Changelog

Replace the full intent catalog injection in the system prompt with a

RAG-based approach. ShellContentProvider now indexes each registered

intent as a separate "intent_action" chunk in the vector index. When

the user sends a chat message, RAG search finds semantically relevant

intents and includes them in context. The ACTION format header is only

injected when intent_action chunks are found, keeping prompts lean for

general questions. This works for both cloud and local models, solving

the issue where local models never received intent instructions.

Three fixes:

1. Thinking animation never started: ChatMessageState.Loading is enum value 0,

which was also the default of _state. Setting State = Loading was a no-op

because the generated setter saw old == new and skipped OnStateChanged.

Fixed by initializing _state = ChatMessageState.Ready so the transition to

Loading is always detected.

2. UI lockup while Duncan thinks: The entire SendChatMessageAsync ran on the UI

thread — RAG vector search (synchronous FFI to Rust), embedding generation,

prompt building, and the AI API call all blocked the UI. Restructured to

capture UI-thread state upfront, then run all heavy work inside Task.Run.

UI updates (streaming tokens, state changes) dispatch back via _dispatcher.Post.

3. Added raw response logging to diagnose when the AI model fails to include

[ACTION] blocks despite the system prompt instructions.

Desktop 1.63.3 → 1.63.4.

Swapped TextBlock → SelectableTextBlock for user label, assistant content,

and error message in the message bubble template so text can be highlighted

and copied.

Added ClipToBounds="True" to all three tab panels (Chat, Intents, History)

to prevent history conversation items from overflowing outside the panel

bounds when the list exceeds the available height.

Desktop 1.63.2 → 1.63.3.

Duncan was claiming to have created tasks/projects without actually executing

anything — the AI responded with natural language confirmations but no [ACTION]

blocks. Root cause: the system prompt instructions were too passive.

Changes:

• Rewrote intent catalog prompt with CRITICAL header and explicit rules:

"WITHOUT an [ACTION] block, NOTHING happens" and "NEVER say 'I've created'

unless you include the [ACTION] block"

• Added instruction to refuse unavailable actions instead of fabricating
• Added action-oriented keywords (create/add/make/log/draft/save) to Medium

tier classifier so token budget is adequate for JSON output

• Added debug logging for action block parsing results and intent catalog

injection to aid future debugging

Desktop 1.63.1 → 1.63.2.