On September 7, 2026, the OpenAtom openKylin (hereinafter referred to as "openKylin") community officially pushed a new round of updates for openKylin 3.0 AgentOS. This update focuses on improving the reliability of agent task completion, fixing dozens of reliability defects in the system agent, CUA (Desktop Operation Agent), unified model inference service, and agent runtime. Additionally, practical features such as voice interaction and human-machine intelligent collaboration have been added, covering multiple core modules including task management, runtime, and model inference services.

• Fixed: Application startup issues, where the open_app action opened the wrong application or had no response after execution, making application startup more precise and reliable
• Optimized: Anti-loop mechanism, added "no-progress action count threshold" monitoring, which actively prompts strategy changes when repeated invalid actions reach the set count, effectively preventing tasks from falling into infinite loops
• Optimized: Mouse up/down scrolling logic, when the interface cannot locate the target element, the system will more actively use scroll wheel actions to search, improving the success rate of element finding
• Optimized: Text input action mechanism, added press_enter and select_all parameters, adapting to complex scenarios such as confirming submission after modification, form filling, and document editing
• Optimized: Log storage strategy, only retaining historical task steps and screenshots, effectively controlling storage usage while facilitating problem tracing
• Added: Human-machine intelligent collaboration mechanism, when task execution steps fail multiple times and the goal still cannot be achieved after strategy adjustment, the system will automatically report the failure reason to the system agent and actively seek user manual intervention, achieving a human-machine collaborative closed loop
• Added: Function to save screenshots and step information during task execution as a dataset, facilitating subsequent analysis, model training, and problem backtracking

• Fixed: Issue where breakpoint markers remained after sending a new message when there was a breakpoint in the task, ensuring session status accurately reflects current progress
• Fixed: Issue where the system agent's token consumption and time consumption were only updated after task completion when resuming a breakpoint
• Fixed: Issue where some session status flags displayed incorrectly in multi-session scenarios
• Fixed: Issue where some session message replies disappeared after restarting the application
• Fixed: Issue where the real-time process position was always at the top in multi-session scenarios
• Added: Task breakpoint discard function, allowing users to flexibly manage breakpoints as needed, avoiding redundant interruptions that affect the process

• Fixed: Issue where scheduled task reply content could only be displayed after restarting the program
• Fixed: Issue where time consumption information was lost after program restart
• Fixed: Issue where the skill self-evolution switch in settings was ineffective
• Optimized: pip installation process in tasks, using the domestic Tsinghua mirror source by default, significantly improving dependency package download speed and avoiding network latency issues
• Optimized: Time from program startup to main interface display, reducing waiting time

• Added: Voice recognition, supports local updates of insertion position text, supports multiple voice models and related configurations on the Qianwen platform
• Added: Voice broadcast, supports multiple voice models and related configurations on the Qianwen platform

• Fixed: False timeout reporting during inference, resolving issues where runtimes error prompts appeared when the model was loading, switching, or inferring
• Fixed: Inaccurate task queue status, waiting tasks clearly enter the queued state, and running statistics no longer count waiting tasks as in-progress tasks
• Optimized: Frontend auto-refresh fault tolerance, retains the last successful data when auto-refresh fails, and prompts status refresh delay, avoiding page status anomalies or false service unavailability due to single interface jitter
• Optimized: Model management information display, the model list hides internal ports, only showing brief information, optimizing the user's business call experience
• Optimized: Automatic information filling when adding a model, automatically fills in the model ID and model type (such as pure text, multimodal, or Embedding) based on model information, also supports manual modification and update
• Added: Task anti-loss and anti-duplicate execution mechanism, when the same task ID is submitted repeatedly during execution, it returns the existing task status; after completion, repeated submission returns the original result, avoiding duplicate execution and result overwriting
• Added: Client disconnection protection, non-streaming requests continue background execution after client interruption, window closure, or network disconnection, and are written to history after completion; streaming requests retain clear failure semantics for easy troubleshooting
• Added: Embedding model usage and configuration, covering model addition, modification, invocation, and error handling, etc., the Embedding model input automatically crops when exceeding context without errors

• Fixed: Issue where skill filtering was ineffective
• Fixed: Issue where Embedding model inference was ineffective
• ．．．．．．