The inspector

Select a node and the inspector shows exactly the fields that node’s type reads — no more, no less. The forms are generated from the same catalog that feeds the linter and the AI assistant, so what you see is what the framework supports.

Anatomy

Part	Purpose
Header	Node icon, editable label, type chip, and a menu (duplicate, mute, delete, copy node JSON)
Issues	The linter’s findings for this node, each linking to the field that caused it
Fields	The node’s parameters, grouped into main and advanced
Run condition	skip_if_false, for every transformation
Worth knowing	The behaviors that usually live only in the framework source
Examples	Valid snippets for this node type
Notes	Free-form annotation, never compiled

Field types

The widget follows the JSON shape the framework expects:

Widget	For	Notes
Text	names, paths, single values
SQL	conditions, expressions, queries	monospace; highlights {param} and {{variable}} placeholders it detects
Number / Toggle	numeric and boolean options
Select	fixed choices (how, mode, method)	only the values the framework accepts
Chips	ordered string lists (columns, by, merge_keys)	paste a comma-separated list to add many at once
Expression rows	ordered maps (agg, with_column.columns)	order is preserved because it is load-bearing
Key/value rows	ordered maps (cast.columns, rename.mappings)	order matters for rename
JSON	nested structures (struct.fields, pivot)	live parse feedback under the box

Ordered widgets exist because order is semantic in the framework: rename applies mappings sequentially, and with_column.columns lets a later expression use a column an earlier one created.

Sources and destinations

Picking a format rewrites the form to that connector’s options — the ones it actually reads, with the framework’s defaults prefilled. Switching format drops options the new one does not define, so a config never carries dead keys from a previous choice.

Destinations add:

• Write mode, restricted to the modes that format supports
• Partition by, hidden for connectors where it means nothing
• Column projection, a toggle plus a chip list
• Options, including the merge fields that appear only in merge mode

Validations

The validations node has its own form: the on_failure policy with a one-line explanation of each mode, a rule list where every rule renders its own fields, and an optional report destination that reuses the full output form.

Run condition

Every transformation carries skip_if_false. The inspector documents the three cases in place: an empty value skips the step, a boolean expression skips it when false, anything else runs it.

Note

The expression sees only literals — the values already substituted — never DataFrame columns. It is a per-run switch, not a row filter.

Worth knowing

Each node type carries the gotchas that cost real debugging time:

• select entries go through F.expr, so odd column names need backticks
• union matches by position unless allow_missing_columns is on
• group_by.agg takes complete SQL expressions with aliases, never a map
• merge needs a unique constraint on the merge keys
• collect triggers a driver action, so it belongs after a checkpoint

They live next to the field they apply to, which is where they are useful.

Jumping from an issue

Issues in the panel — and the badge on a node — link straight to the field that caused them. Studio opens collapsed sections and scrolls the control into view, so a lint result is one click from a fix.