Changelog

Enriched the cloud system prompt with personality traits: honest but

empathetic tone, balanced verbosity, clean dry humor, neutral error

handling that pivots to solutions. Technical depth now adapts to the

user via memory context rather than assuming expertise level.

RAG search results were dominated by the active plugin's data, making

cross-plugin queries (e.g. asking about Notes from Tasks) return no

results from the mentioned plugin. Fix: fetch 3x candidates, then

apply plugin diversity via round-robin interleaving. When the user

mentions a specific plugin by name (e.g. "notes", "calendar"),

half the result slots are reserved for that plugin's data.

The root cause of the AI responding conversationally ("Shall I go ahead?"

"Creating those now!") without actually executing actions was that RAG

returned 0 intent_action chunks. Entity data (the user's actual tasks,

notes, etc.) crowded out intent_action chunks in the general semantic

search because task titles are closer to the query than "ACTION: Create

Task" descriptors.

Fix: BuildRagContextWithIntentsAsync now performs TWO parallel searches —

one general (for entity data + plugin context) and one specifically

filtered to intent_action entity type. Results are deduped and merged,

ensuring the AI always sees relevant available actions alongside the

user's data context.

Also fixed incorrect slot names in ActionFormatHeader for tasks.add_link

(was source_task_id/target_entity_id, corrected to task_id/target_id).

Desktop version 1.65.6 → 1.65.7.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

ShellContentProvider now discovers all IQuickActionProvider plugins at

indexing time and creates individual RAG chunks per quick action. This

replaces the hardcoded shortcut list that was manually maintained and

had drifted out of sync (e.g. listed Cmd+B for transactions instead of

Cmd+Shift+T).

Each quick action gets its own searchable chunk with the action name,

plugin, description, keyboard shortcut, and usage instructions. The

global shortcuts chunk now dynamically appends plugin shortcuts instead

of hardcoding them, so adding a new quick action to any plugin

automatically makes it discoverable by Duncan via semantic search.

Shell-level shortcuts (Cmd+K, Cmd+I, Cmd+M, etc.) remain hardcoded

since they are not plugin-declared.

Desktop version 1.65.5 → 1.65.6.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Three fixes for the AI creating duplicate tasks and inventing fake slots:

1. IntentEngine.ExecuteDirectAsync now validates slot names against the

IntentDescriptor before calling the plugin. Unknown slots (e.g.

"parent_task_id", "parent_id") are stripped and warnings attached to

the IntentResult. This prevents silent data loss.

2. Action execution results are now injected into conversation history

as an [Action Results] block. On the next turn, the AI sees what

succeeded, what failed, and what warnings were raised — enabling

self-correction instead of blindly retrying or duplicating.

3. ActionFormatHeader updated with explicit guidance: only use declared

slot names, use tasks.add_link for linking existing entities instead

of re-creating them, and acknowledge action result feedback.

Desktop version 1.65.4 → 1.65.5.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>