Darkian Studio changelog

Voice-to-text input, user skills, vision, and full internationalization Pre-release 1.0.0-beta.6 · 2026-08-18

Added

Voice-to-text input: speak a prompt; it is transcribed in-app via the Web Speech API and streamed into an inverted top dropdown, then formatted to markdown by the small "big-pickle" model before being sent to chat.
User skills: skills from ~/.agents/skills are now loaded via dsterm's silentRun (no SAF grant required).
Vision: image-capable models can now see images read from files in agent chat.
MCP management screen with an add-dialog and optional description; MCP servers are now wired as agent tools and forward configured environment variables.
GitHub agent tools: issues, pull requests, workflows, and commits.
Full internationalization: hardcoded UI strings wired to AppLocalizations across the marketplace, hamburger menus, and settings; all 16 non-English locales completed (370/370 keys).
Adaptive Material 3 responsive layout system; persistent panels embedded via DsPaneHost docking and DsAdaptiveSheet.
Editor: toggle breakpoints at the cursor (Ctrl+Shift+B); in-document word completions with color icons by kind; CM6 as the single completion UI (Flutter overlay dropped); LSP completion results echoed back via a requestId bridge channel.
Theme: per-theme accent derivation for CM6 bundle themes.
Marketplace: Hugging Face logo on model cards.
Clone: dsterm pre-check and simplified clone input (owner/repo + ssh/https).
Compact (phone) UI: smaller editor/terminal/chat font defaults.
Nav bar and command palette icons migrated to codicons.

Changed

Marketplace install taxonomy clarified; on-disk extensions are the source of truth.
Extension-host message handling replaced with a dispatch table; node/netcat discovery memoized and kill/error helpers de-duplicated.
Extensions nav label renamed to Marketplace.
Welcome screen action block and recent projects centered on wide screens.
Token estimator aligned with opencode; heavy directories excluded from ingestion.
Welcome screen palette kept in sync with darkian/light editor themes.

Fixed

Chat: ref.listen moved into build (startup assertion crash fixed); voice-mode codicon used for the voice button.
LSP: tolerate a null range in content changes (full-document replacement) — startup crash fixed.
Skills indexed by frontmatter/directory name instead of file basename.
UI: left inset no longer reserved behind the nav rail; landscape status strip and nav-gap layout fixed.
Editor: language completions queried via languageDataAt; theme completion popup and color icons by kind; finer CM6 highlighting for installed themes; installed themes mapped to CM6 by most-general scope.
Extensions: reconciled snapshot propagated to installed/catalogue UI; install/uninstall errors logged with stack + step markers.
Theme: canonical syntax color extraction corrected; GitHub dark highlight no longer layered over light themes; gilded accent restored on reopen.
AI: per-id tool-call indices in the local chat transport; tool calls folded into the target in place.
Termux: OpenSSL added so installing curl on a fresh Termux won't break the runtime.
Editor: Exit Workspace moved to the bottom of the workspace submenu behind a divider.
Updates: offer equal-core beta releases instead of reporting "up to date".

Removed

Theme translation probe diagnostic screen (added then removed during beta 6).

View on GitHub Releases →

On-device AI, GitHub cloud workspaces, and the agent chat rework Pre-release 1.0.0-beta.5 · 2026-08-14

Added

Local on-device GGUF inference run entirely through dsterm — load models offline, no API key or network needed for inference.
A full on-device inference stack: protocol client, model resolution, fill-in-the-middle (FIM) with Rust-side template resolution, and tool calling for local GGUF models.
Priority scheduler for local models — per-model queues, concurrency, reference counting, sequence allocation.
Inference performance pass — KV cache reuse, warm-up, batching, streaming, memory live-wiring, and context reuse.
Local-on-device image/context retrieval pipeline (DSPack) with ranking, context injection, and budgeting.
GitHub cloud workspaces — sign in with GitHub OAuth, browse and clone repos, and work on remote files through a GitHub filesystem backend with local caching.
Git panel for remote workspaces via the GitHub Git Data API plus a minimal GitHub Actions integration with per-step job logs.
Remote-safe guards so local-only actions are disabled or clearly identified inside a remote workspace.
Agent chat rework — streaming messages, tool cards, agent state, syntax-highlighted diffs, and OpenCode-style markdown rendering.
Expanded AI tools: workspace search, grep, tree, stat, edit with fuzzy-replacer chain and diff output, web tools (search/fetch/extract), workspace.run, todowrite, askQuestion, and an agent.spawn sub-agent tool.
Token tracking with compaction/pruning, cache policy, and transport retry.
Full internationalization (i18n) infrastructure with 18 languages and a locale picker.
Model marketplace (Hugging Face): search, sort, pagination, README rendering, parameter counts, installed filter, and install/load/offload of GGUF models with real file sizes and a 30-minute download timeout.
Extension marketplace (OpenVSX): browse and open extension details; DSPacks coming-soon placeholder.
PostHog analytics with a safe enabled-guard for CI/test builds.
A remote feature flag (builtin_models_enabled) that can hide the built-in provider server-side without an app update.
In-page eruda developer-tools console toggle in the browser chrome.
Vector GitMark icon set across notes, menus, and the command palette; codicon icon set with tinted marks.
Branded intro animation on the welcome screen before first use.

Changed

Chat syntax highlighting and user bubbles now inherit design tokens and the active editor theme instead of hardcoded colors.
Bottom sheets, dialogs, the command palette, file switcher, and model picker all share a single soft blurred, dimmed backdrop.
The built-in free provider is split out under its own 'Builtin' label; per-model system prompts and a positive-form system prompt with tiered tool retrieval for small models.
AI settings screen reworked; scrollable provider picker; inline model labeling; gated local inference.
Programming-related screens (DAP, host suite, tasks, problems, settings search, GitHub repos) migrated to the DS surface primitives and design tokens.
Darkian Dark is the default first-launch palette; Darkian Light was dropped.
Slash completions load a single bundled skills index instead of many fragile per-directory asset declarations.

Fixed

Remote workspace search matches file names by substring instead of missing them.
Workspace.run returns the real exit code and stderr, applies timeouts, and wraps JSON output without a FormatException.
Tool outputs strip the JSON envelope so users see meaningful content instead of raw payloads.
Reasoning and tool-call rows stay structured per-turn instead of merging into one blob; copy button and ActionsGroup are per-turn.
Model README renders as markdown; real file sizes shown (HF tree endpoint has sizes); nested memory_estimate parsed correctly.
The webview backdrop sizing fixed (moved to didChangeDependencies) and the eruda floating button dropped in favor of the chrome toggle.
The built-in provider feature flag is fail-open when PostHog is unconfigured, so the provider never vanishes in dev/test.

Removed

GitHub Copilot provider, device-code sign-in flow, and probe screens — removed; proved not worth the complexity for the API used.
The unused save button from the editor app-bar chrome (save remains available from the special-keys bar).
Darkian Light theme.

Known limitations

On-device model quality and speed depend on device hardware; very large GGUF models may not fit in memory on low-end devices.
GitHub cloud workspaces require an active network connection and GitHub OAuth; offline work is limited to the local-file cache.
GitHub Actions on remote workspaces covers core run/list/log flows; advanced workflows are not yet surfaced.

View on GitHub Releases →

First public beta Pre-release 1.0.0-beta.1 · Unreleased

Added

First public beta of Darkian Studio (Android APK via GitHub Releases).
Onboarding runtime setup with a single copy-paste command.
install.sh setup script, runtime-aware for Termux, Linux, and macOS, with progress/error output and a post-install verification step.
In-app Check for updates (Settings → About) plus a best-effort startup update check against GitHub Releases.
In-app runtime verification during onboarding (node / dsterm / code-server).
Diagnostics logs and crash reports surfaced entirely inside the app.
Branded intro animation on first launch before the welcome screen.
AI chat agent with tool calling — read/write files, workspace search, and command execution in the runtime — plus streaming responses, thinking-mode toggle, and session persistence.
AI providers: OpenAI-compatible endpoints, a built-in free provider (no key required), OpenRouter, Together, Fireworks, Groq, Mistral, Ollama, and Local GGUF.
Local GGUF model marketplace (Hugging Face): browse, install, load, and unload models with parameter counts, quantisation, and memory estimates.
AI inline completions (fill-in-the-middle) via chat-style and legacy completion endpoints.
Model Context Protocol (MCP) client for AI tool and context integration.
Per-model system prompt controls for local chat models.

Changed

All debugPrint() diagnostics in the app are now routed through the DS logging system so they appear in Diagnostics logs and crash reports.
Darkian Dark is the default first-launch theme palette.

Known limitations

dsterm remote on Windows is not supported in this beta.
No Play Store distribution; APKs are GitHub Releases only.

View on GitHub Releases →

Looking for the download? Grab the latest beta.