feat: context-aware activity suggestions — AI-powered "what next?" in the Activity Palette

[chelslava]

Vision

When a user selects a node on the canvas, the Activity Palette shows a "Suggested Next" section at the top with 3–5 activities that are statistically or semantically likely to follow the selected activity in an RPA flow.

This is a modern RPA trend (seen in Power Automate Copilot, UiPath Autopilot) and aligns with the existing AI infrastructure in RPAForge.

Approach A — Heuristic (no LLM, good first version)

Analyze existing diagram templates (packages/studio/src/templates/) and committed user diagrams to build a co-occurrence frequency table:

P(next = "Click Element" | current = "Navigate") = 0.78
P(next = "Get Element Text" | current = "Click Element") = 0.45

Store the table as a static JSON asset, updated by a build-time script. No API calls, instant suggestions.

Approach B — LLM-powered (extends existing AI infrastructure)

Reuse the ai:generateDiagram IPC infrastructure to ask the LLM for continuations:

Given this partial RPA diagram ending at [activity], what are the 5 most likely next steps?

Cache responses per (current_activity, context_hash) to avoid redundant API calls.

UI Integration

Activity Palette
┌─────────────────────────┐
│ 🔮 Suggested            │
│  ▸ Click Element        │
│  ▸ Get Element Text     │
│  ▸ Take Screenshot      │
├─────────────────────────┤
│ All Activities          │
│  ...existing list...    │
└─────────────────────────┘

The suggestion strip appears only when a canvas node is selected; collapses otherwise.

Relevant Files

• packages/studio/src/components/Designer/ActivityPalette.tsx
• packages/studio/src/stores/ (may need a new suggestionStore)
• packages/studio/src/templates/index.ts (source for co-occurrence mining)
• packages/studio/electron/ai/ (for Approach B)

Acceptance Criteria

• Suggestions appear within 100 ms of node selection (heuristic approach) or with loading indicator (LLM approach)
• Suggestions are dismissible / collapsible per user preference
• Works without an AI provider configured (Approach A as baseline)
• Suggestion quality metric (click-through rate) can be measured via debug logs

[chelslava]

⏳ Deferring to v0.5.0+ roadmap

Analysis

This is an excellent UX enhancement feature, but after review:

Current Status:

• Not in v0.4.0 roadmap
• Low priority (enhancement, not blocking)
• No explicit user demand yet (0 comments, 0 reactions)
• Would require either:
  • Approach A (heuristic): Build co-occurrence table from templates + existing diagrams
  • Approach B (LLM): Extend AI infrastructure with context-aware continuation prompts

Effort Estimate:

• Approach A: ~8-12 hours (data mining + caching)
• Approach B: ~12-16 hours (AI prompt engineering + caching)

Next Steps

Reopening criteria:

• ✅ Explicit user request in a new issue
• ✅ Prioritization in v0.5.0 roadmap planning
• ✅ Available developer capacity

For contributors interested:

• This is a good advanced-feature candidate
• Heuristic approach (A) is safer MVP
• Document co-occurrence patterns in before implementation

Recommendation

Defer to future backlog. Will reopen if user demand grows or v0.5.0 planning includes this. 👋