A cross-platform intelligent task processing method and system based on a cloud-local hybrid architecture

Abstract

This invention discloses a cross-platform intelligent task processing system and method based on a cloud-local hybrid architecture, belonging to the fields of artificial intelligence and automation technology. It includes: user terminals receiving natural language instructions through a unified message gateway and sending them to a task processing hub; the hub dynamically selecting a local or cloud model for semantic understanding and task decomposition according to a strategy, generating atomic operation descriptions; distributing tasks to heterogeneous target terminals based on device capability maps; each terminal converting the atomic operation descriptions into local system calls through a cross-platform atomic operation adapter library and executing them securely, with risky operations executed in a sandbox; communication employing multiple encryption methods; the hub integrating a billing module for cost estimation, real-time billing, and budget control; and the hub feeding back execution results to the original message channel through a unified message gateway. This invention ensures data security through intelligent planning and privacy separation, decouples the core architecture from the deployment form, and supports multiple deployment forms and cross-platform collaboration.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence technology, automation technology, distributed computing, human-computer interaction, and billing management. Specifically, it relates to a method and system that combines cloud intelligence with local execution, integrates a unified message gateway, a cross-platform atomic operation adapter library, and transparent AI resource scheduling and billing management, supports dynamic switching between various local large models and cloud large models, and introduces sandbox isolation, cross-session intelligent memory, private skill repository, deep integration with enterprise business systems, and a multi-agent parallel framework to achieve secure, efficient, and cost-controllable cross-platform automated task processing. Background Technology

[0002] With the widespread adoption of digital office solutions, users often need to repeatedly perform tasks such as file organization, data processing, information summarization, and cross-application operations on multiple devices (such as Windows / Mac computers, Android / iOS / HarmonyOS phones, tablets, etc.). Existing solutions have significant shortcomings:

[0003] 1. Pure cloud-based automation tools: These tools require uploading user data containing sensitive information (such as local files, email texts, and chat logs) to the cloud for processing, posing a serious risk of privacy leaks. They are also affected by network latency, cannot be used offline, and are highly dependent on the network.

[0004] 2. Purely local automation tools (such as desktop scripts and mobile shortcuts): Limited by the computing power and AI capabilities of a single device, they have difficulty understanding complex natural language commands, cannot achieve cross-device collaboration, have poor functional scalability, and require users to have certain programming knowledge.

[0005] 3. Existing hybrid architecture solutions: These solutions typically tightly bind the architecture to specific business deployment models (such as SaaS), resulting in rigid technical solutions that are easily circumvented by changing the deployment form (such as switching to standalone applications, private aPaaS, or LAN master control). They also lack a unified task processing hub, which limits the scope of protection.

[0006] 4. Opaque and Uncontrollable AI Service Costs: After users call large-scale cloud model APIs, they can usually only view aggregated bills in the service provider's backend, unable to perceive the cost of a single call in real time. Complex billing models (such as per token or per request) make it difficult for users to predict costs, easily leading to unexpectedly high bills. Furthermore, users typically manually bind API keys to a particular service provider; when that service experiences high latency, failures, or poor performance for specific tasks, there is no automatic or intelligent switching of requests to better or more economical alternatives, resulting in a lack of global cost and performance trade-offs.

[0007] 5. Billing is disconnected from business operations: Fees cannot be linked to specific business operations and lack insights from a business perspective (such as helping users understand "where the money is spent").

[0008] 6. Difficulty in integrating domestically produced large models: Different manufacturers have different APIs and authentication methods, making it difficult for users to manage and switch between them on the same platform, thus failing to fully leverage the advantages of domestically produced models.

[0009] 7. Dispersed message channel integration: Existing tools cannot achieve unified access across multiple platforms and session continuity. Users need to switch between different applications, resulting in a fragmented experience and increased user burden.

[0010] 8. Weak secure communication mechanisms: Most solutions only use basic transmission encryption and lack multi-layered protection such as end-to-end encryption, session-level encryption, and data anonymization, which poses a risk of data leakage.

[0011] 9. Insufficient security of automated execution: The lack of mechanisms such as isolation sandboxes, operation previews, hierarchical permissions, and operation rollback for high-risk operations makes the automated process uncontrollable for users, which can easily lead to losses due to misoperation.

[0012] 10. Rigid Model Selection: In existing AI services, while domestically developed large-scale models are developing rapidly, integration is difficult, making it hard for users to freely switch between international and domestic models on the same platform. Meanwhile, local model deployment is complex, and ordinary users cannot use them with a single click. Furthermore, existing solutions typically use either large cloud-based models or small local models, failing to dynamically switch between local and cloud-based models based on task complexity, privacy requirements, and cost budgets, thus failing to meet diverse user needs.

[0013] 11. Difficulty in implementing cross-platform atomic operations: lack of a unified, scalable adapter layer to encapsulate the differences in the underlying operating system.

[0014] 12. Pure cloud tools pose privacy risks and rely on the network; pure local tools are limited by device computing power and cannot collaborate across devices; hybrid architectures are often strongly tied to specific deployment models and are easily circumvented; AI service costs are opaque; message channels are fragmented; automated execution lacks security; and model selection is rigid.

[0015] 13. User experience and security need improvement: There are shortcomings in deliverable preview, cross-session memory, sandbox isolation, enterprise skill management, and business system integration, and security needs to be improved.

[0016] Therefore, there is an urgent need for an automated task processing solution that can leverage the powerful AI capabilities of the cloud, strictly guarantee user data privacy, and whose core concept is not limited by specific deployment forms. It should also support full platform coverage, flexible interaction, multi-layered security encryption, controllable local execution, intelligent dynamic model selection and transparent billing management, and integrate multiple message channels and a unified and scalable cross-platform atomic operation adapter library. Summary of the Invention

[0017] This invention aims to overcome the technical problems of insufficient privacy protection, difficulties in cross-platform collaboration, opaque AI service costs, and fragmented message channel integration in existing automation tools. It proposes an innovative cross-platform intelligent task processing method and system with a "cloud-local hybrid architecture." Its core idea is to place the "thinking" process, which requires complex cognition and global planning, within a scalable cloud or local large-scale model, while strictly limiting the "execution" process, which involves users' original sensitive data, to a secure environment on the user's device. By defining a logically centralized and physically flexibly deployable task processing hub, the technical architecture and business model are decoupled. Based on this, a complete cross-platform automated execution system is built, integrating a unified message gateway, cross-platform atomic operation capabilities, and transparent billing management for AI resource scheduling.

[0018] To achieve the above objectives, the present invention adopts the following technical solution:

[0019] A cross-platform intelligent task processing method based on a cloud-local hybrid architecture includes the following steps:

[0020] S1: The first user terminal receives the user's task creation instruction through the unified messaging gateway. The instruction is in natural language form, including but not limited to multimodal input such as voice, text, and screenshot annotations. The unified messaging gateway connects to multiple external instant messaging platforms and node devices, and can be extended to support more chat channels through a plug-in mechanism.

[0021] S2: Send the instruction to a task processing center;

[0022] S3: The task processing center performs semantic understanding and task decomposition on the instruction. Based on the user's preset model usage strategy, it selects to call the local large model or the cloud large model service to complete the understanding and decomposition. If the cloud large model is called, the cost is estimated and user authorization is obtained (or it is automatically executed according to the preset no-confirmation threshold), and a task plan containing at least one platform-independent atomic operation description is generated.

[0023] S4: The task processing center distributes the task to one or more heterogeneous target execution terminals according to the task plan and the capability status of each execution terminal;

[0024] S5: Each target execution terminal, in its local security environment, converts the received platform-independent atomic operation description into a specific operation executable by the local operating system and executes it.

[0025] S6: Summarize the execution status and results of each terminal, feed them back to the task processing center or user, and push the results to the original chat channel used by the user through the unified message gateway.

[0026] A cross-platform intelligent task processing system based on a cloud-local hybrid architecture, used to implement the above method, characterized in that it includes:

[0027] ① Task processing hub, including instruction parsing module (including model routing unit), task planning module, device management module, scheduling and communication module, and billing management module;

[0028] ② Multiple heterogeneous execution terminals, each terminal including a cross-platform atomic operation adapter library, a secure execution engine, a local communication module, and a privacy filter;

[0029] ③ Unified Message Gateway, running as a daemon process, provides a WebSocket interface to connect multiple external instant messaging platforms and node devices, and performs message routing, protocol conversion, session management and media processing;

[0030] ④ Optional enhancement modules, including at least one of the following: real-time delivery preview module, cross-session intelligent memory module, private skill repository, enterprise business system integration module, multi-agent parallel unit, MCP protocol adapter, role adaptation unit, and business data intelligent insight unit.

[0031] Preferably, the physical form of the task processing hub is any one or a combination of the following: software service deployed on a remote cloud server, software running on a dedicated server within the user's local area network, software integrated as a software module in the first user terminal or any target execution terminal, or software service deployed on an edge computing node.

[0032] Preferably, the unified messaging gateway is implemented as a long-running daemon process, with at least one instance running on each host, and is connected to by clients and node devices via a WebSocket interface; it has a built-in multi-protocol adapter, supports at least one third-party instant messaging platform, and provides a plugin mechanism to extend more channels.

[0033] Preferably, the node device includes iOS node, Android node, macOS node and headless device node. When a node connects, it declares its role and capabilities. After pairing and approval, it obtains a device token and can expose at least one of the following capabilities: canvas operation, camera shooting, screen recording, location acquisition, and system command execution.

[0034] Preferably, the local large model includes a local offline model installed via one-click deployment and a locally running web-based large model; the one-click deployment automatically detects hardware configuration, downloads the appropriate model file from the server, and completes the configuration; the local web-based large model accesses the service through automatic discovery of local service interfaces.

[0035] Preferably, the cloud-based large model includes domestically produced large models and mainstream international large models.

[0036] Preferably, in the semantic understanding and task decomposition step, the task processing center selects to call a local large model or a cloud large model service to complete the understanding and decomposition according to the user's preset model usage strategy; the model usage strategy includes at least one of always using a local model, always using a cloud model, automatically selecting according to task complexity, automatically selecting according to cost, and querying each time.

[0037] Preferably, when the target execution terminal performs operations in a local secure environment, operations deemed high-risk are isolated and pre-performed using an operation sandbox, and execution can only continue after the user is requested to confirm.

[0038] Preferably, the first user terminal further includes a privacy filter for locally desensitizing the instruction text before sending the instruction to the task processing center.

[0039] Preferably, when distributing tasks, the task processing center makes decisions based on a dynamically maintained device capability map; the device capability map is dynamically maintained through mechanisms such as device registration, heartbeat reporting, capability change updates, and offline marking, recording the device type, operating system, available atomic operation capabilities, and current status of each execution terminal.

[0040] Preferably, the communication between the task processing center and each target execution terminal adopts at least one encryption method including TLS, end-to-end encryption, and session-level encryption, and the instruction text is locally desensitized.

[0041] Preferably, the process of calling the cloud-based large model also includes a transparent billing management process: the user configures the model usage strategy for the first time; whether user confirmation is required is determined based on the confirmation-free threshold; after the call, the actual usage is measured and billed in real time, and a detailed bill for related business operations is generated; budget control and strategy switching are supported.

[0042] The beneficial effects of this invention are as follows:

[0043] ① Ultimate privacy and security: Sensitive user data does not need to leave the local device. Through multiple layers of protection such as data anonymization, end-to-end encryption, and sandbox isolation, the security of data is ensured throughout its entire lifecycle.

[0044] ② Powerful intelligence and flexible architecture: It fully utilizes the capabilities of large cloud models while supporting offline operation of local models; the task processing hub can be flexibly deployed as cloud SaaS, private aPaaS, standalone version or edge node, and the technical solution has strong universality.

[0045] ③ True cross-platform collaboration: Through platform-independent atomic operation descriptions, unified scheduling of heterogeneous devices such as Windows, macOS, Linux, Android, iOS, and HarmonyOS is achieved.

[0046] ④ Transparent and controllable costs: Integrated transparent billing management supports cost estimation, real-time billing, budget control and policy switching, achieving ultimate cost transparency and intelligent optimization.

[0047] ⑤ Highly secure and controllable: Mechanisms such as operation sandbox, permission verification, operation preview, user confirmation, and operation rollback ensure the safety and reliability of the automation process.

[0048] ⑥ Seamless interactive experience: Supports integration with built-in chat tools and mainstream third-party chat tools, and achieves unified access and session continuity across multiple channels through a unified message gateway.

[0049] ⑦ Unified access and extended ecosystem: The unified message gateway provides advanced functions such as multi-protocol adaptation, node device access, embedded intelligent agents, hook extension, and scheduled tasks; the atomic operation adapter library realizes "describe once, execute everywhere".

[0050] ⑧ Decoupling of core architecture and deployment form: It can be flexibly deployed in various forms (such as cloud SaaS, private aPaaS, standalone version or edge node), effectively preventing plagiarism by changing the deployment form, and forming a three-dimensional protection network through multi-layered claims. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention, showing the overall relationship between the user domain, unified message gateway, task processing hub, heterogeneous execution terminal cluster and AI model service provider.

[0052] Figure 2 This is a flowchart of the core method of the present invention, showing the complete process from user input instructions to final feedback.

[0053] Figure 3 This diagram illustrates the internal logic modules of the task processing hub, showcasing the interaction relationships between modules such as instruction parsing, task planning, device management, scheduling communication, and billing management.

[0054] Figure 4This diagram illustrates the internal structure and workflow of a heterogeneous execution terminal, including the structure and workflow of the local communication module, privacy filter, secure execution engine, and cross-platform atomic operation adapter library.

[0055] Figure 5 This diagram illustrates the mapping between platform-independent atomic operations and platform-specific adapters, showing how a unified description is mapped to various platform adapters through an adapter factory.

[0056] Figure 6 This is a schematic diagram of the cross-platform atomic operation adapter library architecture, showing the class diagram relationship between the abstract interface, platform implementation, and adapter factory.

[0057] Figure 7 This diagram illustrates the example structure of a unified atomic operation description object. It showcases the core methods of each adapter: Execute (executes atomic operations), Executable? (performs a safety pre-check based on the operation description), Rollback (performs an operation rollback based on the operation description), and Call related APIs, achieving "describe once, execute everywhere" operations.

[0058] Figure 8 This is an interaction sequence diagram for Example 1 (cross-device file processing task), showing the message interaction between the mobile phone, cloud hub, computer, and chat tool.

[0059] Figure 9 This is a detailed flowchart of Example 2 (a complex task on a single device, including model selection and billing), which shows the detailed process of model strategy, cost estimation and safe execution in a single device task.

[0060] Figure 10 Deploy a network topology diagram for the cloud-based SaaS hub, demonstrating the network connections and encryption methods between the cloud hub and various terminal devices.

[0061] Figure 11 This is a flowchart for maintaining and updating the equipment capability map, showing the dynamic maintenance process such as equipment registration, heartbeat reporting, capability changes, and offline marking.

[0062] Figure 12 This diagram illustrates the multi-form deployment of the task processing hub, showcasing four deployment forms: cloud-based SaaS, private aPaaS, standalone version, and edge node.

[0063] Figure 13 The accompanying drawings illustrate the core architecture and data flow of this invention. Detailed Implementation

[0064] The following embodiments are merely partial implementations of the present invention, intended to illustrate specific implementations of the present invention in different application scenarios, but are not intended to limit the scope of protection of the present invention. Those skilled in the art can make modifications and extensions based on the teachings of these embodiments without departing from the spirit of the present invention. However, all such modifications and variations should be included within the scope of protection of the present invention. Each embodiment is based on... Figure 1 The system architecture shown is referenced. Figure 2 The core method and process are detailed in the accompanying drawings of each embodiment.

[0065] System operating environment

[0066] This invention's system can be deployed in various computing environments. The task processing hub can run on public cloud servers, enterprise private servers, edge computing nodes, or personal computers; heterogeneous execution terminals encompass various smart devices with the client software of this invention installed, including but not limited to Windows PCs, Mac computers, Linux workstations, Android phones, iPhones / iPads, and HarmonyOS devices. The unified messaging gateway runs as a daemon process, maintaining connections with external instant messaging platforms and node devices. All components communicate securely via the Internet or a local area network.

[0067] Example 1: Cross-device file processing task

[0068] Scenario: A user sends a command via their mobile phone: "Pack all the PPT files that I modified yesterday on my computer desktop and send them to my A chat tool."

[0069] Process (refer to) Figure 1 , Figure 2 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 ):

[0070] 1. Command Reception and Upload: The built-in chat tool on the user's mobile phone (the first terminal) receives voice commands, desensitizes them through a local privacy filter, and then sends them to the task processing center through a unified message gateway.

[0071] 2. Model Selection and Analysis: The central processing unit assesses task complexity based on preset strategies (such as "automatic selection"), calls cloud-based large model analysis commands, and identifies operations requiring "computer desktop," "yesterday's modifications," "PPT file," "compression and packaging," and "sending via chat tool A." The billing management module estimates the cost (e.g., 0.02 yuan) and automatically confirms it (if it's below the no-confirmation threshold).

[0072] 3. Planning and Distribution: The task planning module queries the equipment capability map ( Figure 10The system detected that the user had a Windows computer online that supported file operations, and the phone supported sending messages via chat tool A. A task schedule was then generated.

[0073] (1) Subtask 1 (Computer): Locate and package the files;

[0074] (2) Subtask 2 (Computer → Mobile Phone): Transfer the compressed package;

[0075] (3) Subtask 3 (Mobile): Send to chat tool.

[0076] 4. Safe execution (refer to) Figure 4 ):

[0077] (1) The central task sends subtask 1 to the computer. The computer performs file search and compression through the Windows adapter. The operation is recorded in the sandbox (no high risk, so no user confirmation is required).

[0078] (2) The central coordinator establishes an end-to-end encrypted P2P channel between the computer and mobile phone to transmit compressed packets (such as...). Figure 7 (As shown).

[0079] (3) The central hub sends subtask 3 to the mobile phone, and the mobile phone sends the file by calling the chat tool through the adapter.

[0080] 5. Result feedback: Each terminal sends back the execution result, which is then aggregated by the central hub and pushed to the mobile phone through the gateway to complete the notification.

[0081] Key technical features: cross-platform collaboration, end-to-end encryption, atomic operation adapter, and unified messaging gateway.

[0082] Example 2: Complex Tasks on a Single Device (Including Model Selection and Billing)

[0083] Scenario: A user enters on a Mac: "Organize all images in the Downloads folder into the 'Photos' folder by shooting date." The user's preset model strategy is "Smart Auto," with a confirmation-free threshold of 0.05 yuan.

[0084] Process (refer to) Figure 1 , Figure 2 , Figure 9 , Figure 13 ):

[0085] 1. Command Reception: The local client receives the command, desensitizes it using a privacy filter, and then sends it to the task processing hub (in this embodiment, the hub operates as a local module, i.e., a standalone deployment, such as...). Figure 11 (As shown).

[0086] 2. Model Selection: The complexity of the central assessment task is simple, so we choose to call the local large model for parsing, which is free of charge.

[0087] 3. Analysis and Planning: After analyzing the local model, a task plan is generated: Scan the download folder → Extract the EXIF ​​shooting date → Create a "Year-Month" subfolder → Move the images.

[0088] 4. Safe execution (refer to) Figure 4 ):

[0089] (1) The secure execution engine performs operations through the macOS adapter, displaying an operation preview (such as "Move 25 pictures") before moving files, and executing the operation after user confirmation.

[0090] (2) If the user's policy is "always in the cloud", then the billing process will begin: the estimated cost is 0.02 yuan (below the threshold), the cloud model will be automatically called, the actual usage will be recorded and the bill details will be generated.

[0091] 5. Results Feedback: The compiled results are displayed through the chat window.

[0092] Key technical features: dynamic model selection, threshold without confirmation, local priority execution, and operation preview.

[0093] Example 3: Multi-cloud deployment model

[0094] This embodiment demonstrates various physical deployment forms of the task processing hub, reflecting the decoupling between the core architecture and deployment form (see reference). Figure 11 and Figure 9 ).

[0095] A large enterprise deployed multiple central instances based on its business needs:

[0096] ① Cloud-based SaaS hub ( Figure 10 Deployed on the public cloud, this is a basic automated service available to all employees. Employees access it through a unified message gateway, which automatically routes the data to the cloud hub. The cloud hub then uses a large cloud-based model for semantic understanding and distributes tasks to employees' local terminals.

[0097] ② Privatization of the aPaaS hub ( Figure 12 Deployed within the enterprise intranet, this system processes sensitive data such as financial reimbursements. Data is entirely localized, allowing access to local models or cloud-based models via the enterprise's egress point. Employees access the gateway through the intranet.

[0098] ③ Standalone version of the central hub ( Figure 12 Integrated into personal computers, suitable for business trips or offline scenarios. No network connection is required; it calls local models (such as LLaMA deployed in Ollam), and all operations are performed locally.

[0099] ④ Edge node hub ( Figure 12Deployed in the factory workshop, it connects to IoT devices. Workers issue commands via tablets: "Check the temperature of device A; if it exceeds the threshold, trigger an alarm." The edge hub calls the local model for real-time processing, and the data does not leave the internal network, meeting low latency requirements.

[0100] All central instances implement the same core functional modules. Figure 3 This ensures a consistent user experience. Administrators can monitor the status of each central hub through a unified console.

[0101] Example 4: Dynamic Update of Equipment Capability Map

[0102] This embodiment illustrates the process of maintaining and updating the equipment capability map (refer to...). Figure 11 ).

[0103] 1. Initial Registration: When a new Windows computer launches the client for the first time, it sends a registration request (device type, operating system, hardware configuration, application list, etc.) to the central system. After verification by the central system, the information is stored in the graph and marked as "online".

[0104] 2. Heartbeat Maintenance: The device sends a heartbeat packet (containing load information) every 30 seconds, and the central system updates the "Last Active Time" and "Current Load" in the graph. If no heartbeat is received for 3 consecutive times, it is marked as "Offline".

[0105] 3. Capability Change: When a user installs Photoshop on a Mac computer, the client sends an incremental update, and the central verification process updates the graph.

[0106] 4. Task planning application: When a user initiates the command "Open image with Photoshop", the central system queries the graph and finds that a certain Mac is online and has Photoshop installed, and then distributes the task to that device.

[0107] 5. Clean up expired devices: Devices that have not been online for more than 30 days will be automatically moved to the archive table.

[0108] Example 5: One-click deployment and invocation of local models

[0109] 1. Scenario: The user wants to use the AI ​​offline and clicks "Install local model" in the client.

[0110] 2. Process (refer to) Figure 1 , Figure 2 , Figure 13 ):

[0111] (1) The system automatically detects hardware (CPU, memory, video memory) and recommends a suitable model.

[0112] (2) After user confirmation, the model file is downloaded from CDN (supports breakpoint resume) and automatically loaded into the inference engine (e.g. llama.cpp).

[0113] (3) When the user inputs “summarize this PDF”, the system calls the local model to complete the understanding, and there is no network transmission throughout the process.

[0114] (4) You can switch to other installed local models with one click in the model management interface.

[0115] 3. Key technical features: hardware self-adaptation, one-click deployment, and offline operation.

[0116] Example 6: Cloud-based switching of domestically produced large-scale models

[0117] 1. Scenario: An enterprise uses multiple large models for A / B testing at the same time, and the administrator assigns default models to different departments.

[0118] 2. Process (refer to) Figure 1 , Figure 2 , Figure 13 ):

[0119] (1) Employees can view the current model in the chat using the / model command.

[0120] (2) Send / model Switch to the current model. Subsequent system calls will take effect immediately.

[0121] (3) The usage and cost of each call are recorded separately and included in the department's billing (e.g., in the billing management module). Figure 3 , Figure 9 (As shown).

[0122] (4) Administrators can view the usage statistics and cost comparisons of each model in the background.

[0123] 3. Key technical features: multi-model integration, real-time switching, and revenue sharing management.

[0124] Example 7: Communication Encryption and Data Desensitization

[0125] This embodiment illustrates the secure communication mechanism between the task processing center and each terminal, as well as the local desensitization process.

[0126] like Figure 1 As shown, all communication links employ multiple encryption methods. Control signaling between the central hub and terminals uses TLS 1.3 encryption; large file transfers utilize end-to-end encryption (such as DTLS) via a direct P2P channel. Each session uses a temporarily generated key to ensure forward security.

[0127] Before the instruction is uploaded, the privacy filter on the first user terminal ( Figure 4The command text is locally anonymized to remove sensitive information such as names and ID numbers, retaining only the semantic skeleton. For example, if a user says "Send Zhang San's resume to General Manager Li," the anonymized version becomes "Send [PERSON]'s resume to [PERSON]," effectively protecting privacy.

[0128] Example 8: Safe Execution of High-Risk Operations

[0129] This embodiment illustrates the mechanism by which the target execution terminal performs high-risk operations in a local secure environment, such as... Figure 4 The secure execution engine in [the system].

[0130] 1. Suppose the user instructs: "Delete all temporary files in the download folder." The security execution engine determines this operation is high-risk and sends it to the operation sandbox. The sandbox rehearses the deletion operation, generating an operation preview: "The following 3 files will be deleted: temp1.tmp, temp2.tmp, and temp3.tmp." The preview is sent to the user for confirmation. The user discovers an error in the file list and cancels the execution.

[0131] 2. If the user confirms, the sandbox performs a real deletion and records a snapshot before the operation. If the user discovers that important files have been accidentally deleted after the operation, they can trigger the adapter's rollback method using the rollback button to restore the files from the snapshot.

[0132] Example 9: Collaborative Tasks Between Mobile Phones and Computers

[0133] 1. Scenario: After taking a photo with their mobile phone, the user says, "Transfer the three most recently taken photos from my album to my computer desktop and open the first one in Photoshop."

[0134] 2. Process (refer to) Figure 1 , Figure 2 , Figure 5 , Figure 6 , Figure 7 , Figure 13 ):

[0135] (1) The mobile phone receives the instruction, automatically associates the currently captured photo, and uploads it to the central hub.

[0136] (2) Central parsing instructions to generate task plan: ① Mobile phone to acquire photos; ② Mobile phone to computer transfer; ③ Computer saves and launches Photoshop.

[0137] (3) The mobile phone calls the PhotoKit API through the iOS adapter to obtain photos.

[0138] (4) The central coordination unit establishes an end-to-end encrypted P2P channel between the mobile phone and the computer to transmit photos.

[0139] (5) The computer saves the photo to the desktop via the Windows / macOS adapter and launches Photoshop to open the first photo via the COM interface.

[0140] (6) Results are sent to mobile phone.

[0141] 3. Key technical points: node device integration, cross-platform file transfer, and application automation.

[0142] Example 10: Initiating a task via a third-party chat tool

[0143] 1. Scenario: A company binds its various chat tool accounts to this system. A sales manager sends a message on chat tool A: "Check Zhang San's customer's purchase records from last quarter."

[0144] 2. Process (refer to) Figure 1 , Figure 2 , Figure 13 ):

[0145] (1) The unified message gateway receives messages from chat tool A, identifies the sender (employee associated with mobile phone number) through the message routing engine, and forwards them to the task processing center.

[0146] (2) Central parsing instruction calls the computer where the CRM system is located to perform data query and generate summary report.

[0147] (3) The result is returned to the sales manager's chat tool session via the gateway.

[0148] 3. Key technical features: unified messaging gateway, multi-channel access, and enterprise identity association.

[0149] Example 11: Multi-channel session continuity

[0150] 1. Scenario: A user has a conversation with an AI agent in the morning using chat tool A, and continues the same topic on their phone in the afternoon using chat tool B.

[0151] 2. Process (refer to) Figure 1 , Figure 2 , Figure 13 ):

[0152] (1) The session management module of the unified messaging gateway associates sessions from different channels through a unified user identifier (such as corporate email).

[0153] (2) When a user says “Continue the morning analysis” in chat tool B, the gateway routes to the same agent instance and loads the morning session history.

[0154] (3) The agent continues processing based on the complete context, and the result is returned to the B chat tool.

[0155] 3. Key technical points: cross-channel session management and context persistence.

[0156] Example 12: Operating the Sandbox and Rollback

[0157] 1. Scenario: The user accidentally allows the AI ​​to delete an important folder.

[0158] 2. Process (refer to) Figure 4 ):

[0159] (1) When an agent generates or deletes an operation, the security execution engine determines it to be high-risk and sends it to the operation sandbox.

[0160] (2) Sandbox pre-deletion, generating an operation preview: "The following 3 files will be deleted:..." and requesting user confirmation.

[0161] (3) If the user discovers a mistake, the execution is canceled; if the user confirms, the sandbox performs a real deletion and records a snapshot, which can then be restored with one click via the "rollback" button.

[0162] 3. Key technical features: security sandbox, operation preview, user confirmation, and rollback mechanism.

[0163] Example 13: Hybrid Model Routing

[0164] 1. Scenario: The user sets the strategy to "use local models for simple tasks and domestic cloud models for complex tasks, with cost as the priority".

[0165] 2. Process (refer to) Figure 2 , Figure 3 ):

[0166] (1) When the user inputs “organize the desktop”, the system evaluates it as a simple task and calls the local model.

[0167] (2) When the user inputs “analyze financial statement risks”, the system assesses it as a complex task, selects the lowest-cost domestic model from the cloud model pool, displays the estimated cost, and then executes it automatically (below the confirmation threshold).

[0168] (3) If the cost exceeds the threshold, user confirmation is required.

[0169] 3. Key technical points: task complexity assessment, cost-priority routing, and user confirmation mechanism.

[0170] Example 14: Heartbeat Mechanism for Monitoring the Inbox

[0171] 1. Scenario: Users set their email to check every 30 minutes and send a notification if there are important emails.

[0172] 2. Process (refer to) Figure 2 , Figure 11 ):

[0173] (1) Heartbeat timer triggers the running agent in the main session.

[0174] (2) The agent calls the email tool to check unread emails and filters out emails with the subject line “urgent” or “important”.

[0175] (3) If an email that meets the criteria is found, send a message to the user's chat window: "You have an important email:...".

[0176] (4) If there are no important emails, the agent returns HEARTBEAT_OK and remains silent.

[0177] 3. Key technical points: heartbeat mechanism, periodic sensing, and proactive notification.

[0178] Example 15: Condition-Triggered Task

[0179] 1. Scenario: User-defined rule: "When an email with the subject 'Contract' is received, automatically download the attachment to the 'Contract' folder and send a notification via chat tool A."

[0180] 2. Process (refer to) Figure 2 , Figure 3 , Figure 5 , Figure 6 , Figure 7 ):

[0181] (1) Rule configuration: The user creates a condition-triggered task in the client, specifies the trigger as an email event, the condition as "subject contains 'contract'", and the actions as "download attachment" and "send A chat tool message".

[0182] (2) Continuous monitoring: The client monitors new emails in the background via the email API. When the conditions are met, the task is automatically triggered.

[0183] (3) Task execution:

[0184] ① The email adapter retrieves attachments and saves them to the local sandbox.

[0185] ② The file operation adapter moves the attachments to the "Contracts" folder.

[0186] ③ A chat tool adapter sends a message via API: "Contract attachment downloaded: filename".

[0187] ④ Feedback: The task execution results can be displayed in the chat window.

[0188] 3. Key technical points: conditional triggering, automated response, and multi-adaptor collaboration.

[0189] Example 16: Data Cleaning and Transformation

[0190] 1. Scenario: User instruction: "Convert all PDF invoices in the invoice folder into Excel, extract the amount and date, and summarize them into a single spreadsheet."

[0191] 2. Process (refer to) Figure 2 , Figure 4 ):

[0192] (1) Instruction parsing: The system understands that PDFs need to be processed in batches to extract structured data.

[0193] (2) Task planning: Generate operation sequence: Traverse all PDFs → Call OCR / PDF parsing library to extract amount and date → Write to Excel spreadsheet → Summarize data.

[0194] (3) Execution: Each adapter calls the corresponding library to complete the operation and returns the result to the user.

[0195] 3. Key technical points: batch data processing, OCR integration, and result summarization.

Claims

1. A cross-platform intelligent task processing method based on a cloud-local hybrid architecture, characterized in that, include: 1) The first user terminal receives the user's task creation instruction through the unified messaging gateway; 2) Send the instruction to the task processing center; 3) The task processing center performs semantic understanding and task decomposition on the instructions to generate a task plan containing at least one platform-independent atomic operation description; 4) The task processing center distributes the task to one or more heterogeneous target execution terminals according to the task plan and the capability status of each execution terminal; 5) Each target execution terminal, in its local secure environment, converts the received platform-independent atomic operation description into a specific operation executable by the local operating system and executes it. 6) Summarize the execution results and feed them back through the unified messaging gateway.

2. The method according to claim 1, characterized in that, The physical form of the task processing hub can be any of the following: software service deployed on a remote cloud server, software running on a dedicated server within the user's local area network, software integrated as a software module into the first user terminal or any target execution terminal, or software service deployed on an edge computing node.

3. The method according to claim 1, characterized in that, In the semantic understanding and task decomposition steps, the task processing center selects to call a local large model or a cloud-based large model service to complete the understanding and decomposition based on the user's preset model usage strategy.

4. The method according to claim 1, characterized in that, The unified messaging gateway provides a WebSocket interface for clients and node devices to connect, and supports message sending and receiving from at least one third-party instant messaging platform.

5. The method according to claim 1, characterized in that, The step of "converting the received platform-independent atomic operation description into a specific operation executable by the native operating system" is accomplished by a cross-platform atomic operation adapter library, which contains platform-specific adapters for different operating systems.

6. The method according to claim 1, characterized in that, When the target execution terminal performs operations in a local secure environment, operations deemed high-risk are isolated and rehearsed using an operation sandbox, and the operation can only continue after the user requests confirmation.

7. The method according to claim 1, characterized in that, When distributing tasks, the task processing center makes decisions based on a dynamically maintained device capability map, which records the device type, operating system, available atomic operation capabilities, and current status of each execution terminal.

8. The method according to claim 1, characterized in that, The communication between the task processing center and each target execution terminal adopts at least one encryption method, including TLS and end-to-end encryption.

9. The method according to claim 1, characterized in that, Before step S2, the method further includes: the first user terminal performs local desensitization processing on the instruction text, removes personal identification information, and then sends the desensitized instruction to the task processing center.

10. A cross-platform intelligent task processing system based on a cloud-local hybrid architecture, characterized in that, include: 1) A task processing hub, including an instruction parsing module, a task planning module, a device management module, a scheduling and communication module, and a billing management module; 2) Multiple heterogeneous execution terminals, each terminal including a module for translating platform-independent atomic operation descriptions into specific operations executable by the native operating system, a secure execution engine, a local communication module, and a privacy filter; 3) A unified messaging gateway connects multiple external instant messaging platforms and node devices to perform message routing, protocol conversion, session management, and media processing.

11. The system according to claim 10, characterized in that, The instruction parsing module includes a model routing unit, which is used to select and call a local large model or a cloud-based large model according to the user's preset strategy.

12. The system according to claim 10, characterized in that, The secure execution engine includes a permission verification unit and an operation sandbox unit. The operation sandbox unit provides an isolated execution environment for high-risk operations and supports operation rollback.

13. The system according to claim 10, characterized in that, The unified messaging gateway runs as a daemon process, with at least one instance running on each host. It provides a WebSocket interface for clients and node devices to connect and includes a multi-protocol adapter layer, a message routing engine, a session management module, and a node management module.

14. A computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as claimed in any one of claims 1 to 9.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 9.

Images