Refine a plan#

Prerequisites: an initialized dashboard and a completed planning stage. Outcome: change the intended part of a task, then review all affected outputs.

Choose the right stage#

Change you want	Send feedback to	Downstream work
Reorder or add a subtask	Task Decomposer	Recipes, parameters, and vision are updated
Change controllers or their priority	Skill Generator	Parameters and vision are updated
Change a distance, angle, force, or timing value	Parameter Setter	Updated parameters continue to vision
Correct a point on the displayed captured image	Vision Module	Annotations are refined; regenerate the script

Refinement updates the selected stage and its downstream outputs.

Request a change#

Wait until the workspace is idle.
Describe the desired change in natural language.
Optionally include a subtask, parameter, or exact value if you have one in mind.
Submit to that stage’s feedback control.
Review the updated stage and every affected downstream stage.
Generate a new script only when grounding is current.

Feedback can express a high-level intent without naming a controller or giving an exact value. For example:

Can you lift the cube a bit higher?

If you want a specific adjustment, you can also write:

In subtask 3, increase the lift offset by 0.04 m. Keep the grasp timing and the approach subtask unchanged.

Both forms are supported. Subtask numbers and numeric values are optional; the agents interpret the request using the current plan and conversation context. Review the updated plan to check that the resulting adjustment matches your intent.