Anvia Docs

Use createCompletionStream when you want normalized provider stream events without agent runtime events.

import { createCompletionStream } from "@anvia/core";

for await (const event of createCompletionStream(model, {
  input: "Write a short launch note.",
})) {
  if (event.type === "text_delta") {
    process.stdout.write(event.delta);
  }

  if (event.type === "final") {
    console.log(event.response.usage);
  }
}

createCompletionStream returns the model's CompletionStreamEvent values directly.

Event Types

Direct completion streams can include:

Event	Meaning
`text_delta`	Assistant text arrived
`reasoning_delta`	Reasoning text or summary arrived from a provider that exposes it
`tool_call_delta`	Partial tool call data arrived
`tool_call`	The model emitted a complete tool call
`message_id`	The provider emitted a message id
`final`	The model stream completed with a normalized response
`error`	The stream failed

Raw Model Stream, Not Agent Stream

Direct completion streams do not execute tools and do not emit agent stream events such as:

turn_start
turn_end
tool_result
agent_tool_event
agent final output

For streams with tool execution and turn events, use agent.prompt(...).stream().

Capability Validation

createCompletionStream throws before returning the model stream if streaming or another requested feature is unsupported:

try {
  const stream = createCompletionStream(model, {
    input: "Write a live status update.",
  });

  for await (const event of stream) {
    console.log(event.type);
  }
} catch (error) {
  console.error(error);
}

Streaming Completions

Event Types

Raw Model Stream, Not Agent Stream

Capability Validation