How routing works

Every command goes through a parser (Claude Sonnet 4.6) which returns a task plan as JSON. Tasks can be parallel or sequential — when one task depends on another (e.g. an HTML page that needs generated images), the parser sets a dependsOn field.

The executor then:

1. Runs tasks with no dependencies in parallel.
2. Waits for each child's parents to finish, joins their outputs into the child's context, and runs.
3. Streams every chunk back to the UI as it arrives via Server-Sent Events.

When Claude is the orchestrator (and other models depend on it), it switches into creative-director mode and writes labeled sections like [FLUX]...[/FLUX], [ELEVENLABS]...[/ELEVENLABS] — each section feeds the matching downstream model as its prompt.