Changelog

New SDK message type for dataset insight generation. The Data plugin

sends this message with column metadata and sample rows when the user

clicks "Generate Insights". The shell's DatasetInsightOrchestrator

subscribes and runs AI analysis.

Add email_message and web_clip to EntityTypeMap so the AI tray sends

context for those entity types. For notes containing embedded

datasets/tables, parse dataset_id references from content and fetch

actual row data via IPluginDataSourceProvider. Add fallback entity

fetching through DataSourceProvider for unmapped types like dataset_row.

Fix history tab: clicking a previous conversation now switches to the

chat view by subscribing to SelectedTabIndex PropertyChanged in the

code-behind.

When the user views an item and chats with the AI, cloud models now receive

the full entity JSON (fetched via SDK Read) instead of just the title. This

gives the AI complete knowledge of the note body, task description, contact

details, etc. for the item the user is looking at.

Entity data is fetched async on active item change via SdkMessage Read,

serialized as indented JSON, and capped at 8K chars to avoid prompt bloat.

Local models still get the short "Currently viewing: Title (type)" summary

since their context windows are too small for full entity payloads.

Tabbed AI Panel: Split the single-panel AI tray into Chat, Intents, and History

tabs. Chat shows free-form conversation, Intents shows intent suggestions and

content suggestion cards, History shows browseable past sessions.

Conversation Persistence: New AiConversationStore saves chat sessions to disk

(ai-conversations.json) with debounced writes. Sessions are auto-titled from the

first user message, capped at 100 with oldest auto-pruned.

Token Tracking: Each AI provider now declares ContextWindowTokens per model. The

chat tracks estimated token usage against the active model's context window. At

80% capacity, a warning banner suggests starting a new chat.

Active Context Injection: The AI system prompt now includes what the user is

currently viewing (active plugin item from InfoPanelService) for more relevant

responses. Cloud models get full detail fields, local models get a short summary.

Split ViewModel: AiSuggestionTrayViewModel refactored into partial files — main

orchestrator, Chat (persistence + tokens), Intents (suggestion handlers), and

History (session browsing/deletion).

Version bump: Desktop 1.51.1 → 1.52.0

Added int ContextWindowTokens property to AiModelInfo record in PrivStack.Sdk

to expose per-model context window sizes. This enables the desktop AI chat tray

to track token usage against model limits and warn users when approaching context

exhaustion. Backward-compatible — defaults to 0 for existing consumers.