draw_pin_circle takes a PinSide and centers the circle exactly on the
left or right edge of the node. The reserved Dummy is half-width
(PIN_RADIUS instead of PIN_DIAMETER) so the inside layout stays
compact, and ed::PinRect is set to the full circle so the protruding
half is still grabbable.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Pin radius bumped to 9px (18px diameter). Easier to grab.
- Single neutral pin/cable color across all kinds (data is uniformly
vec4 — coloring by node kind was misleading).
- ed::StyleVar_NodePadding(0,8,0,8): zero left/right padding so the pin
circles sit flush with the node's edges, separated from the title
and controls by 8px gaps.
- Right-click on a pin clears all connections on that pin:
- input pin → clear that single slot's source_id
- output pin → clear every source_id across the pipeline pointing
to this node (fan-out).
- Right-click on a link still deletes that single link (existing).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Three-column node layout: input pins on the left edge, controls in
the middle, output pin on the right edge.
- Pins rendered as 14px filled circles with darker outline. Color is
the node's kind (gen=blue, op=violet, blend=amber, output=red).
- ed::PinPivotAlignment(0,0.5)/(1,0.5) so the cable starts/ends at
the circle center.
- Empty columns (gen with no inputs, output with no output) get a
pin-sized Dummy so column widths stay consistent.
- ed::Link now passes color (= source node kind) and 2.5px thickness.
- Right-click on a link deletes it immediately via
ed::ShowLinkContextMenu (no popup).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previously the cycle validator rejected any link whose source had a
vector index >= target's, which silently killed legitimate connections
between nodes added in the wrong drop order.
Switch to a DFS over source_ids: an edge from->to creates a cycle iff
`from` already (transitively) depends on `to`. topo_sort runs after
each topology change so the vector ends up in a consistent order
regardless of how nodes were inserted.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two bugs:
1. Dropped nodes were pushed to the end of the pipeline, but the Output
node already sat there from startup. The cycle validator and the
compiler only look for sources at indices strictly lower than the
target, so new nodes were invisible to the Output. Fix: insert
dropped nodes before the first Output; topo_sort also stable-moves
Output nodes to the back.
2. ColorEdit3 with default flags rendered RGB text inputs alongside the
swatch; clicking them dragged the node instead of opening the picker.
Fix: NoInputs + NoLabel leaves only the swatch (a single item), and
ed::Suspend/Resume wraps the call so the popup isn't clipped to the
node or captured by the canvas.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous InvisibleButton captured mouse events, so you could drag
from the Functions palette into the canvas, but node dragging and
slider interaction inside the canvas stopped working.
Fix: watch the global drag-drop payload without an explicit target. When
the mouse releases LMB over the DAG window with a "DAG_NODE_TYPE"
payload active, queue an add at that canvas position. No button, no
capture.
Tests (compiled standalone with preprocessor defines):
- dag_compile: 6/6 asserts (empty, single gen, op chain, multi-source
blend, Output-driven fragColor, unconnected-Output fallback).
- dag_catalog: 8/8 asserts (uniqueness, per-kind input invariants,
exactly one Output, body_glsl present & returns, control param
indices valid).
Build with:
g++ -std=c++17 -Icpp/functions -DDAG_COMPILE_TEST cpp/functions/gfx/dag_compile.cpp cpp/functions/gfx/dag_catalog.cpp -o /tmp/dag_compile_test
g++ -std=c++17 -Icpp/functions -DDAG_CATALOG_TEST cpp/functions/gfx/dag_catalog.cpp -o /tmp/dag_catalog_test
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previously SetNodePosition was called after EndNode every frame, which
reset any drag the user had done. Now the initial position is set once
per editor_uid (tracked in a static unordered_set), and the editor owns
the position afterwards. GetNodePosition at end-of-frame keeps step
state in sync for future persistence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
egutierrez