CanvasAnvil is an open source AI diagram maker and flowchart generator - an infinite-canvas diagram editor and self-hosted drawio alternative that writes editable .drawio files. Text to diagram, then fix nodes, layout and labels in place. Also AI presentations, interior design CAD/BOM, posters and infographics.
TypeScript ai-agentsai-diagramai-presentationbomcadcanvasdiagramdiagram-generatordrawioflowchartinfographicspptppt-generatorpresentation-generatorreactsvgedittext-to-diagramtypescriptworkflow
docs: record the chat attachment fix
codingFeng1013a28a738c00f
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 6d84699..ecb0718 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -51,6 +51,7 @@ exist once instead of per workspace.- Fixed text blocks read back from storage reaching the review canvas without their geometry, where they became boxes at undefined coordinates that could be neither seen nor selected.- Fixed the chat upload limits applying to one of the three shapes an attachment can arrive in, so a file sent as `image` or in a `content` array skipped both the size cap and the file count.- Fixed the Flow composer never growing past three lines. The shared `Textarea` did not forward its ref, so the height it measured on every keystroke was never applied and the box stayed at its minimum, with resizing disabled.+- Fixed pictures attached in the chat panel reaching only the branch that edits an existing slide image. Every branch that redraws the slide dropped them -- which is the branch a request to add a picture takes, so the attachment was lost exactly when it was the point of the message. They now travel with all five generation and edit paths, named so the prompt can say where each one goes.- Fixed a slide's material images never reaching the image model. The picture was read from the description, collected and sent, then dropped by the server: the edits route attached only the first reference image, which the template had taken, and the generations route sends no images at all. Because the prompt still named each material and told the model to place it, the slide came back with an invented picture rather than a visible gap. Only the materials a description actually refers to are sent now, and image generation carries up to six references instead of sharing the vision cap of three.- Fixed the chosen template being outvoted by the outline. The planner had no parameter for a template and had never seen one, yet it wrote a background colour and art style into every slide's visual description; that concrete text then beat the template reference at render time, so a deck built on the dark navy template came out white and flat. The planner is now given the template image and writes descriptions that match its palette and art style, and the template still outranks a description that disagrees with it at render time.- Fixed a slide displaying with its text erased after an editable export, from the adapter preferring a render layer's background -- the text-stripped derivative -- over the one the deck view passed.