Probe P01: do region-gated modelling operators really need a GUI?

Note

Generated from docs/discovery/probes/p01-context-override.md.

• Date: 2026-10-08 · Blender: 5.2.2 LTS, background mode (-b --factory-startup --disable-autoexec)
• Question: the S01 trace found 23 modelling operators whose poll demands a GUI region (gate: gui-region). Is that a real dependency, or does the poll only encode where the button lives (ledger L-004)?
• Script: tools/probes/p01_context_override.py. Each operator runs in its own Blender process, so a crash is recorded instead of ending the run. Raw data: ~/newblender-data/probes/p01.json (local only).
• Evidence level reached: L2 (reproduced at runtime).

Method

Scene: the factory cube, subdivided twice (98 vertices). For each operator:

1. Plain poll. Call poll() in the windowless background context.
2. Borrowed context. Take the 3D View area and region from the Layout screen stored in the startup file, and call poll() again under bpy.context.temp_override(screen=…, area=…, region=…). That region has never been drawn: its view data (region.data) is None.
3. Exec. If the poll passes, call the operator with EXEC_DEFAULT, and compare a fingerprint before and after: vertex positions, object matrix, selection, object count and transform orientation.

Operators whose default arguments do nothing were given real ones: a shear angle, an orientation name, a cube matrix, an 8-point brush stroke, and edge or vertex indices for the picking operators. That way “nothing changed” means the call failed, not that it was asked to do nothing.

Result

Outcome	Count	Operators
Plain poll passes headless	0 / 23	(confirms the trace)
Poll passes once any region is supplied	21 / 23	all except transform.delete_orientation (needs an existing custom orientation: a data precondition) and transform.seq_slide (no sequencer area in the startup screens)
Ran and did the job, verified by the fingerprint	5	mesh.loop_select, mesh.edgering_select, mesh.shortest_path_pick (all given element indices), mesh.primitive_cube_add_gizmo, transform.create_orientation
Returned FINISHED, but nothing measurable changed	4	transform.shear (angle 0.5), sculpt.brush_stroke (8-point stroke), mesh.rip_edge, mesh.rip_edge_move
FINISHED; the effect is outside what the fingerprint measures	2	sculpt.set_pivot_position, transform.select_orientation
CANCELLED: the gesture needs a path or points the probe did not supply	6	sculpt.face_set_polyline_gesture, face_set_lasso_gesture, trim_lasso_gesture, trim_polyline_gesture and similar
Error from a data precondition	1	mesh.knife_project (needs a second object with wire edges)
Segmentation fault (Blender crashes)	5	sculpt.face_set_line_gesture, face_set_box_gesture, project_line_gesture, trim_line_gesture, trim_box_gesture, all inside sculpt undo (undo::push_begin_ex, undo::geometry_begin_ex)

What it means

1. Polls check where the button lives, not what the code needs. (Ledger L-004, now at L2.)
   • The three picking operators run correctly with a region whose view data is None. The region only satisfies the poll; the exec never touches it.
   • loop_select’s exec reads object_index and edge_index, as the trace said. Both are hidden properties that default to −1, which cancels the call.
   • In our first run the probe forgot the indices and got CANCELLED. That briefly looked like evidence against the trace; reading the exec body showed the missing input.
   • Lesson: a hidden required input is the same anti-pattern seen from the agent’s side.
2. A status is not a result. (Ledger L-003, new runtime evidence.)
   • Four operators reported FINISHED with real arguments and changed nothing measurable.
   • The likely cause is that they need view data that a borrowed, never-drawn region does not have. Shear, rip and brush strokes all project through the view.
   • The operator gives no sign of this. Only the fingerprint shows it. An agent trusting {'FINISHED'} would believe it had sheared the object.
3. Some code assumes a window exists, and crashes without one. (Ledger L-008 and L-011.)
   • Five sculpt gesture operators crash Blender in their undo push. Sculpt undo assumes the windowed undo system is set up, and in a background session it isn’t.
   • That’s not a poll problem, but a hard dependency on the interactive session buried in the edit path.
   • For an agent this is the worst case: no error to read, just a dead process.

Limits

• No GUI baseline. The same calls were not repeated in a windowed session to confirm that, there, they change the scene. “Did nothing headless” is measured, but “works in the GUI” is assumed from normal use.
• The six CANCELLED gesture operators were not given point lists, so this probe says nothing about them yet.
• Single-object scene. Multi-object editing was not probed.

Next

• Repeat the four silent no-ops with a drawn region (one windowed session) to tell “needs view data” apart from “broken headless”.
• Give the polyline and lasso gestures explicit point lists.
• Probe P02, undo: do scripted calls push undo steps, and does sculpt undo exist at all in -b?