From 02bc9c184087ed46c2b618a18706c1aff16957d7 Mon Sep 17 00:00:00 2001 From: btc-z Date: Thu, 23 Apr 2026 02:10:22 -0400 Subject: [PATCH] =?UTF-8?q?feat:=20=E6=96=B0=E5=A2=9E=E8=AF=AD=E9=9F=B3=20?= =?UTF-8?q?MCP=20=E5=B7=A5=E5=85=B7=20+=20macOS=20=E5=AF=86=E9=92=A5?= =?UTF-8?q?=E6=8F=90=E5=8F=96=E4=BF=AE=E5=A4=8D=20(#53)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * feat: 新增语音 MCP 工具 + macOS 密钥提取修复 - 新增 get_voice_messages / decode_voice / transcribe_voice MCP 工具 - 语音数据存储在 media_0.db VoiceInfo 表（SILK v3 格式） - decode_voice 解码为 WAV 文件（saved to decoded_voices/） - transcribe_voice 通过 Whisper 自动识别语言转录 - 新增 get_chat_history oldest_first 参数，支持从最早消息开始分页 - 修复 macOS 下 check_wechat_running / ensure_keys 逻辑 - 改用 pgrep 检测微信进程，绕过不支持 macOS 的 Python 扫描器 - 无 all_keys.json 时打印清晰引导，提示运行 C 版扫描器 - 新增 Makefile（build / keys / decrypt / web 快捷命令） - .gitignore 补充 find_all_keys_macos 二进制和 decoded_voices/ * fix: 语音查询支持多分片 media DB + 文件名唯一化 解决 PR #53 review 的阻塞项 #1，顺手修 #3、#6。 #1 `_get_media_db_path()` 硬编码 `media_0.db` - 新增模块级 `MEDIA_DB_KEYS`，镜像 `MSG_DB_KEYS` 的分片发现逻辑 - `_fetch_voice_row` 遍历所有分片，按 `(chat_name_id, local_id)` 首个命中即返回；单条语音在 media DB 家族内唯一，命中即可停 - `get_voice_messages` 从每个分片各取 `LIMIT limit`，合并排序后 截断到 `limit`。选择"每分片取 limit 条再合并"而非"按 max(create_time) 排序后逐个取到 limit 即停止"：后者假设分片 间时间不重叠，一旦 WeChat 改分片策略就会静默丢消息；前者工作 量 O(N 分片 × limit)，在任何分片布局下都正确 #3 输出文件名冲突 - `_silk_to_wav` 增加 `local_id` 参数，输出 `{user}_{time}_{lid}.wav`， 同一秒内两条语音不会互相覆盖；两个调用方都已在作用域内持有 `local_id` #6 `_fetch_voice_row` 的 `local_id=None` 死分支 - 随 #1 的重写一并删除，`local_id` 改为必填位置参数 Co-Authored-By: Claude Opus 4.7 * refactor: macOS 密钥提取分层下沉到 find_all_keys.py 解决 PR #53 review 的阻塞项 #2。 review 里提到"跟 PR #51 冲突"实测不存在 —— PR #51 当前 0 文件改动 （fork 分支已与上游同步），但架构建议本身是对的：macOS 处理应集中 在 `find_all_keys.py`，而不是在 `main.py` 提前 return 截胡。 - `main.py:ensure_keys()` 移除 darwin 专属提前返回分支，macOS 走 和其他平台相同的 `extract_keys()` 路径 - `find_all_keys.py:_load_impl()` 在 darwin 分支抛出带 `sudo ./find_all_keys_macos` 操作指引的 RuntimeError；非 macOS 的平台兜底分支保留 - `main.py` 里已有 `except RuntimeError` 会打印并 `sys.exit(1)`， 用户可见行为不变 未来若有 PR 在 `find_all_keys.py` 加 macOS 自动编译 / dispatch， 直接替换这段 RuntimeError 即可，不再需要改 `main.py`。 Co-Authored-By: Claude Opus 4.7 * chore: Makefile 支持 PYTHON 变量覆盖 解决 PR #53 review 的非阻塞项 #7。 原 Makefile 硬编码 `.venv/bin/python3`，没有 venv 的用户跑 `make decrypt` 直接报错。引入 `PYTHON ?= .venv/bin/python3`：默认行为 不变（仍走 venv），想用系统 Python 的用户 `PYTHON=python3 make decrypt` 即可。 Co-Authored-By: Claude Opus 4.7 * docs: 回应 PR #53 review #4 — 澄清 silk-python 与 pysilk 包名关系 验证：本项目 import 的 `pysilk` 实际由 `pip install silk-python` (synodriver/pysilk) 提供；pypi 上另有同名 `pysilk==0.0.1` 是无内容 的占位包，不可用。错误消息里 `pip install silk-python` 已经是对的， 但 reader 看到 `import pysilk` 仍会困惑，所以： - `_silk_to_wav` 的 import 处加一行注释，点名所用的是 synodriver 版本，并提醒 pypi 上还有 pilk / pysilk 两个同类包 - `decode_voice` / `transcribe_voice` 的 docstring 加 "依赖:" 行， 明确 "pip install silk-python (import 名为 pysilk)"，MCP 客户端 读 tool 描述就能看到正确的安装命令 未新增 requirements.txt 条目：voice 支持是可选功能（tool 内 try/except ImportError 懒加载），保持非必需依赖的语义。 Co-Authored-By: Claude Opus 4.7 --------- Co-authored-by: Claude Opus 4.7 --- .gitignore | 4 + Makefile | 16 ++++ find_all_keys.py | 11 ++- main.py | 4 + mcp_server.py | 222 +++++++++++++++++++++++++++++++++++++++++++++-- 5 files changed, 248 insertions(+), 9 deletions(-) create mode 100644 Makefile diff --git a/.gitignore b/.gitignore index db821c3..b10a3d5 100644 --- a/.gitignore +++ b/.gitignore @@ -23,3 +23,7 @@ __pycache__/ # OS .DS_Store Thumbs.db + +# Compiled binaries and output +find_all_keys_macos +decoded_voices/ diff --git a/Makefile b/Makefile new file mode 100644 index 0000000..96a5937 --- /dev/null +++ b/Makefile @@ -0,0 +1,16 @@ +.PHONY: keys decrypt web build + +PYTHON ?= .venv/bin/python3 + +build: + cc -O2 -o find_all_keys_macos find_all_keys_macos.c -framework Foundation + codesign -s - find_all_keys_macos + +keys: + sudo ./find_all_keys_macos + +decrypt: + $(PYTHON) main.py decrypt + +web: + $(PYTHON) main.py diff --git a/find_all_keys.py b/find_all_keys.py index eba2191..316f777 100644 --- a/find_all_keys.py +++ b/find_all_keys.py @@ -12,9 +12,16 @@ def _load_impl(): if system == "linux": import find_all_keys_linux as impl return impl + if system == "darwin": + raise RuntimeError( + "macOS 请先运行 C 版扫描器提取密钥：\n" + "\n" + " sudo ./find_all_keys_macos\n" + "\n" + " 完成后再运行 python main.py decrypt" + ) raise RuntimeError( - f"当前平台暂不支持通过 find_all_keys.py 提取密钥: {platform.system()}\n" - f"macOS 请使用 find_all_keys_macos.c (C 版扫描器)" + f"当前平台暂不支持通过 find_all_keys.py 提取密钥: {platform.system()}" ) diff --git a/main.py b/main.py index 885f975..9ee522c 100644 --- a/main.py +++ b/main.py @@ -6,7 +6,9 @@ python main.py decrypt # 提取密钥 + 解密全部数据库 """ import json import os +import platform import sys +import subprocess import functools print = functools.partial(print, flush=True) @@ -16,6 +18,8 @@ from key_utils import strip_key_metadata def check_wechat_running(): """检查微信是否在运行，返回 True/False""" + if platform.system().lower() == "darwin": + return subprocess.run(["pgrep", "-x", "WeChat"], capture_output=True).returncode == 0 from find_all_keys import get_pids try: get_pids() diff --git a/mcp_server.py b/mcp_server.py index 5c5101d..c464a06 100644 --- a/mcp_server.py +++ b/mcp_server.py @@ -5,7 +5,9 @@ Based on FastMCP (stdio transport), reuses existing decryption. Runs on Windows Python (needs access to D:\ WeChat databases). """ +import io import os, sys, json, time, sqlite3, tempfile, struct, hashlib, atexit, re +import wave import hmac as hmac_mod from contextlib import closing from datetime import datetime @@ -803,18 +805,19 @@ def _build_message_filters(start_ts=None, end_ts=None, keyword=''): return clauses, params -def _query_messages(conn, table_name, start_ts=None, end_ts=None, keyword='', limit=20, offset=0): +def _query_messages(conn, table_name, start_ts=None, end_ts=None, keyword='', limit=20, offset=0, oldest_first=False): if not _is_safe_msg_table_name(table_name): raise ValueError(f'非法消息表名: {table_name}') clauses, params = _build_message_filters(start_ts, end_ts, keyword) where_sql = f"WHERE {' AND '.join(clauses)}" if clauses else '' + order = 'ASC' if oldest_first else 'DESC' sql = f""" SELECT local_id, local_type, create_time, real_sender_id, message_content, WCDB_CT_message_content FROM [{table_name}] {where_sql} - ORDER BY create_time DESC + ORDER BY create_time {order} """ if limit is None: return conn.execute(sql, params).fetchall() @@ -994,14 +997,14 @@ def _history_query_batch_size(candidate_limit): return min(candidate_limit, _HISTORY_QUERY_BATCH_SIZE) -def _page_ranked_entries(entries, limit, offset): - ordered = sorted(entries, key=lambda item: item[0], reverse=True) +def _page_ranked_entries(entries, limit, offset, oldest_first=False): + ordered = sorted(entries, key=lambda item: item[0], reverse=not oldest_first) paged = ordered[offset:offset + limit] paged.sort(key=lambda item: item[0]) return paged -def _collect_chat_history_lines(ctx, names, start_ts=None, end_ts=None, limit=20, offset=0): +def _collect_chat_history_lines(ctx, names, start_ts=None, end_ts=None, limit=20, offset=0, oldest_first=False): collected = [] failures = [] candidate_limit = _candidate_page_size(limit, offset) @@ -1022,6 +1025,7 @@ def _collect_chat_history_lines(ctx, names, start_ts=None, end_ts=None, limit=20 end_ts=end_ts, limit=batch_size, offset=fetch_offset, + oldest_first=oldest_first, ) if not rows: break @@ -1042,7 +1046,7 @@ def _collect_chat_history_lines(ctx, names, start_ts=None, end_ts=None, limit=20 except Exception as e: failures.append(f"{table_ctx['db_path']}: {e}") - paged = _page_ranked_entries(collected, limit, offset) + paged = _page_ranked_entries(collected, limit, offset, oldest_first=oldest_first) return [line for _, line in paged], failures @@ -1348,7 +1352,7 @@ def get_recent_sessions(limit: int = 20) -> str: @mcp.tool() -def get_chat_history(chat_name: str, limit: int = 50, offset: int = 0, start_time: str = "", end_time: str = "") -> str: +def get_chat_history(chat_name: str, limit: int = 50, offset: int = 0, start_time: str = "", end_time: str = "", oldest_first: bool = False) -> str: """获取指定聊天的消息记录。 Args: @@ -1357,6 +1361,7 @@ def get_chat_history(chat_name: str, limit: int = 50, offset: int = 0, start_tim offset: 分页偏移量，默认0 start_time: 起始时间，支持 YYYY-MM-DD / YYYY-MM-DD HH:MM / YYYY-MM-DD HH:MM:SS end_time: 结束时间，支持 YYYY-MM-DD / YYYY-MM-DD HH:MM / YYYY-MM-DD HH:MM:SS + oldest_first: 为 True 时返回最早的消息（默认 False 返回最新消息） """ try: _validate_pagination(limit, offset, limit_max=None) @@ -1378,6 +1383,7 @@ def get_chat_history(chat_name: str, limit: int = 50, offset: int = 0, start_tim end_ts=end_ts, limit=limit, offset=offset, + oldest_first=oldest_first, ) if not lines: @@ -1734,5 +1740,207 @@ def get_chat_images(chat_name: str, limit: int = 20) -> str: return f"{display_name} 的 {len(lines)} 张图片:\n\n" + "\n".join(lines) +# ============ 语音解密 ============ + +DECODED_VOICE_DIR = os.path.join(SCRIPT_DIR, "decoded_voices") + +# media DB 与 message DB 同样会分片（media_0.db、media_1.db…）， +# 每个分片各有独立的 Name2Id / VoiceInfo 表。 +MEDIA_DB_KEYS = sorted([ + k for k in ALL_KEYS + if any(v.startswith("message/") for v in key_path_variants(k)) + and any(re.search(r"media_\d+\.db$", v) for v in key_path_variants(k)) +]) + + +def _iter_media_db_paths(): + for rel_key in MEDIA_DB_KEYS: + path = _cache.get(rel_key) + if path: + yield path + + +def _get_chat_name_id(conn, username): + row = conn.execute( + "SELECT rowid FROM Name2Id WHERE user_name = ?", (username,) + ).fetchone() + return row[0] if row else None + + +def _fetch_voice_row(username, local_id): + """遍历所有 media DB 分片，返回 (voice_data, create_time)；找不到返回 None。""" + for media_db in _iter_media_db_paths(): + with closing(sqlite3.connect(media_db)) as conn: + chat_name_id = _get_chat_name_id(conn, username) + if chat_name_id is None: + continue + row = conn.execute( + "SELECT voice_data, create_time FROM VoiceInfo " + "WHERE chat_name_id = ? AND local_id = ?", + (chat_name_id, local_id), + ).fetchone() + if row: + return row + return None + + +def _silk_to_wav(voice_data, create_time, username, local_id): + """Decode SILK voice blob to WAV file, return output path.""" + # pypi 上有多个 SILK 相关包名（silk-python / pysilk / pilk）， + # 这里用的是 synodriver/pysilk —— 安装包名 silk-python，import 名 pysilk + import pysilk + data = bytes(voice_data) + silk_data = data[1:] if data[0] == 0x02 else data + os.makedirs(DECODED_VOICE_DIR, exist_ok=True) + time_str = datetime.fromtimestamp(create_time).strftime('%Y%m%d_%H%M%S') + out_path = os.path.join(DECODED_VOICE_DIR, f"{username}_{time_str}_{local_id}.wav") + inp = io.BytesIO(silk_data) + out = io.BytesIO() + pysilk.decode(inp, out, 24000) + pcm = out.getvalue() + with wave.open(out_path, 'wb') as wf: + wf.setnchannels(1) + wf.setsampwidth(2) + wf.setframerate(24000) + wf.writeframes(pcm) + return out_path, len(pcm) + + +@mcp.tool() +def get_voice_messages(chat_name: str, limit: int = 20) -> str: + """列出某个聊天中的语音消息。 + + 返回语音的时间、local_id 和大小，可配合 decode_voice 工具解码。 + + Args: + chat_name: 聊天对象的名字、备注名或wxid + limit: 返回数量，默认20 + """ + username = resolve_username(chat_name) + if not username: + return f"找不到聊天对象: {chat_name}" + + names = get_contact_names() + display_name = names.get(username, username) + + if not MEDIA_DB_KEYS: + return "找不到 media DB" + + # 从每个分片各取最多 limit 条后合并再截断：分片若有时间重叠也不会漏最新消息 + rows = [] + for media_db in _iter_media_db_paths(): + with closing(sqlite3.connect(media_db)) as conn: + chat_name_id = _get_chat_name_id(conn, username) + if chat_name_id is None: + continue + rows.extend(conn.execute( + "SELECT local_id, create_time, length(voice_data) FROM VoiceInfo " + "WHERE chat_name_id = ? ORDER BY create_time DESC LIMIT ?", + (chat_name_id, limit), + ).fetchall()) + + if not rows: + return f"{display_name} 无语音消息" + + rows.sort(key=lambda r: r[1], reverse=True) + rows = rows[:limit] + + lines = [] + for local_id, create_time, size in rows: + time_str = datetime.fromtimestamp(create_time).strftime('%Y-%m-%d %H:%M') + lines.append(f"[{time_str}] local_id={local_id} {size/1024:.0f}KB") + + return f"{display_name} 的 {len(lines)} 条语音消息:\n\n" + "\n".join(lines) + + +@mcp.tool() +def decode_voice(chat_name: str, local_id: int) -> str: + """解码微信语音消息为 WAV 文件。 + + 先用 get_voice_messages 获取 local_id，再用此工具解码。 + 输出文件保存在 decoded_voices/ 目录。 + + 依赖: pip install silk-python (import 名为 pysilk) + + Args: + chat_name: 聊天对象的名字、备注名或wxid + local_id: 语音消息的 local_id（从 get_voice_messages 获取） + """ + try: + import pysilk # noqa: F401 + except ImportError: + return "缺少依赖: pip install silk-python" + + username = resolve_username(chat_name) + if not username: + return f"找不到聊天对象: {chat_name}" + + row = _fetch_voice_row(username, local_id) + if row is None: + return f"找不到 local_id={local_id} 的语音消息" + + voice_data, create_time = row + out_path, pcm_len = _silk_to_wav(voice_data, create_time, username, local_id) + duration_s = pcm_len / (24000 * 2) + return ( + f"解码成功!\n" + f" 文件: {out_path}\n" + f" 时长: {duration_s:.1f}秒\n" + f" 大小: {os.path.getsize(out_path):,} bytes" + ) + + +_whisper_model = None + +def _get_whisper_model(model_size="base"): + global _whisper_model + if _whisper_model is None: + import whisper + _whisper_model = whisper.load_model(model_size) + return _whisper_model + + +@mcp.tool() +def transcribe_voice(chat_name: str, local_id: int) -> str: + """将微信语音消息转录为文字（自动检测语言，保留原语言）。 + + 会先解码 SILK 语音为 WAV，再用 Whisper 转录。 + 首次运行会下载 Whisper 模型（约 145MB）。 + + 依赖: pip install silk-python openai-whisper (silk-python 的 import 名为 pysilk) + + Args: + chat_name: 聊天对象的名字、备注名或wxid + local_id: 语音消息的 local_id（从 get_voice_messages 获取） + """ + try: + import whisper # noqa: F401 + except ImportError: + return "缺少依赖: pip install openai-whisper" + try: + import pysilk # noqa: F401 + except ImportError: + return "缺少依赖: pip install silk-python" + + username = resolve_username(chat_name) + if not username: + return f"找不到聊天对象: {chat_name}" + + row = _fetch_voice_row(username, local_id) + if row is None: + return f"找不到 local_id={local_id} 的语音消息" + + voice_data, create_time = row + wav_path, _ = _silk_to_wav(voice_data, create_time, username, local_id) + + model = _get_whisper_model() + result = model.transcribe(wav_path) + lang = result.get("language", "unknown") + text = result.get("text", "").strip() + + time_label = datetime.fromtimestamp(create_time).strftime('%Y-%m-%d %H:%M') + return f"[{time_label}] ({lang})\n{text}" + + if __name__ == "__main__": mcp.run()