mergegate/GOALS.md

Miyabi CLI - Development Goals

Reference Projects

Baseline

• Codex CLI: https://github.com/openai/codex.git

Extension Source

• Miyabi Private: /Users/shunsuke/dev/01-miyabi/_core/miyabi-private

---

Mission

Codex CLI同等の品質を持つ、Production-Ready Miyabi CLIを構築する

---

Gap Analysis Summary

Metric	Codex CLI	Current Miyabi	Target
Total Lines	20,969	1,200	15,000+
Files	50+	10	40+
Features	Complete	Basic	Full Parity

---

Phase 1: Core TUI Foundation ✅ COMPLETED

Status: Done (1,200 lines)

• Cargo workspace setup
• wrapping.rs - Text wrapping with textwrap
• history_cell.rs - Card design UI
• markdown_render.rs - Basic markdown
• app.rs - Main TUI loop
• event.rs - Event handling

---

Phase 2: Advanced Text Rendering 🔄 IN PROGRESS

Target: +1,500 lines

2.1 markdown_stream.rs (690 lines)

• Streaming markdown rendering
• Progressive display during generation
• Cursor position tracking
• Partial code block handling

2.2 diff_render.rs (715 lines)

• Git diff visualization
• Syntax highlighting for diffs
• Line number display
• Add/Remove/Change indicators

---

Phase 3: Chat Composer Separation

Target: +4,000 lines

3.1 chat_composer.rs (3,938 lines)

• Separate input handling from main app
• Multi-line input support
• Command history
• Auto-completion
• Syntax highlighting in input

3.2 textarea.rs (2,213 lines)

• Full-featured text area widget
• Selection support
• Copy/paste
• Undo/redo

---

Phase 4: Approval & Feedback

Target: +1,500 lines

4.1 approval_overlay.rs (656 lines)

• Tool execution approval UI
• Y/N/Always/Never options
• Risk level display
• Preview of changes

4.2 feedback_view.rs (555 lines)

• User feedback collection
• Rating system
• Comment input

---

Phase 5: Advanced Navigation

Target: +2,500 lines

5.1 resume_picker.rs (1,580 lines)

• Session resume selection
• Conversation history browser
• Search functionality
• Preview pane

5.2 pager_overlay.rs (913 lines)

• Full-screen content viewer
• Scroll and search
• Jump to line
• Keyboard shortcuts

---

Phase 6: LLM Integration

Target: +2,000 lines

6.1 Claude API Integration

• Anthropic API client
• Streaming responses
• Token counting
• Rate limiting
• Error handling

6.2 Conversation Management

• Context window management
• Message history
• System prompts
• Tool definitions

---

Phase 7: Tool Execution

Target: +2,500 lines

7.1 A2A Bridge Integration

• Tool registry
• Execution engine
• Result formatting
• Error recovery

7.2 Built-in Tools

• File operations (Read, Write, Edit)
• Bash execution
• Search (Glob, Grep)
• Web fetch

---

Success Criteria

Functional Requirements

• Real Claude API conversations
• Tool execution with approval
• Session persistence and resume
• Markdown rendering with streaming
• Full keyboard navigation

Quality Requirements

• 15,000+ lines of code
• 90%+ test coverage
• < 100ms response latency
• Memory efficient (< 100MB)
• Cross-platform (macOS, Linux)

User Experience

• Intuitive keyboard shortcuts
• Clear visual feedback
• Smooth scrolling
• Proper error messages
• Help system

---

Timeline

Phase	Duration	Deliverable
Phase 1	✅ Done	Core TUI
Phase 2	1 week	Streaming Markdown
Phase 3	2 weeks	Chat Composer
Phase 4	1 week	Approval System
Phase 5	1 week	Navigation
Phase 6	2 weeks	LLM Integration
Phase 7	2 weeks	Tool Execution

Total: ~10 weeks to Production-Ready

---

Next Action

Phase 2.1: Implement markdown_stream.rs

Create streaming markdown renderer that can display partial content during LLM generation.