Kai: all versions and old releases

Features
- Add Requesty and 1min.AI as providers
Fixes
- Keep chat history intact when compaction fails or a tool loop trims context
- Back up an unreadable settings file instead of overwriting it, and make desktop settings writes atomic
- Fix Debian sandbox installs in Android work profiles and cloned apps
- Detect notification listener access on Android 8.0
- Stop or time out shell commands instead of leaving them running in the background
- Cancelling a request no longer shows up as an error
- Rename MCP tools whose names clash with another tool
- Accept OpenAI-compatible base URLs pasted with a trailing /chat/completions
- Show the provider's message for request-too-large errors on non-image requests
- Route GPT-5.6 and newer model families to the Responses API
- Send the OpenCode session header to every opencode.ai OpenAI-compatible instance
- Treat empty scheduled-task triggers and null tool arguments as absent
- Fire kai-ui countdown callbacks only from interactive messages
- Fix a stray %% in the Kai Build download label
- Verify email TLS hostnames and reject line breaks in IMAP/SMTP arguments
- Check every fetch_url redirect hop and cap the bytes it reads
- Limit open_url file:// links on Android to the sandbox's /root
Improvements
- Don't retry provider errors that a retry can't fix
- Return tool errors as structured JSON
- Refresh model catalog Elo scores
- Upgrade dependencies

Features
- Support OpenAI's Responses API for GPT-5.6 models
- Add LFM2.5 1.2B Instruct to the on-device model catalog
Fixes
- Cap stored message size so a large attachment can't brick the app
- Fix the Debian sandbox install failing on the base package step
- Send the OpenCode session header so OpenCode Zen no longer rejects requests
Improvements
- Let on-device models declare their own tool support and sampling defaults, so models without a tool template no longer invent answers
- Refresh the on-device catalog: updated gemma-4-12b-it build with vision, audio, and multi-token prediction
- Refresh model catalog Elo scores
- Upgrade dependencies

Features
- Receive shared text from the Android share sheet
- Add Debian as a sandbox distribution, shared with Kai Build
- Add rename and delete for Kai Build projects
- Add custom model entry for OpenAI-compatible services
- Add an optional API key field when adding the Jina MCP server
Fixes
- Fix model catalog context windows across major families
Improvements
- Respect the system text-to-speech engine and voice
- Improve accessibility
- Improve PTY rendering
- Refresh model catalog Elo scores, free-tier models, and popular MCP servers
- Unify the platform, tool-layer, persistence, provider, and permission abstractions
- Unify the staggered settings layouts
- Upgrade iOS LiteRT-LM to 0.15.0
- Migrate deprecated Gradle DSL and clipboard APIs
- Add missing localizations
- Upgrade dependencies

Features
- Add Claude Code, Grok, and OpenCode support via the Debian Linux terminal
- Add checksum verification for local model downloads
Improvements
- Improve Linux filesystem file handling
- Refresh open files when switching to a Linux system tab
- Upgrade dependencies

Features
- Add free-tier service suggestions when a rate limit is reached
- Add custom local model .litert import support
Improvements
- Refresh popular MCP server list
- Update model catalog Elo scores
- Upgrade dependencies

Features
- Add Perplexity as a service
- Add AI Horde as a service
Fixes
- Align socket timeout with chat request timeout
Improvements
- Improve Alpine package search ranking
- Upgrade Alpine sandbox to 3.22
- Prefer requestScrollToItem for list scrolling
- Update model catalog Elo scores

Features
- Save copies of sent emails to the IMAP Sent folder
- Resolve email tool accounts by id, address, or single account
Fixes
- Fix heartbeat CompletionHandlerException
- Add java.sql to desktop jlink runtime modules
Improvements
- Increase chat request timeout to 180 seconds
- Auto-install bash package during Alpine Linux install
- Update model catalog (Kimi K3, Inkling, Elo scores)
- Upgrade dependencies

Features
- Migrate conversations to SQL storage
- Add Gemma 4 12B local model and missing catalog entries
Fixes
- Fix Android 17 local network permission request
Improvements
- Improve service and tool calling
- Add Atlas Cloud UTM tracking
- Update model catalog
- Upgrade dependencies

Features
- Add per-model image-capability gating so text-only models no longer offer image input
- Add Android assist-gesture launch support
Fixes
- Reload conversations after importing settings
- Restore the Fetch URL tool in the Android tool list
Improvements
- Group sponsored sections in the UI
- More prominent interactive loading animation
- Update model catalog (including DeepSeek V4 1M context window)
- Additional localizations
- Upgrade dependencies

Improvements
- Upgrade dependencies
- Update model catalog

Features
- Introduce Skills — packaged, reusable instructions the assistant can load on demand via the Linux sandbox, plus a built-in "create-skill" workflow to author your own
- Add Atlas Cloud as a featured provider
Fixes
- Fix text selection in chat list
- Fix heartbeat active hours slider ranges
- Fix heartbeat Linux sandbox running in current conversation
- Fix terminal scroll out of bounds
- Fix OpenAI-compatible model loading when only the default URL is set
Improvements
- Upgrade dependencies
- Update model catalog and scores
- Improve free endpoint DTO sanitization
- Extend API error mapping

Features
- Parse inline tool calls
Fixes
- Fix multiple service deletes in a row
- Fix DeepSeek thinking-mode tool calls returning reasoning_content error
- Improve fallback error handling
Improvements
- Improve model sorting
- Update model catalog
- Upgrade SDKs and LiteRT
- Compose performance improvements
- Internal code cleanup

Features
- Collapsable reasoning content UI with cleaned-up handling across services
Improvements
- Improve UI performance
- Add missing localizations
- Upgrade dependencies
- Internal code cleanup

Features
- Allow user-installed certificates
- Desktop AVX2+ CPU compatibility check
Fixes
- Fix MCP tools serializing array and object arguments as strings
- Fix settings crash on Android 8.0
- Fix potential terminal sheet crashes
- Surface real errors from OpenAI-compatible chat
Improvements
- Tolerate non-standard MCP tool schemas
- Improve SSH Alpine integration (requires re-installing Alpine)
- Add honesty, tool-use, and acting sections to chat system prompt
- Make chat input persistent between screen changes
- Upgrade Compose Multiplatform and SDKs
- Update model catalog
- Remove iOS x64 target and deprecated APIs

Fixes
- Fix terminal auto-scroll crash
Improvements
- Update model catalog
- Add dynamic model endpoint for LongCat service
- Add Ollama Gemma4 to no-tool model list
- Support .gsc files and increase upload size limit to 200KB
- Show history delete undo inside modal
- Upgrade dependencies (coroutines, datetime, and others)

Features
- Theme switcher and native desktop UI scale settings
- Linux sandbox icon flash on shell execution
Fixes
- Fix DeepSeek tool execution
- Fix sponsors display
Improvements
- Improve Ollama error handling
- Compose performance improvements
- Add missing localizations
- Upgrade LiteRT
- Upgrade SDKs

Features
- Persistent terminal sessions in conversations
- Notification read integration
Fixes
- Fix Linux sandbox initialization
- Fix Linux sandbox cd command
Improvements
- Show detailed package upgrade info in Linux sandbox
- Show byte size for files smaller than 1 KB
- Material 3 search inputs
- Update model catalog and scores
- Upgrade SDKs

Features
- Linux sandbox file browser with file and folder actions
- Linux sandbox package manager
- Open files from the Linux sandbox
- Themed icon support
- Section selection dialog for exports
Fixes
- Fix markdown regex parsing
Improvements
- Persist Linux sandbox UI state during the session

Fixes
- Fix calendar event time zone issue
- Fix conversation restore ANR
- Fix iOS quick toggle hidden behind soft keyboard
- Fix escape character handling
Improvements
- Add scheduled task details
- Simplify memory, schedule, and heartbeat tool toggles
- Improve email polling and simplify email/SMS tool UI
- Add missing model info
- Migrate deprecated APIs

Features
- Add SMS read and send support to FOSS build
- Add iOS and desktop notifications support
- Add Z.AI Coding Plan service
Fixes
- Fix out-of-memory crash
- Fix crash when deleting models
- Fix crash when downloading models while offline
Improvements
- Improve IMAP tool
- Use existing catalog for model context window info
- Upgrade Kotlin
- Add missing models

Fixes
- Fix kai-ui rendering and restore issues
Improvements
- Make OLED dark mode opt-in
- Improve heartbeat, email polling, and task scheduling reliability
- Prevent incompatible models from being used in interactive chat and automations
- Align general grid UI
- Align localizations
- Upgrade dependencies

Features
- Add LaTeX code block rendering
Improvements
- Pure black backgrounds in dark mode for OLED power savings
- Improve fallback status UI
- Add Mistral 3.5 models to catalog
- Update Kimi model catalog (add K2.6, refresh Arena scores)
- Update free API endpoint
- Upgrade dependencies

Features
- Custom math formula rendering
- Expandable and editable memories
- Zoom and pan for attached images
Fixes
- Hide thinking section on chat restart
Improvements
- Add Alpine Linux package download mirrors
- Pulse animation on interactive chat stop button
- Add missing model info
- Performance improvements
- Replace ABI filter with runtime check
- Remove unnecessary Android version checks

Features
- Add local Qwen 3 0.6B model
Improvements
- Upgrade LiteRT
- Defensive desktop ProGuard rules

Features
- Add custom markdown parser
- Preserve previous Kai UI output and allow editing
Improvements
- Pulse animation on stop button
- Add missing localizations
- Add missing model info

Features
- Add copy-to-clipboard button for code blocks
Fixes
- Filter out Gemini API responses marked as internal thoughts
- Add size safeguards for PDF and image compression failures
Improvements
- Highlight currently selected model in model picker
- Improve Linux sandbox interactive interface
- Adjust Gemma 4 device performance expectations
- Upgrade BouncyCastle and remove version checker dependency
- UI polish and alignment

Features
- Add fast and expert mode toggles for free services
- Add Windows menu entry
- Enrich model catalog with arena scores, release dates, and context windows
- Support multi-file upload for attachments
- Enable simple tools for local Gemma 4
- Select TTS voice based on current locale
Improvements
- Improve conversation restore on app restart and settings navigation
- Upgrade SDKs

Features
- Add local model context configuration
Fixes
- Fix undo snackbars not executing after leaving screen
Improvements
- Improve filetype checks for file attachments

Features
- Add support for local Gemma 4 via LiteRT
- Add PublicAI, Fireworks, DeepInfra, and OpenCode services
Improvements
- Add missing localizations

Fixes
- Fix Bouncycastle and Coil proguard rules on desktop
Improvements
- Simplify and improve chat input UI
- Extract reusable UI components
- Improve heartbeat error messages

Fixes
- Fix toolbar vertical padding
- Fix removeLast crash
Improvements
- Add custom JSON parser for more flexible and resilient UI parsing
- Improve desktop release performance and reduce app size
- Simplify UI spacing by handling padding definitions locally
- General UI/UX improvements

Features
- Add possibility to define service for heartbeat
- Add additional MCP servers
Improvements
- Improve kai-ui prompt and split-block parsing
- Replace Lottie animation with custom implementation
- Improve tool use UI
- Add missing localizations
- Clean up implementations

Fixes
- Fix daemon start crash
- Fix heartbeat thinking in wrong conversation
- Fix leaked path in sandbox
Improvements
- Improve shell integrations
- Add uninstall linux sandbox confirmation dialog
- Add NDK 29 requirement and reproducible build flags

Improvements
- Replace standard tabs with custom piles
- Build proot from source

Improvements
- Improve heartbeat reliability with previous response context and failure logging
- Add proper context compaction for conversations
- Upgrade SDKs

Fixes
- Fix crash when restoring encrypted backups

Features
- Add Splinterlands LLM auto-battle integration
- Add web loading spinner
- List sponsors on free tier UI
- Add bottom scroll pointer icon and rounded corners stop icon
Fixes
- Fix heartbeat and tasks handling
- Fix icon button sizes
Improvements
- Upgrade Kotlin to 2.3.20
- Upgrade Koin

Features
- Add scroll to bottom button
- Show selected model on service rows in settings
Fixes
- Fix model filter
Improvements
- Improve startup time and migrate deprecated APIs
- Improve performance

Features
- Add HuggingFace, LongCat, Moonshot, Together and Venice service support
Improvements
- Sort services in modal alphabetically
- Clean up service quick select and history modal UI

Features
- Add option to reset modified heartbeat
- Add button to cancel ongoing requests
- Add search input to model modal
Fixes
- Fix settings import parsing
- Fix MCP server tool definitions
- Add missing UI pointer