Manager
Manager is the default SI Agent. It is a pass-through orchestrator: it must not rewrite your prompts. It routes to tools. Video / image / audio defaults come from ~/.knoxmedia/defaults.yaml — the agent is instructed never to ask about duration, resolution, or ratio.
Typical turn
- Sign in.
- Type, tap a chip, drop files, or
@a target. - Send. Stop generating (Esc) while streaming.
- Results land in
~/.knoxmedia/{video,image,audio}/and on the timeline. Video is async (Submitted to generation queue); image and audio are synchronous.
Tools
| Group | Tools |
|---|---|
| Library | library_search, library_read, library_write, library_remove |
| Store | memory_read, memory_search |
| Plan | plan_generation (content >15 s per clip) |
| Generate | generate_video, generate_image, generate_audio |
| Audio | isolate_audio, change_voice, force_align, list_audio_isolation, delete_audio_isolation, voices |
| Context | get_role_context, get_scene_context, get_scene_environment, set_scene_environment, set_scene_reference_image, get_timeline_context |
| Place / edit | timeline_place, editor_instruct |
| Clarify | ask_user |
| Director (when that mode is on) | director_* continuity and generate helpers |
Video routing (automatic)
| You attach | Route |
|---|---|
| Text only | Text-to-video |
first_frame / an image | Image-to-video |
| Reference videos / audios or many images | Reference-to-video |
Caps: reference videos up to 3, total ≤ 15 s; reference audios 2–15 s each, up to 3, total ≤ 15 s. Image default size 1920×1080, divisible by 16. GPT Image 2 generate vs edit from reference images.
What Manager will not do
- It does not write documents itself —
library_read/library_write, or hand off to Writer. - Captions / transcripts of existing clips go to Editor (
editor_instruct), notgenerate_video. - Coding is later. Treat source as text.