Software development workflow orchestration method, apparatus, equipment, and media based on agent-based trusted skill scheduling

By performing integrity checks on skill description files and building a two-dimensional index, the problems of skill description file tampering and missing output objects are solved, and automated backup skill scheduling is achieved, improving the efficiency and accuracy of the software development workflow.

CN123114533APending Publication Date: 2026-11-17YUXIANG TECH (HANGZHOU) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611385540.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to detect skill description file tampering and lack of structured output objects in multi-stage agent skill scheduling, making it difficult to balance security and operational efficiency, and lacking an automated backup skill degradation mechanism.

Method used

By performing integrity checks on skill description files, constructing a two-dimensional index, checking the existence of predetermined structured output objects, and automatically calling backup skills when missing, until the predetermined output is generated, the system achieves automated management of file-by-file integrity checks and phased progress of skill description files.

Benefits of technology

It improves the efficiency and accuracy of software development workflow orchestration, ensures the integrity of skill description documents and the reliability of output objects, and reduces the probability of manual intervention and false successes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114533A_ABST
    Figure CN123114533A_ABST
Patent Text Reader

Abstract

This application discloses a method, apparatus, device, and medium for software development workflow orchestration based on agent-based trusted skill scheduling, relating to the field of artificial intelligence technology. The method includes: scanning and parsing a skill catalog associated with the workflow of a target software development task to obtain target skill description files and target stage identifiers; after the integrity verification of the target skill description files passes, obtaining skill output by calling the corresponding main skill based on the target stage identifier and a preset target mapping relationship; if the skill output does not contain a predetermined structured output object corresponding to the target stage identifier, sequentially calling the backup skill corresponding to the target stage identifier based on the target mapping relationship; and regenerating the skill output containing the predetermined structured output object by the target agent according to the skill description file corresponding to the backup skill, and writing it into the workflow context. This improves the efficiency of software development workflow orchestration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a software development workflow orchestration method, apparatus, device, and medium based on agent-based trusted skill scheduling. Background Technology

[0002] Currently, with the popularization of large language models and the rapid development of agent technology, multi-agent software development workflow orchestration is becoming increasingly common. Enterprises typically maintain the skills invoked by each agent at different stages of the responsibility chain using skill description files, and trigger hot reloading of the skill catalog by modifying timestamps.

[0003] For example, when a software development workflow includes dozens of skills and business personnel need to directly edit skill description files, existing solutions default to trusting the content of the skill description files on the disk. They determine the completion of a stage task solely based on whether the large language model call successfully communicates and returns non-empty text, without verifying whether the skill description files have been tampered with or checking whether the pre-defined structured output objects for each stage have actually been generated. To ensure the normal progress of gating at each stage, existing solutions still require manual verification of the consistency of the skill description files and manual retrying to remedy situations where communication is successful but structured output objects are missing. This makes it difficult to detect tampering in a timely manner and to automatically intercept false successes, resulting in a trade-off between security and operational efficiency.

[0004] In summary, in multi-stage agent skill scheduling, the issues of how to verify the integrity of skill description files and automatically trigger the downgrading of backup skills when communication is successful but the pre-defined structured output object for the stage is missing need further resolution. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a software development workflow orchestration method, apparatus, device, and medium based on agent-based trusted skill scheduling, which can improve the efficiency and accuracy of software development workflow orchestration. The specific solution is as follows: In a first aspect, this application discloses a software development workflow orchestration method based on agent-trusted skill scheduling, including: The skill catalog associated with the workflow of the target software development task is scanned and parsed to obtain the target skill description file and the corresponding target stage identifier. The integrity of the target skill description file is then verified based on the manifest file corresponding to the workflow to obtain the corresponding verification results. If the verification result indicates that the verification is successful, then based on the target stage identifier and the preset target mapping relationship, the target agent calls the main skill corresponding to the target stage identifier and obtains the skill output; the target mapping relationship represents the mapping relationship between the target stage identifier and the main skill list and the backup skill list; Check whether the skill output contains a predetermined structured output object corresponding to the target stage identifier; If the predetermined structured output object does not exist, the backup skills corresponding to the target stage identifier are sequentially invoked based on the target mapping relationship. The target agent regenerates the skill output based on the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object, and writes the predetermined structured output object as the stage orchestration output of the current stage into the workflow context.

[0006] Optionally, the step of scanning and parsing the skill catalog associated with the workflow of the target software development task to obtain the target skill description file and the corresponding target stage identifier includes: The skill subdirectories in the skill directory associated with the workflow of the target software development task are traversed, and the corresponding skill description files are obtained to obtain the target skill description file. The target skill description file is parsed, and the corresponding skill category and target stage identifier are extracted; the target stage identifier is a chain of responsibility stage identifier. Based on the skill category and the responsibility chain stage identifier, a two-dimensional index is constructed; wherein, the two-dimensional index includes a category index and a stage inverted index; the category index uses the skill category as the key and the corresponding skill name list as the value; the stage inverted index uses the responsibility chain stage identifier as the key and the corresponding skill name list as the value; the stage inverted index is used to locate the skill set corresponding to the target stage identifier in O(1) time complexity during stage scheduling.

[0007] Optionally, the step of performing an integrity check on the target skill description file based on the manifest file corresponding to the workflow to obtain the corresponding check result includes: Calculate the actual summary of the complete byte stream of the target skill description file, and compare the actual summary with the expected summary; the expected summary is the summary corresponding to the target skill description file recorded in the manifest file corresponding to the workflow. If the actual summary is inconsistent with the expected summary, a verification result indicating that the verification failed is generated. If the actual digest matches the expected digest, a verification result indicating that the verification has passed is generated.

[0008] Optionally, the step of invoking the main skill corresponding to the target stage identifier and obtaining skill output through the target agent based on the target stage identifier and a preset target mapping relationship includes: Read the list of main skills corresponding to the target stage identifier in the target mapping relationship; If the main skill list contains unregistered skill names, then the unregistered skill names are ignored to obtain the processed skill list; The skill description files corresponding to each main skill in the processed skill list are obtained, and each sub-output is loaded and generated sequentially by the target agent. Each sub-output is then written into the context of the workflow to obtain the skill output.

[0009] Optionally, checking whether the skill output contains a predetermined structured output object corresponding to the target stage identifier includes: Obtain the key of the predetermined structured output object corresponding to the target stage identifier, and query whether the key of the predetermined structured output object exists in the context of the workflow; If the key of the predetermined structured output object does not exist, then it is determined that the predetermined structured output object does not exist; If the key of the predetermined structured output object exists, it is determined that the predetermined structured output object exists, and the stage completion flag of the current stage is written into the context of the workflow.

[0010] Optionally, the step of regenerating the skill output by the target agent based on the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object includes: Based on the target stage identifier and the target mapping relationship, obtain the list of backup skills corresponding to the target stage identifier; If the list of backup skills is not empty, then the target skill is determined from the backup skills according to the order in the list of backup skills, and the target skill is invoked. The skill description file corresponding to the target skill is concatenated with the context of the current workflow and input into the target agent, and the regenerated skill output is written into the context of the workflow. Repeat the step of checking whether the predetermined structured output object corresponding to the target stage identifier exists in the skill output until the predetermined structured output object exists or the list of backup skills has been traversed.

[0011] Optionally, the integrity verification is enabled or disabled by a first environment variable; the method further includes: Determine whether the maximum modification timestamp of each skill subdirectory in the skill directory has changed; If the maximum modification timestamp changes, the skill catalog is rescanned to rebuild the skill registry; wherein the integrity check is not triggered during the rebuilding of the skill registry; the rebuilding is controlled by a second environment variable to be enabled or disabled; the first environment variable and the second environment variable are independent of each other.

[0012] Secondly, this application discloses a software development workflow orchestration device based on agent-trusted skill scheduling, comprising: The skill verification module is used to scan and parse the skill catalog associated with the workflow of the target software development task, obtain the target skill description file and the corresponding target stage identifier, and perform integrity verification on the target skill description file based on the manifest file corresponding to the workflow to obtain the corresponding verification result. The first skill output module is used to, if the verification result indicates that the verification is passed, call the main skill corresponding to the target stage identifier and obtain skill output through the target agent based on the target stage identifier and the preset target mapping relationship; the target mapping relationship indicates the mapping relationship between the target stage identifier and the main skill list and the backup skill list; The object checking module is used to check whether the skill output contains a predetermined structured output object corresponding to the target stage identifier; The backup skill invocation module is used to sequentially invoke the backup skills corresponding to the target stage identifier based on the target mapping relationship if the predetermined structured output object does not exist. The second skill output module is used to regenerate the skill output through the target agent according to the skill description file corresponding to the backup skill, until the regenerated skill output contains the predetermined structured output object, and write the predetermined structured output object as the stage orchestration output of the current stage into the context of the workflow.

[0013] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned software development workflow orchestration method based on agent-trusted skill scheduling.

[0014] Fourthly, this application discloses a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the aforementioned software development workflow orchestration method based on agent-trusted skill scheduling.

[0015] As can be seen, in this application, the skill catalog associated with the workflow of the target software development task is scanned and parsed to obtain the target skill description file and the corresponding target stage identifier. The target skill description file is then checked for integrity based on the manifest file corresponding to the workflow to obtain the corresponding verification result. If the verification result indicates that the verification passed, the main skill corresponding to the target stage identifier is called and the skill output is obtained based on the target stage identifier and a preset target mapping relationship through the target agent. The target mapping relationship represents the mapping relationship between the target stage identifier and the main skill list and the backup skill list. It is checked whether a predetermined structured output object corresponding to the target stage identifier exists in the skill output. If the predetermined structured output object does not exist, the backup skills corresponding to the target stage identifier are sequentially called based on the target mapping relationship. The skill output is regenerated by the target agent according to the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object, and the predetermined structured output object is written into the workflow context as the stage orchestration output of the current stage. In this way, on the one hand, by performing integrity checks on each skill description file, it is possible to detect situations where skill description files have been tampered with or where the list has not been updated with deployment; on the other hand, by using the existence of a predetermined structured output object as a criterion for stage advancement, it is possible to automatically trigger the downgrading of backup skills when communication is successful but structured output objects are missing, thereby improving the efficiency and accuracy of software development workflow orchestration. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This application discloses a software development workflow orchestration method based on agent-based trusted skill scheduling. Figure 2 This application discloses a specific signature verification flowchart; Figure 3 This application discloses a specific flowchart of a primary / secondary skill output contract downgrade process; Figure 4 This is a schematic diagram of a software development workflow orchestration device based on agent-trusted skill scheduling disclosed in this application; Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] With the development of large language models and intelligent agent technology, software development workflow orchestration based on multi-agent systems is becoming increasingly popular. Existing solutions default to trusting skill description files on disk, completing only the communication success or non-empty text return determination phase. They neither verify file tampering nor check if the pre-defined structured output objects for each phase are generated. This results in the inability to automatically downgrade when the main skill call is successful but the contract is missing, requiring manual intervention. Furthermore, signature verification and hot reload are coupled, and existing signature schemes are mostly directory-level or package-level, making it difficult to accurately locate single-file anomalies. Therefore, this application will specifically introduce a software development workflow orchestration method based on agent-trusted skill scheduling, which can solve the above problems.

[0020] See Figure 1 As shown in the figure, this application discloses a software development workflow orchestration method based on agent-trusted skill scheduling, including: Step S11: Scan and parse the skill catalog associated with the workflow of the target software development task to obtain the target skill description file and the corresponding target stage identifier, and perform integrity verification on the target skill description file based on the manifest file corresponding to the workflow to obtain the corresponding verification result.

[0021] In this embodiment, the step of scanning and parsing the skill directory associated with the workflow of the target software development task to obtain the target skill description file and the corresponding target stage identifier includes: traversing each skill subdirectory in the skill directory associated with the workflow of the target software development task and obtaining the corresponding skill description file to obtain the target skill description file; parsing the target skill description file and extracting the corresponding skill category and target stage identifier; the target stage identifier is the chain of responsibility stage identifier; constructing a two-dimensional index based on the skill category and the chain of responsibility stage identifier; wherein, the two-dimensional index includes a category index and a stage inverted index; the category index uses the skill category as the key and the corresponding skill name list as the value; the stage inverted index uses the chain of responsibility stage identifier as the key and the corresponding skill name list as the value; the stage inverted index is used to locate the skill set corresponding to the target stage identifier in O(1) time complexity during stage scheduling.

[0022] Specifically, when scanning the skill directory, the system recursively traverses each category subdirectory and locates the corresponding skill description file within each subdirectory using a pre-defined filename discovery strategy. For example, a pre-defined filename might be used as the identifier for the skill description file within that subdirectory. If an unreadable directory is encountered, the overall process is not interrupted; instead, structured failure information is returned for upper-level recording and diagnosis. When parsing the skill description file, one or more comma-separated chain of responsibility stage identifiers are extracted, and the parsing results are written to the mapping table from skill name to skill definition. Unparseable paragraphs are ignored, and corresponding warning messages are recorded to ensure that formatting issues in individual skill description files do not cause the entire registry construction to fail. When constructing the two-dimensional index, the category index uses the skill category as the key and a list of skill names belonging to that category as the value; the stage inverted index uses the chain of responsibility stage identifier as the key and a list of skill names mapped to that stage as the value.

[0023] It should be noted that the reason for maintaining an additional stage inverted index in addition to the category index is that if the skill registry is only indexed by category, a full table scan is required during stage scheduling to match the stage mapping paragraphs in each skill description file. In scenarios with a large number of skills or frequent registry rebuilds after hot reloads, a full table scan will significantly increase the initial response latency of stage execution. By pre-establishing a stage inverted index, the available skill subset for the current stage can be located directly in O(1) time complexity during stage scheduling, so that the stage lookup time does not increase with the total number of skills. For example, in a scenario where the skill catalog contains eighty skills, belonging to five business categories and mapped to seven responsibility chain stages, the stage scheduling can locate the available skill subset for the current stage in O(1) time complexity via the stage inverted index, and its stage lookup time is not affected by the total number of skills.

[0024] In this embodiment, the step of performing integrity verification on the target skill description file based on the manifest file corresponding to the workflow to obtain a corresponding verification result includes: calculating the actual digest of the complete byte stream of the target skill description file and comparing the actual digest with the expected digest; the expected digest is the digest corresponding to the target skill description file recorded in the manifest file corresponding to the workflow; if the actual digest is inconsistent with the expected digest, a verification result indicating that the verification failed is generated; if the actual digest is consistent with the expected digest, a verification result indicating that the verification passed is generated. Figure 2As shown, before performing the digest comparison, the manifest file is located according to a preset priority. The manifest file can be in structured text format or structured data format. If neither format of manifest file exists, a "manifest not found" result is returned. The path of the manifest file is configurable, and the location of the manifest file is decoupled from the skill loading process, so that integrity verification can be performed during runtime along with skill loading, or independently called by the continuous integration pipeline. When calculating the digest, a secure hash digest is calculated on the complete byte stream of the skill description file. The secure hash digest can use a 256-bit digest algorithm, and its hexadecimal representation is compared with the expected digest recorded in the manifest for string equality. If the two are not equal, the skill name and digest information are recorded in the mismatch list. When returning the verification result, the verification result is a structured result, which not only includes verification flags indicating whether the verification passed, but also the list version, the number of items checked, a list of missing directories recorded in the list but not on the disk, and a list of mismatches where the actual summary does not match the expected summary. Various situations such as items with missing summary fields, directories pointed to by the list not existing, and summaries not matching are listed separately in the structured result.

[0025] Since the verification result is a structured object that can be parsed by a machine, the continuous integration pipeline can hard-fail to block merging when a non-empty mismatch list is detected, while the local development environment can only issue a soft warning without preventing loading. This balances the mandatory nature of security auditing with the convenience of local iteration within the same verification logic. Furthermore, this embodiment maintains the expected cryptographic digest for each skill description file in the tree-like skill directory and calculates and compares the actual digest for each file. Therefore, when a single skill description file is replaced, the specific tampered skill can be accurately located, rather than simply detecting an anomaly in the entire skill package or directory. The mismatch list can include the name of the tampered skill and its digest prefix, allowing operations personnel to quickly locate the problem. In terms of time complexity, the overhead of the above integrity verification is approximately O(N×L), where N is the number of skill description files and L is the byte length of a single skill description file.

[0026] Step S12: If the verification result indicates that the verification is successful, then based on the target stage identifier and the preset target mapping relationship, the target agent calls the main skill corresponding to the target stage identifier and obtains the skill output; the target mapping relationship indicates the mapping relationship between the target stage identifier and the main skill list and the backup skill list.

[0027] In this embodiment, the step of calling the main skill corresponding to the target stage identifier and obtaining skill output based on the target stage identifier and a preset target mapping relationship, and through the target agent, includes: reading the main skill list corresponding to the target stage identifier in the target mapping relationship; if the main skill list contains an unregistered skill name, ignoring the unregistered skill name to obtain a processed skill list; obtaining the skill description file corresponding to each main skill in the processed skill list, loading and generating each sub-output sequentially through the target agent, and writing each sub-output into the context of the workflow to obtain the skill output. In this embodiment, the target mapping relationship is a stage-main / backup mapping table, which configures a main skill list and a backup skill list for each responsibility chain stage. The responsibility chain stage may include multiple stages such as requirement mining, scenario restoration, requirement document generation, quotation, contract boundary, review, and archiving. Different stages have significantly different requirements for skill semantics, so it is necessary to configure the main and backup skills at the stage level. When reading the primary skill list, external configuration can override the primary skill list in the stage-primary / backup mapping table to adapt to scheduling strategies in different deployment environments. Skill names in the primary skill list that are not registered in the skill registry are ignored. For each primary skill in the processed skill list, its corresponding skill description file is loaded sequentially, and corresponding sub-outputs are generated. After each call, the sub-output is written to the workflow context. Since the stage-primary / backup mapping table is the sole true source for stage scheduling, if a skill name configured in the mapping table does not exist in the registry, a corresponding configuration error message can be thrown, thus exposing the inconsistency between the primary / backup mapping and the actual skill registration during the scheduling phase.

[0028] Step S13: Check whether the skill output contains a predetermined structured output object corresponding to the target stage identifier.

[0029] In this embodiment, checking whether a predetermined structured output object corresponding to the target stage identifier exists in the skill output includes: obtaining the key of the predetermined structured output object corresponding to the target stage identifier, and querying whether the key of the predetermined structured output object exists in the context of the workflow; if the key of the predetermined structured output object does not exist, it is determined that the predetermined structured output object does not exist; if the key of the predetermined structured output object exists, it is determined that the predetermined structured output object exists, and the stage completion flag of the current stage is written into the context of the workflow.

[0030] That is, the predetermined structured output object is the structured output object that must be written into the workflow context at each stage, and different stages correspond to different structured output object keys. When making a judgment, the key of the predetermined structured output object corresponding to the current chain of responsibility stage is obtained, and the existence of the key is queried in the workflow context with O(1) time complexity; for example, for the requirement mining stage or the scenario restoration stage, the existence of its dedicated structured output object key is checked respectively. It should be noted that this embodiment uses the existence of the predetermined structured output object key as the judgment condition for whether the stage task can be advanced, and this condition takes precedence over the communication success flag called by the large language model and whether the returned text is not empty. This is because the large language model often returns natural language paragraphs rather than the agreed structured data. If only the communication success or the text not being empty is used as the criterion for stage completion, when the model returns non-empty text but does not contain the agreed structured output object, the orchestration layer will misjudge this pseudo-success of communication success but lack of structured object as stage advancement, which will lead to the failure of subsequent stage gating due to lack of contract key, and it will be difficult to locate the root cause. By using the existence of the structured output object key as the criterion for judgment, the aforementioned false successes can be intercepted at the current stage. Compared with schemes that use formal contract expressions to gate tool execution, this embodiment does not rely on complex formal expressions, but uses the existence of the stage-predetermined structured output object key as the judgment condition, making it more suitable for output contracts that are difficult to formally describe in fields such as product management.

[0031] Step S14: If the predetermined structured output object does not exist, then the backup skills corresponding to the target stage identifier are called sequentially based on the target mapping relationship.

[0032] In this embodiment, as Figure 3 As shown, when it is determined that the scene graph output object is missing in the project context, the backup skill downgrade is triggered. That is, the backup skills corresponding to the target stage identifier are called sequentially based on the target mapping relationship so that the skills can be merged in the future for gating evaluation.

[0033] Step S15: The target agent regenerates the skill output according to the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object, and writes the predetermined structured output object as the stage orchestration output of the current stage into the workflow context.

[0034] In this embodiment, the step of regenerating the skill output by the target agent based on the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object includes: obtaining a list of backup skills corresponding to the target stage identifier based on the target stage identifier and the target mapping relationship; if the list of backup skills is not empty, determining the target skill from each of the backup skills according to the order in the list of backup skills, and calling the target skill; concatenating the skill description file corresponding to the target skill with the context of the current workflow and inputting it into the target agent, and writing the regenerated skill output into the context of the workflow; re-executing the step of checking whether the predetermined structured output object corresponding to the target stage identifier exists in the skill output until the predetermined structured output object exists or the list of backup skills has been traversed. When a downgrade is triggered, if the current stage's backup skill list is not empty, the target skills are determined and invoked sequentially according to the order in the backup skill list. The skill description file corresponding to the target skill is concatenated with the current workflow context and input into the target agent. The regenerated skill output is written into the workflow context, and then the output contract determination is re-executed. This process is repeated until the predetermined structured output object exists or the backup skill list is traversed. The backup skill list can be empty. During the downgrade process, the number of downgrades can be recorded for auditing purposes. When the main skill and all backup skills have been executed, but the predetermined structured output object still does not exist, the list of missing fields of the predetermined contract key for this stage is output, and the stage completion flag of the current stage is prevented from being written into the workflow context, thus handing it over to the main chain gating for structured failure diagnosis. In this way, the chain of responsibility gating and the host environment can distinguish between two different root causes: all skills failed and contract key missing. This avoids simply marking the stage as failed after the main and backup chains are exhausted without passing specific contract gap details to the gating layer, thus providing a clear completion path for subsequent manual completion. Through the aforementioned output contract degradation mechanism, even in scenarios where communication with the large language model is successful but the pre-defined structured output object for the stage is lacking, the system can automatically switch to a backup skill for remediation. This can significantly reduce invalid manual retries, thereby lowering the probability of stage gating permanently failing or being mistakenly passed due to false success.

[0035] In this embodiment, the integrity check is controlled to be enabled or disabled by a first environment variable. The method further includes: determining whether the maximum modification timestamp of each skill subdirectory in the skill directory has changed; if the maximum modification timestamp has changed, the skill directory is rescanned to rebuild the skill registry; wherein the integrity check is not triggered during the rebuilding of the skill registry; the rebuilding is controlled to be enabled or disabled by a second environment variable; the first environment variable and the second environment variable are independent of each other. During runtime, the maximum modification timestamp of each skill subdirectory in the skill directory is compared, and when it changes, the skill directory is fully rescanned to rebuild the skill registry; the digest check is not called during the rebuilding process, and after the rebuild is completed, the corresponding change event can be broadcast to notify modules such as stage scheduling to update their indexes, wherein an exception in a single callback does not block the main process. In addition, since the digest check step can be called independently of skill loading, the continuous integration pipeline can perform digest comparison separately in the merge request stage without loading a large language model.

[0036] Because integrity verification is controlled by a first environment variable, while registry reconstruction based on modification timestamps is controlled by a second environment variable, and these two environment variables are independent of each other, the security team can enforce digest verification in the continuous integration pipeline, while the development team can still enable hot reloading for rapid iteration of skill tips. The two are not mutually exclusive within the same process. This avoids the security degradation caused by disabling verification for rapid hot updates, or the iteration efficiency reduction caused by disabling hot reloading to enforce verification.

[0037] It should be noted that in this application, the skill inventory file is used to record the expected cryptographic digests of the complete byte streams of skill description files corresponding to agreed-upon fixed filenames in each skill subdirectory. The inventory file itself does not prove its cryptographic origin within a single skill call path; its credibility is guaranteed by the organization's internal release process. This falls under the same category of release baseline integrity as package manager lock files and container image hash digest inventories: runtime comparison answers whether the current disk content is consistent with the released baseline, rather than re-completing the public key verification of the inventory. In scenarios requiring higher security, the entire inventory file can be asymmetrically digitally signed and distributed with the release package; at runtime, the signer's identity and inventory integrity are verified first, and then a digest comparison is performed on a skill description file-by-skill basis.

[0038] To further illustrate the technical points disclosed in this embodiment, several typical operating scenarios are used to explain this embodiment. It should be noted that the following scenarios represent the operating conditions of this embodiment under different input conditions.

[0039] In the first specific embodiment, the scenario occurs where the verification passes but the main text is maliciously replaced. The attacker replaces the malicious text of a skill description file while simultaneously updating the expected summary corresponding to that skill description file in the list, allowing the summary verification to pass. However, the malicious prompt induces the large language model to return a natural language paragraph instead of writing the pre-defined structured output object for that stage into the workflow context, such as the scene graph output object in the scene reconstruction stage. At this point, the output contract determination step will discover the lack of this structured output object in the workflow context, thus triggering the downgrading of the backup skill. If the backup skill is also tampered with and still cannot produce the structured output object, then a list of missing pre-defined contract keys will ultimately be output. Therefore, the summary verification and runtime shape correctness verification are complementary defenses.

[0040] In the second specific embodiment, the continuous integration pipeline intercepts deployments with outdated inventory lists. If a developer modifies a skill description file locally but deploys it without simultaneously updating the inventory list, the actual summary calculated will not match the expected summary recorded in the inventory list, and the mismatch list in the verification results will not be empty. If the first environment variable for integrity verification is enabled, loading will be blocked accordingly, thereby detecting deployments with outdated inventory lists.

[0041] In the third specific embodiment, the case of reverse lookup in the two-dimensional index stage is described above. When the skill catalog contains eighty skills, belonging to five business categories and mapped to seven responsibility chain stages, the stage inverted index can locate the available skill subset in the current stage in O(1) time complexity during stage scheduling. The stage reverse lookup time is not affected by the total number of skills, and the time to perform full verification of eighty skills is approximately milliseconds to seconds.

[0042] In the fourth specific embodiment, hot reload and verification switching are independent of each other. The development environment enables hot reload for rapid iteration of prompts, while the continuous integration pipeline enforces verification of the manifest summary. At runtime, the index is rebuilt upon detecting a change in the timestamp of a skill catalog modification, but summary verification is not invoked. The continuous integration pipeline, however, separately invokes verification steps such as manifest location, summary calculation, and structured return. Thus, the strategies of the security team and the development team do not conflict within the same process.

[0043] In addition to the aforementioned 256-bit digest algorithm, when calculating the digest of skill description files, national cryptographic digest algorithms or 384-bit digest algorithms can also be used to meet the collision resistance requirements of longer digests or corresponding compliance requirements. In this case, the file-by-file comparison process is still retained. This embodiment can perform asymmetric signing on the entire manifest file and publish it, verifying the signer at runtime to prevent the manifest file itself from being tampered with. This embodiment can use file system events instead of hot reload based on modification timestamps to trigger registry reconstruction for better real-time performance. This method does not replace the digest comparison step but is merely a subordinate variation. In the scenario of an open skill pool, this embodiment can dynamically select skills based on user intent vector retrieval to replace the stage-master-slave mapping table method. In this case, the integrity verification layer is still necessary, but the scheduling semantics are different. This embodiment can merge the structured output objects of the primary and backup skills after parallel invocation according to rules to reduce latency. This embodiment can also perform digest verification only during the continuous integration phase and not online verification at runtime to reduce the verification overhead of the online path, but this method will not be able to detect runtime tampering.

[0044] In its implementation, this embodiment can be implemented using programming languages ​​such as Python; the skill description file can be in Markdown or YAML format; the manifest file can be in YAML or JSON format; the cryptographic digest can be a hexadecimal representation of SHA-256; the large language model is invoked via prompt word templates for primary and backup skills, and its output is written into the project context dictionary, which serves as the workflow context; the continuous integration pipeline can be a corresponding continuous integration service, and the verification step is invoked separately during the merge request phase; this embodiment can be deployed as a model context protocol service or an embedded agent process, and the verification switch is controlled by environment variables.

[0045] In addition, this embodiment can be applied to the following typical scenarios: enterprises maintain a knowledge base containing dozens of skills and need to prevent unauthorized modification of skill description files in the production environment; intelligent agents that manage products in seven stages load different primary and backup skills in each stage and need to ensure that the shape of the output objects in each stage meets the gating requirements; the security team requires that the manifest summary and disk content be consistent before merging, while the development team still needs to perform hot reloading of skills without changing the pipeline; and protection is needed for pseudo-security scenarios where verification passes but stage outputs are unqualified.

[0046] As can be seen, in this embodiment, the skill catalog associated with the workflow of the target software development task is scanned and parsed to obtain the target skill description file and the corresponding target stage identifier. The target skill description file is then checked for integrity based on the manifest file corresponding to the workflow to obtain the corresponding check result. If the check result indicates that the check is successful, the main skill corresponding to the target stage identifier is called and the skill output is obtained based on the target stage identifier and the preset target mapping relationship. The target mapping relationship indicates the mapping relationship between the target stage identifier and the main skill list and the backup skill list. It is checked whether the predetermined structured output object corresponding to the target stage identifier exists in the skill output. If the predetermined structured output object does not exist, the backup skill corresponding to the target stage identifier is called sequentially based on the target mapping relationship. The skill output is regenerated by the target agent according to the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object. The predetermined structured output object is then written into the workflow context as the stage orchestration output of the current stage.

[0047] refer to Figure 4 As shown in the figure, this application also discloses a software development workflow orchestration device based on agent-trusted skill scheduling, including: The skill verification module 11 is used to scan and parse the skill catalog associated with the workflow of the target software development task, obtain the target skill description file and the corresponding target stage identifier, and perform integrity verification on the target skill description file based on the manifest file corresponding to the workflow to obtain the corresponding verification result. The first skill output module 12 is used to, if the verification result indicates that the verification is passed, call the main skill corresponding to the target stage identifier and obtain skill output through the target agent based on the target stage identifier and the preset target mapping relationship; the target mapping relationship indicates the mapping relationship between the target stage identifier and the main skill list and the backup skill list; Object checking module 13 is used to check whether there is a predetermined structured output object corresponding to the target stage identifier in the skill output; The backup skill invocation module 14 is used to sequentially invoke the backup skill corresponding to the target stage identifier based on the target mapping relationship if the predetermined structured output object does not exist. The second skill output module 15 is used to regenerate the skill output through the target agent according to the skill description file corresponding to the backup skill, until the regenerated skill output contains the predetermined structured output object, and write the predetermined structured output object as the stage orchestration output of the current stage into the context of the workflow.

[0048] As can be seen, in this embodiment, on the one hand, by performing integrity verification on a file-by-file basis for skill description files, it is possible to detect situations where skill description files have been tampered with or where the list has not been updated with deployment; on the other hand, by using the existence of a predetermined structured output object as a criterion for stage advancement, it is possible to automatically trigger the downgrading of backup skills when communication is successful but a structured output object is missing, thereby improving the efficiency and accuracy of software development workflow orchestration.

[0049] In some specific embodiments, the skill verification module 11 may specifically include: The file acquisition unit is used to traverse each skill subdirectory in the skill directory associated with the workflow of the target software development task and obtain the corresponding skill description files to obtain the target skill description file. The file parsing unit is used to parse the target skill description file and extract the corresponding skill category and target stage identifier; the target stage identifier is a chain of responsibility stage identifier. An index building unit is used to build a two-dimensional index based on the skill category and the responsibility chain stage identifier; wherein, the two-dimensional index includes a category index and a stage inverted index; the category index uses the skill category as the key and the corresponding skill name list as the value; the stage inverted index uses the responsibility chain stage identifier as the key and the corresponding skill name list as the value; the stage inverted index is used to locate the skill set corresponding to the target stage identifier in O(1) time complexity during stage scheduling.

[0050] In some specific embodiments, the skill verification module 11 may specifically include: The summary comparison unit is used to calculate the actual summary of the complete byte stream of the target skill description file and compare the actual summary with the expected summary; the expected summary is the summary corresponding to the target skill description file recorded in the manifest file corresponding to the workflow; The first verification result generation unit is used to generate a verification result indicating that the verification failed if the actual digest is inconsistent with the expected digest. The second verification result generation unit is used to generate a verification result indicating that the verification has passed if the actual digest is consistent with the expected digest.

[0051] In some specific embodiments, the first skill output module 12 may specifically include: The main skill list acquisition unit is used to read the main skill list corresponding to the target stage identifier in the target mapping relationship; The skill list processing unit is used to ignore unregistered skill names if the main skill list contains such names, so as to obtain a processed skill list. The skill output determination unit is used to obtain the skill description file corresponding to each main skill in the processed skill list, load and generate each sub-output in sequence through the target agent, and write each sub-output into the context of the workflow to obtain the skill output.

[0052] In some specific embodiments, the object inspection module 13 may specifically include: The object determination unit is used to obtain the key of the predetermined structured output object corresponding to the target stage identifier, and query whether the key of the predetermined structured output object exists in the context of the workflow; The first determination result unit is used to determine that the predetermined structured output object does not exist if the key of the predetermined structured output object does not exist. The second determination result unit is used to determine that the predetermined structured output object exists if the key of the predetermined structured output object exists, and to write the stage completion flag of the current stage into the context of the workflow.

[0053] In some specific embodiments, the second skill output module 15 may specifically include: The backup skill list determination unit is used to obtain the backup skill list corresponding to the target stage identifier based on the target stage identifier and the target mapping relationship; A backup skill invocation unit is used to determine a target skill from the backup skills in the order of the backup skill list if the backup skill list is not empty, and to invoke the target skill. The skill splicing unit is used to splice the skill description file corresponding to the target skill with the context of the current workflow and input it into the target agent, and write the regenerated skill output into the context of the workflow; The step jump unit is used to re-execute the step of checking whether the predetermined structured output object corresponding to the target stage identifier exists in the skill output, until the predetermined structured output object exists or the backup skill list has been traversed.

[0054] In some specific embodiments, the software development workflow orchestration device based on agent-trusted skill scheduling may further include: The timestamp determination module is used to determine whether the maximum modification timestamp of each skill subdirectory in the skill directory has changed; The registry rebuilding module is used to rescan the skill directory to rebuild the skill registry if the maximum modification timestamp changes; wherein the integrity check is not triggered during the rebuilding of the skill registry; the rebuilding is controlled to be enabled or disabled by a second environment variable; the first environment variable and the second environment variable are independent of each other.

[0055] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0056] Figure 5 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the software development workflow orchestration method based on agent-trusted skill scheduling disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0057] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0058] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0059] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the software development workflow orchestration method based on agent trusted skill scheduling disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0060] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed software development workflow orchestration method based on agent-trusted skill scheduling. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0062] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0063] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0064] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A software development workflow orchestration method based on agent-trusted skill scheduling, characterized in that, include: The skill catalog associated with the workflow of the target software development task is scanned and parsed to obtain the target skill description file and the corresponding target stage identifier. The integrity of the target skill description file is then verified based on the manifest file corresponding to the workflow to obtain the corresponding verification results. If the verification result indicates that the verification is successful, then based on the target stage identifier and the preset target mapping relationship, the target agent calls the main skill corresponding to the target stage identifier and obtains the skill output. The target mapping relationship represents the mapping relationship between the target stage identifier and the main skill list and the backup skill list; Check whether the skill output contains a predetermined structured output object corresponding to the target stage identifier; If the predetermined structured output object does not exist, the backup skills corresponding to the target stage identifier are sequentially invoked based on the target mapping relationship. The target agent regenerates the skill output based on the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object, and writes the predetermined structured output object as the stage orchestration output of the current stage into the workflow context.

2. The software development workflow orchestration method based on agent-trusted skill scheduling according to claim 1, characterized in that, The step of scanning and parsing the skill catalog associated with the workflow of the target software development task to obtain the target skill description file and the corresponding target stage identifier includes: The skill subdirectories in the skill directory associated with the workflow of the target software development task are traversed, and the corresponding skill description files are obtained to obtain the target skill description file. The target skill description file is parsed, and the corresponding skill category and target stage identifier are extracted; the target stage identifier is a chain of responsibility stage identifier. Based on the skill category and the responsibility chain stage identifier, a two-dimensional index is constructed; wherein, the two-dimensional index includes a category index and a stage inverted index; the category index uses the skill category as the key and the corresponding skill name list as the value; the stage inverted index uses the responsibility chain stage identifier as the key and the corresponding skill name list as the value; the stage inverted index is used to locate the skill set corresponding to the target stage identifier in O(1) time complexity during stage scheduling.

3. The software development workflow orchestration method based on agent-trusted skill scheduling according to claim 1, characterized in that, The process of performing an integrity check on the target skill description file based on the manifest file corresponding to the workflow to obtain the corresponding check result includes: Calculate the actual summary of the complete byte stream of the target skill description file, and compare the actual summary with the expected summary; the expected summary is the summary corresponding to the target skill description file recorded in the manifest file corresponding to the workflow. If the actual summary is inconsistent with the expected summary, a verification result indicating that the verification failed is generated. If the actual digest matches the expected digest, a verification result indicating that the verification has passed is generated.

4. The software development workflow orchestration method based on agent-trusted skill scheduling according to claim 1, characterized in that, The step of invoking the main skill corresponding to the target stage identifier and obtaining skill output through the target agent based on the target stage identifier and a preset target mapping relationship includes: Read the list of main skills corresponding to the target stage identifier in the target mapping relationship; If the main skill list contains unregistered skill names, then the unregistered skill names are ignored to obtain the processed skill list; The skill description files corresponding to each main skill in the processed skill list are obtained, and each sub-output is loaded and generated sequentially by the target agent. Each sub-output is then written into the context of the workflow to obtain the skill output.

5. The software development workflow orchestration method based on agent-trusted skill scheduling according to claim 4, characterized in that, The step of checking whether the skill output contains a predetermined structured output object corresponding to the target stage identifier includes: Obtain the key of the predetermined structured output object corresponding to the target stage identifier, and query whether the key of the predetermined structured output object exists in the context of the workflow; If the key of the predetermined structured output object does not exist, then it is determined that the predetermined structured output object does not exist; If the key of the predetermined structured output object exists, it is determined that the predetermined structured output object exists, and the stage completion flag of the current stage is written into the context of the workflow.

6. The software development workflow orchestration method based on agent-trusted skill scheduling according to claim 5, characterized in that, The step of regenerating the skill output by the target agent based on the skill description file corresponding to the backup skill until the regenerated skill output contains the predetermined structured output object includes: Based on the target stage identifier and the target mapping relationship, obtain the list of backup skills corresponding to the target stage identifier; If the list of backup skills is not empty, then the target skill is determined from the backup skills according to the order in the list, and the target skill is invoked. The skill description file corresponding to the target skill is concatenated with the context of the current workflow and input into the target agent, and the regenerated skill output is written into the context of the workflow. Repeat the step of checking whether the predetermined structured output object corresponding to the target stage identifier exists in the skill output until the predetermined structured output object exists or the list of backup skills has been traversed.

7. The software development workflow orchestration method based on agent-trusted skill scheduling according to any one of claims 1 to 6, characterized in that, The integrity verification is enabled or disabled by a first environment variable; the method further includes: Determine whether the maximum modification timestamp of each skill subdirectory in the skill directory has changed; If the maximum modification timestamp changes, the skill catalog is rescanned to rebuild the skill registry; wherein the integrity check is not triggered during the rebuilding of the skill registry; the rebuilding is controlled by a second environment variable to be enabled or disabled; the first environment variable and the second environment variable are independent of each other.

8. A software development workflow orchestration device based on agent-trusted skill scheduling, characterized in that, include: The skill verification module is used to scan and parse the skill catalog associated with the workflow of the target software development task, obtain the target skill description file and the corresponding target stage identifier, and perform integrity verification on the target skill description file based on the manifest file corresponding to the workflow to obtain the corresponding verification result. The first skill output module is used to, if the verification result indicates that the verification is passed, then based on the target stage identifier and the preset target mapping relationship, call the main skill corresponding to the target stage identifier through the target agent and obtain the skill output; The target mapping relationship represents the mapping relationship between the target stage identifier and the main skill list and the backup skill list; The object checking module is used to check whether the skill output contains a predetermined structured output object corresponding to the target stage identifier; The backup skill invocation module is used to sequentially invoke the backup skills corresponding to the target stage identifier based on the target mapping relationship if the predetermined structured output object does not exist. The second skill output module is used to regenerate the skill output through the target agent according to the skill description file corresponding to the backup skill, until the regenerated skill output contains the predetermined structured output object, and write the predetermined structured output object as the stage orchestration output of the current stage into the context of the workflow.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the software development workflow orchestration method based on agent-trusted skill scheduling as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs, which, when executed by a processor, implement the software development workflow orchestration method based on agent-trusted skill scheduling as described in any one of claims 1 to 7.