Changelog

ResolveWikiLinksAsync was running on the UI thread before Task.Run,

performing up to 5 async database lookups via ILinkableItemProvider

.GetItemByIdAsync — each one blocking the UI. Moved it inside Task.Run.

BuildConversationHistory was also running on UI thread with LINQ over

the ObservableCollection. Replaced with a snapshot taken on the UI

thread (fast — just select/toList) and the filtering/trimming logic

moved inside Task.Run.

This eliminates the brief UI freeze between pressing Send and seeing

the streaming response begin. Bumps Desktop to 1.64.3.

Three chat quality fixes:

1. ACTION block parsing: AI models sometimes omit [/ACTION] closing tags

(e.g. two consecutive [ACTION]{...} blocks with no [/ACTION] between

them). Added UnclosedActionPattern regex that matches blocks terminated

by the next [ACTION], end of string, or a stray [/ACTION]. StrayActionTags

regex strips any remaining tag fragments from the displayed text.

2. Model stop token artifacts: Qwen models output a literal Ċ (U+010A)

character before stop tokens like <|im_end|>. The existing ChatTokenPattern

only stripped the tag itself, leaving the Ċ visible in chat. Updated the

regex to capture preceding U+010A/U+0120 chars, and added

TrailingModelArtifacts regex to clean any remaining instances at end of

output.

3. Cross-plugin action guidance: Updated ActionFormatHeader prompt to

explicitly tell the AI it can use actions from any plugin regardless of

which plugin the user is currently viewing.

Bumps Desktop to 1.64.2.

Three fixes for AI chat:

1. Wiki-link picker: typing [[ in the chat input opens a compact dropdown

that searches all ILinkableItemProvider instances. Keyboard nav (up/down/

enter/esc) and pointer selection insert [[type:id|Title]] into the chat

text, which the existing ResolveWikiLinksAsync then resolves for the AI.

2. Intent ID fuzzy fallback: IntentEngine.ExecuteDirectAsync now tries a

prefix match when exact match fails (e.g. AI sends "tasks.update" →

resolves to "tasks.update_task"). Logged as warning for monitoring.

3. Prompt improvements: ActionFormatHeader now explicitly instructs the AI

to use exact intent IDs, to use cross-plugin actions regardless of the

active plugin, and moves the "can't do that" fallback to only apply when

no matching action exists in the discovered descriptions.

Bumps Desktop to 1.64.1.

After downloading a local model, the settings handler now pre-loads it

into memory immediately instead of requiring a restart. The download

status progresses through "Download complete — loading model..." to

"Ready — model loaded". Also auto-selects the model in settings if none

was previously selected. Added PreloadModelAsync to LocalLlamaProvider

for explicit warm-up without running inference. Bump to 1.63.7.

Chat input bar and token warning banner now only appear on the Chat tab.

Previously they were always visible across all three tabs. Also keeps

ClipToBounds on all panels to prevent content overflow. Bump to 1.63.6.