Summary of preliminary data acquired by the agent during RAG.
Contains lightweight identifiers for each kind of preliminary data that was
loaded into the agent's local context before producing its output. Only the
kinds specified by the Kind type parameter are present.
Number of items completed.
Tracks how many items have been successfully processed so far in the current operation. This value increments as each item is completed, providing real-time progress indication.
The ratio of completed to total gives the completion percentage:
progress = (completed / total) * 100
Timestamp when the event was created.
ISO 8601 formatted date-time string indicating when this event was emitted by the system. This timestamp is crucial for event ordering, performance analysis, and debugging the agent workflow execution timeline.
Format: "YYYY-MM-DDTHH:mm:ss.sssZ" (e.g., "2024-01-15T14:30:45.123Z")
Per-file review results.
Each entry contains the review verdict for a specific file's section content, including whether it was approved, feedback for improvement, and optional revisions.
A unique identifier for the event.
Function calling trial statistics for the operation.
Records the complete trial history of function calling attempts, tracking total executions, successful completions, consent requests, validation failures, and invalid JSON responses. These metrics reveal the reliability and quality of AI agent autonomous operation with tool usage.
Trial statistics are critical for identifying operations where agents struggle with tool interfaces, generate invalid outputs, or require multiple correction attempts through self-healing spiral loops. High failure rates indicate opportunities for system prompt optimization or tool interface improvements.
Retry attempt number for this event.
Starts at 0 for the first attempt. Increments each time some files are rejected and their generation is retried.
Current iteration number of the review process.
Tracks how many cross-file review cycles have been completed.
Detailed token usage metrics for the operation.
Contains comprehensive token consumption data including total usage, input token breakdown with cache hit rates, and output token categorization by generation type (reasoning, predictions). This component-level tracking enables precise cost analysis and identification of operations that benefit most from prompt caching or require optimization.
Token usage directly translates to operational costs, making this metric essential for understanding the financial implications of different operation types and guiding resource allocation decisions.
Total number of items to process.
Represents the complete count of operations, files, endpoints, or other entities that need to be processed in the current workflow step. This value is typically determined at the beginning of an operation and remains constant throughout the process.
Used together with the completed field to calculate progress percentage
and estimate time to completion.
Unique identifier for the event type.
A literal string that discriminates between different event types in the AutoBE system. This field enables TypeScript's discriminated union feature, allowing type-safe event handling through switch statements or conditional checks.
Examples: "analyzeWrite", "databaseSchema", "interfaceOperation", "testScenario"
Event fired during the cross-file review phase of section sections (###) across ALL files.
This event represents the final cross-file quality assurance step where ALL files' section content is reviewed together in a single LLM call. The Cross-File Section Review Agent validates EARS format uniformity, value consistency, terminology alignment, and Mermaid diagram style across the entire set of files before final document assembly.
Review criteria include:
Review outcomes are per-file:
Author
Juntak