Add CodePilot v4: duel toggle, context memory, LeetCode auto-grind

codepilot_v4.py

@@ -0,0 +1,713 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+CodePilot v4 — AI 開發助手 + 自動進化
+======================================
+
+v4 新功能：
+  🔄 /duel on|off  — 雙模型比較開關，開啟後每個問題自動 DPO 配對
+  🧠 上下文記憶     — CODEPILOT.md 專案記憶 + 對話歷史 + 文件快取
+  🏋️ /grind         — LeetCode 自動刷題，無人值守產生訓練數據
+
+Usage:
+    codepilot                                    # 本地模型
+    codepilot --provider openrouter --api-key sk-xxx  # 雲端
+    codepilot --duel --provider openrouter --api-key sk-xxx --adapter ./my-adapter
+    codepilot --grind                            # 自動刷 LeetCode
+    codepilot --grind --provider openrouter --api-key sk-xxx  # 用雲端刷題蒸餾
+"""
+
+import argparse, difflib, json, os, re, shutil, sqlite3, subprocess, sys, torch, time
+from datetime import datetime
+from pathlib import Path
+
+try:
+    import httpx
+except ImportError:
+    httpx = None
+
+DEFAULT_LOCAL_MODEL = "Qwen/Qwen2.5-Coder-3B-Instruct"
+CONFIG_DIR = os.path.expanduser("~/.codepilot")
+DB_PATH = os.path.join(CONFIG_DIR, "feedback.db")
+
+PROVIDER_CONFIGS = {
+    "local": {"name": "Local", "type": "local"},
+    "openai": {"name": "OpenAI", "type": "openai", "base_url": "https://api.openai.com/v1", "default_model": "gpt-4o"},
+    "anthropic": {"name": "Anthropic", "type": "anthropic", "base_url": "https://api.anthropic.com/v1", "default_model": "claude-sonnet-4-20250514"},
+    "openrouter": {"name": "OpenRouter", "type": "openai", "base_url": "https://openrouter.ai/api/v1", "default_model": "anthropic/claude-sonnet-4"},
+    "ollama": {"name": "Ollama", "type": "openai", "base_url": "http://localhost:11434/v1", "default_model": "qwen2.5-coder:3b"},
+}
+
+
+# ============================================================
+#  FEEDBACK DB
+# ============================================================
+class FeedbackDB:
+    def __init__(self):
+        os.makedirs(CONFIG_DIR, exist_ok=True)
+        self.conn = sqlite3.connect(DB_PATH)
+        self.conn.execute("""CREATE TABLE IF NOT EXISTS feedback (
+            id INTEGER PRIMARY KEY, timestamp TEXT, prompt TEXT, completion TEXT,
+            label INTEGER, edited_completion TEXT, project TEXT,
+            source_model TEXT, provider TEXT)""")
+        self.conn.commit()
+
+    def save(self, prompt, completion, label, edited=None, project=None,
+             source_model=None, provider=None):
+        self.conn.execute("INSERT INTO feedback VALUES (NULL,?,?,?,?,?,?,?,?)",
+            (datetime.now().isoformat(), prompt, completion, int(label),
+             edited, project, source_model, provider))
+        self.conn.commit()
+
+    def count(self, provider=None):
+        q = "SELECT COUNT(*), COALESCE(SUM(label),0), SUM(CASE WHEN edited_completion IS NOT NULL THEN 1 ELSE 0 END) FROM feedback"
+        r = self.conn.execute(q + (" WHERE provider=?" if provider else ""), (provider,) if provider else ()).fetchone()
+        return {"total": r[0], "up": int(r[1]), "edits": int(r[2] or 0)}
+
+    def export_sft(self, only_cloud=False):
+        if only_cloud:
+            rows = self.conn.execute("SELECT prompt, completion FROM feedback WHERE label=1 AND provider != 'local' AND provider IS NOT NULL").fetchall()
+        else:
+            rows = self.conn.execute("SELECT prompt, COALESCE(edited_completion, completion) FROM feedback WHERE label=1").fetchall()
+        return [{"messages": [{"role": "user", "content": p}, {"role": "assistant", "content": c}]} for p, c in rows]
+
+    def export_dpo(self):
+        rows = self.conn.execute("""SELECT c.prompt, c.completion, l.completion FROM feedback c
+            JOIN feedback l ON c.prompt = l.prompt WHERE c.provider != 'local' AND c.label = 1
+            AND l.provider = 'local' AND l.label = 0""").fetchall()
+        return [{"prompt": [{"role": "user", "content": p}], "chosen": [{"role": "assistant", "content": c}],
+                 "rejected": [{"role": "assistant", "content": l}]} for p, c, l in rows]
+
+    def export_kto(self):
+        rows = self.conn.execute("SELECT prompt, completion, label FROM feedback").fetchall()
+        return [{"prompt": [{"role": "user", "content": p}], "completion": [{"role": "assistant", "content": c}], "label": bool(l)} for p, c, l in rows]
+
+
+# ============================================================
+#  PROJECT CONTEXT — 記憶系統
+# ============================================================
+class ProjectContext:
+    """專案上下文記憶"""
+
+    def __init__(self, project_dir):
+        self.project_dir = project_dir
+        self.memory_file = os.path.join(project_dir, "CODEPILOT.md")
+        self.session_file = os.path.join(CONFIG_DIR, "sessions",
+            os.path.basename(project_dir) + ".json")
+        os.makedirs(os.path.dirname(self.session_file), exist_ok=True)
+
+    def load_memory(self):
+        """讀取 CODEPILOT.md 專案記憶"""
+        if os.path.exists(self.memory_file):
+            return Path(self.memory_file).read_text(encoding="utf-8")
+        return ""
+
+    def save_memory(self, content):
+        """保存專案記憶"""
+        Path(self.memory_file).write_text(content, encoding="utf-8")
+
+    def load_session(self):
+        """載入上次對話"""
+        if os.path.exists(self.session_file):
+            try:
+                data = json.loads(Path(self.session_file).read_text())
+                # 只保留最近 20 輪對話（防止 context 爆掉）
+                msgs = data.get("messages", [])
+                if len(msgs) > 42:  # system + 20 rounds * 2 + buffer
+                    msgs = [msgs[0]] + msgs[-40:]
+                return msgs
+            except:
+                pass
+        return None
+
+    def save_session(self, messages):
+        """保存當前對話"""
+        # 只保最近 20 輪
+        if len(messages) > 42:
+            messages = [messages[0]] + messages[-40:]
+        Path(self.session_file).write_text(
+            json.dumps({"messages": messages, "timestamp": datetime.now().isoformat()},
+                       ensure_ascii=False))
+
+
+# ============================================================
+#  MODEL BACKENDS
+# ============================================================
+class LocalModel:
+    def __init__(self, model_name=DEFAULT_LOCAL_MODEL, adapter_path=None):
+        from transformers import AutoTokenizer, AutoModelForCausalLM
+        self.name = model_name.split("/")[-1]; self.provider = "local"
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+        if self.tokenizer.pad_token is None: self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
+        if adapter_path and os.path.exists(adapter_path):
+            from peft import PeftModel; self.model = PeftModel.from_pretrained(self.model, adapter_path)
+        self.model.eval()
+
+    def chat(self, messages, max_tokens=4096):
+        text = self.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = self.tokenizer(text, return_tensors="pt").to(self.model.device)
+        with torch.no_grad():
+            out = self.model.generate(**inputs, max_new_tokens=max_tokens, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.1, pad_token_id=self.tokenizer.pad_token_id)
+        return self.tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
+
+
+class CloudModel:
+    def __init__(self, provider_key, api_key, model_name=None):
+        config = PROVIDER_CONFIGS[provider_key]
+        self.provider = provider_key; self.base_url = config["base_url"]
+        self.name = model_name or config["default_model"]; self.api_key = api_key; self.api_type = config["type"]
+
+    def chat(self, messages, max_tokens=4096):
+        if self.api_type == "anthropic": return self._anthropic(messages, max_tokens)
+        else: return self._openai(messages, max_tokens)
+
+    def _openai(self, messages, max_tokens):
+        headers = {"Authorization": f"Bearer {self.api_key}", "Content-Type": "application/json"}
+        if self.provider == "openrouter": headers.update({"HTTP-Referer": "https://codepilot.local", "X-Title": "CodePilot"})
+        resp = httpx.post(f"{self.base_url}/chat/completions", headers=headers,
+            json={"model": self.name, "messages": messages, "max_tokens": max_tokens, "temperature": 0.7}, timeout=120)
+        resp.raise_for_status(); return resp.json()["choices"][0]["message"]["content"]
+
+    def _anthropic(self, messages, max_tokens):
+        system = None; chat_msgs = []
+        for m in messages:
+            if m["role"] == "system": system = m["content"]
+            else: chat_msgs.append(m)
+        data = {"model": self.name, "messages": chat_msgs, "max_tokens": max_tokens, "temperature": 0.7}
+        if system: data["system"] = system
+        resp = httpx.post(f"{self.base_url}/messages", headers={"x-api-key": self.api_key, "Content-Type": "application/json", "anthropic-version": "2023-06-01"}, json=data, timeout=120)
+        resp.raise_for_status(); return resp.json()["content"][0]["text"]
+
+
+# ============================================================
+#  PROJECT TOOLS
+# ============================================================
+class ProjectTools:
+    def __init__(self, project_dir):
+        self.project_dir = os.path.abspath(project_dir); self.cwd = self.project_dir; self.read_cache = {}
+
+    def _resolve(self, path):
+        return path if os.path.isabs(path) else os.path.normpath(os.path.join(self.cwd, path))
+
+    def read_file(self, path, offset=1, limit=200):
+        full = self._resolve(path)
+        if not os.path.exists(full): return f"❌ 不存在: {path}"
+        try:
+            content = Path(full).read_text(encoding="utf-8", errors="replace"); lines = content.splitlines()
+            self.read_cache[full] = {"time": os.path.getmtime(full), "content": content}
+            result = "\n".join(f"{i+offset:4d} │ {line}" for i, line in enumerate(lines[offset-1:offset-1+limit]))
+            if offset + limit < len(lines): result += f"\n... ({len(lines)-offset-limit+1} more)"
+            return result
+        except Exception as e: return f"❌ {e}"
+
+    def edit_file(self, path, old_string, new_string):
+        full = self._resolve(path)
+        if full not in self.read_cache: return "❌ 必須先 read_file"
+        content = Path(full).read_text(encoding="utf-8")
+        if os.path.getmtime(full) != self.read_cache[full]["time"]: return "❌ 文件已被外部修改"
+        count = content.count(old_string)
+        if count == 0: return "❌ 找不到要替換的文字"
+        if count > 1: return f"❌ 找到 {count} 處，請提供更多上下文"
+        new_content = content.replace(old_string, new_string, 1)
+        diff = "".join(difflib.unified_diff(content.splitlines(keepends=True), new_content.splitlines(keepends=True), fromfile=f"a/{path}", tofile=f"b/{path}"))
+        Path(full).write_text(new_content, encoding="utf-8")
+        self.read_cache[full] = {"time": os.path.getmtime(full), "content": new_content}
+        return "✅ 已修改:\n" + diff
+
+    def write_file(self, path, content):
+        full = self._resolve(path); os.makedirs(os.path.dirname(full) or ".", exist_ok=True)
+        is_new = not os.path.exists(full); Path(full).write_text(content, encoding="utf-8")
+        self.read_cache[full] = {"time": os.path.getmtime(full), "content": content}
+        return f"✅ {'建立' if is_new else '覆寫'}: {path}"
+
+    def run_command(self, command, timeout=120):
+        for d in {"rm -rf /", "git push --force", "git reset --hard"}:
+            if d in command: return f"⛔ 危險: {command}"
+        try:
+            r = subprocess.run(command, shell=True, cwd=self.cwd, capture_output=True, text=True, timeout=timeout)
+            return (r.stdout + (f"\nSTDERR:\n{r.stderr}" if r.stderr else ""))[:10000]
+        except subprocess.TimeoutExpired: return "⏰ 超時"
+        except Exception as e: return f"❌ {e}"
+
+    def search_files(self, pattern, glob_pattern=None):
+        rg = shutil.which("rg"); cmd = [rg or "grep", "-rn"]
+        if rg: cmd += ["--color=never", "--max-count=50"]
+        if glob_pattern and rg: cmd += ["--glob", glob_pattern]
+        cmd += [pattern, self.cwd]
+        try: return subprocess.run(cmd, capture_output=True, text=True, timeout=30).stdout[:5000] or "無匹配"
+        except Exception as e: return f"❌ {e}"
+
+    def list_files(self, pattern="*", max_depth=3):
+        files = []
+        for root, dirs, fnames in os.walk(self.cwd):
+            dirs[:] = [d for d in dirs if d not in {".git","node_modules","__pycache__",".venv","dist","build"}]
+            if root.replace(self.cwd, "").count(os.sep) >= max_depth: continue
+            files.extend(os.path.relpath(os.path.join(root, f), self.cwd) for f in fnames if Path(f).match(pattern))
+        return "\n".join(sorted(files)[:100])
+
+    def git_context(self):
+        try:
+            b = subprocess.run(["git","branch","--show-current"], cwd=self.project_dir, capture_output=True, text=True).stdout.strip()
+            s = subprocess.run(["git","status","--short"], cwd=self.project_dir, capture_output=True, text=True).stdout.strip()
+            l = subprocess.run(["git","log","--oneline","-5"], cwd=self.project_dir, capture_output=True, text=True).stdout.strip()
+            return f"Branch: {b}\nStatus:\n{s}\nRecent:\n{l}"
+        except: return "(not a git repo)"
+
+
+TOOL_PATTERN = re.compile(r'<tool>\s*(\w+)\s*\n(.*?)</tool>', re.DOTALL)
+
+def parse_tool_calls(text):
+    calls = []
+    for m in TOOL_PATTERN.finditer(text):
+        try: params = json.loads(m.group(2).strip())
+        except:
+            params = {}
+            for line in m.group(2).strip().split("\n"):
+                if ":" in line: k, v = line.split(":", 1); params[k.strip()] = v.strip().strip('"')
+        calls.append({"tool": m.group(1), "params": params})
+    return calls
+
+def execute_tool(tools, call):
+    n, p = call["tool"], call["params"]
+    try:
+        if n == "read_file": return tools.read_file(p.get("path",""), int(p.get("offset",1)), int(p.get("limit",200)))
+        elif n == "edit_file": return tools.edit_file(p.get("path",""), p.get("old_string",""), p.get("new_string",""))
+        elif n == "write_file": return tools.write_file(p.get("path",""), p.get("content",""))
+        elif n == "run_command": return tools.run_command(p.get("command",""), int(p.get("timeout",120)))
+        elif n == "search_files": return tools.search_files(p.get("pattern",""), p.get("glob"))
+        elif n == "list_files": return tools.list_files(p.get("pattern","*"), int(p.get("max_depth",3)))
+        elif n == "git_status": return tools.git_context()
+        else: return f"❌ 未知: {n}"
+    except Exception as e: return f"❌ {e}"
+
+
+def build_system_prompt(tools, project_memory=""):
+    memory_section = f"\n\n## Project Memory (CODEPILOT.md)\n{project_memory}" if project_memory else ""
+    return f"""You are CodePilot, an expert AI programming assistant working in the user's project.
+
+Working directory: {tools.cwd}
+{tools.git_context()}{memory_section}
+
+## Tools (use <tool>name\n{{json}}</tool>)
+- read_file: {{"path":"...","offset":1,"limit":200}}
+- edit_file: {{"path":"...","old_string":"...","new_string":"..."}} (must read first)
+- write_file: {{"path":"...","content":"..."}}
+- run_command: {{"command":"...","timeout":120}}
+- search_files: {{"pattern":"...","glob":"*.py"}}
+- list_files: {{"pattern":"*","max_depth":3}}
+- git_status: {{}}
+
+Rules: read before edit, old_string must be unique, prefer edit over write, verify changes."""
+
+
+# ============================================================
+#  LEETCODE AUTO-GRIND
+# ============================================================
+def run_grind(args, num_problems=100):
+    """自動刷 LeetCode 題目，產生訓練數據"""
+    from rich.console import Console
+    from rich.progress import Progress
+    console = Console()
+    db = FeedbackDB()
+
+    console.print(f"""
+    ╔════════════════════════════════════════════════════════════╗
+    ║  🏋️ LeetCode Auto-Grind                                   ║
+    ║  自動刷題，無人值守產生訓練數據                                ║
+    ╚════════════════════════════════════════════════════════════╝
+    """)
+
+    # 載入模型
+    provider_key = args.provider or "local"
+    if provider_key == "local":
+        with console.status("[bold green]載入本地模型..."):
+            model = LocalModel(args.model or DEFAULT_LOCAL_MODEL, args.adapter)
+    else:
+        if not args.api_key:
+            console.print("[red]❌ 需要 --api-key[/]"); return
+        cloud_model_name = args.cloud_model or PROVIDER_CONFIGS[provider_key]["default_model"]
+        model = CloudModel(provider_key, args.api_key, cloud_model_name)
+    console.print(f"[green]✅ 模型: {model.name}[/]")
+
+    # 載入 KodCode 題目
+    console.print("📦 載入 KodCode 題庫...")
+    from datasets import load_dataset
+    dataset = load_dataset("KodCode/KodCode-V1", split="train")
+    dataset = dataset.shuffle(seed=int(time.time()) % 10000).select(range(min(num_problems, len(dataset))))
+    console.print(f"   {len(dataset)} 題已載入\n")
+
+    passed = 0
+    failed = 0
+    errors = 0
+
+    with Progress() as progress:
+        task = progress.add_task("[cyan]刷題中...", total=len(dataset))
+
+        for i, problem in enumerate(dataset):
+            question = problem["question"]
+            test_code = problem["test"]
+            solution_ref = problem["solution"]
+
+            prompt = f"Write a Python solution. Provide ONLY the code, no explanation.\n\n{question}"
+            messages = [
+                {"role": "system", "content": "You are an expert Python programmer. Output only clean Python code."},
+                {"role": "user", "content": prompt},
+            ]
+
+            # 生成回答
+            try:
+                response = model.chat(messages, max_tokens=1024)
+            except Exception as e:
+                errors += 1; progress.update(task, advance=1); continue
+
+            # 提取 code
+            code = response
+            if "```python" in code: code = code.split("```python")[1].split("```")[0]
+            elif "```" in code: code = code.split("```")[1].split("```")[0]
+
+            # 執行測試
+            reward = 0.0
+            try:
+                import tempfile
+                with tempfile.TemporaryDirectory() as tmpdir:
+                    Path(os.path.join(tmpdir, "solution.py")).write_text(code)
+                    Path(os.path.join(tmpdir, "test_solution.py")).write_text(test_code)
+                    r = subprocess.run(
+                        [sys.executable, "-m", "pytest", "test_solution.py", "-x", "--tb=no", "-q"],
+                        cwd=tmpdir, capture_output=True, text=True, timeout=15)
+                    if r.returncode == 0:
+                        reward = 1.0; passed += 1
+                    else:
+                        reward = 0.0; failed += 1
+            except:
+                reward = 0.0; failed += 1
+
+            # 記錄數據
+            if reward == 1.0:
+                # 通過測試 → 記為好答案 (SFT + KTO positive)
+                db.save(prompt, code, 1, source_model=model.name,
+                        provider=getattr(model, "provider", provider_key))
+            else:
+                # 失敗 → 記為壞答案，同時記錄正確答案
+                db.save(prompt, code, 0, source_model=model.name,
+                        provider=getattr(model, "provider", provider_key))
+                # 正確答案記為 SFT
+                if solution_ref:
+                    db.save(prompt, solution_ref, 1, source_model="ground_truth",
+                            provider="reference")
+
+            progress.update(task, advance=1,
+                description=f"[cyan]刷題中... ✅{passed} ❌{failed}")
+
+    # 統計
+    total = passed + failed + errors
+    console.print(f"\n{'='*50}")
+    console.print(f"  🏋️ 刷題完成！")
+    console.print(f"  ✅ 通過: {passed}/{total} ({100*passed/max(total,1):.0f}%)")
+    console.print(f"  ❌ 失敗: {failed}/{total}")
+    console.print(f"  ⚠️  錯誤: {errors}")
+    console.print(f"\n  📊 數據統計:")
+    s = db.count()
+    console.print(f"     總數據: {s['total']}")
+    console.print(f"     👍: {s['up']} / 👎: {s['total']-s['up']}")
+    console.print(f"\n  💡 運行 codepilot --train 開始訓練")
+
+
+# ============================================================
+#  MAIN AGENT LOOP
+# ============================================================
+def run_agent_loop(args):
+    from rich.console import Console, Group
+    from rich.markdown import Markdown
+    from rich.panel import Panel
+    from rich.prompt import Prompt
+    from rich.syntax import Syntax
+    from rich.table import Table
+
+    console = Console(); db = FeedbackDB()
+    project_dir = args.project or os.getcwd()
+    tools = ProjectTools(project_dir)
+    ctx = ProjectContext(project_dir)
+    provider_key = args.provider or "local"
+
+    # 載入模型
+    local_model_ref = None; cloud_model_ref = None
+    if provider_key == "local":
+        with console.status("[bold green]載入本地模型..."):
+            model = LocalModel(args.model or DEFAULT_LOCAL_MODEL, args.adapter)
+        local_model_ref = model
+    else:
+        if not args.api_key: console.print(f"[red]❌ 需要 --api-key[/]"); sys.exit(1)
+        model = CloudModel(provider_key, args.api_key, args.cloud_model or PROVIDER_CONFIGS[provider_key]["default_model"])
+        cloud_model_ref = model
+        if args.adapter:
+            try:
+                with console.status("[dim]載入本地模型 (for duel)..."):
+                    local_model_ref = LocalModel(args.model or DEFAULT_LOCAL_MODEL, args.adapter)
+                console.print("[dim]✅ 本地模型已載入[/]")
+            except: pass
+
+    # Duel 模式開關
+    duel_mode = args.duel and local_model_ref and cloud_model_ref
+
+    # 專案記憶
+    project_memory = ctx.load_memory()
+
+    # Banner
+    banner = f"[bold cyan]CodePilot v4[/]"
+    if duel_mode: banner += " [bold yellow]⚔️ Duel ON[/]"
+    banner += f"\n[dim]Model: {model.name}\nProject: {project_dir}[/]"
+    if project_memory: banner += f"\n[dim]📝 CODEPILOT.md loaded ({len(project_memory)} chars)[/]"
+    console.print(Panel.fit(banner, border_style="cyan"))
+
+    git_ctx = tools.git_context()
+    if git_ctx != "(not a git repo)": console.print(Panel(git_ctx, title="📂 Project", border_style="dim"))
+
+    # 嘗試恢復上次對話
+    system_prompt = build_system_prompt(tools, project_memory)
+    prev_session = ctx.load_session()
+    if prev_session and len(prev_session) > 1:
+        messages = prev_session
+        # 更新 system prompt
+        messages[0] = {"role": "system", "content": system_prompt}
+        console.print(f"[dim]🔄 已恢復上次對話 ({(len(messages)-1)//2} 輪)[/]")
+    else:
+        messages = [{"role": "system", "content": system_prompt}]
+
+    console.print("[dim]/duel on|off  /memo  /grind  /ls  /git  /clear  /status  /train  /quit[/]\n")
+
+    while True:
+        try: user_input = Prompt.ask("\n[bold green]🧑 You")
+        except (EOFError, KeyboardInterrupt): break
+        if not user_input.strip(): continue
+        cmd = user_input.strip()
+
+        # ---- 指令 ----
+        if cmd in ("/quit", "/exit"): break
+
+        elif cmd == "/duel on":
+            if local_model_ref and cloud_model_ref:
+                duel_mode = True; console.print("[yellow]⚔️ Duel 模式已開啟 — 每個問題自動雙模型比較[/]")
+            else:
+                console.print("[red]需要同時有本地和雲端模型。啟動: codepilot --duel --provider openrouter --api-key xxx --adapter ./adapter[/]")
+            continue
+
+        elif cmd == "/duel off":
+            duel_mode = False; console.print("[dim]Duel 模式已關閉[/]"); continue
+
+        elif cmd == "/memo":
+            console.print(f"[bold]📝 CODEPILOT.md[/]")
+            console.print("[dim]輸入專案筆記（END 結束），會注入每次對話的 system prompt:[/]")
+            lines = []
+            if project_memory: console.print(f"[dim]目前內容:\n{project_memory[:500]}[/]\n")
+            while True:
+                try:
+                    l = input()
+                    if l.strip() == "END": break
+                    lines.append(l)
+                except EOFError: break
+            if lines:
+                project_memory = "\n".join(lines)
+                ctx.save_memory(project_memory)
+                system_prompt = build_system_prompt(tools, project_memory)
+                messages[0] = {"role": "system", "content": system_prompt}
+                console.print(f"[green]✅ 已保存 CODEPILOT.md ({len(project_memory)} chars)[/]")
+            continue
+
+        elif cmd == "/grind":
+            n = Prompt.ask("刷幾題？", default="50")
+            run_grind(args, int(n)); continue
+
+        elif cmd == "/status":
+            s = db.count()
+            t = Table(title="📊 統計"); t.add_column("", style="cyan"); t.add_column("", style="green")
+            t.add_row("Total", str(s["total"])); t.add_row("👍", str(s["up"]))
+            t.add_row("👎", str(s["total"]-s["up"])); t.add_row("✏️", str(s["edits"]))
+            t.add_row("DPO 對", str(len(db.export_dpo())))
+            t.add_row("Duel", "⚔️ ON" if duel_mode else "OFF")
+            t.add_row("記憶", f"{len(project_memory)} chars" if project_memory else "無")
+            t.add_row("對話輪數", str((len(messages)-1)//2))
+            console.print(t); continue
+
+        elif cmd == "/train": trigger_training(db, console, args); continue
+        elif cmd == "/clear":
+            messages = [{"role": "system", "content": system_prompt}]
+            ctx.save_session(messages); console.print("[dim]已清除[/]"); continue
+        elif cmd == "/git": console.print(Panel(tools.git_context(), title="Git", border_style="dim")); continue
+        elif cmd.startswith("/ls"): console.print(tools.list_files(cmd[3:].strip() or "*")); continue
+        elif cmd == "/switch":
+            new_p = Prompt.ask("切換到", choices=list(PROVIDER_CONFIGS.keys()))
+            if new_p == "local":
+                with console.status("載入..."): model = LocalModel(args.model or DEFAULT_LOCAL_MODEL, args.adapter)
+                local_model_ref = model; provider_key = "local"
+            else:
+                key = args.api_key or Prompt.ask("API Key")
+                cm = Prompt.ask("模型", default=PROVIDER_CONFIGS[new_p]["default_model"])
+                model = CloudModel(new_p, key, cm); cloud_model_ref = model; provider_key = new_p
+            console.print(f"[green]✅ {provider_key}[/]"); continue
+
+        # ---- Duel 模式：自動雙模型比較 ----
+        if duel_mode and local_model_ref and cloud_model_ref:
+            compare_msgs = list(messages) + [{"role": "user", "content": user_input}]
+
+            with console.status("[bold cyan]🏠 本地模型..."):
+                try: local_resp = local_model_ref.chat(compare_msgs)
+                except Exception as e: local_resp = f"(錯誤: {e})"
+
+            with console.status("[bold magenta]☁️ 雲端模型..."):
+                try: cloud_resp = cloud_model_ref.chat(compare_msgs)
+                except Exception as e: cloud_resp = f"(錯誤: {e})"
+
+            console.print(Panel(Markdown(local_resp), title=f"🏠 {local_model_ref.name}", border_style="blue"))
+            console.print(Panel(Markdown(cloud_resp), title=f"☁️ {cloud_model_ref.name}", border_style="magenta"))
+
+            console.print(f"[dim][green]1[/]=🏠本地  [magenta]2[/]=☁️雲端  [yellow]b[/]=都好  [red]x[/]=都差  Enter=跳過[/]")
+            choice = Prompt.ask("  ", choices=["1","2","b","x",""], default="", show_choices=False)
+
+            if choice == "2":
+                db.save(user_input, cloud_resp, 1, source_model=cloud_model_ref.name, provider=cloud_model_ref.provider)
+                db.save(user_input, local_resp, 0, source_model=local_model_ref.name, provider="local")
+                console.print(f"  [magenta]☁️ 雲端勝 → DPO +1 ({len(db.export_dpo())} 對)[/]")
+                messages.append({"role": "user", "content": user_input})
+                messages.append({"role": "assistant", "content": cloud_resp})
+            elif choice == "1":
+                db.save(user_input, local_resp, 1, source_model=local_model_ref.name, provider="local")
+                db.save(user_input, cloud_resp, 0, source_model=cloud_model_ref.name, provider=cloud_model_ref.provider)
+                console.print(f"  [green]🏠 本地勝！[/]")
+                messages.append({"role": "user", "content": user_input})
+                messages.append({"role": "assistant", "content": local_resp})
+            elif choice == "b":
+                db.save(user_input, local_resp, 1, source_model=local_model_ref.name, provider="local")
+                db.save(user_input, cloud_resp, 1, source_model=cloud_model_ref.name, provider=cloud_model_ref.provider)
+                console.print(f"  [yellow]👍 都好[/]")
+                messages.append({"role": "user", "content": user_input})
+                messages.append({"role": "assistant", "content": cloud_resp})
+            elif choice == "x":
+                db.save(user_input, local_resp, 0, source_model=local_model_ref.name, provider="local")
+                db.save(user_input, cloud_resp, 0, source_model=cloud_model_ref.name, provider=cloud_model_ref.provider)
+                console.print(f"  [red]都差[/]")
+            else:
+                messages.append({"role": "user", "content": user_input})
+                messages.append({"role": "assistant", "content": cloud_resp})
+
+            ctx.save_session(messages)
+            continue
+
+        # ---- 正常模式：單模型 + 工具循環 ----
+        messages.append({"role": "user", "content": user_input})
+        full_response = ""
+
+        for rnd in range(10):
+            with console.status(f"[bold cyan]{'思考中' if rnd == 0 else f'工具 round {rnd+1}'}..."):
+                try: response = model.chat(messages)
+                except Exception as e: console.print(f"[red]❌ {e}[/]"); break
+
+            tool_calls = parse_tool_calls(response)
+            text_parts = TOOL_PATTERN.sub("", response).strip()
+            if text_parts:
+                console.print(f"\n[bold blue]🤖 CodePilot:[/]")
+                console.print(Markdown(text_parts))
+            full_response += response + "\n"
+            if not tool_calls: break
+
+            messages.append({"role": "assistant", "content": response})
+            results = []
+            for call in tool_calls:
+                console.print(f"  [dim]🔧 {call['tool']}[/]")
+                result = execute_tool(tools, call)
+                if call["tool"] == "edit_file" and "✅" in result:
+                    d = result.split("\n", 1)[1] if "\n" in result else ""
+                    if d: console.print(Syntax(d, "diff", theme="monokai"))
+                elif call["tool"] == "run_command":
+                    console.print(Panel(result[:500], title="Terminal", border_style="dim"))
+                else: console.print(f"  [dim]{result[:200]}[/]")
+                results.append(f"[{call['tool']}] {result}")
+            messages.append({"role": "user", "content": "Tool results:\n" + "\n\n".join(results)})
+
+        # 回饋
+        console.print(f"\n[dim][green]y[/]=👍 [red]n[/]=👎 [yellow]e[/]=✏️ Enter=跳過[/]")
+        fb = Prompt.ask("  ", choices=["y","n","e",""], default="", show_choices=False)
+        if fb == "y":
+            db.save(user_input, full_response, 1, source_model=getattr(model,"name",""), provider=provider_key)
+            console.print("  [green]👍[/]")
+        elif fb == "n":
+            db.save(user_input, full_response, 0, source_model=getattr(model,"name",""), provider=provider_key)
+            console.print("  [red]👎[/]")
+        elif fb == "e":
+            console.print("  [yellow]貼上修改版(END結束):[/]"); lines = []
+            while True:
+                try:
+                    l = input()
+                    if l.strip() == "END": break
+                    lines.append(l)
+                except EOFError: break
+            edited = "\n".join(lines)
+            if edited.strip():
+                db.save(user_input, full_response, 1, edited=edited, source_model=getattr(model,"name",""), provider=provider_key)
+                console.print("  [yellow]✏️[/]")
+
+        messages.append({"role": "assistant", "content": full_response})
+        ctx.save_session(messages)
+
+    console.print("\n[cyan]👋[/]")
+
+
+# ============================================================
+#  TRAINING
+# ============================================================
+def trigger_training(db, console, args):
+    s = db.count()
+    if s["total"] == 0: console.print("[yellow]⚠️ 無數據[/]"); return
+    cloud_sft = db.export_sft(only_cloud=True); all_sft = db.export_sft(); dpo = db.export_dpo()
+    console.print(f"\n[bold]🚀 數據[/] ⚗️蒸餾SFT:{len(cloud_sft)} 📊DPO:{len(dpo)} 📚全SFT:{len(all_sft)}")
+
+    from datasets import Dataset
+    from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+    from peft import LoraConfig, prepare_model_for_kbit_training
+
+    mn = args.model or DEFAULT_LOCAL_MODEL
+    od = os.path.join(CONFIG_DIR, f"adapter_{datetime.now().strftime('%Y%m%d_%H%M')}")
+    bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
+    pc = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, bias="none", task_type="CAUSAL_LM",
+        target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"])
+    td = cloud_sft or all_sft
+    if td:
+        console.print(f"[bold]📚 {'⚗️蒸餾' if cloud_sft else ''} SFT ({len(td)})...[/]")
+        from trl import SFTTrainer, SFTConfig
+        m = AutoModelForCausalLM.from_pretrained(mn, quantization_config=bnb, device_map="auto", trust_remote_code=True)
+        t = AutoTokenizer.from_pretrained(mn)
+        if t.pad_token is None: t.pad_token = t.eos_token
+        m = prepare_model_for_kbit_training(m)
+        SFTTrainer(model=m, args=SFTConfig(output_dir=od, learning_rate=2e-4, num_train_epochs=3,
+            per_device_train_batch_size=1, gradient_accumulation_steps=8, max_seq_length=1024,
+            gradient_checkpointing=True, bf16=True, optim="paged_adamw_8bit", logging_steps=5,
+            save_total_limit=1, logging_strategy="steps", logging_first_step=True),
+            processing_class=t, train_dataset=Dataset.from_list(td), peft_config=pc).train()
+        m.save_pretrained(od); del m; torch.cuda.empty_cache()
+    console.print(f"\n[bold green]🎉[/] {od}\n   codepilot --adapter {od}")
+
+def show_stats():
+    from rich.console import Console; from rich.table import Table
+    c = Console(); db = FeedbackDB(); s = db.count()
+    t = Table(title="📊 CodePilot"); t.add_column("",style="cyan"); t.add_column("",style="green")
+    t.add_row("Total",str(s["total"])); t.add_row("👍",str(s["up"])); t.add_row("DPO",str(len(db.export_dpo())))
+    c.print(t)
+
+def main():
+    p = argparse.ArgumentParser(description="CodePilot v4")
+    p.add_argument("--model", type=str); p.add_argument("--adapter", type=str)
+    p.add_argument("--project", type=str); p.add_argument("--provider", type=str, choices=list(PROVIDER_CONFIGS.keys()))
+    p.add_argument("--api-key", type=str); p.add_argument("--cloud-model", type=str)
+    p.add_argument("--duel", action="store_true", help="啟動時開啟 Duel 模式")
+    p.add_argument("--distill", action="store_true")
+    p.add_argument("--grind", action="store_true", help="LeetCode 自動刷題")
+    p.add_argument("--grind-count", type=int, default=100, help="刷幾題")
+    p.add_argument("--stats", action="store_true"); p.add_argument("--train", action="store_true")
+    a = p.parse_args()
+    if a.stats: show_stats()
+    elif a.train: from rich.console import Console; trigger_training(FeedbackDB(), Console(), a)
+    elif a.grind: run_grind(a, a.grind_count)
+    else: run_agent_loop(a)
+
+if __name__ == "__main__": main()