Task execution agent phase state auditing, access control method and system

Abstract

The application discloses a task execution agent phase state audit routing and access control method and system, comprising the following steps: receiving a life cycle event issued by a task execution agent phase boundary, atomically replacing and writing a phase state persistence; obtaining an audit depth from a routing matrix by a phase state and a risk level, and applying an independent security lower limit to a high-risk task; taking a cryptography hash of an engineering product text after blank normalization as a content digest, determining repetition in a sliding time window according to a composite key formed by a task identifier, a digest and a phase state, and outputting skipping when a hit and a request depth are not deeper than a recorded depth; writing a final depth into an evidence ledger and outputting a signal, receiving an external audit ruling and merging the ledger; and performing a deterministic check on the evidence ledger before task migration to completion or delivery, and blocking when not ready and configured as a blocking mode to determine an exit code. The application realizes intelligent suppression of audit calls and deterministic control of quality access before delivery.

Description

Technical Field

[0001] This invention relates to the field of computer software automation processing and process control technology, specifically to a method and system for task execution agent stage status auditing routing and access control. Background Technology

[0002] Artificial intelligence task execution agents can execute multiple steps consecutively within a single session, invoke external tools, and produce verifiable engineering deliverables at stage boundaries. In software coding tasks, this agent manifests as continuously modifying multiple files, running tests, and generating audit materials; in non-coding tasks, it may be used for tasks such as compiling solution documents or organizing materials.

[0003] In existing technologies, when a continuous integration pipeline receives a push from a specified branch in the repository, the server executes check jobs sequentially according to the job list specified in the configuration file. The activation or deactivation of each job and its parameters are hardcoded in the configuration. The set of jobs is unrelated to the risk level of the task itself, and its duplication suppression is usually based on the byte-by-byte similarity of the commit identifier or output text. Version control pre-commit hooks operate on a single point in time of local commit operations, do not cross stages, and do not persist the stage state of the task, thus failing to distinguish between duplicate commits within the same stage and normal progress in the next stage. In the code review process, the intensity of review and the closure of comments are not determined by the machine. Workflow engines advance nodes according to a predefined directed graph, and the input of its conditional expressions is usually a business field or the execution result of the previous node, rather than a combination of the task risk level and the current project stage.

[0004] The aforementioned existing technologies have the following drawbacks: the set of inspection tasks is unrelated to the task risk level, leading to high-cost external audit calls that could be avoided for low-risk modifications, while high-risk modifications cannot increase audit depth as needed; duplicate judgments are based on byte-by-byte similarity, and the same content with only changes to indentation, blank lines, or line breaks is not recognized as equivalent content, and the automatic trigger path lacks cross-stage memory, resulting in repeated calls to the external audit process for the same content; the processing and completion judgment of audit conclusions belong to two separate judgment lines, and products with unresolved issues or missing evidence can still be advanced to the completion or delivery state; different engineering environments have their own agreed-upon handling intensity for the same access control, resulting in the same input receiving different exit states in different environments. Summary of the Invention

[0005] The technical problems to be solved by this invention include: how to prevent equivalent content from triggering repeated calls to the external audit process without performing the audit itself; how to make the call depth of the external audit engine determine the value according to the task stage status and risk level, so as to make the number of external audit calls and overhead per unit task controllable; how to ensure that incomplete evidence or unresolved products are rejected from migrating to the completion or delivery status by a determined process exit code when the access control is configured to block mode, and generate a determined alarm output in alarm mode; and how to make the above determinations determined and reproducible under given event input, given project configuration, given snapshot of each task status record and given window determination time point.

[0006] The common technical challenge is to ensure that, without performing the audit itself, the number of calls to the external audit engine and the progression conclusions of the task state machine are both deterministic and reproducible under given event inputs, given project configurations, and given state record snapshots. A shared specific technical feature is a deterministic decision chain centered on the audit depth and its value sources in the evidence ledger for each task.

[0007] This invention provides a method for routing and access control for status auditing during the task execution agent phase, in order to solve the above-mentioned technical problems.

[0008] A first aspect of the present invention provides a deep routing method for auditing the stage status of a task execution agent, comprising: receiving a lifecycle event emitted by the task execution agent at a stage boundary, the event carrying a task identifier, stage status, engineering artifacts, risk level, and user audit signal; the stage status being one of a plurality of stage states proceeding in a predetermined order, the plurality of stage states including at least planning completion and implementation completion; The stage state is written to each task state record, with the task identifier as the unit, using atomic replacement writing. The default audit depth is obtained from the preset routing matrix based on the stage status and the risk level, and the default audit depth is used as the candidate audit depth. The preset routing matrix is ​​arranged with stage status as the row and risk level as the column, and its cell values ​​are comparable order values ​​with a predetermined depth order. The risk level is one of multiple risk levels arranged in a predetermined risk order. For tasks with a risk level belonging to a predetermined high-risk level, a safety lower limit independent of the preset routing matrix is ​​applied to ensure that the candidate depth is not lower than the predetermined lower limit level. The result is the depth of this request. Among them, inputs that explicitly abandon auditing are skipped and a secondary confirmation flag is set in the output structure. The content digest is obtained by using the cryptographic hash of the text representation of the engineering product after blank normalization. A composite key is constructed by the task identifier, the content digest, and the stage status. The composite key includes at least the task identifier, the content digest, and the stage status. Recent audit records are retrieved within a preset sliding time window. If the audit fails to hit, or if the current request depth is deeper than the recorded depth, the current request depth is used as the final output audit depth, and an audit depth signal is output. If the audit succeeds and the current request depth is not deeper than the recorded depth, a skip signal is used as the final output audit depth, and a skip signal is output. The skip signal indicates that no call to the external audit engine will be initiated this time. The skip signal is the shallowest level in the predetermined depth order. Write the audit depth of the final output and its value source into the evidence ledger, and register the composite key and the audit depth of the final output together with the timestamp into the recent audit record.

[0009] A second aspect of the present invention provides a method for accessing the state quality of a task execution agent phase, comprising: before a task is migrated to a completed or delivered state, reading an evidence ledger containing adjudication results written by an audit depth routing step and collected from an external audit engine, and performing a deterministic verification; the deterministic verification includes two independent parallel judgments: first, normalizing each finding in the adjudication material into one of the adjudication states, and verifying the structural integrity of the audit adjudication material and the completeness of the adjudication states for each finding; second, verifying the completeness of predetermined fixed evidence items in the evidence ledger and that their values ​​do not belong to a predetermined set of placeholder values; the predetermined fixed evidence items include the final audit depth written by the audit depth routing step and its value source; for phases where the final audit depth is skipped, verifying that the reused recorded depth is not shallower than the lower limit obtained by the phase state and risk level for this phase, otherwise it is judged as unqualified; the adjudication state is one of a predetermined plurality of adjudication states, the predetermined plurality of adjudication states including at least a state representing acceptance, a state representing rejection, and a state representing a user decision; Read the handling mode specified in the project configuration, where the handling mode is a multi-level enumeration that includes at least alarms and blocking; perform configuration verification on the handling mode: the value of the handling mode must belong to the multi-level enumeration; when the project configuration indicates that the access control is enabled, the handling mode must not be disabled; when the project configuration indicates that the access control is disabled, the handling mode must be disabled; if the configuration verification fails, terminate the current access control operation with a determined configuration error exit code, without using any default value as a fallback. The state transition of the control task is controlled according to the handling mode and the conclusion of the deterministic verification: when configured in alarm mode, if the verification fails, only a risk warning is output and the transition is allowed; when configured in blocking mode, if the verification fails, the state transition is rejected with a determined process exit code; when the configuration verification passes and the handling mode is off, the deterministic verification is bypassed, the state transition is allowed and the predetermined success state is returned; the de-identified evidence result is output.

[0010] Optionally, the atomic replacement writing method includes: first writing all the contents of the state record to a temporary file in the same directory, wherein the temporary file name contains the process identifier of the writing process and a random string; then replacing the target state record file with the temporary file using an atomic renaming operation; and outputting an alarm without rejecting the writing if the stage state of the received event is out of order or rollback relative to the recorded stage state.

[0011] Optionally, the default audit depth is obtained from the preset routing matrix based on the stage status and risk level, and the default audit depth is used as a candidate audit depth; the candidate audit depth can also be replaced by the default audit depth in order of priority by explicit coverage parameters or user audit signal parsing results.

[0012] Optionally, the user audit signal parsing includes: matching text separated by words according to word boundaries, matching text not segmented by whitespace according to substrings, and determining the candidate depth based on the first hit segment of the preset signal vocabulary; for segments pointing to abandoned auditing, performing a negative search within the text segment where the hit position is located, and not judging the input as abandoned auditing when a negative statement appears.

[0013] Optionally, the whitespace normalization selects equivalent rules according to the product type: for text products where whitespace has semantic meaning, it uses original byte or syntax tree level normalization; it only uses whitespace normalization for formats where whitespace has been proven not to affect semantics; it hashes the original binary or deterministically normalized bytes of non-text products; and it incorporates the normalization rule version and product type into the composite key.

[0014] Optionally, the registration of the composite key with the final output audit depth and timestamp occurs when the routing decision is completed, regardless of whether the caller subsequently initiates an external audit.

[0015] Optionally, the engineering product includes textual representations of textual and non-textual products, and the binary content of the non-textual product itself does not participate in the calculation of the content summary.

[0016] Optionally, the structural integrity verification of the audit ruling material includes: the predetermined columns required by the audit ruling material must be complete and not duplicated or missing, the number of cells in rows and columns must be consistent, and cells with blank or placeholder values ​​are judged as unqualified.

[0017] Optionally, the evidence ledger may include the following fixed evidence items: scope of completion, verification commands and their results, audit rulings, persistence status, production environment boundaries, key boundaries and original private data boundaries, and remaining blocking items.

[0018] Optionally, normalizing each finding returned by the external audit engine into one of the decision states includes: normalizing synonymous values ​​expressed in different texts or aliases; each state requiring user decision must contain a non-empty decision subject identifier field and a non-empty pending object reference field.

[0019] Optionally, the determination of whether a value belongs to a predetermined set of placeholder values ​​is performed according to a defined normalization rule: first, remove leading and trailing whitespace and separators from the unit value, then normalize the case, and then compare it precisely with the predetermined set of placeholder values. If the value matches, it is determined to be a placeholder value.

[0020] This invention also provides a deep routing system for auditing the state of a task execution agent phase, comprising: The lifecycle event receiving module is used to receive events and their data emitted by the task execution agent at the stage boundary; The task status parsing module, connected to the lifecycle event receiving module, is used to write the stage status carried by the event into each task status record in an atomic replacement writing manner. The audit depth routing module, connected to the task status parsing module, is used to obtain candidate audit depths from the preset routing matrix based on the stage status and risk level, and to parse user audit signals. The safety lower limit module, connected to the audit depth routing module, is used to apply a safety lower limit independent of the preset routing matrix to high-risk tasks, so that the candidate depth is not lower than the predetermined lower limit level. The deduplication module, connected to the task status parsing module and the security lower limit module, is used to determine duplicates within a preset sliding time window by using the cryptographic hash of the text representation of the engineering product after blank normalization as the content digest, and a composite key composed of the task identifier, content digest and stage status, wherein the composite key includes at least the task identifier, content digest and stage status. The recent audit record registration module, connected to the deduplication module, is used to register the composite key, final audit depth and timestamp into the recent audit record, and filter out expired records according to the preset sliding time window; The evidence ledger module, connected to the deduplication module and the recent audit record registration module, is used to record the final audit depth and its value source, and to collect audit rulings and verification evidence.

[0021] The present invention also provides a task execution agent phase state quality access control system, comprising: The evidence ledger module stores the final audit depth written by the audit depth routing step and the source of its value, and aggregates the adjudication results returned by the external audit engine. The audit ruling verification module, connected to the evidence ledger module, is used to normalize the findings returned by the external audit into one of a number of predetermined ruling states and verify its structural integrity. An evidence ledger verification module, connected to the evidence ledger module, is used to perform deterministic verification on fixed evidence items before closing; the predetermined fixed evidence items include the final audit depth written in the audit depth routing step and its value source; for stages where the final audit depth is skipped, it is verified that the reused recorded depth is not shallower than the lower limit obtained by the stage status and risk level in this stage, otherwise it is judged as unqualified; The strategy configuration module, connected to the evidence ledger verification module, is used to read the disposal mode specified in the project configuration and perform configuration verification on the disposal mode; if the configuration verification fails, it terminates with a determined configuration error exit code, without using the default value as a fallback. The state transition control module, connected to the audit ruling verification module, the evidence ledger verification module, and the policy configuration module, is used to jointly determine whether to allow or deny the state transition of the task based on the two verification results and the handling mode; when the verification fails and the system is configured to block mode, the state transition is rejected based on the determined process exit code.

[0022] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a deep routing method for state auditing during the task execution agent phase, or a state quality gate method for the task execution agent phase.

[0023] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a deep routing method for state auditing during the task execution agent phase, or a state quality gate method for the task execution agent phase.

[0024] Compared with the prior art, the task execution agent stage state audit routing and access control method of the present invention has the following significant improvements: This invention writes the stage state of the task execution agent into the state record of each task in an atomic substitution manner, so that the input for subsequent judgment is a complete snapshot at any reading time, providing a definite input basis for repeated judgment and depth comparison.

[0025] This invention applies a security lower bound independent of the routing matrix to high-risk tasks, ensuring that general degraded user signals do not have a degraded effect on high-risk tasks and guaranteeing the audit strength of critical tasks. At the same time, by using cryptographic hashes with blank normalization and composite key sliding window deduplication, duplicate submissions that only modify blanks are identified as equivalent content. When a match is found and the request depth is not deeper than the recorded depth, a skip signal is output to suppress the current call to the external audit engine, while deeper audit requests can still be made, thereby saving computing power, tokens, and latency resources.

[0026] This invention uses two parallel access control systems to perform deterministic checks on the completeness of the three states of the audit ruling and the non-placeholder nature of the fixed evidence items before the task is closed. The state transition is controlled according to the disposal mode configured at the project level. This ensures that incomplete evidence or unresolved products are rejected by the process exit code under the blocking configuration and generate a deterministic alarm output under the alarm configuration. Moreover, the same event input will result in a deterministic exit state and state transition result under the same project configuration.

[0027] This invention makes the final audit depth output by the routing side and its value source one of the predetermined fixed evidence items for the access control side verification, and performs a lower limit verification on the reuse depth of the skipped stage, so that the two sides make a deterministic decision chain with the audit depth and its value source in the evidence ledger as the hub, ensuring that the advancement conclusion under a given input is certain and reproducible. Attached Figure Description

[0028] 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 some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of the overall process and system module composition of the present invention; Figure 2 This is a flowchart illustrating how the candidate depth, safety lower limit, and deduplication of composite keys in a fixed order jointly determine the final audit depth in this invention. Figure 3 This is a state machine diagram of the present invention for short-term deduplication, audit ruling standardization, and evidence ledger aggregation; Figure 4 A flowchart outlining the categorization of evidence certainty verification and disposal modes prior to the completion or delivery of this invention. Figure 5 This is a hardware structure diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0031] The relevant terms in this invention are defined as follows: A task execution agent is an artificial intelligence program that can continuously execute multiple steps within a single session, invoke external tools, and produce verifiable deliverables at the stage boundaries. The tasks it performs are not limited to software coding; they can also include document preparation, data organization, and other tasks.

[0032] A coding agent is a type of task execution agent that performs software coding tasks. It can modify multiple files, invoke external tools, run tests, and generate audit materials within a single session. This invention uses a coding agent as a preferred embodiment, but is not limited thereto.

[0033] Agent lifecycle events are events generated and emitted by the task execution agent at the stage boundary. They carry task identifier, stage status, engineering artifacts, project configuration, and risk level.

[0034] Engineering artifacts are the products produced by the task execution agent at the phase boundary and submitted for verification along with lifecycle events. They can include text artifacts and non-text artifacts.

[0035] The textual representation of the product is used to calculate the content summary. Textual products are represented by their own text; non-textual products are represented by the textual representation submitted along with the event.

[0036] A stage status represents an enumerated value indicating the current progress of a project task. These stages proceed in a predetermined order, including at least planning completion and implementation completion. The set of stage statuses can be declared in the project configuration file, and their execution order is determined according to the declaration order. The configuration file specifies the stage status sequence used in the project and its sequential relationship in a structured data format. The system reads this configuration upon startup and uses it to initialize the row indexes of the routing matrix and the out-of-order determination rules.

[0037] Each task status record is a persistent file that is independent of the task identifier, and records the current stage status of the task and its recent audit records.

[0038] Atomic substitution writing is a method of writing where the entire content is first written to a temporary file in the same directory, and then the target file is replaced by the temporary file using an atomic renaming operation; the temporary filename contains a process identifier and a random string. It guarantees that the target file will always present a complete before or after image at any given time.

[0039] Risk level is a tiered input determined by the irreversibility of the task, security sensitivity, financial impact, production impact, or data impact.

[0040] Audit depth refers to the level of audit intensity that should be applied to the current engineering product, and it is divided into multiple levels with a predetermined depth order.

[0041] The depth sequence number is the comparable sequence value corresponding to each audit depth level, making the comparison between depths and the selection of the deeper one a calculable operation.

[0042] An external audit engine is a separate process or service located outside this system that actually performs cross-model audits.

[0043] Signal-based routing means that this system only determines whether auditing is required and what level of auditing should be reached, and outputs the corresponding signal. It does not initiate, wait for, or manage external auditing processes, nor does it suspend or cancel initiated external auditing processes.

[0044] The routing matrix is ​​a preset two-dimensional table with stage status as rows, risk level as columns, and cell values ​​representing audit depth.

[0045] User audit signals are inputs expressed in natural language by the user, intended to specify the depth of the audit or to abandon the audit.

[0046] Explicit overriding allows the caller to directly specify the input channel for the audit depth level using command-line options or interface parameters.

[0047] The safety lower limit is a constraint that the audit depth of high-risk tasks must not be reduced below a predetermined level due to general downgraded user audit signals.

[0048] In the secondary confirmation state, when a high-risk task is skipped due to the user explicitly abandoning the audit, the flag bit in the output structure is set along with the output.

[0049] The abstract states that the text representation of the engineering product is first normalized with whitespace and then the value obtained by cryptographic hashing is taken.

[0050] Whitespace normalization selects equivalence rules based on product type: for text products where whitespace has semantic meaning, use raw byte or syntax tree level normalization; whitespace normalization is only used for formats where it has been proven that whitespace does not affect semantics; for non-text products, hash their raw binary or deterministically normalized bytes.

[0051] A composite key, consisting of at least a task identifier, a content summary, and a stage status, is used to determine duplicates within a sliding time window. Other dimensions, such as normalized rule version and product type, can be added to the composite key as needed to improve the accuracy of the hit determination.

[0052] Slide the time window to trace back a predetermined time interval from the current moment.

[0053] Deduplication is achieved when there are duplicate audit records with the same composite key within the sliding time window, and the audit depth of the duplicate record is not less than the depth of the current request.

[0054] The audit log is recorded by registering the selected depth and composite key along with the timestamp in the recent audit log of the task after each route determination is completed.

[0055] Evidence ledger, in a structured format, records the scope of completion, verification commands and their results, audit rulings, persistent status, boundary statements, and remaining blocking items.

[0056] Fixed evidence items are the set of predetermined items that an evidence ledger must contain.

[0057] Placeholder values, which are formally filled in but do not actually provide content, are treated the same as missing items.

[0058] The audit ruling has three states, which are the standardized states obtained for each finding returned by the external audit, including adoption, rejection, and requiring user decision.

[0059] Access control handling mode: The project configuration specifies the handling intensity for the same access control point, which includes multiple levels of enumeration, including at least alarms and blocking.

[0060] The method of this invention is driven by two independent triggering entry points, which are not executed consecutively within the same process call. Entry point one is triggered by a lifecycle event issued by the task execution agent at the stage boundary, performing event reception, state persistence, deep routing, deduplication determination, and signal output; entry point two is triggered by a request for the task to transition to a completed or delivered state, performing audit ruling normalization, evidence certainty verification, and state transition control. The following is combined with... Figures 1 to 5 The specific implementation methods and systems are described in detail.

[0061] like Figure 1As shown, the processing flow at Entry Point 1 begins with the lifecycle event receiving module receiving lifecycle events issued by the task execution agent at the stage boundary. These events carry a task identifier, stage status, engineering deliverables, risk level, and user audit signal. The stage status indicates the current progress of the engineering task, proceeding in a predetermined order, and includes at least planning completion and implementation completion. The risk level is determined by the task's irreversibility, security sensitivity, financial impact, production impact, or data impact. The engineering deliverables are those produced by the task execution agent at the stage boundary and submitted with the lifecycle event for verification.

[0062] Upon receiving an event, the task status parsing module persists the stage status carried by the event to each task status record using an atomic replacement write method. This atomic replacement write method first writes the entire content of the status record to a temporary file in the same directory. The temporary filename contains the process identifier of the writing process and a random string, ensuring that concurrent writing processes do not share the same temporary file. Then, an atomic rename operation replaces the target status record file with this temporary file. This write method guarantees that the target status record presents a complete before or after image at any reading time, without presenting a partially written intermediate state.

[0063] If the stage status of a received event is out of order or rollback compared to the recorded stage status, the system will output an alarm but will not refuse to write the event. Rework and redoing are allowed during the agent's operation, and the status records must be synchronized with the agent's actual progress to provide a basis for subsequent judgments.

[0064] like Figure 2 As shown, the audit depth routing module obtains the default audit depth from a preset routing matrix based on the stage state and the risk level, and uses the default audit depth as a candidate audit depth. The preset routing matrix has stage states as rows and risk levels as columns, with each cell value representing the audit depth. Audit depths are divided into multiple levels, each with a predetermined depth order. This order is represented by comparable order values, making the comparison of depths and the selection of the deeper value a calculable operation. The risk level is one of multiple risk levels arranged in a predetermined risk order.

[0065] The candidate depths come from three independent channels, with the following priority order: explicit coverage parameters, user audit signal parsing results, and the default value of the preset routing matrix. If the former exists, the latter is not used. Explicit coverage is the input channel where the caller directly specifies the audit depth level using command-line options or interface parameters. Its value is taken from a predetermined enumeration of audit depths and does not require natural language parsing.

[0066] For user audit signal parsing, the system parses the natural language input by the user. For text separated by words, it matches based on word boundaries; for text not segmented by whitespace, it matches based on substrings. This allows the same signal vocabulary to be applied to both types of text simultaneously. The signal vocabulary is part of the project configuration and is in the form of an ordered table of fragments and target levels. During parsing, it searches item by item according to the table order; the first matched fragment determines the candidate depth, and the search stops. The table order itself carries the resolution rules, ensuring that the same input has only one parsing result within the same vocabulary.

[0067] For segments indicating abandoned auditing, a negative search is performed within the text segment where the segment is located; if a negative statement appears in that segment, the input is not judged as abandoned auditing. Negative statements containing the word "skip" will also be judged as skipped under positive segment search; if no negative search is performed, they will be mistakenly judged as abandoned auditing, causing the task that should be audited to be silently skipped.

[0068] The safety lower limit module applies a safety lower limit independent of the preset routing matrix to high-risk tasks. When a task's risk level is high, the audit depth obtained through explicit overriding, user audit signal degradation, or matrix default values, if its depth order value is lower than the predetermined lower limit level, is raised to that lower limit level, and a lower limit activation flag is set. General degradation signals therefore do not have a degradation effect on high-risk tasks. Signals from users explicitly abandoning auditing remain valid and can bypass the safety lower limit by outputting a skip; the system sets a secondary acknowledgment flag in the output structure while outputting this skip. The safety lower limit is a second constraint applied independently of the routing matrix; on paths not determined to be explicitly abandoned for auditing and not identified as duplicate hits, the request depth of high-risk tasks will not fall below the lower limit level.

[0069] The audit depth is calculated in a fixed order. The first step is to select candidate depths; the second step is to apply a safety lower bound to these candidate depths, and the result is the current request depth; the third step is to perform a duplicate check within a sliding time window using this request depth and a compound key. If the check is successful (deduplication hit), the final output audit depth is "skipped"; otherwise, the final output is the request depth obtained in the second step. A successful deduplication hit means that the same compound key has already been routed and registered within the sliding time window at a depth no less than the current request depth. Outputting a non-skipped depth in this case would result in repeated calls to the same content.

[0070] like Figure 3As shown, the deduplication module performs duplicate detection on the composite key window. The system first normalizes the text representation of the engineering product according to the product type using equivalence rules: for text products where whitespace has semantic meaning, it uses original byte or syntax tree-level normalization; whitespace normalization is only used for formats where whitespace has been proven not to affect semantics; non-text products are hashed using their original binary or deterministically normalized bytes. The normalization rule version and product type are then incorporated into the composite key. If the original text is hashed directly, the same content can produce different hashes simply by changing indentation, blank lines, or newline conventions, rendering duplicate detection ineffective. After appropriate normalization, such changes do not alter the content digest, and changes to non-whitespace characters typically yield different content digests; the probability of collisions is negligible under the selected cryptographic hash security parameters.

[0071] The system uses a composite key composed of a task identifier, a content summary, and a stage status. This composite key includes at least the task identifier, the content summary, and the stage status. Recorded audit records are retrieved within a sliding time window formed by traversing a predetermined duration from the current time. During the retrieval, expired records outside the window are first filtered by timestamp, and then the composite key is compared for each record within the window. With the stage status incorporated into the key, the progression of the same project deliverable to different stages is considered a normal sequential transition and will not be misjudged as duplicate; only duplicate submissions of the same content within the same stage are suppressed.

[0072] When records with the same composite key exist within the same window, the system further compares the depth ordinal values: only if the ordinal value of the already recorded depth is not lower than the ordinal value of the current request depth is it determined as a deduplication hit and skipped; if the current request depth is deeper than the already recorded depth, it is not determined as a hit, and the current request is not suppressed. Records already routed to the shallow level cannot replace the current requested deep level.

[0073] This record is made immediately after each route determination, and it records the depth selected for that route, regardless of whether the caller subsequently initiates an external audit. The system does not wait for or manage the external audit process.

[0074] The evidence ledger module writes the final audit depth and its source into the evidence ledger, and registers the composite key and the final output audit depth along with the timestamp in the recent audit records for subsequent deduplication. The system then only outputs the aforementioned audit depth signal, which the caller uses to invoke the external audit engine. The actual cross-model audit is performed by this external engine; this system does not initiate, wait for, or manage the audit process, nor does it suspend or cancel initiated audit processes, nor does it participate in the parallel orchestration of audit tasks.

[0075] Coupling mechanism between signals and execution The skip signal represents the admission decision result of the audit call entry point. The call entry point uses this signal as a prerequisite; failure to comply prevents entry. This signal is directly output from the routing decision step to the admission decision interface of the call entry point. The caller can only read the signal value through this interface and cannot bypass it to decide whether to initiate auditing. Therefore, the technical effect of preventing identical content from triggering repeated calls to the external audit process is ensured by the signal output mechanism of this solution and the mandatory prerequisite of the call entry point, rather than relying on the caller's voluntary compliance.

[0076] like Figure 3 As shown on the right, the processing flow of Entry Point Two begins with the audit decision verification module receiving the results returned by the external audit engine. The system normalizes each finding into one of three values: Accept, Reject, or Requires User Decision, and standardizes synonymous values ​​expressed in different words or aliases. The system verifies the structural integrity of the decision material: all required predefined columns must be complete and not duplicated or missing; the number of cells in rows and columns must be consistent; blank or placeholder values ​​are considered unqualified. For each item in the "Requires User Decision" state, its record must contain a non-empty decision subject identifier field and a non-empty pending object reference field; if any of these fields are missing, empty, or belong to the predefined set of placeholder values, the structural integrity is deemed unqualified. If free text such as "Pending Confirmation with Relevant Parties" is allowed without requiring comparable identifiers and references, the access control verification procedure cannot determine whether the item has been processed based on structured fields, and the state transition determination loses its certainty.

[0077] like Figure 4 As shown, before a task migrates to the completed or delivered state, the evidence ledger verification module reads the evidence ledger and performs deterministic verification. This evidence ledger is written by the routing side or is a copy derived from the ledger according to a predetermined structure; the fixed item sets and values ​​of the two correspond one-to-one. This verification requires that the evidence block contain at least the following fixed items: completion scope, verification command and its result, audit ruling, persistence status, production environment boundary, key boundary and original private data boundary, and remaining blocking items. The absence of any fixed item constitutes a failure; even if a fixed item is filled in, if its value belongs to the predetermined set of placeholder values, it is considered a failure, equivalent to the absence of a fixed item.

[0078] The predetermined fixed evidence item also includes the final audit depth written in the audit depth routing step and its value source. For stages where the final audit depth is skipped, it is verified that the reused recorded depth is not shallower than the lower limit obtained according to the stage status and risk level of this stage; otherwise, it is judged as unqualified. Thus, the output of deduplication and routing becomes the object of access control verification, and the determination on both sides forms a deterministic determination chain with the audit depth and its value source in the evidence ledger as the hub.

[0079] The determination of placeholder values ​​follows established normalization rules: first, leading and trailing spaces and separators are removed from the unit values; then, case is normalized; and finally, a precise comparison is made with a predetermined set of placeholder values. Values ​​matching the set are considered placeholder values. If only the existence of a fixed item is checked, filling it with a placeholder will pass the check, and this check will no longer have any discriminatory power regarding the substantial completeness of the evidence.

[0080] like Figure 4 As shown, the deterministic verification before closure includes two independent parallel judgments: the first is the audit ruling verification gate, executed by the audit ruling verification module, which verifies the structural integrity of the audit ruling materials and the completeness of the ruling status for each discovery; the second is the evidence ledger verification gate, executed by the evidence ledger verification module, which verifies the completeness and non-placeholder nature of fixed evidence items. Each independently provides a conclusion of pass or fail. If either fails and its handling mode is configured as a blocker, the state transition will be rejected in the state transition step. The situation of unresolved discoveries is blocked by the ruling verification gate, and the situation of missing evidence or placeholders is blocked by the evidence ledger verification gate. The two are not interchangeable.

[0081] The access control verification program performs this deterministic check and provides the exit code.

[0082] like Figure 4 As shown, the policy configuration module reads the action mode specified for the access control system based on the project configuration, which is a multi-level enumeration including at least alarm and blocking modes. The system first verifies the configuration: the action mode value must belong to the multi-level enumeration; when the project configuration indicates that the access control is enabled, the action mode must not be disabled; when the project configuration indicates that the access control is disabled, the action mode must be disabled. If the configuration verification fails, the system terminates the access control operation with a determined configuration error exit code, without continuing execution with any default value as a fallback, to ensure that the same illegal configuration receives the same level of punishment across different versions.

[0083] The system controls the task's state transition based on the handling mode and verification results: when the handling mode is alarm, a failure to pass verification only outputs a risk warning while allowing the transition; when the handling mode is block, a failure to pass verification is rejected with a determined process exit code; when the handling mode is off, the deterministic verification is bypassed, the state transition is allowed, and the system returns to the predetermined success state. Finally, the system outputs anonymized evidence results, where the values ​​of the involved credentials, keys, and private data are not written out.

[0084] The same access control system can be configured with a specified handling strength at the project level. This allows for switching between alarms and blocking in different engineering environments without modifying the access control system, ensuring that the same event input yields a definite exit code and status transition result under the same project configuration.

[0085] Object eligibility argument This solution modifies the internal execution flow of the computer system (atomic replacement writing and reading of status records, triggering and suppressing audit engine calls, and process termination controlled by process exit codes). It addresses the technical problems of uncontrollable external audit engine call counts and computational resource overhead, as well as the uncertain and unreproducible results under given event inputs and configurations. The result is a reduction in call counts and a more definite and reproducible judgment outcome. The quality gate control section specifically emphasizes "terminating the process with a machine-readable, definitive output of the process exit code, and not using default values ​​as a fallback when the configuration is invalid," avoiding the impression of a manual regulation. The entire document avoids phrases like "improving management level / ensuring product quality."

[0086] The five methods of this invention constitute a unidirectional input-output dependency chain. Atomic substitution is written to the maintained task state record, which is the input shared by matrix value taking and window deduplication; the request depth generated by matrix value taking and safety lower bound is the necessary input for window deduplication to perform depth index comparison; the hit conclusion of window deduplication covers the request depth and determines the audit depth of the final output; the audit depth of the final output and its value source are written to the evidence ledger, and the ruling returned by external audit is also incorporated into the same ledger; this ledger is the read object of the access control verification before closure.

[0087] The stage state dimension in the composite bond and the blank normalization of the content summary are mutually conditional. Incorporating the stage state into the bond prevents the normal transition of the same product at different stages from being unintentionally suppressed; appropriate normalization of the product suppresses duplicate submissions that only modify blanks at the same stage.

[0088] The depth index is a necessary representation for closing the aforementioned dependency chain. The safety lower bound constraint and the deduplication hit determination operate on the same depth value, and their conclusions must be compatible. If the depth is only represented as a parallel label without a comparable index, then the two judgments of being below the lower bound and not shallower than the recorded depth cannot be made on the same scale. The output of the safety lower bound cannot directly serve as the input for deduplication, and the chain is broken.

[0089] Atomic substitution writing is a prerequisite for reliable on-chain input. Window comparison and matrix value retrieval both require reading the state record of each task; if the record may be in a partially written intermediate state, then composite key comparison and depth index comparison will lose their definite input.

[0090] The evidence ledger is the only data path between the routing side and the closing side. When the routing side and the closing side each hold independent records, the decision read during closing may not correspond to the actual depth selected by the routing side. The depth of the decision and whether closing is possible then become two unrelated decision lines. Ledger aggregation closes the second main line of this invention.

[0091] In one improvement process, the early design did not differentiate between levels and audited all at the same depth. Later, it was found that the outputs of the coding agent were sometimes complex and sometimes simple, sometimes important and sometimes general. Auditing all at the same depth caused low-risk tasks to occupy unnecessary high overhead resources. Therefore, it was changed to extract the depth from the preset matrix according to the stage status and risk level, and allowed users to specify the level separately through explicit overlay or audit signals.

[0092] In another improvement process, the early design was purely automatic rating, which automatically output the audit depth based on the stage status and risk level; later it was found that automatic rating could not know the user's special judgment. For example, if the user knew that a medium-risk task was actually important customer material and requested an in-depth audit, the automatic rating would output a standard file. Therefore, it was changed to give priority to explicit coverage parameters and user audit signal parsing results over the default value of the routing matrix.

[0093] Several components of this invention can be replaced by equivalent means, specifically as follows: the generation of lifecycle events can be a local version control hook, a continuous integration adapter, a proxy tool call, an integrated development environment plugin, or a server interface; the determination of risk levels can be a task tag, the affected file path, the data type involved, the deployment environment, or a combination of the above; the content digest can be a blank-normalized cryptographic hash, a change fingerprint, or other representation that can make equivalent content obtain the same value; the evidence ledger can be stored in markup language text, structured data files, databases, or other structured storage; the rejection state transition can be implemented as a process exit code, an interface response state, a state machine rejection transition, or a message notification; the input channel for audit depth can be at least one or a combination of explicit overlay, user audit signals, or default values ​​of the routing matrix.

[0094] This invention also provides a deep routing system for auditing the state of a task execution agent phase. For example... Figure 1 As shown, the system includes a lifecycle event receiving module, a task status parsing module, an audit depth routing module, a security lower limit module, a deduplication module, a recent audit record registration module, and an evidence ledger module.

[0095] A lifecycle event receiving module is configured at the system front end to receive events and their carried data emitted by the task execution agent at stage boundaries. A task status parsing module, connected to the lifecycle event receiving module, writes the stage status carried by the event into each task status record using atomic substitution. An audit depth routing module, connected to the task status parsing module, obtains candidate audit depths from a preset routing matrix based on stage status and risk level, and parses user audit signals. A security lower limit module, connected to the audit depth routing module, applies a security lower limit independent of the preset routing matrix to high-risk tasks, ensuring that the candidate depth is not lower than a predetermined lower limit level. A deduplication module, connected to the task status parsing module and the security lower limit module, uses a cryptographic hash of the text representation of the engineering product (normalized by whitespace) as a content digest, and determines duplicates within a preset sliding time window based on a composite key consisting of the task identifier, content digest, and stage status, where the composite key includes at least the task identifier, content digest, and stage status. The recent audit record registration module is connected to the deduplication module. It is used to register the composite key, final audit depth, and timestamp into the recent audit record and filter out expired records according to the preset sliding time window. The evidence ledger module is connected to the deduplication module and the recent audit record registration module. It is used to record the final audit depth and its value source, and to collect audit decisions and verification evidence.

[0096] This invention also provides a state quality access control system for the task execution agency phase. The system includes an evidence ledger module, an audit ruling verification module, an evidence ledger verification module, a policy configuration module, and a state transition control module.

[0097] The evidence ledger module stores the final audit depth written by the audit depth routing step and its value source, and aggregates the adjudication results returned by the external audit engine. The audit adjudication verification module, connected to the evidence ledger module, normalizes the findings returned by the external audit into one of several predetermined adjudication states and verifies its structural integrity. The evidence ledger verification module, also connected to the evidence ledger module, performs deterministic verification on fixed evidence items before closure; these predetermined fixed evidence items include the final audit depth written by the audit depth routing step and its value source; for stages with a skipped final audit depth, it verifies that the reused recorded depth is not shallower than the lower limit obtained by the stage status and risk level, otherwise it is deemed unqualified. The strategy configuration module, connected to the evidence ledger verification module, reads the disposal mode specified in the project configuration and performs configuration verification on the disposal mode; if the configuration verification fails, it terminates with a determined configuration error exit code, without using the default value as a fallback. The state transition control module is connected to the audit ruling verification module, the evidence ledger verification module, and the policy configuration module. It is used to jointly determine whether to allow or deny the state transition of the task based on the two verification results and the handling mode. When the verification fails and the system is configured to block mode, the state transition is rejected based on the determined process exit code.

[0098] The data flow relationships between the above modules are as follows: After receiving external input, the lifecycle event receiving module transmits the stage status and risk level to the task status parsing module for persistence; the task status parsing module transmits the stage status and risk level to the audit depth routing module; the audit depth routing module transmits the candidate audit depth to the security lower limit module; the security lower limit module transmits the current request depth to the deduplication module; the deduplication module transmits the final audit depth and deduplication result to the evidence ledger module; the evidence ledger module outputs the audit depth signal to the external audit engine and receives the ruling result returned by the external audit engine; the audit ruling verification module normalizes the ruling result and then imports it into the evidence ledger module; the evidence ledger verification module reads the evidence ledger for deterministic verification before state transition; the policy configuration module provides the disposal mode configuration to the evidence ledger verification module; the audit ruling verification module and the evidence ledger verification module respectively transmit their respective verification results to the state transition control module, which decides whether to allow or deny the migration based on the two verification results and the disposal mode.

[0099] The above five methods constitute a unidirectional input-output dependency chain, rather than five parallel steps that can be arbitrarily separated. The dependency chain proceeds as follows: atomic substitution is written to the maintained task state record, which is the input shared by matrix values ​​and window deduplication; the request depth generated by the matrix values ​​and the safety lower bound is the necessary input for window deduplication to compare depth indices; the hit conclusion of window deduplication covers this request depth and determines the audit depth of the final output; the audit depth of the final output and its value source are written to the evidence ledger, and the ruling returned by the external audit is also incorporated into the same ledger; this ledger is the read object of the access control verification before closing.

[0100] The stage state dimension in a composite key and the normalization of the content summary are mutually conditional. Incorporating the stage state into the key prevents the normal transition of the same product at different stages from being unintentionally suppressed; proper normalization of the product suppresses duplicate submissions that only modify blanks at the same stage.

[0101] The depth index is a necessary representation for closing the aforementioned dependency chain. The safety lower bound constraint and the deduplication hit determination operate on the same depth value, and their conclusions must be compatible. If the depth is only represented as a parallel label without a comparable index, then the two judgments of being below the lower bound and not shallower than the recorded depth cannot be made on the same scale. The output of the safety lower bound cannot directly serve as the input for deduplication, and the chain is broken.

[0102] Atomic substitution writing is a prerequisite for reliable on-chain input. Window comparison and matrix value retrieval both require reading the state record of each task; if the record may be in a half-written intermediate state, the read result may be an unparseable record or a partially updated record, and the composite key comparison and depth index comparison will lose their definite input.

[0103] The evidence ledger is the only data path between the routing side and the closing side. Before closing, the access control verification only needs to read the adjudication field and fixed items in the evidence block. Even without the routing side's depth write, the verification itself can still run independently. However, if the routing side and the closing side each hold independent records, the adjudication read during closing and the actual depth selected by the routing side may not correspond to each other. The depth of the adjudication and whether closing is possible then become two unrelated decision lines, and the second main line that this invention aims to solve cannot be established.

[0104] Figure 5 The embodiments of this specification illustrate a hardware structure diagram of an electronic device for a task execution agent phase state audit routing and access control method. For example... Figure 5 As shown, the electronic device may include at least one processor, memory, RAM, and communication interface, and the at least one processor, memory, RAM, and communication interface are connected together via a bus. At least one processor executes at least one computer-readable instruction stored or encoded in the memory.

[0105] It should be understood that the computer-executable instructions stored in memory, when executed, cause at least one processor to perform the above-described combinations in the various embodiments of this specification. Figures 1 to 4 The description includes various operations and functions.

[0106] In the embodiments of this specification, electronic devices may include, but are not limited to: personal computers, server computers, workstations, desktop computers, laptop computers, notebook computers, mobile electronic devices, smartphones, tablet computers, cellular phones, personal digital assistants, handheld systems, messaging devices, wearable electronic devices, consumer electronic devices, etc.

[0107] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code.

[0108] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-mentioned deep routing method for state auditing of the task execution agent stage, or the above-mentioned state quality access control method for the task execution agent stage.

[0109] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0111] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0112] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0113] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A deep routing method for auditing the state of a task execution agent phase, characterized in that, include: Receive lifecycle events issued by the task execution agent at the stage boundary. The events carry task identifier, stage status, engineering artifacts, risk level, and user audit signal. The stage status is one of multiple stage statuses that proceed in a predetermined order. The multiple stage statuses include at least planning completion and implementation completion. The stage state is written to each task state record, with the task identifier as the unit, using atomic replacement writing. The default audit depth is obtained from the preset routing matrix based on the stage status and the risk level, and the default audit depth is used as the candidate audit depth. The preset routing matrix is ​​arranged with stage status as the row and risk level as the column, and its cell values ​​are comparable order values ​​with a predetermined depth order. The risk level is one of multiple risk levels arranged in a predetermined risk order. For tasks with a risk level belonging to a predetermined high-risk level, a safety lower limit independent of the preset routing matrix is ​​applied to ensure that the candidate depth is not lower than the predetermined lower limit level. The result is the depth of this request. Among them, inputs that explicitly abandon auditing are skipped and a secondary confirmation flag is set in the output structure. The content digest is obtained by using the cryptographic hash of the text representation of the engineering product after blank normalization. A composite key is constructed by the task identifier, the content digest, and the stage status. The composite key includes at least the task identifier, the content digest, and the stage status. Recent audit records are retrieved within a preset sliding time window. If the audit fails to hit, or if the current request depth is deeper than the recorded depth, the current request depth is used as the final output audit depth, and an audit depth signal is output. If the audit succeeds and the current request depth is not deeper than the recorded depth, a skip signal is used as the final output audit depth, and a skip signal is output. The skip signal indicates that no call to the external audit engine will be initiated this time. The skip signal is the shallowest level in the predetermined depth order. Write the audit depth of the final output and its value source into the evidence ledger, and register the composite key and the audit depth of the final output together with the timestamp into the recent audit record.

2. The deep routing method for task execution agent phase state auditing according to claim 1, characterized in that, The atomic replacement writing method includes: first, writing all the contents of the status record to a temporary file in the same directory, wherein the temporary file name contains the process identifier of the writing process and a random string; then, replacing the target status record file with the temporary file using an atomic renaming operation; and outputting an alarm without rejecting the writing if the stage status of the received event is out of order or rollback relative to the recorded stage status.

3. The deep routing method for task execution agent phase state auditing according to claim 1, characterized in that, The candidate audit depth can also be replaced by the default audit depth in order of priority by explicit coverage parameters or user audit signal parsing results.

4. The deep routing method for task execution agent phase state auditing according to claim 3, characterized in that, The user audit signal parsing includes: matching text separated by words according to word boundaries, matching text not separated by whitespace according to substrings, and determining the candidate depth based on the first hit segment of the preset signal vocabulary; for segments pointing to abandon auditing, performing a negative search within the text segment where the hit position is located, and not judging the input as abandoning auditing when a negative statement appears.

5. The deep routing method for task execution agent phase state auditing according to claim 1, characterized in that, The blank normalization selects equivalent rules according to the product type: for text products where blanks have semantic meaning, use original byte or syntax tree level normalization; use blank normalization only for formats where blanks have been proven not to affect semantics; hash the original binary or deterministically normalized bytes of non-text products; and incorporate the normalization rule version and product type into the composite key.

6. The deep routing method for task execution agent phase state auditing according to claim 1, characterized in that, The registration of the composite key, the audit depth, and the timestamp of the final output occurs when the routing decision is completed, and is unrelated to whether the caller actually initiates an external audit thereafter.

7. The deep routing method for task execution agent phase state auditing according to claim 1, characterized in that, The engineering products include textual products and textual representations of non-textual products. The binary content of the non-textual products does not participate in the calculation of the content summary.

8. The deep routing method for task execution agent phase state auditing according to claim 1, characterized in that, Also includes: Before outputting the audit depth signal, it is determined whether the size of the engineering product has reached the threshold corresponding to the preset size classification. If it has, a failure status with an over-limit reason identifier is returned.

9. A method for state quality access control during the task execution agent phase, characterized in that, include: Before the task is migrated to the completed or delivered state, the evidence ledger containing the adjudication results written by the audit depth routing step and returned by the external audit engine is read and subjected to deterministic verification. The deterministic verification includes two independent parallel judgments: First, each finding in the adjudication material is normalized into one of the adjudication states, and the structural integrity of the audit adjudication material and the completeness of the adjudication states for each finding are verified. Second, the completeness of the predetermined fixed evidence items in the evidence ledger is verified, and their values ​​do not belong to the predetermined set of placeholder values. The predetermined fixed evidence items include the final audit depth written by the audit depth routing step and its value source. For stages where the final audit depth is skipped, the reused recorded depth is verified to be no shallower than the lower limit obtained by the stage status and risk level of this stage; otherwise, it is judged as unqualified. The adjudication state is one of a predetermined plurality of adjudication states, which at least include a state representing acceptance, a state representing rejection, and a state representing a user decision. Read the handling mode specified in the project configuration, where the handling mode is a multi-level enumeration that includes at least alarms and blocking; perform configuration verification on the handling mode: the value of the handling mode must belong to the multi-level enumeration; when the project configuration indicates that the access control is enabled, the handling mode must not be disabled; when the project configuration indicates that the access control is disabled, the handling mode must be disabled; if the configuration verification fails, terminate the current access control operation with a determined configuration error exit code, without using any default value as a fallback. The state transition of the control task is based on the handling mode and the conclusion of the deterministic verification: when configured in alarm mode, if the verification fails, only a risk warning is output and the transition is allowed. When configured in blocking mode, a failed verification will result in the rejection of the state transition using a determined process exit code. When the configuration verification passes and the processing mode is off, the deterministic verification is bypassed, the state transition is allowed, and the predetermined success state is returned. Output the desensitized evidence results.

10. The task execution agent phase state quality access control method according to claim 9, characterized in that, The structural integrity verification of the audit ruling materials includes: the predetermined columns required by the audit ruling materials must be complete and not duplicated or missing, the number of cells in rows and columns must be consistent, and cells with blank or placeholder values ​​are judged as unqualified.

11. The task execution agent phase state quality access control method according to claim 9, characterized in that, The evidence ledger includes the following fixed evidence items: scope of completion, verification commands and their results, audit rulings, persistence status, production environment boundaries, key boundaries and original private data boundaries, and remaining blocked items.

12. The task execution agent phase state quality access control method according to claim 9, characterized in that, Normalizing each finding returned by the external audit engine into one of the decision states includes: normalizing synonymous values ​​expressed in different texts or aliases; each state requiring user decision must contain a non-empty decision subject identifier field and a non-empty pending object reference field.

13. The task execution agent phase state quality access control method according to claim 9, characterized in that, The determination of whether a value belongs to the predetermined set of placeholder values ​​is performed according to the established normalization rules: first, the leading and trailing whitespace and separators of the unit value are removed, then the case is normalized, and then it is precisely compared with the predetermined set of placeholder values. If the match is found, it is determined to be a placeholder value.

14. A deep routing system for auditing the state of a task execution agent phase, characterized in that, include: The lifecycle event receiving module is used to receive events and their data emitted by the task execution agent at the stage boundary; The task status parsing module, connected to the lifecycle event receiving module, is used to write the stage status carried by the event into each task status record in an atomic replacement writing manner. The audit depth routing module, connected to the task status parsing module, is used to obtain candidate audit depths from the preset routing matrix based on the stage status and risk level, and to parse user audit signals. The safety lower limit module, connected to the audit depth routing module, is used to apply a safety lower limit independent of the preset routing matrix to high-risk tasks, so that the candidate depth is not lower than the predetermined lower limit level. The deduplication module, connected to the task status parsing module and the security lower limit module, is used to determine duplicates within a preset sliding time window by using the cryptographic hash of the text representation of the engineering product after blank normalization as the content digest, and a composite key composed of the task identifier, content digest and stage status, wherein the composite key includes at least the task identifier, content digest and stage status. The recent audit record registration module, connected to the deduplication module, is used to register the composite key, final audit depth and timestamp into the recent audit record, and filter out expired records according to the preset sliding time window; The evidence ledger module, connected to the deduplication module and the recent audit record registration module, is used to record the final audit depth and its value source, and to collect audit rulings and verification evidence.

15. A state quality access control system for the task execution agent phase, characterized in that, include: The evidence ledger module stores the final audit depth written by the audit depth routing step and the source of its value, and aggregates the adjudication results returned by the external audit engine. The audit ruling verification module, connected to the evidence ledger module, is used to normalize the findings returned by the external audit into one of a number of predetermined ruling states and verify its structural integrity. The evidence ledger verification module, connected to the evidence ledger module, is used to perform deterministic verification on fixed evidence items before closing; the predetermined fixed evidence items include the final audit depth written in the audit depth routing step and its value source; for stages where the final audit depth is skipped, it is verified that the reused recorded depth is not shallower than the lower limit obtained by the stage status and risk level in this stage, otherwise it is judged as unqualified; The strategy configuration module, connected to the evidence ledger verification module, is used to read the disposal mode specified in the project configuration and perform configuration verification on the disposal mode; if the configuration verification fails, it terminates with a determined configuration error exit code, without using the default value as a fallback. The state transition control module, connected to the audit ruling verification module, the evidence ledger verification module, and the policy configuration module, is used to jointly determine whether to allow or deny the state transition of the task based on the two verification results and the handling mode; when the verification fails and the system is configured to block mode, the state transition is rejected based on the determined process exit code.

16. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the audit depth routing method of any one of claims 1 to 8, or the quality access control method of any one of claims 9 to 13.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the audit depth routing method of any one of claims 1 to 8, or the quality access control method of any one of claims 9 to 13.

Images