Completion streams

Use createCompletionStream(...) when one direct model call should yield provider-neutral events as they arrive.

Stream text

import { createCompletionStream } from '@anvia/core'

const events = createCompletionStream(model, {
  instructions: 'Write clearly and concisely.',
  input: 'Write a short launch note for Anvia.',
})

for await (const event of events) {
  if (event.type === 'text_delta') {
    process.stdout.write(event.delta)
  }

  if (event.type === 'final') {
    console.log('\nTokens:', event.response.usage.totalTokens)
  }
}

The stream finishes with a final event containing the normalized completion response.

What it can emit

Besides text, a provider may emit reasoning deltas, tool-call deltas, completed tool calls, sources, provider-executed tool calls, a message ID, a final response, or an error.

Direct completion streaming does not execute local tools or continue an agent loop. Use an agent stream when tool calls should run and feed results back to the model.

Check capability support

The selected model must support streaming and every optional request feature. Anvia throws CompletionCapabilityError before transport when the request is incompatible with the model.