Partial Messages

The answer normally reaches the client once the agent has finished writing it. partialMessages adds an earlier stream, carrying it as it is being written.

new MicdropServer(socket, {
  stt,
  agent,
  tts,
  partialMessages: true,
})

It stays off by default, since a page that only displays finished messages has nothing to do with it. Any agent works, since the text is read from the stream the agent writes to.

Reading it in the client

The client keeps the answer being written in its state, next to the conversation:

Micdrop.on('StateChange', (state) => {
  answerText.textContent = state.partialAssistantMessage
})

An event carries the same text, for a page that would rather react than re-render:

Micdrop.on('PartialAssistantMessage', (content) => {
  answerText.textContent = content
})

Each one carries the whole answer so far, replacing the previous one. Display it as it comes, no reassembly needed.

It empties as soon as the settled message lands in conversation, so the two never show the same sentence twice. It also empties when the user interrupts the assistant, and when the answer is skipped.

Where it earns its place

A call without a voice is the clearest case: the answer would otherwise appear all at once, several seconds after the question. Streaming it makes a written answer feel as immediate as a spoken one.

In a spoken call, the voice already carries the answer, so partial messages are worth their bandwidth only when the page displays the text beside it.