Inline AI assistance: Cursor-style Tab completion, ghost previews, and a live suggestions panel

[mitulgarg]

• I agree to follow the Code of Conduct that this project adheres to.
• I have searched the issue tracker for a feature request that matches the one I want to file, without success.

Inline AI assistance: Cursor-style Tab completion, ghost previews, and a live suggestions panel

Is your feature request related to a problem? Please describe.
Diagramming is inherently iterative — you place a shape, adjust alignment, draw a connector, and realise the layout is off. The existing Generate tool (sparkle button) is great for one-shot prompt-to-diagram, but it's too heavy for small in-flight decisions like "this node probably belongs next to that one" or "this branch needs a decision diamond before the merge." There's no way to get lightweight, context-aware suggestions while editing without interrupting the drawing flow entirely.

Describe the solution you'd like
Inline AI assistance that works alongside the user during editing, not instead of it — similar to how Cursor's Tab completion works in code editors.
Tab-to-accept ghost previews. As the user works, the assistant predicts the next likely shape, edge, or label and renders it as a low-opacity ghost on the canvas. Tab accepts, Esc dismisses, any other action ignores it. Triggers after placing a shape, after a short pause near an obvious extension point, or on demand via hotkey. The user controls when suggestions appear.
Live suggestions side panel. An optional right-side panel accepts a context prompt (e.g. "auth flow for a microservice") and surfaces 3–5 ranked suggestions in real time as the diagram evolves. Each is a previewable patch — add / remove / move / relabel / restyle. Hover to ghost-preview on canvas, click to apply.
Iterate-and-edit loop. After accepting a suggestion, the user can refine in natural language or drop straight back into manual editing without losing context.
Auto-fix nudges. Separate from the LLM suggestions: lightweight, deterministic fixes surfaced inline — misaligned nodes, inconsistent spacing, crossed edges with a cleaner routing, unlabeled connectors in an otherwise labeled set. Usable without an API key.

Describe alternatives you've considered
Using the existing Generate / Update / Assist actions — these replace or regenerate the diagram rather than augmenting an in-progress edit. The interaction model is different: you leave the canvas, describe what you want, and get a result back. There's no way to stay in the drawing flow.
Manual editing — fine for simple diagrams, but tedious for anything with consistent layout rules or repeated structural patterns.
Third-party wrappers work around the problem by building a separate UI on top of draw.io, but lose the native editing experience.

Additional context
The existing AI infrastructure in jgraph/drawio — configurable backends (gptApiKey, claudeApiKey), the assist action, and aiGlobals for system prompt customisation — is a natural extension point. A new inline or complete action with a streaming-friendly system prompt could power both the ghost preview and the side panel without a full redesign.
Implementation notes:

The model should receive a compact diff of recent edits + the surrounding sub-graph, not the full , to keep latency and token cost low.
Ghost previews need a render hook that does not pollute undo history until accepted.
Auto-fix should be a separate module, no API key required.

Open questions: debounce/trigger heuristics for ghost predictions; whether the side panel should be plugin-extensible; caching policy to avoid re-calling the LLM on minor canvas changes.

[draw-io-triage-bot[bot]]

👋 Thank you for opening this issue!

---

Summary: The user is requesting a comprehensive inline AI assistance system for draw.io, including Cursor-style Tab-to-accept ghost previews of predicted next shapes/edges, a live AI suggestions side panel with previewable patches, and deterministic auto-fix nudges for alignment/spacing — all designed to augment the editing flow without interrupting it.

---

📝 Additional Information Needed

To help us investigate this more effectively, could you please provide:

• Are you aware that draw.io is primarily a community-driven open-source project and that the AI features (Generate/Assist) rely on external API keys? A feature of this scope (real-time streaming LLM predictions, ghost rendering, side panel) would represent a very large engineering effort — is this something you'd be willing to contribute via a PR, or is this purely a request for the maintainers?
• Have you experimented with the existing Assist action (which can modify a selected portion of the diagram via natural language) to see how close it gets to the 'iterate-and-edit loop' you describe? Understanding the specific gaps might help scope a more incremental proposal.
• For the 'auto-fix nudges' (alignment, spacing, edge routing) — are you aware of the existing alignment/distribute tools and the 'Reset Edge' functionality? Would enhancements to those deterministic tools address part of your need independently of the AI features?

---

---

_{🤖 This is an automated response to help with initial triage. A human maintainer will review this issue. If you've provided all the requested information, no further action is needed - we'll get back to you soon!}