Type Alias TPipelineEvent

Type Declaration

• {
      at: number;
      kind: "pipeline:start";
      pipelineId: string;
      pipelineVersion: string;
  }
• {
      at: number;
      kind: "pipeline:end";
      output: "present" | "null";
      status: "completed" | "failed";
  }
• { at: number; kind: "stage:start"; stageId: string }
• {
      at: number;
      kind: "stage:end";
      stageId: string;
      status: TStageStatus;
      tokenUsage?: TLlmTokenUsage;
  }
• {
      at: number;
      attempt: number;
      kind: "stage:retry";
      reason: string;
      stageId: string;
  }
• {
      at: number;
      attempt: number;
      kind: "stage:llm-request";
      prompts: { system: string; user: string };
      stageId: string;
  }

  • at: number

  • attempt: number

    1, 2, ... — one event per LLM-call attempt, emitted before the provider call resolves. Symmetric with the post-call stage:llm-call.attempt.

  • kind: "stage:llm-request"

  • prompts: { system: string; user: string }

    The prompts about to be sent for this attempt. The user message includes any retry-suffix appended on attempt 2+ after a prior schema-validation failure — identical to the same attempt's stage:llm-call.prompts.user.

  • stageId: string

• {
      at: number;
      attempt: number;
      kind: "stage:llm-response-created";
      responseId: string;
      stageId: string;
  }

  • at: number

  • attempt: number

    1, 2, ... — the attempt index this id belongs to.

  • kind: "stage:llm-response-created"

    Emitted by llmStage as soon as a response id is known. On a streaming call that makes a single request (either streaming mode, with no function tools) the id arrives with the first SSE event, while the call is still running. Only a background-stream response keeps generating after the caller disconnects, so only there is this the point to persist the id for recovery. Otherwise (synchronous, poll-only, or a function-tool loop) the id is known at completion, and this event fires immediately before stage:llm-call on the same attempt, with the same id that event carries.

  • responseId: string

    The OpenAI response id for this attempt.

  • stageId: string

• {
      at: number;
      attempt: number;
      delta: string;
      kind: "stage:llm-text-delta";
      stageId: string;
  }

  • at: number

  • attempt: number

    1, 2, ... — the attempt this chunk belongs to. A retried attempt streams its text again from the beginning.

  • delta: string

    One chunk of assistant output text, as the provider emitted it. Never accumulated.

  • kind: "stage:llm-text-delta"

  • stageId: string

• {
      at: number;
      attempt: number;
      kind: "stage:llm-call";
      output: unknown;
      prompts: { system: string; user: string };
      rawResponseId?: string;
      stageId: string;
      tokenUsage: TLlmTokenUsage;
      validationError?: string;
  }

  • at: number

  • attempt: number

    1, 2, ... — one event per LLM-call attempt.

  • kind: "stage:llm-call"

  • output: unknown

    Raw LLM JSON as returned. May not conform to the stage's outputSchema if validationError is set.

  • prompts: { system: string; user: string }

    The actual prompts sent for this attempt (the user message includes any retry-suffix appended on attempt 2+ after a prior schema-validation failure).

  • OptionalrawResponseId?: string

    The OpenAI response id for this attempt, when the provider surfaces one. Absent for providers that do not expose a response id (e.g. the chat-completions provider) and for attempts where the id could not be captured before completion.

  • stageId: string

  • tokenUsage: TLlmTokenUsage

  • OptionalvalidationError?: string

    Set iff outputSchema rejected this output. When the schema accepted the output the field is undefined.