The Directive Frame Is A Child Frame, Not An Overlay

A directive's answer should be readable and actionable without the session view taking over the window layout, so a directive opens a floating child frame at its source position showing its bound execution session's view. The frame displays the real view buffer, so permissions, Ask, patch review, streaming, and the composer work in it unmodified and no second renderer exists. It takes focus when opened explicitly from the directive's actions and stays unfocused when a dispatch opens it, since a request the user just started should not move point.

An overlay after-string was the obvious cheaper surface and cannot work. Interactions are point-addressed: mevedel--prompt--overlay-at-point resolves its target through (point), and the approve/deny/feedback keymap is an overlay keymap that is live only while point is inside the overlay. Point cannot enter a display string, so an overlay surface fails three ways at once — the keymap never activates, the lookup returns nil even with globally bound keys, and there is no position from which to settle the callback. Making it work would mean re-addressing the interaction layer from point to identity, which is a larger change than the surface is worth. Reading a response never needed point, so a read-only peek would have worked; it is omitted because the frame subsumes it, not because it was impossible.

Filtering the frame to one directive's turns uses invisible text properties and the buffer's invisibility spec, which makes the filter a property of the buffer rather than of the frame. An indirect buffer would give each display its own spec over shared text, but overlays are not shared between an indirect buffer and its base, so every interaction overlay would stay behind and the frame would show no prompts — the same failure as the overlay, reached from the other direction. Rendering a filtered copy into its own buffer is the hybrid 0091 already rejected for producing two renderers and two interaction owners. The frame therefore filters only while it is the sole window showing that view, and otherwise displays the full transcript positioned at the directive's turn.

0105 does not apply. The directive frame is a display geometry for the primary transcript surface, not a cockpit surface, and it adds no fourth archetype. The shared q/?/g key contract is independently impossible here: the view buffer contains an editable composer, so single-letter bindings would break typing. The frame binds only a dismiss and a filter toggle, both prefixed, and leaves the rest of the keymap to the view. Opening the frame enters the directive's composer scope and closing it leaves that scope, so a later ordinary-chat message cannot silently become a directive follow-up.