v0.7.0 — Swiss AI Privacy

Swiss AI Privacy

Mingly v0.7.0 introduces Swiss AI Privacy — automatic on-device PII protection.

Highlights

• 3-Layer PII Detection: Regex + Swiss Patterns + piiranha-v1 NER (400M ONNX)
• 17 PII Categories: Names, addresses, AHV numbers, IBAN, health data, and more
• 4 Privacy Modes: Shield, Vault, Transparent, Local Only
• On-Device NER: <50ms inference, no cloud required
• Red-Team Tested: 143 test cases, 84.9% detection rate, 0 critical failures
• Text Preprocessor: HTML entity decoding, fullwidth normalization, zero-width stripping

Downloads

• macOS: DMG (Apple Silicon) or ZIP
• Windows: Coming soon
• Linux: Coming soon

Full changelog: CHANGELOG.md

v0.5.1 — Bugfix: App-Start Crash behoben

Fixed

• App-Start Crash behoben — nanoid v5 (ESM-only) verursachte ERR_REQUIRE_ESM im Electron Main Process. Downgrade auf nanoid v3.x (CJS-kompatibel) behebt den Fehler.

Downloads

• Mingly-0.5.1-arm64.dmg — macOS Installer (Apple Silicon)
• Mingly-0.5.1-arm64-mac.zip — macOS ZIP (Apple Silicon)

v0.5.0 — Context Engineering + Multi-Ollama Load Balancing

Context Engineering (Manus-inspired)

• Task Progress Recitation — Injiziert kompakte Fortschritts-Summaries nach jedem Agent-Schritt. Verhindert Lost-in-the-Middle bei langen Runs.
• Error Preservation — Volle Fehlerinformationen bleiben im Kontext, damit das LLM aus Fehlern lernt.
• KV-Cache Optimization — Deterministische Tool-Sortierung + stabile Prompt-Struktur fuer bis zu 90% Kostenreduktion.
• File-based Memory — Grosse Tool-Ergebnisse werden in Temp-Dateien externalisiert, Kontext bleibt schlank.

Multi-Ollama Load Balancing

• OllamaLoadBalancer — Verteilt Anfragen ueber mehrere Ollama-Instanzen im lokalen Netzwerk.
• Least-Loaded + Round-Robin — Bevorzugt Backends mit weniger aktiven Requests und niedrigerer Latenz.
• Health Checks — Alle 30s, automatisches Failover nach 3 Fehlern.
• Transparente Integration — OllamaClient nutzt automatisch den Balancer wenn 2+ Backends verfuegbar.

Security Fixes (5)

• URL-Injection-Schutz im Load Balancer
• Path-Traversal-Schutz im AgentContextManager
• Stack Traces nur lokal geloggt (nicht an Cloud-LLMs)
• Timer-Leak-Fix im ToolRegistry
• any-Casts durch typisierte Interfaces ersetzt

Stats

• 1005 Tests in 58 Dateien — alle gruen
• TypeScript strict mode clean
• Build: main + preload + renderer erfolgreich

v0.4.0 — Parallel Intelligence

🚀 Mingly v0.4.0 — Parallel Intelligence

Phase 1: Agentic Mode

• ReAct Agent Loop: Full ReAct (Reason → Act → Observe) loop with automatic tool selection
• Tool-Use: chatWithTools() for Anthropic + OpenAI providers with multi-step tool chains
• ToolRegistry: Centralized tool management (built-in + MCP tools)
• AgentExecutor: Configurable agent runner with step limits, token caps, timeouts
• AgentStepIndicator: Live UI showing thinking → tool calls → results per step
• Feature Gate: Pro+ tier gating for agentic features

Phase 2: Parallel Intelligence

A) Ollama + Generic OpenAI Tool-Use

• Shared openai-tool-use-helper.ts enables function calling for local models via /v1/chat/completions
• Any model with function calling support (llama3.1, qwen2.5, mistral) gets full tool-use
• Works with LM Studio, LocalAI, OpenRouter and other OpenAI-compatible backends

B) Agent Comparison

• Run up to 3 models in parallel, each with its own ReAct loop and tool access
• Live step indicators per slot showing thinking, tool calls, and results
• Provider-colored badges, token/cost/duration tracking per slot
• Ollama multi-slot GPU warning

C) Parallel Subagents

• Master LLM decomposes task into 1-3 independent subtasks via decompose_task tool
• User configures provider/model per subtask (local vs cloud, mix-and-match)
• Parallel execution with full tool access per subtask
• Master LLM synthesizes all results into coherent final answer
• 4-phase UI: Decompose → Configure → Execute → Synthesize

Security Hardening

• IPC input validation on all handler parameters
• AbortControllers properly connected for real cancellation
• Session concurrency limits (max 2 parallel)
• Subtask description length limits (prompt injection mitigation)
• Event listener cleanup preventing memory leaks
• Synthesis error transparency (no silent masking)

Testing

• 956 tests passing across 56 test files
• TypeScript strict mode clean (both configs)
• New test suites: ollama-tool-use, agent-comparison-service, subagent-orchestrator

v0.3.2 — Fix Settings Page Crash

Fixed

• App crash on Settings page — updater-store.ts accessed window.electronAPI.updater without null-guard, causing Cannot read properties of undefined (reading 'updater') crash. All updater store methods now use defensive getUpdaterAPI() accessor.

Full Changelog: v0.3.1...v0.3.2

v0.3.1 — Settings & API Key Persistence Fixes

Fixed

• Settings page infinite loading — Settings page now shows with default values if backend settings load fails, preventing the permanent spinner
• API key persistence across updates — Corrupted encrypted entries (caused by adhoc re-signing) are now auto-cleaned, allowing clean re-entry instead of silent failures
• CSP blocking inline styles — Added 'unsafe-inline' to style-src policy, fixing React inline styles and Radix UI positioning that were silently blocked
• Missing default settings — settings-store.ts now merges loaded settings with defaults, ensuring all UI fields have values even when the stored config is incomplete

Upgrade Notes

If you had API keys configured in v0.3.0, they may need to be re-entered once after updating to v0.3.1 (due to adhoc signature change). After that, keys persist normally.

Full Changelog: v0.3.0...v0.3.1

v0.3.0 — Offline License Validation

What's Changed

Breaking Changes

• Supabase completely removed — No more external service dependency
• License validation mode changed from online/offline to signed/legacy

New

• HMAC-SHA256 license keys — Offline validation, no server required
• Legacy key migration — Old unsigned keys accepted with 90-day grace period
• Platform build scripts — dist:mac, dist:win, dist:linux, dist:publish
• Documentation — Added docs/en/LICENSING.md and docs/en/BUILDING.md

Removed

• Supabase Edge Function license validation
• Supabase Storage download URLs
• Machine fingerprint for license binding

Technical

• Key format: MINGLY-{TIER}-{PAYLOAD}-{HMAC_8_CHARS}
• 880 tests passing (50 test files)
• Downloads now exclusively via GitHub Releases
• Auto-updates via electron-updater (GitHub provider)

Full Changelog: v0.2.1...v0.3.0

v0.2.1 — Bugfix: Black screen, cross-platform, Explore Features

What is Fixed

Critical

• Black screen on dialog open — Opening Settings, starting a conversation, or clicking any dialog resulted in a black/empty screen. Root cause: double h-screen CSS classes causing nested 100vh overflow, plus Suspense fallback={null} silently swallowing lazy-load states.
• Settings page blank screen — SettingsPage returned null while settings were loading, showing nothing instead of a spinner.
• Dark mode flash on app start — Brief white flash before dark theme applied. Now applied synchronously.

UI

• "Explore Features" setup step — Step 3 on the welcome screen now has an "Open Settings" button (previously had no action).
• Silent dialog load failures — All lazy-loaded dialogs now show a spinner while loading instead of rendering nothing.

Cross-Platform (Windows / Linux)

• Window title bar — titleBarStyle: hiddenInset and trafficLightPosition are now macOS-only. Windows and Linux use the default system title bar.
• Sidebar padding — Traffic-light offset padding (pt-10) only applied on macOS; Windows/Linux get standard padding.
• Platform info exposed — Preload bridge now provides platform (darwin | win32 | linux) to the renderer for platform-aware UI.

Tests

• SimpleStore mock pattern — All 4 test files updated from constructor pattern to SimpleStore.create() singleton factory.
• Test isolation — Added _resetForTesting() to SimpleStore for clean test runs.

Full Changelog

See CHANGELOG.md for details.

---

All 878 tests pass | CI green on macOS, Ubuntu, Windows

Mingly v0.2.0 — Service Discovery, Model Comparison & More

Whats New in v0.2.0

New Features

• Model Comparison — Send the same prompt to 2-3 LLMs in parallel, view responses side-by-side, mark a winner
• Service Discovery — Auto-discover RAG and MCP servers across local, network, and cloud
• Updater UI — Check for Updates in Settings with progress bar and tier-aware install
• License Activation — Subscription management in Settings + tier badge in sidebar
• Routing Mode Toggle — Switch between manual model selection and Gemma auto-routing
• Active Model Indicator — See current provider/model in the chat header
• Local LLM Discovery — Auto-detects Ollama, LM Studio, vLLM, LocalAI, Text Gen WebUI, llama.cpp
• Custom RAG Server Name — Customize the display name of your RAG server

Bug Fixes

• Setup wizard re-appearing on restart — Fixed stale-cache issue with SimpleStore singleton pattern
• API key warning on first open — No longer shows No API key before keys are loaded
• macOS traffic lights overlap — Sidebar header pushed below close/minimize/maximize buttons
• Settings not persisting — UPDATE_SETTINGS now returns merged settings correctly

Downloads

• macOS (Apple Silicon): Mingly-0.2.0-arm64.dmg

Upgrading

Existing v0.1.1 installations will detect this update automatically (Pro+ tier) or can download manually.

Full changelog: https://github.com/Baldri/mingly/blob/main/CHANGELOG.md

v0.1.1 — Cross-Platform Release

Mingly v0.1.1

First cross-platform release with builds for macOS, Windows, and Linux.

Downloads

• macOS: DMG installer (Apple Silicon)
• Windows: NSIS installer or portable EXE
• Linux: AppImage or .deb package

Changes since v0.1.0

• Added Windows and Linux builds
• Fixed TypeScript strict mode warnings
• Improved CI pipeline reliability

Note: Builds are unsigned. macOS users may need to right-click > Open on first launch.