Release Notes

v0.9.107 — May 7, 2026

A massive release. Headlines: Claude Code support for Pro/Max subscribers, a complete Image Generation workflow with gallery, Image Tools for editing, Image Annotations, Layered Memory with a three-tier instruction system, YOLO mode, granular per-tool permissions, sound notifications, taskbar alerts, cost-saving Task Agent models, a PDF Reader sub-agent, a brand-new Three.js skill, hooks, local vision models, and a long list of stability and performance gains.

New Features

• Claude Agent (Claude Code Integration) — Sign in with your Claude Pro or Max subscription and run Claude Code directly inside TatsuCode as a first-class provider. Great news for subscribers who want to use their plan without API billing. Supports streaming, durable file diffs, session restore, and Cross-Agent Relay handoffs to other providers mid-conversation
• Image Generation — A complete image generation workflow built into TatsuCode. Generate, edit, inpaint with masks, and use reference images. Configure via the new /image-generation command. Works across providers — OpenAI Direct (GPT Image 2, GPT Image 1.5, GPT Image 1 Mini), Google Direct (Nano Banana 2, Nano Banana Pro), and OpenRouter (GPT-5.4 Image 2, Nano Banana 2 / Pro, Riverflow V2 Pro)
• Image Gallery — Full-featured gallery for browsing every image in your conversations and projects. Stable with 10,000+ images, with quota detection and clean lifecycle handling
• Image Tools — A new built-in skill for editing images on the fly: resize, crop, trim, blur, and more. The agent can use these directly when working on graphics, mockups, or assets
• Image Annotations — Mark up images directly in chat to point things out, circle areas, or add notes for the agent to act on
• YOLO Mode — Session-scoped permission bypass for power users who know what they're doing. Acknowledgement modal on first activation, status-bar badge while active, plus a /yolo toggle
• Granular Per-Tool Permissions — A new /permissions overlay lets you allow individual tool calls from the command line, including unknown tools. Permissions can be granted permanently or just for the current session
• Sound Notifications — New /sounds command adds chimes when a task finishes or when a permission/question prompt needs your attention. Plays only when the TatsuCode window is unfocused so it doesn't interrupt active work
• Windows Taskbar Overlay Badges — When TatsuCode is running in the background, the taskbar icon shows a yellow dot for permission prompts and a red dot for completed turns or alerts. Especially useful when running multiple TatsuCode instances. Cleared on focus
• Elapsed Time per Task — Long agent runs now show how long the task took to finish, and total work time is included in /usage
• PDF Reader Sub-Agent — A dedicated sub-agent that handles PDFs through models with native PDF input and renders pages reliably (charts, diagrams, scanned content), so PDF analysis no longer depends on external Poppler tools being installed
• Cross-Agent Relay — Continue a conversation across providers without losing context. Hand off from one agent to another mid-task with a manifest that preserves files, tools state, and conversation history
• /models-taskagents Command — Assign cost-efficient models to your parallel sub-agents (Web, Code Explorer, Design Reviewer / Architect, YouTube Video, PDF Reader) so scout work runs on smaller/faster models while your main model handles code review, design judgment, and other heavy lifting. Reduces token cost and keeps long sessions going further
• Native Project Folder Picker — Click the folder icon, project label, or welcome banner to open a real Windows folder picker instead of typing paths
• Local Provider Vision Support — Local models like Gemma 4 and Qwen-VL now have full vision support. TatsuCode auto-detects model capabilities (name, context size, vision modality) at runtime by querying the local server, with self-healing retry if the server starts after TatsuCode
• Hooks System — Run scripts at specific lifecycle events (turn start, tool calls, permission prompts, etc.) to customize TatsuCode for your workflow
• Layered Memory — A three-tier instruction system that gives you finer control over how TatsuCode behaves: global preferences in ~/.tatsu/AGENTS.md follow you across every project, project rules live in AGENTS.md / CLAUDE.md / .cursorrules / similar files in the repo, and a new repo-local override in .agents/AGENTS.local.md lets you add machine-specific or experimental rules that win over the broader layers without committing to the shared file. A separate session continuity layer also keeps the chat's current objective and key context alive across compactions and session restores, so long conversations stay coherent even after the transcript has been summarized
• Crash Logger — If TatsuCode ever does crash, a detailed crash log is captured to make debugging and support much easier

New Skills

• Three.js — A complete skill for Three.js / WebGL / WebGPU / React Three Fiber work, backed by 13 deeply-researched code-prompt resources covering core architecture, animation, loaders, materials, physics, and post-processing
• Image Generator — Operational skill that drives the new image generation pipeline (provider-aware presets, prompt recipes, edit/mask workflows)
• Image Tools — Operational skill for the new image editing tool set
• Cross-Engine Gamedev Resources — Seven new shared resources (PBR materials, physics, post-processing, color grading, skeletal animation, texture prompting, visual verification) used by both Unity and Three.js skills so quality stays consistent across engines

New Slash Commands

Improvements

• Massive Streaming and Queue Rewrite — Full rewrite of the message queue and steering system for tighter real-time guidance while the agent is working, plus stronger streaming performance and JSON sanitization
• Office Skills Stabilized and Upgraded — Word, Excel, and PowerPoint now have non-destructive list, extract, replace, and add-image actions. Better handling of locked files (e.g. open in another app), and improved animations and UI for office file results. PowerPoint slide inserts now preserve the deck's theme, layout, and footers; Excel append inherits styles from the prior row
• Compaction Reliability — Smoother auto-compact behavior on long conversations with fewer interrupted responses and better recovery
• Longer, Safer Chat Sessions — Chat sessions are now resilient to system crashes, power loss, and unexpected shutdowns. Long sessions are saved more frequently and restore cleanly
• @-File Search Now Scales to 500k+ Files — File picker is paginated and event-driven, eliminating bridge-cap crashes and spinner flicker on huge projects
• Memory & Performance Hardening — Proactive cleanup for streaming sessions, image cache, DevBrowser state, and per-session caches. Reduced memory growth across long runs and /new resets. Bounded diagnostic logs so noisy plugins don't bloat memory or context
• Image Pipeline Cleanup — Less duplicated images in sessions, better lifecycle handling, and stable rendering for galleries with 10k+ images. XSS prevention hardening across image rendering paths
• Sentence-Level Loop Guard — A new safeguard catches subtle reasoning loops (including character-by-character spell-out loops) without false-positiving on genuinely repetitive topic-focused work
• Improved Computer Use — Better screen capture quality and more reliable click/type/drag flows
• Improved Screen Recording — Higher frame rate and better clarity for recordings the agent reviews
• Improved Mermaid Diagrams — More reliable rendering with sanitizer-based retry, plus prompt guardrails that produce cleaner, more concise diagrams
• Refreshed Tatsu Branding — New TatsuCode icon and updated brand colors throughout the app and website
• Updated Themes — New themes added to /settings, with improved colors, contrast, and readability across the board
• Curated OpenRouter Model List — Removed underperforming models so the default lineup only includes models that actually work well with TatsuCode
• GPT-5.5 Added (OpenAI Codex Provider) — New model added to the Codex provider lineup, plus higher input image resolution support for GPT-5.5 and Claude Opus 4.7
• GPT-5.4 Image 2 — Added to the Image Generation model list
• API Keys Export Improved — Export now includes image generation keys and settings, and OpenRouter first-time setup is smoother
• Smarter Sub-Agents — UI/UX revisions and behavior improvements across the Task Agent, Web Agent, Code Explorer Agent, Design Reviewer / Architect Agent, YouTube Video Agent, and the new PDF Reader Agent
• Plugin & Skill Compatibility — Better cross-provider compatibility for plugins and skills, plus more Unity method aliases so common requests resolve correctly
• Unity Plugin Updates — Improved connection reliability, version updating, and bridge stability. Image captures from Unity no longer balloon context — they route through the image cache like any other image

Bug Fixes

• Fixed application focus bug where TatsuCode could steal focus or leave a stale block cursor when not the foreground app
• Fixed race condition where a cancelled terminal command could still run after a delayed trigger
• Added safety guardrails to prevent creating or opening projects in Windows folders that could be destructive (system directories, drive roots, etc.)
• Fixed Start Menu shortcut duplication on update
• Fixed false-positive "unzip failed" error on first launch
• Fixed Mermaid diagram print rendering bug
• Fixed minor compaction edge cases
• Fixed dropdown rendering glitches when switching providers and models
• Fixed cases where the agent would redundantly re-read images that were already attached
• Fixed local server image rejection by transcoding WebP screenshots to PNG before sending

v0.9.93 — March 12, 2026

New Features

• Plugin System + Unity Editor Integration — TatsuCode can now connect to external applications through a new plugin system. The first built-in plugin connects to a live Unity Editor, letting you create and modify GameObjects, materials, shaders, scenes, animations, UI, VFX, terrain, lighting, and more — all from chat. Manage plugins with the new /plugins command
• Desktop Automation (Computer Use) — A new built-in skill lets TatsuCode interact with any desktop application. Click, type, drag, scroll, and manage windows using native controls. Great for automating repetitive tasks, filling forms, or interacting with apps that don't have APIs
• Screen Recording — Record video clips from any window or screen. TatsuCode can then watch the recording to give feedback, catch issues, or verify behavior — useful for UI reviews and testing workflows
• GPT-5.4 Support — Full support for OpenAI's latest flagship model, including native web search, image detail modes (up to original resolution), and text verbosity control. Available via OpenAI Direct and GitHub Copilot connections
• 2x Tiered Pricing Detection — TatsuCode now warns you when your conversation crosses from base to extended (2x) pricing on supported models. An optional auto-compact mode keeps your context within the cheaper tier automatically
• Image Viewer Upgrade — Images in chat can now be zoomed and navigated through a full image gallery. Browse all images from the conversation without leaving the window
• Session, Project, and Model Deletion — You can now delete old sessions, remove unused projects, and clean up models directly from the UI with confirmation dialogs

New Skills

• Brainstorming — Collaborative design and planning skill that gates implementation until the design is reviewed and approved
• Computer Use — Desktop automation skill with a screenshot → analyze → act → verify workflow
• Windows Control — CLI-first Windows automation using PowerShell, COM objects, and system utilities

New Slash Commands

Improvements

• Updated Default Models — GPT-5.4 and GPT-5.3 Codex are now included as startup models on OpenRouter
• DevBrowser Enhancements — Added network interception, element-level screenshots, and mousewheel scroll control
• Input Area Resizing — Improved text input area behavior when composing longer messages
• Session Restore — Improved reliability when resuming previous chat sessions
• Updated Linting Tools — Refreshed built-in linting tools with improved project detection
• Plugin Stability — Extensive stability work across the plugin system and Unity bridge
• Updated Copilot Model Lineup — Refreshed available models for GitHub Copilot connections
• Native Window Management — Removed legacy Windows.Forms dependency for a cleaner, lighter application

Bug Fixes

• Fixed auto-compact not triggering for GitHub Copilot connections
• Fixed Office document conversion failing in published/distributed builds
• Fixed incorrect icon appearing in the thinking animation after YouTube video analysis
• Fixed PDF and video attachments not persisting correctly across sessions
• Fixed session storage bloat from duplicate attachment files
• Fixed UI truncation issues and search result display glitches
• Fixed minor rendering issues with GPT-5.4 model responses

v0.9.90 — February 26, 2026

New Features

• Native Office Files, End-to-End — You can now read, create, and edit DOCX, XLSX, PPTX, and CSV directly with much stronger output quality for real documents, spreadsheets, and slides
• Targeted Office Editing — New focused actions for Word, Excel, and PowerPoint let you update specific content (sections, ranges, slides, comments, formatting, rows/columns, and more) without rewriting the entire file
• Audio File Input — Audio is now first-class input support for compatible models (.mp3, .wav, .m4a, .aac, .flac, .ogg, .aiff)
• Expanded Video Input — Video support now works across more compatible model routes, including local video files and better per-model validation
• YouTube Analysis Modes — YouTube workflows now support transcript-only, direct video analysis, and isolated deep-analysis mode with improved metadata display

Improvements

• Browser Automation + Performance Upgrade — Browser workflows are substantially expanded (keyboard/mouse controls, drag-and-drop, dialog handling, cookies/storage, source/network inspection), and browser tooling now loads on demand to keep normal coding sessions lighter
• Provider & Model Experience Polished — Updated startup defaults, improved model preselection, and Copilot model refreshes (including Gemini 3.1 Pro as default and Claude Sonnet 4.6 support)
• Error Clarity and Copy UX Improved — Error/warning content is easier to copy, JSON-heavy failures are surfaced in clearer human-readable form, and key provider flows now include faster copy actions

Bug Fixes

• Fixed startup crash scenarios when running commands like /new or /clear before provider setup is complete
• Fixed stale model dropdown/cache issues after provider disconnect or model state changes
• Fixed Gemini thought-signature/loop edge cases that could produce broken or confusing streamed output
• Fixed OpenAI usage/quota and direct-client edge cases for more accurate limits/reporting behavior

v0.9.87 — February 16, 2026

New Features

• GitHub Copilot Provider — Sign in with your GitHub subscription and use Copilot-backed models directly in TatsuCode
• Agent Skills System — New skills workspace with /skills, /skills-template, and /skills-refresh, plus the built-in skill-editor for guided skill creation and editing
• Deep Think Skill (Experimental) — New built-in skill focused on complex planning, debugging, and structured reasoning workflows
• Advanced Drag & Drop Attachments — Drag in images, files, and folders from inside or outside your project; when a file can’t be directly read, TatsuCode now keeps its path as context
• Expanded Model Lineup — Added support for GPT-5.3 Codex (including fast GPT-5.3 Codex Spark preview) and updated high-end Claude model availability

Improvements

• Provider Connections UX Overhaul — /connect is now the primary provider command, onboarding is smoother, first-time model selection is automatic, and disconnect flows are cleaner
• Better Quota Visibility — Improved /usage-quota views across subscription providers, including OpenRouter balance + usage breakdown
• Smarter Auto-Compact — Better recovery when long chats approach context limits, with fewer interrupted responses
• Stronger Real-Time Steering — Messages sent while the AI is already working now keep attachment context and render correctly as queued follow-ups
• Cleaner Chat Experience — Better auto-scroll behavior, always-visible status banner while scrolled up, cleaner diff headers, and horizontally scrollable markdown tables
• Faster Web Research Reads — Web fetch performance for large browser-rendered pages improved significantly (up to ~10x on heavy pages)
• Copilot Reasoning Controls Improved — More reliable reasoning controls for Copilot routes (including Extra High) plus better Claude/Gemini compatibility through Copilot
• Lower Prompt Overhead — Reduced instruction overhead for longer conversations and more efficient context usage

Bug Fixes

• Fixed ESC cancellation reliability across streaming responses and active tool execution
• Fixed model-selection bleed between multiple running TatsuCode instances
• Fixed Codex streaming edge cases where hidden stream errors could appear inside otherwise successful responses
• Fixed PDF upload/reading reliability for Gemini-family models through OpenRouter
• Fixed invalid OpenRouter API key acceptance during validation
• Fixed quota reporting inaccuracies (including Claude overcount in specific cases)
• Fixed session restore edge cases around compact/continuation message handling
• Fixed cases where requests kept retrying when they should fail immediately, now shown with clearer errors

Availability Update

• Google Direct Connections — Temporarily disabled pending provider authorization updates. Gemini models remain available through OpenRouter.

v0.9.82 — February 4, 2026

New Features

• Ctrl+C Process Interrupt — Send Ctrl+C to running shell sessions to stop long-running processes or cancel operations mid-execution

Improvements

• Enhanced Shell Sessions — Ring buffer maintains 5,000 lines of history, paginated reads with offset/limit, regex filtering with context lines, and ANSI escape code stripping
• Longer Chat Sessions — Auto-compact now triggers at 99% context capacity instead of 95%, allowing extended conversations before automatic summarization
• Multi-Instance Support — Run multiple TatsuCode windows simultaneously without browser cache conflicts or LLM task interference
• Session Restore Fidelity — Restored sessions now preserve images, tool call visuals, terminal outputs, and diff views exactly as they appeared
• Command Safety — Added comprehensive terminal safety lists: 18 dangerous commands blocked, 17+ useful utilities whitelisted for project protection

Bug Fixes

• Fixed @ file references not being tracked correctly when restoring saved sessions
• Fixed user messages with images not displaying properly after session restore
• Fixed codefence rendering inside tool call result blocks
• Fixed loading animation showing extra "..." characters
• Fixed cursor position offset in certain text input scenarios
• Fixed UTF-8 BOM issues causing shell session first command failures
• Fixed partial JSON in tool calls that could corrupt session files (now auto-repaired)

v0.9.80 — January 31, 2026

New Features

• Real-Time AI Steering — Send messages while the AI is actively working to guide and redirect it mid-task. Your input is injected immediately for instant course correction
• Chain Failure Auto-Stop — Visual badge counter tracks consecutive failures and automatically stops at 5 attempts with actionable recovery options

Improvements

• Enhanced connection stability during extended thinking sessions with HTTP/2 keepalive
• Improved streaming display with better markdown rendering and collapsible sections
• Improved caching integration across all providers; allowing more sessions and longer ones
• Better Gemini Advanced performance with improved streaming and parallel tool calling
• Auto-compact now works reliably across all providers (Anthropic, OpenAI, Gemini, Local)
• Improved Reasoning / ScratchPad section headers for clearer output
• Clear feedback when operations are cancelled or interrupted
• Startup protection prevents running from ZIP files or protected system directories

Bug Fixes

• Fixed token tracking display for Claude/Anthropic users
• Fixed /usage cache showing incorrect values
• Fixed ESC key sometimes sending double messages
• Fixed Google direct provider "This version is no longer supported." error
• Fixed auto-compact not triggering for some providers
• Fixed memory leaks in LLM client connections

v0.9.78 — January 27, 2026

New Features

• File Undo & History — Revert any file edit during your session with full undo support. View edit history, search across changes, and restore files to any previous state
• Parallel Task Agents — Run up to 5 AI agents simultaneously for faster research and code exploration
• Read-Before-Edit Safety — Files must be read before editing or overwriting, preventing accidental data loss from edits to unseen files

Improvements

• Faster startup when local AI servers (Ollama, LM Studio) are offline — no more freezing
• Better reasoning display with proper formatting while streaming (Claude, OpenAI, Gemini)
• Cross-platform dev tools detection now supports 75+ tools with 40+ common aliases
• Shell commands now display app name badges for clearer command identification
• Config location moved to ~/.tatsu/ for better cross-platform consistency (auto-migrated from old location)
• Dark background during window resize eliminates white flash
• Gemini streaming now shows real-time feedback during file operations

Bug Fixes

• Fixed first screenshot not being visible to AI in some cases
• Fixed reasoning display not updating live for OpenAI subscription users

v0.9.77 — January 24, 2026

New Features

• Update Notifications — A non-intrusive banner now appears on the welcome screen when a new version is available, with quick access to release notes and download links

Improvements

• Refined theme colors and design across all 14+ themes for better contrast and readability

Bug Fixes

• Fixed SaveNote tool failing to save notes correctly

v0.9.76 — January 23, 2026

New Features

• @ File Auto-Read — Files referenced with @ mentions are now pre-read in the background for faster context loading
• Right-Click Paste Images — Paste images directly via the right-click context menu, in addition to Ctrl+V

Improvements

• File dropdown now auto-refreshes stale cache for better accuracy when files change
• Clear error messages when models stop responding or fail to complete (no more silent failures)
• Improved Mermaid diagram rendering for OpenAI models
• Added FileDestructive and DockerDestructive permission categories for safer command execution

Bug Fixes

• Fixed Redis DEL pattern incorrectly matching Windows del command
• Fixed session history corruption from malformed JSON in file contents

v0.9.75 — January 21, 2026

New Features

• Compact Command — New /compact slash command with "Compact Now" and "Auto Compact" menu options for managing conversation context
• Multi-Agent Support — Run up to 3 SubAgents simultaneously for parallel research and code exploration
• Session History — Improved Browse and restore previous chat sessions from the /session slash command
• Code Context Panel — View referenced files and symbols in a dedicated diff view panel

Improvements

• Faster file indexing for large projects (50%+ improvement)
• Better syntax highlighting for 20+ additional languages
• Reduced memory usage when working with many open files
• Improved error messages when API connections fail

Bug Fixes

• Fixed issue where keyboard shortcuts stopped working after closing a modal
• Resolved memory leak when processing large codebases
• Fixed file watcher not detecting changes in nested directories

v0.9.74 — January 15, 2026

New Features

• Updated Browser Automation Tools — More Built-in Chromium browser tools for testing web applications
• Screenshot Analysis — Take screenshots and have the AI analyze UI/UX issues
• Web Research — Search the web and fetch documentation directly from chat

Improvements

• Added keyboard shortcut reference (press ? to view all shortcuts)
• Improved dark theme contrast for better readability
• Faster startup time with lazy-loaded components

Bug Fixes

• Fixed crash when opening files larger than 10MB
• Resolved encoding issues with non-ASCII filenames
• Fixed clipboard paste not working in certain text fields

v0.9.73 — January 8, 2026

New Features

• 200+ AI Models — Access all major providers through OpenRouter integration
• Local Model Support — Connect to Ollama, LM Studio, and llama.cpp servers
• Bring Your Subscription — Use existing ChatGPT Plus or Gemini Advanced accounts

Improvements

• Redesigned provider settings for easier API key management
• Added model search and filtering in the model selector
• Improved streaming response display with proper markdown rendering

Bug Fixes

• Fixed token counting inaccuracies for certain models
• Resolved issue with conversation export to markdown
• Fixed theme not persisting after restart

v0.9.72 — January 1, 2026

Initial Beta Release

• First public beta of TatsuCode
• Core agentic coding capabilities with file operations
• Shell command execution with safety controls
• Git integration for status and diff viewing
• 14 built-in color themes
• Windows x64 and ARM64 support
• Zero installation required — single portable executable

For feature requests and bug reports, reach out on TatsuCode (GitHub).