Changelog

The regex-based ACTION block parser used [^[]* which broke on any JSON

containing literal [ characters (e.g. checklist arrays, tags arrays).

This caused 0 blocks parsed despite valid-looking [ACTION] output,

silently dropping AI actions that would have updated tasks.

Replaced regex approach with brace-depth counting that handles nested

{}, [], and quoted strings correctly. Also added FlattenSlotValue()

which converts JSON arrays to newline-separated strings — so when the

AI sends "add_checklist": ["item1", "item2"], it becomes the

newline-delimited format the intent handler expects.

Added ACTION format header guidance for slot naming (add_checklist not

checklist) and array support.

Version: 1.65.1 -> 1.65.2

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

Cloud providers handle their own context limits and bill per token, so

artificially capping them at 200/1200/2500 caused truncation errors

(MAX_TOKENS from Gemini). Raised cloud token budgets to 1024/4096/8192

— the system prompt's length guidance still controls actual verbosity.

Also bumped AiMemoryExtractor MaxTokens from 100 to 256 to prevent

JSON truncation that caused the JsonReaderException on memory extraction.

Version: 1.65.0 -> 1.65.1

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

Hardware profiling: GPU detection (CUDA/ROCm/Metal), CPU/SIMD capability

assessment (AVX2/AVX-512), and available memory detection. Composite

fitness scoring (0-100) with Green/Yellow/Red tiers determines local AI

viability and drives cloud provider recommendations.

New cloud providers: Mistral AI (EU-based, GDPR-native, HighPrivacy) and

Groq (fast inference, StandardApi). Both use OpenAiCompatibleProviderBase

which extracts shared OpenAI chat format logic.

Privacy tier system: PrivacyTier property on all providers — HighPrivacy

(Anthropic, Mistral, Local) vs StandardApi (OpenAI, Gemini, Groq).

Provider dropdown now shows "Privacy-First" labels.

Smart GPU offload: LocalLlamaProvider uses PlatformDetector.DetectGpu()

to set GpuLayerCount to -1 (all layers) when GPU is available, or 0

(CPU-only) when no accelerator is detected — prevents crashes and

unusable performance on CPU-only systems.

Settings UI: Hardware Report Card replaces single RAM recommendation line

showing RAM/GPU/CPU details, fitness score badge, and cloud-recommended

banner for Yellow/Red tier systems.

Desktop version bumped from 1.64.3 to 1.65.0.

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

Introduces PrivacyTier enum (HighPrivacy, StandardApi) to classify AI

providers by their data handling guarantees. Adds nullable PrivacyTier

property to AiProviderInfo record. Bumps SDK version from 1.64.0 to

1.65.0 to publish updated types before dependent desktop changes.

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

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.