites validate edge cases and maintain expected behavior with various message types, including system and tool messages.
* feat: add MockProvider for LLM compression tests
* chore: remove lock
* ruff fix
* fix
* perf
* feat: enhance context compression with token tracking and logging
* feat: update logging for context compression trigger
* feat: implement context compression logic with dynamic threshold and token tracking
* fix: reorder import statements for consistency
* feat: add token_usage tracking to conversations and update related processing logic
---------
Co-authored-by: kawayiYokami <289104862@qq.com>
* perf: support extended thinking for Anthropic, DeepSeek reasoning mode, and Gemini text part thought signatures to improve multi-turn reasoning performance.
* chore: remove verbose
* perf
* refactor: remove special tools handling for deepseek-reasoner model in openai source
* fix: improve error handling and logging in InternalAgentSubStage processing
* refactor: remove unused reasoning content from Gemini source processing
* refactor: enhance modality determination logic in useProviderSources
Co-authored-by: kawayiYokami <289104862@qq.com>
* fix: improve error handling in tool execution to prevent infinite tool call loops
- Enhanced error handling in `call_local_llm_tool` to provide more informative exceptions for ValueError and TypeError, including detailed parameter information.
- Updated `ToolLoopAgentRunner` to yield appropriate messages for cases with no response or unsupported types, ensuring clearer communication to users.
- Improved logging and messaging consistency across tool execution processes.
* refactor: clean up unused router parameter in message retrieval functions
- Removed the unused `router` parameter from `getSessionMessages` and related function calls in `Chat.vue` and `useMessages.ts`.
- Commented out the `tool_calls` dictionary in `chat.py` for clarity, indicating it is not currently in use.
* fix: enhance exception handling in tool execution for clearer error reporting
- Improved exception handling in `call_local_llm_tool` by chaining exceptions for ValueError and TypeError, providing more context in error messages.
- Ensured that traceback information is preserved in raised exceptions for better debugging.
* perf(agent): add max step limit to prevent infinite tool call loops
* feat: implement max step limit handling in main agent runner
- Enhanced the agent runner to enforce a maximum step limit, logging a warning and forcing a final response when the limit is reached.
- Updated message handling to append a user prompt when the tool call limit is exceeded.
- Refactored tool response handling to yield appropriate messages based on the response type, including handling cases with no response or unsupported types.
- Improved conversation message formatting to ensure consistent output in the assistant's responses.
* chore: ruff format
---------
Co-authored-by: Soulter <905617992@qq.com>
* feat: enhance tool call handling and UI integration for tool calls render
- Added support for tool call messages in the agent runner and webchat event handling.
- Implemented JSON message component for structured tool call data.
- Updated chat route to save tool call information in message history.
- Enhanced frontend to display tool call details in a collapsible format, including status and results.
- Introduced elapsed time tracking for ongoing tool calls in the chat interface.
* fix: improve message handling in agent run utility and tool loop runner
- Refactored message sending logic in `astr_agent_run_util.py` to use `msg_chain` directly for better clarity.
- Added a check in `tool_loop_agent_runner.py` to ensure `tool_call_result_blocks` is not empty before yielding the last tool call result, preventing potential errors.
* refactor: enhance message structure and UI for chat components
- Updated message handling in `MessageList.vue` to support structured message parts, including plain text, images, audio, and files.
- Improved the `Chat.vue` component styles for better visual consistency.
- Refactored message parsing logic in `useMessages.ts` to accommodate new message formats and ensure proper rendering of embedded content.
- Removed deprecated tool call handling from the message structure, streamlining the message display process.
* chore: ruff format
* feat: implement agent statistics tracking and display in chat
- Added `AgentStats` and `TokenUsage` data classes to track agent performance metrics.
- Enhanced `ToolLoopAgentRunner` to collect and update agent statistics during execution.
- Integrated agent statistics sending to webchat for real-time updates.
- Updated chat route to save and display agent statistics in message history.
- Improved frontend components to visualize agent statistics, including token usage and duration metrics.
* fix: improve message handling in Telegram event and agent run utility
- Updated message sending logic in `astr_agent_run_util.py` to send the correct message chain for tool calls.
- Enhanced `tg_event.py` to edit messages during streaming breaks, improving message management and user experience.
- Added error handling for message editing failures to ensure robustness.
* chore: ruff format
LIghtJUNction
whatevertogo
ccsang
zenfun
Gao Jinzhe
Soulter
邹永赫
エイカク
Luna_Dol
Chen
時壹
Li-shi-ling
Dt8333
鸦羽
boushi1111
Soulter
vmoranv
letr
Yokami