Mint v1.5.2

Mint Release Notes

v1.5.2

This release improves desktop readability and adds repository-aware CLI tools for summarizing codebases, indexing symbols, and searching source code semantically.

Repository Intelligence, Semantic Code Search, and CLI Readability

This patch makes Mint more useful inside a project workspace. The CLI can now inspect a repository directly, build a lightweight source symbol index, and create a semantic code index backed by Gemini embeddings.

Highlights

• Desktop UI Polish: Updated the app interface layout and control styling for a cleaner, more organized desktop experience.
• Clearer Conversation Flow: Improved chat presentation so user and AI messages are easier to distinguish while reading.
• Readable AI Replies: Adjusted AI response formatting to produce cleaner, more structured replies with better spacing and less clutter.
• Thai Reply Handling: Improved response behavior for Thai conversations so replies remain natural and easier to read.
• Reduced Visual Noise: Tightened UI states and message rendering so important conversation content stays the focus.
• Repository Summary: Added mint summarize [path] and /summarize [path] to report project structure, package metadata, git state, languages, directories, and important files without requiring a full AI agent turn.
• Symbol Index: Added mint symbols [path] and /symbols [path] to scan JavaScript, TypeScript, Python, and Rust source files for functions, classes, exports, interfaces, types, structs, enums, traits, and related symbols.
• Semantic Code Search: Added mint semantic-code index and mint semantic-code search <query> for embedding source chunks and searching code by meaning.
• Natural Workspace Requests: The interactive CLI can detect plain-language requests such as repository summaries, symbol indexes, and semantic code searches, including Thai phrasing, and route them to the local tools.
• Working Timer: The CLI now shows how long Mint has been working while the agent is thinking.
• Cleaner Streaming: Empty assistant/system chunks are filtered so live replies do not create blank messages.
• Unified Patch Preview: Code-agent patch approvals now show a unified-diff style preview with surrounding context before applying edits.

New CLI Commands

• mint summarize [path]
• mint summarize [path] --json
• mint symbols [path]
• mint symbols [path] --json --limit 120
• mint semantic-code index [path]
• mint semantic-code search "<query>" --path <path> --top-k 5

New Interactive Commands

• /summarize [path] [--json]
• /summary [path] [--json]
• /symbols [path] [--json] [--limit n]
• /symbol-index [path] [--json] [--limit n]
• /semantic-code index [path]
• /semantic-code search <query>
• /semantic index [path]
• /semantic search <query>

Repository Summary

The repository summary tool reports:

• root path
• scanned file count
• package name, version, description, scripts, dependencies, and dev dependencies
• git branch, short status, and diff stat
• top-level directory counts
• language counts by extension
• important files such as README, package metadata, release notes, source files, tests, config files, and CI files

Symbol Index

The symbol indexer scans supported source files locally and reports:

• source files scanned
• files containing symbols
• total symbol count
• counts by symbol kind
• counts by language
• symbol names with file and line references

Supported source extensions:

• .js, .cjs, .mjs, .jsx
• .ts, .tsx
• .py
• .rs

Semantic Code Search

Semantic code search creates workspace-specific indexes under ~/.config/mint/semantic-code.

• Uses gemini-embedding-001 by default.
• Reads the Gemini API key from Mint config or GEMINI_API_KEY.
• Chunks source files with overlap so search results can point back to file and line ranges.
• Includes file path, language, and local symbol names in each embedded chunk.
• Stores file hashes with the index metadata for workspace tracking.
• Supports formatted output and --json output.

Code Agent Patch Preview

Patch approvals now use unified diff previews:

• shows --- a/<path> and +++ b/<path> headers
• includes @@ range metadata
• includes nearby unchanged context
• displays removed and added lines with - and +
• falls back to a clear preview error if a hunk cannot be matched

v1.5.1

This release updates the desktop Live2D experience, CLI response loop, learned skills, and image attachment UX.

Desktop Live2D Model and Interaction UI

This patch adds an optional Live2D assistant model to the desktop UI and improves the top toolbar for easier model control.

Highlights

• Live2D Assistant Model: Added a Shiroko Live2D model panel to the Electron desktop assistant.
• Model Visibility Preference: The model panel can be shown or hidden from the toolbar. New installs start with the model hidden until enabled.
• Expression Cycling: Added a toolbar control for cycling Live2D expressions.
• Expression Toast: The active expression name is shown in a small on-canvas toast after changing expressions.
• Click-to-Chat Reactions: Clicking named model areas can trigger temporary expressions and send a short contextual prompt into the normal chat flow.
• Interaction Area Overlay: Added a transparent guide overlay for clickable regions such as Head Pat, Cheek Poke, Hand Tap, Shoulder Tap, and Careful.
• Safe Missed Clicks: Clicking outside the defined interaction zones no longer changes expression, sends chat prompts, consumes AI tokens, or starts cooldown.
• Interaction Cooldown: Model reactions use a 3-second cooldown to reduce accidental repeated AI calls.
• Live2D Lip Sync: Mint animates mouth parameters while speaking and resets them when speech ends.
• Toolbar Redesign: Desktop header controls are grouped into clearer model, utility, and window-control sections with readable labels.

Live2D Desktop Model

• The model is loaded from models/Shiroko_Model.
• The desktop model can be toggled from the Model toolbar button.
• The first launch defaults to hiding the model panel.
• The visibility setting is stored in localStorage as mint-model-hidden.
• Expression controls use the configured Live2D expression files and reset expression parameters before switching to avoid stuck expressions.
• Temporary click reactions return to the previously selected expression after 2 seconds.

Interaction Zones

The interaction guide overlay labels clickable regions without blocking pointer input:

• Head Pat
• Cheek Poke
• Hand Tap
• Shoulder Tap
• Careful

Only clicks inside these configured regions trigger Live2D reactions. Clicks outside the regions are ignored.

Fast Mode, Live Replies, and Learned Skills

This patch improves the Mint CLI interaction loop and adds persistent skill/instruction learning from local markdown or text files.

Highlights

• Agent Status Label: The interactive CLI now shows [Agent] instead of [Chat] for the normal agent loop.
• Fast Mode: Added /fast, /fast on, /fast off, and /fast status.
• Fast Status: When Fast Mode is enabled, the status bar shows [Fast].
• Quiet Trace Output: Fast Mode keeps ● Mint is thinking... visible but hides Thinking: and intermediate tool/progress trace messages.
• Live Mint Replies: Final Mint responses now render into one live-updating Mint message block instead of appearing all at once or splitting into multiple Mint messages.
• Learned Skills: Added mint learn <path> and /learn <path> for importing local .md or .txt files as persistent skill/instruction context.
• Skill Management: Learned skills can be listed and deleted from both the command line and interactive CLI.
• Expanded Slash Suggestions: /memory skills is now available in the / command suggestions.
• Expanded Tests: Test suite now reports 137 passing tests.

Fast Mode

New interactive commands:

• /fast toggles Fast Mode.
• /fast on enables Fast Mode.
• /fast off disables Fast Mode.
• /fast status shows the current Fast Mode state.

Fast Mode changes presentation only. Agent routing, tool calls, approvals, final answers, and provider/model behavior continue to work normally.

Live Response Rendering

• Thinking: messages can appear before the final answer when Fast Mode is off.
• Final responses stream into one active Mint message block.
• The completed message is stored in history after the response finishes.
• Fast Mode still streams the final answer, but hides the internal Thinking: trace.

Learned Skills

New command-line examples:

• mint learn ./skill.md
• mint learn ./instructions.txt
• mint learn --list
• mint learn --delete <id|path|name>

New interactive CLI examples:

• /learn ./skill.md
• /memory skills
• /memory skills delete <id|path|name>

Learned files are stored in Mint's SQLite memory as persistent skill/instruction documents. Mint injects learned skills into long-term context on later turns.

CLI Image Attachments & Paste UX

This patch improves the interactive Mint CLI input experience, especially for screenshots, clipboard images, and long pasted text.

Highlights

• CLI Image Attachments: Added --image <path> support for mint chat and mint code.
• Interactive Image Paste: Ctrl+V can attach clipboard images inside the interactive CLI without sending immediately.
• Multiple Images: Users can attach several images before pressing Enter; the UI labels them as [Image #1], [Image #2], and so on.
• Attachment Controls: Ctrl+Backspace removes the latest pending image, and Esc clears pending images or pasted content before exiting.
• Long Paste Placeholder: Long or multi-line pasted text is collapsed in the input as [Pasted Content N chars] while preserving the full content for sending.
• Natural Composition Order: Text typed before an image or pasted content, the attachment/content itself, and text typed afterward are sent in the same natural order.
• Vision Context Follow-up: The CLI keeps lightweight context from the latest image turn so follow-up questions can refer back to recently attached images.
• Expanded Tests: Test suite now reports 137 passing tests.

CLI Image Input

• Added src/CLI/image_input.js.
• Supported image formats:
  • .png
  • .jpg
  • .jpeg
  • .webp
  • .gif
• New command-line examples:
  • mint chat --image ./screenshot.png "What is on this screen?"
  • mint code --image ./mockup.png "Build this UI"
• Interactive CLI examples:
  • Ctrl+V to attach a clipboard image.
  • /paste [prompt] to attach a clipboard image through a command.
  • /image ./screenshot.png [prompt] to attach an image file while already inside mint.

Interactive CLI UX

• Clipboard images are now attached first and sent only when the user presses Enter.
• Pending image labels are rendered inside the input box.
• Multiple pending images are shown as compact labels:
  • [Image #1] [Image #2]
• Fixed terminal input edge cases where Ctrl+V could insert stray v characters.
• Long pasted text no longer breaks the input border or terminal layout.
• Pasted content can be combined with typed text:
  • text before paste
  • pasted content
  • text after paste
• Image attachments follow the same composition model:
  • text before image
  • image labels
  • text after image

Provider and Vision Handling

• Gemini image MIME handling now preserves the real image MIME type instead of assuming PNG.
• Chat provider paths can receive multiple images in one request where supported.
• Code Agent multimodal requests now pass attached image context to supported providers.

Current Test Status

• 137 tests passed
• 20 test suites passed