Skip to content

Interface: ComposeOptions

Defined in: aiui-lowering-pipeline/src/types.ts:686

Options for composeIntent.

Properties

cwd?

optional cwd?: string

Defined in: aiui-lowering-pipeline/src/types.ts:693

The agent's working directory: screenshot paths AND source locations under it render relative (shorter, stable across machines); paths outside it stay absolute. Only the channel passes this — the browser has no cwd and its compose is a preview, not the committed prompt.


streaming?

optional streaming?: boolean

Defined in: aiui-lowering-pipeline/src/types.ts:709

Compose a provisional text run for each segment that has transcript -deltas but no final yet — the words you are still speaking. Off by default, and the channel never turns it on: what gets sent is built from finals alone, so in-flight words never reach a prompt or a paid call.

The transcript preview turns it on, and gets one thing for free that it used to fake: with a text run to anchor against, the existing timestamp interleave (below) drops a mid-utterance screenshot where it was taken, live, instead of stacking shots ahead of the segment until the final arrives and reorders everything at once. Streaming transcribers (ElevenLabs Scribe, OpenAI realtime, Gemini Live) carry the deltas that make this possible; a whole-segment REST transcriber has none, so its shots keep their arrival position, exactly as before.