Intelligent agent tool chain dynamic reorganization method and device with autonomous repair capability, electronic equipment and medium

Abstract

This application provides a method, apparatus, electronic device, and medium for dynamic reorganization of an intelligent agent toolchain with autonomous repair capabilities. The method includes: acquiring multimodal temporal execution data corresponding to the current tool call chain of the target intelligent agent; inputting the multimodal temporal execution data into a preset perception network to obtain anomaly node identifiers representing execution anomalies; determining the corresponding failed tool nodes in a global dependency topology graph based on the anomaly node identifiers; isolating the failed tool nodes to obtain a link to be repaired with a topological breakpoint; extracting the breakpoint output features corresponding to the first tool node; obtaining expected state features to generate a feature compensation vector; matching the feature compensation vector with the capability description vectors of candidate replacement nodes to determine the target replacement node; connecting the target replacement node to the topological breakpoint and performing feature alignment processing on the input interface to generate an adaptive toolchain input to the execution unit for processing.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, and more specifically, to a method, apparatus, electronic device, and computer-readable storage medium for dynamic reorganization of an intelligent agent toolchain with autonomous repair capabilities. Background Technology

[0002] With the rapid development of multimodal large-scale models and multi-agent evolution technologies, agents are increasingly relying on dynamic toolchains composed of multiple external tool components to work collaboratively when handling complex tasks. In this scenario of continuous evolution of multiple agents and deep invocation of toolchains, maintaining the continuity of toolchain execution links and the stability of parameter flow is of great significance for ensuring the overall efficiency of agent task execution.

[0003] In existing intelligent agent toolchain exception handling schemes, a static exception capture mechanism based on a rule engine is typically adopted. This mechanism first presets standard error log listeners at the call interfaces of each tool component to collect status code feedback during tool execution; when an exception status code is detected, the system directly triggers external retry logic or performs a complete replacement according to a pre-hard-coded static backup link; finally, the replaced static backup link is re-deployed to the intelligent agent to start executing the task from scratch.

[0004] However, this static capture and replacement scheme has significant technical drawbacks. Because it relies solely on a single status code for anomaly detection, it lacks comprehensive awareness of multi-link operational data, resulting in limitations in locating tool node failures. Furthermore, the hard-coded static backup link replacement method cannot accurately and dynamically repair and match interface features for specific failure points in complex global topologies. This leads to severe loss of context state during toolchain reassembly, making it difficult to meet the needs of intelligent agents for adaptive recovery and dynamic reassembly of toolchain breakpoints in complex scenarios. Summary of the Invention

[0005] This application provides a method, apparatus, electronic device, and computer-readable storage medium for dynamic reorganization of an intelligent agent toolchain with autonomous repair capabilities, so as to at least alleviate the above-mentioned technical problems.

[0006] A method for dynamic reorganization of an intelligent agent toolchain with self-repair capability, comprising: Obtain multimodal time-series runtime data corresponding to the tool call chain of the target agent in the current execution cycle; Multimodal time-series runtime data is input into a pre-defined perception network to obtain abnormal node identifiers that represent execution anomalies. Based on the abnormal node identifier, the corresponding failed tool node is determined in the preset global dependency topology graph. The global dependency topology graph contains multiple tool nodes and the directed connection relationships between the nodes. The failed tool nodes are isolated to obtain the links to be repaired with topological breaks; Extract the breakpoint output features corresponding to the first tool node located upstream of the topological breakpoint in the link to be repaired; Obtain the expected state features corresponding to the current task being performed by the target agent; Based on the breakpoint output features and the expected state features, a feature compensation vector is generated. Obtain the capability description vectors corresponding to each of the multiple candidate filler nodes in the preset component library; The feature compensation vector is matched with the capability description vectors corresponding to each of the multiple candidate replacement nodes to determine the target replacement node. Connect the target filler node to the topology breakpoint and perform feature alignment processing on the input interface of the target filler node to generate an adaptive toolchain.

[0007] Optionally, obtain the multimodal time-series runtime data corresponding to the tool call chain of the target agent in the current execution cycle, including: Obtain the instruction flow time consumption characteristics of each tool node in the tool call chain; Obtain the resource fluctuation characteristics of each tool node during the execution process; The instruction flow time consumption characteristics are time-aligned to obtain aligned time consumption characteristics; The alignment time consumption characteristics and resource fluctuation characteristics are spliced ​​together to generate multimodal time series operation data.

[0008] Optionally, multimodal time-series runtime data is input into a preset perception network to obtain anomaly node identifiers representing execution anomalies, including: Multimodal temporal data is input into the feature mapping layer of the perceptron network to obtain the feature distribution matrix; Calculate the deviation of each feature element in the feature distribution matrix relative to a preset reference element; Extract target feature elements whose deviation values ​​are greater than a preset deviation threshold from the feature distribution matrix; The index information carried by the target feature elements is analyzed, and the identifier corresponding to the index information is determined as the abnormal node identifier.

[0009] Optionally, the failed tool node is isolated to obtain a link to be repaired with a topological break, including: Delete the first input connection edge in the global dependency topology graph that points to the failed tool node; Delete the second output connection edge originating from the failed tool node in the global dependency topology graph; The topology obtained after deleting the first input connection edge and the second output connection edge is identified as the link to be repaired. The starting node corresponding to the first input connection edge is determined as the first tool node, and the position where the first tool node was originally connected to the failed tool node is determined as the topological break.

[0010] Optionally, a feature compensation vector is generated based on the breakpoint output features and the expected state features, including: The difference between the breakpoint output features and the expected state features is calculated to obtain the difference feature vector. The difference feature vector is input into a preset feature mapping network to obtain the mapped feature compensation vector.

[0011] Optionally, the feature compensation vector is matched with the capability description vectors corresponding to each of the multiple candidate replacement nodes to determine the target replacement node, including: For each candidate replacement node among multiple candidate replacement nodes, perform an inner product operation on the capability description vector and feature compensation vector corresponding to the candidate replacement node to obtain the matching value corresponding to the candidate replacement node. Based on the matching values ​​corresponding to each of the multiple candidate filler nodes, the multiple candidate filler nodes are sorted. The candidate replacement node that ranks first in terms of matching value is determined as the target replacement node.

[0012] Optionally, feature alignment processing is performed on the input interface of the target filler node to generate an adaptive toolchain, including: Extract the input data format corresponding to the target padding node; Extract the output data format corresponding to the breakpoint output features; Determine whether the input data format and the output data format are consistent; If the input data format and the output data format are inconsistent, a preset data conversion middleware node is inserted between the first tool node and the target padding node; The breakpoint output features are format-converted by the data conversion middleware node to obtain alignment features that are consistent with the input data format. Based on alignment features, a connection path is established from the first tool node to the target filler node via the data conversion middleware node, generating an adaptive toolchain.

[0013] Optionally, if it is determined that the input data format and the output data format are inconsistent, before inserting a preset data conversion middleware node between the first tool node and the target padding node, the method further includes: If the input data format is determined to be consistent with the output data format, a direct connection path is established from the first tool node to the target padding node; Perform connectivity checks on topologies containing directly connected paths; After the connectivity verification process passes, the topology containing direct connection paths is identified as the adaptive toolchain.

[0014] Optionally, after inputting the adaptive toolchain into the execution unit of the target agent for further processing, the method further includes: Obtain the feedback execution characteristics generated by the execution unit when processing the adaptive toolchain; Calculate the feature similarity between the feedback execution features and the expected state features; If the feature similarity is less than a preset similarity threshold, the feature compensation vector is updated based on the feature similarity. The updated feature compensation vector is associated with the abnormal node identifier; The data after association processing is stored in a pre-defined experience knowledge base.

[0015] A dynamic reconfiguration device for an intelligent agent toolchain with autonomous repair capabilities, comprising: The data acquisition module is used to acquire multimodal time-series runtime data of the target intelligent agent corresponding to the tool call chain in the current execution cycle; The identification and perception module is used to input multimodal time-series runtime data into a preset perception network to obtain abnormal node identifiers that represent execution anomalies. The node location module is used to determine the corresponding failed tool node in a preset global dependency topology graph based on the abnormal node identifier. The global dependency topology graph contains multiple tool nodes and the directed connection relationships between the nodes. The isolation processing module is used to isolate the failed tool nodes to obtain the link to be repaired with topological breaks; The feature extraction module is used to extract the breakpoint output features corresponding to the first tool node located upstream of the topological break in the link to be repaired; The state acquisition module is used to acquire the expected state characteristics corresponding to the current task being performed by the target agent; The vector generation module is used to generate feature compensation vectors based on the breakpoint output features and the expected state features. The capability acquisition module is used to acquire the capability description vectors corresponding to multiple candidate filler nodes in the preset component library. The matching processing module is used to match the feature compensation vector with the capability description vectors corresponding to each of the multiple candidate replacement nodes to determine the target replacement node. The link generation module is used to connect the target replacement node to the topology breakpoint and perform feature alignment processing on the input interface of the target replacement node to generate an adaptive toolchain.

[0016] An electronic device includes a processor and a memory, the memory storing at least one instruction, the at least one instruction being loaded and executed by the processor to perform the operations performed by the autonomously repairable intelligent agent toolchain dynamic reorganization method as described above.

[0017] A computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to perform the operations performed by the autonomously repairable intelligent agent toolchain dynamic reorganization method as described above.

[0018] The technical advantages of the technical solution provided in this application are: This application presents a dynamic reorganization method for intelligent agent toolchains with self-repair capabilities. Addressing the limitations of traditional rule-engine-based static anomaly capture mechanisms, such as perception limitations and severe context loss due to hard-coded replacement, this method obtains anomaly node identifiers by inputting multimodal time-series runtime data into a pre-defined perception network. This solves the problem of insensitive anomaly detection caused by traditional solutions relying solely on single status codes. Compared to traditional solutions limited to a single collection mode of interface log data, this application integrates multimodal time-series runtime data, providing richer, higher-dimensional input to the perception network, resulting in a significant improvement in the sensitivity and coverage of anomaly node location.

[0019] This application identifies and isolates failed tool nodes in the global dependency topology graph based on abnormal node identifiers, solving the problem that traditional hard-coded solutions can only replace the entire link when an anomaly occurs. By logically isolating failed nodes and accurately defining topology breaks, this application preserves the usable topology resources of the original call link, thus largely avoiding the waste of early-stage computational resources compared to a complete replacement solution.

[0020] Furthermore, by extracting breakpoint output features and expected state features to generate a feature compensation vector, and matching it with the capability description vector of candidate replacement nodes to determine the target replacement node, this solves the problem that traditional backup links cannot dynamically adapt to specific breakpoints. Compared to traditional statically preset recovery strategies, the dynamic matching mechanism based on vector space in this application can adaptively select the node that best fits the current breakpoint compensation requirements, and the flexibility and task fit of replacement node selection have significant advantages.

[0021] Finally, by connecting the target replacement node to the topology breakpoint and performing feature alignment processing on the input interface, the problem of context state conflicts and data flow interruptions caused by the forced splicing in traditional methods is solved. Through the alignment processing mechanism, this application realizes the format transition between upstream and downstream nodes, enabling the recombined adaptive toolchain to smoothly inherit the historical context state of the topology breakpoint, effectively avoiding execution crashes caused by parameter gaps, thereby ensuring a high level of consistency in the recombination of the agent toolchain and the stability of subsequent task execution. Attached Figure Description

[0022] Figure 1 This application provides an embodiment of a scenario for the dynamic reorganization of an intelligent agent toolchain with autonomous repair capabilities. Figure 2 This application provides an embodiment of a method for dynamic reorganization of an intelligent agent toolchain with self-repair capabilities. Figure 3 This application provides an embodiment of a dynamic reorganization device for an intelligent agent toolchain with autonomous repair capabilities. Figure 4 This is an electronic device according to an embodiment of the present application.

[0023] Figure 5 This is a computer-readable storage medium according to an embodiment of the present application. Detailed Implementation

[0024] like Figure 1 As shown, this is a scenario of dynamic reorganization of an intelligent agent toolchain with self-repair capabilities according to an embodiment of this application; as Figure 2 As shown, this embodiment of the present application provides a method for dynamic reorganization of an intelligent agent toolchain with self-repair capabilities. The method includes the following steps: Obtain multimodal time-series runtime data corresponding to the tool call chain of the target agent in the current execution cycle; The multimodal time-series execution data is input into a preset perception network to obtain anomaly node identifiers that represent execution anomalies. Based on the abnormal node identifier, the corresponding failed tool node is determined in the preset global dependency topology graph, which contains multiple tool nodes and directed connection relationships between nodes. The failed tool node is isolated to obtain a link to be repaired with a topological break. Extract the breakpoint output features corresponding to the first tool node located upstream of the topological breakpoint in the link to be repaired; Obtain the expected state features corresponding to the current task being performed by the target intelligent agent; Based on the breakpoint output features and the expected state features, a feature compensation vector is generated. Obtain the capability description vectors corresponding to each of the multiple candidate filler nodes in the preset component library; The feature compensation vector is matched with the capability description vector corresponding to each of the multiple candidate replacement nodes to determine the target replacement node; The target filler node is connected to the topology breakpoint, and feature alignment processing is performed on the input interface of the target filler node to generate an adaptive toolchain.

[0025] Optionally, obtaining the multimodal time-series runtime data corresponding to the tool call chain of the target agent in the current execution cycle includes: Obtain the instruction flow time consumption characteristics of each tool node in the tool call chain; Obtain the resource fluctuation characteristics of each tool node during the execution process; The instruction flow time consumption characteristics are subjected to timing alignment processing to obtain aligned time consumption characteristics; The alignment time consumption feature and the resource fluctuation feature are concatenated to generate the multimodal time-series running data.

[0026] Preferably, the specific implementation process for obtaining the multimodal temporal execution data corresponding to the tool call chain of the target agent in the current execution cycle is as follows: First, read the agent resource configuration record, prompt word loading record, and tool call execution record corresponding to the target agent in the current execution cycle, and determine the large language processing model, knowledge base, database, toolset, environment awareness node, and physical execution component that the target agent can call in the current execution cycle based on the agent resource configuration record; then, determine the current execution task, tool call order, and result return requirements that the target agent needs to execute in the current execution cycle based on the prompt word loading record; subsequently, organize the large language processing model, knowledge base, database, toolset, environment awareness node, and physical execution component into a tool call chain according to the actual call order relationship based on the tool call execution record. The technical essence of the tool call chain is the ordered tool node connection relationship formed by the target agent in the current execution cycle to complete the current execution task. The tool call chain is not a static tool list, but an execution chain formed by the large language processing model, knowledge base, database, toolset, environment awareness node, and physical execution component actually called by the target agent according to the data input, processing output, and result return relationship.

[0027] Preferably, each tool node in the tool invocation chain is determined based on the type of resource actually invoked by the target agent within the current execution cycle. Specifically, when the target agent invokes a large language processing model to generate task processing text, the invocation instance corresponding to the large language processing model is determined as a tool node; when the target agent invokes a knowledge base to perform an index query, the query instance corresponding to the knowledge base is determined as a tool node; when the target agent invokes a database to perform structured data query, modification, deletion, or update, the data processing instance corresponding to the database is determined as a tool node; when the target agent invokes a toolset to perform date judgment, statement generation, chart generation, data calculation, or data analysis, the corresponding tool instance in the toolset is determined as a tool node; when the target agent invokes an environmental perception node to collect video, image, voice, or status data, the data collection instance corresponding to the environmental perception node is determined as a tool node; when the target agent invokes a physical execution component to perform movement, voice interaction, display interaction, or mechanical operation, the execution instance corresponding to the physical execution component is determined as a tool node. The technical essence of the tool node is that it is a calling unit in the tool call chain that can receive the output data of the previous tool node, perform local processing actions, and output the processing results to the next tool node. The tool node is written into the tool call execution record through the tool node identifier, so that the time consumption characteristics and resource fluctuation characteristics of subsequent instruction flow can be traced back to the corresponding tool node.

[0028] Preferably, when acquiring the instruction flow time consumption characteristics of each tool node in the tool call chain, the call request timestamp, call reception timestamp, processing start timestamp, processing completion timestamp, and result return timestamp corresponding to each tool node are first read. These timestamps are then written into the node time consumption sampling record according to the tool node arrangement order in the tool call chain. Next, based on the node time consumption sampling record, the entry waiting time consumption between the call request and call reception, the node processing time consumption between processing start and processing completion, and the exit return time consumption between processing completion and result return are calculated. Subsequently, the entry waiting time consumption, node processing time consumption, and exit return time consumption are combined according to the tool node identifier to form the instruction flow time consumption characteristics. The technical essence of these instruction flow time consumption characteristics is a data feature used to describe the time consumed by each tool node in receiving instructions, performing processing, and returning results within the current execution cycle. These instruction flow time consumption characteristics originate from the timestamp data in the tool call execution record and continue to participate in subsequent timing alignment processing.

[0029] Preferably, the instruction flow time consumption feature is not a single total time record, but rather segmented time consumption data bound together with the tool node identifier, call order identifier, and time consumption segment type. The tool node identifier indicates which tool node in the large language processing model, knowledge base, database, toolset, environment-aware node, or physical execution component the instruction flow time consumption feature corresponds to; the call order identifier indicates the sequential position of the corresponding tool node in the tool call chain; and the time consumption segment type distinguishes between entry waiting time consumption, node processing time consumption, and exit feedback time consumption. Through this binding method, the instruction flow time consumption feature can reflect the temporal running status of each tool node in the tool call chain, enabling subsequent alignment of time consumption features to retain the call order relationship between tool nodes, rather than simply expressing the overall time consumption of the current execution cycle.

[0030] Preferably, when acquiring the resource fluctuation characteristics of each tool node during execution, the resource usage sampling record of the corresponding tool node during execution is first read according to the tool node identifier, and the computational resource usage value, memory resource usage value, data access wait value, interface response status value, and queue backlog status value are extracted from the resource usage sampling record; then, according to the tool node identifier and the call order identifier, the computational resource usage value, memory resource usage value, data access wait value, interface response status value, and queue backlog status value are written into the node resource sampling record; subsequently, the node resource sampling record is sorted by time segmentation so that each tool node forms a corresponding resource fluctuation characteristic within the current execution cycle. The technical essence of the resource fluctuation characteristic is to describe the data characteristics of the resource usage intensity, data access wait status, and interface response status of the tool node during execution as they change over time. The resource fluctuation characteristic originates from the resource usage sampling record during the execution of the tool node and continues to participate in the generation of multimodal time-series running data together with the alignment time consumption characteristic.

[0031] Preferably, for different types of tool nodes, the resource fluctuation features adopt sampling content adapted to the running attributes of the tool nodes. For tool nodes corresponding to large language processing models, resource fluctuation features include generation waiting state values, text generation persistence state values, and result return state values; for tool nodes corresponding to knowledge bases, resource fluctuation features include index query waiting state values, similar content return state values, and knowledge base access state values; for tool nodes corresponding to databases, resource fluctuation features include database connection state values, query execution waiting state values, and processing result return state values; for tool nodes corresponding to toolsets, resource fluctuation features include tool startup state values, tool execution state values, and tool result output state values; for tool nodes corresponding to environmental perception nodes, resource fluctuation features include video stream access state values, frame extraction processing state values, voice access state values, and perception result output state values; for tool nodes corresponding to physical execution components, resource fluctuation features include execution queue state values, execution action persistence state values, and execution feedback state values. The generated waiting status value, text generation continuous status value, result return status value, index query waiting status value, similar content return status value, knowledge base access status value, database connection status value, query execution waiting status value, processing result return status value, tool startup status value, tool execution status value, tool result output status value, video stream access status value, frame extraction processing status value, voice access status value, perception result output status value, execution queue status value, execution action continuous status value, and execution feedback status value are all bound to the tool node identifier and sampling timestamp, so that resource fluctuation characteristics can establish a corresponding relationship with instruction flow time consumption characteristics in the same tool call chain.

[0032] Preferably, when performing time-series alignment processing on the instruction flow time consumption characteristics to obtain aligned time consumption characteristics, the call order identifier in the tool call chain is first read, and the sequential relationship between each tool node is determined according to the call order identifier; then, the call request timestamp, processing start timestamp, processing completion timestamp, and result feedback timestamp corresponding to each tool node are read, and the start timestamp of the current execution cycle is used as a unified time base to convert the entry waiting time consumption, node processing time consumption, and exit feedback time consumption of each tool node into time-series position data relative to the unified time base; subsequently, the time-series position data are arranged according to the tool node identifier and the call order identifier to form aligned time consumption characteristics. The technical essence of the aligned time consumption characteristics is to unify the instruction flow time consumption characteristics generated by different tool nodes in the current execution cycle into time-series characteristics after the same time base. The aligned time consumption characteristics are used to enable subsequent splicing processing to combine instruction flow time consumption characteristics and resource fluctuation characteristics under the same time reference.

[0033] Preferably, in the timing alignment process, when there is a result dependency between different tool nodes, the result return timestamp of the previous tool node and the call request timestamp of the next tool node are first read. A node transmission time consumption record is then generated based on the time interval between the result return timestamp of the previous tool node and the call request timestamp of the next tool node. This inter-node transmission time consumption record is then written into the alignment time consumption feature at the positions corresponding to the two adjacent tool nodes, so that the alignment time consumption feature simultaneously expresses the entry waiting time consumption, node processing time consumption, exit return time consumption, and data transmission time consumption between tool nodes within each tool node. The inter-node transmission time consumption record originates from the timestamp differences between adjacent tool nodes and continues to participate in the generation of multimodal timing operation data, used to describe the timing connection status of the tool call chain when transmitting processing results between tool nodes.

[0034] Preferably, when concatenating the alignment time consumption feature and the resource fluctuation feature, firstly, based on the tool node identifier, the entry waiting time consumption, node processing time consumption, exit return time consumption, and inter-node transmission time consumption records in the alignment time consumption feature are mapped at the node level to the computational resource usage value, memory resource usage value, data access waiting value, interface response status value, and queue backlog status value in the resource fluctuation feature; then, based on the sampling timestamp, the mapped alignment time consumption feature and resource fluctuation feature are arranged to the same time series position; subsequently, according to the call order identifier in the tool call chain, the time series position data and resource fluctuation feature corresponding to each tool node are horizontally concatenated to form a node running feature row, and multiple node running feature rows are vertically arranged according to the call order identifier to generate multimodal time series running data. The concatenation process does not involve directly adding or multiplying data of different dimensions, but rather writes the alignment time consumption feature and resource fluctuation feature into the same multimodal time series running data in a field-by-field manner to retain the technical meaning of each of the alignment time consumption feature and resource fluctuation feature.

[0035] Preferably, the technical essence of the multimodal time-series runtime data is structured time-series data used to describe the runtime status of the tool call chain of the target intelligent agent within the current execution cycle. The multimodal time-series runtime data uses tool node identifiers and call order identifiers as node dimensions, sampling timestamps and time-series position data as time dimensions, and alignment time consumption features and resource fluctuation features as runtime status dimensions. Specifically, alignment time consumption features express the instruction flow time status of each tool node in the tool call chain, and resource fluctuation features express the resource occupancy change status of each tool node in the tool call chain. Through this data organization method, the multimodal time-series runtime data can simultaneously retain the execution sequence of tool nodes, the result transmission relationship between nodes, and the resource fluctuation status, providing input data for the subsequent pre-set perception network to identify abnormal node identifiers.

[0036] Preferably, after the multimodal time-series runtime data is generated, integrity verification processing is performed on the multimodal time-series runtime data. The integrity verification process first reads the tool node identifier, call order identifier, sampling timestamp, alignment time consumption feature, and resource fluctuation feature from the multimodal time-series runtime data, and determines whether each tool node identifier simultaneously corresponds to both alignment time consumption feature and resource fluctuation feature. When a tool node identifier lacks an alignment time consumption feature, the corresponding alignment time consumption feature is added based on the timestamp data in the tool call execution record; when a tool node identifier lacks a resource fluctuation feature, the corresponding resource fluctuation feature is added based on the resource usage sampling record. Subsequently, the supplemented multimodal time-series runtime data is written into the runtime data cache record of the current execution cycle. The runtime data cache record is used to temporarily store the multimodal time-series runtime data after integrity verification processing and provides a data reading location when the preset perception network reads the multimodal time-series runtime data. Through the above integrity verification processing, the multimodal time-series runtime data can maintain data continuity in both the tool node dimension and the time dimension before entering the preset perception network.

[0037] Preferably, in a scenario where a target agent sequentially calls a large language processing model, knowledge base, database, toolset, and environmental awareness node within the current execution cycle to complete a park anomaly monitoring and analysis task, a tool call chain containing the aforementioned tool nodes is first generated based on the agent's resource configuration record and tool call execution record. Then, the instruction flow time consumption characteristics of the large language processing model, knowledge base, database, toolset, and environmental awareness node during the call request, processing execution, and result feedback processes are collected, along with the resource fluctuation characteristics during the execution process. Subsequently, the instruction flow time consumption characteristics are time-series aligned to obtain aligned time consumption characteristics. These aligned time consumption characteristics are then concatenated with the resource fluctuation characteristics according to the tool node identifier, call order identifier, and sampling timestamp to generate multimodal time-series runtime data. Through this multimodal time-series runtime data, the subsequently pre-defined perception network can simultaneously read anomalies in both instruction flow time consumption characteristics and resource fluctuation characteristics within the tool call chain, thereby providing structured input for determining anomaly node identifiers.

[0038] Optionally, the step of inputting the multimodal time-series execution data into a preset perception network to obtain anomaly node identifiers representing execution anomalies includes: The multimodal time-series running data is input into the feature mapping layer of the perception network to obtain the feature distribution matrix; Calculate the deviation value of each feature element in the feature distribution matrix relative to the preset reference element; Extract target feature elements whose deviation values ​​are greater than a preset deviation threshold from the feature distribution matrix; The index information carried by the target feature element is parsed, and the identifier corresponding to the index information is determined as the abnormal node identifier.

[0039] Preferably, the specific implementation process of inputting the multimodal time-series runtime data into a preset perception network to obtain anomaly node identifiers representing execution anomalies is as follows: First, read the runtime data cache record of the current execution cycle, and read the multimodal time-series runtime data from the runtime data cache record. The multimodal time-series runtime data includes tool node identifiers, call order identifiers, sampling timestamps, alignment time consumption features, and resource fluctuation features. Then, write the multimodal time-series runtime data into the input buffer of the preset perception network, and perform node order verification processing on the multimodal time-series runtime data in the input buffer according to the tool node identifiers and call order identifiers to form a node order verification record. Subsequently, according to the node order verification record, the multimodal time-series runtime data in the input buffer is sequentially input into the feature mapping layer, benchmark comparison layer, deviation filtering layer, and identifier parsing layer of the preset perception network. The underlying technology of the pre-defined perception network is a data processing network used for node-level anomaly detection of multimodal time-series runtime data in the tool call chain. Specifically, the feature mapping layer converts the multimodal time-series runtime data into a feature distribution matrix; the benchmark comparison layer calculates the deviation of each feature element in the feature distribution matrix relative to a preset benchmark element; the deviation filtering layer extracts target feature elements from the feature distribution matrix whose deviation values ​​exceed a preset deviation threshold; and the identifier parsing layer parses the index information carried by the target feature elements and determines the anomaly node identifier. Thus, the node order verification record limits the tool node arrangement order when the multimodal time-series runtime data enters the feature mapping layer, and the input buffer temporarily stores the multimodal time-series runtime data after node order verification, enabling the subsequent feature mapping layer to generate the feature distribution matrix according to the actual call sequence of the tool call chain.

[0040] Preferably, the preset perception network is pre-configured based on the target agent's agent resource configuration record, the tool node types in the tool call chain, and the multimodal time-series running data formed within the historical execution cycle. Specifically, firstly, the large language processing model, knowledge base, database, toolset, environmental perception node, and physical execution component that the target agent can call are determined according to the agent resource configuration record, and the large language processing model, knowledge base, database, toolset, environmental perception node, and physical execution component are respectively mapped to tool node types; then, the alignment time consumption characteristics and resource fluctuation characteristics corresponding to each tool node type within the historical execution cycle are read, and a historical running status reference record is established according to the tool node identifier, call order identifier, and sampling timestamp; subsequently, the field arrangement method of the feature mapping layer, the preset benchmark element of the benchmark comparison layer, the preset deviation threshold of the deviation filtering layer, and the index parsing method of the identifier parsing layer are configured according to the historical running status reference record. The field arrangement method is used to limit the column arrangement order of the feature mapping layer when generating the feature distribution matrix, and the index parsing method is used to limit the parsing order of the identifier parsing layer in extracting the tool node identifier, call order identifier, sampling timestamp, running status feature field, and matrix position identifier from the index information carried by the target feature element. Therefore, the pre-defined perception network is not an anomaly monitoring structure that only judges status codes, but a data processing network that can perform anomaly perception by combining the timing operation status of tool nodes and the status of resource usage changes.

[0041] Preferably, when the multimodal time-series runtime data is input into the feature mapping layer of the perception network to obtain the feature distribution matrix, the feature mapping layer first reads the tool node identifier and call order identifier in the multimodal time-series runtime data, and determines the row arrangement order of the feature distribution matrix based on the tool node identifier and call order identifier; then it reads the alignment time consumption feature and resource fluctuation feature in the multimodal time-series runtime data, and determines the column arrangement order of the feature distribution matrix based on the entry waiting time consumption, node processing time consumption, exit return time consumption, and inter-node transmission time consumption records in the alignment time consumption feature, and the computational resource occupancy value, memory resource occupancy value, data access waiting value, interface response status value, and queue backlog status value in the resource fluctuation feature; subsequently, the alignment time consumption feature and resource fluctuation feature corresponding to the same tool node identifier and the same sampling timestamp are written into the corresponding intersection position in the feature distribution matrix to obtain the feature distribution matrix. Each row of the feature distribution matrix corresponds to a tool node in the tool call chain, and each column of the feature distribution matrix corresponds to a runtime status feature field. The feature element at the intersection of the row and column is used to represent the sampling value or status value of the corresponding tool node under the corresponding runtime status feature field. The running status feature fields include entry waiting time, node processing time, exit return time, inter-node transmission time records, computing resource usage value, memory resource usage value, data access waiting value, interface response status value, and queue backlog status value, so that the feature elements in the feature distribution matrix can maintain the source correspondence with the original fields in the multimodal time-series running data.

[0042] Preferably, the feature distribution matrix is ​​not a single vector obtained by compressing multimodal time-series runtime data, but rather matrix data obtained by structurally expanding the tool node dimension, time dimension, and runtime state dimension. Specifically, the tool node dimension is jointly defined by the tool node identifier and the call order identifier, enabling the feature distribution matrix to retain the sequential positions of the large language processing model, knowledge base, database, toolset, environment-aware nodes, and physical execution components in the tool call chain; the time dimension is jointly defined by the sampling timestamp and the time-series position data, enabling the feature distribution matrix to represent the runtime sequence position of each tool node in the current execution cycle; the runtime state dimension is jointly defined by the alignment time consumption feature and the resource fluctuation feature, enabling the feature distribution matrix to simultaneously express the instruction flow time consumption state and the resource fluctuation state. Through the above matrix processing, the feature mapping layer can convert multimodal time-series running data into a feature distribution matrix that is convenient for subsequent deviation calculation, while retaining the tool node source and sampling time source corresponding to each feature element; the tool node source is jointly represented by the tool node identifier and the call order identifier, and the sampling time source is jointly represented by the sampling timestamp and the time-series position data, so that the feature elements in the feature distribution matrix can continue to participate in the deviation calculation of the benchmark comparison layer.

[0043] Preferably, the technical essence of each feature element in the feature distribution matrix is ​​the node-level state value of a tool node on a certain running state feature field within the current execution cycle. For feature elements corresponding to entry waiting time consumption, node processing time consumption, exit return time consumption, and inter-node transmission time consumption records, the feature elements represent the time consumption state of the tool node during instruction flow. For feature elements corresponding to computational resource usage value, memory resource usage value, data access waiting value, interface response status value, and queue backlog status value, the feature elements represent the resource usage status, access waiting status, or interface response status of the tool node during execution. Each feature element is bound to a tool node identifier, call order identifier, sampling timestamp, and running state feature field, enabling subsequent deviation values ​​to be traced back to the tool node and running state feature field to which the feature element belongs, thereby avoiding obtaining only an anomaly conclusion without being able to locate the source of the anomaly. The tool node to which the feature element belongs is used to subsequently generate an anomaly node identifier, and the running state feature field to which the feature element belongs is used to subsequently generate anomaly field records.

[0044] Preferably, before calculating the deviation value of each feature element in the feature distribution matrix relative to the preset benchmark element, the benchmark sampling range corresponding to the current tool node type and the current running state feature field is first read from the historical running state reference record according to the tool node identifier and the running state feature field; then, a preset benchmark element is generated according to the benchmark sampling range, and the preset benchmark element is written into the benchmark element table of the benchmark comparison layer; subsequently, the benchmark comparison layer matches each feature element in the feature distribution matrix with the preset benchmark element in the benchmark element table according to the tool node identifier, call order identifier, and running state feature field. The technical essence of the preset benchmark element is to represent the reference state value that a certain running state feature field of the corresponding tool node should fall into under normal execution. The preset benchmark element comes from multimodal time-series running data and agent resource configuration records within the historical execution cycle, rather than temporarily generated ordinary threshold text. The benchmark element table is used to centrally carry the preset benchmark elements corresponding to different tool node types and different running state feature fields, so that the benchmark comparison layer can calculate the deviation value according to the same running state feature field.

[0045] Preferably, when calculating the deviation value, the benchmark comparison layer does not perform mixed operations on feature elements of different running state feature fields, but instead performs deviation calculations separately according to the same running state feature fields. Specifically, the benchmark comparison layer first reads the entry waiting time consumption feature element corresponding to a certain tool node in the feature distribution matrix, and reads the preset benchmark element corresponding to the same tool node type and the same entry waiting time consumption field in the benchmark element table. It then performs the same field difference processing and non-negation processing on the two to form the entry waiting deviation value. Next, it calculates the node processing deviation value, exit feedback deviation value, inter-node transmission deviation value, computing resource usage deviation value, memory resource usage deviation value, data access waiting deviation value, interface response status deviation value, and queue backlog status deviation value in the same way. Finally, it writes the above deviation values ​​into the deviation distribution record. The deviation distribution record is used to represent the degree of deviation of each feature element in the feature distribution matrix relative to the preset benchmark element. The deviation distribution record is then input into the deviation filtering layer, enabling the deviation filtering layer to extract target feature elements based on the deviation distribution record and the preset deviation threshold.

[0046] Preferably, the technical essence of the deviation value is the deviation data of the same field formed by the feature elements in the feature distribution matrix relative to the preset benchmark element. For time-consuming feature elements, the deviation value is used to represent the degree of time deviation of the instruction flow time of a certain tool node in the current execution cycle relative to the normal execution state; for resource-consuming feature elements, the deviation value is used to represent the degree of resource deviation of the resource occupation state, access waiting state, or interface response state of a certain tool node in the current execution cycle relative to the normal execution state. Since the deviation value is always calculated among the feature fields of the same running state, the time-consuming feature elements are not directly added, multiplied, or compared with the resource-consuming feature elements, so that the calculation process of the deviation value meets the processing requirements of data dimension consistency. The time-consuming feature elements include feature elements corresponding to the entry waiting time, node processing time, exit return time, and inter-node transmission time records, and the resource-consuming feature elements include feature elements corresponding to the calculation resource occupation value, memory resource occupation value, data access waiting value, interface response state value, and queue backlog state value.

[0047] Preferably, when extracting target feature elements with deviation values ​​greater than a preset deviation threshold from the feature distribution matrix, the deviation filtering layer first reads the deviation distribution record and then reads the corresponding preset deviation threshold from the threshold configuration record according to the running status feature field; next, it compares each deviation value with the preset deviation threshold under its corresponding running status feature field; when a deviation value is greater than the corresponding preset deviation threshold, the feature element corresponding to that deviation value is extracted from the feature distribution matrix and written into the target feature element record; when a deviation value is less than or equal to the corresponding preset deviation threshold, the original position of the feature element corresponding to that deviation value in the feature distribution matrix is ​​retained. The technical essence of the target feature element is a node-level state value that deviates abnormally from a preset baseline element within the current execution cycle. The target feature element carries the tool node source, the running status feature field source, and the sampling time source. The target feature element record carries the target feature element and the index information carried by the target feature element, and continues to be input into the identifier resolution layer.

[0048] Preferably, the preset deviation threshold is pre-configured based on the tool node type, runtime status feature fields, and the execution requirements of the current task. Specifically, the tool node type is first categorized into large language processing models, knowledge bases, databases, toolsets, environment-aware nodes, and physical execution components. Then, runtime status feature fields are used to distinguish between entry waiting time, node processing time, exit return time, inter-node transmission time records, computational resource usage, memory resource usage, data access waiting time, interface response status, and queue backlog status. Subsequently, based on the result return requirements of the current task, corresponding preset deviation thresholds are configured for different tool node types and runtime status feature fields, and these preset deviation thresholds are written to a threshold configuration record. Therefore, the preset deviation threshold is not a fixed threshold uniformly applied to all feature elements, but rather a filtering reference corresponding to the tool node type and runtime status feature fields, enabling the deviation filtering layer to extract target feature elements according to the runtime attributes of different tool nodes. The threshold configuration record is used to record the correspondence between the preset deviation threshold and the tool node type and runtime status feature fields, and provides a basis for threshold reading when the deviation filtering layer performs size comparisons.

[0049] Preferably, the index information carried by the target feature element is synchronously written by the feature mapping layer when generating the feature distribution matrix. Specifically, when writing the alignment time consumption feature and resource fluctuation feature into the feature distribution matrix, the feature mapping layer simultaneously binds a tool node identifier, a call order identifier, a sampling timestamp, a running status feature field, and a matrix position identifier to each feature element. The tool node identifier indicates which tool node in the tool call chain the feature element originates from; the call order identifier indicates the call position of the tool node in the tool call chain; the sampling timestamp indicates the data acquisition time corresponding to the feature element; the running status feature field indicates which of the following fields the feature element corresponds to: entry waiting time consumption, node processing time consumption, exit return time consumption, inter-node transmission time consumption record, computational resource usage value, memory resource usage value, data access waiting value, interface response status value, or queue backlog status value; and the matrix position identifier indicates the row and column position of the feature element in the feature distribution matrix. The technical essence of the index information is to trace the target feature element back from its matrix position to the corresponding tool node in the tool call chain. The index information is written along with the target feature element into the target feature element record, enabling the identifier parsing layer to determine the abnormal node identifier based on the target feature element record.

[0050] Preferably, when parsing the index information carried by the target feature element, the identifier parsing layer first reads the target feature element record and extracts the tool node identifier, call order identifier, sampling timestamp, running status feature field, and matrix position identifier from the target feature element record; then, it verifies the row and column positions of the target feature element in the feature distribution matrix based on the matrix position identifier, and verifies the sequential position of the tool node corresponding to the target feature element in the tool call chain based on the call order identifier; subsequently, it determines the corresponding tool node based on the tool node identifier, and combines the tool node identifier with the call order identifier to form a node positioning identifier, which is then used as the identifier corresponding to the index information. The technical essence of the node positioning identifier is that it is node positioning data jointly determined by the tool node identifier and call order identifier carried by the target feature element. The node positioning identifier is used to point to the tool node in the tool call chain that has an abnormal deviation. Through the above parsing process, the identifier corresponding to the index information is uniformly expressed as the node positioning identifier in subsequent processing, avoiding the misinterpretation of the matrix position identifier as the abnormal node position in the tool call chain.

[0051] Preferably, when the identifier corresponding to the index information is determined as the abnormal node identifier, the identifier parsing layer first determines whether there are multiple target feature elements under the same node location identifier; when there are multiple target feature elements under the same node location identifier, the running status feature fields corresponding to the multiple target feature elements are written into the abnormal field record, and the node location identifier and the abnormal field record are combined into the abnormal node identifier; when there is only one target feature element under the same node location identifier, the node location identifier corresponding to the target feature element and the running status feature field are combined into the abnormal node identifier. The technical essence of the abnormal node identifier is a data identifier used to represent the tool node in the tool call chain that has an execution exception and the source of its exception field. The abnormal node identifier continues to participate in the subsequent processing of determining the failed tool node in the preset global dependency topology graph. The exception field record is used to represent the exception source field corresponding to the abnormal node identifier, and the exception field record and the node location identifier together define the node source and field source of the abnormal node identifier.

[0052] Preferably, in a scenario, when the target intelligent agent sequentially calls the large language processing model, knowledge base, database, toolset, and environmental perception node to complete the park anomaly monitoring and analysis task within the current execution cycle, the preset perception network first converts the multimodal time-series running data into a feature distribution matrix through a feature mapping layer. Each row of the feature distribution matrix corresponds to a tool node corresponding to the large language processing model, knowledge base, database, toolset, and environmental perception node, respectively. Each column of the feature distribution matrix corresponds to the entry waiting time, node processing time, exit return time, inter-node transmission time, computational resource usage, memory resource usage, data access waiting time, interface response status, and queue backlog status. Subsequently, the benchmark comparison layer calculates the deviation value for each feature element in the feature distribution matrix, the deviation filtering layer extracts target feature elements with deviation values ​​greater than a preset deviation threshold, and the identifier resolution layer determines the node location identifier based on the index information carried by the target feature element, and combines the node location identifier with the anomaly field record to form an anomaly node identifier. Through this process, the abnormal node identifier can point to the tool node in the tool call chain that has experienced time consumption abnormalities or resource fluctuation abnormalities, and provide node-level input for subsequent failure tool node location.

[0053] Preferably, before training the preset perception network, the following are read: multimodal temporal execution data, tool call execution records, abnormal node identifiers, and agent resource configuration records generated within the historical execution cycle. The multimodal temporal execution data generated within the historical execution cycle provides tool node identifiers, call order identifiers, sampling timestamps, alignment time consumption characteristics, and resource fluctuation characteristics. The tool call execution records generated within the historical execution cycle provide the call requests, processing execution, and result feedback status of each tool node within the historical execution cycle. The abnormal node identifiers generated within the historical execution cycle provide the tool nodes confirmed to have execution anomalies and the source of their abnormal fields. The agent resource configuration records provide the large language processing model, knowledge base, database, toolset, environmental perception nodes, and physical execution components that the target agent can call. By using multimodal time-series runtime data, tool call execution records, abnormal node identifiers, and agent resource configuration records generated during historical execution cycles as training configuration data sources, the pre-defined perception network can configure parameters around the node-level time-series state and node-level resource state in the tool call chain, rather than establishing anomaly judgment rules based solely on a single abnormal status code.

[0054] Preferably, when generating training samples, firstly, based on the tool call execution records formed within the historical execution cycle, multimodal time-series runtime data formed within the historical execution cycle is filtered to include complete tool node identifiers, call order identifiers, sampling timestamps, alignment time consumption features, and resource fluctuation features. Then, based on the abnormal node identifiers formed within the historical execution cycle, the tool node identifiers, call order identifiers, and abnormal field records corresponding to the abnormal node identifiers in the multimodal time-series runtime data formed within the historical execution cycle are labeled to form abnormal labeled runtime data. Subsequently, multimodal time-series runtime data formed within the historical execution cycle that is not pointed to by abnormal node identifiers and meets the integrity verification requirements is organized into normal reference runtime data. The abnormal labeled runtime data is used to train a preset perceptual network to identify abnormal node identifiers, and the normal reference runtime data is used to configure preset benchmark elements and preset deviation thresholds, enabling the preset perceptual network to simultaneously obtain sources of abnormal labeled runtime data and normal reference runtime data.

[0055] Preferably, when configuring the feature mapping layer for training, the tool node identifier, call order identifier, sampling timestamp, alignment time consumption feature, and resource fluctuation feature are first read from the abnormal labeled running data and the normal reference running data. Then, the row arrangement rule of the feature distribution matrix is ​​determined according to the tool node identifier and call order identifier, and the column arrangement rule of the feature distribution matrix is ​​determined according to the entry waiting time consumption, node processing time consumption, exit return time consumption, and inter-node transmission time consumption records in the alignment time consumption feature, and the computational resource usage value, memory resource usage value, data access waiting value, interface response status value, and queue backlog status value in the resource fluctuation feature. Subsequently, the row arrangement rule and column arrangement rule are written into the field arrangement configuration record of the feature mapping layer. The field arrangement configuration record enables the feature mapping layer to generate a feature distribution matrix with the same node dimension, time dimension, and running status dimension as the abnormal labeled running data and the normal reference running data when receiving multimodal time-series running data of the current execution cycle, thereby reducing field misalignment between the training configuration stage and the actual perception stage.

[0056] Preferably, the feature mapping layer also generates index binding configuration records based on the anomaly annotation runtime data. Specifically, during the training configuration process, the feature mapping layer reads the tool node identifier, call order identifier, sampling timestamp, and runtime status feature field corresponding to each feature element in the anomaly annotation runtime data, and binds the above information with the matrix position identifier in the feature distribution matrix to form an index binding configuration record. Then, based on the anomaly node identifiers formed within the historical execution cycle, it verifies whether the index binding configuration record can trace back from the target feature element to the corresponding tool node identifier and anomaly field record. When the index binding configuration record can complete the above tracing, it is written to the feature mapping layer. Through the above training configuration, the feature distribution matrix generated by the feature mapping layer not only contains node-level state values ​​but also retains the data location information required by the subsequent identifier parsing layer to determine the anomaly node identifier.

[0057] Preferably, when configuring the benchmark comparison layer for training, the running status feature fields corresponding to different tool node types are first read from the normal reference running data. Then, according to the tool node type and running status feature fields, the reference ranges for entry waiting time, node processing time, exit return time, inter-node transmission time, computing resource usage, memory resource usage, data access waiting, interface response status, and queue backlog status are organized. Preset benchmark elements are then generated based on these reference ranges and written into the benchmark element table of the benchmark comparison layer. These preset benchmark elements originate from the same type of tool node and the same type of running status feature fields in the normal reference running data. They represent the reference status value of the corresponding tool node in its normal execution state, enabling the benchmark comparison layer to perform comparisons within the same field and tool node type range when calculating deviation values.

[0058] Preferably, during the training configuration process, the benchmark comparison layer also configures the deviation distribution for the anomaly-annotated runtime data. Specifically, the benchmark comparison layer first maps the feature distribution matrix obtained by transforming the anomaly-annotated runtime data through the feature mapping layer to the benchmark element table, and then calculates the deviation values ​​for entry waiting, node processing, exit return, inter-node transmission, computational resource usage, memory resource usage, data access waiting, interface response status, and queue backlog status. Subsequently, these deviation values ​​are mapped to the anomaly field records in the anomaly-annotated runtime data to form a training deviation distribution record. The training deviation distribution record is used to represent the deviation patterns of different anomaly field sources on different runtime status feature fields and continues to participate in the training configuration of the deviation filtering layer.

[0059] Preferably, before training and configuring the deviation screening layer, a normal deviation distribution record is formed based on normal reference running data. Specifically, the benchmark comparison layer first maps the feature distribution matrix obtained by transforming the normal reference running data through the feature mapping layer to the benchmark element table, then calculates the deviation value of each feature element in the normal reference running data relative to the preset benchmark element, and organizes the deviation values ​​corresponding to the normal reference running data according to the tool node type and running status feature field to form a normal deviation distribution record. The normal deviation distribution record is used to represent the deviation range of the normal reference running data on different tool node types and different running status feature fields. The normal deviation distribution record continues to participate in the training and configuration of the deviation screening layer together with the training deviation distribution record, enabling the deviation screening layer to configure a preset deviation threshold according to the normal deviation range and the abnormal deviation range.

[0060] Preferably, when configuring the deviation filtering layer for training, the training deviation distribution record and the normal deviation distribution record are read first; then, the deviation distribution boundary between the training deviation distribution record and the normal deviation distribution record is compared according to the tool node type and the running status feature field; subsequently, a preset deviation threshold is configured according to the deviation distribution boundary, and the preset deviation threshold is written into the threshold configuration record of the deviation filtering layer. The threshold configuration record is not a fixed threshold uniformly set for all feature elements, but a filtering reference configured according to the tool node types such as large language processing model, knowledge base, database, toolset, environment-aware node and physical execution component, as well as the running status feature fields such as entry waiting time, node processing time, exit return time, inter-node transmission time, computing resource usage value, memory resource usage value, data access waiting value, interface response status value and queue backlog status value.

[0061] Preferably, the deviation filtering layer further verifies and adjusts the threshold configuration record by extracting target feature elements. Specifically, the deviation filtering layer first extracts target feature elements with deviation values ​​greater than a preset deviation threshold from the feature distribution matrix corresponding to the anomaly labeled running data according to the threshold configuration record. Then, it verifies the correspondence between the index information carried by the target feature element and the anomaly node identifier in the anomaly labeled running data. If the index information carried by the target feature element points to the tool node identifier and anomaly field record in the anomaly node identifier, the current preset deviation threshold is retained. If the index information carried by the target feature element fails to point to the tool node identifier or anomaly field record in the anomaly node identifier, the preset deviation threshold under the corresponding tool node type and the corresponding running status feature field is readjusted according to the training deviation distribution record. Through the above training configuration, the deviation filtering layer can take into account both the source of anomaly fields and the source of tool nodes when extracting target feature elements.

[0062] Preferably, when configuring the identifier resolution layer for training, the following steps are taken: first, read the index binding configuration record formed by the feature mapping layer, the target feature elements extracted by the deviation filtering layer, and the abnormal node identifiers in the abnormal annotation running data; then, read the tool node identifier, call order identifier, sampling timestamp, running status feature field, and matrix position identifier from the index information carried by the target feature elements, and combine the tool node identifier, call order identifier, and running status feature field into a training resolution identifier; subsequently, map the training resolution identifier to the abnormal node identifiers in the abnormal annotation running data to form an identifier resolution configuration record. This identifier resolution configuration record enables the identifier resolution layer to trace the target feature elements back from their matrix positions to the tool nodes in the tool call chain during actual perception, and to identify the tool node and its abnormal field source as the abnormal node identifier.

[0063] Preferably, the identifier resolution layer also performs aggregation training configuration for multiple target feature elements under the same tool node identifier. Specifically, the identifier resolution layer first reads the runtime status feature fields carried by multiple target feature elements under the same tool node identifier, then writes the multiple runtime status feature fields into an anomaly field record, and combines the tool node identifier, call order identifier, and anomaly field record into an aggregated anomaly node identifier; subsequently, the aggregated anomaly node identifier is verified against the anomaly node identifiers formed in the historical execution cycle to determine whether multiple target feature elements should all point to the same tool node. Through the above training configuration, when the same tool node simultaneously exhibits time consumption anomalies and resource fluctuation anomalies, the identifier resolution layer can output an anomaly node identifier containing multiple sources of anomaly fields, rather than generating multiple fragmented anomaly conclusions.

[0064] Preferably, the training of the preset perceptual network further includes inter-layer consistency verification. This inter-layer consistency verification first reads the feature distribution matrix output by the feature mapping layer, the deviation distribution record output by the benchmark comparison layer, the target feature element record output by the deviation filtering layer, and the abnormal node identifier output by the identifier resolution layer. Then, it verifies whether the feature elements in the feature distribution matrix can be found in the deviation distribution record with corresponding deviation values, whether the target feature elements in the target feature element record originate from deviation values ​​in the deviation distribution record that are greater than a preset deviation threshold, and whether the abnormal node identifier can be traced back to the index information carried by the target feature element. When all the above correspondences are satisfied, the training configuration result for this round is written into the perceptual network training configuration record. The perceptual network training configuration record is used to store the field transmission relationships between the feature mapping layer, benchmark comparison layer, deviation filtering layer, and identifier resolution layer, enabling the preset perceptual network to determine abnormal node identifiers according to the same field transmission relationships in the current execution cycle.

[0065] Preferably, in a scenario, when the target agent sequentially calls the large language processing model, knowledge base, database, toolset, and environmental perception node to complete the park anomaly monitoring and analysis task within a historical execution cycle, if the anomaly node identifier formed within the historical execution cycle points to the tool node corresponding to the database, then the multimodal time-series runtime data formed within that historical execution cycle is organized into anomaly-labeled runtime data. During the training configuration process, the feature mapping layer writes the tool nodes corresponding to the large language processing model, knowledge base, database, toolset, and environmental perception node into the feature distribution matrix according to the calling order. The benchmark comparison layer configures preset benchmark elements on the runtime status feature fields such as node processing time consumption, data access waiting value, and interface response status value of the tool node corresponding to the database based on the normal reference runtime data. The deviation filtering layer configures the corresponding preset deviation threshold based on the training deviation distribution record and the normal deviation distribution record. The identifier parsing layer outputs the anomaly node identifier pointing to the tool node corresponding to the database based on the index information carried by the target feature element. Through the above training configuration, the preset perception network can identify the anomaly node identifier in the tool call chain based on the multimodal time-series runtime data in the subsequent current execution cycle, and continue to provide the anomaly node identifier to the subsequent failed tool node location processing.

[0066] Optionally, isolating the failed tool node to obtain a link to be repaired with a topological break includes: Delete the first input connection edge in the global dependency topology graph that points to the failed tool node; Delete the second output connection edge originating from the failed tool node in the global dependency topology graph; The topology obtained after deleting the first input connection edge and the second output connection edge is determined as the link to be repaired; The starting node corresponding to the first input connection edge is determined as the first tool node, and the position where the first tool node was originally connected to the failed tool node is determined as the topological break.

[0067] Preferably, before isolating the failed tool node, a preset global dependency topology graph, abnormal node identifiers, and tool call execution records corresponding to the tool call chain are read. The preset global dependency topology graph is pre-constructed based on the large language processing model, knowledge base, database, toolset, environment-aware nodes, and physical execution components that the target agent can call. Each tool node in the preset global dependency topology graph corresponds to a resource instance that can be called by the target agent. The directed connection relationship between nodes in the preset global dependency topology graph is used to represent the data flow relationship where the output data of the previous tool node is read and processed by the next tool node. Specifically, during construction, the tool node identifier, node input port identifier, and node output port identifier of each tool node are first determined according to the agent resource configuration record. Then, the directed connection relationship between nodes is established according to the tool call order, result return relationship, and data format inheritance relationship within the historical execution cycle. Subsequently, the tool node identifier, node input port identifier, node output port identifier, and directed connection relationship between nodes are written into the same topology graph record, and the directed connection relationship between nodes in the topology graph record is synchronously written into the active connection edge record to form the preset global dependency topology graph. The active connection edge record is used to represent the directed connection relationship between nodes that can currently participate in the data flow of the tool call link. The preset global dependency topology graph is not an ordinary resource list, but a topological data structure used to describe the data input, processing output and result feedback relationship between tool nodes. Subsequent isolation processing can locate the first input connection edge and the second output connection edge that need to be disconnected in the preset global dependency topology graph, rather than replacing the entire tool call link.

[0068] Preferably, when determining a failed tool node in a preset global dependency topology graph based on an anomaly node identifier, the node location identifier and anomaly field record in the anomaly node identifier are first read, and the corresponding tool node identifier is retrieved in the preset global dependency topology graph according to the node location identifier; then, the retrieved tool node identifier is verified against the anomaly field record to determine whether the entry waiting time, node processing time, exit return time, inter-node transmission time, computing resource consumption, memory resource consumption, data access waiting value, interface response status value, or queue backlog status value represented by the anomaly field record belong to the running status feature field corresponding to the tool node identifier; subsequently, the tool node that passes the corresponding verification is determined as a failed tool node. The technical essence of the failed tool node is that it is a tool node pointed to by the anomaly node identifier in the current execution cycle and has a node input port identifier and a node output port identifier in the preset global dependency topology graph. The failed tool node continues to participate in the determination process of the first input connection edge and the second output connection edge, so that the isolation process can be carried out around the input side and the output side of the failed tool node.

[0069] Preferably, when determining the first input connection edge pointing to the failed tool node in the preset global dependency topology graph, the node input port identifier corresponding to the failed tool node is first read, and the directed connection relationship between nodes with the node input port identifier as the termination port is retrieved in the active connection edge record of the preset global dependency topology graph; then, the start port identifier, termination port identifier, connection edge direction identifier, and data receiving field record in the directed connection relationship between the nodes are read, and the directed connection relationship between nodes whose termination port identifier is consistent with the node input port identifier of the failed tool node is determined as the first input connection edge. The technical essence of the first input connection edge is that in the preset global dependency topology graph, it is a data transmission edge from the tool node upstream of the failed tool node to the input side of the failed tool node. The first input connection edge is used to represent the topological connection relationship of the tool node upstream of the failed tool node transmitting the output data of the first tool node to the failed tool node. Through the above determination method, the first input connection edge originates from the existing directed connection relationship between nodes in the preset global dependency topology graph, rather than a connection mark temporarily generated during the isolation process.

[0070] Preferably, when deleting the first input connection edge pointing to the failed tool node in the preset global dependency topology graph, the starting node and the starting port identifier corresponding to the first input connection edge are not deleted. Instead, the first input connection edge is removed from the active connection edge record of the preset global dependency topology graph, and the removed first input connection edge is written into the isolation edge cache record. The isolation edge cache record is used to retain the starting port identifier, ending port identifier, connection edge direction identifier, and data transmission field record of the first input connection edge before isolation processing. By writing the first input connection edge into the isolation edge cache record, the first tool node can be determined based on the starting node corresponding to the first input connection edge, and the topology break point can be determined based on the original position of the first input connection edge connecting to the failed tool node. This allows the isolation processing to cut off the data flow on the input side of the failed tool node while retaining the data source required for break point location.

[0071] Preferably, when determining the second output connection edge issued by the failed tool node in the preset global dependency topology graph, the node output port identifier corresponding to the failed tool node is first read, and the directed connection relationship between nodes with the node output port identifier as the starting port is retrieved in the active connection edge record of the preset global dependency topology graph; then, the starting port identifier, ending port identifier, connection edge direction identifier, and data receiving field record in the directed connection relationship between the nodes are read, and the directed connection relationship between nodes whose starting port identifier is consistent with the node output port identifier of the failed tool node is determined as the second output connection edge. The technical essence of the second output connection edge is that in the preset global dependency topology graph, the output side data transmission edge from the failed tool node to the downstream tool node is used to represent the topology connection relationship in which the failed tool node continues to transmit the processing result to the downstream tool node. Since the second output connection edge directly corresponds to the node output port identifier of the failed tool node, deleting the second output connection edge can block the data path of the failed tool node continuing to output the abnormal processing result to the downstream tool node.

[0072] Preferably, when deleting the second output connection edge issued by the failed tool node in the preset global dependency topology graph, the directed connection relationship between the corresponding nodes is first located from the active connection edge record based on the start port identifier and end port identifier in the second output connection edge; then, the directed connection relationship between the nodes is removed from the active connection edge record, and the removed second output connection edge is written into the isolation edge cache record; subsequently, the deletion status of the first input connection edge and the second output connection edge is recorded according to the isolation edge cache record to form the failed tool node isolation status record. The failed tool node isolation status record is used to indicate that both the input-side data transmission edge and the output-side data transmission edge of the failed tool node have been removed from the active connection edge record, and the failed tool node isolation status record continues to participate in the determination process of the link to be repaired. Through the above deletion method, the second output connection edge is stripped from the directed connection relationship between currently executable nodes, while the original connection information of the second output connection edge is still retained in the isolation edge cache record for subsequent determination of the connection recovery requirements of downstream tool nodes in the link to be repaired.

[0073] Preferably, when the topology obtained after deleting the first input connection edge and the second output connection edge is determined as the link to be repaired, the directed connection relationships between the remaining nodes in the active connection edge record are first read, and the isolated topology is generated by combining the tool node identifiers in the preset global dependency topology graph; then, according to the isolation status record of the failed tool node, the failed tool node is marked as an inactive tool node in the isolated topology, and the node position identifier of the inactive tool node in the preset global dependency topology graph is retained; subsequently, connected segment identification processing is performed on the isolated topology to determine the upstream and downstream link segments that can still flow data along the directed connection relationships between nodes, and the isolated topology containing the upstream link segment, downstream link segment, inactive tool node, and node position identifier is determined as the link to be repaired. The technical essence of the isolated topology is the remaining topology data formed after removing the first input connection edge and the second output connection edge from the preset global dependency topology graph, and the technical essence of the link to be repaired is the tool call link that retains the topology break position, the upstream available tool node, and the downstream tool node to be connected on the basis of the isolated topology. Therefore, the link to be repaired is not the incomplete link text after failure, but a topological data structure that continues to carry the extraction of breakpoint output features and the access processing of target replacement nodes.

[0074] Preferably, the upstream and downstream link segments in the link to be repaired are jointly determined based on the active connection edge record and the isolation edge cache record. Specifically, the first input connection edge is first read from the isolation edge cache record, and the starting port identifier corresponding to the first input connection edge is traced back to the corresponding starting node; then, using the starting node as the termination reference, the tool nodes that are still connected are read in reverse along the active connection edge record to form the upstream link segment; subsequently, the second output connection edge is read from the isolation edge cache record, and the termination port identifier corresponding to the second output connection edge is traced back to the corresponding downstream tool node; then, using the downstream tool node as the starting reference, the tool nodes that are still connected are read in forward along the active connection edge record to form the downstream link segment. The upstream link segment is used to provide the first tool node for subsequent extraction of breakpoint output features, and the downstream link segment is used to provide the continuation direction for the result transmission after the subsequent target replacement node is connected. By retaining both the upstream and downstream link segments in the link to be repaired, the subsequent dynamic reorganization process can utilize the existing available topology resources instead of building a complete tool call link from scratch.

[0075] Preferably, when determining the starting node corresponding to the first input connection edge as the first tool node, the first input connection edge in the isolation edge cache record is first read, and the starting port identifier is extracted from the first input connection edge; then, the corresponding tool node identifier is queried in the preset global dependency topology graph according to the starting port identifier, and the tool node corresponding to the queried tool node identifier is determined as the starting node corresponding to the first input connection edge; subsequently, the most recent result feedback record of the starting node in the tool call execution record is read, and the starting node with the most recent result feedback record is determined as the first tool node. The technical essence of the starting node corresponding to the first input connection edge is the upstream tool node that inputs data to the failed tool node before the failed tool node is isolated; the technical essence of the first tool node is a tool node located upstream of the topology breakpoint in the link to be repaired, which can provide breakpoint output characteristics. By determining the starting node corresponding to the first input connection edge as the first tool node, the subsequent breakpoint output characteristics can directly come from the most recent result feedback record of the first tool node, so that when the target replacement node is connected, it can inherit the upstream data state before the failed tool node is isolated.

[0076] Preferably, when determining the position where the first tool node was originally connected to the failed tool node as the topology break, the first input connection edge in the isolation edge cache record is read first, and the start port identifier of the first tool node, the end port identifier of the failed tool node, and the data transfer field record between them are extracted from the first input connection edge; then, the output-side break position of the first tool node is determined according to the start port identifier of the first tool node, and the input-side break position of the failed tool node is determined according to the end port identifier of the failed tool node; subsequently, the output-side break position of the first tool node, the input-side break position of the failed tool node, and the data transfer field record are written together into the topology break record to determine the topology break. The technical essence of the position where the first tool node was originally connected to the failed tool node is the data transfer position between the output side of the first tool node and the input side of the failed tool node connected by the first input connection edge before isolation processing; the technical essence of the topology break is the position of the data interface to be continued in the link to be repaired after deleting the first input connection edge, and the topology break is used to indicate the toolchain break point position that the target replacement node should connect to.

[0077] Preferably, the topology break is not simply a missing connection edge, but rather the interface data to be repaired expressed by the topology break record. The topology break record includes the tool node identifier of the first tool node, the output-side break position of the first tool node, the node position identifier of the failed tool node, the input-side break position of the failed tool node, a data continuation field record, and a break direction identifier. Specifically, the tool node identifier of the first tool node is used to locate the source of the breakpoint output features; the output-side break position of the first tool node is used to locate the upstream connection end when the target replacement node connects; the node position identifier of the failed tool node is used to preserve the original position of the failed tool node in the preset global dependency topology graph; the input-side break position of the failed tool node is used to indicate the isolated input end reference; the data continuation field record is used to indicate the fields in the output data of the first tool node that need to be read by the target replacement node; and the break direction identifier is used to indicate that the target replacement node should receive data from the first tool node and transmit the results to the downstream link segment. The topology break record continues to participate in breakpoint output feature extraction, target replacement node access, and feature alignment processing, enabling the topology break to serve as a clear data continuation position during subsequent dynamic reassembly.

[0078] Preferably, after the link to be repaired is determined, isolation consistency verification is also performed on the link to be repaired. The isolation consistency verification process first reads the active connection edge record, isolation edge cache record, failed tool node isolation status record, and topology breakpoint record in the link to be repaired; then it verifies whether the active connection edge record still contains a first input connection edge pointing to the failed tool node or a second output connection edge issued by the failed tool node; when the active connection edge record does not contain a first input connection edge or a second output connection edge, it continues to verify whether the output-side breakpoint position of the first tool node in the topology breakpoint record is consistent with the starting port identifier in the first input connection edge; when they are consistent, the link to be repaired is marked as repairable. The repairable state indicates that the link to be repaired has completed the isolation of the failed tool node and retained the topology breakpoint record, and the repairable state continues to participate in subsequent breakpoint output feature extraction processing. Through the above isolation consistency verification process, the link to be repaired can eliminate interference caused by residual connection edges to the access of the target replacement node before entering subsequent dynamic reassembly.

[0079] Preferably, in a scenario, the target agent sequentially calls the large language processing model, knowledge base, database, toolset, and environmental perception node to complete the park anomaly monitoring and analysis task within the current execution cycle. When the anomaly node identifier points to the tool node corresponding to the database, the tool node corresponding to the database is first identified as the failed tool node in the preset global dependency topology graph. Then, the first input connection edge pointing to the tool node corresponding to the database is read, and the starting node corresponding to the first input connection edge is identified as the tool node corresponding to the knowledge base, thus identifying the tool node corresponding to the knowledge base as the first tool node. Subsequently, the first input connection edge pointing to the tool node corresponding to the database is deleted, and the second output connection edge emitted by the tool node corresponding to the database is also deleted, to obtain the repair link containing the tool node corresponding to the large language processing model, the tool node corresponding to the knowledge base, the tool node corresponding to the toolset, and the tool node corresponding to the environmental perception node. Finally, the position where the tool node corresponding to the knowledge base was originally connected to the tool node corresponding to the database is identified as the topology breakpoint. Through this process, the repair link can retain the data retrieval and task understanding results completed before the anomaly occurred, and provide topology location basis for subsequent extraction of the breakpoint output features corresponding to the first tool node and access to the target replacement node.

[0080] Optionally, generating a feature compensation vector based on the breakpoint output features and the expected state features includes: The difference between the breakpoint output features and the expected state features is calculated to obtain a difference feature vector. The difference feature vector is input into a preset feature mapping network to obtain the mapped feature compensation vector.

[0081] Preferably, before generating the feature compensation vector based on the breakpoint output features and the expected state features, the topology breakpoint record, the most recent result feedback record of the first tool node, and the task state reference record corresponding to the current task being executed by the target agent are read from the link to be repaired. The topology breakpoint record provides the output-side breakpoint location, data receiving field record, and breakpoint direction identifier of the first tool node. The most recent result feedback record of the first tool node provides intermediate result data that the first tool node output before the failed tool node was isolated but has not yet been stably processed by the failed tool node. The task state reference record provides the expected data state, expected interface state, and expected result state that the target agent needs to achieve when continuing to execute its current task. Then, based on the data receiving field record in the topology breakpoint record, the output field values, output field types, output timestamps, and output source identifiers related to the topology breakpoint are extracted from the most recent result feedback record of the first tool node, and these values ​​are organized into breakpoint output features. The technical essence of the breakpoint output feature is that the first tool node forms a structured feature upstream of the topology breakpoint that can represent the current breakpoint data state of the link to be repaired. The breakpoint output feature is not all the output content of the first tool node, but rather the upstream data state that needs to be taken over by the subsequent target replacement node when it is connected, based on the topology breakpoint record. The upstream data state comes from the intermediate result data in the most recent result feedback record of the first tool node and continues to participate in subsequent feature field alignment processing and difference processing.

[0082] Preferably, the expected state features are generated based on the task state reference record corresponding to the current task being executed by the target agent. Specifically, the task processing objective, result return requirements, data fields that downstream link segments need to receive, and data formats that downstream link segments can recognize are first read from the task state reference record. Then, based on the data fields that downstream link segments need to receive, the expected field value range, expected field type, expected interface state, and expected result state that the target supplementary node should output after completing the supplementary processing are determined. Subsequently, the expected field value range, expected field type, expected interface state, and expected result state are organized into expected state features. The technical essence of the expected state features is the data state reference required when the current task being executed by the target agent continues to flow after the topology breakpoint. The expected state features and the breakpoint output features are established in a corresponding relationship under the same topology breakpoint record, so that subsequent difference processing can be carried out around the difference between the data state already output by the first tool node and the data state that the current task being executed by the target agent needs to achieve.

[0083] Preferably, before performing the difference processing on the breakpoint output features and the expected state features, feature field alignment processing is performed first. Feature field alignment processing first reads the output field type, output field value, and data continuation field record from the breakpoint output features, and reads the expected field type, expected field value range, and data fields that the downstream link segment needs to receive from the expected state features. Then, based on the data continuation field record, the output field type in the breakpoint output features is matched with the expected field type in the expected state features to form a field correspondence record. Subsequently, based on the field correspondence record, the output field value in the breakpoint output features and the expected field value range in the expected state features are arranged to the same field position to form a breakpoint field sequence and an expected field sequence. The breakpoint field sequence represents the field state already formed by the first tool node at the topology breakpoint, and the expected field sequence represents the field state that needs to be achieved when the topology breakpoint is subsequently reconnected. By first forming the breakpoint field sequence and the expected field sequence, the difference processing can be performed between fields of the same type and at the same field position, avoiding data meaning confusion caused by direct comparison between different field types.

[0084] Preferably, when performing subtraction processing on the breakpoint output features and the expected state features, the breakpoint field sequence and the expected field sequence are first read, and the output field values ​​in the breakpoint field sequence are compared with the expected field value ranges in the expected field sequence item by item according to the corresponding records of the fields. For numeric fields, the deviation between the output field values ​​and the expected field value ranges is organized into numeric difference components. For text fields, the semantic category differences between the output field values ​​and the expected field value ranges are organized into text difference components. For interface fields, the format inconsistencies between the output field type and the expected field type are organized into interface difference components. For time-series fields, the deviations between the output timestamp and the time-series requirements corresponding to the expected interface state are organized into time-series difference components. Subsequently, the numeric difference components, text difference components, interface difference components, and time-series difference components are arranged according to the field order in the corresponding records of the fields to obtain the difference feature vector. The technical essence of the difference feature vector is that it is a vectorized data used to express the compensation requirements between the breakpoint output features and the expected state features in terms of field content, field type, interface state and timing state. Each component in the difference feature vector can be traced back to a field position in the corresponding record.

[0085] Preferably, the difference feature vector is not a simple set of differences obtained by subtracting the breakpoint output features from the expected state features, but rather an expression of the breakpoint compensation requirement formed under the constraints of corresponding field records. Specifically, the difference feature vector includes at least field content difference components, field type difference components, interface state difference components, temporal state difference components, and downstream continuation difference components; wherein, the field content difference components are obtained by organizing the numerical difference components and text difference components according to the corresponding field records, and are used to represent the deviation state between the output content of the first tool node and the content required by the target agent for the current task; the field type difference components are obtained by organizing the interface difference components according to the corresponding field records, and are used to represent the adaptation state between the output field type of the first tool node and the received field type of the downstream link segment; the interface state difference components are used to represent the connection state between the input interface and the output interface at the topological breakpoint; the temporal state difference components are obtained by organizing the temporal difference components according to the corresponding field records, and are used to represent the connection state between the output time of the first tool node and the continued processing time of the downstream link segment; the downstream continuation difference components are used to represent the data content processing capability that should be supplemented after the target supplementary node is connected. The above components together describe the technical content that needs to be compensated at the topological break, so that the difference feature vector can continue to be input into the preset feature mapping network for compensation mapping processing.

[0086] Preferably, the preset feature mapping network is pre-configured based on the agent resource configuration record, the capability description vectors corresponding to multiple candidate supplementary nodes in the preset component library, the difference feature vectors formed during the historical execution cycle, and the feature compensation vectors formed during the historical execution cycle. Specifically, firstly, the tool node types corresponding to the large language processing model, knowledge base, database, toolset, environmental perception node, and physical execution component that the target agent can call are determined according to the agent resource configuration record; then, the capability description vectors corresponding to multiple candidate supplementary nodes in the preset component library are read, and the field types, interface types, data processing actions, and result output types that the candidate supplementary nodes can process are organized according to the capability description vectors; subsequently, based on the correspondence between the difference feature vectors formed during the historical execution cycle and the target supplementary node access results formed during the historical execution cycle, the input normalization layer, field completion layer, difference encoding layer, compensation semantic mapping layer, and output constraint layer in the preset feature mapping network are configured. Thus, the preset feature mapping network is used to convert the difference feature vectors into feature compensation vectors that can match the capability description vectors, rather than simply classifying the difference feature vectors.

[0087] Preferably, the input normalization layer is used to perform field order verification and dimension normalization on the difference feature vector. Specifically, the input normalization layer first reads the field content difference components, field type difference components, interface status difference components, time sequence status difference components, and downstream continuity difference components in the difference feature vector, and reads the field order in the corresponding field records; then, it verifies whether each component in the difference feature vector corresponds to the field position in the breakpoint output features and expected state features according to the corresponding field records; subsequently, it performs same-type scaling on the difference components belonging to the same field type to form a normalized difference feature vector. The normalized difference feature vector is used to ensure that difference feature vectors from different tool nodes have a consistent field arrangement rule before entering the field completion layer. The normalized difference feature vector still retains the tool node identifier, topology breakpoint record, and field correspondence record corresponding to each difference component, so that the subsequent compensation mapping process will not lose the difference source.

[0088] Preferably, the field completion layer is used to complete missing fields in the normalized difference feature vector based on the topology break record and downstream link segments. Specifically, the field completion layer first reads the data acceptance field record in the topology break record and then reads the data fields that the downstream link segments need to receive; then it determines whether the normalized difference feature vector is missing a difference component corresponding to the data field that the downstream link segments need to receive; when the normalized difference feature vector is missing a corresponding difference component, it generates a completion difference component based on the expected field type and expected interface state in the expected state feature, and writes the completion difference component into the corresponding field position in the normalized difference feature vector to form a completion difference feature vector. The completion difference feature vector is used to express the existing difference state at the topology break and the compensation state required for the downstream link segments to continue execution, so that the subsequent difference encoding layer can perform encoding processing within a relatively complete field range.

[0089] Preferably, the differential encoding layer is used to convert the padded differential feature vector into a breakpoint compensation encoding record. Specifically, the differential encoding layer first reads the field content differential components in the padded differential feature vector and generates content compensation codes based on the field content differential components; then it reads the field type differential components and interface status differential components and generates interface compensation codes based on the field type differential components and interface status differential components; subsequently, it reads the timing status differential components and downstream continuity differential components and generates timing continuity compensation codes based on the timing status differential components and downstream continuity differential components; finally, it arranges the content compensation codes, interface compensation codes, and timing continuity compensation codes according to the breakpoint direction identifier in the topology breakpoint record to form a breakpoint compensation encoding record. The technical essence of the breakpoint compensation encoding record is to convert the compensation requirements in the differential feature vector into intermediate encoded data that can be read by the compensation semantic mapping layer. The breakpoint compensation encoding record continues to retain the topology breakpoint record, the tool node identifier of the first tool node, and the continuity direction of the downstream link segment.

[0090] Preferably, the compensation semantic mapping layer is used to generate the main content of the feature compensation vector based on the break-point compensation coding record. Specifically, the compensation semantic mapping layer first reads the content compensation coding in the break-point compensation coding record and maps the content compensation coding to a content compensation component that represents the data content processing capabilities that the target replacement node should supplement; then it reads the interface compensation coding and maps the interface compensation coding to an interface compensation component that represents the input interface acceptance capabilities and output interface acceptance capabilities that the target replacement node should possess; subsequently, it reads the timing continuity compensation coding and maps the timing continuity compensation coding to a timing compensation component that represents the result acceptance timing and result return timing that the target replacement node should possess. The content compensation component, interface compensation component, and timing compensation component together form the initial feature compensation vector, which is used to describe the capability requirements that need to be supplemented by the candidate replacement node at the topological break point.

[0091] Preferably, the output constraint layer is used to perform component library adaptation processing on the initial feature compensation vector to obtain the mapped feature compensation vector. The output constraint layer first reads the capability description vectors corresponding to multiple candidate filler nodes in the preset component library, and reads the processable field types, input interface types, output result types, and acceptable time series ranges from the capability description vectors. Then, it performs field-level mapping between the content compensation component, interface compensation component, and time series compensation component in the initial feature compensation vector and the processable field types, input interface types, output result types, and acceptable time series ranges, respectively. Subsequently, components that cannot be read by any candidate filler node in the preset component library are converted into compensation expressions consistent with the capability description vector fields to form the mapped feature compensation vector. The technical essence of the mapped feature compensation vector is a compensation requirement vector formed based on the difference between the breakpoint output features and the expected state features, capable of matching with the capability description vectors corresponding to each candidate filler node. The mapped feature compensation vector continues to participate in the determination of the target filler node.

[0092] Preferably, the mapped feature compensation vector includes at least a content compensation component, an interface compensation component, a timing compensation component, and a connection direction component. The content compensation component represents the data content processing capability that the target replacement node needs to supplement the output data of the first tool node; the interface compensation component represents the input and output interfaces that the target replacement node needs to adapt to; the timing compensation component represents the result feedback timing that the target replacement node needs to meet after connecting to the topology breakpoint; and the connection direction component represents the breakpoint output features that the target replacement node should read from the output breakpoint location of the first tool node and transmit the processing results to the downstream link segment. Each of these components originates from the field correspondence between the differential feature vector, the topology breakpoint record, and the capability description vectors corresponding to multiple candidate replacement nodes in the preset component library. This ensures that the mapped feature compensation vector can express the breakpoint compensation requirements that can be used for candidate replacement node matching, rather than simply expressing the ordinary distance between the breakpoint output features and the expected state features.

[0093] Preferably, after generating the mapped feature compensation vector, a compensation consistency verification process is performed on the mapped feature compensation vector. The compensation consistency verification process first reads the content compensation component, interface compensation component, timing compensation component, and connection direction component from the mapped feature compensation vector, and reads the data acceptance field record and the breakpoint direction identifier from the topology breakpoint record; then it verifies whether the content compensation component corresponds to the field type in the data acceptance field record, whether the interface compensation component corresponds to the output-side breakpoint position of the first tool node and the input-side connection position of the downstream link segment, and whether the connection direction component corresponds to the breakpoint direction identifier; when all the above correspondences are satisfied, the mapped feature compensation vector is written into the compensation vector cache record. The compensation vector cache record is used to store feature compensation vectors that can enter subsequent matching processing, so that when the feature compensation vector is subsequently matched with the capability description vectors corresponding to multiple candidate replacement nodes, the feature compensation vector that has passed the compensation consistency verification process can be directly read.

[0094] Preferably, in a scenario, the target agent sequentially calls the large language processing model, knowledge base, database, toolset, and environmental perception node to complete the park anomaly monitoring and analysis task within the current execution cycle. When the tool node corresponding to the database is determined to be a failed tool node and forms a topological break, the first tool node can be the tool node corresponding to the knowledge base. At this time, the breakpoint output features corresponding to the data transmission field record of the topological break record are first extracted from the most recent result feedback record of the tool node corresponding to the knowledge base. Then, the expected state features are generated according to the data fields that the target agent needs to continue to transmit to the toolset and environmental perception node according to the current execution task. Subsequently, the breakpoint output features and expected state features are aligned and subtracted to obtain the difference feature vector, and the difference feature vector is input into the preset feature mapping network. The pre-defined feature mapping network processes the differential feature vectors sequentially through an input normalization layer, a field completion layer, a differential encoding layer, a compensation semantic mapping layer, and an output constraint layer to obtain the mapped feature compensation vector. This allows the mapped feature compensation vector to express the data reading capability, input interface acceptance capability, output interface acceptance capability, and result feedback capability that the target filler node needs to supplement after the corresponding tool node in the database is isolated.

[0095] Optionally, the step of matching the feature compensation vector with the capability description vectors corresponding to each of the plurality of candidate replacement nodes to determine the target replacement node includes: For each of the plurality of candidate replacement nodes, the capability description vector corresponding to the candidate replacement node is subjected to an inner product operation with the feature compensation vector to obtain the matching value corresponding to the candidate replacement node. Based on the matching values ​​corresponding to each of the multiple candidate replacement nodes, the multiple candidate replacement nodes are sorted. The candidate replacement node that ranks first in the matching value is determined as the target replacement node.

[0096] Preferably, before matching the feature compensation vector with the capability description vectors corresponding to the plurality of candidate replacement nodes, the compensation vector cache record, the preset component library, and the topology breakpoint record in the link to be repaired are read first. The compensation vector cache record provides the feature compensation vector that has passed the compensation consistency check; the preset component library provides multiple candidate replacement nodes, node registration information, interface description information, historical access records, and execution feedback records; and the topology breakpoint record provides the output-side breakpoint location, data acceptance field record, and breakpoint direction identifier of the first tool node. Then, based on the data acceptance field record in the topology breakpoint record, candidate replacement nodes that can read the corresponding field type or accept the corresponding field type through the input interface are selected from the preset component library, and the selected candidate replacement nodes are written into the candidate replacement node record. The candidate replacement node record is used to store candidate replacement nodes that have a field connection relationship with the data connection field record in the topology break record. The candidate replacement node record continues to participate in capability description vector reading and matching processing, so that the matching processing revolves around the field connection requirements that need to be filled at the topology break, rather than performing an indiscriminate search on all components in the preset component library.

[0097] Preferably, the capability description vectors corresponding to each of the multiple candidate replacement nodes are pre-generated based on node registration information, interface description information, historical access records, and execution feedback records in a preset component library. Specifically, firstly, the candidate replacement node identifier, the resource types that the candidate replacement node can call, and the corresponding execution entry point of the candidate replacement node are read from the node registration information; then, the input interface types that the candidate replacement node can receive, the output result types, the field types that can be processed, the time sequence range that can be accepted, and the result output direction are read from the interface description information; subsequently, combined with the historical access records and execution feedback records, the acceptance status of the candidate replacement node for the same field type, the same interface type, the same result output type, and the same result output direction in the historical execution cycle is sorted out to form the capability description vectors corresponding to the candidate replacement nodes. The technical essence of the capability description vector is a structured vector data used to describe the callable capabilities of the candidate replacement node in terms of content processing, interface acceptance, time sequence acceptance, result feedback, and result output direction. The capability description vector is used for same-dimensional matching processing with the feature compensation vector.

[0098] Preferably, the capability description vector includes at least a content processing capability component, an input interface acceptance capability component, an output interface acceptance capability component, a timing acceptance capability component, a result feedback capability component, and a result output direction component; wherein, the content processing capability component is used to indicate which data fields the candidate replacement node can perform reading, conversion, querying, analysis, or result generation processing on; the input interface acceptance capability component is used to indicate what input data format the candidate replacement node can receive from the output-side breakpoint of the first tool node; the output interface acceptance capability component is used to indicate what output data format the candidate replacement node can provide when transmitting processing results to downstream link segments; the timing acceptance capability component is used to indicate the result reception timing and continued processing timing that the candidate replacement node can accept after accessing the topology breakpoint; the result feedback capability component is used to indicate what processing result the candidate replacement node can return to downstream link segments after completing replacement processing; and the result output direction component is used to indicate the data flow direction in which the candidate replacement node can transmit the processing result to downstream link segments after completing replacement processing. All of the above components are bound to the candidate replacement node identifier, so that the capability description vector can not only express the callable capabilities of the candidate replacement node, but also backtrack to the specific candidate replacement node in the preset component library after the matching process.

[0099] Preferably, before performing the inner product operation between the capability description vector corresponding to the candidate replacement node and the feature compensation vector, vector field alignment processing is first performed on the feature compensation vector and the capability description vector. The vector field alignment processing first reads the content compensation component, interface compensation component, timing compensation component, and connection direction component from the feature compensation vector, and reads the content processing capability component, input interface acceptance capability component, output interface acceptance capability component, timing acceptance capability component, result feedback capability component, and result output direction component from the capability description vector. Then, according to the data acceptance field record in the topology break record, the content compensation component and the content processing capability component are arranged to the same component position; the interface compensation component is arranged to the same component position as the input interface acceptance capability component and the output interface acceptance capability component, respectively; the timing compensation component is arranged to the same component position as the timing acceptance capability component and the result feedback capability component. Subsequently, according to the break direction identifier, the connection direction component is matched with the running entry point and the result output direction component of the candidate replacement node to form a matching field correspondence record. The matching field corresponds to the record used to limit the relationship of the same component in the subsequent inner product operation, so as to avoid the mixed calculation of components with different technical meanings, such as content processing capability components and timing compensation components.

[0100] Preferably, for each of the plurality of candidate replacement nodes, when performing an inner product operation between the capability description vector corresponding to the candidate replacement node and the feature compensation vector, the corresponding record of the matching field is first read, and each compensation component in the feature compensation vector and each capability component in the capability description vector are read in the same position according to the component order in the corresponding record of the matching field; then, the compensation component and capability component read in the same position are quantized on the same scale, so that the compensation component represents the required strength of the topological break, and the capability component represents the capability strength of the candidate replacement node to cover the corresponding compensation requirement; subsequently, the correspondence degree calculation is performed on each group of corresponding components, and the results of multiple correspondence degree calculations are summarized according to the component order of the corresponding record of the matching field to obtain the matching value corresponding to the candidate replacement node. The technical essence of the inner product operation is to evaluate the correspondence between the capability coverage strength of the candidate replacement node and the compensation requirement strength of the topological break at the same component position, rather than performing unconstrained numerical superposition of two vectors from different sources.

[0101] Preferably, the technical essence of the matching value is a quantitative matching result used to characterize the degree of coverage of the candidate replacement node's ability to compensate for topological breaks. Specifically, the matching values ​​are formed by content matching sub-values, interface matching sub-values, timing matching sub-values, and connection direction matching sub-values. The content matching sub-values ​​are derived from the correspondence between the content compensation component and the content processing capability component, and are used to indicate whether the candidate replacement node can make up for the data content difference between the output data of the first tool node and the expected state characteristics. The interface matching sub-values ​​are derived from the correspondence between the interface compensation component and the input interface carrying capacity component and the output interface carrying capacity component, and are used to indicate whether the candidate replacement node can take over the output side break position of the first tool node and output identifiable results to the downstream link segment. The timing matching sub-values ​​are derived from the correspondence between the timing compensation component and the timing carrying capacity component and the result feedback capacity component, and are used to indicate whether the candidate replacement node can meet the timing requirements for the continued processing of the link to be repaired after access. The connection direction matching sub-values ​​are derived from the correspondence between the connection direction component and the running entry and result output direction components of the candidate replacement node, and are used to indicate whether the access direction of the candidate replacement node is consistent with the break direction identifier in the topology break record.

[0102] Preferably, after obtaining the matching value corresponding to each candidate replacement node, a validity verification process is performed on the matching value. The validity verification process first reads the matching value, candidate replacement node identifier, capability description vector, and corresponding record of the matching field corresponding to the candidate replacement node; then it verifies whether the matching value simultaneously includes content matching sub-values, interface matching sub-values, time sequence matching sub-values, and continuity direction matching sub-values, and verifies whether the content matching sub-values, interface matching sub-values, time sequence matching sub-values, and continuity direction matching sub-values ​​can all be traced back to the corresponding capability components in the capability description vector; when a matching value lacks a corresponding matching sub-value, the candidate replacement node corresponding to that matching value is marked as a candidate replacement node to be reviewed; when a matching value can be completely traced back to the corresponding capability component in the capability description vector, the matching value is written into the matching value record. The matching value record is used to store matching values ​​that can participate in the sorting process, enabling subsequent sorting processes to be performed based on matching values ​​with complete field sources.

[0103] Preferably, when sorting the candidate replacement nodes based on their respective matching values, the process first reads the matching value record and initially arranges the candidate replacement nodes in descending order of matching value to form an initial sorting record. Then, it reads the data transmission field record and the breakpoint direction identifier from the topology breakpoint record, and performs a breakpoint adaptation verification on candidate replacement nodes with the same or similar matching values ​​in the initial sorting record. The breakpoint adaptation verification determines whether a candidate replacement node is suitable for accessing the output-side breakpoint position of the first tool node and continuing to transmit the processing result downstream based on the interface matching sub-value, timing matching sub-value, and connection direction matching sub-value. Subsequently, the initial sorting record of candidate replacement nodes is adjusted according to the breakpoint adaptation verification result to form a final sorting record. The candidate replacement node sorting record represents the access priority relationship of multiple candidate replacement nodes relative to the same topology breakpoint, rather than simply representing the static arrangement relationship of multiple candidate replacement nodes in a preset component library.

[0104] Preferably, when determining the candidate replacement node with the highest matching value as the target replacement node, the candidate replacement node sorting record is first read, and the identifier of the candidate replacement node at the top of the sorting record is extracted. Then, based on the candidate replacement node identifier, the corresponding candidate replacement node, the capability description vector corresponding to the candidate replacement node, the input interface description information corresponding to the candidate replacement node, and the output interface description information corresponding to the candidate replacement node are read from a preset component library. Subsequently, a correspondence is established between the candidate replacement node and the topology break record, the feature compensation vector, and the matching value to form a target replacement node record, and the candidate replacement nodes in the target replacement node record are determined as the target replacement nodes. The technical essence of the target replacement node is that it is a tool node among multiple candidate replacement nodes that has a high degree of matching with the compensation requirements of the feature compensation vector and can access the topology break to replace the failed tool node to complete the break compensation processing. The target replacement node continues to participate in the subsequent access to the topology break and input interface feature alignment processing.

[0105] Preferably, the target filler node record includes a target filler node identifier, a capability description vector corresponding to the target filler node, a matching value, a content matching sub-value, an interface matching sub-value, a timing matching sub-value, a connection direction matching sub-value, input interface description information, output interface description information, and a topology break record. The target filler node identifier is used to locate the target filler node in a preset component library; the capability description vector corresponding to the target filler node is used to indicate the capability source from which the target filler node is identified; the matching value is used to indicate the overall correspondence between the target filler node and the feature compensation vector; the input interface description information and the output interface description information are used for subsequent feature alignment processing of the input interface of the target filler node; and the topology break record is used to connect the target filler node to the output-side break position of the first tool node. By writing the target replacement node identifier, the capability description vector corresponding to the target replacement node, the matching value, the content matching sub-value, the interface matching sub-value, the timing matching sub-value, the connection direction matching sub-value, the input interface description information, the output interface description information, and the topology break record into the target replacement node record, the determination result of the target replacement node can be directly entered into the subsequent adaptive toolchain generation process.

[0106] Preferably, after determining the target replacement node, a pre-access consistency verification process is performed on the target replacement node. The pre-access consistency verification process first reads the target replacement node record, the topology break record, and the feature compensation vector; then it verifies whether the input interface description information in the target replacement node record can read the data transmission field record in the topology break record, whether the output interface description information in the target replacement node record can provide the output field type required by the expected state characteristics to the downstream link segment, and whether the connection direction matching sub-value in the target replacement node record corresponds to the break direction identifier in the topology break record; when all the above correspondences are satisfied, the target replacement node is marked as accessible. The accessible state indicates that the target replacement node has the conditions to enter the subsequent feature alignment process, enabling the target replacement node to complete the corresponding verification of capability source, interface source, and break direction source before accessing the topology break.

[0107] Preferably, in a scenario, the target agent sequentially calls a large language processing model, knowledge base, database, toolset, and environmental perception node within the current execution cycle to complete the park anomaly monitoring and analysis task. When the tool node corresponding to the database is determined to be a failed tool node and forms a topological break, the feature compensation vector is used to represent the data reading capability, input interface acceptance capability, output interface acceptance capability, and result feedback capability that the target replacement node needs to supplement. At this time, multiple candidate replacement nodes capable of performing data query, structured field reading, or cached result retrieval are first read from the preset component library, and the capability description vectors corresponding to each candidate replacement node are read. Then, the capability description vector corresponding to each candidate replacement node is processed by the inner product operation with the feature compensation vector to obtain the matching value corresponding to each candidate replacement node. Subsequently, the multiple candidate replacement nodes are sorted according to the matching value, and the candidate replacement node with the first matching value and passing the pre-access consistency verification is determined as the target replacement node. Thus, the target replacement node can be selected according to the data acceptance requirements at the topological break and continue to participate in the target replacement node access to the topological break and input interface feature alignment processing.

[0108] Optionally, the step of performing feature alignment processing on the input interface of the target filler node to generate an adaptive toolchain includes: Extract the input data format corresponding to the target padding node; Extract the output data format corresponding to the breakpoint output features; Determine whether the input data format and the output data format are consistent; If it is determined that the input data format and the output data format are inconsistent, a preset data conversion middleware node is inserted between the first tool node and the target padding node; The breakpoint output feature is format-converted by the data conversion middleware node to obtain an alignment feature that is consistent with the input data format. Based on the alignment feature, a connection path is established from the first tool node to the target filler node via the data conversion middleware node, thereby generating the adaptive toolchain.

[0109] Preferably, before performing feature alignment processing on the input interface of the target filler node, the target filler node record, topology break record, the most recent result feedback record of the first tool node, the compensation vector cache record, and the preset component library are read first. The target filler node record provides the target filler node identifier, the input interface description information of the target filler node, the output interface description information of the target filler node, and the capability description vector corresponding to the target filler node. The topology break record provides the output-side break location, data transmission field record, and break direction identifier of the first tool node. The most recent result feedback record of the first tool node provides the source data of the breakpoint output features. The compensation vector cache record provides the feature compensation vector that has passed the compensation consistency check. The preset component library provides the node registration information corresponding to the target filler node, the interface description information corresponding to the target filler node, and the data conversion middleware node configuration record. By using the target filler node record, topology break record, the most recent result feedback record of the first tool node, compensation vector cache record, and the preset component library as input sources, subsequent input data format extraction, output data format extraction, format consistency judgment, data conversion middleware node insertion, and connection path establishment can all revolve around the same topology break. This allows the input interface of the target filler node and the output break position of the first tool node to form a continuous data connection relationship through the input data format, output data format, and data conversion middleware node.

[0110] Preferably, when extracting the input data format corresponding to the target filler node, firstly, the input interface description information of the target filler node is read from the target filler node record according to the target filler node identifier, and the interface description information corresponding to the target filler node is read from a preset component library according to the target filler node identifier; then, the input field name, input field type, input field arrangement order, input field encapsulation method, input field mandatory status, and input field parsing method are extracted from the input interface description information and the interface description information corresponding to the target filler node; subsequently, the input field name, input field type, input field arrangement order, input field encapsulation method, input field mandatory status, and input field parsing method are organized into an input data format. The technical essence of the input data format is the description of the input-side data structure that the target filler node can read and parse after accessing the topology breakpoint. The input data format is used to limit the target filler node's requirements for receiving the field name, field type, field arrangement order, field encapsulation method, input field mandatory status, and input field parsing method of the breakpoint output features.

[0111] Preferably, the input data format does not simply represent the file type format tags that the target complement node can receive, but rather the interface acceptance rules that the target complement node must satisfy when receiving breakpoint output features. Specifically, the input field name represents the data field that the target complement node's input interface can recognize; the input field type represents the data type that the target complement node's input interface can parse; the input field order represents the arrangement relationship when multiple input fields enter the target complement node's input interface; the input field encapsulation method represents the encapsulation structure when multiple input fields are organized into an input payload; the mandatory input field status represents the fields that the target complement node needs to read when completing the complement processing; and the input field parsing method represents the processing method by which the target complement node decomposes the input payload into individual input fields. All of the above constitute the input data format, enabling subsequent format consistency judgment to be performed item by item on the input field name, input field type, input field order, input field encapsulation method, mandatory input field status, and input field parsing method.

[0112] Preferably, when extracting the output data format corresponding to the breakpoint output features, the breakpoint output features are first read from the most recent result feedback record of the first tool node based on the data succession field record in the topology break record; then, the output field name, output field type, output field order, output field encapsulation method, output field availability status, and output field source identifier are extracted from the breakpoint output features; subsequently, the output field name, output field type, output field order, output field encapsulation method, output field availability status, and output field source identifier are organized into an output data format. The technical essence of the output data format is a description of the output-side data structure that has been formed by the first tool node upstream of the topology break and can be transmitted to the target replacement node. The output data format is used to define the field organization form of the breakpoint output features at the output-side break position of the first tool node.

[0113] Preferably, there is a source correspondence between the output data format and the breakpoint output feature. The breakpoint output feature is used to express the upstream data state that the first tool node has formed before the failed tool node is isolated. The output data format is used to express the field presentation method of this upstream data state at the breakpoint position on the output side of the first tool node. The output field name is used to indicate the data field that can be read in the breakpoint output feature. The output field type is used to indicate the type of each data field in the breakpoint output feature. The output field arrangement order is used to indicate the output sequence of multiple data fields in the breakpoint output feature. The output field encapsulation method is used to indicate the organization method when the breakpoint output feature is written to the result feedback payload. The output field availability status is used to indicate whether the corresponding output field can enter the subsequent format conversion process. The output field source identifier is used to indicate that the corresponding output field comes from the most recent result feedback record of the first tool node. Through the above output data format, the subsequent data conversion middleware node can determine which output fields should be read from the breakpoint output feature and how to convert these output fields.

[0114] Preferably, when determining whether the input data format and the output data format are consistent, the following steps are taken: First, the input field name, input field type, input field order, input field encapsulation method, required input field status, and input field parsing method in the input data format are read. Then, the output field name, output field type, output field order, output field encapsulation method, available output field status, and output field source identifier in the output data format are read. Next, based on the data succession field record in the topology breakpoint record, the input field name is matched with the output field name; the input field type is matched with the output field type; the input field order is matched with the output field order; the input field encapsulation method is matched with the output field encapsulation method; and the required input field status is matched with the available output field status. Subsequently, a format consistency judgment record is generated based on the above correspondences. This format consistency judgment record represents the correspondence between the input data format and the output data format in terms of field name, field type, field order, field encapsulation method, and field availability. This format consistency judgment record continues to participate in the data conversion middleware node insertion process.

[0115] Preferably, the determination of whether the input data format and the output data format are consistent essentially determines whether the breakpoint output feature at the output-side breakpoint location of the first tool node can be directly read by the input interface of the target supplementary node. If the format consistency judgment record indicates that the input field name corresponds to the output field name, the input field type corresponds to the output field type, the input field arrangement order corresponds to the output field arrangement order, the input field encapsulation method corresponds to the output field encapsulation method, and the mandatory state of the input field corresponds to the available state of the output field, then it means that the breakpoint output feature can enter the target supplementary node according to the input interface requirements of the target supplementary node; if the format consistency judgment record indicates that at least one of the field name, field type, field arrangement order, field encapsulation method, or field available state cannot correspond, then it means that the breakpoint output feature needs to undergo format conversion processing first. Through the above format consistency judgment, the subsequent insertion of the data conversion middleware node is based on the field-level differences between the input data format and the output data format, rather than a general judgment on whether the target supplementary node can access the topology breakpoint.

[0116] Preferably, when it is determined that the input data format and the output data format are inconsistent, before inserting a preset data conversion middleware node between the first tool node and the target padding node, a format difference record is first generated based on the format consistency judgment record. The format difference record includes field name difference items, field type difference items, field arrangement order difference items, field encapsulation method difference items, and field mandatory status difference items. Specifically, the field name difference item indicates fields where the input field name and output field name do not correspond; the field type difference item indicates fields where the input field type and output field type do not correspond; the field arrangement order difference item indicates the arrangement difference between the input field arrangement order and the output field arrangement order; the field encapsulation method difference item indicates the encapsulation difference between the input field encapsulation method and the output field encapsulation method; and the field mandatory status difference item indicates input fields that the target padding node needs to read but which are not directly provided by the breakpoint output features. The format difference record continues to participate in the data conversion middleware node's reading, enabling the data conversion middleware node to perform the corresponding format conversion processing according to the format difference record.

[0117] Preferably, the preset data conversion middleware node is pre-configured based on the data conversion middleware node configuration record in the preset component library. Specifically, the convertible field types, field name mapping rules, field order rearrangement rules, field encapsulation method conversion rules, missing field completion rules, and conversion result verification rules are first pre-written into the data conversion middleware node configuration record; then, a correspondence is established between the convertible field types, field name mapping rules, field order rearrangement rules, field encapsulation method conversion rules, missing field completion rules, and conversion result verification rules and the input interface description information of the target completion node; subsequently, the data conversion middleware node capable of completing the conversion of corresponding format difference records is registered in the preset component library. The technical essence of the data conversion middleware node is that it is an intermediate data processing node located between the first tool node and the target completion node, used to convert the output data format of the breakpoint output features into the input data format required by the target completion node.

[0118] Preferably, when inserting a preset data conversion middleware node between the first tool node and the target filler node, the following steps are taken: First, based on the format difference record, data conversion middleware nodes capable of handling field name differences, field type differences, field order differences, field encapsulation method differences, and field mandatory status differences are read from a preset component library. Then, an input-side connection is established between the data conversion middleware node and the output-side breakpoint of the first tool node, and an output-side connection is established between the data conversion middleware node and the input interface of the target filler node. Subsequently, the output-side breakpoint of the first tool node, the input port of the data conversion middleware node, the output port of the data conversion middleware node, and the input interface of the target filler node are written into the middleware insertion record. The middleware insertion record indicates that the data conversion middleware node has been set between the first tool node and the target filler node, and the middleware insertion record continues to participate in format conversion processing and connection path establishment processing.

[0119] Preferably, when performing format conversion processing on the breakpoint output features through the data conversion middleware node, the breakpoint output features are first read from the breakpoint position on the output side of the first tool node according to the middleware insertion record, and the output field names, output field types, output field order, and output field encapsulation methods in the breakpoint output features are parsed according to the output data format. Then, field name mapping processing is performed according to the field name difference items in the format difference record, field type conversion processing is performed according to the field type difference items, field order reordering processing is performed according to the field order difference items, field encapsulation method conversion processing is performed according to the field encapsulation method difference items, and missing field completion processing is performed according to the field mandatory status difference items. Subsequently, the breakpoint output features after the above processing are organized into aligned features consistent with the input data format. The technical essence of the format conversion processing is to convert the field names, field types, field order, and field encapsulation methods of the breakpoint output features into an input data format that can be read by the input interface of the target padding node, while retaining the source and meaning of the breakpoint output feature data.

[0120] Preferably, the technical essence of the alignment feature is a structured input feature formed by the breakpoint output feature after format conversion processing by the data conversion middleware node, which can be read by the input interface of the target padding node. The alignment feature includes an alignment field name, alignment field type, alignment field arrangement order, alignment field encapsulation method, alignment field source identifier, and alignment field verification status corresponding to the input data format. Specifically, the alignment field name originates from field name mapping processing, the alignment field type from field type conversion processing, the alignment field arrangement order from field arrangement order rearrangement processing, the alignment field encapsulation method from field encapsulation method conversion processing, the alignment field source identifier is used to retain the correspondence between the output field source identifier in the breakpoint output feature and the first tool node, and the alignment field verification status indicates whether the alignment feature meets the input data format corresponding to the target padding node. By setting the alignment field source identifier, the alignment feature can still trace back to the most recent result feedback record of the first tool node after being read by the target padding node.

[0121] Preferably, after obtaining the alignment feature consistent with the input data format, an alignment consistency verification process is performed on the alignment feature. The alignment consistency verification process first reads the alignment field name, alignment field type, alignment field order, alignment field encapsulation method, and alignment field verification status from the alignment feature, and then reads the input field name, input field type, input field order, input field encapsulation method, and input field mandatory status from the input data format. Next, it determines whether the alignment field name corresponds to the input field name, whether the alignment field type corresponds to the input field type, whether the alignment field order corresponds to the input field order, whether the alignment field encapsulation method corresponds to the input field encapsulation method, and whether all fields corresponding to the mandatory status of the input field exist in the alignment feature. When all the above correspondences are satisfied, the alignment feature is written to the alignment feature cache record. The alignment feature cache record is used to store the alignment features of the input interface that can enter the target padding node, so that subsequent connection path establishment can directly read the alignment features that have passed the alignment consistency verification process.

[0122] Preferably, when establishing the connection path from the first tool node to the target complement node via the data conversion middleware node based on the alignment features, the topology break record, middleware insertion record, and alignment feature cache record are read first. Then, based on the output-side break position of the first tool node in the topology break record, a first connection relationship is established between the first tool node and the data conversion middleware node, enabling the breakpoint output feature to be transmitted from the first tool node to the data conversion middleware node. Subsequently, based on the output port of the data conversion middleware node and the input interface of the target complement node in the middleware insertion record, a second connection relationship is established between the data conversion middleware node and the target complement node, enabling the alignment feature to be transmitted from the data conversion middleware node to the target complement node. Finally, the first and second connection relationships are written into the connection path record. The connection path record is used to represent the data flow order between the first tool node, the data conversion middleware node, and the target complement node, and the connection path record continues to participate in the adaptive toolchain generation process.

[0123] Preferably, the connection path from the first tool node to the target replacement node via the data conversion middleware node is not a normal node connection, but a data flow path carrying data format conversion relationships and field source relationships. The connection path record includes the tool node identifier of the first tool node, the output-side break point location of the first tool node, the node identifier of the data conversion middleware node, the input port of the data conversion middleware node, the output port of the data conversion middleware node, the identifier of the target replacement node, the input interface of the target replacement node, an alignment feature cache record, and a break point direction identifier. The alignment feature cache record provides the data content transmitted along the connection path, and the break point direction identifier specifies that the connection path should point from the first tool node to the data conversion middleware node and continue to the target replacement node. Through the above connection path record, after the target replacement node accesses the topology break point, it can read the alignment features along the original direction of the link to be repaired, rather than operating independently without accessing the topology break point.

[0124] Preferably, when generating the adaptive toolchain, the following steps are taken: first, read the link to be repaired, the connection path record, the target replacement node record, and the isolation status record of the failed tool node; then, retain the upstream link fragment, downstream link fragment, inactive tool node, and topology break record in the link to be repaired, and write the first and second connection relationships from the connection path record into the active connection edge record of the link to be repaired; subsequently, insert the target replacement node from the target replacement node record into the break position corresponding to the topology break record, and write the data conversion middleware node as the data format conversion node between the first tool node and the target replacement node into the active connection edge record; finally, form the adaptive toolchain based on the updated active connection edge record, target replacement node record, data conversion middleware node, alignment feature cache record, and downstream link fragment. The technical essence of the adaptive toolchain is that, after deleting the data flow impact of the failed tool node on the link to be repaired, a continuing executable tool call link is formed by reconnecting the topology break through the target replacement node and the data conversion middleware node.

[0125] Preferably, the adaptive toolchain includes an upstream link segment, a first tool node, a data conversion middleware node, a target replacement node, a downstream link segment, a connection path record, and an alignment feature cache record. The upstream link segment retains usable processing results generated before the failed tool node was isolated. The first tool node provides breakpoint output features. The data conversion middleware node converts the breakpoint output features into alignment features. The target replacement node performs breakpoint compensation processing based on the alignment features. The downstream link segment continues to receive processing results output by the target replacement node. The connection path record defines the data flow order between the first tool node, the data conversion middleware node, and the target replacement node. The alignment feature cache record provides input data that the target replacement node can read. Through this composition, the adaptive toolchain can, while retaining usable upstream processing results in the link to be repaired, connect the target replacement node to the topology breakpoint and continue transmitting processing results to the downstream link segment.

[0126] Preferably, after generating the adaptive toolchain, a link executability verification process is performed on the adaptive toolchain. The link executability verification process first reads the active connection edge record, connection path record, alignment feature cache record, target complement node record, and downstream link fragment in the adaptive toolchain; then it verifies whether the first connection relationship from the first tool node to the data conversion middleware node can transmit breakpoint output features, whether the second connection relationship from the data conversion middleware node to the target complement node can transmit alignment features, and whether the output interface description information of the target complement node corresponds to the input-side connection position of the downstream link fragment; when all the above correspondences are satisfied, the adaptive toolchain is marked as being ready to continue execution. The ready-to-continue execution state indicates that the adaptive toolchain has completed topology breakpoint connection, format conversion, and target complement node access, and the ready-to-continue execution state continues to participate in the target agent's execution unit's continued processing of the adaptive toolchain.

[0127] Preferably, in a scenario, the target agent sequentially calls the large language processing model, knowledge base, database, toolset, and environmental perception node to complete the park anomaly monitoring and analysis task within the current execution cycle. When the tool node corresponding to the database is determined to be a failed tool node and forms a topological break, the first tool node can be the tool node corresponding to the knowledge base, and the target replacement node can be a candidate replacement node capable of performing cached result retrieval or structured field reading. At this time, the input data format corresponding to the target replacement node is first extracted from the target replacement node record, and the output data format is extracted from the breakpoint output features of the tool node corresponding to the knowledge base; then, the consistency between the input data format and the output data format is judged. When the input data format and the output data format are inconsistent, a data conversion middleware node is inserted between the tool node corresponding to the knowledge base and the target replacement node. The breakpoint output features are converted into alignment features consistent with the input data format through the data conversion middleware node; then, a connection path is established from the tool node corresponding to the knowledge base to the target replacement node through the data conversion middleware node based on the alignment features, and the connection path is written into the active connection edge record of the link to be repaired to generate an adaptive toolchain. This adaptive toolchain can take over the upstream data state that has been formed by the tool nodes corresponding to the knowledge base, and enable the target complement node to read the alignment features according to the input data format and continue to pass the processing results to the toolset and environment awareness nodes.

[0128] Optionally, before inserting a preset data conversion middleware node between the first tool node and the target padding node when it is determined that the input data format and the output data format are inconsistent, the method further includes: If it is determined that the input data format is consistent with the output data format, a direct connection path is established from the first tool node to the target padding node; Perform connectivity verification on the topology containing the directly connected paths; After the connectivity verification process passes, the topology containing the direct connection path is determined as the adaptive toolchain.

[0129] Preferably, before establishing a direct connection path from the first tool node to the target replacement node after determining that the input data format and the output data format are consistent, the following steps are taken: First, the link to be repaired, the topology break record, the format consistency judgment record, the target replacement node record, and the most recent result feedback record of the first tool node are read. Specifically, the link to be repaired provides upstream link segments, downstream link segments, and active connection edge records after the isolated and failed tool node; the topology break record provides the output-side break location, data receiving field record, and break direction identifier of the first tool node; the format consistency judgment record provides the corresponding status of the input data format and output data format in terms of input field name and output field name, input field type and output field type, input field order and output field order, input field encapsulation method and output field encapsulation method, and input field mandatory status and output field available status; the target replacement node record provides the target replacement node identifier, the target replacement node's input interface, and the target replacement node's output interface; and the most recent result feedback record of the first tool node provides breakpoint output characteristics. The establishment of the direct connection path is based on the topological location of the link to be repaired, the data acceptance requirements of the topological break record, the format correspondence result of the format consistency judgment record, and the interface location of the target replacement node record. This allows the direct connection path to connect the first tool node and the target replacement node when the input data format and the output data format are already consistent.

[0130] Preferably, the consistency between the input data format and the output data format means that the input field names in the format consistency judgment record correspond to the output field names, the input field types correspond to the output field types, the input field order corresponds to the output field order, the input field encapsulation method corresponds to the output field encapsulation method, and the mandatory state of the input fields corresponds to the available state of the output fields. When the format consistency judgment record indicates that all the above correspondences are satisfied, the breakpoint output feature can be output from the output-side breakpoint position of the first tool node according to the output data format, and can be read by the input interface of the target filler node according to the input data format. In this case, the breakpoint output feature does not need to go through the data conversion middleware node for field name mapping, field type conversion, field order rearrangement, field encapsulation method conversion, or missing field completion processing; the direct connection path can directly undertake the data flow task of transferring the breakpoint output feature from the first tool node to the target filler node.

[0131] Preferably, when establishing a direct connection path from the first tool node to the target replacement node, the starting connection end of the direct connection path is first determined based on the output-side break position of the first tool node in the topology break record; then, the ending connection end of the direct connection path is determined based on the input interface of the target replacement node in the target replacement node record; subsequently, based on the data receiving field record in the topology break record, the output field name, output field type, output field arrangement order, and output field encapsulation method in the breakpoint output feature are bound to the data receiving field position of the direct connection path, and the data flow direction of the direct connection path is determined based on the break direction identifier in the topology break record. The technical essence of the direct connection path is that, when the input data format and output data format are consistent, the data flow connection relationship directly points from the output-side break position of the first tool node to the input interface of the target replacement node. The direct connection path is used to allow the breakpoint output feature to enter the target replacement node along the original data flow direction of the link to be repaired.

[0132] Preferably, after the direct connection path is established, the tool node identifier of the first tool node, the output-side breakpoint location of the first tool node, the target replacement node identifier, the input interface of the target replacement node, the data receiving field record, the breakpoint direction identifier, the breakpoint output feature source identifier, and the format consistency judgment record are written into the direct connection path record. The direct connection path record indicates that a data flow relationship has been formed between the first tool node and the target replacement node without going through a data conversion middleware node; wherein, the breakpoint output feature source identifier points to the most recent result feedback record of the first tool node, the format consistency judgment record indicates the basis for judging that the direct connection path does not need to perform format conversion processing, and the data receiving field record limits the range of fields transmitted along the direct connection path. Through the direct connection path record, subsequent topologies containing the direct connection path can read the port source, field source, and direction source of the direct connection path, instead of only storing the abstract connection result between the first tool node and the target replacement node.

[0133] Preferably, when determining the topology containing the direct connection path, firstly, the upstream link segment, downstream link segment, inactive tool node, topology break record, and active connection edge record in the link to be repaired are read; then, the direct connection path record is written into the active connection edge record, so that the output-side break position of the first tool node forms a directed connection relationship between the nodes through the direct connection path and the input interface of the target replacement node; subsequently, the target replacement node in the target replacement node record is inserted into the break position corresponding to the topology break record, and the failed tool node is kept in an inactive tool node state in the link to be repaired, so as to form a topology containing the direct connection path. The technical essence of the topology containing the direct connection path is that, based on the link to be repaired, a topological data structure is formed after replacing the break connection relationship corresponding to the first input connection edge with the direct connection path. The topology containing the direct connection path retains the data flow relationship between the upstream link segment, the first tool node, the target replacement node, the downstream link segment, and the direct connection path record.

[0134] Preferably, when performing connectivity verification on the topology containing the direct connection path, the active connection edge record, direct connection path record, topology break record, target replacement node record, and downstream link segment in the topology containing the direct connection path are first read; then, it is verified whether the starting connection end in the direct connection path record corresponds to the output-side break position of the first tool node, and whether the terminating connection end in the direct connection path record corresponds to the input interface of the target replacement node; subsequently, it is verified whether the break direction indicator indicates that data flows from the first tool node to the target replacement node, and whether there is still a first input connection edge pointing to the failed tool node or a second output connection edge issued by the failed tool node in the active connection edge record. The above connectivity verification process is used to determine whether the topology containing the direct connection path has established a data continuity relationship from the upstream link segment to the target replacement node via the direct connection path after isolating the failed tool node.

[0135] Preferably, the connectivity verification process also verifies the readability of the breakpoint output features entering the target replacement node along the direct connection path. Specifically, firstly, the data receiving field record and format consistency judgment record are read from the direct connection path record, and the breakpoint output features are read from the most recent result feedback record of the first tool node; then, based on the data receiving field record, it is verified whether the output fields involved in the transmission in the breakpoint output features correspond to the input fields required by the input interface of the target replacement node, and based on the format consistency judgment record, it is verified whether the input field name, input field type, input field arrangement order, input field encapsulation method, and input field mandatory status still correspond to the output field name, output field type, output field arrangement order, output field encapsulation method, and output field available status; subsequently, it is verified whether the output interface of the target replacement node can form a result transmission relationship with the input side continuation position of the downstream link segment. Through this connectivity verification process, the topology containing the direct connection path is not only continuous in topological connection, but also continuous in the field receiving of the breakpoint output features.

[0136] Preferably, the criteria for determining whether connectivity verification passes include topology port correspondence, connection direction correspondence, field acceptance correspondence, and downstream continuation correspondence. Topology port correspondence indicates that the output-side break point of the first tool node has been connected to the input interface of the target complement node via a direct connection path; connection direction correspondence indicates that the data flow direction of the direct connection path is consistent with the break point direction identifier in the topology break point record; field acceptance correspondence indicates that the output data format of the breakpoint output characteristics can be read by the target complement node according to the input data format; downstream continuation correspondence indicates that the output interface of the target complement node can continue to transmit processing results to downstream link segments. When topology port correspondence, connection direction correspondence, field acceptance correspondence, and downstream continuation correspondence are all satisfied, the topology structure containing the direct connection path is marked as having passed connectivity verification; the connectivity verification passed state continues to participate in the determination process of the adaptive toolchain.

[0137] Preferably, after the connectivity verification process passes, when determining the topology containing the direct connection path as the adaptive toolchain, the topology containing the direct connection path that is in a connectivity verification passed state is first read, and the upstream link segment, first tool node, direct connection path record, target replacement node, downstream link segment, and active connection edge record in the topology containing the direct connection path are read; then, the direct connection path record is used as the connection basis for the target replacement node to access the topology breakpoint, and the direct connection path in the active connection edge record is retained as the directed connection relationship between currently executable nodes; subsequently, the upstream link segment, first tool node, direct connection path record, target replacement node, downstream link segment, and active connection edge record are jointly written into the adaptive toolchain record to determine the topology containing the direct connection path as the adaptive toolchain. The adaptive toolchain record is used to express the node composition, connection relationship, field inheritance relationship, and execution direction of the adaptive toolchain in the direct connection case.

[0138] Preferably, the adaptive toolchain formed when the input and output data formats are consistent has a different data continuation method than the adaptive toolchain formed after inserting a data conversion middleware node. In the direct connection case, the adaptive toolchain includes an upstream link segment, a first tool node, a direct connection path record, a target complement node, and a downstream link segment. The first tool node provides breakpoint output features, the direct connection path record defines the port relationships and field inheritance relationships from the first tool node to the target complement node, the target complement node performs breakpoint compensation processing based on the breakpoint output features, and the downstream link segment continues to receive the processing results output by the target complement node. Since the direct connection path record has already passed format consistency judgment and connectivity verification, the adaptive toolchain can continue to inherit the upstream data state of the repair link without inserting a data conversion middleware node.

[0139] Preferably, after determining the adaptive toolchain, a direct connection status recording process is also performed on the adaptive toolchain. The direct connection status recording process first reads the adaptive toolchain record, the direct connection path record, and the connectivity verification pass status; then, it writes the output-side break point position of the first tool node, the input interface of the target replacement node, the data inheritance field record, the break point direction identifier, and the format consistency judgment record from the direct connection path record into the direct connection status record; subsequently, it establishes a correspondence between the direct connection status record and the target replacement node record. The direct connection status record indicates that the adaptive toolchain uses a direct connection path to connect topology break points, and that the breakpoint output features can enter the input interface of the target replacement node in the output data format. The direct connection status record continues to participate in the target agent's execution unit's continued processing of the adaptive toolchain, enabling the target agent's execution unit to read the direct connection path and execute subsequent tool calls based on the direct connection status record.

[0140] Preferably, in a scenario, the target agent sequentially calls the large language processing model, knowledge base, database, toolset, and environmental awareness node within the current execution cycle to complete the park anomaly monitoring and analysis task. When the tool node corresponding to the database is determined to be a failed tool node and forms a topological break, the first tool node is the tool node corresponding to the knowledge base, and the target replacement node is a candidate replacement node that can directly read the output fields of the tool node corresponding to the knowledge base. At this time, if the format consistency judgment record indicates that the output data format formed by the tool node corresponding to the knowledge base is consistent with the input data format corresponding to the target replacement node, a direct connection path is established from the tool node corresponding to the knowledge base to the target replacement node, and the direct connection path is written into the active connection edge record of the link to be repaired to obtain a topology containing the direct connection path; then, connectivity verification processing is performed on the topology containing the direct connection path. After the connectivity verification processing passes, the topology containing the direct connection path is determined as the adaptive toolchain. This adaptive toolchain can enable the breakpoint output features of the tool node corresponding to the knowledge base to directly enter the target replacement node without inserting a data conversion middleware node, and continue to transmit the processing results to the toolset and environmental awareness node.

[0141] Optionally, the method further includes: Obtain the feedback execution characteristics generated by the execution unit when processing the adaptive toolchain; Calculate the feature similarity between the feedback execution feature and the expected state feature; If the feature similarity is less than a preset similarity threshold, the feature compensation vector is updated based on the feature similarity. The updated feature compensation vector is associated with the abnormal node identifier; The data after association processing is stored in a pre-defined experience knowledge base.

[0142] Preferably, after the adaptive toolchain is input to the execution unit of the target agent for further processing, before obtaining the feedback execution features generated by the execution unit when processing the adaptive toolchain, the adaptive toolchain record, the target filler node record, the connection relationship record corresponding to the adaptive toolchain, the alignment feature cache record, and the direct continuation state record are read first. The adaptive toolchain record provides records of upstream link segments, first tool nodes, target filler nodes, downstream link segments, and active connection edges in the adaptive toolchain. The target filler node record provides the target filler node identifier, the input interface of the target filler node, and the output interface of the target filler node. The connection relationship record corresponding to the adaptive toolchain includes a connection path record or a direct connection path record. The connection path record provides the data flow relationship when accessed via a data conversion middleware node, and the direct connection path record provides the data flow relationship when no data conversion middleware node is inserted. The alignment feature cache record provides the alignment features that the target filler node can read. The direct continuation state record provides the continuation state of the breakpoint output features entering the input interface of the target filler node in the output data format. Through the above reading method, the data generated by the execution unit when processing the adaptive toolchain can establish a correspondence with the node composition of the adaptive toolchain, the connection relationship record corresponding to the adaptive toolchain, the input source and the output destination, so that the subsequent feedback execution features can express the actual execution status of the adaptive toolchain in the process of continued processing.

[0143] Preferably, when acquiring the feedback execution features generated by the execution unit during the processing of the adaptive toolchain, the following steps are taken: First, the target filler node is triggered to read the breakpoint output feature or alignment feature according to the adaptive toolchain record, and an input reading feedback record is generated when the target filler node reads the breakpoint output feature or alignment feature; then, the target filler node result feedback record is read according to the output interface of the target filler node after the target filler node completes the breakpoint compensation processing, and the downstream receiving feedback record is read according to the downstream link segment after receiving the target filler node result feedback record; subsequently, the node call status, interface response status, result feedback timestamp, and exception return status during the continued processing of the adaptive toolchain are read according to the active connection edge record, and the input reading feedback record, target filler node result feedback record, downstream receiving feedback record, node call status, interface response status, result feedback timestamp, and exception return status are organized into feedback execution features. The technical essence of the feedback execution feature is the actual running state feature formed after the execution unit continues to process the adaptive toolchain. The feedback execution feature is used to indicate whether the target filler node can read the breakpoint output feature or alignment feature after accessing the topology breakpoint, complete the breakpoint compensation process, and transmit the processing result to the downstream link segment.

[0144] Preferably, the feedback execution feature is not a simple success or failure marker, but rather structured feedback data bound to the node position in the adaptive toolchain, the connection relationship record corresponding to the adaptive toolchain, and the field inheritance relationship. Specifically, the input read feedback record indicates whether the target filler node has read the breakpoint output feature or alignment feature according to the input interface of the target filler node; the target filler node result return record indicates the output field name, output field type, output field arrangement order, and output field encapsulation method formed after the target filler node performs breakpoint compensation processing on the breakpoint output feature or alignment feature; the downstream receive feedback record indicates whether the downstream link segment can read the target filler node result return record; the node call status indicates whether each tool node in the adaptive toolchain continues to run according to the active connection edge record; the interface response status indicates the response status of the input interface and output interface of the target filler node during the continued processing; the result return timestamp indicates the time position of the target filler node's output processing result; and the exception return status indicates whether a new exception return occurs during the continued processing of the adaptive toolchain. The above elements together constitute the feedback execution feature, enabling the feedback execution feature to continue participating in the feature similarity calculation of the expected state feature.

[0145] Preferably, before calculating the feature similarity between the feedback execution feature and the expected state feature, the task state reference record is read first, and the expected state feature corresponding to the target agent's current task is read from the task state reference record. The expected state feature is used to represent the data state reference required when the target agent's current task continues to flow after the topological breakpoint. The expected state feature includes the expected field value range, expected field type, expected interface state, and expected result state. Subsequently, based on the data succession field record in the topological breakpoint record, the target replacement node result feedback record, downstream received feedback record, interface response state, and result feedback timestamp in the feedback execution feature are respectively matched with the expected field value range, expected field type, expected interface state, and expected result state in the expected state feature to form a feedback field correspondence record. The feedback field correspondence record is used to limit the same field comparison relationship between the feedback execution feature and the expected state feature, so that the feature similarity calculation can be carried out around the same field type, the same interface state, and the same result state.

[0146] Preferably, when calculating the feature similarity between the feedback execution feature and the expected state feature, firstly, based on the records corresponding to the feedback fields, the actual field value corresponding to the target padding node result return record in the feedback execution feature is compared with the expected field value range in the expected state feature to form a content similarity component; then, the actual field type corresponding to the target padding node result return record in the feedback execution feature is compared with the expected field type in the expected state feature to form a field type similarity component; subsequently, the interface response state in the feedback execution feature is compared with the expected interface state in the expected state feature to form an interface state similarity component; and the result return timestamp in the feedback execution feature is compared with the timing requirement corresponding to the expected result state in the expected state feature to form a timing similarity component. Finally, based on the field order in the records corresponding to the feedback fields, the content similarity component, field type similarity component, interface state similarity component, and timing similarity component are organized into similar scales to obtain the feature similarity. The technical essence of feature similarity is a quantitative result used to represent the degree of closeness between the actual running state formed after the adaptive toolchain continues processing and the data state reference required by the current task.

[0147] Preferably, the calculation of feature similarity does not involve directly adding or multiplying feedback data of different dimensions. Instead, it first establishes a correspondence between fields with the same technical meaning, and then organizes the similarity components formed by similar fields. For the comparison between the actual field value and the expected field value range corresponding to the result return record of the target supplementary node, feature similarity is used to express whether the output content of the target supplementary node falls within the field range required by the currently executed task; for the comparison between the actual field type and the expected field type corresponding to the result return record of the target supplementary node, feature similarity is used to express whether the output field of the target supplementary node can be read by downstream link segments; for the comparison between the interface response status and the expected interface status, feature similarity is used to express whether the target supplementary node has completed the acceptance of the input and output interfaces after access; for the comparison between the result return timestamp and the timing requirements corresponding to the expected result status, feature similarity is used to express whether the result return of the target supplementary node meets the timing connection requirements for the continued execution of the adaptive toolchain. Through the above-mentioned comparison method of the same field, feature similarity can objectively reflect the continued processing effect of the adaptive toolchain, rather than solely relying on whether a result is returned as the judgment criterion.

[0148] Preferably, after calculating the feature similarity, a preset similarity threshold is first read from the similarity threshold configuration record. This preset similarity threshold is pre-configured based on the tool node type, the result return requirements of the current task, and the continued processing records of the adaptive toolchain within the historical execution cycle. The preset similarity threshold represents the minimum level of proximity that the feedback execution feature and the expected state feature should achieve. Then, the feature similarity is compared with the preset similarity threshold. When the feature similarity is greater than or equal to the preset similarity threshold, the continued processing result corresponding to the adaptive toolchain is written into the normal continuation record. When the feature similarity is less than the preset similarity threshold, the feature similarity, feedback execution feature, expected state feature, and feature compensation vector are jointly written into the compensation update trigger record. The compensation update trigger record provides the data basis for subsequent update processing of the feature compensation vector.

[0149] Preferably, when the feature similarity is less than a preset similarity threshold, when updating the feature compensation vector based on the feature similarity, the compensation update trigger record is first read, and the field positions where the feedback execution feature and the expected state feature do not reach the preset similarity threshold are extracted from the compensation update trigger record; then, according to the field positions, the corresponding content similarity component, field type similarity component, interface state similarity component, and temporal similarity component are read from the corresponding record of the feedback field, and the similarity components that do not reach the preset similarity threshold are organized into feedback difference components; subsequently, the feedback difference components are mapped to the content compensation component, interface compensation component, temporal compensation component, and continuation direction component in the original feature compensation vector to form a compensation component adjustment record. The compensation component adjustment record is used to indicate which compensation components in the original feature compensation vector failed to make the target filler node result return record meet the expected state feature, thereby providing a component-level basis for the update process.

[0150] Preferably, when updating the feature compensation vector, the original feature compensation vector, compensation component adjustment record, and target filler node record are read first. Then, based on the compensation component adjustment record, the content compensation component in the original feature compensation vector is adjusted for content compensation intensity to form an updated content compensation component; the interface compensation component in the original feature compensation vector is adjusted for interface acceptance requirements to form an updated interface compensation component; the timing compensation component in the original feature compensation vector is adjusted for result feedback requirements to form an updated timing compensation component; and the continuity direction component in the original feature compensation vector is verified for breakpoint direction to form an updated continuity direction component. Subsequently, the updated content compensation component, updated interface compensation component, updated timing compensation component, and updated continuity direction component are rearranged to form the updated feature compensation vector. The technical essence of the updated feature compensation vector is that, based on the feedback execution characteristics after further processing by the adaptive toolchain, the original breakpoint compensation requirement is corrected to form a compensation requirement vector. The updated feature compensation vector continues to be used as a compensation reference for subsequent candidate filler node matching or when the same abnormal node identifier reappears.

[0151] Preferably, the update process does not directly replace the original feature compensation vector. Instead, it retains the source of the topological break, the source of the abnormal node identifier, and the source of the corresponding field in the original feature compensation vector, and then makes targeted adjustments to the compensation components that do not reach the preset similarity threshold. Specifically, if the actual field value in the result feedback record of the target replacement node shown by the feedback execution feature does not fall within the expected field value range, the content compensation component is updated; if the actual field type in the result feedback record of the target replacement node shown by the feedback execution feature cannot be read by the downstream link segment, the interface compensation component is updated; if the timestamp of the result feedback shown by the feedback execution feature does not meet the timing requirements corresponding to the expected result state, the timing compensation component is updated; if the output direction of the target replacement node shown by the feedback execution feature does not correspond to the break direction identifier, the continuation direction component is updated. Through the above component correspondence method, the updated feature compensation vector can retain the traceable source of the original compensation requirement, while introducing the actual deviation reflected by the feedback execution feature.

[0152] Preferably, when associating the updated feature compensation vector with the abnormal node identifier, the process first reads the abnormal node identifier, the updated feature compensation vector, the topological break record, the target replacement node record, and the feature similarity; then, it reads the tool node identifier, the call order identifier, and the abnormal field record from the abnormal node identifier, and uses these as the primary association index; subsequently, it writes the updated feature compensation vector, the topological break record, the target replacement node record, and the feature similarity into the data location corresponding to the primary association index to form an abnormal compensation association record. This abnormal compensation association record represents the correspondence between the break location, compensation requirement, target replacement node, and continued processing feedback corresponding to a certain abnormal node identifier, enabling the read-updated feature compensation vector to be retrieved from the abnormal compensation association record when the same abnormal node identifier or similar abnormal field record reappears.

[0153] Preferably, the association processing further verifies the field source between the abnormal node identifier and the updated feature compensation vector. The field source verification first reads the abnormal field record in the abnormal node identifier, and then reads the updated content compensation component, updated interface compensation component, updated timing compensation component, and updated connection direction component from the updated feature compensation vector. Next, it determines whether the entry waiting time, node processing time, exit return time, inter-node transmission time, computational resource usage, memory resource usage, data access waiting value, interface response status value, or queue backlog status value pointed to by the abnormal field record can establish a field association with the corresponding compensation component in the updated feature compensation vector. When the field association is successful, the field source verification status is written to the abnormal compensation association record. Through field source verification, the abnormal compensation association record not only records the correspondence between the abnormal node identifier and the updated feature compensation vector, but also records the abnormal field source upon which this correspondence is based.

[0154] Preferably, when storing the associated data in a preset experience knowledge base, the abnormal compensation association record is first read, and the abnormal node identifier, topological break record, updated feature compensation vector, target replacement node record, feedback execution feature, expected state feature, feature similarity, and field source verification status are extracted from the abnormal compensation association record. Then, according to the tool node identifier and abnormal field record in the abnormal node identifier, an index directory in the experience knowledge base is generated. Subsequently, the topological break record, updated feature compensation vector, target replacement node record, feedback execution feature, expected state feature, feature similarity, and field source verification status are written into the index directory to form experience knowledge base records. The technical essence of the preset experience knowledge base is a structured knowledge storage space used to store the correspondence between abnormal node identifiers, break compensation requirements, target replacement node selection results, and continued processing feedback. The experience knowledge base records are used to provide historical references for perception, isolation, compensation mapping, and target replacement node matching in subsequent execution cycles.

[0155] Preferably, the pre-configured experience knowledge base is obtained based on agent resource configuration records, abnormal node identifiers formed within historical execution cycles, feature compensation vectors formed within historical execution cycles, and feedback execution features formed within historical execution cycles. Specifically, a node type directory of the experience knowledge base is first established according to tool node types, then an abnormal field directory is established according to abnormal field records, and subsequently a breakpoint location directory is established according to topological breakpoint records. The abnormal compensation association records are then written to the data location pointed to by the node type directory, abnormal field directory, and breakpoint location directory. Through this storage method, the experience knowledge base can retrieve the updated feature compensation vectors according to tool node type, abnormal field source, and topological breakpoint location. This allows the experience knowledge base records to be read as reference data for feature compensation vector generation and target replacement node matching when encountering the same or similar toolchain abnormalities in the current execution cycle.

[0156] Preferably, in a scenario, the target agent sequentially calls a large language processing model, a knowledge base, a database, a toolset, and an environmental perception node within the current execution cycle to complete a park anomaly monitoring and analysis task. Once the tool node corresponding to the database is isolated and forms an adaptive toolchain, the execution unit continues processing the adaptive toolchain. If, after the target supplementary node is connected, it can read the breakpoint output features provided by the tool node corresponding to the knowledge base, but the output field type in the result feedback record of the target supplementary node fails to meet the input requirements of the tool node corresponding to the toolset, then the field type similarity component in the feedback execution features will be lower than the reference requirement corresponding to the preset similarity threshold. At this time, a compensation component adjustment record is generated based on the field type similarity component, and the interface compensation component in the feature compensation vector is updated based on the compensation component adjustment record to obtain the updated feature compensation vector. Subsequently, the updated feature compensation vector is associated with the anomaly node identifier pointing to the tool node corresponding to the database, and the associated data is stored in a preset experience knowledge base. This experience knowledge base record can provide data reference for the target agent to reselect a target supplementary node that meets the input requirements of the tool node corresponding to the toolset when the same anomaly node identifier appears subsequently.

[0157] like Figure 3 As shown, this is an embodiment of a dynamic reorganization device for an intelligent agent toolchain with autonomous repair capabilities, the device comprising: The data acquisition module is used to acquire multimodal time-series runtime data of the target intelligent agent corresponding to the tool call chain in the current execution cycle; The identification and perception module is used to input the multimodal time-series running data into a preset perception network to obtain anomaly node identifiers that represent execution anomalies. The node location module is used to determine the corresponding failed tool node in a preset global dependency topology graph based on the abnormal node identifier. The global dependency topology graph includes multiple tool nodes and directed connection relationships between nodes. An isolation processing module is used to isolate the failed tool node to obtain a link to be repaired with a topological break. The feature extraction module is used to extract the breakpoint output features corresponding to the first tool node located upstream of the topological breakpoint in the link to be repaired; The state acquisition module is used to acquire the expected state features corresponding to the current task being performed by the target intelligent agent; The vector generation module is used to generate a feature compensation vector based on the breakpoint output features and the expected state features. The capability acquisition module is used to acquire the capability description vectors corresponding to multiple candidate filler nodes in the preset component library. The matching processing module is used to match the feature compensation vector with the capability description vectors corresponding to each of the plurality of candidate replacement nodes to determine the target replacement node. The link generation module is used to connect the target replacement node to the topology breakpoint and perform feature alignment processing on the input interface of the target replacement node to generate an adaptive toolchain.

[0158] like Figure 4 As shown, an electronic device according to an embodiment of this application includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the intelligent agent toolchain dynamic reorganization method with self-repair capability as described in any one of the claims of this application.

[0159] like Figure 5 As shown, this is a computer-readable storage medium according to an embodiment of the present application. The computer-readable storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device implements the intelligent agent toolchain dynamic reorganization method with autonomous repair capability as described in any one of the present applications.

[0160] The above Figures 3-5 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for dynamic reorganization of an intelligent agent toolchain with self-repair capability, characterized in that, The method includes: Obtain multimodal time-series runtime data corresponding to the tool call chain of the target agent in the current execution cycle; The multimodal time-series execution data is input into a preset perception network to obtain anomaly node identifiers that represent execution anomalies. Based on the abnormal node identifier, the corresponding failed tool node is determined in the preset global dependency topology graph, which contains multiple tool nodes and directed connection relationships between nodes. The failed tool node is isolated to obtain a link to be repaired with a topological break. Extract the breakpoint output features corresponding to the first tool node located upstream of the topological breakpoint in the link to be repaired; Obtain the expected state features corresponding to the current task being performed by the target intelligent agent; Based on the breakpoint output features and the expected state features, a feature compensation vector is generated. Obtain the capability description vectors corresponding to each of the multiple candidate filler nodes in the preset component library; The feature compensation vector is matched with the capability description vector corresponding to each of the multiple candidate replacement nodes to determine the target replacement node; The target filler node is connected to the topology breakpoint, and feature alignment processing is performed on the input interface of the target filler node to generate an adaptive toolchain.

2. The method according to claim 1, characterized in that, The acquisition of multimodal time-series runtime data corresponding to the tool call chain of the target agent in the current execution cycle includes: Obtain the instruction flow time consumption characteristics of each tool node in the tool call chain; Obtain the resource fluctuation characteristics of each tool node during the execution process; The instruction flow time consumption characteristics are subjected to timing alignment processing to obtain aligned time consumption characteristics; The alignment time consumption feature and the resource fluctuation feature are concatenated to generate the multimodal time-series running data.

3. The method according to claim 1, characterized in that, The step of inputting the multimodal time-series execution data into a preset perception network to obtain anomaly node identifiers representing execution anomalies includes: The multimodal time-series running data is input into the feature mapping layer of the perception network to obtain the feature distribution matrix; Calculate the deviation value of each feature element in the feature distribution matrix relative to the preset reference element; Extract target feature elements whose deviation values ​​are greater than a preset deviation threshold from the feature distribution matrix; The index information carried by the target feature element is parsed, and the identifier corresponding to the index information is determined as the abnormal node identifier.

4. The method according to claim 1, characterized in that, The process of isolating the failed tool node to obtain a link to be repaired with a topological break includes: Delete the first input connection edge in the global dependency topology graph that points to the failed tool node; Delete the second output connection edge originating from the failed tool node in the global dependency topology graph; The topology obtained after deleting the first input connection edge and the second output connection edge is determined as the link to be repaired; The starting node corresponding to the first input connection edge is determined as the first tool node, and the position where the first tool node was originally connected to the failed tool node is determined as the topological break.

5. The method according to claim 1, characterized in that, The step of generating a feature compensation vector based on the breakpoint output features and the expected state features includes: The difference between the breakpoint output features and the expected state features is calculated to obtain a difference feature vector. The difference feature vector is input into a preset feature mapping network to obtain the mapped feature compensation vector.

6. The method according to claim 1, characterized in that, The step of matching the feature compensation vector with the capability description vectors corresponding to each of the plurality of candidate filler nodes to determine the target filler node includes: For each of the plurality of candidate replacement nodes, the capability description vector corresponding to the candidate replacement node is subjected to an inner product operation with the feature compensation vector to obtain the matching value corresponding to the candidate replacement node. Based on the matching values ​​corresponding to each of the multiple candidate replacement nodes, the multiple candidate replacement nodes are sorted. The candidate replacement node that ranks first in the matching value is determined as the target replacement node.

7. The method according to claim 1, characterized in that, The step of performing feature alignment processing on the input interface of the target filler node to generate an adaptive toolchain includes: Extract the input data format corresponding to the target padding node; Extract the output data format corresponding to the breakpoint output features; Determine whether the input data format and the output data format are consistent; If it is determined that the input data format and the output data format are inconsistent, a preset data conversion middleware node is inserted between the first tool node and the target padding node; The breakpoint output feature is format-converted by the data conversion middleware node to obtain an alignment feature that is consistent with the input data format. Based on the alignment feature, a connection path is established from the first tool node to the target filler node via the data conversion middleware node, thereby generating the adaptive toolchain.

8. A dynamic reorganization device for an intelligent agent toolchain with autonomous repair capabilities, characterized in that, The device includes: The data acquisition module is used to acquire multimodal time-series runtime data of the target intelligent agent corresponding to the tool call chain in the current execution cycle; The identification and perception module is used to input the multimodal time-series running data into a preset perception network to obtain anomaly node identifiers that represent execution anomalies. The node location module is used to determine the corresponding failed tool node in a preset global dependency topology graph based on the abnormal node identifier. The global dependency topology graph includes multiple tool nodes and directed connection relationships between nodes. An isolation processing module is used to isolate the failed tool node to obtain a link to be repaired with a topological break. The feature extraction module is used to extract the breakpoint output features corresponding to the first tool node located upstream of the topological breakpoint in the link to be repaired; The state acquisition module is used to acquire the expected state features corresponding to the current task being performed by the target intelligent agent; The vector generation module is used to generate a feature compensation vector based on the breakpoint output features and the expected state features. The capability acquisition module is used to acquire the capability description vectors corresponding to multiple candidate filler nodes in the preset component library. The matching processing module is used to match the feature compensation vector with the capability description vectors corresponding to each of the plurality of candidate replacement nodes to determine the target replacement node. The link generation module is used to connect the target replacement node to the topology breakpoint and perform feature alignment processing on the input interface of the target replacement node to generate an adaptive toolchain.

9. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the dynamic reorganization method of the intelligent agent toolchain with self-repair capability as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the dynamic reorganization method of the intelligent agent toolchain with autonomous repair capability as described in any one of claims 1 to 7.

Images