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")
A unique identifier for the event.
Raw function calling arguments from the new replacement write submission.
Contains the full arguments object the LLM passed in its latest write
call, which replaces the previous output stored in oldbie. This becomes
the candidate output that will either be finalized via complete or
replaced again by yet another rewrite if the write limit has not been
reached.
Raw function calling arguments from the previous write submission.
Contains the full arguments object the LLM passed in its prior write call
(e.g., { type: "write", plan: "...", definition: {...} } for database
schema generation). The structure varies by orchestrator type. Comparing
oldbie and newbie reveals exactly what the agent changed during its
self-review rewrite cycle.
Source orchestrator whose write loop produced this rewrite.
Identifies which pipeline operation (e.g., database schema generation, interface operation design, test scenario creation) the LLM was working on when it decided to rewrite its previous output. Cannot be "facade" or "preliminaryAcquire" as those are not content-producing orchestrators.
The LLM's chain-of-thought reasoning for why the previous write was insufficient and what the new write improves.
Extracted from the thinking field of the LLM's function calling
arguments. Documents the agent's self-review process: what it found lacking
in the previous output and what design decisions changed in the
replacement. Useful for debugging output quality and understanding the
agent's iterative refinement process.
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 when the LLM agent submits a subsequent
writecall during the preliminary RAG write loop, replacing its previous output.The preliminary system operates as a write-review-rewrite cycle: after the LLM incrementally loads context via RAG requests and submits its first
write, the system allows it to self-review the output. If the LLM is satisfied, it callscompleteto finalize (IAutoBePreliminaryComplete). If it wants to improve, it submits anotherwrite— and that subsequent write triggers this event.This event captures the full before-and-after picture of a rewrite: the agent's reasoning for why the previous output was insufficient (
thinking), the raw arguments of the previous write (oldbie), and the raw arguments of the new replacement write (newbie). The number of rewrites is capped byAutoBeConfigConstant.PRELIMINARY_WRITE_LIMIT.Unlike AutoBePreliminaryAcquireEvent which tracks incremental context loading (RAG data requests), this event specifically tracks the moment the agent decides its prior output needs improvement and submits a replacement.
Author
Samchon