Skales Roadmap

A due schedule starts within a minute, “Run now” is instant, and every run’s result lands in Notifications and the schedule’s own execution log. One scheduler, one truth.

Task results, approvals, the stand-up, schedule outcomes: all follow your chosen channel instead of defaulting to Telegram. Quiet hours only mute messages while Friend Mode is on - they never pause work.

The WhatsApp assistant keeps a private log of what your allowed contacts wrote, so “yes, add her appointment” works from the app, WhatsApp or Telegram - without ever exposing your data to them.

Per-model profiles on by default, one-time detection of models without tool support, self-compacting long chats, and saner memory use for local models.

For a multi-step task Skales keeps a checklist pinned at the top of the chat that ticks off in real time, the way Claude Code tracks its to-dos. Works in any chat (not only Code mode), collapses out of the way, and tidies itself away when the plan is done.

Ask for parallel work ("run these in parallel", "leg ein multitask an") and Skales fans it out to background sub-agents that report back into the chat, with a live running badge. A multitask is no longer mistaken for a day-planner entry.

After changing code, Skales runs the project's own check (it detects npm / cargo / go / pytest), reads failures, fixes them, and re-runs until green; in Auto mode it does this on its own.

Inbound from your own number works again (modern linked-device ids resolve to the real phone). A permitted contact can message Skales; instead of answering them it pings you, and once you reply with what to do it acts with your tools and answers them for you. On your own thread your instruction is the approval. Owner commands on the road: /help, /pending.

Every proactive update - a finished task, a scheduled run's result, a reminder, a Friend Mode check-in - now lands on the Notifications page with a sound and an unread dot, and pops a toast (important ones stay until you dismiss them). It works even with the desktop buddy off and no Telegram connected. The scheduler is honest too: it logs "Queued" when a job fires and delivers the result to Notifications, and never claims its tools do not exist.

Pairing two desktops with a 6-digit code now shows the accept prompt inside the code window, so the connection is confirmed in place and the team actually forms.

A Chat / Code / Plan / Auto switch under the composer (or /code, or "/code <task>") binds a chat to a folder on disk. Plan is read-only and proposes; Code edits directly; Auto runs the whole task on its own after a one-time consent. Each chat keeps its own folder, so a normal chat is unchanged. If the folder is outside the file-access policy, Skales asks before widening it.

edit_file makes exact-string patch edits instead of rewriting whole files. git_status / git_diff / git_commit / git_push work in the bound folder with your own git identity (no added attribution), so commit and push run end to end. test_run detects and runs the project's test framework (npm/jest/vitest, pytest, cargo, go).

The old hard 30-second command limit killed installs, builds, git pushes and Firebase/Vercel deploys mid-run. Commands now run up to 2 minutes by default, configurable to 10 in Settings > Chat & Code, and keep 10MB of output, so real work completes and the error is visible if it fails.

In Auto mode a folder-bound chat pre-approves the in-folder file and shell tools after a one-time consent, so a long task does not pause on every step; outside tools (email, WhatsApp, browser) still ask and dangerous commands stay blocked. MiniMax (M2 / M2.7) added as a built-in LLM Profile.

When the agent needs you to choose between options or confirm a direction, it shows a small slide-up form (clickable single or multi-select options, one to a few steps) instead of a long question, then continues with your answers.

A dedicated Chat & Code settings tab: Deep reasoning (xhigh) asks any model to think step by step before acting, even models without a native reasoning mode; a Code model runs code work on a strong cloud model while chat stays on your default (pick a provider, fetch its models, choose one); and deploy_project deploys the bound folder (Firebase / Vercel / Netlify / npm) using your existing CLI login.

Skales Mobile is publicly available on the Google Play Store for Android phones and tablets. Pair your phone via QR for full access to Desktop's tool surface through an end-to-end encrypted relay, or run standalone with 27 native mobile tools. Install from https://play.google.com/store/apps/details?id=app.skales.mobile. iOS is in App Store review and lands the moment review clears.

The setting formerly known as Token Compressor is now called Memory Mode. Three modes: Always Remember, Compact, Minimal. Minimal moved behind an Advanced disclosure to prevent accidental selection. When the active mode is non-default, a small amber Brain badge appears in the chat header to surface the state and provide a one-click jump back to settings. Hidden in Mini Mode, matching the established pattern for Voice, Call, Share, and Incognito.

Resume banner replaces the previous Continue button as the canonical resume affordance for paused or interrupted Codework sessions, with Resume and Dismiss actions that persist. A new header button toggles the file tree pane visibility, with state retained across restarts. Recent sessions now sort by last-touched time instead of filename, and status badges expand to five states (IN PROGRESS, DONE, ERR, STOPPED, or none for unknown).

Concurrent writers (chat page mid-conversation, buddy poll, mobile bridge, Telegram inbound, Spotlight) could previously clobber each other through a last-writer-wins pattern. A new per-session in-process mutex now serializes all writers, and the mobile bridge re-loads the session at write time instead of overwriting with a pre-call snapshot. Race window shrinks from seconds-to-minutes down to milliseconds, effectively race-free for everyday use. Telemetry warns in DevTools console when an unexpected shrink occurs; Manual Compact logs an "intentional shrink" marker so the two are distinguishable.

Assistant turns that carried tool_calls but no captured tool results (interrupted mid-execution) used to vanish from the chat render entirely; they now render as an italic indicator listing the attempted tool names. The sidebar now filters sessions by the currently selected agent so clicking a session no longer reroutes to a different agent's context. Codework projects gained an Open in Finder / Open in Explorer button in the project header. Long task results no longer truncate silently — click any task to open a full-text modal. OpenRouter is now the default provider for new installs, persona persists across conversation restarts, and three API routes that were inadvertently ignored are now tracked correctly.

v10.2.6 detected o1/o3/o4 and bare gpt-5 but missed every dot-version (gpt-5.1 through gpt-5.5 and all suffix variants like -mini, -nano, -pro, -codex, -codex-spark, -chat-latest, -thinking, -instant) plus every dated snapshot. v10.2.7 ships a regex that covers the verified model lineup. Detection is now provider-agnostic, so GPT-5.x routed through OpenRouter, Custom Provider, or any OpenAI-compatible relay gets the correct request shape (max_completion_tokens, no temperature, system-message folding limited to o1 family per current OpenAI documentation).

Previously the Custom Provider builder unconditionally appended /v1/chat/completions to whatever base URL you saved. Any URL that already carried a version segment (Z.ai coding /api/coding/paas/v4, OpenAI-compat /api/openai/v1, etc.) ended up with a duplicate path and returned 404. v10.2.7 detects three cases at fetch time: the URL already ends in /chat/completions (use as-is), the URL ends in a numbered version segment like /v1 or /v4 (append /chat/completions), or the URL ends in /openai or a bare hostname (append /v1/chat/completions). Same logic applies to /models discovery. Verified against OpenAI, OpenRouter, Z.ai, Groq, Mistral, DeepSeek, xAI documentation.

Friend Mode, Identity Maintenance, daily standup, and cron task completion notifications previously built their own request bodies via separate code paths. When a GPT-5.x reasoning model was the default, in-app chat got the v10.2.6 fix while the Telegram-channel paths kept sending max_tokens and returned provider 400s. v10.2.7 routes every OpenAI-compatible call through a single body builder, so reasoning-model handling, system-message folding, and tool-array formatting stay consistent across the chat path, the orchestrator ReAct loop, the autonomous-task fallback, and the code builder.

Each provider card in Settings → AI Providers has a Refresh button that hits /v1/models (or the vendor equivalent). Anthropic, OpenAI, Google Gemini, Groq, DeepSeek, Mistral, xAI, Together, MiniMax, Cloudflare, NVIDIA, SambaNova, and Cerebras supported. Result persists to settings.modelCache; every model dropdown prefers the cached list over the hardcoded baseline. New models become usable without a Skales release.

Settings → AI Providers → Override Model Limits. Add per-(provider, model) override rows for context and output token caps. Use * as the model name for a provider-wide wildcard. Resolves ahead of the static lib/model-limits registry via a 5s in-process cache of settings.modelLimits. Useful for newly released models whose limits differ from the built-in registry.

v10.2.0 declared the feature but the dispatcher was not reaching the fetch calls — the standalone Next.js build did not include the undici package, so the runtime require returned undefined silently. Fixed by making undici an explicit dependency, switching the dispatcher to a proper import, and externalizing it in next.config.mjs so Webpack does not parse undici's private class fields.

v10.2.0 trimmed the in-memory message array but did not propagate the deletion to disk. Fixed by saving the trimmed session explicitly from the delete handler. Buddy-poll dedup baseline updated in step so the poll does not re-inject removed messages.

v10.2.0 surfaced the Branch hover button on chat bubbles but the slice index was wrong — the new session contained a different subset than the user expected. Fixed end-to-end: clicking Branch on the Nth message creates a new session with messages 1 through N inclusive. Snapshot is captured up-front so re-renders cannot shift the slice mid-handler.

A handful of chat.bubble.* locale keys shipped without translations in v10.2.0 — the Branch toast, the Delete tooltip, and the Branch error showed raw key strings. Real translations added across all 12 locales (en, de, es, fr, hr, ja, ko, pt, ru, tr, vi, zh).

Settings → Per-Mode Model Overrides previously showed a hardcoded curated list regardless of what the user had configured. Now uses the same getCuratedModels() helper as the chat header and Playground pickers — provider dropdown lists only enabled providers with API keys, model dropdown reflects configured + curated models for the chosen provider, plus a free-text fallback for off-list models.

Playground on-page model picker and Settings Per-Mode Override now share the same source of truth (settings.modeOverrides.playground). A small window-event bus fires after each save so the other surface re-hydrates instead of going stale until reload.

Chat header and Playground header model pickers hide below 768px. Matches the existing Mini Mode behavior — at iPhone widths the picker cluttered the chat header without serving a clear purpose.

The "Skales agents at a glance" trigger on the Settings page was removed. It belongs on the Agents page only, which is unchanged.

Context window and max output tokens now read from a per-provider, per-model registry. Replaces the previous hardcoded 32K context fallback and 2048-4096 output ceilings across orchestrator, chat, autopilot, code-builder, browser-control, autonomous-runner. Live HF Router context_length values can override the static entry at call time.

When effective context exceeds 75% of budget, older turns are LLM-summarized via a single low-cost call instead of being truncated to 280 chars per message. Falls back to the truncation behaviour on summary failure. Upstream slice(-N) caps in chat.ts and chat/page.tsx replaced with byte-budgeted slices so compaction is actually reached.

customProviders[] array supports more than one custom endpoint. Legacy providers.custom keeps working and is mirrored as customProviders[0] via a one-shot migration. Settings UI exposes label, base URL, API key, model, enabled toggle, tool-calling and vision toggles per entry.

Resolver chain: caller override → settings.modeOverrides[mode] → settings.activeProvider/model. Wired into Playground, Buddy, and exposed in Settings as a "Per-Mode Model Overrides" panel covering Chat, Codework, Organization, Studio, Playground, Buddy, Spotlight.

Compact icon-only button in the chat header opens an absolute overlay popup listing installed providers and curated models. Layout never shifts. Solid background respects light/dark theme. Custom agent provider/model wins over the global default. Removed entirely in Mini Mode.

Raw provider errors become friendly toasts with one-click recovery: continue / retry on Ollama crashes and timeouts, compact on context-exceeded, switch-fallback on rate limits and quota errors, open-settings on auth failures. Pattern rules for Ollama, OpenRouter, Anthropic, OpenAI; generic fallback annotates with provider+status for debug.

create_calendar_event description anchors against natural-language calendar phrasing ("in Google Cal", "in den Kalender", "gcal", "schedule a meeting at TIME"). planner_create_task description clarifies it is for autonomous Skales prompts, not human appointments. TASK ROUTING RULES gain a Calendar branch above Planner.

Toggling the master switch from the Autopilot page (not just Settings) now persists isAutonomousMode atomically alongside heartbeat start/stop. Activate-on-Start works from a clean restart regardless of which UI surface the toggle came from.

Provider configs accept an optional proxy: { enabled, url } field. When set, requests to that provider go through an Undici ProxyAgent dispatcher. Cached per URL. Wired into chat.ts and orchestrator.ts hot paths. Schema is additive — existing settings.json files load unchanged.

Hover actions on user and assistant message bubbles. Delete trims trailing tool messages so no orphan tool results remain. Branch creates a new session populated with messages up to the chosen point.

When a model returns finish_reason: 'length' with no further tool calls, the orchestrator injects "Continue from where you left off." and re-enters the ReAct loop instead of dropping the partial answer.

In-memory cooldown timestamps are now persisted to settings and hydrated on restart, so app updates no longer reset proactive-message cooldowns. Every early-return point in the tick emits a structured console.warn for diagnostics.

skalesinsider Text Posts now use a single text field (max 4000 chars) matching what the feed renderer expects. Link Posts use og_url / og_description / og_source. No more crashes from undefined fields.

Entries missing text, og_url, or gif_url no longer crash the feed for the user. Optional chaining and length-guards wrap every field access.

Gemma 3, Gemma 4, LLaVA, Pixtral, Qwen-VL, and MiniCPM-V now correctly route image inputs to the configured vision model. Explicit user-set Vision Model is always respected, even if the model is unknown to auto-detection.

Messages sent from Skales Mobile show a small phone icon next to the timestamp. Read-only visual cue, no behavioural change.

Admin dashboard now distinguishes Desktop, Mobile (iOS / Android), and Cross-Platform users. Discover gamertags carry created_via metadata for platform-of-creation tracking.

Three activation modes alongside the master switch: Manual (legacy default), On Startup (auto-start at app launch), and Time Window (active during a daily 24-hour window). Overnight windows like 22:00 - 06:00 are supported. Manual pauses inside an active window are honoured.

Generate HTML/CSS/Tailwind designs from prompts. 7 starter templates, fullscreen preview, refine drawer for iteration, recent designs persist locally. Type a prompt, get a design, open in your editor.

SDXL, FLUX, and others working again via the Inference Providers Router. View Transitions API powers smooth tab crossfades across Studio.

Approval Gates (auto-approve writes / exec / all). Forbidden command denylist (17+ patterns). Test loop with 3-failure progress guardrail. Preview Mode with diff-based accept/reject. Repo-map indexing scales to 500+ files. Long-context tier adaptation, token usage tracking, commit-message generator.

HF Spaces and MCP servers are now usable in Chat, Codework, and Studio. New “Active Tools Across Skales” Settings panel. Set up once. Use anywhere in Skales.

Lio gets recursive project snapshot (max depth 10, max 2000 files) and improved plan context assembly. Chat token-pruning widened to 120 chars, gpt-tokenizer integration honors provider TPM ceilings, update toast renders the changelog inline in 12 languages.

writeJsonSafe / readJsonSafe across 65+ files. Hierarchy clarity in Organizations, sidebar trailing-slash fix. Codework runtime extracted to actions/codework.ts (route trimmed 488 → 102 LOC). Auto-updater stable since v10.0.3. 12-language locale parity verified.

Outbox foreground sync via AppState listener. 10s sweepStale during active chat use. Debug log noise removed. /lio and /lizard slash-commands removed.

Standalone Android companion + Remote Mode. Pair via QR; phone reaches THIS desktop's 139+ tools through an end-to-end encrypted relay (TweetNaCl box). 27 native mobile tools, image upload, shared Discover Feed.

New Studio video provider. Text→video + image→video, standard & fast variants, 5/10 second clips, native 9:16 portrait, $0.06/sec at 1080p. Shares the existing Cloud Render pipeline.

OpenAI Speech added (alloy/echo/fable/onyx/nova/shimmer). Per-message speaker icon on every AI reply. "Read responses aloud" auto-TTS toggle. Groq-key hint for free Whisper STT.

```html fenced blocks render live in a sandboxed iframe. Show Code / Download HTML / Save as Image (via Electron capturePage) / Mute / Hide. Global mute+hide persist across chats.

Framer Motion spring entrance per message (AnimatePresence + initial=false). Smooth wave typing indicator. Scroll-to-bottom FAB. Global `prefers-reduced-motion` guard.

Buddy drag on Windows, Agent delete fix, Playwright dynamic chromium detection, Telegram auto-restart watchdog, gamertag update-in-place. 12 locales at perfect 3838-key parity.

readJsonSafe/writeJsonSafe used across entire codebase. Fixes Windows crashes caused by UTF-8 BOM in JSON config files.

Complete injection rewrite fixes dead buttons in all Spaces. Bridge API now guards against double-injection and injects before body scripts.

Download button now appears correctly when an update is available. Resolves UI state issue in the update page.

Advisor now validates model availability with fallback chain. Eliminates 400 errors on OpenRouter when configured model is unavailable.

Identity Maintenance preserved, Dreaming staggered to 3:30 AM, Knowledge Graph shows helpful empty-state message.

Browser logins now survive restarts. Persistent context storage for Playwright automation sessions.

Skill loading shows batch summary instead of 28 individual lines. Integrity check downgraded to info-level.

Null-safe rendering prevents crashes on malformed posts. Admin posting and delete restored.

One-click install for Gemma 3, Llama 3.3, DeepSeek R1, Mistral, Phi-4, Qwen 3, Codestral. Download progress bar. Auto-detection of ComfyUI and Stable Diffusion WebUI.

Record browser sessions as replayable workflows. Auto-captures navigations, clicks, and inputs. Schedule as recurring Planner tasks. playbook_run and playbook_list agent tools.

YouTube and LinkedIn OAuth. Post directly from Studio Export. social_post and social_upload agent tools. Export presets for TikTok 9:16, YouTube 16:9, Instagram 1:1, LinkedIn 1.91:1.

Enterprise knowledge graph builds entity relationships as you work. Projects, people, tools, preferences. knowledge_graph_query / update / delete agent tools. Toggle + reset in Settings.

2800+ translation keys across 12 languages. All Settings strings converted from hardcoded to i18n (450+ keys). Bootstrap wizard, module pages, and error messages fully translated.

Render progress tracking, gallery file path resolution, template click pre-fill, FFmpeg path resolution, path traversal hardening (CWE-22), Dashboard i18n keys, Planner calendar events, fallback chain persistence, Studio model names premium refresh.

Save and load reusable project configurations with team, description, and default instructions. Instant pre-fill for the Execute tab.

Team-level memory pool separate from personal chat memory. Agents read and write shared context that persists across execution runs.

Organization agents respect Safe Mode. Destructive tools are blocked via system prompt enforcement — emails saved as files, commands described but not run.

Conversations automatically compress when approaching the token limit. Older messages are summarized to free space while preserving key context.

Prevents runaway agent loops. Max 2 identical tool calls and 20 total tool calls per agent turn. Stops infinite loops and excessive API usage.

Organization agents can send emails as part of workflows. Works with the approval gate — Safe Mode saves drafts, Unrestricted sends immediately.

Native File System Access API for folder selection in Codework and Organization. Automatic fallback to Electron dialog for compatibility.

New Settings section for destructive actions. Delete My Server Data with clear warnings and confirmation. Separate button text from description for clarity.

Google Calendar, Apple Calendar, Outlook, and any CalDAV server. Unified in Planner with conflict detection.

Notion, Todoist, Spotify, Smart Home (Home Assistant), Google Drive, GitHub, Google Docs. All configurable in Settings.

Developer Mode: API Playground, Debug Panel, CLI, comprehensive tool/skill/provider reference documentation.

Import from ChatGPT, Claude, Copilot, Gemini, OpenClaw, Hermes. Settings > Import from Another Tool.

Model Context Protocol for external tools. 5 built-in templates.

New default: deep navy + emerald. 5 themes total. Light mode default on first launch.

Slack and Signal integration alongside existing Telegram, Discord, WhatsApp.

CEO, CTO, Marketing Director, Research Analyst, Project Manager, and more. Each with 200-400 word system prompts.

Lightweight notifications now appear directly in the Skales interface with customizable timing. Less intrusive than system popups and fully respecting your privacy by staying within the app.

Dynamic prompt tuning adapts to your usage patterns and context window constraints. Skales automatically optimizes system prompts for maximum reasoning capability with fewer errors.

Enhanced access control, improved audit logging, and stricter permission validation for multi-user deployments. Better security boundaries between roles.

Custom Skills now run in isolated iframe sandboxes for better security. Prevents accidental or malicious interference with the main app while allowing safe skill development.

Global activity feed showing what the Skales community is building. Gamertag system, upvotes, category filters, blurred preview for non-members. Privacy-first: no personal data collected.

Floating search bar accessible from anywhere on your desktop via Cmd/Ctrl+Shift+S. Ask Skales anything without opening the main window. Eye button captures your screen for visual context — ask about anything visible on your desktop.

Compact always-on-top chat window. Toggle from the chat header. Stays on top of other apps so Skales is always reachable.

Dedicated page for all notifications with read/unread state, category filters, and admin broadcast support. Replaces the old inline notification cards.

Hub-and-spoke node visualization, task history, quick delegate, and chat integration hints. mDNS state now syncs reliably with the Settings toggle.

Audible feedback when tasks complete, notifications arrive, or Swarm tasks finish. Theme-aware sounds, configurable in Settings.

Full month grid with event previews and click-to-navigate. Download your plan as a .ics calendar file without connecting a provider.

Both proactive systems now fire independently of Autonomous Mode. Morning greetings, idle check-ins, and meeting reminders work as configured.

Reduced token usage by moving dynamic context to on-demand tool calls. Free-tier models no longer hit rate limits on simple messages.

Autonomous web browsing with DOM settle delay, auto/manual approval modes, dangerous action blacklist, CAPTCHA detection, and real-time action log.

LAN peer discovery via mDNS. Skales instances on the same network can discover each other and delegate tasks with ping-pong prevention.

Admin broadcasts and bug report replies delivered to users. Dashboard shows dismissable notification cards with read state tracking.

Permanently fixed the recurring approval loop. BLOCKED tool results replaced correctly after approval. System jobs bypass approval gates.

Electron-driven 60-second tick ensures cron jobs actually execute. Overdue catch-up on app launch. isSystemJob bypass for Identity Maintenance.

Hamburger menus for Snowfield and Neon themes on mobile. Obsidian tablet overlap fixed at 1024px breakpoint. Custom Skills in all mobile menus.

Fixed "Invalid Date" for Google Calendar events. Planner sync errors resolved. Weather widget uses device geolocation with Vienna fallback.

Email subscriptions now transmitted to server. Opt-in with explicit consent. 9 languages, 1745 translation keys fully synced.

Customizable drag-and-drop dashboard with resizable widgets (clock, weather, system stats, quick actions, recent chats, tasks, notes, pomodoro). Powered by react-grid-layout with per-widget settings.

Natural-language and cron-syntax task scheduling with job management UI, execution logs, and retry support. Jobs stored and managed with full visibility.

Built-in Puppeteer browser for web research and interaction. Sandboxed with user approval flow and screenshot capture for autonomous web tasks.

Background agent that monitors and executes scheduled tasks autonomously. Toggle in Settings > Advanced.

Full-screen VAD-based continuous voice conversation. STT > LLM > TTS pipeline with barge-in support, waveform visualization, and end-call keyword detection.

LAN peer discovery via mDNS. Discover other Skales instances on the network, view status, and send tasks. Hidden behind Settings > Advanced > Experimental.

Progressive Web App manifest with installable icons. /mobile page with 4-step Tailscale setup wizard and QR code for phone access.

View your own bug reports with status badges (open, in progress, closed). Optional email field on bug reports. Newsletter opt-in in onboarding and settings.

KoboldCpp, LM Studio, and vLLM are now first-class provider options in onboarding. No longer hidden under "Custom Endpoint."

Voice settings now support local TTS servers (KoboldCpp, XTTS-API-Server) and Whisper for local speech-to-text transcription.

Configurable image generation endpoint alongside Replicate. Run image models locally without cloud APIs.

Fixed bot process crash on end-user machines. Bot now uses Electron's built-in Node runtime instead of requiring system Node.js installation.

Fixed chat becoming unresponsive after vision model error. Session history is now sanitized before every API call. Vision fallback strips images gracefully when models don't support vision.

Added 60-second inactivity timeout to prevent chat UI from hanging permanently on broken API responses.

Rule-based buddy intelligence observes calendar, email, tasks, and idle time. Meeting reminders, end-of-day summaries, idle check-ins, morning greetings. Respects quiet hours with no LLM calls.

AI-powered daily scheduling with 8-step wizard. Learns work patterns, generates time-blocked plans from calendar events, pushes them back to your calendar.

Google Calendar, Apple Calendar (CalDAV/iCloud), and Outlook (Microsoft Graph API). All three work simultaneously. Planner AI reads from all providers.

Upload Lio AI projects to any FTP server. Per-project deploy config, incremental upload (only changed files), test connection, 4 website starter templates.

English, German, Spanish, French, Russian, Chinese (Simplified), Japanese. Full UI translation including onboarding and system prompts.

Three modes: Unrestricted, Workspace Only, Custom Folders. Enforced on all file tools for safer file access control.

Daily digest of calendar events, pending tasks, and unread emails delivered via Telegram and chat.

7-step wizard with Cloud/Local/Custom provider cards, Ollama auto-detect, buddy picker, and safety mode selection.

Fixed endless Telegram approval loop. 9 tools missing from TOOL_SAFETY map now properly handled. Read-only tools no longer require repeated approval.

TOOL_SAFETY fallback changed from 'confirm' to 'auto'. Session history now preserves tool results with orphan protection.

Fixed hardcoded German language setting. Now uses user's configured nativeLanguage/locale for TTS generation.

Fixed telemetry key mismatch. /api/settings endpoint now exposes telemetry_enabled. Feedback page uses correct URL. Feature Request textarea is now editable.

Skin/language change now prompts for restart with Electron relaunch support. Chat input focus preserved across sessions.

White screen on mobile Chrome tab switch via Tailscale fixed. Privacy section consolidated into Security & Privacy tab.

Master Plan now generates cron jobs for recurring goals. "Check my email every morning at 8am" creates an actual scheduled task, not a one-shot.

New tab in Autopilot dashboard shows real-time agent reasoning, tool calls, and results as they happen. Watch Skales think in real-time.

Approve or reject Autopilot tasks directly from Telegram. Reply `approve ` or `reject `, or simple `yes`/`no` for single pending tasks.

Meet Bubbles — a playful blue liquid blob that morphs into different shapes. Selectable in Settings → Desktop Buddy alongside the original Skales gecko.

New /feedback page with 3 sections: Rate Skales (4 emoji ratings), Report a Bug, Request a Feature. GDPR compliant data handling.

Telegram approval gate respects safety mode. Inline keyboard buttons replaced with text-based approval. Agent no longer hallucinates tool execution. Telemetry opt-in enforcement at API level.

Orphaned tool_result blocks no longer crash API. Message sanitization in agentDecide() removes tool_results referencing non-existent tool_calls.

Reset All button purges Telegram data. Replicate images now save to workspace/files/images/ correctly. Bubbles mascot skin added.

Safety Mode simplified to Safe + Unrestricted (removed confusing Advanced mode). Approval flow no longer stalls. Agent loop continues after approval (ReAct continuation).

Telegram bot shows proper messages instead of raw translation keys. Markdown formatting cleaned before sending. Inline buttons appear correctly.

Agent executes immediately instead of explaining. Platform-aware system prompt with PowerShell vs bash rules. Enhanced anonymous telemetry with opt-in control.

Full UI translation: English, Deutsch, Español, Français. Language picker as first onboarding screen. System messages, approval prompts, and error messages translated.

BYOK access to 50+ image and video AI models with one API key. Bring-your-own-key for image generation.

Supports llama.cpp, LM Studio, vLLM, koboldcpp, text-generation-webui. Tool calling toggle for custom endpoints (on/off for local models).

Smooth 150ms crossfade transitions. Dynamic folder system. Full skin system. Buddy can execute tools: write files, send emails, browse, manage calendar.

Anonymous telemetry opt-in (disabled by default). Bug report button with system info. Feedback sent to developer via collect.php.

Destructive actions (send email, delete file, calendar changes, tweets) now require explicit user confirmation. Browser blacklist prevents bypasses. Unrestricted Mode properly bypasses approval gate.

ARIA labels on all interactive UI elements. Full keyboard navigation with visible focus indicators. Screen reader support via aria-live regions. Compatible with NVDA and VoiceOver.

Friendly error messages. Video transition flickering fixed. Honest responses when tools unavailable. Button reliability improvements.

Floating mascot with finite state machine animations. Glassmorphism input field for Spotlight quick action. AI response bubbles with auto-dismiss. Settings toggle for always-on desktop companion.

Deep-dive interview learning your goals. Master Plan generation with task list. OODA self-correction loop. Human-in-the-loop approval gates. Anti-loop protocol with retry tracking. Daily stand-up report. Execution Board Kanban interface.

Microphone button for voice input. Fullscreen voice chat mode with status labels. Groq/OpenAI Whisper transcription. ElevenLabs TTS with browser fallback.

ZIP upload to install custom capabilities. AI scaffolding generates skill definitions automatically. Skills page with enable/disable, metadata view, deletion. Sandboxed execution.

Word (.docx), PDF, and Excel (.xlsx) generation from natural language. Fully-formatted documents saved to DATA_DIR/documents/.

Nearby search, place details, geocoding, directions, and photo URLs. Pure REST API implementation without native binaries.

English, German, Spanish, French, Russian, Chinese, Japanese, Korean, Portuguese. Full UI translations with proper locale formatting.

Connect KoboldCpp, LM Studio, vLLM, and other OpenAI-compatible local LLM servers directly to Skales.

Offline text-to-speech, speech-to-text, and image generation powered by local models. No external API calls required.

Chat with Skales across your favorite messaging platforms. Unified conversation history and skill access.

Skales (gecko), Bubbles (blob), and Capy mascots. Animated, interactive desktop companions with customizable behaviors.

LLM-generated schedules with Google Calendar, Apple CalDAV, and Outlook integration. One-click sync to calendar.

Build web applications with AI, then deploy directly to FTP or SFTP servers. Supports all major hosting providers.

Controlled file access with read-only, read+write, and full modes. Protect your system with granular permissions.

Enable autonomous task execution with full reasoning transparency. Skales breaks down complex tasks and executes them step-by-step.

Read, compose, and send emails through your IMAP and SMTP accounts. Seamless email integration within Skales.

Skales remembers your preferences, past conversations, and important context. Build on knowledge across sessions.

Interactive approve/reject buttons for destructive tool calls during Organization execution. State machine architecture breaks execution into segments for true non-blocking approval.

Refactored execution engine that breaks multi-agent workflows into resumable segments. Enables pause, resume, approval gates, and retry without blocking the server thread.

Full create, rename, archive, restore, and continue workflow for Organization projects. Switch between saved configurations instantly from the Execute tab.

Decide how often and which proactive notifications Skales may show. Discover activity, unread email digests and similar pings get their own frequency and type controls under Settings > Notifications, including off. Planned for the days right after v11.3.0.

Hook the v10.2.2 live model fetch into the chat header picker so users can refresh the model list without opening Settings. Same cache, same fallback to the hardcoded baseline.

Per-provider proxy is currently configurable only via settings.json edits. v10.3 surfaces it as a UI toggle and URL field on each provider card — same dispatcher path that v10.2.2 hardened.

Wire the v10.2.0 customProviders[] array into the existing fallback chain so power users can chain multiple OpenAI-compatible endpoints alongside cloud providers.

Persistent Goal system, autonomous runner on the heartbeat, long-term memory upgrades and the messaging gateways. Landed across the v11 releases above.

iOS Beta progressing through Apple Review. Android Closed Testing live and expanding. Same Skales pairing flow, same provider stack.

Personalized daily digest with weather, calendar events, tasks, and memory highlights at a glance.

Stay informed with curated newsletters and product updates delivered directly to Skales.

Enhanced voice synthesis and more natural conversational responses from your Desktop Buddy.

Access Skales from any mobile device on your network through a Progressive Web App.

Connect with other Skales users, share skills, and collaborate on automation ideas.

EV code signing certificate for Windows builds. No more SmartScreen warnings on first launch.

Public marketplace for browsing, publishing, and installing Agent Skills. One-click import into Skales.

Flexible database backend selection. Choose PostgreSQL for cloud deployments or SQLite for lightweight local setups. Easy migration tools included.

Discover and install community-created plugins with one click. Rate, review, and manage your installed plugins from a centralized marketplace.

Native support for Model Context Protocol servers. Connect advanced tools and services directly to your Skales instance for enhanced capabilities.

Social feed, Skales Wrapped, Token Compressor, Friend Mode, Safety Mode improvements, Spotlight full tool access. See v8.0.0 section above.

Google Veo 2, Kling AI, Runway Gen-4, and Replicate (MiniMax/Seedance) fully integrated in Studio. Generate cinematic video from text prompts with real-time progress polling and gallery save.

Inline connected/not-connected status indicators for Google Calendar, Apple Calendar, and Outlook on the Settings page. Green dot when linked, red when missing.

13 new slash commands: /memory, /skills, /provider, /version, /export, /theme, /language, /settings, /discover, /studio, /codework, /wordpress, /status. Translations synced across all 12 locales.

React Native companion app for iOS and Android. Full remote control of your desktop instance: chat, trigger tasks, review briefings, and approve tool executions from your phone.

Automated daily briefing at your configured time. Pulls calendar events, pending tasks, unread emails, weather, and top news into a structured summary delivered in Chat.

Kling AI, Veo 2, MiniMax Video, and Seedance integration in Studio. Generate high-quality cinematic video from text prompts using frontier cloud video models.

Complete the social media suite with Meta platforms. Post images and videos directly from Studio Export to Instagram, TikTok, and Facebook Pages via OAuth.

Install and manage local TTS voice models: Coqui TTS, Bark, Piper, OpenVoice. Browse voices, preview samples, and set defaults per language — all without an API key.

Daily usage streaks, weekly activity scores, and milestone badges. Visualized on the Dashboard and in Skales Wrapped. Friendly gamification to keep your AI habits consistent.

Community challenges, trending skills leaderboard, push notifications for Sparks and replies, and verified creator badges. Making Discover the most active AI community feed.

Hosted version of Skales. Same features, zero local setup. End-to-end encrypted. Deferred while local performance improvements are prioritized.

Native iOS and Android app for managing tasks, reviewing briefings, and controlling Autopilot from anywhere.

Discover, install, and manage community-built skills, integrations, and custom agents from a curated marketplace.

End-to-end encrypted sync of memory, settings, and custom skills across all your devices. Optional and never required.

Hosted version of Skales for users who prefer not to run locally. Same features, zero setup. Your data encrypted at rest and in transit.

Native mobile app for iOS and Android. Full chat, memory, and agent capabilities in your pocket. Syncs with your desktop instance.