winter-super-cli 2026.6.11 → 2026.6.13

package/README.md

@@ -2,17 +2,16 @@
 
 > **Winter** is a powerful, AI-driven CLI coding assistant designed to supercharge your development workflow directly from the terminal. Inspired by Claude Code and Codex CLI, Winter brings a beautiful Cyberpunk aesthetic, smart session management, autonomous agent capabilities, and 25+ AI tools to your local environment.
 >
-> *Winter la mot tro ly lap trinh CLI manh me duoc dieu khien boi AI, giup tang toc quy trinh phat trien truc tiep tu terminal. Lay cam hung tu Claude Code va Codex CLI, Winter mang den giao dien Cyberpunk dep mat, quan ly phien lam viec thong minh, kha nang agent tu tri, va hon 25 cong cu AI.*
+> *Winter là một trợ lý lập trình CLI mạnh mẽ được điều khiển bởi AI, giúp tăng tốc quy trình phát triển trực tiếp từ terminal. Lấy cảm hứng từ Claude Code và Codex CLI, Winter mang đến giao diện Cyberpunk đẹp mắt, quản lý phiên làm việc thông minh, khả năng agent tự trị, và hơn 25 công cụ AI.*
 
 ---
 
 ## Key Features
 
 | Feature | Description |
-|---|---|
+|---------|-------------|
 | **26 AI Tools** | Read, Write, Edit, Bash, Glob, Grep, Notebook, Todo, Scheduler, Agent, MCP & more |
 | **Multi-Provider** | Anthropic, OpenAI, Ollama, Groq -- with **smart context routing** |
-| **26 Tools** | Full parity with Claude Code + extra tools (WebArchive, HtmlEffectiveness, Notebook, Todo, Scheduler, InsertText, StrReplaceAll) |
 | **Context Router** | Auto-selects best provider per task (Claude for code, OpenAI for docs, Groq for speed, Ollama for privacy) |
 | **Session Management** | Isolated sessions with persistent memory, plans, and context |
 | **MCP Support** | Model Context Protocol -- connect VS Code, GitHub, databases & more |
@@ -30,20 +29,22 @@
 | **Clipboard Integration** | `/paste` command reads clipboard directly |
 | **Image Support** | Analyze images & screenshots via `/image` |
 | **Cross-Platform** | Windows, macOS, Linux -- full shell support |
+| **Codebase Index** | Fast semantic search with optional CodeGraph integration |
+| **Embedded Context Corpus** | Built-in knowledge base from 7,900+ embedded documents |
 
 ---
 
 ## Project Stats
 
 | Metric | Value |
-|---|---|
-| **Total Lines of Code** | ~13,500 |
-| **Source Files** | 40+ |
+|--------|-------|
+| **Total Source Files** | 125 |
+| **Source Lines of Code** | 26,080 |
+| **Source Code Size** | ~1 MB |
 | **Tools** | **26** |
-| **Tests** | **213** (all passing) |
-| **Exported Symbols** | 78 |
-| **AI Providers** | 4 (Anthropic, OpenAI, Ollama, Groq) |
 | **Slash Commands** | 50+ |
+| **AI Providers** | 4 (Anthropic, OpenAI, Ollama, Groq) |
+| **Bundled Resources** | 8,578 files (design guides, agent docs, karpathy tools, page agents) |
 | **Node.js** | >= 18.0.0 |
 
 ---
@@ -99,7 +100,7 @@ winter chat "explain this" --no-interactive
 Winter supports **4 AI providers** with automatic context routing:
 
 | Provider | Config Key | Default Model | Best For |
-|---|---|---|---|
+|----------|------------|---------------|----------|
 | **Anthropic** | `anthropic` | `claude-sonnet-4-20250514` | Code review, refactoring, debugging |
 | **OpenAI** | `openai` | `gpt-4-turbo` | Documentation, explanations |
 | **Ollama** (local) | `ollama` | `llama3` | Privacy, offline use |
@@ -228,7 +229,7 @@ GROQ_API_KEY=gsk-...
 Winter provides **26 tools** -- fully compatible with Claude Code's tool interface + extras:
 
 | # | Tool | Description |
-|---|---|---|
+|----|------|-------------|
 | 1 | **Read** | Read files with line numbers |
 | 2 | **Write** | Create new files |
 | 3 | **Edit** | Surgical string replacement (search & replace) |
@@ -272,12 +273,51 @@ Grep tool supports:
 
 ---
 
+## Bundled Resources
+
+Winter ships with **7 integrated resource corpora** (8,578 files total):
+
+| Resource | Files | Size | Description |
+|----------|-------|------|-------------|
+| **codex** | 206 | 116 MB | Codex CLI documentation & examples |
+| **page-agent** | 239 | - | Alibaba GUI agent framework |
+| **awesome-design-md** | 142 | 14 MB | 71 design systems (Airbnb, Apple, BMW, Claude, Coinbase, Cursor, Figma...) |
+| **claude** | 795 | 3 MB | Claude Code prompts & patterns |
+| **agents.md** | 62 | 2 MB | Agent documentation & examples |
+| **karpathy-tools** | 2 | 1.3 MB | Karpathy's coding principles (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) |
+| **ecc** | 7,932 | - | Embedded Context Corpus - searchable knowledge base |
+
+### Resource Commands
+
+```bash
+# Browse karpathy-tools
+/karpathy
+
+# Browse codex resources
+/codex
+
+# Browse claude resources
+/claude
+
+# Browse agents.md
+/agents
+
+# Show all bundled resources
+/resources
+
+# Search ECC (Embedded Context Corpus)
+winter ecc search "error handling"
+winter ecc browse skills/error-handling
+```
+
+---
+
 ## Slash Commands
 
 ### Project & Session
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/project` | Show/set current project |
 | `/pwd` | Show current directory |
 | `/cd <path>` | Change directory |
@@ -288,7 +328,7 @@ Grep tool supports:
 ### Memory & Plans
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/remember <text>` | Store in memory |
 | `/memories` | Show stored memories |
 | `/forget [pattern]` | Clear memories |
@@ -301,7 +341,7 @@ Grep tool supports:
 ### Git & Automation
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/commit` | Auto-generate commit message & commit |
 | `/review` | AI code review on current diff |
 | `/diff` | Preview git diff |
@@ -316,7 +356,7 @@ Grep tool supports:
 ### Tool Shortcuts
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/read <file>` | Quick file read |
 | `/write <file>` | Quick file write (via editor) |
 | `/edit <file>` | Quick file edit |
@@ -329,7 +369,7 @@ Grep tool supports:
 ### Resources & Design
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/codex` | Browse ~/.codex resources |
 | `/claude` | Browse ~/.claude resources |
 | `/karpathy` | Browse karpathy-tools |
@@ -340,6 +380,8 @@ Grep tool supports:
 | `/design add <brand>` | Add design system |
 | `/design list` | List design systems |
 | `/design preview <brand>` | Preview design system |
+| `/ecc search <query>` | Search Embedded Context Corpus |
+| `/ecc browse <path>` | Browse ECC section |
 | `/skill list` | List skills |
 | `/skill enable <name>` | Enable a skill |
 | `/skill create <name>` | Create a custom skill |
@@ -349,7 +391,7 @@ Grep tool supports:
 ### Provider & Config
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/provider [name]` | Show/switch provider |
 | `/providers` | List all providers |
 | `/model [id]` | Show/set model |
@@ -364,7 +406,7 @@ Grep tool supports:
 ### Help & Exit
 
 | Command | Description |
-|---|---|
+|---------|-------------|
 | `/help` or `/?` | Show help |
 | `/exit` or `/quit` | Exit Winter |
 
@@ -415,8 +457,8 @@ Tools with the `mcp__` prefix are automatically routed to the appropriate MCP se
 
 ```
 mcp__vscode__openFile  ->  VS Code extension server
-mcp__github__createPR  ->  GitHub integration server
-mcp__db__query         ->  Database server
+mcp__github__createPR ->  GitHub integration server
+mcp__db__query        ->  Database server
 ```
 
 ---
@@ -443,7 +485,7 @@ winter session list      # List all sessions
 ### Three-Tier Memory System
 
 | Level | Scope | Purpose |
-|---|---|---|
+|-------|-------|---------|
 | **Working Memory** | Current session | Task context, conversation history |
 | **Project Memory** | Project-wide | Project-specific learnings & rules |
 | **Long-term Memory** | Cross-project | Reusable patterns & knowledge |
@@ -457,7 +499,7 @@ Winter supports hot-reloadable skills for specialized tasks.
 ### Built-in Skills
 
 | Skill | Description |
-|---|---|
+|-------|-------------|
 | `coding` | Code analysis, generation, review |
 | `design` | Design system integration |
 | `debug` | Debugging assistance |
@@ -493,6 +535,31 @@ winter plugin remove <name>            # Remove plugin
 
 ---
 
+## Embedded Context Corpus (ECC)
+
+Winter includes a **searchable knowledge base** with 7,932 embedded documents covering:
+
+- Error handling patterns
+- Testing strategies
+- Security best practices
+- Performance optimization
+- And more...
+
+### ECC Commands
+
+```bash
+# Search the corpus
+winter ecc search "error handling"
+winter ecc search "async await best practices"
+
+# Browse sections
+winter ecc browse skills
+winter ecc browse skills/error-handling
+winter ecc browse testing/patterns
+```
+
+---
+
 ## Architecture
 
 ```
@@ -502,52 +569,69 @@ winter/
 |- src/
 |   |- agent/
 |   |   - swe-agent.js       # SWE-agent integration
+|   |   - runtime.js         # Agent runtime
+|   |   - agent-definitions.js
 |   |- ai/
 |   |   - providers.js       # AI provider management + streaming
+|   |   - orchestrator.js   # AI orchestration
+|   |   - reasoning.js       # Reasoning patterns
+|   |   - small-model-amplifier.js
+|   |   - workflow-selector.js
 |   |- cache/
-|   |   - system.js          # Caching layer
+|   |   - system.js         # Caching layer
+|   |   - embeddings.js     # Embedding cache
 |   |- cli/
-|   |   - commands.js        # CLI command parser (50+ commands)
-|   |   - config.js          # Config loader + secret management
-|   |   - conversation-format.js  # Utility functions
-|   |   - markdown-format.js # Markdown rendering
-|   |   - prompt-builder.js  # System prompt assembly
-|   |   - repl-commands.js   # Slash command handlers (extracted)
-|   |   - repl.js            # Interactive REPL loop
-|   |   - secret-env.js      # Env file loader + redaction
-|   |   - slash-commands.js  # Command definitions
-|   |   - snowflake-logo.js  # Cyberpunk UI branding
-|   |   - spinner.js         # Terminal spinner
-|   |   - terminal-ui.js     # Terminal rendering utilities
+|   |   - commands.js       # CLI command parser (50+ commands)
+|   |   - config.js         # Config loader + secret management
+|   |   - conversation-format.js
+|   |   - markdown-format.js
+|   |   - prompt-builder.js
+|   |   - repl-commands.js
+|   |   - repl.js           # Interactive REPL loop
+|   |   - secret-env.js     # Env file loader + redaction
+|   |   - slash-commands.js
+|   |   - snowflake-logo.js # Cyberpunk UI branding
+|   |   - spinner.js
+|   |   - terminal-ui.js
+|   |   - context-loader.js # Resource loading
+|   |   - ecc.js            # Embedded Context Corpus
+|   |   - codebase-index/  # Fast semantic search
 |   |- context/
-|   |   - compress.js        # Conversation compression
-|   |   - router.js          # Smart provider routing
+|   |   - compress.js       # Conversation compression
+|   |   - router.js         # Smart provider routing
+|   |   - resource-loader.js
 |   |- design/
-|   |   - commands.js        # Design system commands
+|   |   - commands.js       # Design system commands
 |   |- mcp/
-|   |   - client.js          # MCP client
-|   |   - protocol.js        # MCP protocol implementation
+|   |   - client.js         # MCP client
+|   |   - protocol.js       # MCP protocol implementation
 |   |- plugins/
-|   |   - manager.js         # Plugin manager
+|   |   - manager.js        # Plugin manager
 |   |- session/
-|   |   - manager.js         # Session manager
+|   |   - manager.js        # Session manager
 |   |- skills/
-|   |   - manager.js         # Skills manager
+|   |   - manager.js        # Skills manager
 |   - tools/
-|       - executor.js        # Tool executor (25 tools)
-|       - notebook.js        # Jupyter notebook tools
-|       - todo.js            # Todo tools
-|       - scheduler.js       # Schedule wakeup tool
-|       - interactive.js     # Ask user tool
-|       - agent.js           # Sub-agent tool
-|       - insert-text.js     # Insert text tool
+|       - executor.js       # Tool executor (25 tools)
+|       - notebook.js       # Jupyter notebook tools
+|       - todo.js           # Todo tools
+|       - scheduler.js      # Schedule wakeup tool
+|       - interactive.js    # Ask user tool
+|       - agent.js          # Sub-agent tool
+|       - insert-text.js    # Insert text tool
 |       - str-replace-all.js # Batch replace tool
-|       - web-archive.js     # Web archive tool
-|       - permission.js      # Permission manager
-|       - analytics.js       # Tool usage tracking
-|       - retry.js           # Retry utility
+|       - web-archive.js    # Web archive tool
+|       - permission.js     # Permission manager
+|- resources/local/
+|   |- codex/              # Codex CLI resources (206 files)
+|   |- page-agent/         # GUI automation (239 files)
+|   |- awesome-design-md/  # Design systems (142 files)
+|   |- claude/             # Claude patterns (795 files)
+|   |- agents.md/          # Agent docs (62 files)
+|   |- karpathy-tools/     # Coding principles (2 files)
+|   |- ecc/                # Embedded Context Corpus (7,932 files)
 |- README.md
-|- WINTER.md                  # Project rules for AI
+|- WINTER.md               # Project rules for AI
 |- package.json
 ```
 
@@ -555,17 +639,17 @@ winter/
 
 ## Core Philosophy
 
-Winter operates on three core principles:
+Winter operates on **four core principles**:
 
 ### 1. Think Before Coding
-*Nghi truoc khi code*
+*Nghi trước khi code*
 
 - State assumptions explicitly. If uncertain, ask.
 - Surface tradeoffs and alternatives -- don't pick silently.
 - If a simpler approach exists, say so. Push back when warranted.
 
 ### 2. Simplicity First
-*Don gian la tren het*
+*Đơn giản là trên hết*
 
 - Minimum code that solves the problem. Nothing speculative.
 - No features beyond what was asked.
@@ -573,13 +657,22 @@ Winter operates on three core principles:
 - No error handling for impossible scenarios.
 
 ### 3. Surgical Changes
-*Sua doi chinh xac*
+*Sửa đổi chính xác*
 
 - Touch only what you must. Clean up only your own mess.
 - Don't "improve" adjacent code, comments, or formatting.
 - Don't refactor things that aren't broken.
 - Match existing style, even if you'd do it differently.
 
+### 4. Goal-Driven Execution
+*Thực thi theo mục tiêu*
+
+- Define success criteria clearly.
+- Loop until verified -- don't claim completion without tool results.
+- Verify each step with the closest command available.
+
+> *These principles are integrated from [Karpathy Tools](resources/local/karpathy-tools/CLAUDE.md) and applied throughout Winter's operation.*
+
 ---
 
 ## Streaming & Real-time
@@ -600,7 +693,7 @@ User input -> REPL -> collectAssistantStream()
 ## Cross-Platform Support
 
 | Platform | Shell Support | Status |
-|---|---|---|
+|----------|----------------|--------|
 | **Windows** | cmd, PowerShell, Git Bash | Full support |
 | **macOS** | zsh, bash | Full support |
 | **Linux** | bash, sh, zsh | Full support |
@@ -611,21 +704,6 @@ Winter auto-detects your platform and shell:
 
 ---
 
-## Testing
-
-```bash
-# Run all tests (213)
-npm test
-
-# Run specific test file
-node --test src/tools/executor.test.js
-
-# Run tests with watch mode
-node --test --watch src/**/*.test.js
-```
-
----
-
 ## License
 
 MIT License -- see [LICENSE](LICENSE) for details.
@@ -638,4 +716,4 @@ MIT License -- see [LICENSE](LICENSE) for details.
 
 ---
 
-*Built with care -- Winter is the AI assistant that never sleeps.*
+*Built with care -- Winter is the AI assistant that never sleeps.*

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "winter-super-cli",
-  "version": "2026.6.11",
+  "version": "2026.6.13",
   "description": "❄️ AI-Powered Development CLI with Interactive REPL",
   "type": "module",
   "main": "bin/winter.js",

package/src/agent/runtime.js

@@ -9,7 +9,7 @@ export class AgentRuntime {
     this.repl = repl;
   }
 
-  async runConversation(messages, label = 'Thinking', tools = null) {
+  async runConversation(messages, label = 'Thinking... (Đang suy nghĩ, tìm cách giải quyết)', tools = null) {
     const repl = this.repl;
     repl.spinner = new Spinner(label + '...');
     repl.spinner.start();
@@ -46,12 +46,17 @@ export class AgentRuntime {
     try {
       for (let i = 0; i < maxToolTurns; i++) {
         if (repl.isCancelled) throw new Error('AbortError');
+        if (repl.spinner) {
+          repl.spinner.update(`${label}...`);
+          repl.spinner.start();
+        }
         const turn = await repl.requestAssistantTurn(messages, {
           provider: executionProfile.provider,
           model: executionProfile.model,
           enableTools: true,
           toolPromptOnly: forceTextToolFallback,
           requireToolEvidence: requireToolEvidence && !usedTools,
+          usedMutatingTools: usedMutatingTools,
         }, startedAt, totalUsage);
 
         const assistantMsg = turn.assistantMsg || {};
@@ -64,12 +69,15 @@ export class AgentRuntime {
         if (toolCalls.length === 0) {
           if (turn.finishReason === 'tool_evidence_required') {
             noToolActionRetries++;
-            if (noToolActionRetries > 2) {
-              finalContent = 'Chưa thực hiện được: model trả lời mà không dùng tool nên Winter đã chặn để tránh báo xạo.';
+            if (noToolActionRetries > 3) {
+              finalContent = 'Chưa thực hiện được: model trả lời mà không dùng tool nên Winter đã chặn để tránh báo xạo. Hãy thử lại hoặc dùng model mạnh hơn.';
               console.log(`\n${colors.yellow}${finalContent}${colors.reset}\n`);
               reachedToolLimit = false;
               break;
             }
+            if (noToolActionRetries >= 2) {
+              console.log(`\n${colors.yellow}! Model không chịu dùng tool (lần ${noToolActionRetries}/3). Đang ép buộc lại...${colors.reset}`);
+            }
             messages.push({
               role: 'assistant',
               content: assistantMsg.content || '',
@@ -177,9 +185,14 @@ export class AgentRuntime {
           } else if (!proceed) {
             result = { success: false, error: 'User denied permission to execute this command.' };
           } else {
+            if (repl.spinner) {
+              repl.spinner.update(`Executing ${canonicalToolName}... (Đang chạy lệnh)`);
+              repl.spinner.start();
+            }
             result = toolName
               ? await repl.tools.execute(canonicalToolName, enrichedArgs, { cwd: repl.projectPath })
               : { success: false, error: 'Tool call is missing a tool name' };
+            if (repl.spinner) repl.spinner.stop();
           }
           const promptToolResult = await repl.buildPromptToolResultForModel(canonicalToolName, result);
           messages.push({

package/src/cli/prompt-builder.js

@@ -90,6 +90,7 @@ export class PromptBuilder {
         `Debug: reproduce or locate the failing path first, read the exact failing file/log, patch the smallest cause, then run the closest test/build/smoke command.`,
         `Design/UI: inspect existing UI and design resources first; deliver polished, responsive, non-generic interfaces, not placeholder layouts.`,
         `Images: if the user attaches or pastes an image, analyze it directly and connect findings to project files when relevant.`,
+        `CRITICAL: You MUST call tools (Read/Write/Edit/Bash) to do real work. NEVER write code in markdown and claim done. Winter blocks fake completions.`,
         `Tool fallback when native calls are unavailable: <invoke name="Read"><parameter name="path">README.md</parameter></invoke> OR {"tool":"Read","arguments":{"path":"README.md"}} OR CALL_TOOL Read {"path":"README.md"}.`,
         `Session: cwd=${this.projectPath}; id=${this.session?.getSessionId?.()?.substring(0, 8) || 'unknown'}`,
         `${requiredResourcesStr}${memoryStr}${plansStr}${skillsStr}${workflowStr}${blueprintStr}${startupPlanStr}${sessionSignalsStr}`,
@@ -114,6 +115,7 @@ export class PromptBuilder {
       `## Tool Usage`,
       `Use tools when they materially help. For coding tasks: inspect first, edit second, verify third.`,
       `Prefer Read/Grep/Glob before editing. Use Write/Edit for file changes.`,
+      `CRITICAL: When the user asks you to fix/create/edit/run/modify anything, you MUST call tools (Read, Write, Edit, Bash, etc.) to actually do it. NEVER just write code in a markdown code block and claim it is done. Winter will detect and block fake completions. If you say "đã sửa/đã tạo/done/fixed" without a tool call, your response will be rejected.`,
       `Tool call compatibility: if native tool calls are unavailable, output exactly one of these forms and no prose: <invoke name="Read"><parameter name="path">README.md</parameter></invoke> OR {"tool":"Read","arguments":{"path":"README.md"}} OR CALL_TOOL Read {"path":"README.md"}.`,
       `Browser capability: You CAN browse URLs! Use WebFetch to fetch page content (text extraction) or BrowserDebug for Chrome automation (JS rendering, screenshots). If user shares a URL or asks to view a website, use these tools automatically.`,
       `When a task touches coding, agents, UI, brand, or design, inspect the relevant required local resource in depth before deciding.`,

package/src/cli/repl-commands.js

@@ -3,6 +3,8 @@ import { promises as fs } from 'fs';
 import { colors } from './snowflake-logo.js';
 import { SLASH_COMMANDS } from './slash-commands.js';
 
+import { DesignCommands } from '../design/commands.js';
+
 function getPageAgentRoot(repl) {
   return repl.contextLoader?.getResourcePaths?.()?.pageAgent || path.join(repl.projectPath, 'resources', 'local', 'page-agent');
 }
@@ -735,11 +737,11 @@ export async function handleSlashCommand(repl, input) {
       break;
 
     // Design system
-    case '/design':
-      console.log(`${colors.yellow}Design commands:${colors.reset}`);
-      console.log(`  /design add <name> — Add a design system`);
-      console.log(`  /design search <query> — Search design systems`);
+    case '/design': {
+      const designCmd = new DesignCommands(repl);
+      await designCmd.execute(args[0], args.slice(1));
       break;
+    }
     case '/designs':
       await repl.showDesignSystems();
       break;

package/src/cli/repl.js

@@ -1706,11 +1706,19 @@ ${colors.reset}
     const text = this.getLatestUserText(messages).toLowerCase();
     if (!text.trim()) return false;
 
-    const actionPattern = /\b(fix|repair|bug|debug|implement|create|write|edit|modify|update|delete|remove|refactor|run|test|build|commit|push|publish|install|check|inspect|read|scan|grep|search|change|apply|patch|sua|lam|tao|ghi|doc|xoa|chay|kiem tra|cai|them|doi|review|tim|sửa|làm|tạo|đọc|xóa|xoá|chạy|kiểm tra|cài|thêm|đổi|tìm)\b/i;
-    const targetPattern = /\b(file|repo|project|code|src|test|build|git|npm|node|folder|directory|cli|tool|provider|model|config|readme|package\.json|du an|thu muc|tap tin|loi|chuc nang|dự án|thư mục|tập tin|lỗi|chức năng)\b|[A-Za-z]:[\\/]|\.js\b|\.ts\b|\.tsx\b|\.json\b|\.md\b/i;
-    const pureQuestionPattern = /^(what|why|how|when|where|is|are|can|could|should|would|tai sao|vi sao|la gi|co nen|co phai|tại sao|vì sao|là gì|có nên|có phải)\b/i;
+    // Direct action verbs (EN + VN)
+    const actionPattern = /\b(fix|repair|bug|debug|implement|create|write|edit|modify|update|delete|remove|refactor|run|test|build|commit|push|publish|install|check|inspect|read|scan|grep|search|change|apply|patch|add|move|rename|copy|migrate|deploy|setup|configure|generate|format|clean|reset|revert|undo|merge|split|extract|inject|insert|replace|append|prepend|convert|transform|compile|execute|verify|validate|optimize|improve|enhance|upgrade|downgrade|enable|disable|toggle|set|connect|send|fetch|download|upload|open|close|start|stop|restart|sua|lam|tao|ghi|doc|xoa|chay|kiem tra|cai|them|doi|review|tim|viet|trien khai|cap nhat|xay dung|cau hinh|chinh|bo sung|loai bo|sửa|làm|tạo|đọc|xóa|xoá|chạy|kiểm tra|cài|thêm|đổi|tìm|viết|triển khai|cập nhật|xây dựng|cấu hình|chỉnh|bổ sung|loại bỏ|di chuyển|đổi tên|nâng cấp|tối ưu|kết nối|gửi|tải|mở|đóng|khởi động|dừng)\b/i;
+    // Target patterns (filenames, paths, project terms)
+    const targetPattern = /\b(file|repo|project|code|src|test|build|git|npm|node|folder|directory|cli|tool|provider|model|config|readme|package\.json|component|module|class|function|api|endpoint|route|page|style|css|html|template|schema|database|server|client|app|service|hook|context|store|reducer|middleware|controller|handler|view|layout|du an|thu muc|tap tin|loi|chuc nang|dự án|thư mục|tập tin|lỗi|chức năng|giao diện|trang|dịch vụ|thành phần|mô hình)\b|[A-Za-z]:[\\/]|\.jsx?\b|\.tsx?\b|\.json\b|\.md\b|\.py\b|\.css\b|\.html\b|\.vue\b|\.svelte\b|\.rs\b|\.go\b|\.java\b|\.c(pp)?\b|\.rb\b/i;
+    // Pure questions that don't need tool action
+    const pureQuestionPattern = /^(what|why|how|when|where|is|are|can|could|should|would|explain|describe|tell me|compare|giải thích|mô tả|so sánh|tai sao|vi sao|la gi|co nen|co phai|tại sao|vì sao|là gì|có nên|có phải|nhu the nao|như thế nào|khi nào)\b/i;
 
     if (pureQuestionPattern.test(text) && !actionPattern.test(text)) return false;
+    
+    // Even without explicit target, some verbs are strong enough on their own
+    const strongActionAlone = /\b(fix|debug|deploy|build|test|commit|install|run|refactor|sửa|chạy|cài|triển khai|xây dựng)\b/i;
+    if (strongActionAlone.test(text)) return true;
+    
     return actionPattern.test(text) && targetPattern.test(text);
   }
 
@@ -1728,25 +1736,109 @@ ${colors.reset}
     const text = String(content || '').toLowerCase();
     if (!text.trim()) return false;
 
-    const clarification = /(?:cần thêm|cho mình|vui lòng|please provide|which file|what file|need more|clarify|không rõ|chưa rõ|file nào|thư mục nào|c?n th?m|cho m?nh|vui l?ng|kh?ng r?|ch?a r?|file n?o|th? m?c n?o)/i;
+    // If model is asking for clarification, that's legitimate - no tool needed
+    const clarification = /(?:cần thêm thông tin|cho mình biết|vui lòng cung cấp|please provide|which file|what file|need more info|clarify|không rõ|chưa rõ|file nào|thư mục nào|bạn muốn|you want me to|could you specify|can you tell)/i;
     if (clarification.test(text)) return false;
+
     return true;
   }
 
+  detectFakeCompletion(content = '') {
+    const text = String(content || '').toLowerCase();
+    if (!text.trim()) return false;
+
+    // Detect fake completion claims - model says it did something without using tools
+    const fakeCompletionClaims = /(?:đã (?:sửa|tạo|viết|xóa|cập nhật|thêm|chỉnh|xong|hoàn thành|fix|update|edit|write|create|delete|remove|modify|change|apply|deploy|push)|i(?:'ve| have) (?:fixed|created|written|updated|added|modified|changed|edited|applied|deployed|deleted|removed|patched|implemented|refactored)|done!|xong rồi|hoàn thành|đã hoàn tất|hoàn tất|the (?:fix|change|update|edit|modification) (?:has been|is) (?:applied|done|completed|made)|here(?:'s| is) the (?:fix|update|change|solution|implementation|code)|file (?:has been|was) (?:updated|created|modified|written|changed)|changes? (?:have been|has been|were) (?:made|applied|saved)|successfully (?:updated|created|modified|fixed|applied|changed|written))/i;
+    if (fakeCompletionClaims.test(text)) return true;
+
+    // Detect code blocks that pretend to show "changes" without tool use
+    const codeBlockWithFilePath = /```[\s\S]*?(?:[\/\\][\w.-]+\.(?:js|ts|py|css|html|json|md|jsx|tsx|vue|go|rs|java|c|cpp|rb|sh))[\s\S]*?```/i;
+    const claimsFileChange = /(?:here(?:'s| is)|below|sau đây|dưới đây|như sau|updated|modified|changed|new|fixed)/i;
+    if (codeBlockWithFilePath.test(text) && claimsFileChange.test(text)) return true;
+
+    return false;
+  }
+
   buildToolEvidenceCorrection(messages = []) {
     const request = this.getLatestUserText(messages);
     return [
-      'Runtime correction: the user requested an action that requires tool evidence.',
-      'Your previous response did not use any tool, so it was blocked to avoid falsely claiming completion.',
-      'Now use the available tools to inspect/edit/run/check as needed. Do not say the task is done until a tool result proves it.',
-      'If native tool calls are not supported by this model/provider, output exactly one fallback tool call and no prose, for example:',
+      '⚠️ RUNTIME ENFORCEMENT: Your previous response was BLOCKED because you did not use any tool.',
+      '',
+      'The user requested an action that requires REAL tool execution. You MUST:',
+      '1. Call Read/Grep/Glob to inspect the relevant files FIRST',
+      '2. Call Write/Edit to make changes',
+      '3. Call Bash to run/test/verify',
+      '',
+      'DO NOT write code in a code block and claim it is done. That is a hallucination.',
+      'DO NOT say "I have updated/created/fixed" without a tool call proving it.',
+      'DO NOT describe what you would do. Actually DO IT with tool calls.',
+      '',
+      'Available tools: Read, Write, Edit, Bash, Glob, Grep, BrowserDebug, WebFetch, WebSearch.',
+      '',
+      'If native tool calls are not supported, output exactly one fallback tool call:',
       '<invoke name="Read"><parameter name="path">README.md</parameter></invoke>',
       '{"tool":"Read","arguments":{"path":"README.md"}}',
       'CALL_TOOL Read {"path":"README.md"}',
+      '',
       `Original user request: ${request}`,
     ].join('\n');
   }
 
+  /**
+   * Smart Tool Routing: phân tích câu lệnh user và gợi ý tool phù hợp.
+   * Giúp model yếu/nhỏ chọn đúng tool thay vì dùng Bash cho mọi thứ.
+   */
+  buildToolRoutingHint(userMessage = '') {
+    const text = String(userMessage || '').toLowerCase();
+    if (!text.trim()) return null;
+
+    const hints = [];
+
+    // Detect file reading requests
+    const hasPath = /[A-Za-z]:[\\/][\w.\\/\\-]+/i.test(text) || /(?:^|\s)[.~]?\/[\w.\/-]+/i.test(text);
+    const readVerbs = /\b(đọc|doc|read|xem|view|mở|open|show|hiện|hiển thị|cat|type)\b/i;
+    const readFilePatterns = /\b(đọc|doc|read|xem|view|mở|open|show|hiện|cat|type)\b.*\.(?:js|ts|py|json|md|css|html|txt|yaml|yml|toml|cfg|ini|env|sh|bat|ps1|xml|vue|svelte|go|rs|java|c|cpp|rb|php)\b/i;
+    const dirPatterns = /\b(đọc|doc|read|xem|liệt kê|list|ls|dir|show|hiện)\b.*\b(thư mục|folder|directory|dir)\b/i;
+
+    if (readFilePatterns.test(text)) {
+      hints.push('TOOL HINT: To read a file, call tool Read with {"file_path": "<path>"}. Do NOT use Bash with cat/type.');
+    } else if (hasPath && readVerbs.test(text)) {
+      // Path detected + read verb, could be file or directory
+      hints.push('TOOL HINT: To read a file, call tool Read with {"file_path": "<path>"}. To list a directory, call Glob with {"pattern": "*", "cwd": "<path>"}. Do NOT use Bash with ls/dir/cat/type.');
+    } else if (dirPatterns.test(text)) {
+      hints.push('TOOL HINT: To list directory contents, call tool Glob with {"pattern": "*", "cwd": "<path>"}. Do NOT use Bash with ls/dir.');
+    }
+
+    // Detect file writing/creating requests
+    if (/\b(tạo|tao|create|viết|viet|write|ghi)\b.*\b(file|tập tin|tap tin)\b/i.test(text)) {
+      hints.push('TOOL HINT: To create/write a file, call tool Write with {"file_path": "<path>", "content": "<content>"}. Do NOT use Bash with echo/cat.');
+    }
+
+    // Detect file editing requests
+    if (/\b(sửa|sua|edit|chỉnh|chinh|thay|replace|đổi|doi|modify|update|cập nhật)\b.*(?:file|\.(?:js|ts|py|json|md|css|html)\b)/i.test(text)) {
+      hints.push('TOOL HINT: To edit a file, first Read it, then call Edit with {"file_path": "<path>", "old_string": "<exact text>", "new_string": "<replacement>"}.');
+    }
+
+    // Detect search/find requests
+    if (/\b(tìm|tim|find|search|kiếm|kiem|grep|ở đâu|o dau|where)\b/i.test(text) && !(/\b(web|google|online|internet)\b/i.test(text))) {
+      hints.push('TOOL HINT: To search in code, call tool Grep with {"pattern": "<search term>", "path": "<directory>"}. Do NOT use Bash with grep/findstr.');
+    }
+
+    // Detect test/run requests
+    if (/\b(chạy|chay|run|test|execute|thực thi|thuc thi|build|compile|npm|node|python|pip)\b/i.test(text) && !readPatterns.test(text)) {
+      hints.push('TOOL HINT: To run commands, call tool Bash with {"command": "<command>"}.');
+    }
+
+    // Detect URL/web requests  
+    if (/\b(https?:\/\/[^\s]+|url|website|trang web|web page)\b/i.test(text)) {
+      hints.push('TOOL HINT: To fetch a URL, call tool WebFetch with {"url": "<url>"}. For browser debugging, use BrowserDebug.');
+    }
+
+    if (hints.length === 0) return null;
+
+    return `[Tool Router] ${hints.join(' | ')}`;
+  }
+
   withCurrentAbortSignal(options = {}) {
     const signal = options.signal || options.abortSignal || this.currentAbortController?.signal;
     return signal ? { ...options, signal } : options;
@@ -1810,8 +1902,11 @@ ${colors.reset}
     }
     const finishReason = response.choices?.[0]?.finish_reason;
 
+    if (this.spinner) this.spinner.stop();
+
     if (assistantMsg.content && toolCalls.length === 0) {
-      if (options?.requireToolEvidence && this.responseNeedsToolEvidence(assistantMsg.content)) {
+      const isFakeCompletion = !options?.usedMutatingTools && this.detectFakeCompletion(assistantMsg.content);
+      if ((options?.requireToolEvidence && this.responseNeedsToolEvidence(assistantMsg.content)) || isFakeCompletion) {
         return { assistantMsg, toolCalls, finalContent: '', finishReason: 'tool_evidence_required' };
       }
       this.printAssistantAnswer(assistantMsg.content, startedAt, totalUsage);
@@ -1823,6 +1918,8 @@ ${colors.reset}
 
   async collectAssistantStream(messages, options, startedAt, totalUsage) {
     let content = '';
+    let streamBuffer = '';
+    let printedLines = 0;
     const toolCallParts = [];
     let finishReason = null;
     let printed = false;
@@ -1878,6 +1975,18 @@ ${colors.reset}
 
       if (chunk.content) {
         content += chunk.content;
+        if (bufferToolModeContent && this.spinner && process.env.NODE_ENV !== 'test') {
+          streamBuffer += chunk.content;
+          let newlineIdx;
+          const terminalCols = process.stdout.columns || 80;
+          while ((newlineIdx = streamBuffer.indexOf('\n')) !== -1) {
+            const line = streamBuffer.slice(0, newlineIdx).replace(/\r$/, '');
+            streamBuffer = streamBuffer.slice(newlineIdx + 1);
+            const visibleLength = line.replace(/\x1b\[[0-9;]*m/g, '').length + 2;
+            printedLines += Math.max(1, Math.ceil(visibleLength / terminalCols));
+            process.stdout.write(`\r\x1b[K${colors.dim}│ ${colors.reset}${colors.dim}${line}${colors.reset}\n`);
+          }
+        }
       }
 
       if (this.spinner && printed === false) {
@@ -1887,16 +1996,20 @@ ${colors.reset}
             const args = lastCall.function.arguments || '';
             let summary = args.replace(/\s+/g, ' ');
             if (summary.length > 60) summary = '...' + summary.slice(-60);
-            this.spinner.update(`Calling ${lastCall.function.name}... ${summary}`);
+            this.spinner.update(`Calling ${lastCall.function.name}... (Chuẩn bị gọi tool) ${summary}`);
           }
-        } else if (content.length > 0 && bufferToolModeContent) {
-          let summary = content.replace(/\s+/g, ' ');
-          if (summary.length > 60) summary = '...' + summary.slice(-60);
-          this.spinner.update(`Generating... ${summary}`);
         }
       }
     }
 
+    if (streamBuffer.length > 0 && this.spinner) {
+      const line = streamBuffer.replace(/\r$/, '');
+      const terminalCols = process.stdout.columns || 80;
+      const visibleLength = line.replace(/\x1b\[[0-9;]*m/g, '').length + 2;
+      printedLines += Math.max(1, Math.ceil(visibleLength / terminalCols));
+      process.stdout.write(`\r\x1b[K${colors.dim}│ ${colors.reset}${colors.dim}${line}${colors.reset}\n`);
+    }
+
     if (this.spinner) this.spinner.stop();
 
     const inlineToolExtraction = this.extractInlineToolCalls(content);
@@ -1911,7 +2024,15 @@ ${colors.reset}
     const visibleContent = inlineToolExtraction.content || content;
 
     if (toolCalls.length === 0 && visibleContent) {
-      if (options?.requireToolEvidence && this.responseNeedsToolEvidence(visibleContent)) {
+      if (printedLines > 0 && bufferToolModeContent) {
+        const linesToErase = Math.min(printedLines, (process.stdout.rows || 24) - 2);
+        if (linesToErase > 0) {
+          process.stdout.write(`\r\x1b[K\x1b[${linesToErase}A\x1b[J`);
+        }
+      }
+
+      const isFakeCompletion = !options?.usedMutatingTools && this.detectFakeCompletion(visibleContent);
+      if ((options?.requireToolEvidence && this.responseNeedsToolEvidence(visibleContent)) || isFakeCompletion) {
         return {
           assistantMsg: { content: visibleContent },
           toolCalls,
@@ -2067,6 +2188,11 @@ ${colors.reset}
     ];
 
     try {
+      if (this.spinner) {
+        this.spinner.update('Thinking (final answer)... (Đang viết câu trả lời cuối)');
+        this.spinner.start();
+      }
+
       if (typeof this.ai.streamRequest === 'function') {
         return await this.streamFinalAnswer(finalMessages, startedAt, totalUsage, executionProfile);
       }
@@ -2079,11 +2205,15 @@ ${colors.reset}
       });
       this.addUsage(totalUsage, response.usage);
       const content = response.choices?.[0]?.message?.content || '';
+      
+      if (this.spinner) this.spinner.stop();
+      
       if (content) {
         this.printAssistantAnswer(content, startedAt, totalUsage);
       }
       return content;
     } catch (error) {
+      if (this.spinner) this.spinner.stop();
       if (this.isAbortError(error)) throw new Error('AbortError');
       const fallback = this.buildToolFallbackAnswer(toolSummaries, error.message);
       console.log(`\n${colors.yellow}${fallback}${colors.reset}\n`);
@@ -2110,6 +2240,9 @@ ${colors.reset}
         if (chunk.content) {
           content += chunk.content;
         }
+        if (this.spinner && isFirst === false) {
+          this.spinner.update('Generating...');
+        }
       }
 
       if (this.spinner) this.spinner.stop();
@@ -2450,6 +2583,13 @@ ${colors.reset}
       }
 
       const tools = this.getAgentTools('general');
+
+      // Smart tool routing: inject a hint for weak models
+      const toolHint = this.buildToolRoutingHint(message);
+      if (toolHint) {
+        messages.push({ role: 'system', content: toolHint });
+      }
+
       const { finalContent, usedMutatingTools } = await this.runConversation(messages, 'Thinking', tools);
 
       const allToolCalls = [];

package/src/cli/slash-commands.js

@@ -35,7 +35,7 @@ export const SLASH_COMMANDS = [
   { cmd: '/grep', desc: 'Search files', usage: '/grep <pattern>' },
   { cmd: '/bash', desc: 'Run command', usage: '/bash <command>' },
   { cmd: '/image', desc: 'Analyze image/screenshot or clipboard image', usage: '/image [file] [question]' },
-  { cmd: '/design', desc: 'Design commands', sub: ['search', 'add', 'list', 'preview'] },
+  { cmd: '/design', desc: 'Design commands', sub: ['search', 'add', 'apply', 'list', 'preview'] },
   { cmd: '/designs', desc: 'List/search awesome-design-md systems', usage: '/designs [query]' },
   { cmd: '/skill', desc: 'Skills management', sub: ['list', 'enable', 'create'] },
   { cmd: '/skills', desc: 'List local Winter/Codex/Claude skills' },

package/src/cli/spinner.js

@@ -14,6 +14,11 @@ export class Spinner {
     const reset = this.colors.reset || '';
     const dim = this.colors.dim || '';
     this.startTime = Date.now();
+    this.lastLines = 0;
+    
+    // Make sure we're on a clean line
+    process.stdout.write('\r\x1b[K');
+    
     this.interval = setInterval(() => {
       const elapsed = Math.floor((Date.now() - this.startTime) / 1000);
       let timeStr = '';
@@ -24,25 +29,17 @@ export class Spinner {
       const cols = process.stdout.columns || 80;
       let fullStr = `${this.frames[this.frameIndex]} ${this.text}${timeStr}`;
       
-      // Clean ANSI for length calculation (approximate)
-      const visibleLen = fullStr.replace(/\x1b\[[0-9;]*m/g, '').length;
-      
-      if (visibleLen >= cols - 1) {
-        // Truncate the text part, leave room for timeStr and frame
-        const timeStrLen = timeStr.length;
-        const availableSpaceForText = cols - timeStrLen - 6; 
-        if (availableSpaceForText > 10) {
-          const truncText = this.text.length > availableSpaceForText 
-            ? '...' + this.text.slice(-(availableSpaceForText - 3))
-            : this.text;
-          fullStr = `${this.frames[this.frameIndex]} ${truncText}${timeStr}`;
-        } else {
-          // Terminal is extremely narrow, just truncate everything
-          fullStr = fullStr.slice(0, cols - 4) + '...';
-        }
+      // Clear previous frame
+      if (this.lastLines > 0) {
+        process.stdout.write(`\r\x1b[${this.lastLines}A\x1b[J`);
+      } else {
+        process.stdout.write('\r\x1b[K');
       }
+      
+      const visibleLen = fullStr.replace(/\x1b\[[0-9;]*m/g, '').length;
+      this.lastLines = Math.max(0, Math.ceil(visibleLen / cols) - 1);
 
-      process.stdout.write(`\r\x1b[K${cyan}${fullStr.slice(0, 2)}${reset}${dim}${fullStr.slice(2)}${reset}`);
+      process.stdout.write(`${cyan}${fullStr.slice(0, 2)}${reset}${dim}${fullStr.slice(2)}${reset}`);
       this.frameIndex = (this.frameIndex + 1) % this.frames.length;
     }, 80);
   }
@@ -51,7 +48,16 @@ export class Spinner {
     if (!this.interval) return;
     clearInterval(this.interval);
     this.interval = null;
-    process.stdout.write(`\r\x1b[K${finalText ? `${finalText}\n` : ''}`);
+    
+    if (this.lastLines > 0) {
+      process.stdout.write(`\r\x1b[${this.lastLines}A\x1b[J`);
+    } else {
+      process.stdout.write('\r\x1b[K');
+    }
+    
+    if (finalText) {
+      process.stdout.write(`${finalText}\n`);
+    }
   }
 
   update(text) {

package/src/design/commands.js

@@ -11,9 +11,10 @@ import { colors, statusIcons } from '../cli/snowflake-logo.js';
 const packageRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', '..');
 
 export class DesignCommands {
-  constructor(session, config) {
-    this.session = session;
-    this.config = config;
+  constructor(repl) {
+    this.repl = repl;
+    this.session = repl.session;
+    this.config = repl.config;
     this.brandsDir = path.join(packageRoot, 'resources', 'local', 'awesome-design-md', 'design-md');
   }
 
@@ -26,6 +27,9 @@ export class DesignCommands {
       case 'add':
         await this.addBrand(args[0]);
         break;
+      case 'apply':
+        await this.applyBrand(args[0]);
+        break;
       case 'list':
         await this.listBrands();
         break;
@@ -109,6 +113,51 @@ export class DesignCommands {
     }
   }
 
+  async applyBrand(brand) {
+    if (!brand) {
+      console.log(`${colors.yellow}Usage: winter design apply <brand>${colors.reset}`);
+      return;
+    }
+
+    try {
+      const brandDir = path.join(this.brandsDir, brand);
+      let fileContent = null;
+      let fileName = null;
+
+      const designPath = path.join(brandDir, 'DESIGN.md');
+      const readmePath = path.join(brandDir, 'README.md');
+
+      if (await fs.access(designPath).then(() => true).catch(() => false)) {
+        fileContent = await fs.readFile(designPath, 'utf8');
+        fileName = 'DESIGN.md';
+      } else if (await fs.access(readmePath).then(() => true).catch(() => false)) {
+        fileContent = await fs.readFile(readmePath, 'utf8');
+        fileName = 'README.md';
+      }
+
+      if (!fileContent) {
+        console.log(`${colors.red}${statusIcons.error} Brand "${brand}" not found${colors.reset}`);
+        return;
+      }
+
+      console.log(`${colors.cyan}${statusIcons.info} Analyzing and applying ${brand} design system...${colors.reset}`);
+      
+      const prompt = `Please act as a Senior UI/UX Engineer. Analyze the following design system (${brand}) and completely refactor the UI and styles in this project to match its specifications. Focus on colors, typography, border radiuses, interactive states, and overall visual aesthetics as defined in the document.
+
+<design_system>
+${fileContent}
+</design_system>
+
+Start by reviewing the codebase, especially tailwind configs or global css, then rewrite the main components. Create a plan if needed.`;
+      
+      // Inject the task to the AI REPL loop
+      await this.repl.chat(prompt);
+      
+    } catch (error) {
+      console.log(`${colors.red}${statusIcons.error} Error: ${error.message}${colors.reset}`);
+    }
+  }
+
   async listBrands() {
     try {
       const brands = await fs.readdir(this.brandsDir);