Start and Finish Parts

Overview

When streamText runs, each generation step emits a step-start part at the beginning and a step-finish part at the end. A “step” is one LLM call — in a multi-step scenario with tool calls, a single streamText invocation can produce multiple steps (controlled by maxSteps).

The step-finish part carries finishReason (why the step ended), usage (tokens for this step), and isContinued (whether the next step continues the same text). This information is critical for monitoring, cost tracking, and understanding multi-step behavior.

Two callbacks provide the same data in a more convenient form: onStepFinish fires after each individual step, and onFinish fires once after all steps are complete with aggregated totals.

Stream Part Types

step-start

Emitted when a generation step begins. In fullStream:

for await (const part of result.fullStream) {
  if (part.type === 'step-start') {
    console.log('New generation step started');
  }
}

step-finish

Emitted when a generation step completes:

for await (const part of result.fullStream) {
  if (part.type === 'step-finish') {
    console.log('Reason:', part.finishReason);
    console.log('Tokens:', part.usage.totalTokens);
    console.log('Continued:', part.isContinued);
  }
}

finishReason Values

Value	Meaning
`'stop'`	LLM finished naturally
`'length'`	Token limit reached (response may be truncated)
`'content-filter'`	Content was filtered by the provider
`'tool-calls'`	LLM wants to call tools (step continues)
`'error'`	An error occurred
`'other'`	Provider-specific reason

Callbacks: onStepFinish and onFinish

import { streamText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';

const result = streamText({
  model: anthropic('claude-sonnet-4-5-20250514'),
  prompt: 'What is the weather in Berlin and Tokyo?',
  tools: { weather: weatherTool },
  maxSteps: 5,

  // Called after EACH step
  onStepFinish({ stepType, finishReason, usage }) {
    console.log(`Step (${stepType}): ${finishReason}`);
    console.log(`  Tokens: ${usage.totalTokens}`);
  },

  // Called once after ALL steps
  onFinish({ text, totalUsage, steps, finishReason }) {
    console.log(`Done. Reason: ${finishReason}`);
    console.log(`Total tokens: ${totalUsage.totalTokens}`);
    console.log(`Steps: ${steps.length}`);
  },
});

Callback Parameters

onStepFinish

Property	Type	Description
`stepType`	`'initial' \| 'continue' \| 'tool-result'`	What triggered this step
`finishReason`	`FinishReason`	Why the step ended
`usage`	`{ promptTokens, completionTokens, totalTokens }`	Token usage for this step
`text`	`string`	Text generated in this step
`toolCalls`	`ToolCall[]`	Tool calls made in this step
`toolResults`	`ToolResult[]`	Tool results from this step

onFinish

Property	Type	Description
`text`	`string`	Complete generated text across all steps
`finishReason`	`FinishReason`	Final finish reason
`usage`	`TokenUsage`	Token usage of the last step
`totalUsage`	`TokenUsage`	Aggregated tokens across all steps
`steps`	`Step[]`	Array of all step results