Runora v2.1.4 - Smart GGUF Chat Routing & Engine Fixes

Runora v2.1.4 - Smart GGUF Chat Routing & Engine Fixes

Added:

• Running GGUF Auto-Detection: Chat Playground connects automatically to active running GGUF model servers.
• No Server Guidance Card: Clear prompt directing users to the Launch tab ([2] or [Enter]) when no GGUF server is running.
• Multi-Instance Model Selector: Interactive card to pick which active GGUF instance to chat with when multiple servers are running.
• Modifier-Based Hotkeys: Switched chat hotkeys to dedicated modifier shortcuts ([Ctrl+S] Model Switcher, [Ctrl+K] Compact, [Ctrl+P] Parameters, [Ctrl+Y] Copy) eliminating typing interference.
• Slash Commands: Type commands directly in chat: /compact, /params, /model, /clear, /help.

Fixed:

• Flash Attention Argument: Fixed server launch failure caused by bare --flash-attn flag by providing explicit --flash-attn on expected by modern llama-server builds.
• Streamlined Model Inspector: Removed redundant manual task cycling ([E] Task). Models execute under their intrinsic engines based on format (.gguf with llama.cpp, .onnx with ONNX Runtime).

Runora v2.1.3 - Integrated Terminal Chat Playground

Runora v2.1.3 - Integrated Terminal Chat Playground & Updater Fixes

Added

• [7] Chat Playground: Direct conversational playground inside the terminal without external clients or curl.
• SSE Token Streaming: Real-time token streaming with live throughput telemetry (tokens/sec) and prompt/generation counting.
• 3-Layer Hybrid Context Compaction:
  • Layer 1: System prompt (never compacted).
  • Layer 2: Compaction Checkpoints summary stack.
  • Layer 3: Recent verbatim tail (20% context budget).
  • On-demand ([K]) and automatic (>85% pressure) history compaction with lossless original message preservation on disk.
• Multiple Named Sessions: Conversation histories saved atomically under the user data directory with create ([N]), delete ([D]), and rename ([R]) controls.
• Parameters Overlay ([P]): Dynamic tuning of Temperature, Top-P, Top-K, and Context Size.
• Inline Model Launcher: Seamless launcher displayed when entering Chat without an active server.
• Clipboard Extraction ([C]): Instantaneous copy of assistant answers to clipboard.
• Multiline Prompts: Full textarea editing with Enter to send and Ctrl+Enter for newlines.

Fixed

• Windows Self-Updater Asset Matching: Fixed a bug where darwin release packages matched the win substring token, causing macOS binaries to be downloaded on Windows.
• Platform Classifiers: Added strict OS discrimination (isWindowsAsset, isDarwinAsset, isLinuxAsset) ensuring exact OS and architecture matching.
• Target Executable Resolution: Enhanced executable path resolution to locate and upgrade the true binary in system PATH (%USERPROFILE%\go\bin\runora.exe).

Runora v2.1.2

What's Changed in v2.1.2

• Settings Target Scoping: Fixed isolated runtime inspector updates so checking/updating llama.cpp, ONNX, or App never interfere with one another.
• Two-Stage Enter Key Workflow: First Enter checks for updates, and second Enter immediately downloads and installs available updates.
• Normalized Version Checks: Stripped leading 'v' and 'b' prefixes during semantic comparison to prevent false positive update notifications.
• Cleaned App Inspector: Removed duplicate theme switcher from Settings App card.
• Full Mouse & Keyboard Navigation: Full suite support across all screens.

Runora v2.1.1: Automatic Library Migration and Direct Self-Updater

Runora v2.1.1 introduces automatic assistant library migration from Ollama and LM Studio models, alongside direct binary self-updating from GitHub Releases.

What is New in v2.1.1

• Automatic Assistant Library Migration:
  • Ollama Manifest-Aware Resolver: Automatically discovers models located in Ollama registries across Windows, macOS, and Linux, mapping tags and layers to raw GGUF blobs.
  • LM Studio Recursive Scanner: Discovers all .gguf and .onnx models within default and custom LM Studio directories.
  • Zero Disk Duplication: Indexes models by reference in the Runora catalog without copying multi-gigabyte weight files.
  • Onboarding & Settings Integration: Dedicated onboarding step (StepModelSources) and Settings Inspector ([4] Model Sources) with interactive toggle and rescan capabilities.
• Direct GitHub Release Self-Updater:
  • Standalone Binary Downloads: Checks �pi.github.com/repos/BIJJUDAMA/runora/releases/latest, downloads the matching OS/architecture package, and swaps binaries in-place.
  • No Go Toolchain Required: Users running precompiled binaries can update directly from GitHub Releases with live download percentage progress bars.

Runora v2.1.0: Full Mouse Navigation and Cross-Platform OS Keyring

Runora v2.1.0 introduces complete mouse and touchpad navigation across the entire Bubble Tea TUI, alongside secure native OS Keyring integration for API credentials.

What is New in v2.1.0

• Spatial Mouse Hit-Testing Subsystem: Full spatial event routing with Z-Index layers, click detection, rapid double-click support (<450ms), context-aware mouse wheel scrolling, and automatic click-away modal dismissal.
• Clickable Global Navigation: Click any top header tab ([1] Models, [2] Launch, [3] Monitor, [4] Downloads, [5] Benchmarks, [6] Settings) to jump between views.
• Interactive Screen Hitboxes:
  • Model Explorer: Click model rows to select, double-click to immediately open the Launch Dashboard, and scroll wheel to browse large libraries.
  • Launch Dashboard: Click profile cards in the 5x5 Bento grid to select, double-click to immediately launch inference, and click [C] to copy the generated CLI command.
  • Server Monitor: Click running instance rows to inspect telemetry, double-click to stream live logs, and click action buttons ([R], [S], [Ctrl+K], [L]).
  • Downloader and Settings: Click input fields to focus and type, and click component list items to inspect sections.
  • Theme Picker: Click themes to preview, double-click to apply and dismiss, or click outside the dialog to cancel.
• Native OS Keyring Credentials: Encrypted, cross-platform credential storage using Windows Credential Manager, macOS Keychain, and Linux SecretService/libsecret.
• Automated Credentials Migration: Automatic migration of API tokens from plaintext config.json to the OS keyring with automatic token sanitization.

v2.0.0: Bento Architecture, 5x5 Profiles, Flash Attention & Hardened Test Suite

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.0.0] - 2026-08-23

Added

• Persistent Global Navigation Header across all views with 1-6 numeric hotkeys (1: Models, 2: Dashboard, 3: Monitor, 4: Downloads, 5: Benchmark, 6: Settings) and bidirectional Tab / Shift+Tab and [ / ] cycling.
• Persistent live hardware header telemetry displaying running server instance count and dynamic GPU VRAM meter.
• Bento Card layout architecture across all screens with dynamic height clamping and vertical alignment.
• 10 curated accessible themes (Dracula, Sunset, Nord, Cyberpunk, Forest, Monochrome, Solarized Light, Paper Light, and High Contrast WCAG AAA) with linear RGB gradient text and gauge rendering.
• Interactive Theme Picker modal ([Y]) with live swatches and descriptions.
• Non-intrusive floating toast notification system with ANSI compositing.
• Real-time Log Streamer ([L]) with 250ms tailing, auto-scroll, regex filter ([/]), pause/resume ([Space]), syntax highlighting, and multi-instance tab switching ([Tab]).
• 5-per-row Bento execution profile grid supporting up to 25 profiles with dynamic downward vertical expansion and 2D keyboard navigation ([←/→/↑/↓]).
• Flash Attention enabled by default across all llama.cpp server invocations and profile presets (--flash-attn).
• Quantized KV Cache support with --cache-type-k and --cache-type-v flags (f16, q8_0, q4_0, fp8).
• Raw Custom CLI arguments field in profiles allowing arbitrary user-supplied arguments passed directly to llama-server.
• Interactive 8-field Profile Creator and Editor supporting Name, Context, Threads, GPU Layers, Port, Flash Attention toggle, KV Quantization cycler, and Custom CLI Arguments.
• Unrestricted profile deletion allowing removal of any custom or built-in default profile from disk.
• Runtime version slots under llama.cpp/versions/<tag>/ enabling side-by-side installations, instant version switching, listing, and cleanup.
• Release channel selector supporting Stable (vX.Y.Z releases) and Nightly (upstream continuous tags) with explicit backend selection (CUDA 12, CUDA 13, Vulkan, CPU, ROCm, Metal).
• Multi-GPU enumeration, total VRAM aggregation, and TensorSplitAdvisor GCD integer ratio calculation (e.g. 24GB + 16GB + 8GB -> 3,2,1).
• Physical CPU core vs logical thread topology detection across Windows, Linux, and macOS.
• Apple Silicon Metal piecewise unified memory curve (67% to 92%).
• Multi-part GGUF shard auto-grouping (model-00001-of-00004.gguf) into consolidated single model entries with aggregate sizes.
• Multi-directory model discovery scanning primary and secondary paths (Paths.ModelDirectories).
• Headless CLI flags: --list-models (with --json), --status (with --json), --data-dir <path>, --models <path>, --version, --reset-onboarding.
• Comprehensive behavioral test suite hardening with boundary condition and error recovery validation.

Changed

• Refactored model browser, launch dashboard, server monitor, downloader, benchmarks, and settings to full-screen Bento card layouts.
• Updated Settings screen left panel to clean component hierarchy: API Token, llama.cpp, ONNX Runtime, Runora App.
• Improved onboarding wizard with an 86-cell wide layout to eliminate awkward line wraps and support direct API credential configuration.
• Enforced strict zero emoji invariant across all views, headers, footers, badges, and notifications.

Fixed

• Fixed Windows file lock race condition in atomic file writing by adding exponential backoff retry loops during high-concurrency renames.
• Fixed invisible UTF-8 BOM headers in Go source files that prevented statement coverage profiling.
• Fixed single-owner cmd.Wait() race conditions in process supervisor during multi-threaded instance termination.
• Fixed download queue range resumption and .part file cleanup on cancellation or completion.

[1.1.1] - 2026-08-22

Added

• Modular Theme Class architecture with CSS-variable-style semantic tokens and centralized stylesheet (global.css pattern).
• Support for llama.cpp semantic versioning releases (vX.Y.Z) with automated nightly build tag resolution via nightly-tag.txt.
• In-app configuration and environment variable support for GitHub API token (G hotkey) to increase release check limits from 60 to 5,000 req/hour.
• Direct numeric hotkeys (1, 2, 3) to switch runtime focus in the Settings view.
• Explicit inline available actions indicator for each runtime option card.

Fixed

• Eliminated all hardcoded colors from UI components in favor of dynamic theme tokens.
• Resolved lifecycle message channel cross-contamination between runtime engines and application update checks.
• Resolved Windows CUDA asset matching for upstream continuous builds.
• Fixed column vertical separator alignment in the Preferences and Hardware info panel.
• Clean opening of Settings view without triggering automatic background downloads.
• Improved GitHub API rate limit error detection and diagnostic feedback on 403 Forbidden responses.

[1.1.0] - 2026-08-22

Added

• Multi-runtime abstraction architecture supporting both llama.cpp and ONNX Runtime engines.
• Unified Lifecycle and Settings screen with component focus navigation using Tab and Arrow keys.
• Universal action hotkeys for checking (C / Enter), updating (U / Space), and rolling back (R) runtime components.
• Automated backup creation and rollback restoration for ONNX Runtime library installations.
• Magic-byte archive format sniffing supporting ZIP, Nupkg, Tar.gz, and TGZ packages without relying on file extensions.
• Support for searching and downloading .onnx model files directly from Hugging Face repositories.
• Automatic configuration directory migration from legacy paths to the runora application directory.
• Dynamic port allocation and socket collision mitigation for concurrent multi-model server instances.
• Onboarding walkthrough tour and visual gradient theme switcher.

Changed

• Rebranded project, Go module path, CLI command, and configuration namespace from llama-manager/llmgr to Runora/runora.
• Streamlined settings footer keybindings into an intuitive unified layout.
• Updated Hugging Face search filters to query both GGUF and ONNX format repositories.

Fixed

• Fixed ONNX runtime installation error caused by unsupported .archive temporary file extensions.
• Resolved race conditions in download task queue scheduling during concurrent model operations.
• Resolved socket race condition during rapid sequential server deployments on identical ports.

[1.0.0] - 2026-08-20

Added

• Core TUI model browser with real-time fuzzy search and model metadata inspection.
• Hardware discovery engine detecting CPU, RAM, and GPU capabilities with VRAM estimation.
• Custom launch profiles supporting custom context lengths, GPU layer offloading, and thread counts.
• Integrated benchmark suite measuring token generation throughput and memory footprints.
• Server monitoring dashboard displaying instance uptime, resident memory RSS, and status.
• Hugging Face repository model downloader with pause, resume, and cancellation support.

v1.1.1 - Multi-Runtime Support & Dynamic Theme Architecture

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.1.1] - 2026-08-22

Added

• Modular Theme Class architecture with CSS-variable-style semantic tokens and centralized stylesheet (global.css pattern).
• Support for llama.cpp semantic versioning releases (vX.Y.Z) with automated nightly build tag resolution via nightly-tag.txt.
• In-app configuration and environment variable support for GitHub API token (G hotkey) to increase release check limits from 60 to 5,000 req/hour.
• Direct numeric hotkeys (1, 2, 3) to switch runtime focus in the Settings view.
• Explicit inline available actions indicator for each runtime option card.

Fixed

• Eliminated all hardcoded colors from UI components in favor of dynamic theme tokens.
• Resolved lifecycle message channel cross-contamination between runtime engines and application update checks.
• Resolved Windows CUDA asset matching for upstream continuous builds.
• Fixed column vertical separator alignment in the Preferences and Hardware info panel.
• Clean opening of Settings view without triggering automatic background downloads.
• Improved GitHub API rate limit error detection and diagnostic feedback on 403 Forbidden responses.

[1.1.0] - 2026-08-22

Added

• Multi-runtime abstraction architecture supporting both llama.cpp and ONNX Runtime engines.
• Unified Lifecycle and Settings screen with component focus navigation using Tab and Arrow keys.
• Universal action hotkeys for checking (C / Enter), updating (U / Space), and rolling back (R) runtime components.
• Automated backup creation and rollback restoration for ONNX Runtime library installations.
• Magic-byte archive format sniffing supporting ZIP, Nupkg, Tar.gz, and TGZ packages without relying on file extensions.
• Support for searching and downloading .onnx model files directly from Hugging Face repositories.
• Automatic configuration directory migration from legacy paths to the runora application directory.
• Dynamic port allocation and socket collision mitigation for concurrent multi-model server instances.
• Onboarding walkthrough tour and visual gradient theme switcher.

Changed

• Rebranded project, Go module path, CLI command, and configuration namespace from llama-manager/llmgr to Runora/runora.
• Streamlined settings footer keybindings into an intuitive unified layout.
• Updated Hugging Face search filters to query both GGUF and ONNX format repositories.

Fixed

• Fixed ONNX runtime installation error caused by unsupported .archive temporary file extensions.
• Resolved race conditions in download task queue scheduling during concurrent model operations.
• Resolved socket race condition during rapid sequential server deployments on identical ports.

[1.0.0] - 2026-08-20

Added

• Core TUI model browser with real-time fuzzy search and model metadata inspection.
• Hardware discovery engine detecting CPU, RAM, and GPU capabilities with VRAM estimation.
• Custom launch profiles supporting custom context lengths, GPU layer offloading, and thread counts.
• Integrated benchmark suite measuring token generation throughput and memory footprints.
• Server monitoring dashboard displaying instance uptime, resident memory RSS, and status.
• Hugging Face repository model downloader with pause, resume, and cancellation support.

v1.1.0 - Multi-Runtime Architecture & Unified Lifecycle

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.1.0] - 2026-08-22

Added

• Multi-runtime abstraction architecture supporting both llama.cpp and ONNX Runtime engines.
• Unified Lifecycle and Settings screen with component focus navigation using Tab and Arrow keys.
• Universal action hotkeys for checking (C / Enter), updating (U / Space), and rolling back (R) runtime components.
• Automated backup creation and rollback restoration for ONNX Runtime library installations.
• Magic-byte archive format sniffing supporting ZIP, Nupkg, Tar.gz, and TGZ packages without relying on file extensions.
• Support for searching and downloading .onnx model files directly from Hugging Face repositories.
• Automatic configuration directory migration from legacy paths to the runora application directory.
• Dynamic port allocation and socket collision mitigation for concurrent multi-model server instances.
• Onboarding walkthrough tour and visual gradient theme switcher.

Changed

• Rebranded project, Go module path, CLI command, and configuration namespace from llama-manager/llmgr to Runora/runora.
• Streamlined settings footer keybindings into an intuitive unified layout.
• Updated Hugging Face search filters to query both GGUF and ONNX format repositories.

Fixed

• Fixed ONNX runtime installation error caused by unsupported .archive temporary file extensions.
• Resolved race conditions in download task queue scheduling during concurrent model operations.
• Resolved socket race condition during rapid sequential server deployments on identical ports.

[1.0.0] - 2026-08-20

Added

• Core TUI model browser with real-time fuzzy search and model metadata inspection.
• Hardware discovery engine detecting CPU, RAM, and GPU capabilities with VRAM estimation.
• Custom launch profiles supporting custom context lengths, GPU layer offloading, and thread counts.
• Integrated benchmark suite measuring token generation throughput and memory footprints.
• Server monitoring dashboard displaying instance uptime, resident memory RSS, and status.
• Hugging Face repository model downloader with pause, resume, and cancellation support.

v1.0.0

Llama Manager is a Windows terminal user interface for managing, launching, monitoring, benchmarking, and updating llama.cpp models from a single interface.

Features

Model Management

• Automatically discovers GGUF models in the local models directory.
• Extracts and caches metadata such as architecture, parameter count, quantization, and context length.
• Supports custom tags and notes for model organization.

Memory Estimation

• Detects local hardware, including CPU, RAM, GPU, VRAM, and CUDA support.
• Estimates model and KV cache memory usage before launch.
• Provides color-coded fit indicators for selected configurations.

Launch Profiles

• Includes preset and custom launch configurations.
• Displays the exact llama.cpp command before execution.
• Starts and manages llama.cpp server instances directly from the TUI.

Downloads

• Download models from URLs or Hugging Face repositories.
• Supports selecting specific GGUF files from repositories with multiple weights.
• Download queue with pause, resume, and cancel controls.

Monitoring & Updates

• Real-time monitoring of running servers, including throughput, uptime, and resource usage.
• Built-in llama.cpp update, backup, and rollback management.

Customization

• Multiple built-in themes with customizable terminal appearance.