v1.7.0

Features

• Multi-fork support — install, manage, and run multiple llama.cpp forks alongside upstream:
  • koboldcpp — with variant-based backends (CUDA, CPU, old GPU, Metal), custom CLI flag mapping, KoboldAI-specific preset fields
  • beellama.cpp — full upstream preset compatibility, extended backends (SYCL, TurboQuant)
  • ik_llama.cpp — manual install support (no prebuilt releases)
  • llamacpp-rocm — AMD ROCm fork with 7 GPU target variants (gfx120X, gfx1151, gfx1150, gfx110X, gfx103X, gfx90a, gfx908)
• Fork registry (src/lib/forks.ts) — data-driven per-fork metadata: binary names, asset naming, backend variants, preset compatibility, CLI field mappings
• Fork selector in VersionsTab — cycle forks when browsing releases, selection persists via defaultFork config option
• Fork-aware presets — SettingsPanel filters preset categories and fields per fork (e.g., hides incompatible flags for koboldcpp, shows KoboldAI-specific fields)
• Fork labels — displayed in local versions list and DashboardTab version display
• Field mapping system — per-fork CLI flag remapping with support for value transforms, negate inversion, and fork-specific fields

v1.6.0

Build

• Switched from tsc to tsup bundler for single-file ESM output (no .js extension issues)
• Publish to npm

Features

• Added update check on startup, every 6h, and via Ctrl+U
• Added Copy button to update modal with npm update command
• Sort profile list ascending and version list descending

Fixes

• Clean stale folder on retry instead of throwing "already installed"
• Critical UI framework issues (layout, rendering, control lifecycle)

v1.5.3

Improvements:

• Dashboard profile/version selectors - Ability to pick profile and version directly from the dashboard view (PR #41)
• Device memory usage tracking - Added device memory usage tracking to dashboard and model info (PR #40)

Fixes:

• Derive free memory from total - self instead of trusting Vulkan pool accounting
• Sum KV cache sizes for MTP models
• Only capture first KV cache line (skip MTP draft context)
• Update device memory in currentModel for post-load breakdown lines
• Hide unused devices from loaded model memory table
• Handle negative unaccounted in memory breakdown regex

Refactor:

• Reorganize modal files into ui/specialized/

v1.5.0

Breaking

• Major codebase restructuring:
  • src/components/ui/ moved to src/framework/ — the UI framework is now a self-contained module
  • src/components/ moved to src/ui/ — tabs, specialized components, and MainControl
  • getConfig() removed from RenderContext — now only available on TabContext

Added

• Application base class — reusable app lifecycle manager with render loop, input/mouse handling, resize, cursor visibility, and modal rendering
• LlamaManagerApp — app-specific class extending Application with theme handling, config management, and key routing
• Comprehensive framework documentation — updated src/framework/README.md with full API reference for all classes, widgets, and types

Changed

• Event system upgraded to multi-listener — Control.on() now supports multiple handlers per event and returns the bound handler for removal
• Control.render() split into render() (base, handles bg clear, clipping, children) and draw() (subclass override for custom content)
• ModalManager render context no longer includes getConfig() — matches updated RenderContext interface
• All import paths updated across the codebase to reflect new directory structure

v1.4.6

Improvements

• Stopping server modal — visual modal shown while the server is stopping (c35a840)
• Spinner animation on loading model message (66907d6)

Fixes

• TextInput cursor not advancing after character input — cursor position was reset on every value setter call (ca49422)
• Hide search box when browsing model files (85665e8)
• Back button added to VersionsControl for navigation (30c2eb4)
• OptionsControl now correctly extends Control instead of Column (0f06505)
• Removed shebang from ServerTab, fixed LogsTab handleKey, removed empty draw (ef784e7)

Refactors

• Removed all custom renderers — List/Table now use default rendering (8737927)
• Deduplicated formatSize into utils.ts, removed inline fmtNum (91a2767)
• Standardized focusable, button rows, and ESC back navigation across ServerTab, ModelsTab, TasksTab, and VersionsTab (702208a)
• Back button moved to far-left in ServerTab and VersionsTab (423adde)

v1.4.5

Improvements

• Added device selector modal for GPU/CPU device selection
• Open profile editor on Enter for highlighted profile, show profile name in section title
• Show Back button in editor, hide CRUD buttons while editing
• Independent scroll offset and mouse wheel scrolling for profile list
• Promoted 13 commonly used settings from advanced to default visibility
• Added Delete key to restore default value in SettingsPanel

Fixes

• Cursor flicker — only emit show/hide escape sequences on actual visibility change
• TextInput viewport scroll when value is set before layout
• Preserve ProfileList selection when returning from editor
• Allow mouse wheel to scroll selection out of viewport in EditableList

v1.4.4

Improvements

• Empty state messages — Show empty state messages inside sections and move Search button after input
• Search box layout — Move search box to button row and remove TextInput borders

Fixes

• Accurate GGUF search — Filter non-LLM models and use filter=gguf for accurate results
• HF API fixes — Update fetch URL in listFiles and correct filter parameter in browseModels
• Search query parameter — Use library_name=gguf instead of appending to search query

Refactoring

• Dead code cleanup — Remove api.ts, unused widgets (Group, HelpBar, Box, AlertDialog), and unused imports/exports

v1.4.3

Features

• Precise progress bars — context and progress bars now use partial-fill blocks for accurate visualization
• Checkpoint density bar — new checkpoint bar aligned with context bar, showing checkpoint count vs context size
• Prompt progress percentage — shows PP (prompt processing) speed with a percentage indicator
• Animated download spinner — download modal status line now shows a spinning indicator with percentage during model downloads

Fixes

• Prompt progress fraction — corrected 0-1 fraction to display properly (was treated as percentage)
• Checkpoint bar alignment — uses nCtxSlot to align with context bar
• Checkpoint stale state — slot.checkpoints now cleared on slot release, matching server behavior
• ProgressBar width — uses full available width instead of reserving unused space

Refactoring

• ProgressBar simplified — reduced to single-row bar-only widget, removed unused percentage label
• Checkpoint display simplified — removed visual bar, now shows numeric count and total size only
• Dead code removed — deleted unused ProgressDialog widget, removed dead ProgressBar from VersionsTab
• Single-state context bar — simplified to show only prompt tokens
• Removed unused ProgressBar from ModelsTab

v1.4.2

Features

• Timestamped log files — each server session writes to logs/server.<timestamp>.log instead of a single server.log. Old logs are preserved across restarts instead of being truncated.

Improvements

• MetricsPanel — streamlined rendering, fixed label widths (5 chars), adjusted context bar to 33 chars

v1.4.1

Fixed

• Metrics panel context bar restored — llama.cpp changed log function names from update_slots to operator(), causing regex-based parsing to fail. Context bar, cached tokens, and checkpoint erasures now parse correctly again.

Added

• Scrollable MetricsPanel — metrics panel now inherits from Scrollable, with a shared scrollbar for long slot lists.
• Scrollable ChangelogView — changelog panel now scrolls with a scrollbar.
• Responsive breakpoint system — new width and height breakpoints for adaptive layouts.
• Spinner animation — active slots show an animated spinner; spinner logic centralized in utils.ts.
• Model loading indicator — visual feedback while the model is loading on startup.