feat(attachment): support image extraction from local chat data (#57)

Cargo.lock

@@ -105,6 +105,12 @@ version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c08606f8c3cbf4ce6ec8e28fb0014a2c086708fe954eaa885384a6165172e7e8"
 
+[[package]]
+name = "base64"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
+
 [[package]]
 name = "bitflags"
 version = "2.11.1"
@@ -1311,6 +1317,7 @@ version = "0.1.11"
 dependencies = [
  "aes",
  "anyhow",
+ "base64",
  "cbc",
  "chrono",
  "clap",

Cargo.toml

@@ -50,6 +50,9 @@ dirs = "5"
 # MD5 (联系人表名 Msg_<md5>)
 md5 = "0.7"
 
+# 附件 ID 编码（base64url）
+base64 = "0.22"
+
 # 正则表达式
 regex = "1"
 roxmltree = "0.20"

README.md

@@ -211,6 +211,35 @@ wx biz-articles --json | jq '.[].url'             # 下游消费 URL
 
 每条返回：`account` / `account_username` / `title` / `url` / `digest` / `cover_url` / `time` / `timestamp` / `recv_time_str`。多图文推送会展开成多行。
 
+### 附件提取（图片）
+
+聊天里的附件本体存在 `xwechat_files/<wxid>/msg/attach/...` 下的 `.dat` 文件，需要按消息所在 `message_resource.db` 的 md5 + 平台相关 image key 解码才能拿到原图。
+
+```bash
+# 1) 列出会话里的图片附件，先拿到不透明的 attachment_id
+wx attachments "张三"
+wx attachments "AI群" --kind image -n 100
+wx attachments "AI群" --since 2026-04-01 --until 2026-04-15
+
+# 2) 把单个 attachment_id 解密写出去（扩展名建议保留 .jpg / .mp4 等）
+wx extract <attachment_id> -o ~/Desktop/photo.jpg
+wx extract <attachment_id> -o /tmp/x.jpg --overwrite
+```
+
+`attachments` 输出每条带：`attachment_id` / `kind` / `type` / `local_id` / `timestamp` / `time`，群聊里还有 `sender`。当前 `kind` 固定为 `image`；命令名保留成 `attachments` 是为了后续扩到其他附件类型时不 break CLI。
+
+`extract` 输出报告里带：`md5` / `dat_path` / `dat_size` / `output` / `output_size` / `format`（实际识别出的图片格式：jpg / png / gif / webp / hevc 等）/ `decoder`（实际选用的解码器：`legacy_xor` / `v1_aes` / `v2`）。
+
+支持的解码档位：
+- **legacy XOR**：早期单字节 XOR，无 magic（按文件首字节探测格式自动反推）
+- **V1 fixed-AES**（`07 08 V1 08 07`）：AES-128-ECB + 固定 key `cfcd208495d565ef`
+- **V2 AES + XOR**（`07 08 V2 08 07`）：AES-128-ECB + raw + XOR；AES key 平台派生
+
+V2 image key 提取：
+- **macOS**：`kvcomm` cache（`key_<uin>_*.statistic` 文件名取 uin → `md5(str(uin) + wxid)[:16]`）+ brute-force fallback（`md5(str(uin))[:4] == wxid_suffix` 枚举 2^24）；xor_key = `uin & 0xff`，**不是硬编码 0x88**
+- **Windows**：扫 `Weixin.exe` 内存匹配 `[A-Za-z0-9]{32|16}` 候选，按 V2 template ciphertext-block 反验
+- **Linux**：上游空白，遇到 V2 .dat 会报 unsupported
+
 ### 联系人 & 群组
 
 ```bash

SKILL.md

@@ -242,6 +242,34 @@ wx biz-articles --since 2026-05-10 --json | jq '.[].url'
 
 每条返回的字段：`account` / `account_username`（`gh_*`）/ `title` / `url`（`mp.weixin.qq.com` 链接）/ `digest` / `cover_url` / `time` + `timestamp`（文章发布时间）/ `recv_time_str` + `recv_time`（微信接收推送的时间）。多图文推送会展开为多行。
 
+### 附件提取（图片）
+
+聊天里的图片本体在 `xwechat_files/<wxid>/msg/attach/...` 下加密存储（`.dat`），需要按消息所在 `message_resource.db` 的 md5 + 平台相关 image key 才能解码。两步走：
+
+```bash
+# 1) 先列出图片附件，拿到不透明的 attachment_id
+wx attachments "张三"
+wx attachments "AI群" --kind image -n 100
+wx attachments "AI群" --since 2026-04-01 --until 2026-04-15
+
+# 2) 用 attachment_id 把单个资源解密写到指定路径
+wx extract <attachment_id> -o ~/Desktop/photo.jpg
+wx extract <attachment_id> -o /tmp/x.jpg --overwrite
+```
+
+`attachments` 输出每条带：`attachment_id` / `kind`（当前固定 `image`）/ `type` / `local_id` / `timestamp` / `time`，群聊里另带 `sender`。命令名保留成 `attachments` 是为了后续扩到其他附件类型时不 break CLI。
+
+`extract` 报告里带：`md5` / `dat_path` / `dat_size` / `output` / `output_size` / `format`（实际识别出的图片格式：jpg / png / gif / webp / hevc 等）/ `decoder`（实际选用的解码器：`legacy_xor` / `v1_aes` / `v2`）。
+
+支持的解码档位：
+- **legacy XOR**：早期单字节 XOR，无 magic（按文件首字节探测格式自动反推）
+- **V1 fixed-AES**（`07 08 V1 08 07`）：AES-128-ECB + 固定 key `cfcd208495d565ef`
+- **V2 AES + XOR**（`07 08 V2 08 07`）：AES-128-ECB + raw + XOR；AES key 平台派生
+
+V2 image key 提取（macOS / Windows 自动；Linux 暂不支持）：
+- macOS：`kvcomm` cache（`key_<uin>_*.statistic` 文件名取 uin → `md5(str(uin) + wxid)[:16]`）+ brute-force fallback；`xor_key = uin & 0xff`
+- Windows：扫 `Weixin.exe` 内存匹配 `[A-Za-z0-9]{32|16}` 候选，按 V2 template ciphertext-block 反验
+
 ### 收藏与统计
 
 ```bash

src/attachment/attachment_id.rs

@@ -0,0 +1,153 @@
+//! 不透明附件 ID — 跨 CLI / IPC 的圆 trip 句柄。
+//!
+//! 编码：`base64url_no_pad(serde_json(payload))`。
+//! 选择 base64url(json) 而不是紧凑 bit-pack：
+//! - phase 1 求稳，不发明二进制协议
+//! - 后面加字段（`resource_md5` / `decoder_hint` 之类）老 CLI 不 break
+//! - debug 直接 base64 -d | jq 看字段
+//!
+//! ⚠️ `local_id` 在同一 chat 内会被 WeChat 复用（实测同 chat 最多 7 条同 local_id），
+//! 所以 `(chat, local_id, create_time)` 三元组才是定位资源行的最小集。
+
+use anyhow::{anyhow, Context, Result};
+use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum AttachmentKind {
+    Image,
+    Video,
+    File,
+    Voice,
+}
+
+impl AttachmentKind {
+    /// 从 message.local_type 推 attachment kind（只覆盖 phase 1 关心的几种）。
+    /// 高 32 bit 是版本/会话 flag，要先 mask 到低 32 bit。
+    pub fn from_local_type(local_type: i64) -> Option<Self> {
+        let lo = (local_type as u64) & 0xFFFF_FFFF;
+        match lo {
+            3 => Some(AttachmentKind::Image),
+            34 => Some(AttachmentKind::Voice),
+            43 => Some(AttachmentKind::Video),
+            // type=49 是 appmsg，里面 subtype=6 才是文件；这里偏宽松返回 File，
+            // 由 resolver 进一步根据 appmsg subtype 决定是否真的能 extract
+            49 => Some(AttachmentKind::File),
+            _ => None,
+        }
+    }
+
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            AttachmentKind::Image => "image",
+            AttachmentKind::Video => "video",
+            AttachmentKind::File => "file",
+            AttachmentKind::Voice => "voice",
+        }
+    }
+}
+
+/// 附件 ID payload（序列化后 base64url 编码）。
+///
+/// `v` 是版本字段，将来 schema 变了可以走分支兼容。当前 v=1。
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AttachmentId {
+    /// payload schema version
+    pub v: u32,
+    /// 会话 username（同时用于 ChatName2Id 查 chat_id 和拼 attach 路径）
+    pub chat: String,
+    /// 消息行的 local_id
+    pub local_id: i64,
+    /// 消息行的 create_time（unix 秒）— 用于 disambiguate 同 chat 内 local_id 复用
+    pub create_time: i64,
+    /// 附件类别
+    pub kind: AttachmentKind,
+    /// 可选 hint：消息所在 message_N.db 的 N。给定时 resolver 可跳过 shard 扫描；
+    /// 缺省时 resolver 会按 `find_msg_tables` 逻辑全量扫
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub db: Option<u8>,
+}
+
+impl AttachmentId {
+    pub fn encode(&self) -> Result<String> {
+        let json = serde_json::to_vec(self).context("序列化 AttachmentId")?;
+        Ok(URL_SAFE_NO_PAD.encode(json))
+    }
+
+    pub fn decode(s: &str) -> Result<Self> {
+        let bytes = URL_SAFE_NO_PAD
+            .decode(s.trim())
+            .map_err(|e| anyhow!("attachment_id 不是合法 base64url: {}", e))?;
+        let id: AttachmentId =
+            serde_json::from_slice(&bytes).context("attachment_id payload 非合法 JSON")?;
+        if id.v != 1 {
+            return Err(anyhow!("不支持的 attachment_id 版本 v={}", id.v));
+        }
+        Ok(id)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn round_trip_minimal() {
+        let id = AttachmentId {
+            v: 1,
+            chat: "wxid_abc".to_string(),
+            local_id: 12345,
+            create_time: 1_715_678_901,
+            kind: AttachmentKind::Image,
+            db: None,
+        };
+        let s = id.encode().unwrap();
+        let back = AttachmentId::decode(&s).unwrap();
+        assert_eq!(back.chat, id.chat);
+        assert_eq!(back.local_id, id.local_id);
+        assert_eq!(back.create_time, id.create_time);
+        assert_eq!(back.kind, id.kind);
+        assert_eq!(back.db, id.db);
+    }
+
+    #[test]
+    fn round_trip_with_db_hint() {
+        let id = AttachmentId {
+            v: 1,
+            chat: "1234@chatroom".to_string(),
+            local_id: 42,
+            create_time: 1,
+            kind: AttachmentKind::Image,
+            db: Some(2),
+        };
+        let s = id.encode().unwrap();
+        assert!(!s.contains('=')); // base64url no-pad
+        let back = AttachmentId::decode(&s).unwrap();
+        assert_eq!(back.db, Some(2));
+    }
+
+    #[test]
+    fn local_type_mask_high_bits() {
+        // monitor_web.py 里 image push 路径：高位带 flag，低 32 bit 是 3
+        let high_flag = (0xDEAD_BEEFu64 << 32) as i64 | 3;
+        assert_eq!(
+            AttachmentKind::from_local_type(high_flag),
+            Some(AttachmentKind::Image)
+        );
+    }
+
+    #[test]
+    fn rejects_unknown_version() {
+        let id = AttachmentId {
+            v: 99,
+            chat: "x".to_string(),
+            local_id: 0,
+            create_time: 0,
+            kind: AttachmentKind::Image,
+            db: None,
+        };
+        let s = id.encode().unwrap();
+        assert!(AttachmentId::decode(&s).is_err());
+    }
+}

src/attachment/decoder/mod.rs

@@ -0,0 +1,122 @@
+//! `.dat` 文件解码：根据 6B header magic 分发到具体 decoder。
+//!
+//! 三档：
+//! | header[0..6]            | decoder           | 备注                                    |
+//! |-------------------------|-------------------|-----------------------------------------|
+//! | `07 08 V2 08 07`        | `v2`              | AES-128-ECB + XOR 混合，需要 image AES key |
+//! | `07 08 V1 08 07`        | `v1_aes`          | 固定 AES key `cfcd208495d565ef`         |
+//! | (其他, 通常无 magic)    | `v1_xor`          | legacy single-byte XOR，magic 自动探测  |
+//!
+//! 决策点放在 `dispatch`，让上层（`resolver` / CLI extract 命令）只跟一个入口打交道。
+
+use anyhow::{anyhow, Result};
+
+pub mod v1_xor;
+pub mod v2;
+
+/// 完整 V2 magic：`\x07\x08V2\x08\x07`
+pub const V2_MAGIC: [u8; 6] = [0x07, 0x08, b'V', b'2', 0x08, 0x07];
+/// 完整 V1 magic：`\x07\x08V1\x08\x07`
+pub const V1_MAGIC: [u8; 6] = [0x07, 0x08, b'V', b'1', 0x08, 0x07];
+
+/// 解码后的产物 + 探测出的图片格式
+#[derive(Debug)]
+pub struct DecodedImage {
+    pub data: Vec<u8>,
+    /// 推断出的图片扩展名（不带点），由 magic 决定。例如 "jpg" / "png" / "gif" / "webp" /
+    /// "tif" / "bmp" / "hevc"（wxgf 容器）/ "bin"（未识别）
+    pub format: &'static str,
+    /// 解码器名称（"legacy_xor" / "v1_aes" / "v2"），用于 CLI 调试输出
+    pub decoder: &'static str,
+}
+
+/// 由 caller 提供的 V2 image AES key（codex 的 `image_key` 模块负责拿到）。
+/// 缺省时遇到 V2 文件会返回 `Err`，caller 可以拿到具体错误信息再处理。
+#[derive(Debug, Clone, Copy, Default)]
+pub struct V2KeyMaterial<'a> {
+    pub aes_key: Option<&'a [u8; 16]>,
+    /// XOR key — WeChat 4.x 默认 0x88，可 override
+    pub xor_key: u8,
+}
+
+impl<'a> V2KeyMaterial<'a> {
+    pub fn with_aes(key: &'a [u8; 16]) -> Self {
+        Self { aes_key: Some(key), xor_key: 0x88 }
+    }
+}
+
+/// 根据 `dat_bytes` 头部 magic 自动分发到对应 decoder。
+///
+/// `v2_key` 仅在文件是 V2 magic 时被消费。
+pub fn dispatch(dat_bytes: &[u8], v2_key: V2KeyMaterial<'_>) -> Result<DecodedImage> {
+    if dat_bytes.len() >= 6 {
+        let head: &[u8; 6] = dat_bytes[..6].try_into().unwrap();
+        if head == &V2_MAGIC {
+            return v2::decode(dat_bytes, v2_key);
+        }
+        if head == &V1_MAGIC {
+            // V1 fixed-AES: 固定 key = md5("0")[:16] = "cfcd208495d565ef"
+            let fixed_key: [u8; 16] = *b"cfcd208495d565ef";
+            return v2::decode(
+                dat_bytes,
+                V2KeyMaterial { aes_key: Some(&fixed_key), xor_key: v2_key.xor_key },
+            )
+            .map(|mut d| {
+                d.decoder = "v1_aes";
+                d
+            });
+        }
+    }
+    if dat_bytes.is_empty() {
+        return Err(anyhow!("空 .dat 文件"));
+    }
+    v1_xor::decode(dat_bytes)
+}
+
+/// 从解密后的字节流头部探测图片格式扩展名。
+///
+/// 与上游 `decode_image.py::detect_image_format` 一致；新增 wxgf (HEVC 裸流) 的探测，
+/// 因为 V2 解码后产物可能直接是 wxgf 容器。
+pub fn detect_image_format(bytes: &[u8]) -> &'static str {
+    if bytes.len() >= 4 && &bytes[..4] == b"wxgf" {
+        return "hevc";
+    }
+    if bytes.len() >= 3 && bytes[..3] == [0xFF, 0xD8, 0xFF] {
+        return "jpg";
+    }
+    if bytes.len() >= 4 && bytes[..4] == [0x89, 0x50, 0x4E, 0x47] {
+        return "png";
+    }
+    if bytes.len() >= 3 && &bytes[..3] == b"GIF" {
+        return "gif";
+    }
+    if bytes.len() >= 12 && &bytes[..4] == b"RIFF" && &bytes[8..12] == b"WEBP" {
+        return "webp";
+    }
+    if bytes.len() >= 4 && bytes[..4] == [0x49, 0x49, 0x2A, 0x00] {
+        return "tif";
+    }
+    if bytes.len() >= 2 && &bytes[..2] == b"BM" {
+        return "bmp";
+    }
+    "bin"
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn detect_basic_formats() {
+        assert_eq!(detect_image_format(&[0xFF, 0xD8, 0xFF, 0xE0]), "jpg");
+        assert_eq!(detect_image_format(&[0x89, 0x50, 0x4E, 0x47]), "png");
+        assert_eq!(detect_image_format(b"GIF89a"), "gif");
+        assert_eq!(detect_image_format(b"BM\0\0\0\0\0\0\0\0\0\0\0\0"), "bmp");
+        let mut webp = b"RIFF\0\0\0\0WEBP".to_vec();
+        webp.extend_from_slice(&[0; 4]);
+        assert_eq!(detect_image_format(&webp), "webp");
+        assert_eq!(detect_image_format(&[0x49, 0x49, 0x2A, 0x00]), "tif");
+        assert_eq!(detect_image_format(b"wxgfXXXX"), "hevc");
+        assert_eq!(detect_image_format(&[0, 0, 0, 0]), "bin");
+    }
+}

src/attachment/decoder/v1_xor.rs

@@ -0,0 +1,166 @@
+//! Legacy single-byte XOR decoder（无 magic 头的旧 .dat）
+//!
+//! 算法：用已知图片 magic 反推 XOR key —— `key = file[0] ^ magic[0]`。
+//! 然后用同一个 key 校验 `file[i] ^ key == magic[i]`，全部命中才接受这个 key。
+//!
+//! 优先级（按 magic 长度降序，避免短 magic 假阳性）：
+//!   PNG (4) > GIF (4) > TIF (4) > WEBP (4, RIFF) > JPG (3) > BMP (2, 需额外校验)
+//!
+//! BMP 只有 2 字节 magic，假阳性高；额外用 BMP file header 里的
+//! `bf_size`（offset 2, u32 LE）和 `bf_offset`（offset 10, u32 LE）做合理性校验：
+//!   - `|bf_size - file_size| < 1024`（允许微小 padding 差）
+//!   - `14 <= bf_offset <= 1078`（最大调色板 256*4 + header 14 = 1038，留点余量）
+
+use anyhow::{anyhow, Result};
+
+use super::{detect_image_format, DecodedImage};
+
+const PNG: &[u8] = &[0x89, 0x50, 0x4E, 0x47];
+const GIF: &[u8] = &[0x47, 0x49, 0x46, 0x38];
+const TIF: &[u8] = &[0x49, 0x49, 0x2A, 0x00];
+const WEBP_RIFF: &[u8] = &[0x52, 0x49, 0x46, 0x46];
+const JPG: &[u8] = &[0xFF, 0xD8, 0xFF];
+const BMP: &[u8] = &[0x42, 0x4D];
+
+/// 在 `header` 上尝试一个固定 magic：返回 `Some(key)` 当且仅当所有字节都对得上。
+fn try_magic(header: &[u8], magic: &[u8]) -> Option<u8> {
+    if header.len() < magic.len() {
+        return None;
+    }
+    let key = header[0] ^ magic[0];
+    for i in 1..magic.len() {
+        if header[i] ^ key != magic[i] {
+            return None;
+        }
+    }
+    Some(key)
+}
+
+/// 探测 XOR key。失败返回 `None`（caller 决定是不是错）。
+pub fn detect_key(file_bytes: &[u8]) -> Option<u8> {
+    if file_bytes.len() < 4 {
+        return None;
+    }
+    let header = &file_bytes[..file_bytes.len().min(16)];
+
+    // 先试 3+ 字节 magic
+    for magic in [PNG, GIF, TIF, WEBP_RIFF, JPG] {
+        if let Some(k) = try_magic(header, magic) {
+            return Some(k);
+        }
+    }
+
+    // 最后试 BMP（只有 2B magic，需额外校验）
+    if let Some(k) = try_magic(header, BMP) {
+        if header.len() >= 14 {
+            // 解 BMP file header 14 字节
+            let mut dec = [0u8; 14];
+            for i in 0..14 {
+                dec[i] = header[i] ^ k;
+            }
+            let bmp_size = u32::from_le_bytes([dec[2], dec[3], dec[4], dec[5]]);
+            let bmp_offset = u32::from_le_bytes([dec[10], dec[11], dec[12], dec[13]]);
+            let file_size = file_bytes.len() as u32;
+            // 允许 1024 字节 padding 差；offset 在合理范围
+            if file_size.abs_diff(bmp_size) < 1024 && (14..=1078).contains(&bmp_offset) {
+                return Some(k);
+            }
+        }
+    }
+
+    None
+}
+
+/// XOR 解码整个 `.dat` 内容。
+pub fn decode(file_bytes: &[u8]) -> Result<DecodedImage> {
+    let key =
+        detect_key(file_bytes).ok_or_else(|| anyhow!("legacy XOR: 无法识别图片 magic（key 探测失败）"))?;
+    let data: Vec<u8> = file_bytes.iter().map(|b| b ^ key).collect();
+    let format = detect_image_format(&data);
+    if format == "bin" {
+        return Err(anyhow!("legacy XOR: 解出 key=0x{:02x} 但产物 magic 不识别", key));
+    }
+    Ok(DecodedImage { data, format, decoder: "legacy_xor" })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// 把一段 plaintext 用单字节 key XOR 加密，模拟 .dat 文件
+    fn xor_encrypt(plain: &[u8], key: u8) -> Vec<u8> {
+        plain.iter().map(|b| b ^ key).collect()
+    }
+
+    #[test]
+    fn detect_jpg_key() {
+        let plain = vec![0xFF, 0xD8, 0xFF, 0xE0, 0x00, 0x10, 0x4A, 0x46, 0x49, 0x46];
+        let enc = xor_encrypt(&plain, 0x3C);
+        assert_eq!(detect_key(&enc), Some(0x3C));
+    }
+
+    #[test]
+    fn detect_png_key() {
+        let mut plain = vec![0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A];
+        plain.extend_from_slice(&[0; 16]);
+        let enc = xor_encrypt(&plain, 0xA5);
+        assert_eq!(detect_key(&enc), Some(0xA5));
+    }
+
+    #[test]
+    fn detect_gif_key() {
+        let mut plain = b"GIF89a".to_vec();
+        plain.extend_from_slice(&[0; 16]);
+        let enc = xor_encrypt(&plain, 0x77);
+        assert_eq!(detect_key(&enc), Some(0x77));
+    }
+
+    #[test]
+    fn detect_webp_riff_key() {
+        let mut plain = b"RIFF\x00\x00\x00\x00WEBP".to_vec();
+        plain.extend_from_slice(&[0; 8]);
+        let enc = xor_encrypt(&plain, 0x12);
+        assert_eq!(detect_key(&enc), Some(0x12));
+    }
+
+    #[test]
+    fn detect_tif_key() {
+        let mut plain = vec![0x49, 0x49, 0x2A, 0x00, 0x08, 0x00, 0x00, 0x00];
+        plain.extend_from_slice(&[0; 16]);
+        let enc = xor_encrypt(&plain, 0xC3);
+        assert_eq!(detect_key(&enc), Some(0xC3));
+    }
+
+    #[test]
+    fn detect_bmp_with_valid_header() {
+        // BMP 14B header: 'BM' + size(u32 LE) + reserved(2*u16) + offset(u32 LE)
+        let mut plain = Vec::new();
+        plain.extend_from_slice(b"BM");
+        plain.extend_from_slice(&100u32.to_le_bytes()); // file_size = 100
+        plain.extend_from_slice(&[0; 4]); // reserved
+        plain.extend_from_slice(&54u32.to_le_bytes()); // pixel data offset = 54
+        plain.resize(100, 0); // 整个文件 100 字节，匹配 file_size
+        let enc = xor_encrypt(&plain, 0x55);
+        assert_eq!(detect_key(&enc), Some(0x55));
+    }
+
+    #[test]
+    fn reject_random_bytes() {
+        // 全 0 文件：BMP 检测会算出 key = 0x42 ^ 0 = 0x42，
+        // 但解密出的 BMP file_size = 0 vs file_size = 100，差距 > 1024 →
+        // 应该 reject
+        let bytes = vec![0u8; 100];
+        assert_eq!(detect_key(&bytes), None);
+    }
+
+    #[test]
+    fn decode_round_trip_jpg() {
+        let mut plain = vec![0xFF, 0xD8, 0xFF, 0xE0];
+        plain.extend_from_slice(b"JFIF padding here");
+        let enc = xor_encrypt(&plain, 0xAB);
+        let out = decode(&enc).unwrap();
+        assert_eq!(out.format, "jpg");
+        assert_eq!(out.decoder, "legacy_xor");
+        assert_eq!(out.data, plain);
+    }
+}

src/attachment/decoder/v2.rs

@@ -0,0 +1,130 @@
+//! V2 .dat 解码：`AES-128-ECB(PKCS7) + raw + XOR` 三段拼接。
+//!
+//! 文件结构（来自上游 `decode_image.py::v2_decrypt_file`）：
+//!   `[6B magic V2/V1] [4B aes_size LE] [4B xor_size LE] [1B padding]`
+//!   `[aligned_aes_size bytes AES-ECB ciphertext]`
+//!   `[len - aligned_aes_size - xor_size bytes raw_data (不加密)]`
+//!   `[xor_size bytes XOR (单字节 key)]`
+//!
+//! `aligned_aes_size`：把 `aes_size` 向上对齐到 16 的倍数；当 `aes_size` 本身是
+//! 16 的倍数时，PKCS7 还会再加一整块 padding，所以再 +16。等价于
+//! `aes_size + (16 - aes_size % 16)`。
+//!
+//! ⚠️ 此模块由 codex 落地完整 V2 实现 + image key 模块。当前只提供一个
+//! `decode` 入口骨架，方便 v1_aes 路径（固定 key）和 dispatch 一起编译过。
+//! `aes_key=None` 时返回带具体诊断信息的错误。
+
+use anyhow::{anyhow, bail, Result};
+
+use super::{detect_image_format, DecodedImage, V2KeyMaterial, V1_MAGIC, V2_MAGIC};
+
+const HEADER_SIZE: usize = 15;
+
+pub fn decode(file_bytes: &[u8], key: V2KeyMaterial<'_>) -> Result<DecodedImage> {
+    if file_bytes.len() < HEADER_SIZE {
+        bail!("V2 .dat: 文件过短（{} < {} 字节）", file_bytes.len(), HEADER_SIZE);
+    }
+    let magic: &[u8; 6] = file_bytes[..6].try_into().unwrap();
+    if magic != &V2_MAGIC && magic != &V1_MAGIC {
+        bail!("V2 .dat: header magic 不匹配 V1/V2");
+    }
+
+    let aes_key = key.aes_key.ok_or_else(|| {
+        anyhow!("V2 .dat: 需要 image AES key（codex 的 image_key 模块尚未填充）")
+    })?;
+
+    let aes_size = u32::from_le_bytes(file_bytes[6..10].try_into().unwrap()) as usize;
+    let xor_size = u32::from_le_bytes(file_bytes[10..14].try_into().unwrap()) as usize;
+
+    // PKCS7 对齐：aes_size 不是 16 的倍数 → 向上对齐；是 16 的倍数 → 再加一整块
+    let aligned_aes_size = aes_size + (16 - (aes_size % 16));
+
+    let aes_end = HEADER_SIZE.checked_add(aligned_aes_size).ok_or_else(|| anyhow!("aes 段长度溢出"))?;
+    if aes_end > file_bytes.len() {
+        bail!(
+            "V2 .dat: 头部宣称 aes_size={} (aligned={}) 超过文件长度 {}",
+            aes_size,
+            aligned_aes_size,
+            file_bytes.len()
+        );
+    }
+    let raw_end = file_bytes.len().checked_sub(xor_size).ok_or_else(|| {
+        anyhow!("V2 .dat: 头部宣称 xor_size={} 超过文件长度 {}", xor_size, file_bytes.len())
+    })?;
+    if aes_end > raw_end {
+        bail!(
+            "V2 .dat: aes_end={} > raw_end={}（aes/xor 段重叠）",
+            aes_end,
+            raw_end
+        );
+    }
+
+    // === AES-128-ECB 解密 + PKCS7 unpad ===
+    let aes_data = &file_bytes[HEADER_SIZE..aes_end];
+    let dec_aes = aes_ecb_decrypt_pkcs7(aes_key, aes_data)?;
+
+    // === Raw 段（未加密） ===
+    let raw_data = &file_bytes[aes_end..raw_end];
+
+    // === XOR 段 ===
+    let xor_data: Vec<u8> = file_bytes[raw_end..].iter().map(|b| b ^ key.xor_key).collect();
+
+    let mut out = Vec::with_capacity(dec_aes.len() + raw_data.len() + xor_data.len());
+    out.extend_from_slice(&dec_aes);
+    out.extend_from_slice(raw_data);
+    out.extend_from_slice(&xor_data);
+
+    let format = detect_image_format(&out);
+    if format == "bin" {
+        bail!("V2 .dat: AES 解密成功但产物 magic 不识别（key 可能错）");
+    }
+    Ok(DecodedImage { data: out, format, decoder: "v2" })
+}
+
+/// AES-128-ECB 解密 + PKCS7 unpad。失败时返回 `Err`，不返回半结果。
+///
+/// 不引第三方 ECB 包；ECB 本身就是 block-by-block，手工跑就行。
+/// PKCS7 padding 由本函数最后一段做 strict 校验：长度 1..=16，且尾部全是同值字节。
+fn aes_ecb_decrypt_pkcs7(key: &[u8; 16], cipher: &[u8]) -> Result<Vec<u8>> {
+    use aes::cipher::{generic_array::GenericArray, BlockDecrypt, KeyInit};
+    if cipher.is_empty() || cipher.len() % 16 != 0 {
+        bail!("AES 输入长度 {} 不是 16 的倍数", cipher.len());
+    }
+    let aes = aes::Aes128::new(key.into());
+    let mut out = Vec::with_capacity(cipher.len());
+    for chunk in cipher.chunks_exact(16) {
+        let mut block = GenericArray::clone_from_slice(chunk);
+        aes.decrypt_block(&mut block);
+        out.extend_from_slice(&block);
+    }
+    let pad = *out.last().ok_or_else(|| anyhow!("AES PKCS7: 空输出"))? as usize;
+    if pad == 0 || pad > 16 || pad > out.len() {
+        bail!("AES PKCS7: 非法 padding 长度 {}", pad);
+    }
+    let tail = &out[out.len() - pad..];
+    if !tail.iter().all(|&b| b as usize == pad) {
+        bail!("AES PKCS7: padding 字节不一致");
+    }
+    out.truncate(out.len() - pad);
+    Ok(out)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn rejects_short_file() {
+        let r = decode(&[0u8; 4], V2KeyMaterial::default());
+        assert!(r.is_err());
+    }
+
+    #[test]
+    fn rejects_v2_without_key() {
+        let mut buf = V2_MAGIC.to_vec();
+        buf.extend_from_slice(&[0u8; HEADER_SIZE - 6]);
+        let r = decode(&buf, V2KeyMaterial::default());
+        let err = r.unwrap_err().to_string();
+        assert!(err.contains("AES key"), "{}", err);
+    }
+}

src/attachment/image_key/linux.rs

@@ -0,0 +1,11 @@
+use anyhow::{bail, Result};
+
+use super::{ImageKeyMaterial, ImageKeyProvider};
+
+pub struct LinuxImageKeyProvider;
+
+impl ImageKeyProvider for LinuxImageKeyProvider {
+    fn get_key(&self, _wxid: &str) -> Result<ImageKeyMaterial> {
+        bail!("Linux V2 图片 key 当前未实现；请先用 legacy/V1 图片或在 README 中标注 unsupported")
+    }
+}

src/attachment/image_key/macos.rs

@@ -0,0 +1,423 @@
+//! macOS V2 image AES key 提取。
+//!
+//! 主路径：从 `key_<uin>_*.statistic` 文件名拿 uin，然后
+//! `md5(str(uin) + normalize(wxid)).hex()[:16]` 派生 AES key。
+//!
+//! fallback：通过 `md5(str(uin))[:4] == wxid_suffix` + `uin & 0xff == xor_key`
+//! 把搜索空间压到 2^24，再用 V2 模板反验 AES key。
+
+use anyhow::{bail, Context, Result};
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{mpsc, Arc, Mutex};
+
+use crate::config;
+
+use super::{
+    attach_root_for_db_dir, configured_db_dir_for_wxid, derive_xor_key_from_v2_dat,
+    find_v2_template_ciphertexts, join_components, normalize_wxid, verify_aes_key, wxid_from_db_dir,
+    ImageKeyMaterial, ImageKeyProvider,
+};
+
+pub struct MacosImageKeyProvider {
+    configured_db_dir: Result<PathBuf, String>,
+    cache: Mutex<HashMap<String, ImageKeyMaterial>>,
+}
+
+impl MacosImageKeyProvider {
+    pub fn from_current_config() -> Self {
+        let configured_db_dir = config::load_config()
+            .map(|cfg| cfg.db_dir)
+            .map_err(|err| err.to_string());
+        Self {
+            configured_db_dir,
+            cache: Mutex::new(HashMap::new()),
+        }
+    }
+}
+
+impl ImageKeyProvider for MacosImageKeyProvider {
+    fn get_key(&self, wxid: &str) -> Result<ImageKeyMaterial> {
+        let cache_key = normalize_wxid(wxid);
+        if let Some(found) = self.cache.lock().unwrap().get(&cache_key).copied() {
+            return Ok(found);
+        }
+
+        let configured_db_dir = self
+            .configured_db_dir
+            .as_ref()
+            .map_err(|err| anyhow::anyhow!("读取 config.db_dir 失败: {}", err))?;
+        let db_dir = configured_db_dir_for_wxid(configured_db_dir, wxid);
+        let attach_dir = attach_root_for_db_dir(&db_dir);
+        let key = derive_key_for_paths(&db_dir, &attach_dir)?;
+        self.cache.lock().unwrap().insert(cache_key, key);
+        Ok(key)
+    }
+}
+
+fn derive_key_for_paths(db_dir: &Path, attach_dir: &Path) -> Result<ImageKeyMaterial> {
+    let templates = find_v2_template_ciphertexts(attach_dir, 3, 64)?;
+    if templates.is_empty() {
+        bail!("在 {} 下找不到 V2 模板文件", attach_dir.display());
+    }
+
+    if let Some(found) = find_via_kvcomm(db_dir, &templates)? {
+        return Ok(found);
+    }
+
+    let (wxid_full, wxid_norm, suffix) =
+        extract_wxid_parts(db_dir).context("db_dir 不含可用于 fallback 的 wxid 4 位后缀")?;
+    let (xor_key, _votes, _total) = derive_xor_key_from_v2_dat(attach_dir, 10, 3)?
+        .context("V2 .dat 样本不足，无法投票反推 xor_key")?;
+
+    for wxid in preferred_wxid_candidates(&wxid_full, &wxid_norm) {
+        if let Some(aes_key) = bruteforce_aes_key(xor_key, &suffix, wxid, &templates)? {
+            return Ok(ImageKeyMaterial { aes_key, xor_key });
+        }
+    }
+
+    bail!("macOS V2 图片 key 派生失败")
+}
+
+fn find_via_kvcomm(db_dir: &Path, templates: &[[u8; 16]]) -> Result<Option<ImageKeyMaterial>> {
+    let Some(kvcomm_dir) = find_existing_kvcomm_dir(db_dir) else {
+        return Ok(None);
+    };
+
+    let codes = collect_kvcomm_codes(&kvcomm_dir)?;
+    if codes.is_empty() {
+        return Ok(None);
+    }
+    let wxids = collect_wxid_candidates(db_dir);
+    if wxids.is_empty() {
+        return Ok(None);
+    }
+
+    for wxid in wxids {
+        for code in &codes {
+            let candidate = derive_image_key_material(*code, &wxid);
+            if verify_aes_key(&candidate.aes_key, templates) {
+                return Ok(Some(candidate));
+            }
+        }
+    }
+    Ok(None)
+}
+
+fn derive_image_key_material(code: u32, wxid: &str) -> ImageKeyMaterial {
+    let xor_key = (code & 0xFF) as u8;
+    let digest = format!("{:x}", md5::compute(format!("{}{}", code, wxid)));
+    let mut aes_key = [0u8; 16];
+    aes_key.copy_from_slice(&digest.as_bytes()[..16]);
+    ImageKeyMaterial { aes_key, xor_key }
+}
+
+fn collect_wxid_candidates(db_dir: &Path) -> Vec<String> {
+    let Some(raw) = wxid_from_db_dir(db_dir) else {
+        return Vec::new();
+    };
+    let mut out = vec![raw.clone()];
+    let normalized = normalize_wxid(&raw);
+    if normalized != raw {
+        out.push(normalized);
+    }
+    out
+}
+
+fn extract_wxid_parts(db_dir: &Path) -> Option<(String, String, String)> {
+    let raw = wxid_from_db_dir(db_dir)?;
+    let idx = raw.rfind('_')?;
+    let suffix = &raw[idx + 1..];
+    if suffix.len() != 4 || !suffix.bytes().all(|byte| byte.is_ascii_hexdigit()) {
+        return None;
+    }
+    Some((raw.clone(), normalize_wxid(&raw), suffix.to_ascii_lowercase()))
+}
+
+fn preferred_wxid_candidates<'a>(raw: &'a str, normalized: &'a str) -> Vec<&'a str> {
+    if raw == normalized {
+        vec![raw]
+    } else {
+        vec![normalized, raw]
+    }
+}
+
+fn derive_kvcomm_dir_candidates(db_dir: &Path) -> Vec<PathBuf> {
+    let parts: Vec<String> = db_dir
+        .components()
+        .map(|component| component.as_os_str().to_string_lossy().into_owned())
+        .collect();
+
+    let mut candidates = Vec::new();
+    if let Some(idx) = parts.iter().position(|part| part == "xwechat_files") {
+        let documents_root = join_components(&parts[..idx]);
+        candidates.push(documents_root.join("app_data/net/kvcomm"));
+        candidates.push(documents_root.join("xwechat/net/kvcomm"));
+        if idx >= 1 {
+            let container_root = join_components(&parts[..idx - 1]);
+            candidates.push(
+                container_root
+                    .join("Library/Application Support/com.tencent.xinWeChat/xwechat/net/kvcomm"),
+            );
+            candidates.push(
+                container_root.join("Library/Application Support/com.tencent.xinWeChat/net/kvcomm"),
+            );
+        }
+    }
+    if let Some(home) = dirs::home_dir() {
+        candidates.push(
+            home.join("Library/Containers/com.tencent.xinWeChat/Data/Documents/app_data/net/kvcomm"),
+        );
+    }
+
+    let mut dedup = Vec::new();
+    for candidate in candidates {
+        if !dedup.contains(&candidate) {
+            dedup.push(candidate);
+        }
+    }
+    dedup
+}
+
+fn find_existing_kvcomm_dir(db_dir: &Path) -> Option<PathBuf> {
+    derive_kvcomm_dir_candidates(db_dir)
+        .into_iter()
+        .find(|path| path.is_dir())
+}
+
+fn collect_kvcomm_codes(kvcomm_dir: &Path) -> Result<Vec<u32>> {
+    let mut codes = std::collections::BTreeSet::new();
+    for entry in std::fs::read_dir(kvcomm_dir)? {
+        let entry = entry?;
+        let Some(name) = entry.file_name().to_str().map(|value| value.to_string()) else {
+            continue;
+        };
+        let Some(rest) = name.strip_prefix("key_") else {
+            continue;
+        };
+        let Some((code, _)) = rest.split_once('_') else {
+            continue;
+        };
+        if let Ok(code) = code.parse::<u32>() {
+            codes.insert(code);
+        }
+    }
+    Ok(codes.into_iter().collect())
+}
+
+fn bruteforce_aes_key(
+    xor_key: u8,
+    suffix_hex: &str,
+    wxid: &str,
+    templates: &[[u8; 16]],
+) -> Result<Option<[u8; 16]>> {
+    let suffix = hex_prefix_to_bytes(suffix_hex)?;
+    let workers = std::thread::available_parallelism()
+        .map(|count| count.get())
+        .unwrap_or(1)
+        .max(1);
+    let total = 1u32 << 24;
+    let chunk = total / workers as u32;
+    let stop = Arc::new(AtomicBool::new(false));
+    let (tx, rx) = mpsc::channel();
+    let wxid = Arc::new(wxid.as_bytes().to_vec());
+    let templates = Arc::new(templates.to_vec());
+
+    std::thread::scope(|scope| {
+        for idx in 0..workers {
+            let start = idx as u32 * chunk;
+            let end = if idx + 1 == workers {
+                total
+            } else {
+                (idx as u32 + 1) * chunk
+            };
+            let stop = Arc::clone(&stop);
+            let tx = tx.clone();
+            let wxid = Arc::clone(&wxid);
+            let templates = Arc::clone(&templates);
+            scope.spawn(move || {
+                for upper in start..end {
+                    if stop.load(Ordering::Relaxed) {
+                        break;
+                    }
+                    let uin = (upper << 8) | xor_key as u32;
+                    let uin_ascii = uin.to_string();
+                    let digest = md5::compute(uin_ascii.as_bytes());
+                    if digest.0[0] != suffix[0] || digest.0[1] != suffix[1] {
+                        continue;
+                    }
+
+                    let mut input = Vec::with_capacity(uin_ascii.len() + wxid.len());
+                    input.extend_from_slice(uin_ascii.as_bytes());
+                    input.extend_from_slice(&wxid);
+                    let aes_hex = format!("{:x}", md5::compute(input));
+                    let mut aes_key = [0u8; 16];
+                    aes_key.copy_from_slice(&aes_hex.as_bytes()[..16]);
+                    if verify_aes_key(&aes_key, &templates) {
+                        stop.store(true, Ordering::Relaxed);
+                        let _ = tx.send(aes_key);
+                        break;
+                    }
+                }
+            });
+        }
+    });
+    drop(tx);
+    Ok(rx.try_iter().next())
+}
+
+fn hex_prefix_to_bytes(hex: &str) -> Result<[u8; 2]> {
+    if hex.len() != 4 {
+        bail!("wxid suffix 不是 4 位 hex: {}", hex);
+    }
+    let hi = u8::from_str_radix(&hex[..2], 16)?;
+    let lo = u8::from_str_radix(&hex[2..], 16)?;
+    Ok([hi, lo])
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{derive_key_for_paths, find_existing_kvcomm_dir};
+    use super::collect_wxid_candidates;
+    use crate::attachment::image_key::normalize_wxid;
+    use aes::cipher::{generic_array::GenericArray, BlockEncrypt, KeyInit};
+    use aes::Aes128;
+    use std::fs;
+    use std::path::Path;
+
+    fn temp_dir(label: &str) -> std::path::PathBuf {
+        let mut dir = std::env::temp_dir();
+        dir.push(format!(
+            "wx-cli-image-key-macos-{}-{:?}",
+            label,
+            std::thread::current().id()
+        ));
+        let _ = fs::remove_dir_all(&dir);
+        fs::create_dir_all(&dir).unwrap();
+        dir
+    }
+
+    fn write_v2_template(path: &Path, aes_key: &[u8; 16], xor_key: u8, plaintext: &[u8; 16]) {
+        let cipher = Aes128::new(aes_key.into());
+        let mut block = GenericArray::clone_from_slice(plaintext);
+        cipher.encrypt_block(&mut block);
+
+        let mut data = Vec::new();
+        data.extend_from_slice(&crate::attachment::decoder::V2_MAGIC);
+        data.extend_from_slice(&0u32.to_le_bytes());
+        data.extend_from_slice(&0u32.to_le_bytes());
+        data.push(0);
+        data.extend_from_slice(&block);
+        data.push(0);
+        data.push(0xD9 ^ xor_key);
+        fs::create_dir_all(path.parent().unwrap()).unwrap();
+        fs::write(path, data).unwrap();
+    }
+
+    #[test]
+    fn normalize_wxid_matches_expected_shapes() {
+        assert_eq!(normalize_wxid("wxid_abc_def"), "wxid_abc");
+        assert_eq!(normalize_wxid("your_wxid_a1b2"), "your_wxid");
+        assert_eq!(normalize_wxid("plain"), "plain");
+    }
+
+    #[test]
+    fn kvcomm_path_detection_works() {
+        let dir = temp_dir("kvcomm");
+        let db_dir = dir.join(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/your_wxid_a1b2/db_storage",
+        );
+        let kvcomm = dir.join(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/app_data/net/kvcomm",
+        );
+        fs::create_dir_all(&db_dir).unwrap();
+        fs::create_dir_all(&kvcomm).unwrap();
+        assert_eq!(find_existing_kvcomm_dir(&db_dir), Some(kvcomm));
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn derives_key_via_kvcomm() {
+        let dir = temp_dir("via-kvcomm");
+        let db_dir = dir.join(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/your_wxid_a1b2/db_storage",
+        );
+        let attach = dir.join(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/your_wxid_a1b2/msg/attach/chat/2026-05/Img",
+        );
+        let kvcomm = dir.join(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/app_data/net/kvcomm",
+        );
+        fs::create_dir_all(&db_dir).unwrap();
+        fs::create_dir_all(&kvcomm).unwrap();
+        fs::write(kvcomm.join("key_42_x.statistic"), b"").unwrap();
+
+        let digest = format!("{:x}", md5::compute("42your_wxid"));
+        let mut aes_key = [0u8; 16];
+        aes_key.copy_from_slice(&digest.as_bytes()[..16]);
+        write_v2_template(
+            &attach.join("sample_t.dat"),
+            &aes_key,
+            42,
+            b"\xFF\xD8\xFFtemplate-001!",
+        );
+
+        let derived = derive_key_for_paths(&db_dir, db_dir.parent().unwrap().join("msg/attach").as_path())
+            .unwrap();
+        assert_eq!(derived.aes_key, aes_key);
+        assert_eq!(derived.xor_key, 42);
+
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn derives_key_via_bruteforce_fallback() {
+        let dir = temp_dir("via-fallback");
+        let suffix = format!("{:x}", md5::compute("42"))
+            .chars()
+            .take(4)
+            .collect::<String>();
+        let raw_wxid = format!("mywxid_{}", suffix);
+        let db_dir = dir.join(format!(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/{}/db_storage",
+            raw_wxid
+        ));
+        let attach = dir.join(format!(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/{}/msg/attach/chat/2026-05/Img",
+            raw_wxid
+        ));
+        fs::create_dir_all(&db_dir).unwrap();
+
+        let digest = format!("{:x}", md5::compute("42mywxid"));
+        let mut aes_key = [0u8; 16];
+        aes_key.copy_from_slice(&digest.as_bytes()[..16]);
+        for idx in 0..3 {
+            write_v2_template(
+                &attach.join(format!("sample{}_t.dat", idx)),
+                &aes_key,
+                42,
+                b"\xFF\xD8\xFFtemplate-001!",
+            );
+        }
+
+        let derived = derive_key_for_paths(&db_dir, db_dir.parent().unwrap().join("msg/attach").as_path())
+            .unwrap();
+        assert_eq!(derived.aes_key, aes_key);
+        assert_eq!(derived.xor_key, 42);
+
+        let _ = fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn collects_raw_and_normalized_wxid() {
+        let dir = temp_dir("wxid");
+        let db_dir = dir.join(
+            "Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/your_wxid_a1b2/db_storage",
+        );
+        fs::create_dir_all(&db_dir).unwrap();
+        let wxids = collect_wxid_candidates(&db_dir);
+        assert_eq!(wxids, vec!["your_wxid_a1b2".to_string(), "your_wxid".to_string()]);
+        let _ = fs::remove_dir_all(dir);
+    }
+}

src/attachment/image_key/mod.rs

@@ -0,0 +1,342 @@
+//! V2 image AES key 提取 — 平台相关。
+//!
+//! 路径：
+//! - macOS：磁盘派生（`key_<uin>_*.statistic` 文件名拿 uin → `md5(str(uin) + wxid)[:16]`）
+//!   + brute-force fallback（`md5(str(uin))[:4] == wxid_suffix` 枚举 2^24）
+//! - Windows：扫 `Weixin.exe` 内存，匹配 `[a-zA-Z0-9]{32}` 候选，按已知 AES ciphertext-block
+//!   反验（`find_image_key.py` / `find_image_key.c` 已写实）
+//! - Linux：上游空白；当前不实现，遇到 V2 .dat 返回 unsupported 错误
+
+#[cfg(target_os = "linux")]
+pub mod linux;
+#[cfg(target_os = "macos")]
+pub mod macos;
+#[cfg(target_os = "windows")]
+pub mod windows;
+
+use anyhow::Result;
+use regex::bytes::Regex;
+use std::collections::HashSet;
+use std::fs;
+use std::path::{Path, PathBuf};
+use std::sync::OnceLock;
+
+use crate::attachment::decoder::{detect_image_format, V2_MAGIC};
+
+/// V2 图片真正需要的是两份材料：
+/// - 16 字节 ASCII AES key
+/// - XOR key（macOS 上来自 uin & 0xff，不是总能硬编码成 0x88）
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ImageKeyMaterial {
+    pub aes_key: [u8; 16],
+    pub xor_key: u8,
+}
+
+/// 单个 wxid 的 V2 image key 提取接口。
+///
+/// 实现者负责跨调用缓存（一台机器上同一 wxid 的 image key 在微信不重启时通常稳定）。
+pub trait ImageKeyProvider {
+    fn get_key(&self, wxid: &str) -> Result<ImageKeyMaterial>;
+
+    fn get_aes_key(&self, wxid: &str) -> Result<[u8; 16]> {
+        Ok(self.get_key(wxid)?.aes_key)
+    }
+
+    fn get_xor_key(&self, wxid: &str) -> Result<u8> {
+        Ok(self.get_key(wxid)?.xor_key)
+    }
+}
+
+/// 平台默认实现。
+pub fn default_provider() -> Option<Box<dyn ImageKeyProvider + Send + Sync>> {
+    #[cfg(target_os = "macos")]
+    {
+        return Some(Box::new(macos::MacosImageKeyProvider::from_current_config()));
+    }
+    #[cfg(target_os = "windows")]
+    {
+        return Some(Box::new(windows::WindowsImageKeyProvider::from_current_config()));
+    }
+    #[cfg(target_os = "linux")]
+    {
+        return Some(Box::new(linux::LinuxImageKeyProvider));
+    }
+    #[cfg(not(any(target_os = "macos", target_os = "windows", target_os = "linux")))]
+    {
+        None
+    }
+}
+
+pub(crate) fn configured_db_dir_for_wxid(configured_db_dir: &Path, requested_wxid: &str) -> PathBuf {
+    if requested_wxid.trim().is_empty() {
+        return configured_db_dir.to_path_buf();
+    }
+
+    let configured_leaf = wxid_from_db_dir(configured_db_dir);
+    if let Some(leaf) = configured_leaf.as_deref() {
+        if same_wxid(leaf, requested_wxid) {
+            return configured_db_dir.to_path_buf();
+        }
+    }
+
+    xwechat_files_root(configured_db_dir)
+        .map(|root| root.join(requested_wxid).join("db_storage"))
+        .unwrap_or_else(|| configured_db_dir.to_path_buf())
+}
+
+pub(crate) fn wxid_from_db_dir(db_dir: &Path) -> Option<String> {
+    let mut components = db_dir
+        .components()
+        .map(|component| component.as_os_str().to_string_lossy().into_owned());
+    while let Some(component) = components.next() {
+        if component == "xwechat_files" {
+            return components.next();
+        }
+    }
+    None
+}
+
+pub(crate) fn xwechat_files_root(db_dir: &Path) -> Option<PathBuf> {
+    let parts: Vec<_> = db_dir
+        .components()
+        .map(|component| component.as_os_str().to_string_lossy().into_owned())
+        .collect();
+    let idx = parts.iter().position(|part| part == "xwechat_files")?;
+    Some(join_components(&parts[..=idx]))
+}
+
+pub(crate) fn normalize_wxid(raw: &str) -> String {
+    let raw = raw.trim();
+    if raw.is_empty() {
+        return String::new();
+    }
+    if let Some(stripped) = raw.strip_prefix("wxid_") {
+        let head = stripped.split('_').next().unwrap_or(stripped);
+        return format!("wxid_{}", head);
+    }
+    if let Some((base, suffix)) = raw.rsplit_once('_') {
+        if suffix.len() == 4 && suffix.bytes().all(|byte| byte.is_ascii_hexdigit()) {
+            return base.to_string();
+        }
+    }
+    raw.to_string()
+}
+
+pub(crate) fn same_wxid(a: &str, b: &str) -> bool {
+    a == b || normalize_wxid(a) == normalize_wxid(b)
+}
+
+pub(crate) fn join_components(parts: &[String]) -> PathBuf {
+    let mut out = if parts.first().map(|part| part.is_empty()).unwrap_or(false) {
+        PathBuf::from("/")
+    } else {
+        PathBuf::new()
+    };
+    for part in parts {
+        if part.is_empty() {
+            continue;
+        }
+        out.push(part);
+    }
+    out
+}
+
+pub(crate) fn attach_root_for_db_dir(db_dir: &Path) -> PathBuf {
+    db_dir
+        .parent()
+        .map(|base| base.join("msg").join("attach"))
+        .unwrap_or_else(|| PathBuf::from("msg/attach"))
+}
+
+pub(crate) fn find_v2_template_ciphertexts(
+    attach_dir: &Path,
+    max_templates: usize,
+    max_files: usize,
+) -> Result<Vec<[u8; 16]>> {
+    if !attach_dir.is_dir() {
+        return Ok(Vec::new());
+    }
+
+    let mut out = collect_templates_with_suffix(attach_dir, "_t.dat", max_templates, max_files)?;
+    if out.is_empty() {
+        out = collect_templates_with_suffix(attach_dir, ".dat", max_templates, max_files)?;
+    }
+    Ok(out)
+}
+
+pub(crate) fn derive_xor_key_from_v2_dat(
+    attach_dir: &Path,
+    sample: usize,
+    min_samples: usize,
+) -> Result<Option<(u8, usize, usize)>> {
+    if !attach_dir.is_dir() {
+        return Ok(None);
+    }
+    let mut votes = Vec::new();
+    visit_files(attach_dir, &mut |path| -> Result<bool> {
+        let Some(name) = path.file_name().and_then(|value| value.to_str()) else {
+            return Ok(false);
+        };
+        if !name.ends_with(".dat") {
+            return Ok(false);
+        }
+
+        let meta = fs::metadata(path)?;
+        if meta.len() < 0x20 {
+            return Ok(false);
+        }
+
+        let bytes = fs::read(path)?;
+        if bytes.starts_with(&V2_MAGIC) {
+            let last = *bytes.last().unwrap();
+            votes.push(last ^ 0xD9);
+            if votes.len() >= sample {
+                return Ok(true);
+            }
+        }
+        Ok(false)
+    })?;
+
+    if votes.len() < min_samples {
+        return Ok(None);
+    }
+
+    let mut counts = [0usize; 256];
+    for vote in &votes {
+        counts[*vote as usize] += 1;
+    }
+    let (xor_key, top_votes) = counts
+        .iter()
+        .enumerate()
+        .max_by_key(|(_, count)| *count)
+        .map(|(idx, count)| (idx as u8, *count))
+        .expect("votes 非空");
+    Ok(Some((xor_key, top_votes, votes.len())))
+}
+
+pub(crate) fn verify_aes_key(aes_key: &[u8; 16], templates: &[[u8; 16]]) -> bool {
+    !templates.is_empty()
+        && templates
+            .iter()
+            .all(|template| decrypt_template_block(aes_key, template).is_some())
+}
+
+pub(crate) fn ascii_alnum_candidates<'a>(buf: &'a [u8], len: usize) -> Vec<&'a [u8]> {
+    let re = match len {
+        16 => regex16(),
+        32 => regex32(),
+        _ => return Vec::new(),
+    };
+
+    re.find_iter(buf)
+        .filter_map(|matched| {
+            let start = matched.start();
+            let end = matched.end();
+            let left_ok = start == 0 || !buf[start - 1].is_ascii_alphanumeric();
+            let right_ok = end == buf.len() || !buf[end].is_ascii_alphanumeric();
+            (left_ok && right_ok).then_some(&buf[start..end])
+        })
+        .collect()
+}
+
+fn collect_templates_with_suffix(
+    dir: &Path,
+    suffix: &str,
+    max_templates: usize,
+    max_files: usize,
+) -> Result<Vec<[u8; 16]>> {
+    let mut out = Vec::new();
+    let mut seen = HashSet::new();
+    let mut examined = 0usize;
+    visit_files(dir, &mut |path| -> Result<bool> {
+        let Some(name) = path.file_name().and_then(|value| value.to_str()) else {
+            return Ok(false);
+        };
+        if !name.ends_with(suffix) {
+            return Ok(false);
+        }
+        examined += 1;
+        let bytes = fs::read(path)?;
+        if bytes.len() >= 0x1F && bytes.starts_with(&V2_MAGIC) {
+            let template: [u8; 16] = bytes[0x0F..0x1F].try_into().unwrap();
+            if seen.insert(template) {
+                out.push(template);
+                if out.len() >= max_templates {
+                    return Ok(true);
+                }
+            }
+        }
+        Ok(examined >= max_files && !out.is_empty())
+    })?;
+    Ok(out)
+}
+
+fn visit_files<F>(dir: &Path, f: &mut F) -> Result<bool>
+where
+    F: FnMut(&Path) -> Result<bool>,
+{
+    let mut entries: Vec<PathBuf> = fs::read_dir(dir)?
+        .flatten()
+        .map(|entry| entry.path())
+        .collect();
+    entries.sort();
+
+    for path in entries {
+        if path.is_dir() {
+            if visit_files(&path, f)? {
+                return Ok(true);
+            }
+            continue;
+        }
+        if f(&path)? {
+            return Ok(true);
+        }
+    }
+    Ok(false)
+}
+
+fn decrypt_template_block(aes_key: &[u8; 16], ciphertext: &[u8; 16]) -> Option<&'static str> {
+    use aes::cipher::{generic_array::GenericArray, BlockDecrypt, KeyInit};
+
+    let cipher = aes::Aes128::new(aes_key.into());
+    let mut block = GenericArray::clone_from_slice(ciphertext);
+    cipher.decrypt_block(&mut block);
+    let block: [u8; 16] = block.as_slice().try_into().ok()?;
+    let format = detect_image_format(&block);
+    (format != "bin").then_some(format)
+}
+
+fn regex16() -> &'static Regex {
+    static RE: OnceLock<Regex> = OnceLock::new();
+    RE.get_or_init(|| Regex::new(r"[A-Za-z0-9]{16}").unwrap())
+}
+
+fn regex32() -> &'static Regex {
+    static RE: OnceLock<Regex> = OnceLock::new();
+    RE.get_or_init(|| Regex::new(r"[A-Za-z0-9]{32}").unwrap())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{ascii_alnum_candidates, normalize_wxid, same_wxid};
+
+    #[test]
+    fn regex_candidates_respect_boundaries() {
+        let buf = b"xx 0123456789ABCDef yy";
+        let hits = ascii_alnum_candidates(buf, 16);
+        assert_eq!(hits, vec![&buf[3..19]]);
+    }
+
+    #[test]
+    fn regex_candidates_ignore_embedded_runs() {
+        let buf = b"x0123456789ABCDefz";
+        assert!(ascii_alnum_candidates(buf, 16).is_empty());
+    }
+
+    #[test]
+    fn wxid_normalization_matches_expected_forms() {
+        assert_eq!(normalize_wxid("wxid_abc_def"), "wxid_abc");
+        assert_eq!(normalize_wxid("your_wxid_a1b2"), "your_wxid");
+        assert!(same_wxid("your_wxid_a1b2", "your_wxid"));
+    }
+}

src/attachment/image_key/windows.rs

@@ -0,0 +1,238 @@
+//! Windows V2 image AES key 提取。
+//!
+//! 扫 `Weixin.exe` 进程内存，匹配模式 `[A-Za-z0-9]{32}` / `[A-Za-z0-9]{16}`，
+//! 然后用 V2 模板 AES block 反验，控制 false positive。
+
+use anyhow::{bail, Context, Result};
+use std::collections::{HashMap, HashSet};
+use std::path::PathBuf;
+use std::sync::Mutex;
+
+use windows::Win32::Foundation::{CloseHandle, HANDLE};
+use windows::Win32::System::Diagnostics::Debug::ReadProcessMemory;
+use windows::Win32::System::Diagnostics::ToolHelp::{
+    CreateToolhelp32Snapshot, Process32First, Process32Next, PROCESSENTRY32, TH32CS_SNAPPROCESS,
+};
+use windows::Win32::System::Memory::{
+    VirtualQueryEx, MEMORY_BASIC_INFORMATION, MEM_COMMIT, PAGE_EXECUTE_READWRITE,
+    PAGE_EXECUTE_WRITECOPY, PAGE_GUARD, PAGE_NOCACHE, PAGE_NOACCESS, PAGE_READWRITE,
+    PAGE_WRITECOMBINE, PAGE_WRITECOPY,
+};
+use windows::Win32::System::Threading::{OpenProcess, PROCESS_QUERY_INFORMATION, PROCESS_VM_READ};
+
+use crate::config;
+
+use super::{
+    ascii_alnum_candidates, attach_root_for_db_dir, configured_db_dir_for_wxid,
+    derive_xor_key_from_v2_dat, find_v2_template_ciphertexts, verify_aes_key, ImageKeyMaterial,
+    ImageKeyProvider,
+};
+
+const CHUNK_SIZE: usize = 2 * 1024 * 1024;
+const MAX_REGION_SIZE: usize = 50 * 1024 * 1024;
+
+pub struct WindowsImageKeyProvider {
+    configured_db_dir: Result<PathBuf, String>,
+    cache: Mutex<HashMap<String, ImageKeyMaterial>>,
+}
+
+impl WindowsImageKeyProvider {
+    pub fn from_current_config() -> Self {
+        let configured_db_dir = config::load_config()
+            .map(|cfg| cfg.db_dir)
+            .map_err(|err| err.to_string());
+        Self {
+            configured_db_dir,
+            cache: Mutex::new(HashMap::new()),
+        }
+    }
+}
+
+impl ImageKeyProvider for WindowsImageKeyProvider {
+    fn get_key(&self, wxid: &str) -> Result<ImageKeyMaterial> {
+        let cache_key = wxid.trim().to_string();
+        if let Some(found) = self.cache.lock().unwrap().get(&cache_key).copied() {
+            return Ok(found);
+        }
+
+        let configured_db_dir = self
+            .configured_db_dir
+            .as_ref()
+            .map_err(|err| anyhow::anyhow!("读取 config.db_dir 失败: {}", err))?;
+        let db_dir = configured_db_dir_for_wxid(configured_db_dir, wxid);
+        let attach_dir = attach_root_for_db_dir(&db_dir);
+        let key = derive_key_for_paths(&attach_dir)?;
+        self.cache.lock().unwrap().insert(cache_key, key);
+        Ok(key)
+    }
+}
+
+fn derive_key_for_paths(attach_dir: &std::path::Path) -> Result<ImageKeyMaterial> {
+    let templates = find_v2_template_ciphertexts(attach_dir, 3, 64)?;
+    if templates.is_empty() {
+        bail!("在 {} 下找不到 V2 模板文件", attach_dir.display());
+    }
+    let xor_key = derive_xor_key_from_v2_dat(attach_dir, 10, 3)?
+        .map(|(key, _, _)| key)
+        .unwrap_or(0x88);
+
+    let pid = find_wechat_pid().context("找不到 Weixin.exe 进程，请确认微信正在运行")?;
+    let process = unsafe {
+        OpenProcess(PROCESS_VM_READ | PROCESS_QUERY_INFORMATION, false, pid)
+            .context("OpenProcess 失败，请以管理员权限运行")?
+    };
+
+    let aes_key = scan_memory_for_key(process, &templates);
+    unsafe {
+        let _ = CloseHandle(process);
+    }
+
+    Ok(ImageKeyMaterial {
+        aes_key: aes_key?,
+        xor_key,
+    })
+}
+
+fn find_wechat_pid() -> Option<u32> {
+    let snapshot = unsafe { CreateToolhelp32Snapshot(TH32CS_SNAPPROCESS, 0).ok()? };
+    let mut entry = PROCESSENTRY32 {
+        dwSize: std::mem::size_of::<PROCESSENTRY32>() as u32,
+        ..Default::default()
+    };
+
+    unsafe {
+        if Process32First(snapshot, &mut entry).is_err() {
+            let _ = CloseHandle(snapshot);
+            return None;
+        }
+        loop {
+            let name =
+                std::ffi::CStr::from_ptr(entry.szExeFile.as_ptr() as *const i8).to_string_lossy();
+            if name.eq_ignore_ascii_case("Weixin.exe") {
+                let pid = entry.th32ProcessID;
+                let _ = CloseHandle(snapshot);
+                return Some(pid);
+            }
+            if Process32Next(snapshot, &mut entry).is_err() {
+                break;
+            }
+        }
+        let _ = CloseHandle(snapshot);
+    }
+    None
+}
+
+fn scan_memory_for_key(process: HANDLE, templates: &[[u8; 16]]) -> Result<[u8; 16]> {
+    let mut seen = HashSet::<[u8; 16]>::new();
+    let mut address = 0usize;
+
+    loop {
+        let mut mbi = MEMORY_BASIC_INFORMATION::default();
+        let ret = unsafe {
+            VirtualQueryEx(
+                process,
+                Some(address as *const _),
+                &mut mbi,
+                std::mem::size_of::<MEMORY_BASIC_INFORMATION>(),
+            )
+        };
+        if ret == 0 {
+            break;
+        }
+
+        let base = mbi.BaseAddress as usize;
+        let size = mbi.RegionSize;
+        if mbi.State == MEM_COMMIT && is_candidate_page(mbi.Protect.0) && size <= MAX_REGION_SIZE {
+            if let Some(aes_key) = scan_region(process, base, size, templates, &mut seen)? {
+                return Ok(aes_key);
+            }
+        }
+
+        address = base.saturating_add(size);
+        if address == 0 {
+            break;
+        }
+    }
+
+    bail!("Windows 进程内存里没有找到可验证的 V2 AES key")
+}
+
+fn scan_region(
+    process: HANDLE,
+    base: usize,
+    size: usize,
+    templates: &[[u8; 16]],
+    seen: &mut HashSet<[u8; 16]>,
+) -> Result<Option<[u8; 16]>> {
+    let overlap = 31usize;
+    let mut offset = 0usize;
+
+    while offset < size {
+        let chunk_size = std::cmp::min(CHUNK_SIZE, size - offset);
+        let addr = base + offset;
+        let mut buf = vec![0u8; chunk_size];
+        let mut bytes_read = 0usize;
+
+        let ok = unsafe {
+            ReadProcessMemory(
+                process,
+                addr as *const _,
+                buf.as_mut_ptr() as *mut _,
+                chunk_size,
+                Some(&mut bytes_read),
+            )
+            .is_ok()
+        };
+
+        if ok && bytes_read > 0 {
+            buf.truncate(bytes_read);
+            if let Some(key) = scan_candidate_buffer(&buf, templates, seen) {
+                return Ok(Some(key));
+            }
+        }
+
+        offset += if chunk_size > overlap {
+            chunk_size - overlap
+        } else {
+            chunk_size
+        };
+    }
+
+    Ok(None)
+}
+
+fn scan_candidate_buffer(
+    buf: &[u8],
+    templates: &[[u8; 16]],
+    seen: &mut HashSet<[u8; 16]>,
+) -> Option<[u8; 16]> {
+    for candidate in ascii_alnum_candidates(buf, 32) {
+        let mut key = [0u8; 16];
+        key.copy_from_slice(&candidate[..16]);
+        if seen.insert(key) && verify_aes_key(&key, templates) {
+            return Some(key);
+        }
+    }
+    for candidate in ascii_alnum_candidates(buf, 16) {
+        let mut key = [0u8; 16];
+        key.copy_from_slice(candidate);
+        if seen.insert(key) && verify_aes_key(&key, templates) {
+            return Some(key);
+        }
+    }
+    None
+}
+
+fn is_candidate_page(protect: u32) -> bool {
+    if protect == PAGE_NOACCESS.0 || (protect & PAGE_GUARD.0) != 0 {
+        return false;
+    }
+    let base = protect & !(PAGE_GUARD.0 | PAGE_NOCACHE.0 | PAGE_WRITECOMBINE.0);
+    matches!(
+        base,
+        value if value == PAGE_READWRITE.0
+            || value == PAGE_WRITECOPY.0
+            || value == PAGE_EXECUTE_READWRITE.0
+            || value == PAGE_EXECUTE_WRITECOPY.0
+    )
+}

src/attachment/mod.rs

@@ -0,0 +1,28 @@
+//! 聊天附件提取链路（图片 / 视频 / 语音 / 文件本体的本地解码）
+//!
+//! 整条链：
+//!   message_N.db (Msg_<md5>) → message_resource.db (ChatName2Id + MessageResourceInfo)
+//!     → packed_info protobuf md5 提取 → xwechat_files/<wxid>/msg/attach/.../Img/<md5>[_t|_h].dat
+//!     → magic 分发 (legacy XOR / V1 fixed-AES / V2 AES+XOR) → 写出实际图片
+//!
+//! 模块切分：
+//! - `attachment_id`：跨 IPC / CLI 的不透明 ID（base64url(json)）
+//! - `resolver`：从 `attachment_id` 反查 message_resource.db，定位本地 .dat
+//! - `decoder`：根据文件 magic 分发到具体解码器（V1 / V2 等）
+//! - `image_key`：V2 image AES key 提取（macOS / Windows）
+//!
+//! V2 / image_key 模块由 codex 落地，先放空 stub 以便 V1 / resolver / CLI 不被 block。
+
+// 此模块由分多个 PR/commit 增量启用：
+// 1) 先落 attachment_id / decoder / resolver / image_key 骨架（本 commit）
+// 2) IPC + CLI + daemon route 把它们串起来（后续 commit）
+// 3) image_key 平台实现（codex 后续 commit）
+// 在 step 1 完成、step 2 未到时，大量公开 API 仍未被引用，#[allow(dead_code)] 抑制噪音
+#![allow(dead_code)]
+
+pub mod attachment_id;
+pub mod decoder;
+pub mod resolver;
+pub mod image_key;
+
+pub use attachment_id::{AttachmentId, AttachmentKind};

src/attachment/resolver.rs

@@ -0,0 +1,439 @@
+//! 把 `AttachmentId` 翻译成本地 `.dat` 路径。
+//!
+//! 流程：
+//!   1. `chat` username → `ChatName2Id.rowid`（资源库）
+//!   2. `(chat_id, local_id)` + `ORDER BY message_create_time DESC LIMIT 1` →
+//!      `MessageResourceInfo.packed_info`
+//!   3. 从 `packed_info` (protobuf) 提取 32 字节 ASCII hex MD5
+//!   4. 在 `<wxchat_base>/msg/attach/<md5(chat)>/<YYYY-MM>/Img/<md5>[_t|_h].dat`
+//!      下找对应文件，按 full > _h > _t 优先级选一个
+//!
+//! `<wxchat_base>` 由 daemon 已知（同 `db_dir` 的父目录），路径 layout 平台差异：
+//! - Linux: `~/Documents/xwechat_files/<wxid>`
+//! - macOS: `~/Library/Containers/com.tencent.xinWeChat/Data/Documents/xwechat_files/<wxid>`
+//!   ⚠️  msg/attach/... 子树 layout 待我用真实账号验证；上游 docstring 只写了 Windows
+//! - Windows: `<root>\xwechat_files\<wxid>`（root 从 `%APPDATA%\Tencent\xwechat\config\*.ini` 读）
+
+use anyhow::{anyhow, Context, Result};
+use chrono::TimeZone;
+use rusqlite::Connection;
+use std::path::{Path, PathBuf};
+
+use super::AttachmentId;
+
+/// 单条 attachment 在资源库 + 本地 attach 树下的解析结果。
+#[derive(Debug, Clone)]
+pub struct ResolvedAttachment {
+    pub id: AttachmentId,
+    /// 从 `packed_info` 提取出的资源 MD5（小写 hex）
+    pub md5: String,
+    /// 命中的本地 .dat 路径（按 full > _h > _t 优先级选一个）
+    pub dat_path: PathBuf,
+    /// 文件 size（字节）
+    pub size: u64,
+}
+
+/// 仅 schema lookup（不去找本地 .dat）。
+/// 用于 `wx attachments` 列表时填 `md5` 字段——文件可能根本不在本地。
+#[derive(Debug, Clone)]
+pub struct AttachmentMetadata {
+    pub md5: String,
+}
+
+/// 用 `(chat, local_id)` 查 message_resource.db 拿 file md5。
+///
+/// 调用方传已经解密好的 `message_resource.db` 路径（由 daemon 的 `DBCache` 准备）。
+/// 同步函数 — caller 在 `spawn_blocking` 里跑。
+pub fn lookup_md5_blocking(
+    resource_db_path: &Path,
+    chat: &str,
+    local_id: i64,
+    create_time: i64,
+    msg_local_type_lo32: i64,
+) -> Result<Option<AttachmentMetadata>> {
+    let conn = Connection::open_with_flags(
+        resource_db_path,
+        rusqlite::OpenFlags::SQLITE_OPEN_READ_ONLY | rusqlite::OpenFlags::SQLITE_OPEN_URI,
+    )
+    .with_context(|| format!("打开 message_resource.db {:?}", resource_db_path))?;
+
+    // 1) ChatName2Id: user_name -> rowid
+    let chat_id: Option<i64> = conn
+        .query_row(
+            "SELECT rowid FROM ChatName2Id WHERE user_name = ?1",
+            [chat],
+            |row| row.get(0),
+        )
+        .ok();
+    let Some(chat_id) = chat_id else {
+        return Ok(None);
+    };
+
+    // 2) MessageResourceInfo:
+    //    同 chat 内 local_id 会复用，所以先用 create_time 精确命中；
+    //    若资源库里的时间戳跟 message_N.db 不完全对齐，再 fallback 到“同 local_id/type 取最新”
+    //    message_local_type 高 32 bit 是版本/会话 flag，低 32 bit 才是真实类型
+    let packed_exact: Option<Vec<u8>> = conn
+        .query_row(
+            "SELECT packed_info FROM MessageResourceInfo
+             WHERE chat_id = ?1
+               AND message_local_id = ?2
+               AND (message_local_type = ?3 OR message_local_type % 4294967296 = ?3)
+               AND message_create_time = ?4
+             ORDER BY rowid DESC
+             LIMIT 1",
+            rusqlite::params![chat_id, local_id, msg_local_type_lo32, create_time],
+            |row| row.get(0),
+        )
+        .ok();
+
+    let packed: Option<Vec<u8>> = packed_exact.or_else(|| conn
+        .query_row(
+            "SELECT packed_info FROM MessageResourceInfo
+             WHERE chat_id = ?1
+               AND message_local_id = ?2
+               AND (message_local_type = ?3 OR message_local_type % 4294967296 = ?3)
+             ORDER BY message_create_time DESC
+             LIMIT 1",
+            rusqlite::params![chat_id, local_id, msg_local_type_lo32],
+            |row| row.get(0),
+        )
+        .ok());
+
+    let Some(blob) = packed else {
+        return Ok(None);
+    };
+    Ok(extract_md5_from_packed_info(&blob).map(|md5| AttachmentMetadata { md5 }))
+}
+
+/// 从 `MessageResourceInfo.packed_info` (protobuf) 提取 32 字节 ASCII hex md5。
+///
+/// 主路径：搜 4 字节 marker `12 22 0a 20`（field=2 LEN, length=34, sub field=1 LEN, length=32），
+/// 紧跟 32 字节 ASCII hex。
+/// Fallback：扫整个 blob 找连续 32 字节合法 hex 字符。
+pub fn extract_md5_from_packed_info(blob: &[u8]) -> Option<String> {
+    const MARKER: &[u8; 4] = &[0x12, 0x22, 0x0A, 0x20];
+
+    // 主路径
+    if let Some(pos) = find_subslice(blob, MARKER) {
+        let start = pos + MARKER.len();
+        if start + 32 <= blob.len() {
+            if let Ok(s) = std::str::from_utf8(&blob[start..start + 32]) {
+                if s.chars().all(|c| c.is_ascii_hexdigit()) {
+                    return Some(s.to_ascii_lowercase());
+                }
+            }
+        }
+    }
+
+    // Fallback：连续 32 字节合法 hex
+    if blob.len() >= 32 {
+        for start in 0..=blob.len() - 32 {
+            let chunk = &blob[start..start + 32];
+            if let Ok(s) = std::str::from_utf8(chunk) {
+                if s.chars().all(|c| c.is_ascii_hexdigit()) {
+                    return Some(s.to_ascii_lowercase());
+                }
+            }
+        }
+    }
+    None
+}
+
+/// 简单的子串扫描（避免拉 memchr/memmem 依赖；blob 通常 < 1KB）
+fn find_subslice(haystack: &[u8], needle: &[u8]) -> Option<usize> {
+    if needle.is_empty() || needle.len() > haystack.len() {
+        return None;
+    }
+    haystack
+        .windows(needle.len())
+        .position(|w| w == needle)
+}
+
+/// 在 `<attach_root>/<md5(chat)>/<YYYY-MM>/Img/<md5>[_t|_h].dat` 下找文件。
+///
+/// 优先级：full > `_h`（HD thumbnail）> `_t`（thumbnail）。返回最优的一个；
+/// 找不到返回 None。
+///
+/// `attach_root` = `<wxchat_base>/msg/attach`。
+/// `create_time` 用于先定位 `<YYYY-MM>` 子目录；找不到时再 fallback 全月份扫描，
+/// 因为 WeChat 的 `YYYY-MM` 目录有时跟消息时间差 1 个月（按收到时间归档）。
+pub fn find_dat_file(
+    attach_root: &Path,
+    chat: &str,
+    file_md5: &str,
+    create_time: i64,
+) -> Option<PathBuf> {
+    let chat_hash = format!("{:x}", md5::compute(chat.as_bytes()));
+    let chat_dir = attach_root.join(&chat_hash);
+    if !chat_dir.is_dir() {
+        return None;
+    }
+
+    // 第一步：试 create_time 当月 + 前后各一个月（共 3 个候选目录）
+    let candidates_ym: Vec<String> = three_month_candidates(create_time);
+    for ym in &candidates_ym {
+        let img_dir = chat_dir.join(ym).join("Img");
+        if let Some(p) = pick_best_in_img_dir(&img_dir, file_md5) {
+            return Some(p);
+        }
+    }
+
+    // 第二步 fallback：扫整个 chat_dir 的所有月份子目录
+    let entries = std::fs::read_dir(&chat_dir).ok()?;
+    let mut all_months: Vec<PathBuf> = entries
+        .filter_map(|e| e.ok())
+        .map(|e| e.path())
+        .filter(|p| p.is_dir())
+        .collect();
+    // 已经试过的 3 个候选可以跳过，但成本极小；保留全量扫
+    all_months.sort();
+    for month_dir in all_months {
+        let img_dir = month_dir.join("Img");
+        if let Some(p) = pick_best_in_img_dir(&img_dir, file_md5) {
+            return Some(p);
+        }
+    }
+    None
+}
+
+fn pick_best_in_img_dir(img_dir: &Path, file_md5: &str) -> Option<PathBuf> {
+    if !img_dir.is_dir() {
+        return None;
+    }
+    let full = img_dir.join(format!("{}.dat", file_md5));
+    if full.is_file() {
+        return Some(full);
+    }
+    let hd = img_dir.join(format!("{}_h.dat", file_md5));
+    if hd.is_file() {
+        return Some(hd);
+    }
+    let thumb = img_dir.join(format!("{}_t.dat", file_md5));
+    if thumb.is_file() {
+        return Some(thumb);
+    }
+    None
+}
+
+fn three_month_candidates(unix_ts: i64) -> Vec<String> {
+    use chrono::{Datelike, Duration};
+    let dt = match chrono::Local.timestamp_opt(unix_ts, 0).single() {
+        Some(d) => d,
+        None => return Vec::new(),
+    };
+    let prev = dt - Duration::days(31);
+    let next = dt + Duration::days(31);
+    [prev, dt, next]
+        .iter()
+        .map(|d| format!("{:04}-{:02}", d.year(), d.month()))
+        .collect()
+}
+
+/// 把 `<wxchat_base>` （即 `db_storage` 父目录）拼成 `<base>/msg/attach`。
+pub fn attach_root_for(wxchat_base: &Path) -> PathBuf {
+    wxchat_base.join("msg").join("attach")
+}
+
+/// 完整流程：用 `attachment_id` 拿 md5 + 找 .dat。失败返回带具体诊断信息的 `Err`。
+///
+/// `resource_db_path` 由 daemon 提供（DBCache 已经解密好）；
+/// `attach_root` 由 caller 拼好（`attach_root_for(wxchat_base)`）。
+/// 同步函数 — caller 在 `spawn_blocking` 里跑。
+pub fn resolve_blocking(
+    id: &AttachmentId,
+    resource_db_path: &Path,
+    attach_root: &Path,
+) -> Result<ResolvedAttachment> {
+    let lo32_type: i64 = match id.kind {
+        super::AttachmentKind::Image => 3,
+        super::AttachmentKind::Voice => 34,
+        super::AttachmentKind::Video => 43,
+        super::AttachmentKind::File => 49,
+    };
+
+    let meta = lookup_md5_blocking(
+        resource_db_path,
+        &id.chat,
+        id.local_id,
+        id.create_time,
+        lo32_type,
+    )?
+        .ok_or_else(|| {
+            anyhow!(
+                "message_resource.db 中找不到 chat={} local_id={} type={} 的资源行（可能是非附件消息或资源库未同步）",
+                id.chat,
+                id.local_id,
+                lo32_type
+            )
+        })?;
+
+    let dat_path = find_dat_file(attach_root, &id.chat, &meta.md5, id.create_time).ok_or_else(
+        || {
+            anyhow!(
+                "找不到本地 .dat（md5={} chat={} create_time={}）— 微信可能尚未下载该附件，或附件已被清理",
+                meta.md5,
+                id.chat,
+                id.create_time
+            )
+        },
+    )?;
+    let size = std::fs::metadata(&dat_path).map(|m| m.len()).unwrap_or(0);
+
+    Ok(ResolvedAttachment { id: id.clone(), md5: meta.md5, dat_path, size })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn extract_md5_main_path() {
+        // 构造一段含 12 22 0a 20 marker 的 blob
+        let mut blob = vec![0xAA, 0xBB, 0xCC];
+        blob.extend_from_slice(&[0x12, 0x22, 0x0A, 0x20]);
+        blob.extend_from_slice(b"deadbeefcafebabe1234567890abcdef");
+        blob.extend_from_slice(&[0xFF, 0xFF]);
+        assert_eq!(
+            extract_md5_from_packed_info(&blob),
+            Some("deadbeefcafebabe1234567890abcdef".to_string())
+        );
+    }
+
+    #[test]
+    fn extract_md5_fallback_no_marker() {
+        // 没有 marker，但 blob 里有合法 32 字节 hex
+        let mut blob = vec![0xFF, 0x00];
+        blob.extend_from_slice(b"00112233445566778899aabbccddeeff");
+        blob.extend_from_slice(&[0x01]);
+        assert_eq!(
+            extract_md5_from_packed_info(&blob),
+            Some("00112233445566778899aabbccddeeff".to_string())
+        );
+    }
+
+    #[test]
+    fn extract_md5_uppercase_normalized_to_lower() {
+        let mut blob = vec![0x12, 0x22, 0x0A, 0x20];
+        blob.extend_from_slice(b"DEADBEEFCAFEBABE1234567890ABCDEF");
+        // 上游/CI/本地 file md5 都是 lowercase；强制小写化避免大小写不一致导致命中失败
+        assert_eq!(
+            extract_md5_from_packed_info(&blob),
+            Some("deadbeefcafebabe1234567890abcdef".to_string())
+        );
+    }
+
+    #[test]
+    fn extract_md5_returns_none_on_garbage() {
+        let blob = vec![0; 16];
+        assert!(extract_md5_from_packed_info(&blob).is_none());
+    }
+
+    #[test]
+    fn lookup_md5_prefers_exact_create_time_over_latest_reuse() {
+        let dir = tempdir_for_test();
+        let db_path = dir.join("message_resource.db");
+        let conn = Connection::open(&db_path).unwrap();
+        conn.execute(
+            "CREATE TABLE ChatName2Id (user_name TEXT)",
+            [],
+        )
+        .unwrap();
+        conn.execute(
+            "INSERT INTO ChatName2Id (rowid, user_name) VALUES (1, 'room@chatroom')",
+            [],
+        )
+        .unwrap();
+        conn.execute(
+            "CREATE TABLE MessageResourceInfo (
+                chat_id INTEGER,
+                message_local_id INTEGER,
+                message_local_type INTEGER,
+                message_create_time INTEGER,
+                packed_info BLOB
+            )",
+            [],
+        )
+        .unwrap();
+
+        let old_blob = {
+            let mut blob = vec![0x12, 0x22, 0x0A, 0x20];
+            blob.extend_from_slice(b"11111111111111111111111111111111");
+            blob
+        };
+        let new_blob = {
+            let mut blob = vec![0x12, 0x22, 0x0A, 0x20];
+            blob.extend_from_slice(b"22222222222222222222222222222222");
+            blob
+        };
+
+        conn.execute(
+            "INSERT INTO MessageResourceInfo
+             (chat_id, message_local_id, message_local_type, message_create_time, packed_info)
+             VALUES (?1, ?2, ?3, ?4, ?5)",
+            rusqlite::params![1i64, 7i64, 3i64, 1000i64, old_blob],
+        )
+        .unwrap();
+        conn.execute(
+            "INSERT INTO MessageResourceInfo
+             (chat_id, message_local_id, message_local_type, message_create_time, packed_info)
+             VALUES (?1, ?2, ?3, ?4, ?5)",
+            rusqlite::params![1i64, 7i64, 3i64, 2000i64, new_blob],
+        )
+        .unwrap();
+
+        let old = lookup_md5_blocking(&db_path, "room@chatroom", 7, 1000, 3)
+            .unwrap()
+            .unwrap();
+        let new = lookup_md5_blocking(&db_path, "room@chatroom", 7, 2000, 3)
+            .unwrap()
+            .unwrap();
+        assert_eq!(old.md5, "11111111111111111111111111111111");
+        assert_eq!(new.md5, "22222222222222222222222222222222");
+    }
+
+    #[test]
+    fn three_month_candidates_includes_prev_curr_next() {
+        // 2025-08-15 (mid-month) → 2025-07, 2025-08, 2025-09
+        let ts = chrono::Local
+            .with_ymd_and_hms(2025, 8, 15, 12, 0, 0)
+            .unwrap()
+            .timestamp();
+        let v = three_month_candidates(ts);
+        assert!(v.contains(&"2025-07".to_string()));
+        assert!(v.contains(&"2025-08".to_string()));
+        assert!(v.contains(&"2025-09".to_string()));
+    }
+
+    #[test]
+    fn pick_best_prefers_full_then_h_then_t() {
+        let tmp = tempdir_for_test();
+        let img = tmp.join("Img");
+        std::fs::create_dir_all(&img).unwrap();
+        let md5 = "abcd1234";
+        std::fs::write(img.join(format!("{}_t.dat", md5)), b"thumb").unwrap();
+        std::fs::write(img.join(format!("{}_h.dat", md5)), b"hd").unwrap();
+        // 只有 _t / _h 时取 _h
+        assert_eq!(
+            pick_best_in_img_dir(&img, md5).unwrap().file_name().unwrap(),
+            format!("{}_h.dat", md5).as_str()
+        );
+        // 加 full 后取 full
+        std::fs::write(img.join(format!("{}.dat", md5)), b"full").unwrap();
+        assert_eq!(
+            pick_best_in_img_dir(&img, md5).unwrap().file_name().unwrap(),
+            format!("{}.dat", md5).as_str()
+        );
+    }
+
+    fn tempdir_for_test() -> PathBuf {
+        let pid = std::process::id();
+        let nanos = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos();
+        let p = std::env::temp_dir().join(format!("wx-cli-attach-test-{}-{}", pid, nanos));
+        std::fs::create_dir_all(&p).unwrap();
+        p
+    }
+}

src/cli/attachments.rs

@@ -0,0 +1,42 @@
+use anyhow::Result;
+
+use crate::ipc::Request;
+use super::history::{parse_time, parse_time_end};
+use super::output::{print_value, resolve};
+use super::transport;
+
+/// `wx attachments` — 列出指定会话的附件消息（默认 image，可多选）。
+///
+/// 输出每条 `attachment_id`，再传给 `wx extract` 才真正读 message_resource.db
+/// 与本地 .dat 解码。这一步只查 `Msg_<chat>` 表，几千条群聊也能秒返。
+pub fn cmd_attachments(
+    chat: String,
+    kinds: Vec<String>,
+    limit: usize,
+    offset: usize,
+    since: Option<String>,
+    until: Option<String>,
+    json: bool,
+) -> Result<()> {
+    let since_ts = since.as_deref().map(parse_time).transpose()?;
+    let until_ts = until.as_deref().map(parse_time_end).transpose()?;
+
+    // CLI 收上来的 Vec<String> 为空时按默认（image）走，让 daemon 决定 fallback。
+    let kinds_param = if kinds.is_empty() { None } else { Some(kinds) };
+
+    let req = Request::Attachments {
+        chat,
+        kinds: kinds_param,
+        limit,
+        offset,
+        since: since_ts,
+        until: until_ts,
+    };
+    let resp = transport::send(req)?;
+    let data = resp
+        .data
+        .get("attachments")
+        .cloned()
+        .unwrap_or(serde_json::Value::Array(vec![]));
+    print_value(&data, &resolve(json))
+}

src/cli/extract.rs

@@ -0,0 +1,25 @@
+use anyhow::Result;
+
+use crate::ipc::Request;
+use super::output::{print_value, resolve};
+use super::transport;
+
+/// `wx extract` — 把单个 `attachment_id` 对应的资源解密写到指定路径。
+///
+/// daemon 端：解析 `attachment_id` → 查 `message_resource.db` 拿 file md5 →
+/// 在 `<wxchat_base>/msg/attach/...` 找 .dat → 按 magic 分发到 v1/v2 解码器 →
+/// 写出真实图片/文件。
+pub fn cmd_extract(
+    attachment_id: String,
+    output: String,
+    overwrite: bool,
+    json: bool,
+) -> Result<()> {
+    let req = Request::Extract {
+        attachment_id,
+        output,
+        overwrite,
+    };
+    let resp = transport::send(req)?;
+    print_value(&resp.data, &resolve(json))
+}

src/cli/mod.rs

@@ -1,5 +1,7 @@
 mod init;
+pub mod attachments;
 pub mod biz_articles;
+pub mod extract;
 pub mod sessions;
 pub mod history;
 pub mod search;
@@ -262,6 +264,44 @@ enum Commands {
         #[arg(long)]
         json: bool,
     },
+    /// 列出某会话的图片附件，返回不透明 attachment_id
+    Attachments {
+        /// 会话名称（联系人显示名 / wxid / @chatroom username 都可以）
+        chat: String,
+        /// 类型（当前仅支持 image）
+        #[arg(long = "kind", value_name = "KIND",
+              value_parser = ["image", "img"])]
+        kinds: Vec<String>,
+        /// 显示数量
+        #[arg(short = 'n', long, default_value = "50")]
+        limit: usize,
+        /// 分页偏移
+        #[arg(long, default_value = "0")]
+        offset: usize,
+        /// 起始时间 YYYY-MM-DD
+        #[arg(long)]
+        since: Option<String>,
+        /// 结束时间 YYYY-MM-DD
+        #[arg(long)]
+        until: Option<String>,
+        /// 输出 JSON（默认 YAML）
+        #[arg(long)]
+        json: bool,
+    },
+    /// 把单个 attachment_id 对应的资源解密写到指定文件路径
+    Extract {
+        /// 由 `wx attachments` 输出的不透明 ID（base64url 字符串）
+        attachment_id: String,
+        /// 输出文件路径（绝对或相对当前工作目录均可；扩展名建议保留为 .jpg 等）
+        #[arg(short = 'o', long)]
+        output: String,
+        /// 目标已存在时覆盖
+        #[arg(long)]
+        overwrite: bool,
+        /// 输出 JSON（默认 YAML）
+        #[arg(long)]
+        json: bool,
+    },
     /// 管理 wx-daemon
     Daemon {
         #[command(subcommand)]
@@ -329,6 +369,12 @@ fn dispatch(cli: Cli) -> Result<()> {
         Commands::BizArticles { limit, account, since, until, unread, json } => {
             biz_articles::cmd_biz_articles(limit, account, since, until, unread, json)
         }
+        Commands::Attachments { chat, kinds, limit, offset, since, until, json } => {
+            attachments::cmd_attachments(chat, kinds, limit, offset, since, until, json)
+        }
+        Commands::Extract { attachment_id, output, overwrite, json } => {
+            extract::cmd_extract(attachment_id, output, overwrite, json)
+        }
         Commands::Daemon { cmd } => daemon_cmd::cmd_daemon(cmd),
     }
 }

src/daemon/cache.rs

@@ -54,6 +54,12 @@ impl DbCache {
         Ok(cache)
     }
 
+    /// 数据库根目录（即 `<wxchat_base>/db_storage`）。
+    /// 上层（attachment resolver）需要 `db_dir.parent()` 来定位 `msg/attach/...` 解密图片。
+    pub fn db_dir(&self) -> &Path {
+        &self.db_dir
+    }
+
     fn cache_file_path(&self, rel_key: &str) -> PathBuf {
         let hash = format!("{:x}", md5::compute(rel_key.as_bytes()));
         self.cache_dir.join(format!("{}.db", hash))

src/daemon/query.rs

@@ -3285,6 +3285,316 @@ pub async fn q_biz_articles(
     Ok(json!({ "count": results.len(), "articles": results }))
 }
 
+// ─── 附件（当前先支持图片）查询与提取 ─────────────────────────────────
+//
+// 设计要点：
+// - `q_attachments` 只走 `Msg_<chat_md5>` 表，按 `local_type & 0xFFFFFFFF IN (...)` 过滤
+//   出附件消息行，再编出 `attachment_id`。**不**去翻 `message_resource.db`，因为列出动作
+//   要可枚举几千条；resource lookup 留到 `q_extract` 才做。
+// - `q_extract` 走完整链：`AttachmentId` → `message_resource.db` 查 md5 →
+//   `<wxchat_base>/msg/attach/...` 找 .dat → 按 magic 分发到 v1/v2 decoder → 写盘。
+// - V2 image AES key 通过 `image_key::default_provider()` 拿（codex 后续填实现）。
+//   缺 key 时 V2 解码会返回明确错误，CLI 直接抛给用户。
+
+/// 列出某会话内的附件消息（当前仅 image）。返回每条的 `attachment_id`，
+/// 后续传给 `Extract` 才真正读 message_resource.db + 解密 .dat。
+pub async fn q_attachments(
+    db: &DbCache,
+    names: &Names,
+    chat: &str,
+    kinds: Option<Vec<String>>,
+    limit: usize,
+    offset: usize,
+    since: Option<i64>,
+    until: Option<i64>,
+) -> Result<Value> {
+    use crate::attachment::{AttachmentId, AttachmentKind};
+
+    let username = resolve_username(chat, names)
+        .with_context(|| format!("找不到联系人: {}", chat))?;
+    let display = names.display(&username);
+    let chat_type = chat_type_of(&username, names);
+    let is_group = chat_type == "group";
+
+    // 解析 kinds → 低 32 bit local_type 集合
+    let kind_filters: Vec<(AttachmentKind, i64)> = parse_attachment_kinds(kinds.as_deref())?;
+    if kind_filters.is_empty() {
+        anyhow::bail!("kinds 为空 — 当前至少传一种 image");
+    }
+    let lo32_types: Vec<i64> = kind_filters.iter().map(|(_, t)| *t).collect();
+    // local_type → AttachmentKind 反查（mask 完后定 kind）
+    let type_to_kind: HashMap<i64, AttachmentKind> = kind_filters.iter()
+        .map(|(k, t)| (*t, *k))
+        .collect();
+
+    let tables = find_msg_tables(db, names, &username).await?;
+    if tables.is_empty() {
+        anyhow::bail!("找不到 {} 的消息记录", display);
+    }
+
+    // 群聊需要 sender 显示名
+    let group_nicknames = if is_group {
+        load_group_nicknames(db, &username).await.unwrap_or_default()
+    } else {
+        HashMap::new()
+    };
+
+    let mut all_rows: Vec<(i64, i64, i64, i64, String, i64, i64)> = Vec::new();
+    // 元组：(local_id, local_type_lo32, create_time, real_sender_id, sender_label, ts_for_sort, db_idx)
+    for (db_idx, (db_path, table_name)) in tables.iter().enumerate() {
+        let path = db_path.clone();
+        let tname = table_name.clone();
+        let uname = username.clone();
+        let is_group2 = is_group;
+        let names_map = names.map.clone();
+        let group_nicknames2 = group_nicknames.clone();
+        let lo32_types2 = lo32_types.clone();
+        let since2 = since;
+        let until2 = until;
+        // per-DB 软上限避免巨群全量加载
+        let per_db_cap = (offset + limit).max(limit) * 2;
+        let db_idx2 = db_idx as i64;
+
+        let rows: Vec<(i64, i64, i64, i64, String, i64, i64)> =
+            tokio::task::spawn_blocking(move || {
+                let conn = Connection::open(&path)?;
+                let id2u = load_id2u(&conn);
+
+                // local_type 在 DB 里可能带高位 flag，过滤要 mask 低 32 bit
+                let placeholders = lo32_types2.iter().map(|_| "?").collect::<Vec<_>>().join(",");
+                let mut clauses: Vec<String> = vec![
+                    format!("(local_type & 4294967295) IN ({})", placeholders),
+                ];
+                let mut params: Vec<Box<dyn rusqlite::types::ToSql>> = lo32_types2.iter()
+                    .map(|t| Box::new(*t) as Box<dyn rusqlite::types::ToSql>)
+                    .collect();
+                if let Some(s) = since2 {
+                    clauses.push("create_time >= ?".into());
+                    params.push(Box::new(s));
+                }
+                if let Some(u) = until2 {
+                    clauses.push("create_time <= ?".into());
+                    params.push(Box::new(u));
+                }
+                let where_clause = format!("WHERE {}", clauses.join(" AND "));
+
+                let sql = format!(
+                    "SELECT local_id, local_type, create_time, real_sender_id,
+                            message_content, WCDB_CT_message_content
+                     FROM [{}] {} ORDER BY create_time DESC LIMIT ?",
+                    tname, where_clause
+                );
+                params.push(Box::new(per_db_cap as i64));
+
+                let params_ref: Vec<&dyn rusqlite::types::ToSql> =
+                    params.iter().map(|p| p.as_ref()).collect();
+                let mut stmt = conn.prepare(&sql)?;
+                let rows: Vec<(i64, i64, i64, i64, String, i64, i64)> = stmt
+                    .query_map(params_ref.as_slice(), |row| {
+                        let local_id: i64 = row.get(0)?;
+                        let raw_type: i64 = row.get(1)?;
+                        let lo32 = (raw_type as u64 & 0xFFFFFFFF) as i64;
+                        let ts: i64 = row.get(2)?;
+                        let real_sender_id: i64 = row.get(3)?;
+                        let content_bytes = get_content_bytes(row, 4);
+                        let ct: i64 = row.get::<_, i64>(5).unwrap_or(0);
+                        let content = decompress_message(&content_bytes, ct);
+                        let sender = if is_group2 {
+                            sender_label(real_sender_id, &content, true, &uname,
+                                &id2u, &names_map, &group_nicknames2)
+                        } else {
+                            String::new()
+                        };
+                        Ok((local_id, lo32, ts, real_sender_id, sender, ts, db_idx2))
+                    })?
+                    .filter_map(|r| r.ok())
+                    .collect();
+                Ok::<_, anyhow::Error>(rows)
+            })
+            .await??;
+        all_rows.extend(rows);
+    }
+
+    // 全局按 ts DESC 排序后分页
+    all_rows.sort_by_key(|r| std::cmp::Reverse(r.5));
+    let paged: Vec<_> = all_rows.into_iter().skip(offset).take(limit).collect();
+
+    // 翻成 JSON
+    let mut results: Vec<Value> = Vec::with_capacity(paged.len());
+    for (local_id, lo32, ts, _real_sender_id, sender, _ts2, _db_idx) in paged {
+        let kind = type_to_kind.get(&lo32).copied()
+            .unwrap_or(AttachmentKind::Image); // 理论不会 fallthrough
+        let id = AttachmentId {
+            v: 1,
+            chat: username.clone(),
+            local_id,
+            create_time: ts,
+            kind,
+            db: None,
+        };
+        let id_str = id.encode()?;
+
+        let mut row = json!({
+            "attachment_id": id_str,
+            "kind": kind.as_str(),
+            "type": fmt_type(lo32),
+            "local_id": local_id,
+            "timestamp": ts,
+            "time": fmt_time(ts, "%Y-%m-%d %H:%M"),
+        });
+        if is_group && !sender.is_empty() {
+            row["sender"] = Value::String(sender);
+        }
+        results.push(row);
+    }
+
+    Ok(json!({
+        "chat": display,
+        "username": username,
+        "is_group": is_group,
+        "chat_type": chat_type,
+        "count": results.len(),
+        "attachments": results,
+    }))
+}
+
+/// 解码 attachment_id → 查 message_resource.db → 找本地 .dat → 解密 → 写盘。
+pub async fn q_extract(
+    db: &DbCache,
+    _names: &Names,
+    attachment_id: &str,
+    output: &str,
+    overwrite: bool,
+) -> Result<Value> {
+    use crate::attachment::{
+        attachment_id::AttachmentId,
+        decoder::{self, V2KeyMaterial},
+        image_key,
+        resolver,
+    };
+
+    let id = AttachmentId::decode(attachment_id)
+        .context("解析 attachment_id 失败（不是合法 base64url(json)？）")?;
+
+    let output_path = std::path::PathBuf::from(output);
+    if output_path.exists() && !overwrite {
+        anyhow::bail!(
+            "目标已存在：{}（加 --overwrite 覆盖）",
+            output_path.display()
+        );
+    }
+    if let Some(parent) = output_path.parent() {
+        if !parent.as_os_str().is_empty() {
+            tokio::fs::create_dir_all(parent).await
+                .with_context(|| format!("创建输出目录失败：{}", parent.display()))?;
+        }
+    }
+
+    // 1) 拿 message_resource.db
+    let resource_path = db.get("message/message_resource.db").await?
+        .context("无法解密 message_resource.db（请确认 all_keys.json 包含该 DB 的密钥）")?;
+
+    // 2) 推 wxchat_base = db_dir.parent()，再拼 attach_root
+    let wxchat_base = db.db_dir().parent()
+        .ok_or_else(|| anyhow::anyhow!("db_dir 没有 parent，无法推断 xwechat_files 根目录"))?
+        .to_path_buf();
+    let attach_root = resolver::attach_root_for(&wxchat_base);
+
+    // 3) blocking pool 跑 resolver + 读盘 + 解码
+    let id_for_task = id.clone();
+    let resource_path2 = resource_path.clone();
+    let attach_root2 = attach_root.clone();
+    let wxchat_base2 = wxchat_base.clone();
+    let output_path2 = output_path.clone();
+
+    let report: Value = tokio::task::spawn_blocking(move || -> Result<Value> {
+        let resolved = resolver::resolve_blocking(&id_for_task, &resource_path2, &attach_root2)?;
+
+        let dat_bytes = std::fs::read(&resolved.dat_path)
+            .with_context(|| format!("读取 .dat 失败：{}", resolved.dat_path.display()))?;
+
+        // V2 image key — 平台相关。`ImageKeyMaterial` 同时给 aes_key + xor_key。
+        // xor_key 不能硬编码 0x88：实测 macOS 真实账号上是 `uin & 0xff` 派生的（0xa2 等），
+        // 所以这里桥接时必须把 provider 的 xor_key 透传给 V2KeyMaterial。
+        // 缺 key 时让 decoder 自己抛带诊断的错。
+        let provider = image_key::default_provider();
+        let key_material = if let Some(p) = provider.as_ref() {
+            // 从 wxchat_base 末段拿 wxid
+            let wxid = wxchat_base2.file_name()
+                .and_then(|s| s.to_str())
+                .unwrap_or_default()
+                .to_string();
+            if wxid.is_empty() {
+                None
+            } else {
+                match p.get_key(&wxid) {
+                    Ok(km) => Some(km),
+                    Err(e) => {
+                        eprintln!("[extract] image key 提取失败 (wxid={}): {} — V2 文件将无法解码", wxid, e);
+                        None
+                    }
+                }
+            }
+        } else {
+            None
+        };
+        let v2_key = match key_material.as_ref() {
+            Some(km) => V2KeyMaterial { aes_key: Some(&km.aes_key), xor_key: km.xor_key },
+            None => V2KeyMaterial::default(),
+        };
+
+        let decoded = decoder::dispatch(&dat_bytes, v2_key)?;
+
+        // 写盘
+        std::fs::write(&output_path2, &decoded.data)
+            .with_context(|| format!("写出文件失败：{}", output_path2.display()))?;
+
+        // 注意：不要在这里塞 `ok: true`。dispatch 会用 Response::ok(v) 包一层，
+        // Response 的 `data: Value` 字段是 #[serde(flatten)] 写出的，本 payload
+        // 的 `ok` 会和 Response 自带的 `ok` 在线上拼成两个同名 key，CLI 反序列化时
+        // serde_json 直接报 "duplicate field"，业务请求看上去像 daemon 解析失败。
+        Ok(json!({
+            "kind": id_for_task.kind.as_str(),
+            "md5": resolved.md5,
+            "dat_path": resolved.dat_path.display().to_string(),
+            "dat_size": resolved.size,
+            "output": output_path2.display().to_string(),
+            "output_size": decoded.data.len(),
+            "format": decoded.format,
+            "decoder": decoded.decoder,
+        }))
+    }).await??;
+
+    Ok(report)
+}
+
+/// 解析 `kinds` 参数到 `(AttachmentKind, lo32_local_type)` 列表。
+/// 当前只支持 image；命令名保留成 `attachments` 是为了后续扩到其他附件类型时不 break CLI。
+fn parse_attachment_kinds(
+    kinds: Option<&[String]>,
+) -> Result<Vec<(crate::attachment::AttachmentKind, i64)>> {
+    use crate::attachment::AttachmentKind;
+    let raw = kinds.unwrap_or(&[]);
+    if raw.is_empty() {
+        return Ok(vec![(AttachmentKind::Image, 3)]);
+    }
+    let mut out: Vec<(AttachmentKind, i64)> = Vec::with_capacity(raw.len());
+    let mut seen = HashSet::<&'static str>::new();
+    for k in raw {
+        let (kind, t): (AttachmentKind, i64) = match k.to_ascii_lowercase().as_str() {
+            "image" | "img" => (AttachmentKind::Image, 3),
+            "voice" | "audio" | "video" | "file" => {
+                anyhow::bail!("当前只支持 image 提取；video/file/voice 的资源路径与 decoder 还没接通")
+            }
+            other => anyhow::bail!("未知附件类型：{}（当前仅支持 image）", other),
+        };
+        if seen.insert(kind.as_str()) {
+            out.push((kind, t));
+        }
+    }
+    Ok(out)
+}
+
 #[cfg(test)]
 mod biz_tests {
     use super::*;

src/daemon/server.rs

@@ -240,5 +240,17 @@ async fn dispatch(
                 Err(e) => Response::err(e.to_string()),
             }
         }
+        Attachments { chat, kinds, limit, offset, since, until } => {
+            match query::q_attachments(db, &names_arc, &chat, kinds, limit, offset, since, until).await {
+                Ok(v) => Response::ok(v),
+                Err(e) => Response::err(e.to_string()),
+            }
+        }
+        Extract { attachment_id, output, overwrite } => {
+            match query::q_extract(db, &names_arc, &attachment_id, &output, overwrite).await {
+                Ok(v) => Response::ok(v),
+                Err(e) => Response::err(e.to_string()),
+            }
+        }
     }
 }

src/ipc.rs

@@ -131,6 +131,32 @@ pub enum Request {
     },
     /// 重新加载配置和密钥（init --force 后 daemon 不会自动重读）
     ReloadConfig,
+    /// 列出某个会话里的图片附件
+    /// 输出每条带 `attachment_id`（不透明 base64url 句柄），传给 `Extract` 时取回本体
+    Attachments {
+        chat: String,
+        /// 类型过滤：当前仅支持 image
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        kinds: Option<Vec<String>>,
+        #[serde(default = "default_limit_50")]
+        limit: usize,
+        #[serde(default)]
+        offset: usize,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        since: Option<i64>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        until: Option<i64>,
+    },
+    /// 提取（解密）单个附件的本体到指定路径
+    Extract {
+        /// `Attachments` 返回的不透明 ID
+        attachment_id: String,
+        /// 写入的绝对路径（daemon 直接写盘，不经 socket 传 binary）
+        output: String,
+        /// 已存在时是否覆盖
+        #[serde(default)]
+        overwrite: bool,
+    },
 }
 
 

src/main.rs

@@ -4,6 +4,7 @@ mod crypto;
 mod scanner;
 mod daemon;
 mod cli;
+mod attachment;
 
 fn main() {
     if std::env::var("WX_DAEMON_MODE").is_ok() {