v2.0.2

Release Notes

• Vision runtime is included — no separate download needed.
• Tesseract OCR is auto-installed or uses bundled tesseract.js fallback.
• Use Download Base Models after install to enable AI features.
• See the Install Guide for Windows & Mac — one download, no CLI.

[2.0.2] - 2026-02-15

Changed

• Release pipeline run: Prepared a patch release to trigger fresh Windows and macOS release
  workflows with the current packaging pipeline and artifact conventions.

Release Checklist

See docs/RELEASING.md for versioning, runtime staging, checksums, and release
notes.

v2.0.1

Release Notes

• Vision runtime is included — no separate download needed.
• Tesseract OCR is auto-installed or uses bundled tesseract.js fallback.
• Use Download Base Models after install to enable AI features.
• See the Install Guide for Windows & Mac — one download, no CLI.

[2.0.1] - 2026-02-13

Fixed

• macOS startup stability: Prevented app launch failure when lz4-napi native bindings are
  unavailable in packaged environments by adding a shared codec wrapper with graceful fallback.
• Persistence startup path hardening: Updated vector DB, analysis history, and embedding queue
  persistence modules to use safe LZ4 codec loading instead of hard-failing at module import time.

Full Changelog: v2.0.0...v2.0.1

v2.0.0

Release Notes

• Vision runtime is included — no separate download needed.
• Tesseract OCR is auto-installed or uses bundled tesseract.js fallback.
• Use Download Base Models after install to enable AI features.
• See the Install Guide for Windows & Mac — one download, no CLI.

[2.0.0] - 2026-02-09

Changed (BREAKING)

• In-Process AI Engine: Replaced external Ollama server with node-llama-cpp for fully
  in-process LLM inference. All text, vision, and embedding models now run on-device via GGUF files
  with automatic GPU acceleration (Metal, CUDA, Vulkan) and CPU fallback. No external services
  required.
• In-Process Vector Database: Replaced external ChromaDB server with Orama for in-process
  vector storage and hybrid search. BM25 full-text and vector similarity are combined via Reciprocal
  Rank Fusion (RRF). Indexes persist atomically with LZ4 compression.
• Electron 40+: Upgraded from Electron 35 to Electron 40 with modern security defaults.
• Zero External Dependencies: The application is now fully self-contained. No Ollama, ChromaDB,
  or any other external server is needed. All processing happens on-device with zero data leaving
  the machine.

Removed

• Ollama Integration: All Ollama client code, connection management, and configuration removed.
• ChromaDB Integration: All ChromaDB client code, server health checks, and configuration
  removed.
• Deprecated Settings: Legacy ollamaUrl, chromadbUrl, and related connection settings are
  ignored on upgrade (cleaned up by SettingsMigrator).

Added

• LlamaService: Centralized in-process LLM inference service with model switching, memory
  management, and GPU/CPU backend selection.
• LlamaResilience: Circuit breaker, retry, and timeout guard for all AI inference calls.
• ModelAccessCoordinator: Coordinates concurrent model access to prevent conflicts during model
  switching.
• Content Selector: Intelligent document content selection for LLM analysis with token budget
  awareness.
• Crash Reporting: Sentry integration (opt-in via SENTRY_DSN) with local
  electron-crashReporter fallback.
• Test Coverage Enforcement: 50% global thresholds (branches/functions/lines/statements)
  enforced in CI.

Improved

• Code Quality: Consolidated 11 categories of duplicate, conflicting, and dead code:
  • Unified temp-file patterns into single canonical set in performanceConstants.js
  • Unified file-ID generators (getSemanticFileId delegates to getCanonicalFileId)
  • Disambiguated path normalization variants (sanitizePathValue vs resolveFsPath)
  • Replaced 15+ inline setTimeout delays with shared delay() utility
  • Centralized error classifiers (lifecycle delegates to errorClassifier.js)
  • Unified fallback analysis factories (analysisUtils delegates to fallbackUtils)
  • Replaced hardcoded RRF constant with SEARCH.RRF_K from performanceConstants
  • Replaced hardcoded retry counts with RETRY.* constants across 6 files
  • Consolidated duplicate Semaphore classes (single implementation in RateLimiter.js)
  • Removed dead no-op functions and commented-out code
  • Replaced ad-hoc pLimit with Semaphore-based concurrency limiter
• Production Builds: Console logs stripped from production bundles via Terser drop_console.
• Security Documentation: Updated known production gaps in CLAUDE.md to reflect resolved items
  (crash reporting, test coverage enforcement).