Demand automatic review method and system based on multi-agent cooperation

Abstract

The application discloses a demand automatic evaluation method and system based on multi-agent cooperation, and the method comprises the following steps: acquiring project identification data and project demand evaluation data, constructing project evaluation tasks and configurations, and decomposing the data to obtain multiple atomic demands, scanning evidence and role evaluation context files; blind area detection is performed by calling each evaluation role agent, and the blind area detection results are merged to generate the evaluation log of the current evaluation round; based on the evaluation log, the conflict detection between roles, cross-interrogation, coverage review and user clarification are performed on multiple evaluation roles; the complete evaluation materials of the current round are acquired and rescanned, and whether the next round of evaluation is added is judged according to the rescan result and a preset convergence rule; if not, the formal demand document is generated and output after the preset access check item is checked and passed. The application can realize the automatic evaluation of demand evaluation, does not need too much manual participation, and improves the auditing efficiency and quality.

Description

Technical Field

[0001] This invention relates to a requirements review system for software requirements engineering, and more particularly to an automatic requirements review method and system based on multi-agent collaboration. Background Technology

[0002] Requirements review is a crucial link in the software delivery chain. Its quality directly determines the rework rate and launch cycle of software engineering. Currently, requirements review is mainly carried out manually, such as manual review through multi-transaction meetings, document circulation review, and instant messaging group chat review. This review method has drawbacks such as long review time, many blind spots, difficulty in collecting review conclusions, easy loss of key conclusions, inconsistent review quality, and inability to trace review omissions. Summary of the Invention

[0003] In order to overcome the shortcomings of the existing technology, one of the objectives of this invention is to provide an automatic requirement review method based on multi-agent collaboration, which can solve the problems of long review time, difficulty in collecting review conclusions, inconsistent review quality, and easy omissions in the review that cannot be traced.

[0004] The second objective of this invention is to provide an automatic requirements review system based on multi-agent collaboration, which can solve the problems of long review time, difficulty in collecting review conclusions, inconsistent review quality, and easy omissions in the review that cannot be traced.

[0005] One of the objectives of this invention is achieved through the following technical solution: An automatic requirement review method based on multi-agent collaboration, the automatic requirement review method includes: Review initialization steps: Obtain project identification data and project requirement review data, and then initialize the project according to the review process template and project requirement review data to build project review tasks and configure the constraints of each review task and the dependencies between review tasks; Data preparation steps: The project requirements review data is structured to decompose it into multiple atomic requirements, scan evidence, and role review context files; the project requirements review data includes at least project requirements, project prototypes, project code, project output documents, and external materials; Multi-role review process: By calling the intelligent agent of each review role to perform blind spot detection on the corresponding role's review context file, a review role output file is obtained; the review role output file includes blind spot discovery items of different severity levels; Cross-questioning steps: merge the review role output files of all review role agents to generate the review log for the current review round, perform conflict detection between multiple review roles based on the review log for the current round, construct a cross-questioning queue based on the conflict detection results, perform cross-questioning based on the cross-questioning queue to generate cross-questioning results, and rewrite the cross-questioning results back into the review log for the current round. Coverage audit steps: The auditing agent performs coverage audit on the coverage status and blind spot state of each atomic requirement based on preset multi-dimensional coverage to obtain audit results; and generates clarification questions based on the review log and audit results of the current round and feeds the clarification questions back to the user, then receives the user's answer to each clarification question, and when the user fills in the clarification question completely and correctly, writes the user's answer into the decision ledger corresponding to the current review stage; the review materials include all materials involved in the review within the current round; Dispute Decision-Making Steps: Obtain complete review data for the current round and rescan the complete review data to obtain the rescan results. Then, based on the rescan results of the current round and the preset convergence rules, determine whether to add a review for the next round. If yes, add a review for the next round; if no, execute the output step. The rescan results include: whether there are any new serious or important issues in the current round. Output steps: Based on preset admission check items, conduct admission checks on converged atomic requirement items, review logs, requirements of different levels, and decision ledgers within the system. After the admission checks are passed, generate a draft requirement document and then generate a formal requirement document based on the draft requirement document. The formal requirement document includes a structured review findings list, cross-questioning records, coverage design report, dispute resolution records, action items and responsible persons, and review structure.

[0006] Furthermore, the step of determining whether to add a next round of review based on the rescan results of the current round and the preset convergence rules specifically includes: when the rescan results of the current round pass verification, determining whether the current round is the first round; if so, returning to the multi-role review step to add a next round of review; if not, executing the output step when the rescan results of the current round pass verification; and when the rescan results of the current round fail verification, returning to the multi-role review step to add a next round of review; the rescan results passing verification means that no serious or important issues have occurred in the current round.

[0007] Furthermore, the data preparation steps specifically include: input data archiving step: obtaining the user's original words, current working directory, display files or links, requirement subject and candidate resource entry points according to the requirement review request, normalizing them and generating session input data; Atomic requirement decomposition steps: Extract the session context based on the session input data to decompose multiple atomic requirements in real time; each atomic requirement item includes requirement number, requirement topic, source, description, responsible role, related roles, status, requirement document section, acceptance number, and omission risk; Evidence scanning steps: Check each item of the context directory, requirement document, prototype, code repository and links listed for each review task, archive them and assign evidence numbers, and build an evidence index for each piece of evidence; Role Context Generation Steps: Generate a role review context file for each review role based on atomic requirements, evidence index, and session input file; each role review context file includes the role's concerns, evidence index, blind spots, risks, and reading order.

[0008] Furthermore, the multi-role review step specifically includes: reading the input archived data, the corresponding role review context file for the current round, and the decision ledger confirmed in the current stage through the product role agent, front-end role agent, back-end role agent, UI role agent, and test role agent respectively to conduct blind spot reviews, thereby deriving the review role output file; the review role output file includes blind spot discovery items of different levels; the levels include severity level, importance level, and recommendation level.

[0009] Furthermore, the cross-questioning step specifically includes: firstly, the review coordinator role agent reads the review role output files of the five review role agents and merges them into the review log of the current round; the review log includes closed blind spots, blind spots not yet closed, newly added blind spots, newly added decision points, coverage descriptions, confidence levels, and blocking issues; Based on the review logs of the current round, professional conflict detection is performed on multiple review roles to identify multiple professional conflicts, and these conflicts are prioritized according to their severity to construct a cross-consultation queue. Each professional conflict in the constructed cross-consultation queue is executed sequentially in three steps: the claimant's explanation, the questioner's response, and the adjudicator's judgment, to obtain the cross-consultation result for each professional conflict. The cross-consultation conclusion for each professional conflict is then written back to the review log of the current round. The cross-consultation result includes: the claimant's explanation, the questioner's response, the adjudication conclusion, the adjudication status, the impact of the requirements document, minority opinions, and candidates for user clarification. Candidates for user clarification include issues related to business scope, priority issues, and risk acceptance issues. Each professional conflict is allowed only one question, one answer, and one adjudication. The adjudication conclusion includes at least an internal adjudication, initiation of an evidence task, user decision required, and rejection.

[0010] Furthermore, the audit steps specifically include: verifying the coverage completeness, blind spot closure rate, and review quality evidence of each requirement item sequentially through the audit role intelligent agent according to multi-dimensional coverage, thereby generating audit results; wherein, the multi-dimensional coverage includes roles, pages, user flows, interfaces, data fields, permissions, state machines, abnormal flows, acceptance criteria, test data, non-functional requirements, and online migration compatibility risks; wherein, the verification status of each dimension is limited to covered, unclear, inapplicable, delayed, and accepted risk.

[0011] Furthermore, when the audit results contain markers requiring user decision-making, a unified user clarification file for the current round is generated and fed back to the user to obtain the user's answer results for each clarification question in the user clarification file, and the user's answer results are written into the decision ledger of the audit phase; the clarification file includes multiple clarification questions, and the multiple clarification questions are sorted according to the priority of severity level and pushed to the user answer entry, so that the user can answer each clarification question in sequence according to the priority of severity level.

[0012] Furthermore, the output step specifically includes: Access check steps: Based on preset access check items, perform access checks on converged atomic requirement items, review logs, audit results, and decision ledgers, and determine whether the access check passes. If so, execute the draft document generation step. The preset access check items include two consecutive complete scans with no new serious or important issues, issues meeting the final checkpoint requirements, no unclear items in the coverage matrix, complete decision ledger at the current stage, and full playback of user decisions. Draft document generation steps: Generate a draft requirement document based on the converged atomic requirement items, review log, coverage matrix, and replayed decision ledger; the draft requirement document includes a product review draft, a front-end review draft, a back-end review draft, a UI review draft, and a test review draft. Red Team Audit Steps: Verify the draft requirements document based on preset discovery and remediation items; the preset discovery and remediation items include: implementable remediation, direct correction, requiring technical adjudication, and requiring user supplementary decision-making. Delivery Steps: When the document verification passes, determine whether the software requirements audit should be delivered based on preset check conditions. If so, generate a requirements document. The preset check conditions include: all critical issues have been closed, postponed, or the risk has been accepted; all important issues have clear solutions or risk descriptions; two consecutive complete scans show no new critical or important issues; the draft review is unobstructed; the red team audit has passed; the traceability table is complete; and no unconfirmed facts are written into the facts section of the final delivery package. The formal requirements document includes: the final requirements document, UI interaction specifications, development contract, test acceptance package, traceability table, and conservatively written back project-level knowledge. The preset check conditions include: all critical issues have been closed, postponed, or the risk has been accepted; all important issues have clear solutions or risk descriptions; two consecutive complete scans show no new critical or important issues; the draft review is unobstructed; the red team audit has passed; the traceability table is complete; and no unconfirmed facts are written into the facts section of the final delivery package.

[0013] The second objective of this invention is achieved through the following technical solution: An automated requirements review system based on multi-agent collaboration includes: The review initialization module is used to obtain project identification data and project requirement review data, and then initialize the project according to the review process template and project requirement review data to build project review tasks and configure the constraints of each review task and the dependencies between review tasks. The data preparation module is used to perform structured processing on project requirement review data to decompose it into multiple atomic requirements, scan evidence, and role review context files; the project requirement review data includes at least project requirements, project prototypes, project code, project output documents, and external materials; A multi-role review module is used to perform blind spot detection on the corresponding role's review context file by calling each review role's intelligent agent to obtain the review role's output file; the review role's output file includes blind spot discovery items of different severity levels; The cross-questioning module is used to merge the review role output files of all review role agents to generate the review log of the current review round, and to perform conflict detection between multiple review roles based on the review log of the current round, and then construct a cross-questioning queue based on the conflict detection results, and perform cross-questioning based on the cross-questioning queue to generate cross-questioning results, and rewrite the cross-questioning results back to the review log of the current round. The coverage audit module is used to perform coverage audits on the coverage status and blind spot status of each atomic requirement based on preset multi-dimensional coverage by an audit role intelligent agent to obtain audit results; and to generate clarification questions based on the review logs and audit results of the current round and feed the clarification questions back to the user, and then receive the user's answers to each clarification question and write the user's answers into the decision ledger corresponding to the current review stage when the user fills in the clarification question completely and correctly; the review materials include all materials involved in the review within the current round; The dispute decision module is used to obtain the complete review data for the current round and rescan the complete review data to obtain the rescan results. Then, based on the rescan results of the current round and the preset convergence rules, it determines whether to add the next round of review. If yes, the next round of review is added; if no, the output step is executed. The rescan results include: whether there are any new serious or important issues in the current round. The output module is used to perform admission checks on converged atomic requirement items, review logs, requirements of different levels, and decision ledgers within the system according to preset admission check items. After the admission check is passed, a draft requirement document is generated, and then a formal requirement document is generated based on the draft requirement document. The formal requirement document includes a structured review findings list, cross-questioning records, coverage design report, dispute resolution records, action items and responsible persons, and review structure.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: By employing a requirement blind spot review method and system that involves multiple intelligent agents with independent responsibility boundaries working together, and driven by a dual-track hard criterion of continuous multi-round complete re-scanning and unreplayed user decisions, requirement blind spots are discovered from multiple perspectives, professional conflicts are explicitly adjudicated, independent audit verification is closed, and decisions are traceable and replayed. Ultimately, an executable development and acceptance contract is produced, thus automating the requirement review process. Compared with traditional manual review, its review efficiency and review quality have been greatly improved. Attached Figure Description

[0015] Figure 1 The flowchart of the software requirements review method based on multi-agent provided by the present invention. Detailed Implementation

[0016] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments. Example 1

[0017] This invention constructs a requirement blind spot review method with multiple intelligent agent roles having independent responsibility boundaries that collaboratively execute requirements. This method enables the requirement blind spots in software engineering to be discovered from multiple perspectives, professional conflicts to be explicitly resolved, independent audit verification to be closed, and decisions to be traceable and revisited, thereby achieving automatic review of software functions and significantly shortening the review time.

[0018] Specifically, such as Figure 1 As shown, the present invention provides a preferred embodiment of a software requirements review method based on multi-agent systems, comprising: Step S1: Obtain project identification data and project requirement review data, and then initialize the project according to the review process template and project requirement review data to build project review tasks and configure the constraints of each review task and the dependencies between review tasks.

[0019] Specifically, when a user initiates a review request, the system can obtain project identification data and project requirement review data based on the information in the review request. The system then executes project verification through the constructed review coordinator agent. If the current project name matches the requirement review, the system creates the project directory in the project-level file mailbox or reads an existing project directory. Then, the system initializes the project-level file, and all subsequent project-related files are mounted to this initial project set file.

[0020] Furthermore, the constraints of a review task include the task's read range and the allowed output set. The task's read range includes the user-restricted file region that the task can read; the allowed output set specifies the unique set of files that the task can write to. The dependencies between review tasks refer to the execution order of these tasks.

[0021] Step S2: Structure the project requirements review data to decompose it into multiple atomic requirements, scan evidence, and role review context files.

[0022] The project requirements review data includes at least project requirements, project prototypes, project code, project output documents, and external materials. Specifically, step S2 also includes: Step 1: Input Data Archiving Steps: Based on the requirements review request, obtain the user's original words, current working directory, displayed files or links, the main body of the requirement, and candidate resource entry points, normalize them, and generate session input data.

[0023] Step 2: Extract the session context based on the session input data to break down multiple atomic requirements in real time; each atomic requirement item includes requirement number, requirement topic, source, description, responsible role, related roles, status, requirement document section, acceptance number, and omission risk.

[0024] Step 3: Check each item in the context directory, requirement document, prototype, code repository and links listed for each review task, archive them and assign evidence numbers, and build an evidence index for each piece of evidence; Step 4: Generate a role review context file for each review role based on atomic requirements, evidence index, and session input file; each role review context file includes the role's concerns, evidence index, blind spots, risks, and reading order.

[0025] The above content is used to organize the user input data to construct the role review context file for each review role, in preparation for subsequent reviews.

[0026] Step S3: By calling each review role agent to perform blind spot detection on the corresponding role review context file to obtain the review role output file; the review role output file includes blind spot discovery items of different severity levels.

[0027] The product, front-end, back-end, UI, and testing roles each read the input archives, the corresponding role review context file for the current round, and the confirmed decision ledger for the current stage to perform blind spot reviews, thus generating a review role output file. The review role output file includes blind spot findings at different levels: severity, importance, and recommendation. More specifically, the review role agents include product, front-end, back-end, UI, and testing roles. These five review role agents execute in parallel, each reading the input data summary, the role context file for the current round, and the confirmed decision ledger for the current stage within its task scope to perform blind spot reviews. Once blind spots are found in any of the data, the review role output file is generated, categorizing the blind spots in this file according to a set level, including severity, importance, and recommendation. In other words, by constructing each role in the review process into a role-based intelligent agent, the role-based intelligent agent reviews the corresponding materials according to the role review mechanism to determine whether there are blind spots in the corresponding materials, and classifies the blind spots into levels such as serious, important, or recommended according to preset levels. In addition, the review role-based intelligent agent is prohibited from scanning the code repository, prototype directory, historical review directory, or external materials without boundaries when performing the review; at the same time, the same missing item is verified a maximum of twice, once by index search and once by bounded range search, and if no result is found, it is written as an evidence gap; the role must not be blocked due to insufficient evidence, and an evidence gap section must be added to the official document and completed.

[0028] More specifically, the review role agents in this system are constructed based on the configurations corresponding to the roles participating in the review. Each role has its responsibility boundaries, focus dimensions, non-focus dimensions, and output format defined by an independent behavior rule file; a corresponding agent is constructed for each role. In addition to the five roles directly involved in the review mentioned above, the roles in this system also include a review coordinator role and an audit role. The review coordinator role, as the sole host of the entire system, is responsible for project directory, Kanban task diagram, blind spot merging, clarifying issues, outputting requirement documents, final checkpoints, and project-level knowledge rewriting. The product review role is responsible for business objectives, user roles, scope boundaries, priorities, success metrics, and business acceptance criteria. The front-end review role is responsible for page status, interaction, components, front-end capabilities, front-end and back-end contracts, and front-end performance. The back-end review role is responsible for interfaces, data models, permissions, security, auditing, concurrency, caching, background tasks, and external systems. The UI review role is responsible for information architecture, page paths, usability, status feedback, and prototype conflicts. The test review role is responsible for acceptance criteria, test matrix, exceptions, boundaries, test data, and regression risks. The audit review role is responsible for covering matrices, traceability tables, checking for omissions, and auditing the feasibility of requirements documents, but does not make professional decisions.

[0029] Correspondingly, the role-based intelligent agents include the Review Coordinator Review Role-based Intelligent Agent, the Product Role-based Intelligent Agent, the Front-end Review Role-based Intelligent Agent, the Back-end Review Role-based Intelligent Agent, the UI Role-based Intelligent Agent, the Test Review Role-based Intelligent Agent, and the Audit Review Role-based Intelligent Agent. The Review Coordinator Review Role-based Intelligent Agent coordinates the various role-based intelligent agents throughout the system. The Product Role-based Intelligent Agent addresses blind spots regarding business objectives, scope boundaries, and priorities; the Front-end Review Role-based Intelligent Agent addresses blind spots regarding page status, interaction flow, and front-end / back-end contracts; the Back-end Review Role-based Intelligent Agent addresses blind spots regarding interface definitions, data models, and permission security; the UI Role-based Intelligent Agent addresses blind spots regarding information architecture, usability, and status feedback; and the Test Review Role-based Intelligent Agent addresses blind spots regarding acceptance criteria, exception boundaries, and test data. The Audit Review Role does not participate in professional decision-making; it is only used to audit the output results of other role-based intelligent agents.

[0030] Step S4: Merge the review role output files of all review role agents to generate the review log for the current review round, and perform conflict detection between multiple review roles based on the review log for the current round. Then, construct a cross-questioning queue based on the conflict detection results, perform cross-questioning based on the cross-questioning queue to generate cross-questioning results, and rewrite the cross-questioning results back into the review log for the current round.

[0031] Once blind spot detection is complete, the review coordinator agent first reads the review role output files from the five review role agents and merges them into the review log for the current round. The review log includes information on closed blind spots, blind spots not yet closed, newly added blind spots, newly added decision points, coverage descriptions, confidence levels, and blocking issues.

[0032] Then, based on the review logs of the current round, professional conflict detection is performed on multiple review roles to identify multiple professional conflicts. These conflicts are then prioritized according to their severity to construct a cross-consultation queue. Specifically, constructing the cross-consultation queue includes: performing professional conflict detection on multiple review roles based on the review logs of the current round to identify multiple professional conflicts, and prioritizing these conflicts according to their severity to construct the cross-consultation queue. Here, professional conflict refers to different decisions made by different agents regarding the same issue.

[0033] For each professional conflict within the constructed cross-consultation queue, a three-step sequential process is performed: the claimant's explanation, the questioner's response, and the adjudicator's judgment, to arrive at the cross-consultation result for each professional conflict. The cross-consultation conclusion for each professional conflict is then written back to the review log of the current round. The cross-consultation result includes: the claimant's explanation, the questioner's response, the adjudication conclusion, the adjudication status, the impact of the requirements document, minority opinions, and candidates awaiting user clarification. Candidates awaiting user clarification include issues related to business scope, priority, and risk acceptance. Each professional conflict is allowed only one question, one answer, and one adjudication. The adjudication conclusion includes at least an internal adjudication, initiation of an evidence task, requirement for user decision, and rejection.

[0034] Preferably, during cross-examination, a decision is made based on the candidate flags for each professional conflict. Specifically, when a flag indicates a decision based on evidence or role, the decision is set as an internal decision; when a flag indicates insufficient technical facts or evidence, the decision is set as initiating an evidence task as the next round of evidence gaps; when a flag indicates that a business decision-maker must choose and cannot be internally defaulted to, the decision is set as requiring a user decision; and when a duplicate flag indicates that the decision is discarded. Furthermore, technical facts, evidence gaps, and code implementation status must not be converted into user clarification.

[0035] Step S5: The auditing role intelligent agent performs coverage audit on the coverage and blind spot status of each atomic requirement based on the preset multi-dimensional coverage to obtain the audit results; and generates clarification questions based on the review log and audit results of the current round and feeds the clarification questions back to the user, and then receives the user's answer to each clarification question. When the user fills in the clarification question completely and correctly, the user's answer is written into the decision ledger corresponding to the current review stage.

[0036] The review materials include all materials involved in the review within the current round.

[0037] The audit role-based intelligent agent performs coverage auditing on the coverage and blind spot status of each atomic requirement based on preset multi-dimensional coverage. Specifically, this includes: the audit role-based intelligent agent sequentially verifies the coverage completeness, blind spot closure rate, and review quality evidence of each requirement item according to the multi-dimensional coverage. The multi-dimensional coverage includes roles, pages, user flows, interfaces, data fields, permissions, state machines, abnormal flows, acceptance criteria, test data, non-functional requirements, and deployment / migration compatibility risks. The verification status of each dimension is limited to covered, unclear, inapplicable, delayed, and accepted risk. When the verification status of a dimension is unclear, the atomic requirement condition must enter a clarified state.

[0038] More specifically, when audit results contain markers requiring user decisions, a unified user clarification document for the current round is generated and provided to users to obtain their responses to each clarification question. These responses are then written into the audit phase's decision ledger. The clarification document includes multiple clarification questions, which are prioritized according to severity level and then pushed to the user response portal, allowing users to answer each question sequentially based on severity. Generally, a clarification document includes a maximum of seven mandatory clarification questions, prioritized according to their severity level. Technical facts, evidentiary gaps, role-decidable items, and safe defaults that constitute a serious or important business decision are excluded from the user clarification process.

[0039] When obtaining the user's answer to each clarification question, the system replays the responses to determine if the clarification document completed by the user is complete. If so, it is written into the audit decision ledger.

[0040] Step S6: Obtain the complete review data for the current round and rescan the complete review data to obtain the rescan results. Then, based on the rescan results of the current round and the preset convergence rules, determine whether to add the next round of review. If yes, add the next round of review; otherwise, proceed to step S7.

[0041] Once the user's clarification document is written into the audit decision ledger, a complete re-scan is performed by rereading the corresponding role's review context file through the review roles participating in the current round. This re-scanning involves repeatedly performing a full re-scan of the role's context file by the intelligent agents of the roles participating in this round, not just processing the questions the user just answered. During the re-scan, the level of any newly added questions, open questions, closed items, and whether the requirement of no new questions being added in the previous two rounds is met are identified. The decision ledger is confirmed to have a higher priority than the initial input materials entered into the system during the re-scan. In other words, the re-scan results include whether any serious or important questions have been added in the current round.

[0042] More specifically, step S6 also includes: when the rescan result of the current round passes the verification, determining whether the current round is the first round; if so, returning to step S3 to add the review of the next round; if not, when the rescan result of the previous round passes the verification, executing step S7.

[0043] And if the verification of the previous round of rescan results fails, return to the multi-role review step to add the next round of review.

[0044] Step S7: Perform admission checks on the converged atomic requirement items, review logs, requirements of different levels, and decision ledgers within the system according to the preset admission check items. After the admission check is passed, generate a draft requirement document and then generate a formal requirement document based on the draft requirement document. The formal requirement document includes a structured review findings list, cross-questioning records, coverage design report, dispute resolution records, action items and responsible persons, and review structure.

[0045] More preferably, step S7 further includes: firstly, performing an admission check on the converged atomic requirement items, review logs, audit results, and decision ledger according to preset admission check items, and determining whether the admission check passes. The preset admission check items include: no new serious or important issues after two consecutive complete rescans; issues meeting the final checkpoint requirements; no unclear items in the coverage matrix; a complete decision ledger at the current stage; and full playback of user decisions. Once the access check is passed, a draft requirement document is generated based on the converged atomic requirement items, review logs, coverage matrix, and replayed decision ledger. This draft requirement document includes product review drafts, front-end review drafts, back-end review drafts, UI review drafts, and test review drafts.

[0046] Next, the draft requirements document is validated based on the preset identified remediation items. These preset remediation items include: implementable fixes, direct corrections, those requiring technical adjudication, and those requiring user supplementary decisions.

[0047] Finally, when the document verification passes, the system determines whether the software requirements audit should be delivered based on preset check conditions. If so, a requirements document is generated.

[0048] The preset inspection conditions include: all critical issues have been closed, postponed, or the risk has been accepted; all important issues have clear solutions or risk statements; two consecutive complete re-scans show no new critical or important issues; draft review is unobstructed; red team audit is passed; the traceability table is complete; and no unconfirmed facts are included in the fact section of the final deliverable package. Formal requirements documents include: final requirements document, UI interaction specifications, development contract, test and acceptance package, traceability table, and conservatively written back project-level knowledge.

[0049] Preset inspection conditions include: all critical issues have been closed, postponed, or risk accepted; all important issues have clear solutions or risk statements; two consecutive complete rescans show no new critical or important issues; draft review is unobstructed; red team audit is passed; traceability table is complete; and no unconfirmed facts are included in the fact section of the final delivery package.

[0050] This invention can be applied not only to software requirements review, but also to requirements review in various application scenarios, such as technical solution review, code review, architecture review, security review, test case review, UI or UX review, and compliance review. The input materials for different reviews can be in various formats, including requirements documents, prototype diagrams, design drafts, front-end code, back-end code, interface documents, test cases, and operation manuals. Example 2

[0051] Based on Embodiment 1, the present invention also provides another embodiment, a software requirements review system based on multi-agent systems, comprising: The review initialization module is used to obtain project identification data and project requirement review data, and then initialize the project according to the review process template and project requirement review data to build project review tasks and configure the constraints of each review task and the dependencies between review tasks. The data preparation module is used to structure project requirement review data to break it down into multiple atomic requirements, scan evidence, and role review context files; project requirement review data includes at least project requirements, project prototypes, project code, project output documents, and external materials; The multi-role review module is used to perform blind spot detection on the corresponding role's review context file by calling each review role's intelligent agent to obtain the review role's output file; the review role's output file includes blind spot discovery items of different severity levels; The cross-questioning module is used to merge the review role output files of all review role agents to generate the review log of the current review round, and to perform conflict detection between multiple review roles based on the review log of the current round, and then construct a cross-questioning queue based on the conflict detection results, and perform cross-questioning based on the cross-questioning queue to generate cross-questioning results, and rewrite the cross-questioning results back to the review log of the current round. The coverage audit module is used to perform coverage audits on the coverage status and blind spot status of each atomic requirement based on preset multi-dimensional coverage by the audit role intelligent agent to obtain audit results; and to generate clarification questions based on the review logs and audit results of the current round and feed the clarification questions back to the user, and then receive the user's answers to each clarification question. When the user fills in the clarification question completely and correctly, the user's answer is written into the decision ledger corresponding to the current review stage; the review materials include all materials involved in the review within the current round; The dispute decision module is used to obtain the complete review data for the current round and rescan the complete review data to obtain the rescan results. Then, based on the rescan results of the current round and the preset convergence rules, it determines whether to add the next round of review. If yes, the next round of review is added; if no, the output steps are executed. The rescan results include: whether there are any new serious or important issues in the current round. The output module is used to perform admission checks on converged atomic requirement items, review logs, requirements of different levels, and decision ledgers within the system based on preset admission check items. After the admission check is passed, a draft requirement document is generated, and then a formal requirement document is generated based on the draft requirement document. The formal requirement document includes a structured review findings list, cross-questioning records, coverage design report, dispute resolution records, action items and responsible persons, and review structure.

[0052] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for automatic requirement review based on multi-agent collaboration, characterized in that, The automatic requirements review method includes: Review initialization steps: Obtain project identification data and project requirement review data, and then initialize the project according to the review process template and project requirement review data to build project review tasks and configure the constraints of each review task and the dependencies between review tasks; Data preparation steps: The project requirements review data is structured to decompose it into multiple atomic requirements, scan evidence, and role review context files; the project requirements review data includes at least project requirements, project prototypes, project code, project output documents, and external materials; Multi-role review process: By calling the intelligent agent of each review role to perform blind spot detection on the corresponding role's review context file, a review role output file is obtained; the review role output file includes blind spot discovery items of different severity levels; Cross-questioning steps: merge the review role output files of all review role agents to generate the review log for the current review round, perform conflict detection between multiple review roles based on the review log for the current round, construct a cross-questioning queue based on the conflict detection results, perform cross-questioning based on the cross-questioning queue to generate cross-questioning results, and rewrite the cross-questioning results back into the review log for the current round. Coverage audit steps: The auditing agent performs coverage audit on the coverage status and blind spot state of each atomic requirement based on preset multi-dimensional coverage to obtain audit results; and generates clarification questions based on the review log and audit results of the current round and feeds the clarification questions back to the user, then receives the user's answer to each clarification question, and when the user fills in the clarification question completely and correctly, writes the user's answer into the decision ledger corresponding to the current review stage; the review materials include all materials involved in the review within the current round; Dispute Decision-Making Steps: Obtain complete review data for the current round and rescan the complete review data to obtain the rescan results. Then, based on the rescan results of the current round and the preset convergence rules, determine whether to add a review for the next round. If yes, add a review for the next round; if no, execute the output step. The rescan results include: whether there are any new serious or important issues in the current round. Output steps: Based on preset admission check items, conduct admission checks on converged atomic requirement items, review logs, requirements of different levels, and decision ledgers within the system. After the admission checks are passed, generate a draft requirement document and then generate a formal requirement document based on the draft requirement document. The formal requirement document includes a structured review findings list, cross-questioning records, coverage design report, dispute resolution records, action items and responsible persons, and review structure.

2. The automatic requirement review method based on multi-agent collaboration according to claim 1, characterized in that, The dispute decision-making step, which determines whether to add a next round of review based on the current round's rescan results and preset convergence rules, specifically includes: when the current round's rescan results pass verification, determining whether the current round is the first round; if so, returning to the multi-role review step to add a next round of review; if not, executing the output step when the current round's rescan results pass verification; and when the current round's rescan results fail verification, returning to the multi-role review step to add a next round of review; the passing of the rescan results verification means that no serious or important issues have occurred in the current round.

3. The automatic requirement review method based on multi-agent collaboration according to claim 1, characterized in that, The data preparation steps specifically include: Input data archiving step: Based on the requirement review request, obtain the user's original words, current working directory, display file or link, requirement subject and candidate resource entry, normalize them and generate session input data; Atomic requirement decomposition steps: Extract the session context based on the session input data to decompose multiple atomic requirements in real time; each atomic requirement item includes requirement number, requirement topic, source, description, responsible role, related roles, status, requirement document section, acceptance number, and omission risk; Evidence scanning steps: Check each item of the context directory, requirement document, prototype, code repository and links listed for each review task, archive them and assign evidence numbers, and build an evidence index for each piece of evidence; Role Context Generation Steps: Generate a role review context file for each review role based on atomic requirements, evidence index, and session input file; each role review context file includes the role's concerns, evidence index, blind spots, risks, and reading order.

4. The automatic requirement review method based on multi-agent collaboration according to claim 1, characterized in that, The multi-role review process specifically includes: using product role agents, front-end role agents, back-end role agents, UI role agents, and testing role agents to read the input archived data, the corresponding role review context file for the current round, and the decision ledger confirmed in the current stage to conduct blind spot reviews, thereby generating review role output files; the review role output files include blind spot findings at different levels; the levels include severity level, importance level, and recommendation level.

5. The automatic requirement review method based on multi-agent collaboration according to claim 1, characterized in that, The cross-questioning step specifically includes: firstly, the review coordinator role agent reads the review role output files of the five review role agents and merges them into the review log of the current round; the review log includes closed blind spots, blind spots that are not yet closed, newly added blind spots, newly added decision points, coverage descriptions, confidence levels, and blocking issues; Based on the review logs of the current round, professional conflict detection is performed on multiple review roles to identify multiple professional conflicts, and these conflicts are prioritized according to their severity to construct a cross-consultation queue. Each professional conflict in the constructed cross-consultation queue is executed sequentially in three steps: the claimant's explanation, the questioner's response, and the adjudicator's judgment, to obtain the cross-consultation result for each professional conflict. The cross-consultation conclusion for each professional conflict is then written back to the review log of the current round. The cross-consultation result includes: the claimant's explanation, the questioner's response, the adjudication conclusion, the adjudication status, the impact of the requirements document, minority opinions, and candidates for user clarification. Candidates for user clarification include issues related to business scope, priority issues, and risk acceptance issues. Each professional conflict is allowed only one question, one answer, and one adjudication. The adjudication conclusion includes at least an internal adjudication, initiation of an evidence task, user decision required, and rejection.

6. The automatic requirement review method based on multi-agent collaboration according to claim 1, characterized in that, The audit steps specifically include: verifying the coverage completeness, blind spot closure rate, and review quality evidence of each requirement item sequentially through the audit role intelligent agent according to multi-dimensional coverage, and then generating audit results; wherein, the multi-dimensional coverage includes roles, pages, user flows, interfaces, data fields, permissions, state machines, abnormal flows, acceptance criteria, test data, non-functional requirements, and online migration compatibility risks; wherein, the verification status of each dimension is limited to covered, unclear, inapplicable, delayed, and accepted risk.

7. The automatic requirement review method based on multi-agent collaboration according to claim 6, characterized in that, When the audit results contain markers that require user decision-making, a user clarification file for the current round is generated and fed back to the user to obtain the user's answer results for each clarification question in the user clarification file. The user's answer results are then written into the decision ledger of the audit phase. The clarification file includes multiple clarification questions. These multiple clarification questions are sorted according to their severity priority and then pushed to the user answer entry point so that the user can answer each clarification question in sequence according to the severity priority.

8. The automatic requirement review method based on multi-agent collaboration according to claim 1, characterized in that, The output step specifically includes: Access check steps: Based on preset access check items, perform access checks on converged atomic requirement items, review logs, audit results, and decision ledgers, and determine whether the access check passes. If so, execute the draft document generation step. The preset access check items include two consecutive complete scans with no new serious or important issues, issues meeting the final checkpoint requirements, no unclear items in the coverage matrix, complete decision ledger at the current stage, and full playback of user decisions. Draft document generation steps: Generate a draft requirement document based on the converged atomic requirement items, review log, coverage matrix, and replayed decision ledger; the draft requirement document includes a product review draft, a front-end review draft, a back-end review draft, a UI review draft, and a test review draft. Red Team Audit Steps: Verify the draft requirements document based on preset discovery and remediation items; the preset discovery and remediation items include: implementable remediation, direct correction, requiring technical adjudication, and requiring user supplementary decision-making. Delivery Steps: When the document verification passes, determine whether the software requirements audit should be delivered based on preset check conditions. If so, generate a requirements document. The preset check conditions include: all critical issues have been closed, postponed, or the risk has been accepted; all important issues have clear solutions or risk descriptions; two consecutive complete scans show no new critical or important issues; the draft review is unobstructed; the red team audit has passed; the traceability table is complete; and no unconfirmed facts are written into the facts section of the final delivery package. The formal requirements document includes: the final requirements document, UI interaction specifications, development contract, test acceptance package, traceability table, and conservatively written back project-level knowledge. The preset check conditions include: all critical issues have been closed, postponed, or the risk has been accepted; all important issues have clear solutions or risk descriptions; two consecutive complete scans show no new critical or important issues; the draft review is unobstructed; the red team audit has passed; the traceability table is complete; and no unconfirmed facts are written into the facts section of the final delivery package.

9. An automatic requirements review system based on multi-agent collaboration, characterized in that, include: The review initialization module is used to obtain project identification data and project requirement review data, and then initialize the project according to the review process template and project requirement review data to build project review tasks and configure the constraints of each review task and the dependencies between review tasks. The data preparation module is used to perform structured processing on project requirement review data to decompose it into multiple atomic requirements, scan evidence, and role review context files; the project requirement review data includes at least project requirements, project prototypes, project code, project output documents, and external materials; A multi-role review module is used to perform blind spot detection on the corresponding role's review context file by calling each review role's intelligent agent to obtain the review role's output file; the review role's output file includes blind spot discovery items of different severity levels; The cross-questioning module is used to merge the review role output files of all review role agents to generate the review log of the current review round, and to perform conflict detection between multiple review roles based on the review log of the current round, and then construct a cross-questioning queue based on the conflict detection results, and perform cross-questioning based on the cross-questioning queue to generate cross-questioning results, and rewrite the cross-questioning results back to the review log of the current round. The coverage audit module is used to perform coverage audit on the coverage status and blind spot status of each atomic requirement based on the preset multi-dimensional coverage by the audit role intelligent agent to obtain the audit results; The system generates clarification questions based on the review logs and audit results of the current round and provides these questions to users. It then receives user responses to each clarification question and writes the user's response into the decision ledger corresponding to the current review stage when the user completes the clarification question in a complete and accurate manner. The review materials include all materials involved in the review within the current round. The dispute decision module is used to obtain the complete review data for the current round and rescan the complete review data to obtain the rescan results. Then, based on the rescan results of the current round and the preset convergence rules, it determines whether to add the next round of review. If so, the next round of review is added. If not, proceed with the output step; The rescan results include: whether any new serious or important issues have been added in the current round; The output module is used to perform admission checks on converged atomic requirement items, review logs, requirements of different levels, and decision ledgers within the system according to preset admission check items. After the admission check is passed, a draft requirement document is generated, and then a formal requirement document is generated based on the draft requirement document. The formal requirement document includes a structured review findings list, cross-questioning records, coverage design report, dispute resolution records, action items and responsible persons, and review structure.

Images