from __future__ import annotations import asyncio import json from datetime import datetime from pathlib import Path from typing import Any, Dict, List, Optional, Tuple SUMMARY_PROMPT = ( "由于当前对话过长，系统正在自动压缩。请你基于已有上下文输出一份可继续执行的工作总结，要求：\n" "1) 任务目标与用户真实诉求\n" "2) 已完成工作（按时间顺序）\n" "3) 关键决策与原因\n" "4) 修改过的文件与核心变更点\n" "5) 工具调用中的重要结果/错误与修复\n" "6) 当前未完成事项与下一步计划（可直接执行）\n" "7) 风险与注意事项\n" "请使用中文，结构清晰，尽量具体，不要省略关键上下文。\n" "不要考虑过往对话，只考虑当前对话任务。\n" "禁止调用任何工具，必须直接输出总结内容。" ) GUIDE_USER_MESSAGE_TEMPLATE = "当前对话已经被自动压缩（第{count}次）。请阅读{path}并继续工作" def _emit(sender, event_type: str, payload: Dict[str, Any]): if not callable(sender): return try: sender(event_type, payload) except Exception: pass def _normalize_deep_compression_records(metadata: Dict[str, Any]) -> List[Dict[str, Any]]: records = metadata.get("deep_compression_records") if not isinstance(records, list): return [] normalized: List[Dict[str, Any]] = [] for item in records: if not isinstance(item, dict): continue try: count = int(item.get("count", 0) or 0) except Exception: count = 0 path = str(item.get("compact_file") or "").strip() if count <= 0 or not path: continue normalized.append({ "count": count, "compact_file": path, "created_at": item.get("created_at"), "source_conversation_id": item.get("source_conversation_id"), "compressed_conversation_id": item.get("compressed_conversation_id"), }) normalized.sort(key=lambda x: (int(x.get("count") or 0), str(x.get("created_at") or ""))) deduped: List[Dict[str, Any]] = [] seen = set() for rec in normalized: key = (int(rec.get("count") or 0), str(rec.get("compact_file") or "")) if key in seen: continue seen.add(key) deduped.append(rec) return deduped def _build_guide_message(*, compression_index: int, compact_file: str, previous_records: List[Dict[str, Any]]) -> str: """生成文件模式的引导语：仅提示压缩文件位置，由模型自行阅读。""" base = GUIDE_USER_MESSAGE_TEMPLATE.format(count=compression_index, path=compact_file) if not previous_records: return base lines = [base, "", "此前压缩摘要文件位置："] for rec in previous_records: count = int(rec.get("count") or 0) path = str(rec.get("compact_file") or "").strip() if count <= 0 or not path: continue lines.append(f"- 第{count}次：{path}") return "\n".join(lines).strip() def _read_compact_file_content(project_path: Path, relative_path: str) -> str: """读取 compact 文件全文，读取失败时返回空串。""" rel = str(relative_path or "").strip() if not rel: return "" try: target = (Path(project_path) / rel).resolve() return target.read_text(encoding="utf-8").strip() except Exception: return "" def _build_inject_guide_message( *, project_path: Path, compression_index: int, current_record: Dict[str, Any], previous_records: List[Dict[str, Any]], ) -> str: """生成直接注入模式的引导语：把历次压缩文件全文按顺序拼入正文，不提及文件位置。""" lines: List[str] = [f"当前对话已被第{compression_index}次压缩，以下为历次压缩的完整工作总结，请据此继续工作。"] all_records = list(previous_records) + [current_record] for rec in all_records: count = int(rec.get("count") or 0) rel_path = str(rec.get("compact_file") or "").strip() if count <= 0 or not rel_path: continue content = _read_compact_file_content(project_path, rel_path) lines.append("") lines.append(f"第{count}次压缩：") lines.append(content or "（压缩文件内容读取失败）") return "\n".join(lines).strip() def _extract_text_only(content: Any) -> str: if content is None: return "" if isinstance(content, str): return content if isinstance(content, list): parts: List[str] = [] for item in content: if not isinstance(item, dict): continue if item.get("type") == "text": txt = item.get("text") if isinstance(txt, str) and txt.strip(): parts.append(txt) return "\n".join(parts).strip() return str(content) def _extract_tool_arg_map(messages: List[Dict[str, Any]]) -> Dict[str, Dict[str, Any]]: tool_map: Dict[str, Dict[str, Any]] = {} for msg in messages: if msg.get("role") != "assistant": continue for tc in msg.get("tool_calls") or []: tc_id = tc.get("id") or tc.get("tool_call_id") if not tc_id: continue func = tc.get("function") or {} name = func.get("name") or tc.get("name") or "unknown_tool" args_raw = func.get("arguments") args_obj: Any = args_raw if isinstance(args_raw, str): try: args_obj = json.loads(args_raw) except Exception: args_obj = args_raw tool_map[tc_id] = {"name": name, "arguments": args_obj} return tool_map def _collect_last_tool_entries(messages: List[Dict[str, Any]], limit: int = 5, max_content_chars: int = 3000) -> List[Dict[str, Any]]: tool_arg_map = _extract_tool_arg_map(messages) entries: List[Dict[str, Any]] = [] for msg in messages: if msg.get("role") != "tool": continue tc_id = msg.get("tool_call_id") or msg.get("id") mapping = tool_arg_map.get(tc_id, {}) content_text = _extract_text_only(msg.get("content")) if len(content_text) > max_content_chars: content_text = content_text[:max_content_chars] + "\n...(已截断)" entries.append({ "tool_call_id": tc_id, "tool_name": msg.get("name") or mapping.get("name") or "unknown_tool", "arguments": mapping.get("arguments"), "content": content_text, }) return entries[-max(1, limit):] def _collect_user_texts(messages: List[Dict[str, Any]]) -> List[str]: result: List[str] = [] for msg in messages: if msg.get("role") != "user": continue metadata = msg.get("metadata") if isinstance(msg.get("metadata"), dict) else {} if str(metadata.get("message_source") or "").strip().lower() != "user": continue text = _extract_text_only(msg.get("content")) if text.strip(): result.append(text.strip()) return result async def _generate_summary(web_terminal, prompt: str, retries: int = 5) -> Tuple[str, Optional[str]]: last_reason: Optional[str] = None context = web_terminal.build_context() messages = web_terminal.build_messages(context, prompt) # build_messages 不会自动附加 user_input，压缩总结必须显式注入 if prompt and isinstance(prompt, str): messages = list(messages) + [{"role": "user", "content": prompt}] # 关键：总结请求必须与"用户正常发一条消息"完全无差别，才能 100% 命中前缀缓存。 # 因此 tools 必须照常传入（缺失 tools 字段会改变请求前缀、破坏缓存）。 # 模型即便返回 tool_calls 也会被忽略——我们只取 content；prompt 已要求其直接输出总结。 tools = web_terminal.define_tools() for _ in range(max(1, retries)): try: response_text = "" async for chunk in web_terminal.api_client.chat(messages, tools=tools, stream=False): if not isinstance(chunk, dict): continue choices = chunk.get("choices") or [] if choices: msg = choices[0].get("message") or {} content = msg.get("content") if isinstance(content, str): response_text = content if response_text.strip(): return response_text.strip(), None last_reason = "模型返回空内容" except Exception as exc: last_reason = str(exc) await asyncio.sleep(0.2) return f"生成总结失败（{last_reason or '未知原因'}）", last_reason def _write_compact_file( project_path: Path, *, compression_index: int, summary_text: str, user_inputs: List[str], last_tools: List[Dict[str, Any]], latest_user_input: str, ) -> str: compact_dir = project_path / ".agents" / "compact_result" compact_dir.mkdir(parents=True, exist_ok=True) filename = f"compact_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{compression_index:03d}.md" file_path = compact_dir / filename lines: List[str] = [ f"# 对话已被第{compression_index}次压缩", "", "## 工作总结", summary_text or "生成总结失败", "", "## 用户的所有输入（仅文字）", ] if user_inputs: for idx, text in enumerate(user_inputs, start=1): lines.append(f"{idx}. {text}") else: lines.append("- （无）") lines.extend(["", "## 最近5条工具调用（参数 + 结果）"]) if last_tools: for idx, item in enumerate(last_tools, start=1): lines.extend([ f"### {idx}. {item.get('tool_name')}", f"- tool_call_id: {item.get('tool_call_id')}", "- 参数：", "```json", json.dumps(item.get("arguments"), ensure_ascii=False, indent=2), "```", "- 结果：", "```text", str(item.get("content") or ""), "```", "", ]) else: lines.append("- （无）") lines.extend(["", "## 用户最新的一次输入"]) if (latest_user_input or "").strip(): lines.append(latest_user_input.strip()) else: lines.append("（无）") file_path.write_text("\n".join(lines).strip() + "\n", encoding="utf-8") return str(file_path.relative_to(project_path)) def _mark_history_compacted(history: List[Dict[str, Any]], *, round_index: int, now: str) -> int: """把当前对话历史中所有尚未标记的消息打上 deep_compacted 标记（in-place）。 深压缩按"整段前缀"处理：本次压缩时，历史里所有还没被标记的消息整体视为已压缩前缀。 这天然保证 assistant.tool_calls 与其 tool 响应被一并标记/排除，不会出现配对悬空。 原文保留在 metadata 中，持久化与重新加载时照常显示，仅在 build_messages 构建请求时被跳过。 返回本次新标记的消息条数。 """ if not isinstance(history, list): return 0 marked = 0 for msg in history: if not isinstance(msg, dict): continue metadata = msg.get("metadata") if not isinstance(metadata, dict): metadata = {} if metadata.get("deep_compacted"): continue metadata["deep_compacted"] = True metadata["deep_compacted_round"] = round_index metadata["deep_compacted_at"] = now msg["metadata"] = metadata marked += 1 return marked async def run_deep_compression( *, web_terminal, workspace, conversation_id: str, mode: str, sender=None, ) -> Dict[str, Any]: cm = web_terminal.context_manager conv_data = cm.conversation_manager.load_conversation(conversation_id) if not conv_data: return {"success": False, "error": f"对话不存在: {conversation_id}"} metadata = conv_data.get("metadata", {}) or {} if metadata.get("compression_in_progress"): return {"success": False, "error": "对话正在压缩中", "in_progress": True} # 读取个性化压缩设置： # - compress_form: file（生成文件，引导语提示位置） / inject（把历次压缩文件全文注入引导语） # - compress_behavior: continue（注入引导语并触发请求） / wait（仅注入引导语，等待用户） # compress_behavior 仅作用于手动压缩；自动深压缩永远继续工作。 try: from modules.personalization_manager import load_personalization_config pconfig = load_personalization_config(workspace.data_dir) or {} except Exception: pconfig = {} compress_form = str(pconfig.get("deep_compress_form") or "file").strip().lower() if compress_form not in ("file", "inject"): compress_form = "file" compress_behavior = str(pconfig.get("deep_compress_behavior") or "continue").strip().lower() if compress_behavior not in ("continue", "wait"): compress_behavior = "continue" if mode != "manual": compress_behavior = "continue" # in-place 压缩全程操作"当前对话"的内存历史：总结、标记、状态写入都依赖 # cm.conversation_history / cm.conversation_metadata。若当前对话不是目标对话， # 必须先切换过去，否则总结会基于错误历史、压缩状态也会写错对话。 if getattr(cm, "current_conversation_id", None) != conversation_id: try: web_terminal.load_conversation(conversation_id) except Exception as exc: return {"success": False, "error": f"加载目标对话失败: {exc}"} compression_count = int(metadata.get("compression_count", 0) or 0) previous_records = _normalize_deep_compression_records(metadata) previous_max_count = max([int(item.get("count") or 0) for item in previous_records], default=0) target_count = max(compression_count, previous_max_count) + 1 job_id = f"cmp_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}" cm.set_compression_state( in_progress=True, mode=mode, stage="generating_summary", job_id=job_id, resume_payload={"conversation_id": conversation_id, "mode": mode}, ) _emit(sender, "compression_state", { "conversation_id": conversation_id, "in_progress": True, "mode": mode, "stage": "generating_summary", "job_id": job_id, }) summary_text, summary_fail_reason = await _generate_summary(web_terminal, SUMMARY_PROMPT, retries=5) if summary_fail_reason: _emit(sender, "system_message", {"content": f"自动压缩总结失败，将使用失败占位文本：{summary_fail_reason}"}) cm.set_compression_state( in_progress=True, mode=mode, stage="writing_compact", job_id=job_id, ) _emit(sender, "compression_state", { "conversation_id": conversation_id, "in_progress": True, "mode": mode, "stage": "writing_compact", "job_id": job_id, }) messages = conv_data.get("messages") or [] user_inputs = _collect_user_texts(messages) latest_user_input = user_inputs[-1] if user_inputs else "" last_tools = _collect_last_tool_entries(messages, limit=5, max_content_chars=3000) relative_compact_path = _write_compact_file( Path(workspace.project_path), compression_index=target_count, summary_text=summary_text, user_inputs=user_inputs, last_tools=last_tools, latest_user_input=latest_user_input, ) cm.set_compression_state( in_progress=True, mode=mode, stage="marking_history", job_id=job_id, ) # === in-place 压缩：不创建/切换新对话，只把当前对话历史前缀打上 deep_compacted 标记 === # 当前对话已在函数开头对齐为目标对话，这里直接对内存历史打标。 now_iso = datetime.now().isoformat() marked_count = _mark_history_compacted( cm.conversation_history or [], round_index=target_count, now=now_iso, ) # 标记后立即持久化历史（标记写在每条消息的 metadata 中）。 try: cm.save_current_conversation() except Exception as exc: _emit(sender, "system_message", {"content": f"压缩标记保存失败：{exc}"}) # 关键：重置 current_context_tokens，避免自动压缩续接后阈值判断仍读到压缩前的大值而陷入死循环。 # 真实上下文长度会在下一次 API 响应后被重新写入。 try: cm.conversation_manager.update_token_statistics( conversation_id, input_tokens=0, output_tokens=0, total_tokens=0, current_context_tokens=0, ) except Exception as exc: _emit(sender, "system_message", {"content": f"压缩后重置上下文统计失败：{exc}"}) current_record = { "count": target_count, "compact_file": relative_compact_path, "created_at": now_iso, "source_conversation_id": conversation_id, "compressed_conversation_id": conversation_id, } all_records = previous_records + [current_record] # 构建引导语（按压缩形式）。inject 模式读取历次压缩文件全文，文件仍会生成，只是不提及位置。 if compress_form == "inject": guide_message = _build_inject_guide_message( project_path=Path(workspace.project_path), compression_index=target_count, current_record=current_record, previous_records=previous_records, ) else: guide_message = _build_guide_message( compression_index=target_count, compact_file=relative_compact_path, previous_records=previous_records, ) # 更新对话 metadata：压缩记录 + 清理压缩状态标记（同一对话，无切换）。 meta_updates = { "compression_count": target_count, "deep_compression_records": all_records, "last_deep_compression_record": current_record, "compression_in_progress": False, "compression_mode": None, "compression_stage": None, "compression_job_id": None, "compression_error": summary_fail_reason, "compression_resume_payload": None, "is_ultra_long_conversation": False, } cm.conversation_manager.update_conversation_metadata(conversation_id, meta_updates) # 同步内存中的 metadata，保证后续在同一对话内的判断读到最新值。 try: if getattr(cm, "current_conversation_id", None) == conversation_id and isinstance(cm.conversation_metadata, dict): cm.conversation_metadata.update(meta_updates) except Exception: pass _emit(sender, "compression_finished", { "source_conversation_id": conversation_id, "conversation_id": conversation_id, "in_progress": False, "compact_file": relative_compact_path, "marked_count": marked_count, "compress_form": compress_form, "compress_behavior": compress_behavior, "job_id": job_id, }) return { "success": True, "in_place": True, "compressed_conversation_id": conversation_id, "compact_file": relative_compact_path, "marked_count": marked_count, "compress_form": compress_form, "compress_behavior": compress_behavior, "summary_failed": bool(summary_fail_reason), "guide_message": guide_message, }