You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Seamless On-Demand Headless Browser Automation (CLI, Desktop, and Web)
Zero-Setup Headless Browsing: Mint Agent automatically spawns an isolated Chromium instance (--headless=new, --disable-gpu, --no-sandbox, 1440x900 viewport) on demand whenever web inspection or navigation is requested, eliminating the need to run mint auto in a separate terminal.
Default Browser Tools: Enabled browser_open, browser_screenshot, browser_read, browser_click, and browser_observe by default across all platforms without requiring manual configuration.
Prompt Guidance & Anti-Scraping Rules: Updated system prompt instructions to prioritize browser_open for web URLs / SPAs and strictly forbid writing ad-hoc Python scraping scripts (urllib, requests) or raw curl when inspecting web pages.
Age Gate (18+) & Cookie Consent Protocols: Automatically dismisses standard Age Verification gates ("I am 18 or older", "Enter Site") and cookie consent banners via browser_click; prompts for user confirmation with visual screenshots (browser_screenshot + ask_user) for sensitive content warnings.
Settings Toggle Parity: Added the Browser automation tools toggle switch in the Automation tab across Desktop and Web Settings (defaulting to enabled).
Interactive Slash Commands (/browser & /browser status): Added the /browser command across CLI TUI, Desktop, and Web. Pressing Enter opens an interactive prompt choice to toggle tools (on/off), switch between headless Browser and headed Automation, or view browser status and detected Chrome binary via /browser status.
Queued Messages & Steering (CLI, Desktop, and Web)
Queued Prompts Tray & Auto-Dispatching: Users can now queue instructions and multi-modal attachments while the agent is actively executing (sending). Queued messages are displayed in a dedicated tray above the composer and auto-dispatched sequentially as soon as the current turn concludes.
Stop with Paused Queue (Safe Interruption): Clicking the Stop button halts the active turn while preserving all queued messages in a ⏸ Queue paused state, allowing users to inspect, resume, edit, or clear queued items safely.
Interrupt & Steer (->): Clicking the Send Now arrow button on any queued prompt immediately interrupts the running turn and steers the agent directly to the selected queued message.
Composer Edit & Recall (✎): Clicking the edit button removes the prompt from the queue and restores its full text and attachments back into the composer for immediate revision.
Full-Screen TUI Visual Parity (crates/mint-cli): Pressing Enter while a turn is running queues the prompt with live border feedback (Queued (N): "..." · Ctrl+D drop), and Ctrl+D immediately pops/drops the last queued message.
Desktop & Web UI / UX
Clean File Changes Cards: Extracted repetitive common directory paths to the card header with folder icons, replaced [NEW FILE] brackets with sleek rounded green chip badges (New file), and redesigned the expander button with a subtle dashed border and rotating chevron icon.
Adaptive Artifact Preview Panel:
Automatically adapts header tabs based on file capabilities: displays [Preview | Code] dual-mode toggles only for renderable formats (HTML, SVG, Markdown), while presenting direct code view with language chip badges (e.g. [<FileCode /> Python], [<FileCode /> Rust]) for pure source code and data files.
Image files display dedicated format chip badges ([🖼️ WEBP], [🖼️ PNG], [🖼️ JPEG]) along with live image resolution badges (e.g. 1080 × 1620 px).
Restricted "Open in Mint Browser" action strictly to HTML/Web documents, avoiding misleading browser launches on scripts, code, and images.
Live Tool Activity Retention: Fixed an issue where older tool execution groups (e.g. run_shell) were temporarily trimmed during live agent turns once non-thought events exceeded 20; preserved all tool lifecycle and orchestration events throughout live streaming without dropping earlier progress.
Action-First Tool Activity Parsing: Eliminated null/undefined placeholder strings from displayed tool arguments across CLI, Desktop, and Web by prioritizing action-specific fields and sanitizing dummy strings in mint-core and frontend utilities.
Resilient Media & Image Proxy: Added local image proxy (/api/proxy-image) with disk LRU caching and SSRF protection to bypass ad-blocker false positives; auto-heals unclosed markdown image links (, normalizes Twitter/X banner URLs, and displays a graceful fallback card on load errors.
Provider Brand Badges & Logo Picker: Vector SVG brand logos in chat message footers across 13+ providers; redesigned custom provider logo picker with preset SVGs, emoji selection, and an upload modal with pan/crop/rotate controls.
Theme Refresh (Studio Dark & Charcoal Onyx): Introduced IDE-inspired "Studio Dark" palette (#101010), refined dark themes with balanced Charcoal Onyx styling, suppressed colored aurora background glow blobs in Studio Dark for a solid pure black chat canvas (#101010), and added sleek neutral gray buttons with subtle mint hover glow.
Configurable Step Limits (/steps): Control max tool execution steps per turn (10–200, default 40) via /steps slash command or Settings across Desktop, Web, and Full-Screen TUI.
Trajectory Export & Diagnostics: One-click SWE-bench/Agent Protocol v1.0 trajectory export (.json); dedicated high-contrast callout styling for self-correction guidance; persistent activity drawer accessible across completed and interrupted turns.
Full-Screen TUI Onboarding Wizard (mint onboard & /onboard):
Overhauled to native Ratatui alternate-screen wizard with responsive edge-to-edge layout and Mint TUI styling.
Step 1 (Core AI): Unified Gemini setup with masked key input (Ctrl+H), live model discovery, and integrated model picker.
Step 2 (Services): Grouped checklist with arrow navigation, spacebar toggles, select-all (a), and invert (i).
Step 3 (Configurations): Paged per-service cards featuring bundled API credentials and real-time model pickers (Anthropic, OpenAI, DeepSeek, Ollama, etc.).
Step 4 (Summary & Launch): Visual configuration overview with direct launch (Save & Launch Mint) or exit options.
Added /onboard slash command and first-run detection across CLI, Desktop, and Web.
Terminal Polish: Compact web search citations (top 5 with Ctrl+O view web_search smart jump, matched tool-card ANSI color styling, and structured results formatting), automatic source citation reconstruction on session resume (/resume), unified command badges, external browser link handling, and non-blocking desktop titlebar during settings.
Performance & Frontend Optimization
Re-render Isolation: Decoupled message input state to eliminate full-page re-renders while typing.
Component Splitting: Isolated streaming elapsed timers into leaf components and lazy-loaded secondary modals.
Zero-Allocation Buffers: Replaced 32 MiB buffers with lightweight UTF-8 estimators and compact attachment fingerprinting.
Adaptive Polling: Switched background jobs and status polling to idle-aware intervals that pause on inactive tabs.
GPU Compositing Relief: Removed backdrop blur filters on scrollable chat containers for smooth virtualized scrolling.
Core Engine & Web API Server
Step Budget Warning & Wrap-Up Synthesis: Proactive warning at max_steps - 5 and graceful synthesis pass (tools: None) ensuring a complete final answer instead of abrupt termination.
Structured Self-Correction Diagnostics: Diagnostic engine (diagnostics.rs) detecting patch indentation drift, line numbering offsets, and compiler errors (Rust, Python, TypeScript) with actionable recovery suggestions.
Standardized Trajectory Schema v1.0: Versioned, typed trajectory runs (StandardTrajectoryRun) with step tokens, thoughts, and replay support via /trajectory.
Head-Tail Truncation: Dual-bound log clipping (first 35 lines context, last 75 lines error output), cutting token overhead by up to 70% while preserving root-cause errors.
Repetition & Loop Guard: Detects duplicate tool executions and oscillating loops (A→B→A→B), injecting self-correction guidance without aborting the turn.
Context Compaction Role Alternation: Alternating Assistant/User format for summaries, fixing HTTP 400 crashes on reasoning models (DeepSeek, Gemini).
Engine Reliability: Non-blocking SQLite persistence via spawn_blocking, low-latency TCP sockets (TCP_NODELAY), standardized workspace path resolution, and terminal API noise suppression for high-frequency Library thumbnails (/thumbnail), asset content (/content), and MCP artifacts.
