feat(agent): 抽出 agent-core 并新增 Linux 被控端（X11 + XTest）

- 新增 agent-core：平台无关会话层（QUIC/TLS 握手、验证码认证、H.264 编码下行、
  latest-wins 丢帧、输入分发），平台差异经 ScreenCapturer / InputInjector 两个
  trait 注入；Windows agent 一并改接过来，避免两端各写一份握手/认证造成协议漂移
- 新增 linux-agent：X11 GetImage 采集 + 整帧变更检测；XTest 注入 + 扫描码→X11
  keycode 映射（主键区 = scancode+8，E0 扩展键必须查表，含单元测试）
- 协议零改动：复用 Set-1 扫描码与虚拟桌面归一化坐标，viewer 无需任何修改
- 修复：静止桌面下关键帧请求无人响应——解码端一旦出错将永远无法恢复。
  现在保留最近一帧裸像素，收到 KeyFrameReq 时重编一个真关键帧
- 修复：错误验证码时 AuthFail 送不达——进程退出会带走 QUIC 缓冲区内尚未发出的
  数据，客户端只能看到 connection lost；现等对端挂断再退出
- 新增端到端冒烟测试 tests/smoke.rs 与 scripts/linux-smoke.sh（自建 Xvfb，无需
  图形桌面）：覆盖采集→编码→QUIC 解帧、认证（含错误码拒绝）、Ping/Pong、
  静止桌面关键帧恢复，以及输入注入实证（注入后从另一条 X11 连接读回指针坐标）
- 该关键帧用例已做变异测试：退回修复会红，恢复后转绿，确认断言有效
- 手册：修正已失效的工具链路径，补充 cargo 镜像说明与 Linux 章节

验证：
- Windows: cargo build --release 通过，产物 agent.exe 9.32MB / viewer.exe 8.73MB
- Linux(容器): cargo test -p linux-agent 通过，2 passed

.gitattributes

@@ -0,0 +1,15 @@
+# 统一换行符，避免跨平台踩坑
+# shell 脚本必须 LF：Windows 上若被 checkout 成 CRLF，容器/Linux 里执行会报
+# "$'\r': command not found"，或被 sed/grep 当成带 \r 的内容
+*.sh text eol=lf
+
+# Windows 侧脚本保持 CRLF
+*.ps1 text eol=crlf
+*.bat text eol=crlf
+
+# 二进制/生成物不参与换行转换
+*.exe binary
+*.dll binary
+*.png binary
+*.jpg binary
+*.ico binary

Cargo.lock

@@ -5,16 +5,25 @@ version = 4
 [[package]]
 name = "agent"
 version = "0.1.0"
+dependencies = [
+ "agent-core",
+ "anyhow",
+ "common",
+ "tokio",
+ "windows",
+]
+
+[[package]]
+name = "agent-core"
+version = "0.1.0"
 dependencies = [
  "anyhow",
  "common",
- "lz4_flex",
  "openh264",
  "quinn",
  "rcgen",
  "rustls",
  "tokio",
- "windows",
 ]
 
 [[package]]
@@ -303,6 +312,16 @@ dependencies = [
  "slab",
 ]
 
+[[package]]
+name = "gethostname"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1bd49230192a3797a9a4d6abe9b3eed6f7fa4c8a8a4947977c6f80025f92cbd8"
+dependencies = [
+ "rustix",
+ "windows-link",
+]
+
 [[package]]
 name = "getrandom"
 version = "0.2.17"
@@ -452,6 +471,19 @@ dependencies = [
  "redox_syscall 0.9.5",
 ]
 
+[[package]]
+name = "linux-agent"
+version = "0.1.0"
+dependencies = [
+ "agent-core",
+ "anyhow",
+ "common",
+ "quinn",
+ "rustls",
+ "tokio",
+ "x11rb",
+]
+
 [[package]]
 name = "linux-raw-sys"
 version = "0.12.1"
@@ -1673,6 +1705,23 @@ dependencies = [
  "pkg-config",
 ]
 
+[[package]]
+name = "x11rb"
+version = "0.13.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9993aa5be5a26815fe2c3eacfc1fde061fc1a1f094bf1ad2a18bf9c495dd7414"
+dependencies = [
+ "gethostname",
+ "rustix",
+ "x11rb-protocol",
+]
+
+[[package]]
+name = "x11rb-protocol"
+version = "0.13.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ea6fc2961e4ef194dcbfe56bb845534d0dc8098940c7e5c012a258bfec6701bd"
+
 [[package]]
 name = "xcursor"
 version = "0.3.11"

Cargo.toml

@@ -1,6 +1,10 @@
 [workspace]
 resolver = "2"
-members = ["common", "agent", "viewer"]
+# agent-core：平台无关的被控端会话层（Windows/Linux 共用）；linux-agent：Linux 被控端
+members = ["common", "agent-core", "agent", "linux-agent", "viewer"]
+# 默认成员保持 Windows 三件套，`cargo build --release` 依旧只出 agent.exe / viewer.exe；
+# Linux 被控端显式构建：`cargo build --release -p linux-agent`
+default-members = ["common", "agent", "viewer"]
 
 [workspace.dependencies]
 anyhow = "1"

agent-core/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "agent-core"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+common = { path = "../common" }
+anyhow.workspace = true
+tokio.workspace = true
+quinn.workspace = true
+rustls.workspace = true
+rcgen.workspace = true
+openh264 = { version = "0.9.9", features = ["source"] }

agent-core/src/lib.rs

@@ -0,0 +1,398 @@
+//! 平台无关的被控端会话层：QUIC/TLS 握手、验证码认证、H.264 编码下行、输入分发。
+//!
+//! 从 `agent/`（Windows）抽出来，供 Windows 与 Linux 被控端共用，避免两端各写一份
+//! QUIC 握手/认证/丢帧逻辑造成协议漂移。平台差异只通过两个 trait 注入：
+//!
+//! - [`ScreenCapturer`]：屏幕采集（Windows = DXGI Desktop Duplication，Linux = X11）
+//! - [`InputInjector`]：输入注入（Windows = SendInput，Linux = XTest）
+//!
+//! 线格式完全复用 `common`，**不改协议**：键盘沿用 Set-1 扫描码 + extended 标志，
+//! 鼠标沿用虚拟桌面归一化坐标 0..=65535。
+
+use anyhow::Result;
+use common::{
+    encode_video_frame, read_control, write_control, ControlMsg, InputEvent, MouseButton,
+    VideoFrameHeader,
+};
+use openh264::encoder::{
+    BitRate, Encoder, EncoderConfig, FrameRate, IntraFramePeriod, RateControlMode, UsageType,
+};
+use openh264::formats::{BgraSliceU8, YUVBuffer};
+use openh264::OpenH264API;
+use rustls::pki_types::PrivatePkcs8KeyDer;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+use tokio::sync::{mpsc, watch};
+
+/// 屏幕采集后端。
+///
+/// 语义刻意对齐 DXGI Desktop Duplication，使 Linux 侧只需实现同一套模型：
+/// `acquire()` 阻塞至多 `timeout_ms`，返回是否拿到**新**帧；`copy_frame()` 再取出像素。
+pub trait ScreenCapturer: Send {
+    /// 桌面尺寸（物理像素）
+    fn size(&self) -> (u32, u32);
+
+    /// 阻塞至多 `timeout_ms`。
+    /// `Ok(true)` = 有新帧（已拷进内部缓冲，可由 `copy_frame` 取出）；
+    /// `Ok(false)` = 超时且画面未变（静止桌面）。
+    fn acquire(&mut self, timeout_ms: u32) -> Result<bool>;
+
+    /// 取出最近一帧，BGRA 紧凑排列（已去掉 row pitch），供 openh264 直接编码。
+    fn copy_frame(&mut self) -> Result<Vec<u8>>;
+}
+
+/// 输入注入后端。实现需可在异步运行时中被多任务共享，故要求 `Send + Sync`。
+pub trait InputInjector: Send + Sync + 'static {
+    fn mouse_move(&self, x: u16, y: u16);
+    fn mouse_button(&self, b: MouseButton, down: bool);
+    fn wheel(&self, dy: i32);
+    fn key(&self, scancode: u16, extended: bool, down: bool);
+}
+
+/// 把一条协议输入事件分发到注入后端。
+pub fn dispatch_input(inj: &dyn InputInjector, ev: InputEvent) {
+    match ev {
+        InputEvent::MouseMove { x, y } => inj.mouse_move(x, y),
+        InputEvent::MouseDown(b) => inj.mouse_button(b, true),
+        InputEvent::MouseUp(b) => inj.mouse_button(b, false),
+        InputEvent::Wheel { dy } => inj.wheel(dy),
+        InputEvent::KeyDown { scancode, extended } => inj.key(scancode, extended, true),
+        InputEvent::KeyUp { scancode, extended } => inj.key(scancode, extended, false),
+    }
+}
+
+/// 会话参数
+#[derive(Debug, Clone)]
+pub struct SessionConfig {
+    pub port: u16,
+    pub code: String,
+}
+
+/// 编码档位。默认值即 Windows 现有档位，改动需两端同步评估。
+#[derive(Debug, Clone, Copy)]
+pub struct EncoderSettings {
+    pub bitrate_bps: u32,
+    pub max_fps: f32,
+    pub intra_period: u32,
+}
+
+impl Default for EncoderSettings {
+    fn default() -> Self {
+        // 对标 UU：屏幕内容实时档；关键帧周期 300 帧 ≈ 10s
+        Self {
+            bitrate_bps: 5_000_000,
+            max_fps: 30.0,
+            intra_period: 300,
+        }
+    }
+}
+
+/// 解析 `--port` / `--code`（两端共用，保持命令行一致）
+pub fn parse_args() -> Result<(u16, String)> {
+    let mut port = 47800u16;
+    let mut code = "123456".to_string();
+    let a: Vec<String> = std::env::args().collect();
+    let mut i = 1;
+    while i + 1 < a.len() {
+        match a[i].as_str() {
+            "--port" => port = a[i + 1].parse()?,
+            "--code" => code = a[i + 1].clone(),
+            _ => {}
+        }
+        i += 2;
+    }
+    Ok((port, code))
+}
+
+fn make_encoder(enc: &EncoderSettings) -> Result<Encoder> {
+    Ok(Encoder::with_api_config(
+        OpenH264API::from_source(),
+        EncoderConfig::new()
+            .bitrate(BitRate::from_bps(enc.bitrate_bps))
+            .max_frame_rate(FrameRate::from_hz(enc.max_fps))
+            .rate_control_mode(RateControlMode::Bitrate)
+            .usage_type(UsageType::ScreenContentRealTime)
+            .intra_frame_period(IntraFramePeriod::from_num_frames(enc.intra_period)),
+    )?)
+}
+
+enum LoopEnd {
+    /// 下行通道关闭 —— 会话结束，采集线程退出
+    ChannelClosed,
+    /// 采集后端失效（Windows 的 ACCESS_LOST / X11 连接断开）—— 重建后继续
+    Rebuild,
+}
+
+/// 编码一帧并推入下行通道。返回 false 表示通道已关闭（会话结束）。
+fn encode_send(
+    encoder: &mut Encoder,
+    raw: &[u8],
+    (w, h): (u32, u32),
+    ts_ms: u64,
+    tx: &watch::Sender<Option<Vec<u8>>>,
+    last_frame: &mut Option<Vec<u8>>,
+) -> bool {
+    let header = VideoFrameHeader {
+        width: w,
+        height: h,
+        ts_ms,
+    };
+    let yuv = YUVBuffer::from_bgra8_source(BgraSliceU8::new(raw, (w as usize, h as usize)));
+    match encoder.encode(&yuv) {
+        Ok(ebs) => {
+            let frame = encode_video_frame(&header, &ebs.to_vec());
+            if tx.send(Some(frame.clone())).is_err() {
+                return false;
+            }
+            *last_frame = Some(frame);
+            true
+        }
+        Err(e) => {
+            eprintln!("[capture] encode failed: {e}");
+            true
+        }
+    }
+}
+
+/// 单次采集会话：抓帧 → 编码 → 推入 watch 单槽。
+fn capture_session(
+    cap: &mut dyn ScreenCapturer,
+    tx: &watch::Sender<Option<Vec<u8>>>,
+    keyreq: &AtomicBool,
+    enc: &EncoderSettings,
+) -> LoopEnd {
+    let mut encoder = match make_encoder(enc) {
+        Ok(e) => e,
+        Err(e) => {
+            eprintln!("[capture] encoder init failed: {e}");
+            return LoopEnd::ChannelClosed;
+        }
+    };
+
+    let (w, h) = cap.size();
+    eprintln!("[capture] started {w}x{h}");
+    let start = Instant::now();
+    let min_frame = Duration::from_secs_f32(1.0 / enc.max_fps);
+    let mut last_frame: Option<Vec<u8>> = None;
+    // 保留最近一帧裸像素：静止桌面收到关键帧请求时，仍能重编一个真关键帧
+    let mut last_raw: Option<Vec<u8>> = None;
+    let mut last_sent = Instant::now() - Duration::from_secs(10);
+
+    loop {
+        let iter_start = Instant::now();
+        match cap.acquire(80) {
+            Ok(true) => match cap.copy_frame() {
+                Ok(raw) => {
+                    // 主控端请求关键帧（解码异常恢复）
+                    if keyreq.swap(false, Ordering::Relaxed) {
+                        encoder.force_intra_frame();
+                    }
+                    let ts = start.elapsed().as_millis() as u64;
+                    if !encode_send(&mut encoder, &raw, (w, h), ts, tx, &mut last_frame) {
+                        return LoopEnd::ChannelClosed; // 会话结束
+                    }
+                    last_raw = Some(raw);
+                    last_sent = Instant::now();
+                    // 帧间隔节流到 max_fps
+                    let el = iter_start.elapsed();
+                    if el < min_frame {
+                        std::thread::sleep(min_frame - el);
+                    }
+                }
+                Err(e) => {
+                    eprintln!("[capture] frame copy failed: {e}");
+                    return LoopEnd::Rebuild;
+                }
+            },
+            Ok(false) => {
+                let force = keyreq.swap(false, Ordering::Relaxed);
+                if force && last_raw.is_some() {
+                    // 静止桌面下的关键帧恢复：画面没变不代表可以不给关键帧——
+                    // 解码端一旦出错，若只会收到旧 P 帧就永远恢复不了
+                    encoder.force_intra_frame();
+                    let raw = last_raw.as_ref().expect("上面已判断 is_some");
+                    let ts = start.elapsed().as_millis() as u64;
+                    if !encode_send(&mut encoder, raw, (w, h), ts, tx, &mut last_frame) {
+                        return LoopEnd::ChannelClosed;
+                    }
+                    last_sent = Instant::now();
+                } else if last_sent.elapsed() > Duration::from_millis(500) && last_frame.is_some() {
+                    // 桌面静止：低频重发最后一帧（重复 P 帧，解码端画面不变）
+                    let f = last_frame.as_ref().expect("上面已判断 is_some");
+                    if tx.send(Some(f.clone())).is_err() {
+                        return LoopEnd::ChannelClosed;
+                    }
+                    last_sent = Instant::now();
+                } else {
+                    std::thread::sleep(Duration::from_millis(8));
+                }
+            }
+            Err(e) => {
+                eprintln!("[capture] acquire failed: {e}");
+                return LoopEnd::Rebuild;
+            }
+        }
+    }
+}
+
+/// 采集主循环：独立线程运行；后端失效时经 `factory` 重建，通道关闭时退出。
+/// 通道为 watch 单槽（latest-wins）：发送即覆盖，网络慢时不产生积压。
+pub fn run_capture_loop<F>(
+    tx: watch::Sender<Option<Vec<u8>>>,
+    keyreq: Arc<AtomicBool>,
+    mut cap: Box<dyn ScreenCapturer>,
+    mut factory: F,
+    enc: EncoderSettings,
+) where
+    F: FnMut() -> Result<Box<dyn ScreenCapturer>>,
+{
+    loop {
+        match capture_session(cap.as_mut(), &tx, keyreq.as_ref(), &enc) {
+            LoopEnd::ChannelClosed => return,
+            LoopEnd::Rebuild => {}
+        }
+        std::thread::sleep(Duration::from_millis(300));
+        loop {
+            match factory() {
+                Ok(c) => {
+                    cap = c;
+                    break;
+                }
+                Err(e) => {
+                    eprintln!("[capture] init failed: {e}, retry in 2s");
+                    std::thread::sleep(Duration::from_secs(2));
+                }
+            }
+        }
+    }
+}
+
+/// 跑一次完整的被控端会话：监听 → 认证 → 视频下行 + 输入上行。
+///
+/// `capture_factory` 可被多次调用（采集后端失效时重建），故需为 `FnMut`。
+pub async fn serve<F>(
+    cfg: SessionConfig,
+    enc: EncoderSettings,
+    mut capture_factory: F,
+    injector: Arc<dyn InputInjector>,
+) -> Result<()>
+where
+    F: FnMut() -> Result<Box<dyn ScreenCapturer>> + Send + 'static,
+{
+    rustls::crypto::ring::default_provider()
+        .install_default()
+        .ok();
+
+    // 自签名证书（MVP：客户端跳过校验；正式版改为固定公钥/账号体系）
+    let ck = rcgen::generate_simple_self_signed(vec!["localhost".into()])?;
+    let cert = ck.cert.der().clone();
+    let key = PrivatePkcs8KeyDer::from(ck.key_pair.serialize_der());
+
+    // 传输参数调优：
+    // - 收发窗口放大，避免高 RTT 下吞吐被 窗口/RTT 死锁（默认 512KB 流窗口）
+    // - 发送缓冲上限 2MB：QUIC 内部排队有界（约 1 帧多），配合上游 latest-wins 丢帧
+    //   把端到端排队延迟压在 ~0.4s 内
+    let mut transport = quinn::TransportConfig::default();
+    transport.receive_window(quinn::VarInt::from_u32(16 * 1024 * 1024));
+    transport.stream_receive_window(quinn::VarInt::from_u32(8 * 1024 * 1024));
+    transport.send_window(2 * 1024 * 1024);
+    let mut server_config = quinn::ServerConfig::with_single_cert(vec![cert], key.into())?;
+    server_config.transport_config(Arc::new(transport));
+
+    let endpoint = quinn::Endpoint::server(server_config, format!("0.0.0.0:{}", cfg.port).parse()?)?;
+    println!(
+        "[agent] listening 0.0.0.0:{}, verify code = {}",
+        cfg.port, cfg.code
+    );
+
+    let conn = endpoint
+        .accept()
+        .await
+        .ok_or_else(|| anyhow::anyhow!("endpoint closed"))?
+        .await?;
+    println!("[agent] peer connected: {}", conn.remote_address());
+
+    // 流 1：认证 + 视频下行
+    let (mut send, mut recv) = conn.accept_bi().await?;
+
+    let first = read_control(&mut recv).await?;
+    let code_ok = matches!(&first, ControlMsg::Auth { code: c } if *c == cfg.code);
+    if !code_ok {
+        write_control(&mut send, &ControlMsg::AuthFail("bad code".into())).await?;
+        send.finish()?;
+        // 必须等对端读到 AuthFail 并挂断再退出：直接 return 会让进程带走 QUIC
+        // 缓冲区里尚未发出的数据，客户端只会看到 "connection lost" 而不是 AuthFail
+        let _ = tokio::time::timeout(Duration::from_secs(5), conn.closed()).await;
+        anyhow::bail!("auth failed");
+    }
+
+    // 采集后端先建一次：既拿桌面尺寸回给主控端，也避免白建一个 duplication 只为读尺寸
+    let first_cap = capture_factory()?;
+    let (w, h) = first_cap.size();
+    write_control(&mut send, &ControlMsg::AuthOk { width: w, height: h }).await?;
+    println!("[agent] auth ok, desktop {w}x{h}");
+
+    // 流 2：键鼠输入 + 控制上行（与视频分离，避免被视频积压阻塞）
+    let (mut isend, mut irecv) = conn.accept_bi().await?;
+
+    // 采集线程 → 视频帧（watch 单槽：新帧覆盖旧帧，latest-wins）
+    let (tx, mut frames_rx) = watch::channel::<Option<Vec<u8>>>(None);
+    let keyreq = Arc::new(AtomicBool::new(false));
+    let keyreq_capture = keyreq.clone();
+    let enc_settings = enc;
+    std::thread::spawn(move || {
+        run_capture_loop(tx, keyreq_capture, first_cap, capture_factory, enc_settings)
+    });
+
+    // 上行控制消息（Pong）经通道交给主循环写流 2，避免多任务争抢
+    let (ctl_tx, mut ctl_rx) = mpsc::unbounded_channel::<ControlMsg>();
+    let t0 = Instant::now();
+
+    // 读循环（流 2）：输入事件 + Ping
+    let keyreq_reader = keyreq.clone();
+    let reader = tokio::spawn(async move {
+        loop {
+            match read_control(&mut irecv).await {
+                Ok(ControlMsg::Input(ev)) => dispatch_input(injector.as_ref(), ev),
+                Ok(ControlMsg::KeyFrameReq) => {
+                    keyreq_reader.store(true, Ordering::Relaxed);
+                }
+                Ok(ControlMsg::Ping(t)) => {
+                    let agent_ts = t0.elapsed().as_millis() as u64;
+                    let _ = ctl_tx.send(ControlMsg::Pong { t, agent_ts });
+                }
+                Ok(_) => {}
+                Err(_) => break,
+            }
+        }
+    });
+
+    loop {
+        tokio::select! {
+            changed = frames_rx.changed() => {
+                if changed.is_err() {
+                    break; // 采集线程退出
+                }
+                if let Some(data) = frames_rx.borrow().clone() {
+                    if send.write_all(&data).await.is_err() {
+                        break;
+                    }
+                }
+            }
+            ctl = ctl_rx.recv() => {
+                match ctl {
+                    Some(m) => {
+                        if write_control(&mut isend, &m).await.is_err() {
+                            break;
+                        }
+                    }
+                    None => break,
+                }
+            }
+        }
+    }
+    reader.abort();
+    println!("[agent] session closed");
+    Ok(())
+}

agent/Cargo.toml

@@ -4,14 +4,11 @@ version = "0.1.0"
 edition = "2021"
 
 [dependencies]
+# 会话层（QUIC/TLS/认证/编码/丢帧）在 agent-core，与 linux-agent 共用
+agent-core = { path = "../agent-core" }
 common = { path = "../common" }
 anyhow.workspace = true
 tokio.workspace = true
-quinn.workspace = true
-rustls.workspace = true
-rcgen.workspace = true
-lz4_flex.workspace = true
-openh264 = { version = "0.9.9", features = ["source"] }
 windows = { workspace = true, features = [
     "Win32_Foundation",
     "Win32_Graphics_Direct3D",

agent/src/capture.rs

@@ -1,25 +1,19 @@
-//! DXGI Desktop Duplication 采集（Windows 8+）。
+//! Windows 屏幕采集：DXGI Desktop Duplication（Windows 8+）。
-//! 帧处理：BGRA → openh264 H.264 编码（屏幕内容实时档，5Mbps）→ 线格式。
+//!
-//! 支持主控端 KeyFrameReq 强制关键帧（解码异常恢复）。
+//! 采集循环、H.264 编码、丢帧与下行都在 `agent-core`（两端共用），本模块只负责
+//! DXGI 的「取新帧 → 拷到 staging 纹理 → 映射内存」这三步。
+//!
+//! `ScreenCapturer` 的语义与 Linux 端一致：`acquire()` 返回 false 即代表超时且画面未变，
+//! 由 agent-core 决定低频重发上一帧。
 
+use agent_core::ScreenCapturer;
 use anyhow::{anyhow, Result};
-use common::{encode_video_frame, VideoFrameHeader};
-use openh264::encoder::{
-    BitRate, Encoder, EncoderConfig, FrameRate, IntraFramePeriod, RateControlMode, UsageType,
-};
-use openh264::formats::{BgraSliceU8, YUVBuffer};
-use openh264::OpenH264API;
-use std::sync::atomic::{AtomicBool, Ordering};
-use std::sync::Arc;
-use tokio::sync::watch::Sender;
 use windows::core::Interface;
 use windows::Win32::Foundation::HMODULE;
-use windows::Win32::Graphics::Direct3D::{
+use windows::Win32::Graphics::Direct3D::{D3D_DRIVER_TYPE_HARDWARE, D3D_FEATURE_LEVEL_11_0};
-    D3D_DRIVER_TYPE_HARDWARE, D3D_FEATURE_LEVEL_11_0,
-};
 use windows::Win32::Graphics::Direct3D11::{
     D3D11CreateDevice, ID3D11Device, ID3D11DeviceContext, ID3D11Resource, ID3D11Texture2D,
-    D3D11_CREATE_DEVICE_FLAG, D3D11_CPU_ACCESS_READ, D3D11_MAP, D3D11_MAP_READ,
+    D3D11_CREATE_DEVICE_FLAG, D3D11_CPU_ACCESS_READ, D3D11_MAP_READ,
     D3D11_MAPPED_SUBRESOURCE, D3D11_SDK_VERSION, D3D11_TEXTURE2D_DESC, D3D11_USAGE_STAGING,
 };
 use windows::Win32::Graphics::Dxgi::Common::{DXGI_FORMAT_B8G8R8A8_UNORM, DXGI_SAMPLE_DESC};
@@ -28,8 +22,6 @@ use windows::Win32::Graphics::Dxgi::{
     DXGI_OUTDUPL_DESC, DXGI_OUTDUPL_FRAME_INFO,
 };
 
-use std::time::{Duration, Instant};
-
 pub struct Duplication {
     _device: ID3D11Device,
     context: ID3D11DeviceContext,
@@ -105,9 +97,15 @@ impl Duplication {
             })
         }
     }
+}
+
+impl ScreenCapturer for Duplication {
+    fn size(&self) -> (u32, u32) {
+        (self.width, self.height)
+    }
 
     /// true=拿到新帧并已复制到 staging；false=超时无更新
-    pub fn acquire(&mut self, timeout_ms: u32) -> Result<bool> {
+    fn acquire(&mut self, timeout_ms: u32) -> Result<bool> {
         unsafe {
             let mut info = DXGI_OUTDUPL_FRAME_INFO::default();
             let mut resource: Option<IDXGIResource> = None;
@@ -126,7 +124,7 @@ impl Duplication {
     }
 
     /// 把 staging 纹理映射到内存并去掉行间距
-    pub fn copy_frame(&mut self) -> Result<Vec<u8>> {
+    fn copy_frame(&mut self) -> Result<Vec<u8>> {
         unsafe {
             let mut mapped = D3D11_MAPPED_SUBRESOURCE::default();
             self.context
@@ -145,112 +143,7 @@ impl Duplication {
     }
 }
 
-#[allow(dead_code)]
+/// 采集器工厂：agent-core 在采集后端失效时会重新调用
-pub fn desktop_size() -> Result<(u32, u32)> {
+pub fn new_capturer() -> Result<Box<dyn ScreenCapturer>> {
-    let d = Duplication::new()?;
+    Ok(Box::new(Duplication::new()?))
-    Ok((d.width, d.height))
-}
-
-/// 采集主循环：独立线程运行；连接断开（通道关闭）时退出。
-/// 通道为 watch 单槽（latest-wins）：发送即覆盖，网络慢时不产生积压。
-pub fn run(tx: Sender<Option<Vec<u8>>>, keyreq: Arc<AtomicBool>) {
-    let start = Instant::now();
-    loop {
-        let mut dup = match Duplication::new() {
-            Ok(d) => d,
-            Err(e) => {
-                eprintln!("[capture] init failed: {e}, retry in 2s");
-                std::thread::sleep(Duration::from_secs(2));
-                continue;
-            }
-        };
-        eprintln!(
-            "[capture] duplication started {}x{}",
-            dup.width, dup.height
-        );
-
-        // H.264 编码器：屏幕内容实时档 + 码率控制（对标 UU：3.2Mbps@60fps）
-        let mut encoder = match Encoder::with_api_config(
-            OpenH264API::from_source(),
-            EncoderConfig::new()
-                .bitrate(BitRate::from_bps(5_000_000))
-                .max_frame_rate(FrameRate::from_hz(30.0))
-                .rate_control_mode(RateControlMode::Bitrate)
-                .usage_type(UsageType::ScreenContentRealTime)
-                .intra_frame_period(IntraFramePeriod::from_num_frames(300)),
-        ) {
-            Ok(e) => e,
-            Err(e) => {
-                eprintln!("[capture] encoder init failed: {e}");
-                return;
-            }
-        };
-
-        let mut last_frame: Option<Vec<u8>> = None;
-        let mut last_sent = Instant::now() - Duration::from_secs(10);
-        let min_frame = Duration::from_millis(33); // 30fps 上限
-        loop {
-            let iter_start = Instant::now();
-            match dup.acquire(80) {
-                Ok(true) => match dup.copy_frame() {
-                    Ok(raw) => {
-                        // 主控端请求关键帧（解码异常恢复）
-                        if keyreq.swap(false, Ordering::Relaxed) {
-                            encoder.force_intra_frame();
-                        }
-                        let header = VideoFrameHeader {
-                            width: dup.width,
-                            height: dup.height,
-                            ts_ms: start.elapsed().as_millis() as u64,
-                        };
-                        let yuv = YUVBuffer::from_bgra8_source(BgraSliceU8::new(
-                            &raw,
-                            (dup.width as usize, dup.height as usize),
-                        ));
-                        match encoder.encode(&yuv) {
-                            Ok(ebs) => {
-                                let payload = ebs.to_vec();
-                                let frame = encode_video_frame(&header, &payload);
-                                if tx.send(Some(frame.clone())).is_err() {
-                                    return; // 会话结束
-                                }
-                                last_frame = Some(frame);
-                                last_sent = Instant::now();
-                            }
-                            Err(e) => {
-                                eprintln!("[capture] encode failed: {e}");
-                            }
-                        }
-                        // 帧间隔节流到 ~30fps
-                        let el = iter_start.elapsed();
-                        if el < min_frame {
-                            std::thread::sleep(min_frame - el);
-                        }
-                    }
-                    Err(e) => {
-                        eprintln!("[capture] frame copy failed: {e}");
-                        break; // 重建 duplication
-                    }
-                },
-                Ok(false) => {
-                    // 桌面静止：低频重发最后一帧（重复 P 帧，解码端画面不变）
-                    if last_sent.elapsed() > Duration::from_millis(500) {
-                        if let Some(f) = last_frame.as_ref() {
-                            if tx.send(Some(f.clone())).is_err() {
-                                return;
-                            }
-                            last_sent = Instant::now();
-                        }
-                    } else {
-                        std::thread::sleep(Duration::from_millis(8));
-                    }
-                }
-                Err(e) => {
-                    eprintln!("[capture] acquire failed: {e}");
-                    break; // ACCESS_LOST 等，重建
-                }
-            }
-        }
-        std::thread::sleep(Duration::from_millis(300));
-    }
 }

agent/src/input.rs

@@ -1,6 +1,11 @@
-//! 输入注入：SendInput 实现（鼠标归一化绝对坐标 + 扫描码键盘）。
+//! Windows 输入注入：SendInput（鼠标虚拟桌面绝对坐标 + 扫描码键盘）。
+//!
+//! 协议里的坐标/扫描码语义两端一致，只是落地 API 不同：
+//! - 鼠标：`MOUSEEVENTF_ABSOLUTE | MOUSEEVENTF_VIRTUALDESK`，直接吃 0..65535 归一化值；
+//! - 键盘：`KEYEVENTF_SCANCODE` 直接送 Set-1 扫描码（Linux 端则需换算成 X11 keycode）。
 
-use common::{InputEvent, MouseButton};
+use agent_core::InputInjector;
+use common::MouseButton;
 use windows::Win32::UI::Input::KeyboardAndMouse::{
     SendInput, INPUT, INPUT_0, INPUT_KEYBOARD, INPUT_MOUSE, KEYBDINPUT, KEYEVENTF_EXTENDEDKEY,
     KEYEVENTF_KEYUP, KEYEVENTF_SCANCODE, MOUSEEVENTF_ABSOLUTE, MOUSEEVENTF_MIDDLEDOWN,
@@ -73,31 +78,49 @@ pub fn virtual_desktop() -> (i32, i32, i32, i32) {
     }
 }
 
-pub fn handle(ev: InputEvent) {
+/// SendInput 注入后端（无状态，直接调 Win32）
-    match ev {
+pub struct WinInput;
-        InputEvent::MouseMove { x, y } => {
+
-            // 主控端已按虚拟桌面归一化，这里直接透传 0..65535
+impl WinInput {
-            send_mouse(
+    pub fn new() -> Self {
-                MOUSEEVENTF_ABSOLUTE.0 | MOUSEEVENTF_VIRTUALDESK.0 | MOUSEEVENTF_MOVE.0,
+        Self
-                x as i32,
+    }
-                y as i32,
+}
-                0,
+
-            );
+impl Default for WinInput {
-        }
+    fn default() -> Self {
-        InputEvent::MouseDown(b) => match b {
+        Self::new()
-            MouseButton::Left => send_mouse(MOUSEEVENTF_LEFTDOWN, 0, 0, 0),
+    }
-            MouseButton::Right => send_mouse(MOUSEEVENTF_RIGHTDOWN.0, 0, 0, 0),
+}
-            MouseButton::Middle => send_mouse(MOUSEEVENTF_MIDDLEDOWN.0, 0, 0, 0),
+
-        },
+impl InputInjector for WinInput {
-        InputEvent::MouseUp(b) => match b {
+    fn mouse_move(&self, x: u16, y: u16) {
-            MouseButton::Left => send_mouse(MOUSEEVENTF_LEFTUP, 0, 0, 0),
+        // 主控端已按虚拟桌面归一化，这里直接透传 0..65535
-            MouseButton::Right => send_mouse(MOUSEEVENTF_RIGHTUP.0, 0, 0, 0),
+        send_mouse(
-            MouseButton::Middle => send_mouse(MOUSEEVENTF_MIDDLEUP.0, 0, 0, 0),
+            MOUSEEVENTF_ABSOLUTE.0 | MOUSEEVENTF_VIRTUALDESK.0 | MOUSEEVENTF_MOVE.0,
-        },
+            x as i32,
-        InputEvent::Wheel { dy } => {
+            y as i32,
-            send_mouse(MOUSEEVENTF_WHEEL.0, 0, 0, dy as u32);
+            0,
-        }
+        );
-        InputEvent::KeyDown { scancode, extended } => send_key(scancode, extended, false),
+    }
-        InputEvent::KeyUp { scancode, extended } => send_key(scancode, extended, true),
+
+    fn mouse_button(&self, b: MouseButton, down: bool) {
+        let flags = match (b, down) {
+            (MouseButton::Left, true) => MOUSEEVENTF_LEFTDOWN,
+            (MouseButton::Left, false) => MOUSEEVENTF_LEFTUP,
+            (MouseButton::Right, true) => MOUSEEVENTF_RIGHTDOWN.0,
+            (MouseButton::Right, false) => MOUSEEVENTF_RIGHTUP.0,
+            (MouseButton::Middle, true) => MOUSEEVENTF_MIDDLEDOWN.0,
+            (MouseButton::Middle, false) => MOUSEEVENTF_MIDDLEUP.0,
+        };
+        send_mouse(flags, 0, 0, 0);
+    }
+
+    fn wheel(&self, dy: i32) {
+        send_mouse(MOUSEEVENTF_WHEEL.0, 0, 0, dy as u32);
+    }
+
+    fn key(&self, scancode: u16, extended: bool, down: bool) {
+        send_key(scancode, extended, !down);
     }
 }

agent/src/main.rs

@@ -1,142 +1,28 @@
-//! 被控端 Agent：DXGI 桌面复制采集 → LZ4 压缩 → QUIC 下行视频；上行接收输入事件注入。
+//! 被控端 Agent（Windows）：DXGI 桌面复制采集 + SendInput 键鼠注入。
-//! P0 优化：视频帧经 watch 单槽通道（latest-wins，网络慢时只发最新帧，不积压）；
+//!
-//! 键鼠/控制走独立 QUIC 双向流，不与视频互相阻塞。
+//! 会话层（QUIC/TLS 握手、验证码认证、H.264 编码下行、latest-wins 丢帧、输入分发）在
+//! `agent-core`，与 `linux-agent` 共用同一份实现；本 crate 只提供平台相关的采集与注入，
+//! 避免两端各写一份握手/认证造成协议漂移。
 
 mod capture;
 mod input;
 
+use agent_core::{parse_args, serve, EncoderSettings, InputInjector, SessionConfig};
 use anyhow::Result;
-use common::{read_control, write_control, ControlMsg};
-use rustls::pki_types::PrivatePkcs8KeyDer;
-use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::Arc;
-use tokio::io::AsyncWriteExt;
-use tokio::sync::{mpsc, watch};
-
-fn parse_args() -> Result<(u16, String)> {
-    let mut port = 47800u16;
-    let mut code = "123456".to_string();
-    let a: Vec<String> = std::env::args().collect();
-    let mut i = 1;
-    while i < a.len() {
-        match a[i].as_str() {
-            "--port" => port = a[i + 1].parse()?,
-            "--code" => code = a[i + 1].clone(),
-            _ => {}
-        }
-        i += 2;
-    }
-    Ok((port, code))
-}
 
 #[tokio::main]
 async fn main() -> Result<()> {
     let (port, code) = parse_args()?;
-    rustls::crypto::ring::default_provider()
-        .install_default()
-        .ok();
 
-    // 自签名证书（MVP：客户端跳过校验；正式版改为固定公钥/账号体系）
+    let injector: Arc<dyn InputInjector> = Arc::new(input::WinInput::new());
-    let ck = rcgen::generate_simple_self_signed(vec!["localhost".into()])?;
-    let cert = ck.cert.der().clone();
-    let key = PrivatePkcs8KeyDer::from(ck.key_pair.serialize_der());
 
-    // 传输参数调优：
+    serve(
-    // - 收发窗口放大，避免高 RTT 下吞吐被 窗口/RTT 死锁（默认 512KB 流窗口）
+        SessionConfig { port, code },
-    // - 发送缓冲上限 2MB：QUIC 内部排队有界（约 1 帧多），配合上游 latest-wins 丢帧
+        EncoderSettings::default(),
-    //   把端到端排队延迟压在 ~0.4s 内
+        // DXGI 后端失效（ACCESS_LOST 等）时 agent-core 会重新调用本工厂
-    let mut transport = quinn::TransportConfig::default();
+        capture::new_capturer,
-    transport.receive_window(quinn::VarInt::from_u32(16 * 1024 * 1024));
+        injector,
-    transport.stream_receive_window(quinn::VarInt::from_u32(8 * 1024 * 1024));
+    )
-    transport.send_window(2 * 1024 * 1024);
+    .await
-    let mut server_config = quinn::ServerConfig::with_single_cert(vec![cert], key.into())?;
-    server_config.transport_config(Arc::new(transport));
-
-    let endpoint = quinn::Endpoint::server(
-        server_config,
-        format!("0.0.0.0:{port}").parse()?,
-    )?;
-    println!("[agent] listening 0.0.0.0:{port}, verify code = {code}");
-
-    let conn = endpoint
-        .accept()
-        .await
-        .ok_or_else(|| anyhow::anyhow!("endpoint closed"))?
-        .await?;
-    println!("[agent] peer connected: {}", conn.remote_address());
-
-    // 流 1：认证 + 视频下行
-    let (mut send, mut recv) = conn.accept_bi().await?;
-
-    // 认证
-    let first = read_control(&mut recv).await?;
-    let code_ok = matches!(&first, ControlMsg::Auth { code: c } if *c == code);
-    if !code_ok {
-        write_control(&mut send, &ControlMsg::AuthFail("bad code".into())).await?;
-        send.finish()?;
-        anyhow::bail!("auth failed");
-    }
-    let (w, h) = capture::desktop_size()?;
-    write_control(&mut send, &ControlMsg::AuthOk { width: w, height: h }).await?;
-    println!("[agent] auth ok, desktop {w}x{h}");
-
-    // 流 2：键鼠输入 + 控制上行（与视频分离，避免被视频积压阻塞）
-    let (mut isend, mut irecv) = conn.accept_bi().await?;
-
-    // 采集线程 → 视频帧（watch 单槽：新帧覆盖旧帧，latest-wins）
-    let (tx, mut frames_rx) = watch::channel::<Option<Vec<u8>>>(None);
-    let keyreq = Arc::new(AtomicBool::new(false));
-    let keyreq_capture = keyreq.clone();
-    std::thread::spawn(move || capture::run(tx, keyreq_capture));
-
-    // 上行控制消息（Pong）经通道交给主循环写流 2，避免多任务争抢
-    let (ctl_tx, mut ctl_rx) = mpsc::unbounded_channel::<ControlMsg>();
-    let t0 = std::time::Instant::now();
-
-    // 读循环（流 2）：输入事件 + Ping
-    let keyreq_reader = keyreq.clone();
-    let reader = tokio::spawn(async move {
-        loop {
-            match read_control(&mut irecv).await {
-                Ok(ControlMsg::Input(ev)) => input::handle(ev),
-                Ok(ControlMsg::KeyFrameReq) => {
-                    keyreq_reader.store(true, Ordering::Relaxed);
-                }
-                Ok(ControlMsg::Ping(t)) => {
-                    let agent_ts = t0.elapsed().as_millis() as u64;
-                    let _ = ctl_tx.send(ControlMsg::Pong { t, agent_ts });
-                }
-                Ok(_) => {}
-                Err(_) => break,
-            }
-        }
-    });
-
-    loop {
-        tokio::select! {
-            changed = frames_rx.changed() => {
-                if changed.is_err() {
-                    break; // 采集线程退出
-                }
-                if let Some(data) = frames_rx.borrow().clone() {
-                    if send.write_all(&data).await.is_err() {
-                        break;
-                    }
-                }
-            }
-            ctl = ctl_rx.recv() => {
-                match ctl {
-                    Some(m) => {
-                        if write_control(&mut isend, &m).await.is_err() {
-                            break;
-                        }
-                    }
-                    None => break,
-                }
-            }
-        }
-    }
-    reader.abort();
-    println!("[agent] session closed");
-    Ok(())
 }

linux-agent/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "linux-agent"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+agent-core = { path = "../agent-core" }
+common = { path = "../common" }
+anyhow.workspace = true
+tokio.workspace = true
+x11rb = { version = "0.13", features = ["xtest"] }
+
+[dev-dependencies]
+# 冒烟测试里要扮演主控端，自己发起 QUIC/TLS 连接
+quinn.workspace = true
+rustls.workspace = true

linux-agent/src/capture.rs

@@ -0,0 +1,147 @@
+//! X11 屏幕采集。
+//!
+//! MVP 走核心协议 `GetImage`：零扩展依赖、零 unsafe，先把「采集 → 编码 → QUIC」整条链路打通。
+//! 变更检测用整帧字节比较，桌面静止时不上行视频（省掉编码开销）。
+//!
+//! 二期优化锚点（已在 `ScreenCapturer` 语义内预留，无需改 agent-core）：
+//! 1. **MIT-SHM**：共享内存替代 `GetImage` 的 socket 传输，省掉 X server 一次全画布拷贝；
+//! 2. **XDamage**：只取变化矩形 + 只编码脏区，静止时零拷贝。
+
+use agent_core::ScreenCapturer;
+use anyhow::{anyhow, Result};
+use std::time::{Duration, Instant};
+use x11rb::connection::Connection;
+use x11rb::protocol::xproto::{ConnectionExt as _, ImageFormat, VisualClass, Window};
+use x11rb::rust_connection::RustConnection;
+
+/// 帧节流上限，与 agent-core 默认编码档位（30fps）对齐
+const FRAME_INTERVAL: Duration = Duration::from_millis(33);
+
+pub struct X11Capturer {
+    conn: RustConnection,
+    root: Window,
+    width: u16,
+    height: u16,
+    /// 最近一帧 BGRA（既是变更检测基线，也是 copy_frame 的数据源）
+    last: Vec<u8>,
+    last_capture: Option<Instant>,
+}
+
+impl X11Capturer {
+    pub fn new() -> Result<Self> {
+        let (conn, screen_num) = x11rb::connect(None).map_err(|e| {
+            anyhow!(
+                "连接 X11 失败（DISPLAY={:?}）: {e}",
+                std::env::var("DISPLAY").ok()
+            )
+        })?;
+
+        // 先把需要的屏幕参数拷出来：随后 conn 要被移进结构体，不能再持有对它的借用
+        let (root, width, height, depth, visual_class, bpp) = {
+            let setup = conn.setup();
+            let screen = &setup.roots[screen_num];
+
+            // bits_per_pixel 不在 visual 上，而在 Setup 的 pixmap_formats 里（按 depth 索引）
+            let fmt = setup
+                .pixmap_formats
+                .iter()
+                .find(|f| f.depth == screen.root_depth)
+                .ok_or_else(|| anyhow!("找不到 depth={} 对应的像素格式", screen.root_depth))?;
+
+            let visual_class = screen
+                .allowed_depths
+                .iter()
+                .flat_map(|d| d.visuals.iter())
+                .find(|v| v.visual_id == screen.root_visual)
+                .map(|v| v.class)
+                .ok_or_else(|| anyhow!("找不到 root visual"))?;
+
+            (
+                screen.root,
+                screen.width_in_pixels,
+                screen.height_in_pixels,
+                screen.root_depth,
+                visual_class,
+                fmt.bits_per_pixel,
+            )
+        };
+
+        anyhow::ensure!(width > 0 && height > 0, "屏幕尺寸异常 {width}x{height}");
+        anyhow::ensure!(
+            bpp == 32,
+            "当前仅支持 32bpp 真彩采集（depth={depth}，实际 {bpp}bpp）"
+        );
+        // BGRA 字节序假设只在 TrueColor 下成立
+        anyhow::ensure!(
+            visual_class == VisualClass::TRUE_COLOR,
+            "当前仅支持 TrueColor 视觉（实际 {visual_class:?}）"
+        );
+
+        println!("[capture] X11 screen {width}x{height}, depth={depth}, {bpp}bpp");
+
+        Ok(Self {
+            conn,
+            root,
+            width,
+            height,
+            last: Vec::new(),
+            last_capture: None,
+        })
+    }
+
+    /// 抓一帧整屏到 `last`，返回画面是否发生变化
+    fn grab(&mut self) -> Result<bool> {
+        let reply = self
+            .conn
+            .get_image(
+                ImageFormat::Z_PIXMAP,
+                self.root,
+                0,
+                0,
+                self.width,
+                self.height,
+                !0,
+            )
+            .map_err(|e| anyhow!("get_image 请求失败: {e}"))?
+            .reply()
+            .map_err(|e| anyhow!("get_image 应答失败: {e}"))?;
+
+        let expect = self.width as usize * self.height as usize * 4;
+        anyhow::ensure!(
+            reply.data.len() == expect,
+            "帧数据长度异常：期望 {expect} 字节，实际 {}",
+            reply.data.len()
+        );
+
+        let changed = reply.data != self.last;
+        if changed {
+            self.last = reply.data;
+        }
+        Ok(changed)
+    }
+}
+
+impl ScreenCapturer for X11Capturer {
+    fn size(&self) -> (u32, u32) {
+        (self.width as u32, self.height as u32)
+    }
+
+    /// X11 核心协议没有「阻塞至新帧」语义，故这里按帧槽节流后主动抓帧：
+    /// 睡到下一个 1/30s 槽位 → 抓帧 → 与上一帧逐字节比较。
+    /// 因此 `timeout_ms` 在 Linux 实现中被忽略（agent-core 的 min_frame 节流也随之自然满足）。
+    fn acquire(&mut self, _timeout_ms: u32) -> Result<bool> {
+        if let Some(t) = self.last_capture {
+            let elapsed = t.elapsed();
+            if elapsed < FRAME_INTERVAL {
+                std::thread::sleep(FRAME_INTERVAL - elapsed);
+            }
+        }
+        self.last_capture = Some(Instant::now());
+        self.grab()
+    }
+
+    fn copy_frame(&mut self) -> Result<Vec<u8>> {
+        anyhow::ensure!(!self.last.is_empty(), "尚无帧数据可拷贝");
+        Ok(self.last.clone())
+    }
+}

linux-agent/src/input.rs

@@ -0,0 +1,249 @@
+//! X11 输入注入：XTest 扩展（Xorg / XWayland / Xvfb 均支持）。
+//!
+//! **键码映射是本模块的关键**：协议里键盘沿用 Windows 口味的 PS/2 Set-1 扫描码，
+//! 而 X11 要的是 keycode，两套编号并非处处等价：
+//! - 主键区：`keycode = scancode + 8`（evdev 偏移），已逐键核对；
+//! - E0 扩展键：两条链路编号**完全不同**（如方向键 Set-1 是 0x48，evdev 侧 X11 keycode 是 111），必须查表。
+//!
+//! 鼠标沿用虚拟桌面归一化坐标 0..=65535：X11 根窗口天然横跨全部显示器，
+//! 与 Windows 的 `MOUSEEVENTF_VIRTUALDESK` 语义一致，故映射到根窗口像素即可。
+
+use agent_core::InputInjector;
+use anyhow::{anyhow, Result};
+use common::MouseButton;
+use std::sync::Mutex;
+use x11rb::connection::Connection;
+use x11rb::protocol::xproto::{ConnectionExt as _, Window};
+use x11rb::protocol::xtest::ConnectionExt as _;
+use x11rb::rust_connection::RustConnection;
+
+// X11 核心事件码（协议固定值）
+const KEY_PRESS: u8 = 2;
+const KEY_RELEASE: u8 = 3;
+const BUTTON_PRESS: u8 = 4;
+const BUTTON_RELEASE: u8 = 5;
+const MOTION_NOTIFY: u8 = 6;
+
+// X11 按钮编号：1=左 2=中 3=右 4=滚轮上 5=滚轮下
+const BTN_LEFT: u8 = 1;
+const BTN_MIDDLE: u8 = 2;
+const BTN_RIGHT: u8 = 3;
+const BTN_WHEEL_UP: u8 = 4;
+const BTN_WHEEL_DOWN: u8 = 5;
+
+/// Windows 滚轮一格 = 120（协议沿用该语义）
+const WHEEL_DELTA: i32 = 120;
+
+/// 单次滚轮最多合成的格数，防止异常值刷爆事件队列
+const WHEEL_MAX_CLICKS: i32 = 20;
+
+pub struct X11Injector {
+    /// `Mutex` 只为满足 `InputInjector: Sync`（X11 连接本身非 Sync），无竞争
+    conn: Mutex<RustConnection>,
+    root: Window,
+    width: i32,
+    height: i32,
+}
+
+impl X11Injector {
+    pub fn new() -> Result<Self> {
+        let (conn, screen_num) = x11rb::connect(None).map_err(|e| {
+            anyhow!(
+                "连接 X11 失败（DISPLAY={:?}）: {e}",
+                std::env::var("DISPLAY").ok()
+            )
+        })?;
+
+        // XTest 缺失就没有注入能力，启动即失败好过连上后才报错
+        let ext = conn
+            .query_extension(b"XTEST")
+            .map_err(|e| anyhow!("查询 XTEST 扩展失败: {e}"))?
+            .reply()
+            .map_err(|e| anyhow!("查询 XTEST 扩展应答失败: {e}"))?;
+        anyhow::ensure!(ext.present, "X server 未提供 XTEST 扩展，无法注入输入");
+
+        // 先拷出参数，避免借用 conn 的同时又把它移进结构体
+        let (root, width, height) = {
+            let screen = &conn.setup().roots[screen_num];
+            (
+                screen.root,
+                screen.width_in_pixels as i32,
+                screen.height_in_pixels as i32,
+            )
+        };
+        println!("[input] XTEST ready, root {width}x{height}");
+
+        Ok(Self {
+            conn: Mutex::new(conn),
+            root,
+            width,
+            height,
+        })
+    }
+
+    fn fake(&self, type_: u8, detail: u8, x: i32, y: i32) {
+        let conn = match self.conn.lock() {
+            Ok(c) => c,
+            Err(poisoned) => poisoned.into_inner(),
+        };
+        if let Err(e) = conn.xtest_fake_input(
+            type_,
+            detail,
+            0,
+            self.root,
+            x as i16,
+            y as i16,
+            0, // deviceid：0 = 核心设备
+        ) {
+            eprintln!("[input] xtest_fake_input 失败: {e}");
+            return;
+        }
+        let _ = conn.flush();
+    }
+
+    fn click(&self, button: u8) {
+        self.fake(BUTTON_PRESS, button, 0, 0);
+        self.fake(BUTTON_RELEASE, button, 0, 0);
+    }
+}
+
+impl InputInjector for X11Injector {
+    fn mouse_move(&self, x: u16, y: u16) {
+        // detail=0 → 绝对坐标
+        let px = scale(x, self.width);
+        let py = scale(y, self.height);
+        self.fake(MOTION_NOTIFY, 0, px, py);
+    }
+
+    fn mouse_button(&self, b: MouseButton, down: bool) {
+        let button = match b {
+            MouseButton::Left => BTN_LEFT,
+            MouseButton::Middle => BTN_MIDDLE,
+            MouseButton::Right => BTN_RIGHT,
+        };
+        self.fake(
+            if down { BUTTON_PRESS } else { BUTTON_RELEASE },
+            button,
+            0,
+            0,
+        );
+    }
+
+    fn wheel(&self, dy: i32) {
+        // 协议沿用 Windows 语义：dy > 0 向上滚，一格 = 120
+        if dy == 0 {
+            return;
+        }
+        let clicks = (dy.abs() / WHEEL_DELTA).clamp(1, WHEEL_MAX_CLICKS);
+        let button = if dy > 0 { BTN_WHEEL_UP } else { BTN_WHEEL_DOWN };
+        for _ in 0..clicks {
+            self.click(button);
+        }
+    }
+
+    fn key(&self, scancode: u16, extended: bool, down: bool) {
+        let Some(keycode) = keycode_of(scancode, extended) else {
+            eprintln!("[input] 未映射的扫描码 scancode=0x{scancode:02X} extended={extended}");
+            return;
+        };
+        self.fake(
+            if down { KEY_PRESS } else { KEY_RELEASE },
+            keycode,
+            0,
+            0,
+        );
+    }
+}
+
+/// 虚拟桌面归一化坐标（0..=65535）→ 根窗口像素坐标
+fn scale(v: u16, extent: i32) -> i32 {
+    if extent <= 1 {
+        return 0;
+    }
+    ((v as i64 * (extent as i64 - 1)) / 65535) as i32
+}
+
+/// PS/2 Set-1 扫描码 → X11 keycode。
+///
+/// 非扩展键恒为 `scancode + 8`（当前 viewer 的 keymap 只发该区间内的键）；
+/// E0 扩展键查表，编号与主键区规则无关。
+fn keycode_of(scancode: u16, extended: bool) -> Option<u8> {
+    if extended {
+        // 已按 Xorg evdev 标准 keycode 分配核对
+        let kc = match scancode {
+            0x1C => 104, // 小键盘 Enter
+            0x1D => 105, // 右 Ctrl
+            0x35 => 106, // 小键盘 /
+            0x37 => 107, // PrintScreen
+            0x38 => 108, // 右 Alt
+            0x47 => 110, // Home
+            0x48 => 111, // ↑
+            0x49 => 112, // PageUp
+            0x4B => 113, // ←
+            0x4D => 114, // →
+            0x4F => 115, // End
+            0x50 => 116, // ↓
+            0x51 => 117, // PageDown
+            0x52 => 118, // Insert
+            0x53 => 119, // Delete
+            0x5B => 133, // 左 Super(Win)
+            0x5C => 134, // 右 Super(Win)
+            0x5D => 135, // Menu
+            _ => return None,
+        };
+        Some(kc)
+    } else {
+        // 0x01..=0x53 主键区，以及 0x57/0x58 的 F11/F12；其余为 Set-1 未定义位
+        match scancode {
+            0x01..=0x53 | 0x57 | 0x58 => Some((scancode + 8) as u8),
+            _ => None,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn main_block_uses_evdev_offset() {
+        assert_eq!(keycode_of(0x1E, false), Some(38)); // A
+        assert_eq!(keycode_of(0x02, false), Some(10)); // 1
+        assert_eq!(keycode_of(0x39, false), Some(65)); // Space
+        assert_eq!(keycode_of(0x1C, false), Some(36)); // Enter
+        assert_eq!(keycode_of(0x1D, false), Some(37)); // 左 Ctrl
+        assert_eq!(keycode_of(0x2A, false), Some(50)); // 左 Shift
+        assert_eq!(keycode_of(0x58, false), Some(96)); // F12
+    }
+
+    #[test]
+    fn extended_keys_are_not_offset() {
+        // 若误用 +8 规则，方向键会落到小键盘区（0x48+8=80=KP_8），这正是要防的坑
+        assert_eq!(keycode_of(0x48, true), Some(111)); // ↑
+        assert_ne!(keycode_of(0x48, true), Some(0x48 + 8));
+        assert_eq!(keycode_of(0x50, true), Some(116)); // ↓
+        assert_eq!(keycode_of(0x4B, true), Some(113)); // ←
+        assert_eq!(keycode_of(0x4D, true), Some(114)); // →
+        assert_eq!(keycode_of(0x47, true), Some(110)); // Home
+        assert_eq!(keycode_of(0x4F, true), Some(115)); // End
+        assert_eq!(keycode_of(0x49, true), Some(112)); // PageUp
+        assert_eq!(keycode_of(0x51, true), Some(117)); // PageDown
+        assert_eq!(keycode_of(0x53, true), Some(119)); // Delete
+        assert_eq!(keycode_of(0x5B, true), Some(133)); // 左 Super
+    }
+
+    #[test]
+    fn undefined_scancodes_are_rejected() {
+        assert_eq!(keycode_of(0x54, false), None);
+        assert_eq!(keycode_of(0x7F, false), None);
+        assert_eq!(keycode_of(0x99, true), None);
+    }
+
+    #[test]
+    fn coordinate_scaling_covers_full_range() {
+        assert_eq!(scale(0, 1920), 0);
+        assert_eq!(scale(65535, 1920), 1919);
+        assert_eq!(scale(32768, 1920), 960);
+        assert_eq!(scale(12345, 0), 0); // 退化尺寸不 panic
+    }
+}

linux-agent/src/main.rs

@@ -0,0 +1,31 @@
+//! Linux 被控端：X11 采集 + XTest 注入；会话层（QUIC/认证/编码/丢帧）复用 `agent-core`。
+//!
+//! 与 Windows 端的差异仅在这两个模块里，线格式与主控端完全一致，viewer 不需要任何改动。
+//! 依赖 `DISPLAY` 环境变量（Xorg / XWayland 均可）。
+
+mod capture;
+mod input;
+
+use agent_core::{parse_args, serve, EncoderSettings, InputInjector, ScreenCapturer, SessionConfig};
+use anyhow::Result;
+use std::sync::Arc;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    let (port, code) = parse_args()?;
+
+    let display = std::env::var("DISPLAY").unwrap_or_else(|_| "(未设置)".to_string());
+    println!("[linux-agent] DISPLAY={display}");
+
+    // 注入端连接常驻（启动即校验 XTEST 扩展是否可用，缺了就直接报错退出）
+    let injector: Arc<dyn InputInjector> = Arc::new(input::X11Injector::new()?);
+
+    serve(
+        SessionConfig { port, code },
+        EncoderSettings::default(),
+        // 采集端可被重建（X11 连接断开等），故用工厂闭包
+        || Ok(Box::new(capture::X11Capturer::new()?) as Box<dyn ScreenCapturer>),
+        injector,
+    )
+    .await
+}

linux-agent/tests/smoke.rs

@@ -0,0 +1,434 @@
+//! Linux 被控端端到端冒烟测试。
+//!
+//! 需要 X server（Xvfb 即可），通过 `DISPLAY` 指定；推荐用 `scripts/linux-smoke.sh` 一键跑。
+//!
+//! 覆盖链路：X11 采集 → openh264 编码 → QUIC/TLS 下行 → 主控端解帧；
+//! 外加验证码认证、Ping/Pong、关键帧请求，以及**输入注入的实证断言**
+//! （注入鼠标移动后，从另一条 X11 连接查询指针真实落点，而不是只看进程没崩）。
+
+use anyhow::{bail, Context, Result};
+use common::{read_control, read_video_frame, write_control, ControlMsg, InputEvent};
+use rustls::client::danger::{HandshakeSignatureValid, ServerCertVerified, ServerCertVerifier};
+use rustls::pki_types::{CertificateDer, ServerName, UnixTime};
+use rustls::{DigitallySignedStruct, SignatureScheme};
+use std::io::{BufRead, BufReader};
+use std::net::{TcpListener, ToSocketAddrs};
+use std::process::{Child, Command, Stdio};
+use std::sync::{mpsc, Arc};
+use std::time::{Duration, Instant};
+use x11rb::connection::Connection;
+use x11rb::protocol::xproto::{ConnectionExt as _, Window};
+use x11rb::rust_connection::RustConnection;
+
+/// 与 viewer 一致：MVP 阶段跳过自签名证书校验
+#[derive(Debug)]
+struct SkipVerify;
+
+impl ServerCertVerifier for SkipVerify {
+    fn verify_server_cert(
+        &self,
+        _end_entity: &CertificateDer<'_>,
+        _intermediates: &[CertificateDer<'_>],
+        _server_name: &ServerName<'_>,
+        _ocsp_response: &[u8],
+        _now: UnixTime,
+    ) -> Result<ServerCertVerified, rustls::Error> {
+        Ok(ServerCertVerified::assertion())
+    }
+    fn verify_tls12_signature(
+        &self,
+        _message: &[u8],
+        _certificate: &CertificateDer<'_>,
+        _dss: &DigitallySignedStruct,
+    ) -> Result<HandshakeSignatureValid, rustls::Error> {
+        Ok(HandshakeSignatureValid::assertion())
+    }
+    fn verify_tls13_signature(
+        &self,
+        _message: &[u8],
+        _certificate: &CertificateDer<'_>,
+        _dss: &DigitallySignedStruct,
+    ) -> Result<HandshakeSignatureValid, rustls::Error> {
+        Ok(HandshakeSignatureValid::assertion())
+    }
+    fn supported_verify_schemes(&self) -> Vec<SignatureScheme> {
+        vec![
+            SignatureScheme::RSA_PKCS1_SHA256,
+            SignatureScheme::ECDSA_NISTP256_SHA256,
+            SignatureScheme::ED25519,
+            SignatureScheme::RSA_PSS_SHA256,
+        ]
+    }
+}
+
+/// 被测进程的 RAII 句柄：测试无论成功失败都会收尸
+struct AgentProc {
+    child: Child,
+}
+
+impl Drop for AgentProc {
+    fn drop(&mut self) {
+        let _ = self.child.kill();
+        let _ = self.child.wait();
+    }
+}
+
+fn expected_screen() -> (u32, u32) {
+    let s = std::env::var("SMOKE_SCREEN").unwrap_or_else(|_| "1280x720".to_string());
+    let (w, h) = s
+        .split_once('x')
+        .unwrap_or_else(|| panic!("SMOKE_SCREEN 需形如 1280x720，实际 {s:?}"));
+    (w.parse().expect("宽"), h.parse().expect("高"))
+}
+
+fn free_port() -> Result<u16> {
+    let l = TcpListener::bind("127.0.0.1:0")?;
+    Ok(l.local_addr()?.port())
+}
+
+fn spawn_agent(port: u16, code: &str) -> Result<AgentProc> {
+    let mut child = Command::new(env!("CARGO_BIN_EXE_linux-agent"))
+        .args(["--port", &port.to_string(), "--code", code])
+        .stdout(Stdio::piped())
+        .stderr(Stdio::inherit())
+        .spawn()
+        .context("启动 linux-agent 失败")?;
+
+    let stdout = child.stdout.take().expect("stdout 已 piped");
+    let (tx, rx) = mpsc::channel();
+    std::thread::spawn(move || {
+        for line in BufReader::new(stdout).lines().map_while(|l| l.ok()) {
+            println!("[agent] {line}");
+            let _ = tx.send(line);
+        }
+    });
+
+    // 等 QUIC 端口就绪
+    let deadline = Instant::now() + Duration::from_secs(30);
+    loop {
+        let left = deadline.saturating_duration_since(Instant::now());
+        if left.is_zero() {
+            bail!("等待 agent 监听超时（DISPLAY 是否已设置？）");
+        }
+        match rx.recv_timeout(left) {
+            Ok(line) if line.contains("listening") => return Ok(AgentProc { child }),
+            Ok(_) => continue,
+            Err(_) => bail!("等待 agent 监听超时，进程可能已退出"),
+        }
+    }
+}
+
+async fn connect(port: u16) -> Result<(quinn::Endpoint, quinn::Connection)> {
+    let tls = rustls::ClientConfig::builder()
+        .dangerous()
+        .with_custom_certificate_verifier(Arc::new(SkipVerify))
+        .with_no_client_auth();
+    let qc: quinn::crypto::rustls::QuicClientConfig = Arc::new(tls).try_into()?;
+
+    let mut transport = quinn::TransportConfig::default();
+    transport.receive_window(quinn::VarInt::from_u32(16 * 1024 * 1024));
+    transport.stream_receive_window(quinn::VarInt::from_u32(8 * 1024 * 1024));
+
+    let mut client_config = quinn::ClientConfig::new(Arc::new(qc));
+    client_config.transport_config(Arc::new(transport));
+
+    let mut endpoint = quinn::Endpoint::client("0.0.0.0:0".parse()?)?;
+    endpoint.set_default_client_config(client_config);
+
+    let addr = ("127.0.0.1", port)
+        .to_socket_addrs()?
+        .next()
+        .ok_or_else(|| anyhow::anyhow!("地址解析失败"))?;
+    let conn = endpoint.connect(addr, "localhost")?.await?;
+    // endpoint 必须与 conn 同生命周期，故一并返回
+    Ok((endpoint, conn))
+}
+
+fn has_start_code(payload: &[u8]) -> bool {
+    let head = &payload[..payload.len().min(64)];
+    head.windows(4).any(|w| w == [0, 0, 0, 1]) || head.windows(3).any(|w| w == [0, 0, 1])
+}
+
+/// H.264 NAL 类型：5 = IDR 片（关键帧）
+const NAL_IDR: u8 = 5;
+
+/// 按 Annex-B 起始码扫描，判断载荷里是否含指定类型的 NAL
+fn contains_nal(payload: &[u8], nal_type: u8) -> bool {
+    let mut i = 0usize;
+    while i + 3 <= payload.len() {
+        let sc_len = if payload[i..].starts_with(&[0, 0, 0, 1]) {
+            4
+        } else if payload[i..].starts_with(&[0, 0, 1]) {
+            3
+        } else {
+            i += 1;
+            continue;
+        };
+        if i + sc_len >= payload.len() {
+            return false;
+        }
+        if payload[i + sc_len] & 0x1F == nal_type {
+            return true;
+        }
+        i += sc_len + 1;
+    }
+    false
+}
+
+/// 在根窗口左上角画一个**小**矩形，制造一次细微画面变化。
+///
+/// 刻意不用「整屏刷白」那类大变化：大变化会触发编码器的场景切换自动 IDR，
+/// 于是「重发旧帧」和「真重编关键帧」拿到同样的字节，用例就失去区分度了。
+fn perturb_screen(conn: &RustConnection, root: Window) -> Result<()> {
+    use x11rb::protocol::xproto::{CreateGCAux, Rectangle};
+    let gc = conn.generate_id()?;
+    conn.create_gc(
+        gc,
+        root,
+        &CreateGCAux::new().foreground(0x00FF_FFFF).graphics_exposures(0),
+    )?;
+    conn.poly_fill_rectangle(
+        root,
+        gc,
+        &[Rectangle {
+            x: 32,
+            y: 32,
+            width: 64,
+            height: 64,
+        }],
+    )?;
+    conn.free_gc(gc)?;
+    conn.flush()?;
+    println!("[smoke] 已扰动屏幕（根窗口画 64x64 白块）");
+    Ok(())
+}
+
+/// 读帧直到拿到一个非关键帧（跳过静止桌面的低频重发），返回该帧载荷
+async fn read_until_p_frame(recv: &mut quinn::RecvStream, timeout: Duration) -> Result<Vec<u8>> {
+    let deadline = Instant::now() + timeout;
+    while Instant::now() < deadline {
+        let (_h, p) = tokio::time::timeout(Duration::from_secs(5), read_video_frame(recv))
+            .await
+            .context("等待扰动帧超时")??;
+        if !contains_nal(&p, NAL_IDR) {
+            return Ok(p);
+        }
+    }
+    bail!("扰动后未拿到非关键帧，用例前置条件不成立（扰动量是否太大？）")
+}
+
+/// 请求关键帧后，容错地等待一个 IDR：期间允许出现静止桌面的 P 帧重发。
+/// 返回是否在超时内拿到 IDR。
+async fn wait_for_idr(recv: &mut quinn::RecvStream, timeout: Duration) -> Result<bool> {
+    let deadline = Instant::now() + timeout;
+    while Instant::now() < deadline {
+        match tokio::time::timeout(Duration::from_secs(2), read_video_frame(recv)).await {
+            Ok(Ok((_h, p))) => {
+                if contains_nal(&p, NAL_IDR) {
+                    return Ok(true);
+                }
+            }
+            _ => break,
+        }
+    }
+    Ok(false)
+}
+
+/// 归一化坐标 → 根窗口像素（与 linux-agent/src/input.rs 的 scale 完全一致）
+fn scale(v: u16, extent: u32) -> i16 {
+    if extent <= 1 {
+        return 0;
+    }
+    ((v as i64 * (extent as i64 - 1)) / 65535) as i16
+}
+
+fn open_display() -> Result<(RustConnection, Window)> {
+    let (conn, screen_num) = x11rb::connect(None)
+        .map_err(|e| anyhow::anyhow!("连接 X11 失败（DISPLAY 是否已设置？）: {e}"))?;
+    let root = conn.setup().roots[screen_num].root;
+    Ok((conn, root))
+}
+
+fn pointer_position(conn: &RustConnection, root: Window) -> Result<(i16, i16)> {
+    let r = conn.query_pointer(root)?.reply()?;
+    Ok((r.root_x, r.root_y))
+}
+
+fn wait_pointer(
+    conn: &RustConnection,
+    root: Window,
+    want: (i16, i16),
+    timeout: Duration,
+) -> Result<(i16, i16)> {
+    let deadline = Instant::now() + timeout;
+    let mut last = pointer_position(conn, root)?;
+    while Instant::now() < deadline {
+        last = pointer_position(conn, root)?;
+        if (last.0 - want.0).abs() <= 2 && (last.1 - want.1).abs() <= 2 {
+            return Ok(last);
+        }
+        std::thread::sleep(Duration::from_millis(50));
+    }
+    Ok(last)
+}
+
+#[tokio::test]
+async fn end_to_end_capture_encode_transport_input() -> Result<()> {
+    rustls::crypto::ring::default_provider()
+        .install_default()
+        .ok();
+
+    let (sw, sh) = expected_screen();
+    let port = free_port()?;
+    let _agent = spawn_agent(port, "smoke-code")?;
+    let (_endpoint, conn) = connect(port).await?;
+
+    // 流 1：认证，并核对被控端上报的桌面尺寸就是 Xvfb 的尺寸
+    let (mut send, mut recv) = conn.open_bi().await?;
+    write_control(
+        &mut send,
+        &ControlMsg::Auth {
+            code: "smoke-code".into(),
+        },
+    )
+    .await?;
+    match read_control(&mut recv).await? {
+        ControlMsg::AuthOk { width, height } => {
+            assert_eq!(
+                (width, height),
+                (sw, sh),
+                "agent 上报桌面尺寸与实际屏幕不符"
+            );
+        }
+        other => bail!("认证应成功，却收到 {other:?}"),
+    }
+
+    // 流 2：输入 + 控制
+    let (mut isend, mut irecv) = conn.open_bi().await?;
+
+    // Ping/Pong（延迟测量通道）
+    write_control(&mut isend, &ControlMsg::Ping(4242)).await?;
+    match read_control(&mut irecv).await? {
+        ControlMsg::Pong { t, agent_ts } => {
+            assert_eq!(t, 4242, "Pong 回显的时间戳不一致");
+            println!("[smoke] Pong 收到，agent_ts={agent_ts}");
+        }
+        other => bail!("期望 Pong，收到 {other:?}"),
+    }
+
+    // 首帧：证明采集 + 编码 + 线格式整条链路通了
+    let (h1, p1) = tokio::time::timeout(Duration::from_secs(15), read_video_frame(&mut recv))
+        .await
+        .context("等待首个视频帧超时")??;
+    assert_eq!((h1.width, h1.height), (sw, sh), "视频帧头尺寸不符");
+    assert!(!p1.is_empty(), "视频载荷为空");
+    println!(
+        "[smoke] 首帧 {}x{} ts={} 编码后 {} 字节，头部 {:02X?}",
+        h1.width,
+        h1.height,
+        h1.ts_ms,
+        p1.len(),
+        &p1[..p1.len().min(12)]
+    );
+    assert!(
+        has_start_code(&p1),
+        "载荷不像 H.264（前 16 字节 {:02X?}）",
+        &p1[..p1.len().min(16)]
+    );
+
+    // 关键帧恢复：先在**静止**桌面上请求关键帧。
+    // 为了有区分度，先扰动一下屏幕，让缓存里落入一个非关键帧（P 帧），否则
+    // 「重发旧帧」和「真重编关键帧」两种行为拿到的字节是一样的，测不出来。
+    let (probe, root) = open_display()?;
+    perturb_screen(&probe, root)?;
+    let pp = read_until_p_frame(&mut recv, Duration::from_secs(10)).await?;
+    println!(
+        "[smoke] 扰动帧 {} 字节，头部 {:02X?}（应为 P 帧，无 IDR）",
+        pp.len(),
+        &pp[..pp.len().min(12)]
+    );
+    // 前置条件自检：没有这一步，用例可能因为「扰动本身产生了 IDR」而虚假通过
+    assert!(
+        !contains_nal(&pp, NAL_IDR),
+        "扰动帧仍是 IDR，本用例无法区分重发旧帧与真重编关键帧"
+    );
+    tokio::time::sleep(Duration::from_millis(800)).await; // 等画面重新静止
+
+    write_control(&mut isend, &ControlMsg::KeyFrameReq).await?;
+    assert!(
+        wait_for_idr(&mut recv, Duration::from_secs(3)).await?,
+        "静止桌面请求关键帧后 3s 内未收到 IDR：解码端一旦出错将永远无法恢复"
+    );
+    println!("[smoke] 静止桌面请求关键帧 → 收到 IDR");
+
+    // 输入注入实证：注入到屏幕中心，再从独立 X11 连接查询指针真实落点
+    let before = pointer_position(&probe, root)?;
+    write_control(
+        &mut isend,
+        &ControlMsg::Input(InputEvent::MouseMove {
+            x: 32768,
+            y: 32768,
+        }),
+    )
+    .await?;
+    let want = (scale(32768, sw), scale(32768, sh));
+    let got = wait_pointer(&probe, root, want, Duration::from_secs(5))?;
+    println!("[smoke] 指针 {before:?} → {got:?}（期望 ≈{want:?}）");
+    assert!(
+        (got.0 - want.0).abs() <= 2 && (got.1 - want.1).abs() <= 2,
+        "XTest 注入未生效：期望指针 ≈{want:?}，实际 {got:?}"
+    );
+
+    // 键盘：链路不得因此崩溃（无窗口可断言落点）
+    write_control(
+        &mut isend,
+        &ControlMsg::Input(InputEvent::KeyDown {
+            scancode: 0x1E,
+            extended: false,
+        }),
+    )
+    .await?;
+    write_control(
+        &mut isend,
+        &ControlMsg::Input(InputEvent::KeyUp {
+            scancode: 0x1E,
+            extended: false,
+        }),
+    )
+    .await?;
+
+    tokio::time::sleep(Duration::from_millis(300)).await;
+    assert!(
+        conn.close_reason().is_none(),
+        "会话被意外关闭：{:?}",
+        conn.close_reason()
+    );
+
+    println!("[smoke] 全部断言通过");
+    Ok(())
+}
+
+#[tokio::test]
+async fn rejects_wrong_verification_code() -> Result<()> {
+    rustls::crypto::ring::default_provider()
+        .install_default()
+        .ok();
+
+    let port = free_port()?;
+    let _agent = spawn_agent(port, "right-code")?;
+    let (_endpoint, conn) = connect(port).await?;
+
+    let (mut send, mut recv) = conn.open_bi().await?;
+    write_control(
+        &mut send,
+        &ControlMsg::Auth {
+            code: "wrong-code".into(),
+        },
+    )
+    .await?;
+
+    match read_control(&mut recv).await? {
+        ControlMsg::AuthFail(_) => Ok(()),
+        other => bail!("错误验证码竟被接受：{other:?}"),
+    }
+}

scripts/linux-smoke.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# Linux 被控端冒烟测试一键脚本：自建 Xvfb 虚拟显示 → 跑端到端测试 → 收摊。
+#
+# 用法（容器内或任意有 Xvfb 的 Linux）：
+#   bash scripts/linux-smoke.sh
+# 可用环境变量覆盖：
+#   SMOKE_SCREEN=1920x1080   虚拟屏分辨率（默认 1280x720）
+#   SMOKE_DISPLAY=:99        X display 编号（默认 :99）
+set -euo pipefail
+
+SCREEN="${SMOKE_SCREEN:-1280x720}"
+DISPLAY_NUM="${SMOKE_DISPLAY:-:99}"
+
+export DISPLAY="$DISPLAY_NUM"
+export SMOKE_SCREEN="$SCREEN"
+
+XVFB_PID=""
+cleanup() {
+  if [ -n "$XVFB_PID" ]; then
+    kill "$XVFB_PID" 2>/dev/null || true
+    wait "$XVFB_PID" 2>/dev/null || true
+  fi
+}
+trap cleanup EXIT
+
+echo "[smoke] 启动 Xvfb $DISPLAY_NUM (${SCREEN}x24)"
+Xvfb "$DISPLAY_NUM" -screen 0 "${SCREEN}x24" -nolisten tcp >/tmp/xvfb.log 2>&1 &
+XVFB_PID=$!
+
+ready=0
+for _ in $(seq 1 50); do
+  if xdpyinfo -display "$DISPLAY_NUM" >/dev/null 2>&1; then
+    ready=1
+    break
+  fi
+  sleep 0.2
+done
+
+if [ "$ready" -ne 1 ]; then
+  echo "[smoke] Xvfb 启动失败，日志："
+  cat /tmp/xvfb.log 2>/dev/null || true
+  exit 1
+fi
+echo "[smoke] X server 就绪：$(xdpyinfo -display "$DISPLAY_NUM" | grep -E 'dimensions' | head -1)"
+
+# 两个测试共用同一块虚拟屏与指针，必须串行
+cargo test -p linux-agent --test smoke -- --nocapture --test-threads=1
+echo "[smoke] 通过"

构建运行手册.md

@@ -1,7 +1,7 @@
 # MVP 远程控制 Demo · 构建运行手册
 
-> 项目路径：`D:\wordkb\2026-10-22-15-52-47\mvp-remote\`
+> 项目路径：`D:\code\remote\`
-> 形态：Rust workspace，三个 crate：`common`（协议）、`agent`（被控端）、`viewer`（主控端）
+> 形态：Rust workspace：`common`（协议）、`agent-core`（平台无关会话层）、`agent`（Windows 被控端）、`linux-agent`（Linux 被控端）、`viewer`（主控端）
 
 ## 1. 功能范围（量级一 Demo）
 
@@ -10,14 +10,27 @@
 - ✅ 验证码认证（默认 123456）
 - ✅ 键鼠远程控制（SendInput 注入，虚拟桌面绝对坐标 + 扫描码）
 - ✅ 桌面静止时低频重发（保持画面新鲜）
-- ❌ 暂无：硬件编码、P2P 打洞/中继（当前仅局域网/可直连地址）、音频、文件传输、Linux 被控
+- ✅ Linux 被控端（X11 采集 + XTest 注入，与 Windows 端共用同一套协议与会话层）
+- ❌ 暂无：硬件编码、P2P 打洞/中继（当前仅局域网/可直连地址）、音频、文件传输
 
 ## 2. 环境要求
 
 - Windows 8+（被控端需要 DXGI Duplication；主控端 Win7 亦可）
-- 工具链已全部部署在 D 盘（无需再装）：
+- Rust（GNU host：`x86_64-pc-windows-gnu`）+ MinGW-w64，已验证可用组合：
-  - Rust (GNU)：`uurev/tools/rust/cargo/bin`
+  - `rustup default stable-x86_64-pc-windows-gnu`，装到 `%USERPROFILE%\.cargo\bin`
-  - MinGW-w64 14.2：`uurev/tools/mingw64/bin`
+  - MinGW-w64：`D:\mingw64\bin`（gcc 16.2.0）
+  - 构建时 PATH 需同时包含上面两个目录（缺 MinGW 会在链接 openh264 的 C++ 时失败）
+- **cargo 源**：国内直连 crates.io 会超时（cargo 不读系统代理），需配置镜像。已在本机
+  `%USERPROFILE%\.cargo\config.toml` 写入 USTC 源：
+
+  ```toml
+  [source.crates-io]
+  replace-with = "ustc"
+  [source.ustc]
+  registry = "sparse+https://mirrors.ustc.edu.cn/crates.io-index/"
+  ```
+
+> 注：旧版本手册写的 `D:\wordkb\...\uurev\tools\rust` 路径已不存在，按上面重新准备工具链即可。
 
 ## 3. 构建
 
@@ -74,3 +87,72 @@ viewer.exe --addr <被控端IP> --port 47800 --code 888888
 3. QUIC 多流：视频/控制/键鼠分离流（键鼠零阻塞）
 4. P2P 打洞 + coturn 中继（自建信令服务器）
 5. 被控端服务化（Windows Service + 会话切换处理）
+
+## 8. Linux 被控端（linux-agent）
+
+### 8.1 架构：会话层抽到 agent-core
+
+平台差异只留在采集/注入两个模块里，QUIC 握手、认证、编码、丢帧、输入分发全部共用，
+**主控端 viewer 不需要任何改动**：
+
+| crate / 文件 | 职责 | 平台 |
+|---|---|---|
+| `agent-core/src/lib.rs` | QUIC/TLS 握手、验证码认证、H.264 编码下行、latest-wins 丢帧、输入分发 | 共用 |
+| `agent/src/capture.rs` | DXGI Desktop Duplication 采集 | Windows |
+| `agent/src/input.rs` | SendInput 注入 | Windows |
+| `linux-agent/src/capture.rs` | X11 `GetImage` 采集 + 整帧变更检测 | Linux |
+| `linux-agent/src/input.rs` | XTest 注入 + 扫描码→X11 keycode 映射 | Linux |
+
+两端靠 `ScreenCapturer` / `InputInjector` 两个 trait 注入，编码档位（5Mbps / 30fps / 关键帧周期 300）
+与线格式完全一致；协议**未做任何改动**。
+
+### 8.2 构建
+
+Windows 侧照旧（`default-members` 已排除 linux-agent，产物仍是 agent.exe / viewer.exe）：
+
+```bash
+export PATH="$USERPROFILE/.cargo/bin:/d/mingw64/bin:$PATH"   # Git Bash 写法
+cargo build --release
+```
+
+Linux 侧显式指定包（交叉编译不推荐，用容器最省事）：
+
+```bash
+cargo build --release -p linux-agent
+```
+
+### 8.3 运行
+
+```bash
+# 被控 Linux 机器（Xorg 或 XWayland 均可，依赖 DISPLAY 环境变量）
+./linux-agent --port 47800 --code 123456
+# 主控端（Windows）命令行不变
+viewer.exe --addr <被控机IP> --port 47800 --code 123456
+```
+
+### 8.4 冒烟测试（不需要图形桌面）
+
+```bash
+bash scripts/linux-smoke.sh
+```
+
+脚本自建 Xvfb 虚拟屏后跑端到端测试，覆盖：X11 采集 → openh264 编码 → QUIC 下行解帧、
+验证码认证（含**错误验证码被拒**）、Ping/Pong、静止桌面下的关键帧恢复（断言拿到 IDR），
+以及**输入注入实证**——注入鼠标移动后，从另一条 X11 连接读回指针真实坐标。
+
+### 8.5 键码映射（两端最关键的差异）
+
+协议里键盘沿用 Windows 口味的 PS/2 Set-1 扫描码，而 X11 要的是 keycode，两者**并非处处等价**：
+
+- 主键区：满足 `keycode = scancode + 8`；
+- **E0 扩展键编号体系完全不同**（`↑` 是 0x48，但 X11 keycode 是 111 而不是 0x50）——必须查表。
+
+映射表在 `linux-agent/src/input.rs::keycode_of`，带单元测试防回归。
+
+### 8.6 Linux 侧已知限制
+
+1. 仅支持 32bpp TrueColor 视觉（绝大多数桌面满足），其他格式启动即明确报错而非静默出错；
+2. 采集走核心协议 `GetImage`，每帧经 X socket 全量传输；**MIT-SHM 共享内存 + XDamage 脏区**
+   是下一步优化（锚点已写在 `linux-agent/src/capture.rs` 顶部注释）；
+3. 仅覆盖 X11 / XWayland；纯 Wayland（无 XWayland）需另走 PipeWire 门户；
+4. 键位表按 Xorg evdev 标准 keycode 编写，自定义键盘布局可能需要调整。