JOJO
310d82f4b8
- 新增 modules/memory_debug.py 统一内存监控模块 - 在子智能体任务、主对话 build_messages、多智能体消息注入、 API 请求准备、深压缩等关键路径记录 RSS/VMS 和消息规模 - 日志输出到 ~/.astrion/astrion/host/logs/memory_debug.log
538 lines
20 KiB
Python
538 lines
20 KiB
Python
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import json
|
||
from datetime import datetime
|
||
from pathlib import Path
|
||
from typing import Any, Dict, List, Optional, Tuple
|
||
|
||
from modules.memory_debug import log_memory_event, estimate_messages_size
|
||
|
||
|
||
def _load_summary_prompt(web_terminal) -> str:
|
||
"""从 prompts/deep_compression_summary.txt 加载压缩总结提示词。"""
|
||
try:
|
||
return web_terminal.load_prompt("deep_compression_summary").strip()
|
||
except Exception:
|
||
return (
|
||
"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。"
|
||
)
|
||
|
||
|
||
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
|
||
try:
|
||
user_inputs_before = int(item.get("user_inputs_before", 0) or 0)
|
||
except Exception:
|
||
user_inputs_before = 0
|
||
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"),
|
||
"user_inputs_before": user_inputs_before,
|
||
"summary": str(item.get("summary") or ""),
|
||
})
|
||
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) -> 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。"""
|
||
return f"当前对话已经被第{compression_index}次压缩。请阅读 {compact_file} 并继续工作。"
|
||
|
||
|
||
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 _read_summary_from_record(record: Dict[str, Any]) -> str:
|
||
"""从 deep_compression_records 条目中读取保存的总结内容。"""
|
||
summary = record.get("summary")
|
||
return str(summary).strip() if isinstance(summary, str) else ""
|
||
|
||
|
||
def _build_user_inputs_section(
|
||
user_inputs: List[str],
|
||
previous_records: List[Dict[str, Any]],
|
||
current_count: int,
|
||
) -> str:
|
||
"""构建'用户的所有输入'区块，按历史压缩轮次插入分段标记。"""
|
||
if not user_inputs:
|
||
return "（无）"
|
||
breakpoints: Dict[int, int] = {}
|
||
for rec in previous_records:
|
||
count = int(rec.get("count") or 0)
|
||
before = int(rec.get("user_inputs_before") or 0)
|
||
if count > 0 and before > 0:
|
||
breakpoints[before] = count
|
||
sorted_breaks = sorted(breakpoints.items(), key=lambda x: x[0])
|
||
break_iter = iter(sorted_breaks)
|
||
next_break = next(break_iter, None)
|
||
lines: List[str] = []
|
||
last_break_index = 0
|
||
for idx, text in enumerate(user_inputs, start=1):
|
||
lines.append(f"{idx}. {text}")
|
||
if next_break and idx == next_break[0]:
|
||
lines.append(f"<第{next_break[1]}次压缩>")
|
||
last_break_index = idx
|
||
next_break = next(break_iter, None)
|
||
# 当前压缩之后还有新增输入时，追加当前压缩标记
|
||
if len(user_inputs) > last_break_index:
|
||
lines.append(f"<当前触发的第{current_count}次压缩>")
|
||
return "\n".join(lines)
|
||
|
||
|
||
def _build_summaries_section(
|
||
previous_records: List[Dict[str, Any]],
|
||
current_summary: str,
|
||
current_count: int,
|
||
) -> List[str]:
|
||
"""构建'历次压缩总结'区块，按顺序列出每次压缩的总结。"""
|
||
lines: List[str] = []
|
||
for rec in previous_records:
|
||
count = int(rec.get("count") or 0)
|
||
if count <= 0:
|
||
continue
|
||
summary = _read_summary_from_record(rec)
|
||
lines.append(f"### 第{count}次的总结")
|
||
lines.append(summary or "（读取失败）")
|
||
lines.append("")
|
||
lines.append(f"### 第{current_count}次的总结")
|
||
lines.append(current_summary or "（生成失败）")
|
||
return lines
|
||
|
||
|
||
def _build_inject_guide_message(
|
||
*,
|
||
compression_index: int,
|
||
current_record: Dict[str, Any],
|
||
previous_records: List[Dict[str, Any]],
|
||
user_inputs: List[str],
|
||
latest_user_input: str,
|
||
) -> str:
|
||
"""生成直接注入模式的引导语：把历次压缩总结、用户输入按顺序拼入正文。"""
|
||
lines: List[str] = [
|
||
f"当前对话已被第{compression_index}次压缩。以下为按时间顺序汇总的用户输入、历次压缩总结以及最近一次输入，请据此继续工作。",
|
||
"",
|
||
"用户的所有输入",
|
||
]
|
||
lines.append(_build_user_inputs_section(user_inputs, previous_records, compression_index))
|
||
lines.append("")
|
||
lines.append("历次压缩总结")
|
||
lines.append("")
|
||
summary_lines = _build_summaries_section(
|
||
previous_records,
|
||
current_record.get("summary", ""),
|
||
compression_index,
|
||
)
|
||
lines.extend(summary_lines)
|
||
lines.append("")
|
||
lines.append("用户的最近一次输入：")
|
||
if (latest_user_input or "").strip():
|
||
lines.append(latest_user_input.strip())
|
||
else:
|
||
lines.append("（无）")
|
||
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 _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],
|
||
latest_user_input: str,
|
||
previous_records: List[Dict[str, Any]],
|
||
) -> str:
|
||
compact_dir = project_path / ".astrion" / "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",
|
||
"",
|
||
"## 用户的所有输入",
|
||
]
|
||
lines.append(_build_user_inputs_section(user_inputs, previous_records, compression_index))
|
||
lines.append("")
|
||
lines.append("## 历次压缩总结")
|
||
lines.append("")
|
||
summary_lines = _build_summaries_section(
|
||
previous_records,
|
||
summary_text,
|
||
compression_index,
|
||
)
|
||
lines.extend(summary_lines)
|
||
lines.append("")
|
||
lines.append("## 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]:
|
||
log_memory_event(
|
||
"run_deep_compression_enter",
|
||
conversation_id=conversation_id,
|
||
mode=mode,
|
||
)
|
||
cm = web_terminal.context_manager
|
||
target_manager = (
|
||
cm._get_conversation_manager_for_id(conversation_id)
|
||
if hasattr(cm, "_get_conversation_manager_for_id")
|
||
else cm.conversation_manager
|
||
)
|
||
conv_data = target_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（把历次压缩内容注入引导语）
|
||
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 固定规则（无个性化开关）：
|
||
# - 手动压缩：只生成压缩消息（引导语），不自动续接，等待用户继续发送消息才工作。
|
||
# - 自动深压缩：当前任务尚未完成才触发，压缩后必须继续工作。
|
||
compress_behavior = "wait" if mode == "manual" else "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, _load_summary_prompt(web_terminal), 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 ""
|
||
user_inputs_before = len(user_inputs)
|
||
|
||
relative_compact_path = _write_compact_file(
|
||
Path(workspace.project_path),
|
||
compression_index=target_count,
|
||
summary_text=summary_text,
|
||
user_inputs=user_inputs,
|
||
latest_user_input=latest_user_input,
|
||
previous_records=previous_records,
|
||
)
|
||
|
||
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 ===
|
||
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:
|
||
target_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,
|
||
"user_inputs_before": user_inputs_before,
|
||
"summary": summary_text,
|
||
}
|
||
all_records = previous_records + [current_record]
|
||
|
||
# 构建引导语（按压缩形式）。
|
||
if compress_form == "inject":
|
||
guide_message = _build_inject_guide_message(
|
||
compression_index=target_count,
|
||
current_record=current_record,
|
||
previous_records=previous_records,
|
||
user_inputs=user_inputs,
|
||
latest_user_input=latest_user_input,
|
||
)
|
||
else:
|
||
guide_message = _build_guide_message(
|
||
compression_index=target_count,
|
||
compact_file=relative_compact_path,
|
||
)
|
||
|
||
# 更新对话 metadata：压缩记录 + 清理压缩状态标记（同一对话，无切换）。
|
||
# 同时清除 frozen prompt 缓存，使压缩后下一次请求自动重新加载动态内容。
|
||
REBUILD_FROZEN_KEYS = (
|
||
"frozen_main_system_prompt",
|
||
"frozen_permission_prompt",
|
||
"frozen_execution_prompt",
|
||
"frozen_recent_conversations_prompt",
|
||
"frozen_personalization_prompt",
|
||
"frozen_workspace_prompt",
|
||
"frozen_agents_md_prompt",
|
||
"frozen_skills_prompt",
|
||
"frozen_memory_prompt",
|
||
"frozen_custom_system_prompt",
|
||
"frozen_disabled_tools_prompt",
|
||
)
|
||
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,
|
||
}
|
||
for frozen_key in REBUILD_FROZEN_KEYS:
|
||
meta_updates[frozen_key] = None
|
||
target_manager.update_conversation_metadata(conversation_id, meta_updates)
|
||
# 同步内存中的 metadata，清除 frozen 缓存
|
||
try:
|
||
if getattr(cm, "current_conversation_id", None) == conversation_id and isinstance(cm.conversation_metadata, dict):
|
||
for frozen_key in REBUILD_FROZEN_KEYS:
|
||
cm.conversation_metadata.pop(frozen_key, None)
|
||
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,
|
||
})
|
||
log_memory_event(
|
||
"run_deep_compression_exit",
|
||
conversation_id=conversation_id,
|
||
mode=mode,
|
||
marked_count=marked_count,
|
||
history_size=estimate_messages_size(cm.conversation_history),
|
||
)
|
||
|
||
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,
|
||
}
|