P0: /init + tool budget reduction + error recovery + auto-verify + hooks

codepilot_v4.py

@@ -404,6 +404,23 @@ class ProjectTools:
 
 TOOL_PATTERN = re.compile(r'<tool>\s*(\w+)\s*\n(.*?)</tool>', re.DOTALL)
 
+# ============================================================
+#  P0-2: TOOL RESULT BUDGET REDUCTION（工具結果截斷）
+# ============================================================
+MAX_TOOL_RESULT_CHARS = 12000  # ~3000 tokens
+
+def truncate_tool_result(result, max_chars=MAX_TOOL_RESULT_CHARS):
+    """Claude Code 的 Budget Reduction — 限制每個工具結果大小"""
+    if len(result) <= max_chars:
+        return result
+    head = max_chars * 2 // 3
+    tail = max_chars // 3
+    truncated_lines = len(result) - max_chars
+    return (result[:head]
+            + f"\n\n... ⚠️ Output truncated ({len(result):,} chars total, {truncated_lines:,} chars omitted) ...\n\n"
+            + result[-tail:])
+
+
 def parse_tool_calls(text):
     calls = []
     for m in TOOL_PATTERN.finditer(text):
@@ -418,15 +435,227 @@ def parse_tool_calls(text):
 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}"
+        if n == "read_file": result = tools.read_file(p.get("path",""), int(p.get("offset",1)), int(p.get("limit",200)))
+        elif n == "edit_file": result = tools.edit_file(p.get("path",""), p.get("old_string",""), p.get("new_string",""))
+        elif n == "write_file": result = tools.write_file(p.get("path",""), p.get("content",""))
+        elif n == "run_command": result = tools.run_command(p.get("command",""), int(p.get("timeout",120)))
+        elif n == "search_files": result = tools.search_files(p.get("pattern",""), p.get("glob"))
+        elif n == "list_files": result = tools.list_files(p.get("pattern","*"), int(p.get("max_depth",3)))
+        elif n == "git_status": result = tools.git_context()
+        else: result = f"❌ 未知: {n}"
+    except Exception as e: result = f"❌ {e}"
+    return truncate_tool_result(result)  # P0-2: 自動截斷
+
+
+# ============================================================
+#  P0-1: /init 自動產生 CODEPILOT.md
+# ============================================================
+def cmd_init(tools, model, console):
+    """掃描專案結構，用模型自動產生 CODEPILOT.md"""
+    console.print("\n[bold]🔍 掃描專案結構...[/]")
+
+    # 收集專案資訊
+    file_list = tools.list_files("*", max_depth=2)
+    git = tools.git_context()
+
+    # 嘗試讀取關鍵檔案
+    key_files = {}
+    for f in ["README.md", "README.rst", "package.json", "pyproject.toml",
+              "requirements.txt", "Cargo.toml", "go.mod", "Makefile",
+              "docker-compose.yml", "Dockerfile", ".gitignore"]:
+        full = os.path.join(tools.project_dir, f)
+        if os.path.exists(full):
+            try:
+                content = Path(full).read_text(encoding="utf-8", errors="replace")[:3000]
+                key_files[f] = content
+            except: pass
+
+    key_files_text = "\n\n".join(f"--- {k} ---\n{v}" for k, v in key_files.items())
+
+    prompt = f"""Analyze this project and generate a CODEPILOT.md configuration file.
+
+## Project Files (top 2 levels)
+{file_list[:3000]}
+
+## Git Info
+{git}
+
+## Key Config Files
+{key_files_text[:6000]}
+
+## Instructions
+Generate a markdown file with these sections:
+1. **Project Overview** — one-line description
+2. **Tech Stack** — languages, frameworks, databases
+3. **Code Style** — formatting tools, naming conventions
+4. **Testing** — test framework, how to run tests
+5. **Key Commands** — build, run, test, lint commands
+6. **Architecture** — key directories and their purpose
+7. **Rules** — important rules for AI to follow (e.g., "always write tests", "use TypeScript strict mode")
+
+Be concise. Use bullet points. Write in the language matching the project (Chinese if README is Chinese, English otherwise)."""
+
+    with console.status("[bold cyan]分析專案中..."):
+        result = model.chat([{"role": "user", "content": prompt}], max_tokens=2048)
+
+    codepilot_path = os.path.join(tools.project_dir, "CODEPILOT.md")
+    Path(codepilot_path).write_text(result, encoding="utf-8")
+    console.print(f"\n[green]✅ 已產生 CODEPILOT.md[/]")
+    console.print(f"[dim]{result[:500]}...[/]")
+    console.print(f"\n[dim]檢查並編輯: {codepilot_path}[/]")
+    return result
+
+
+# ============================================================
+#  P0-3: ERROR RECOVERY（錯誤自動恢復）
+# ============================================================
+MAX_RETRIES = 3
+
+def chat_with_recovery(model, messages, ctx=None, console=None, fallback_model=None):
+    """帶自動恢復的 model.chat — 重試 + 壓縮 + fallback"""
+    last_error = None
+
+    for attempt in range(MAX_RETRIES):
+        try:
+            return model.chat(messages)
+        except Exception as e:
+            last_error = e
+            error_str = str(e).lower()
+
+            if console:
+                console.print(f"  [yellow]⚠️ 嘗試 {attempt+1}/{MAX_RETRIES}: {type(e).__name__}[/]")
+
+            # 策略 1: context 太長 → 壓縮
+            if any(k in error_str for k in ["too long", "too_long", "context_length", "max_tokens", "prompt_too_long"]):
+                if ctx and hasattr(ctx, 'check_compact'):
+                    if console: console.print("  [dim]🔄 壓縮對話歷史...[/]")
+                    messages = ctx.check_compact(messages, model_chat_fn=model.chat)
+                    continue
+                else:
+                    # 手動截斷
+                    if len(messages) > 6:
+                        messages = [messages[0]] + messages[-4:]
+                        continue
+
+            # 策略 2: rate limit → 等待重試
+            if any(k in error_str for k in ["rate_limit", "429", "too many"]):
+                wait = 2 ** attempt * 5  # 5s, 10s, 20s
+                if console: console.print(f"  [dim]⏳ Rate limit, 等待 {wait}s...[/]")
+                time.sleep(wait)
+                continue
+
+            # 策略 3: 伺服器錯誤 → 等待重試
+            if any(k in error_str for k in ["500", "502", "503", "server", "timeout", "connection"]):
+                wait = 2 ** attempt * 3
+                if console: console.print(f"  [dim]⏳ 伺服器錯誤, 等待 {wait}s...[/]")
+                time.sleep(wait)
+                continue
+
+            # 策略 4: 切換 fallback model
+            if fallback_model and attempt == MAX_RETRIES - 1:
+                if console: console.print(f"  [yellow]🔄 切換到 fallback 模型...[/]")
+                try:
+                    return fallback_model.chat(messages)
+                except: pass
+
+            # 其他錯誤直接 break
+            break
+
+    raise last_error or RuntimeError("chat failed after retries")
+
+
+# ============================================================
+#  P0-4: VERIFICATION SUB-AGENT（驗證子代理）
+# ============================================================
+def run_verification(model, tools, console, edited_files=None):
+    """完成修改後自動跑測試驗證"""
+    console.print("\n[bold]🔍 Verification Agent[/]")
+
+    checks = []
+
+    # 1. 語法檢查修改過的 Python 文件
+    if edited_files:
+        for f in edited_files:
+            if f.endswith(".py") and os.path.exists(f):
+                try:
+                    content = Path(f).read_text()
+                    compile(content, f, "exec")
+                    checks.append(f"  ✅ {os.path.basename(f)} 語法正確")
+                except SyntaxError as e:
+                    checks.append(f"  ❌ {os.path.basename(f)} 語法錯誤: {e.msg} (line {e.lineno})")
+
+    # 2. 嘗試跑 pytest / npm test
+    test_commands = []
+    if os.path.exists(os.path.join(tools.project_dir, "pytest.ini")) or \
+       os.path.exists(os.path.join(tools.project_dir, "tests")) or \
+       os.path.exists(os.path.join(tools.project_dir, "test")):
+        test_commands.append(("pytest", f"{sys.executable} -m pytest --tb=short -q"))
+
+    if os.path.exists(os.path.join(tools.project_dir, "package.json")):
+        test_commands.append(("npm test", "npm test --if-present 2>&1 | head -30"))
+
+    if os.path.exists(os.path.join(tools.project_dir, "Makefile")):
+        # 檢查是否有 test target
+        makefile = Path(os.path.join(tools.project_dir, "Makefile")).read_text(errors="replace")
+        if "test:" in makefile:
+            test_commands.append(("make test", "make test 2>&1 | tail -20"))
+
+    for name, cmd in test_commands:
+        console.print(f"  [dim]🧪 Running {name}...[/]")
+        result = tools.run_command(cmd, timeout=60)
+        # 判斷通過/失敗
+        result_lower = result.lower()
+        if any(k in result_lower for k in ["passed", "ok", "success", "0 error"]):
+            passed_match = re.search(r'(\d+) passed', result)
+            n = passed_match.group(1) if passed_match else ""
+            checks.append(f"  ✅ {name}: {n} passed" if n else f"  ✅ {name}: OK")
+        elif any(k in result_lower for k in ["failed", "error", "fail"]):
+            # 只顯示最後幾行
+            last_lines = "\n".join(result.strip().split("\n")[-5:])
+            checks.append(f"  ❌ {name}: FAILED\n{last_lines}")
+        else:
+            checks.append(f"  ⚠️ {name}: {result[:200]}")
+
+    if not checks:
+        checks.append("  [dim]沒有找到測試框架[/]")
+
+    for c in checks:
+        console.print(c)
+
+    return checks
+
+
+# ============================================================
+#  P0-BONUS: HOOKS SYSTEM（post-edit 自動格式化）
+# ============================================================
+class Hooks:
+    """簡易 Hooks 系統 — 讀取 .codepilot/hooks.json"""
+
+    def __init__(self, project_dir):
+        self.project_dir = project_dir
+        self.hooks = {}
+        hooks_file = os.path.join(project_dir, ".codepilot", "hooks.json")
+        if os.path.exists(hooks_file):
+            try:
+                self.hooks = json.loads(Path(hooks_file).read_text())
+            except: pass
+
+    def run(self, event, context=None):
+        """執行 hook。context = {"file": "path/to/file.py"} 等"""
+        cmd_template = self.hooks.get(event)
+        if not cmd_template:
+            return None
+        # 替換變數
+        cmd = cmd_template
+        if context:
+            for k, v in context.items():
+                cmd = cmd.replace(f"{{{k}}}", str(v))
+        try:
+            result = subprocess.run(cmd, shell=True, cwd=self.project_dir,
+                capture_output=True, text=True, timeout=30)
+            return result.stdout + result.stderr if result.returncode != 0 else None
+        except:
+            return None
 
 
 def build_system_prompt(tools, project_memory=""):
@@ -644,7 +873,10 @@ def run_agent_loop(args):
     else:
         messages = [{"role": "system", "content": system_prompt}]
 
-    console.print("[dim]/duel on|off  /memo  /grind  /ls  /git  /clear  /status  /train  /quit[/]\n")
+    hooks = Hooks(project_dir)  # P0-Bonus: Hooks 系統
+    edited_files_this_session = []  # 追蹤修改過的文件
+
+    console.print("[dim]/init /duel on|off /memo /verify /grind /ls /git /clear /status /train /quit[/]\n")
 
     while True:
         try: user_input = Prompt.ask("\n[bold green]🧑 You")
@@ -655,6 +887,17 @@ def run_agent_loop(args):
         # ---- 指令 ----
         if cmd in ("/quit", "/exit"): break
 
+        elif cmd == "/init":
+            result = cmd_init(tools, model, console)
+            # 重建 system prompt
+            system_prompt = ctx.build_system_prompt(tools.git_context())
+            messages[0] = {"role": "system", "content": system_prompt}
+            continue
+
+        elif cmd == "/verify":
+            run_verification(model, tools, console, edited_files_this_session)
+            continue
+
         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[/]")
@@ -786,14 +1029,21 @@ def run_agent_loop(args):
             ctx.save_session(messages)
             continue
 
-        # ---- 正常模式：單模型 + 工具循環 ----
+        # ---- 正常模式：單模型 + 工具循環 + 錯誤恢復 ----
         messages.append({"role": "user", "content": user_input})
         full_response = ""
+        tools_used_this_turn = []  # 追蹤這輪用了哪些工具
 
         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
+                try:
+                    # P0-3: 使用帶恢復的 chat
+                    response = chat_with_recovery(
+                        model, messages, ctx=ctx, console=console,
+                        fallback_model=local_model_ref if provider_key != "local" else None)
+                except Exception as e:
+                    console.print(f"[red]❌ 所有重試失敗: {e}[/]")
+                    break
 
             tool_calls = parse_tool_calls(response)
             text_parts = TOOL_PATTERN.sub("", response).strip()
@@ -807,7 +1057,22 @@ def run_agent_loop(args):
             results = []
             for call in tool_calls:
                 console.print(f"  [dim]🔧 {call['tool']}[/]")
-                result = execute_tool(tools, call)
+                result = execute_tool(tools, call)  # 已含 P0-2 截斷
+                tools_used_this_turn.append(call["tool"])
+
+                # 追蹤修改的文件
+                if call["tool"] in ("edit_file", "write_file") and "✅" in result:
+                    fpath = call["params"].get("path", "")
+                    if fpath:
+                        full_path = os.path.join(tools.cwd, fpath) if not os.path.isabs(fpath) else fpath
+                        if full_path not in edited_files_this_session:
+                            edited_files_this_session.append(full_path)
+                        # P0-Bonus: 觸發 post-edit hook
+                        hook_result = hooks.run(f"post_{call['tool']}", {"file": full_path})
+                        if hook_result:
+                            console.print(f"  [dim]🪝 Hook: {hook_result[:100]}[/]")
+
+                # 顯示結果
                 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"))
@@ -817,6 +1082,12 @@ def run_agent_loop(args):
                 results.append(f"[{call['tool']}] {result}")
             messages.append({"role": "user", "content": "Tool results:\n" + "\n\n".join(results)})
 
+        # P0-4: 自動驗證 — 如果這輪有修改文件，自動跑測試
+        if any(t in ("edit_file", "write_file") for t in tools_used_this_turn):
+            if edited_files_this_session:
+                console.print(f"\n[dim]🔍 Auto-verify ({len(edited_files_this_session)} files modified)...[/]")
+                run_verification(model, tools, console, edited_files_this_session)
+
         # 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)