One canvas. Follow-ups, missing inputs and re-runs go into the active workflow_id; an unrelated goal starts a new one.
Solid edge, fixed asset. A canvas output reaches its consumer through an edge; an asset_id input is outside media or a result the user picked.
cawcut asset upload <path|url|asset_id>… --json
cawcut workflow nodes [--capability <c>] [--asset <asset_id>…] [--model <id> --schema | --has-port <selector>] --json
cawcut workflow quick "<prompt>" --capability <c> --model <id> [--workflow <id>] [--param k=v] [--image|--video|--audio <asset_id>] --estimate --json
cawcut workflow create --file - --json
cawcut workflow edit <id> --file - --json
cawcut workflow set <id> --node <n> key=value --json
cawcut workflow connect <id> <from> <to> --json
cawcut workflow upscale <id> --node <draft_node> --json
cawcut workflow prepare <id> [--target <n>] --json
cawcut workflow run <id> [--target <n>] --wait --json
cawcut workflow results <id> --node <n> --download --json
--json replies {v,ok,data} or {ok:false,error}; follow error.next. Ask with the host question tool, recommended option first; independent settings may share one call up to its question limit. Questions never arrive as a block of numbered prose; a numbered list is for a choice longer than the tool allows, or when no tool exists.
0. Bootstrap
Silent guardrail, once per AI session: a session that already ran a cawcut command successfully starts at step 1.
Otherwise run cawcut upgrade check --json.
"update_available": true→ runcawcut upgradeyourself; the user never types it. Then continue."ahead_of_registry": true→ a local build; continue.- Command missing, or any other check failure → failures. Done: the check ran, any update is installed, and no bootstrap text reaches the user.
1. Route
| Request | Owner |
|---|---|
| Named App, "what apps exist" | cawcut-app-run |
| Running or named Ad Video, Story Video or avatar project | that skill |
| Story: multi-shot narrative or dialogue video | story on a canvas; /cawcut-story-video on request |
| Script, storyboard or shot list only (ad scripts too) | script-writing |
| Ad or ad video, or a product-page URL alone | cawcut-ad-video; an ad asked for on a canvas → product |
| Product URL to scrape or extract | ask once: Ad Video wizard or canvas with a Product node → product |
| Cut, join, trim, timeline | cawcut-vn |
| Anything else new | cawcut app list --schema --json; offer an App scoring ≥70 (the rubric is in cawcut-app-run) beside "Build a canvas"; else canvas |
Done: the user agreed to an App, a pipeline or a canvas.
2. Prepare media
Outside media (file, URL, asset id): cawcut asset upload; relay its notice once. A canvas output ("the video above") is a solid edge, never an upload.
Done: outside inputs have an asset_id; canvas outputs have edges.
3. Plan nodes
nodes → capabilities; --capability → models; --model --schema → ports and parameters. Offer motion-transfer and object-swap only when the user names them. Scripts are text_input text, never a script node.
- Audio or video references, sound, frames, edits of existing media,
{{Image N}}→ plan-rules. - Video or audio input → add
--asset <asset_id>…to everynodes --capabilitycall; a model whose duration limit rejects it → plan-rules. - Story → story.
- Change an existing canvas, or "what next" → edit.
Model and settings (each paid node, before pricing)
- Model: the first one
nodesreturns (the default sits there), or the one the user named. - Read
nodes --capability <c> --model <id> --schemafor that exact model, again after every model change. Settings, defaults andoptionscome only from that reply; one model's options never carry to another (Seedance 2.5 Draft allows480ponly, Seedance 2.5 allows more). Settings menu, one question: the model and its key settings bylabel(ratio or aspect, resolution or quality, duration, count, audio; only those in its schema). Options: Use current settings (recommended) · Adjust settings · Change model. - Adjust settings: one question per key setting in a single call, options from the parameter's labelled
options, the default marked recommended. Other parameters keep their defaults. Write the answer under the parameter's key (quick --param key=value, orseton a canvas). - Change model (the schema read in step 2 repeats for the new model; settings the new model lacks or no longer allows reset to its defaults): one question with the first three models
nodesreturns (name, price) and More models; More models → every model as a numbered list (name, price), the turn's final reply, answered by number. Both keep the CLI's order. Then back to 2.
Done: each node has an input source, and each paid node a model and settings the user accepted.
4. Build, confirm, run
- Build and price:
- Active
workflow_id(this AI session ranquickorcreate; a script, an image or a video already sits on it): edit adds the new nodes and edges to it, thenprepare <id> --target <n>. Stages of one goal share one canvas: image → script → video. - No canvas yet, several nodes:
create→set/connect→prepare <id>(runscheck, returns credits). - No canvas yet, one node:
quick … --estimate(runscheck; returnsworkflow_id,node_id, credits).
- Active
- Ask one question: model, settings by label, scope, credits. Options: Confirm · Change model · Modify settings · Cancel. Each later paid run gets its own. Change model or Modify settings → the step 3 menus, then
switch-model <id> --node <n> --model <m>orset, andprepare --target. - After Confirm:
run <id> --target <n> --waiton the same workflow.
Done: Confirm for this scope, check passed, task_id read to its end state. Other cases → failures.
5. Deliver
delivery.
Done: outputs, actual cost, Portal link reach the user; cawcut-vn gets every path, images too.