jarvis-ai-assistant 0.3.30__py3-none-any.whl → 0.7.6__py3-none-any.whl

jarvis/jarvis_rag/retriever.py

@@ -10,7 +10,6 @@ from langchain.text_splitter import RecursiveCharacterTextSplitter
 from rank_bm25 import BM25Okapi  # type: ignore
 
 from .embedding_manager import EmbeddingManager
-from jarvis.jarvis_utils.output import OutputType, PrettyOutput
 
 
 class ChromaRetriever:
@@ -42,9 +41,8 @@ class ChromaRetriever:
         self.collection = self.client.get_or_create_collection(
             name=self.collection_name
         )
-        PrettyOutput.print(
-            f"ChromaDB 客户端已在 '{db_path}' 初始化，集合为 '{collection_name}'。",
-            OutputType.SUCCESS,
+        print(
+            f"✅ ChromaDB 客户端已在 '{db_path}' 初始化，集合为 '{collection_name}'。"
         )
 
         # BM25索引设置
@@ -58,15 +56,15 @@ class ChromaRetriever:
     def _load_or_initialize_bm25(self):
         """从磁盘加载BM25索引或初始化一个新索引。"""
         if os.path.exists(self.bm25_index_path):
-            PrettyOutput.print("正在加载现有的 BM25 索引...", OutputType.INFO)
+            print("ℹ️ 正在加载现有的 BM25 索引...")
             with open(self.bm25_index_path, "rb") as f:
                 data = pickle.load(f)
                 self.bm25_corpus = data["corpus"]
                 self.bm25_index = BM25Okapi(self.bm25_corpus)
-            PrettyOutput.print("BM25 索引加载成功。", OutputType.SUCCESS)
+            print("✅ BM25 索引加载成功。")
         else:
-            PrettyOutput.print(
-                "未找到 BM25 索引，将初始化一个新的。", OutputType.WARNING
+            print(
+                "⚠️ 未找到 BM25 索引，将初始化一个新的。"
             )
             self.bm25_corpus = []
             self.bm25_index = None
@@ -74,10 +72,10 @@ class ChromaRetriever:
     def _save_bm25_index(self):
         """将BM25索引保存到磁盘。"""
         if self.bm25_index:
-            PrettyOutput.print("正在保存 BM25 索引...", OutputType.INFO)
+            print("ℹ️ 正在保存 BM25 索引...")
             with open(self.bm25_index_path, "wb") as f:
                 pickle.dump({"corpus": self.bm25_corpus, "index": self.bm25_index}, f)
-            PrettyOutput.print("BM25 索引保存成功。", OutputType.SUCCESS)
+            print("✅ BM25 索引保存成功。")
 
     def _load_manifest(self) -> Dict[str, Dict[str, Any]]:
         """加载已索引文件清单，用于变更检测。"""
@@ -97,7 +95,7 @@ class ChromaRetriever:
             with open(self.manifest_path, "w", encoding="utf-8") as f:
                 json.dump(manifest, f, ensure_ascii=False, indent=2)
         except Exception as e:
-            PrettyOutput.print(f"保存索引清单失败: {e}", OutputType.WARNING)
+            print(f"⚠️ 保存索引清单失败: {e}")
 
     def _compute_md5(
         self, file_path: str, chunk_size: int = 1024 * 1024
@@ -136,8 +134,8 @@ class ChromaRetriever:
                 continue
         if updated > 0:
             self._save_manifest(manifest)
-            PrettyOutput.print(
-                f"已更新索引清单，记录 {updated} 个源文件状态。", OutputType.INFO
+            print(
+                f"ℹ️ 已更新索引清单，记录 {updated} 个源文件状态。"
             )
 
     def _detect_changed_or_deleted(self) -> Dict[str, List[str]]:
@@ -203,7 +201,8 @@ class ChromaRetriever:
         lines.append(
             "提示：请使用 'jarvis-rag add <路径>' 重新索引相关文件，以更新向量库与BM25索引。"
         )
-        PrettyOutput.print("\n".join(lines), OutputType.WARNING)
+        joined_lines = '\n'.join(lines)
+        print(f"⚠️ {joined_lines}")
 
     def detect_index_changes(self) -> Dict[str, List[str]]:
         """
@@ -225,8 +224,8 @@ class ChromaRetriever:
                 removed += 1
         if removed > 0:
             self._save_manifest(manifest)
-            PrettyOutput.print(
-                f"已从索引清单中移除 {removed} 个已删除的源文件记录。", OutputType.INFO
+            print(
+                f"ℹ️ 已从索引清单中移除 {removed} 个已删除的源文件记录。"
             )
 
     def update_index_for_changes(self, changed: List[str], deleted: List[str]) -> None:
@@ -254,7 +253,8 @@ class ChromaRetriever:
             except Exception as e:
                 delete_errors.append(f"删除源 '{src}' 时出错: {e}")
         if delete_errors:
-            PrettyOutput.print("\n".join(delete_errors), OutputType.WARNING)
+            joined_errors = '\n'.join(delete_errors)
+            print(f"⚠️ {joined_errors}")
 
         # 再处理变更（重建）
         docs_to_add: List[Document] = []
@@ -275,14 +275,15 @@ class ChromaRetriever:
             except Exception as e:
                 rebuild_errors.append(f"重建源 '{src}' 内容时出错: {e}")
         if rebuild_errors:
-            PrettyOutput.print("\n".join(rebuild_errors), OutputType.WARNING)
+            joined_errors = '\n'.join(rebuild_errors)
+            print(f"⚠️ {joined_errors}")
 
         if docs_to_add:
             try:
                 # 复用现有拆分与嵌入逻辑
                 self.add_documents(docs_to_add)
             except Exception as e:
-                PrettyOutput.print(f"添加变更文档到索引时出错: {e}", OutputType.ERROR)
+                print(f"❌ 添加变更文档到索引时出错: {e}")
 
         # 重建BM25索引，确保删除后的语料被清理
         try:
@@ -292,7 +293,7 @@ class ChromaRetriever:
             self.bm25_index = BM25Okapi(self.bm25_corpus) if self.bm25_corpus else None
             self._save_bm25_index()
         except Exception as e:
-            PrettyOutput.print(f"重建BM25索引失败: {e}", OutputType.WARNING)
+            print(f"⚠️ 重建BM25索引失败: {e}")
 
         # 更新manifest：变更文件更新状态；删除文件从清单中移除
         try:
@@ -301,11 +302,10 @@ class ChromaRetriever:
             if deleted:
                 self._remove_sources_from_manifest(deleted)
         except Exception as e:
-            PrettyOutput.print(f"更新索引清单时出错: {e}", OutputType.WARNING)
+            print(f"⚠️ 更新索引清单时出错: {e}")
 
-        PrettyOutput.print(
-            f"索引已更新：变更 {len(changed)} 个，删除 {len(deleted)} 个。",
-            OutputType.SUCCESS,
+        print(
+            f"✅ 索引已更新：变更 {len(changed)} 个，删除 {len(deleted)} 个。"
         )
 
     def add_documents(
@@ -319,9 +319,8 @@ class ChromaRetriever:
         )
         chunks = text_splitter.split_documents(documents)
 
-        PrettyOutput.print(
-            f"已将 {len(documents)} 个文档拆分为 {len(chunks)} 个块。",
-            OutputType.INFO,
+        print(
+            f"ℹ️ 已将 {len(documents)} 个文档拆分为 {len(chunks)} 个块。"
         )
 
         if not chunks:
@@ -341,9 +340,8 @@ class ChromaRetriever:
             documents=chunk_texts,
             metadatas=cast(Any, metadatas),
         )
-        PrettyOutput.print(
-            f"成功将 {len(chunks)} 个块添加到 ChromaDB 集合中。",
-            OutputType.SUCCESS,
+        print(
+            f"✅ 成功将 {len(chunks)} 个块添加到 ChromaDB 集合中。"
         )
 
         # 更新并保存BM25索引

jarvis/jarvis_sec/__init__.py

@@ -0,0 +1,305 @@
+# -*- coding: utf-8 -*-
+"""
+Jarvis 安全分析套件
+
+当前版本概述：
+- 关键路径：直扫（direct_scan）→ 单Agent逐条验证（只读工具：read_code/execute_script）→ 聚合输出（JSON + Markdown）
+- 目标范围：内存管理、缓冲区操作、错误处理等基础安全问题识别
+- 约束：不修改核心框架文件，保持最小侵入；严格只读分析
+
+集成方式：
+- 复用 jarvis.jarvis_agent.Agent 与工具注册系统（jarvis.jarvis_tools.registry.ToolRegistry）
+- 提供入口：
+  - run_security_analysis(entry_path, ...)：直扫 + 单Agent逐条验证 + 聚合
+
+  - workflow.direct_scan(entry_path, ...)：仅启发式直扫
+
+说明：
+- 已移除 MultiAgent 编排与相关提示词；不存在"阶段一"等表述
+- 模块化重构：将功能拆分为多个模块（prompts, parsers, utils, agents, clustering, analysis, verification, review）
+"""
+
+from typing import Dict, List, Optional
+
+import typer
+
+from jarvis.jarvis_agent import Agent  # noqa: F401
+from jarvis.jarvis_sec.workflow import direct_scan, run_with_agent
+
+# 导入模块化后的函数（用于触发模块加载）
+from jarvis.jarvis_sec.prompts import (  # noqa: F401
+    build_summary_prompt as _build_summary_prompt,
+    build_verification_summary_prompt as _build_verification_summary_prompt,
+    get_review_system_prompt as _get_review_system_prompt,
+    get_review_summary_prompt as _get_review_summary_prompt,
+    get_cluster_system_prompt as _get_cluster_system_prompt,
+    get_cluster_summary_prompt as _get_cluster_summary_prompt,
+)
+from jarvis.jarvis_sec.parsers import (  # noqa: F401
+    parse_clusters_from_text as _parse_clusters_from_text,
+    try_parse_summary_report as _try_parse_summary_report,
+)
+from jarvis.jarvis_sec.utils import (  # noqa: F401
+    git_restore_if_dirty as _git_restore_if_dirty,
+    get_sec_dir as _get_sec_dir,
+    initialize_analysis_context as _initialize_analysis_context,
+    load_or_run_heuristic_scan as _load_or_run_heuristic_scan,
+    compact_candidate as _compact_candidate,
+    prepare_candidates as _prepare_candidates,
+    group_candidates_by_file as _group_candidates_by_file,
+    create_report_writer as _create_report_writer,
+    sig_of as _sig_of,
+    load_processed_gids_from_issues as _load_processed_gids_from_issues,
+    count_issues_from_file as _count_issues_from_file,
+    load_all_issues_from_file as _load_all_issues_from_file,
+    load_processed_gids_from_agent_issues as _load_processed_gids_from_agent_issues,
+)
+from jarvis.jarvis_sec.agents import (  # noqa: F401
+    subscribe_summary_event as _subscribe_summary_event,
+    create_analysis_agent as _create_analysis_agent,
+    create_review_agent as _create_review_agent,
+    create_cluster_agent as _create_cluster_agent,
+)
+from jarvis.jarvis_sec.clustering import (  # noqa: F401
+    load_existing_clusters as _load_existing_clusters,
+    restore_clusters_from_checkpoint as _restore_clusters_from_checkpoint,
+    create_cluster_snapshot_writer as _create_cluster_snapshot_writer,
+    collect_candidate_gids as _collect_candidate_gids,
+    collect_clustered_gids as _collect_clustered_gids,
+    # supplement_missing_gids_for_clustering已移除，不再需要
+    handle_single_alert_file as _handle_single_alert_file,
+    validate_cluster_format as _validate_cluster_format,
+    extract_classified_gids as _extract_classified_gids,
+    build_cluster_retry_task as _build_cluster_retry_task,
+    build_cluster_error_guidance as _build_cluster_error_guidance,
+    run_cluster_agent_direct_model as _run_cluster_agent_direct_model,
+    validate_cluster_result as _validate_cluster_result,
+    check_cluster_completeness as _check_cluster_completeness,
+    run_cluster_agent_with_retry as _run_cluster_agent_with_retry,
+    process_cluster_results as _process_cluster_results,
+    supplement_missing_gids as _supplement_missing_gids,
+    build_cluster_task as _build_cluster_task,
+    extract_input_gids as _extract_input_gids,
+    build_gid_to_item_mapping as _build_gid_to_item_mapping,
+    process_cluster_chunk as _process_cluster_chunk,
+    filter_pending_items as _filter_pending_items,
+    process_file_clustering as _process_file_clustering,
+    # check_and_supplement_missing_gids已移除，完整性检查已移至process_clustering_phase中
+    initialize_clustering_context as _initialize_clustering_context,
+    check_unclustered_gids as _check_unclustered_gids,
+    execute_clustering_for_files as _execute_clustering_for_files,
+    record_clustering_completion as _record_clustering_completion,
+    fallback_to_file_based_batches as _fallback_to_file_based_batches,
+    process_clustering_phase as _process_clustering_phase,
+)
+from jarvis.jarvis_sec.review import (  # noqa: F401
+    build_review_task as _build_review_task,
+    process_review_batch_items as _process_review_batch_items,
+    reinstated_candidates_to_cluster_batches as _reinstated_candidates_to_cluster_batches,
+    process_review_phase as _process_review_phase,
+    build_gid_to_review_mapping as _build_gid_to_review_mapping,
+    process_review_batch as _process_review_batch,
+    run_review_agent_with_retry as _run_review_agent_with_retry,
+    is_valid_review_item as _is_valid_review_item,
+)
+from jarvis.jarvis_sec.analysis import (  # noqa: F401
+    build_analysis_task_context as _build_analysis_task_context,
+    build_validation_error_guidance as _build_validation_error_guidance,
+    run_analysis_agent_with_retry as _run_analysis_agent_with_retry,
+    expand_and_filter_analysis_results as _expand_and_filter_analysis_results,
+    valid_items as _valid_items,
+)
+from jarvis.jarvis_sec.verification import (  # noqa: F401
+    build_gid_to_verification_mapping as _build_gid_to_verification_mapping,
+    merge_verified_items as _merge_verified_items,
+    merge_verified_items_without_verification as _merge_verified_items_without_verification,
+    process_verification_batch as _process_verification_batch,
+    is_valid_verification_item as _is_valid_verification_item,
+    run_verification_agent_with_retry as _run_verification_agent_with_retry,
+    process_verification_phase as _process_verification_phase,
+)
+
+
+# 注：当前版本不使用 MultiAgent 编排，已移除默认多智能体配置与创建函数。
+# 请使用 run_security_analysis（单Agent逐条验证）或 workflow.direct_scan + format_markdown_report（直扫基线）。
+# 注意：部分函数已迁移到模块化文件中（prompts.py, parsers.py, utils.py, agents.py, clustering.py, analysis.py, verification.py, review.py），
+# 本文件中保留了这些函数的别名导入，以便向后兼容。
+
+
+
+
+
+
+
+
+
+def run_security_analysis(
+    entry_path: str,
+    languages: Optional[List[str]] = None,
+    llm_group: Optional[str] = None,
+    report_file: Optional[str] = None,
+    cluster_limit: int = 50,
+    exclude_dirs: Optional[List[str]] = None,
+    enable_verification: bool = True,
+    force_save_memory: bool = False,
+    output_file: Optional[str] = None,
+) -> str:
+    """
+    运行安全分析工作流（混合模式）。
+
+    改进：
+    - 即使在 agent 模式下，也先进行本地正则/启发式直扫，生成候选问题；
+      然后将候选问题拆分为子任务，交由多Agent进行深入分析与聚合。
+    
+    注意：此函数会在发生异常时更新状态文件为 error 状态。
+
+    参数：
+    - entry_path: 待分析的根目录路径
+    - languages: 限定扫描的语言扩展（例如 ["c", "cpp", "h", "hpp", "rs"]），为空则使用默认
+
+    返回：
+    - 最终报告（字符串），由 Aggregator 生成（JSON + Markdown）
+
+    其他：
+    - llm_group: 模型组名称（仅在当前调用链内生效，不覆盖全局配置），将直接传入 Agent 用于选择模型
+    - report_file: 增量报告文件路径（JSONL）。当每个子任务检测到 issues 时，立即将一条记录追加到该文件；
+      若未指定，则默认写入 entry_path/.jarvis/sec/agent_issues.jsonl
+    - cluster_limit: 聚类时每批次最多处理的告警数（默认 50），当单个文件告警过多时按批次进行聚类
+    - exclude_dirs: 要排除的目录列表（可选），默认已包含测试目录（test, tests, __tests__, spec, testsuite, testdata）
+    - enable_verification: 是否启用二次验证（默认 True），关闭后分析Agent确认的问题将直接写入报告
+    - 断点续扫: 默认开启。会基于 .jarvis/sec/candidates.jsonl、clusters.jsonl 和 analysis.jsonl 文件进行状态恢复。
+    """
+
+    langs = languages or ["c", "cpp", "h", "hpp", "rs"]
+
+    # 状态管理器（不再使用 status.json，使用空对象）
+    class DummyStatusManager:
+        def update_pre_scan(self, **kwargs): pass
+        def update_clustering(self, **kwargs): pass
+        def update_review(self, **kwargs): pass
+        def update_verification(self, **kwargs): pass
+        def mark_completed(self, **kwargs): pass
+        def mark_error(self, **kwargs): pass
+    
+    status_mgr = DummyStatusManager()
+
+    # 初始化分析上下文
+    sec_dir, progress_path, _progress_append = _initialize_analysis_context(
+        entry_path, status_mgr
+    )
+
+    # 1) 启发式扫描（支持断点续扫）
+    candidates, summary = _load_or_run_heuristic_scan(
+        entry_path, langs, exclude_dirs, sec_dir, status_mgr, _progress_append
+    )
+
+    # 2) 将候选问题精简为子任务清单，控制上下文长度
+    compact_candidates = _prepare_candidates(candidates)
+    
+    # 3) 保存候选到新的 candidates.jsonl 文件（包含gid）
+    from jarvis.jarvis_sec.file_manager import save_candidates, get_candidates_file
+    try:
+        save_candidates(sec_dir, compact_candidates)
+        _progress_append({
+            "event": "candidates_saved",
+            "path": str(get_candidates_file(sec_dir)),
+            "issues_count": len(compact_candidates),
+        })
+    except Exception:
+        pass
+    
+    # 记录批次选择信息（可选，用于日志）
+    try:
+        groups = _group_candidates_by_file(compact_candidates)
+        if groups:
+            selected_file, items = max(groups.items(), key=lambda kv: len(kv[1]))
+            try:
+                typer.secho(f"[jarvis-sec] 批次选择: 文件={selected_file} 数量={len(items)}", fg=typer.colors.BLUE)
+            except Exception:
+                pass
+            _progress_append({
+                "event": "batch_selection",
+                "selected_file": selected_file,
+                "selected_count": len(items),
+                "total_in_file": len(items),
+            })
+    except Exception:
+        pass
+    
+    # 创建报告写入函数
+    _append_report = _create_report_writer(sec_dir, report_file)
+
+    # 3) 处理聚类阶段
+    cluster_batches, invalid_clusters_for_review = _process_clustering_phase(
+        compact_candidates,
+        entry_path,
+        langs,
+        cluster_limit,
+        llm_group,
+        sec_dir,
+        status_mgr,
+        _progress_append,
+        force_save_memory=force_save_memory,
+    )
+
+    # 4) 处理验证阶段
+    meta_records: List[Dict] = []
+    all_issues = _process_verification_phase(
+        cluster_batches,
+        entry_path,
+        langs,
+        llm_group,
+        sec_dir,
+        status_mgr,
+        _progress_append,
+        _append_report,
+        enable_verification=enable_verification,
+        force_save_memory=force_save_memory,
+    )
+    
+    # 5) 使用统一聚合器生成最终报告（JSON + Markdown）
+    try:
+        from jarvis.jarvis_sec.report import build_json_and_markdown
+        result = build_json_and_markdown(
+            all_issues,
+            scanned_root=summary.get("scanned_root"),
+            scanned_files=summary.get("scanned_files"),
+            meta=meta_records or None,
+            output_file=output_file,
+        )
+        # 标记分析完成
+        status_mgr.mark_completed(
+            total_issues=len(all_issues),
+            message=f"安全分析完成，共发现 {len(all_issues)} 个问题"
+        )
+        return result
+    except Exception as e:
+        # 发生错误时更新状态
+        error_msg = str(e)
+        status_mgr.mark_error(
+            error_message=error_msg,
+            error_type=type(e).__name__
+        )
+        raise
+    finally:
+        # 清理LSP客户端资源，防止文件句柄泄露
+        try:
+            from jarvis.jarvis_tools.lsp_client import LSPClientTool
+            LSPClientTool.cleanup_all_clients()
+        except Exception:
+            pass  # 清理失败不影响主流程
+
+
+
+
+
+
+
+
+__all__ = [
+    
+    "run_security_analysis",
+
+    "direct_scan",
+    "run_with_agent",
+]

jarvis/jarvis_sec/agents.py

@@ -0,0 +1,143 @@
+# -*- coding: utf-8 -*-
+"""Agent创建和订阅模块"""
+
+from typing import Dict, Optional
+from jarvis.jarvis_agent import Agent
+from jarvis.jarvis_tools.registry import ToolRegistry
+from jarvis.jarvis_sec.prompts import (
+    build_summary_prompt,
+    get_review_system_prompt,
+    get_review_summary_prompt,
+    get_cluster_system_prompt,
+    get_cluster_summary_prompt,
+)
+
+
+def subscribe_summary_event(agent: Agent) -> Dict[str, str]:
+    """订阅Agent摘要事件"""
+    summary_container: Dict[str, str] = {"text": ""}
+    try:
+        from jarvis.jarvis_agent.events import AFTER_SUMMARY as _AFTER_SUMMARY
+    except Exception:
+        _AFTER_SUMMARY = None
+    
+    if _AFTER_SUMMARY:
+        def _on_after_summary(**kwargs):
+            try:
+                summary_container["text"] = str(kwargs.get("summary", "") or "")
+            except Exception:
+                summary_container["text"] = ""
+        try:
+            agent.event_bus.subscribe(_AFTER_SUMMARY, _on_after_summary)
+        except Exception:
+            pass
+    return summary_container
+
+
+def create_analysis_agent(task_id: str, llm_group: Optional[str], force_save_memory: bool = False) -> Agent:
+    """创建分析Agent"""
+    system_prompt = """
+# 单Agent安全分析约束
+- 你的核心任务是评估代码的安全问题，目标：针对本候选问题进行证据核实、风险评估与修复建议补充，查找漏洞触发路径，确认在某些条件下会触发；以此来判断是否是漏洞。
+- **必须进行调用路径推导**：
+  - 对于每个候选问题，必须明确推导从可控输入到缺陷代码的完整调用路径。
+  - 调用路径推导必须包括：
+    1. 识别可控输入的来源（例如：用户输入、网络数据、文件读取、命令行参数等）
+    2. 追踪数据流：从输入源开始，逐步追踪数据如何传递到缺陷代码位置
+    3. 识别调用链：明确列出从入口函数到缺陷代码的所有函数调用序列（例如：main() -> parse_input() -> process_data() -> vulnerable_function()）
+    4. 分析每个调用点的数据校验情况：检查每个函数是否对输入进行了校验、边界检查或安全检查
+    5. 确认触发条件：明确说明在什么条件下，未校验或恶意输入能够到达缺陷代码位置
+  - 如果无法推导出完整的调用路径，或者所有调用路径都有充分的保护措施，则应该判定为误报。
+  - 调用路径推导必须在分析过程中明确展示，不能省略或假设。
+- 工具优先：使用 read_code 读取目标文件附近源码（行号前后各 ~50 行），必要时用 execute_script 辅助检索。
+- **调用路径追溯要求**：
+  - 必须向上追溯所有可能的调用者，查看完整的调用路径，以确认风险是否真实存在。
+  - 使用 read_code 和 execute_script 工具查找函数的调用者（例如：使用 grep 搜索函数名，查找所有调用该函数的位置）。
+  - 对于每个调用者，必须检查其是否对输入进行了校验。
+  - 如果发现任何调用路径未做校验，必须明确记录该路径。
+  - 例如：一个函数存在空指针解引用风险，必须检查所有调用者。如果所有调用者均能确保传入的指针非空，则该风险在当前代码库中可能不会实际触发；但如果存在任何调用者未做校验，则风险真实存在。
+- 若多条告警位于同一文件且行号相距不远，可一次性读取共享上下文，对这些相邻告警进行联合分析与判断；但仍需避免无关扩展与大范围遍历。
+- 禁止修改任何文件或执行写操作命令（rm/mv/cp/echo >、sed -i、git、patch、chmod、chown 等）；仅进行只读分析与读取。
+- 每次仅执行一个操作；等待工具结果后再进行下一步。
+- **记忆使用**：
+  - 在分析过程中，充分利用 retrieve_memory 工具检索已有的记忆，特别是与当前分析函数相关的记忆。
+  - 如果有必要，使用 save_memory 工具保存每个函数的分析要点，使用函数名作为 tag（例如：函数名、文件名等）。
+  - 记忆内容示例：某个函数的指针已经判空、某个函数已有输入校验、某个函数的调用路径分析结果等。
+  - 这样可以避免重复分析，提高效率，并保持分析的一致性。
+- 完成对本批次候选问题的判断后，主输出仅打印结束符 <!!!COMPLETE!!!> ，不需要汇总结果。
+""".strip()
+    
+    agent_kwargs: Dict = dict(
+        system_prompt=system_prompt,
+        name=task_id,
+        auto_complete=True,
+        need_summary=True,
+        summary_prompt=build_summary_prompt(),
+        non_interactive=True,
+        in_multi_agent=False,
+        use_methodology=False,
+        use_analysis=False,
+        output_handler=[ToolRegistry()],
+        force_save_memory=force_save_memory,
+        use_tools=["read_code", "execute_script", "save_memory", "retrieve_memory"],
+    )
+    if llm_group:
+        agent_kwargs["model_group"] = llm_group
+    return Agent(**agent_kwargs)
+
+
+def create_review_agent(
+    current_review_num: int,
+    llm_group: Optional[str],
+) -> Agent:
+    """创建复核Agent"""
+    review_system_prompt = get_review_system_prompt()
+    review_summary_prompt = get_review_summary_prompt()
+    
+    review_task_id = f"JARVIS-SEC-Review-Batch-{current_review_num}"
+    review_agent_kwargs: Dict = dict(
+        system_prompt=review_system_prompt,
+        name=review_task_id,
+        auto_complete=True,
+        need_summary=True,
+        summary_prompt=review_summary_prompt,
+        non_interactive=True,
+        in_multi_agent=False,
+        use_methodology=False,
+        use_analysis=False,
+        output_handler=[ToolRegistry()],
+        use_tools=["read_code", "execute_script", "retrieve_memory", "save_memory"],
+    )
+    if llm_group:
+        review_agent_kwargs["model_group"] = llm_group
+    return Agent(**review_agent_kwargs)
+
+
+def create_cluster_agent(
+    file: str,
+    chunk_idx: int,
+    llm_group: Optional[str],
+    force_save_memory: bool = False,
+) -> Agent:
+    """创建聚类Agent"""
+    cluster_system_prompt = get_cluster_system_prompt()
+    cluster_summary_prompt = get_cluster_summary_prompt()
+    
+    agent_kwargs_cluster: Dict = dict(
+        system_prompt=cluster_system_prompt,
+        name=f"JARVIS-SEC-Cluster::{file}::batch{chunk_idx}",
+        auto_complete=True,
+        need_summary=True,
+        summary_prompt=cluster_summary_prompt,
+        non_interactive=True,
+        in_multi_agent=False,
+        use_methodology=False,
+        use_analysis=False,
+        output_handler=[ToolRegistry()],
+        force_save_memory=force_save_memory,
+        use_tools=["read_code", "execute_script", "save_memory", "retrieve_memory"],
+    )
+    if llm_group:
+        agent_kwargs_cluster["model_group"] = llm_group
+    return Agent(**agent_kwargs_cluster)
+