diff --git a/Cargo.lock b/Cargo.lock
index e6decc6..d950809 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -78,9 +78,9 @@ dependencies = [
 
 [[package]]
 name = "anyhow"
-version = "1.0.102"
+version = "1.0.103"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
+checksum = "2a4385e2e34eb35d6b3efe798b9eb88096925d87726c0798709bf56d9ed84af3"
 
 [[package]]
 name = "asn1-rs"
@@ -599,7 +599,7 @@ dependencies = [
 
 [[package]]
 name = "etr"
-version = "0.6.3"
+version = "0.6.4"
 dependencies = [
  "clap",
  "clap_complete",
@@ -619,6 +619,7 @@ dependencies = [
  "serde",
  "tokio",
  "toml",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
diff --git a/Cargo.toml b/Cargo.toml
index 2fce79c..2dd6332 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "etr"
-version = "0.6.3"
+version = "0.6.4"
 edition = "2024"
 description = "A Rust implementation of Eternal Terminal (et)"
 license = "GPL-3.0-only"
@@ -36,6 +36,14 @@ libc = "0.2"
 serde = { version = "1", features = ["derive"] }
 toml = "1"
 
+# Windows-only: put the console into virtual-terminal mode so the client speaks
+# the same key/ANSI byte conventions a Unix PTY expects (Backspace → DEL, etc.).
+[target.'cfg(windows)'.dependencies]
+windows-sys = { version = "0.61", features = [
+    "Win32_Foundation",
+    "Win32_System_Console",
+] }
+
 [dev-dependencies]
 criterion = { version = "0.8", features = ["async_tokio"] }
 
diff --git a/NOTES.md b/NOTES.md
index 7d19fc7..1789b46 100644
--- a/NOTES.md
+++ b/NOTES.md
@@ -9,7 +9,51 @@ the link drops.  This project uses **QUIC** (via the `quinn` crate) for the tran
 layer, which provides reliable, ordered, multiplexed streams with congestion control
 and TLS 1.3 built-in.
 
-## Current state: v0.6.3 — project logo
+## Current state: v0.6.4 — Windows Backspace fix
+
+New in v0.6.4:
+- **Windows client Backspace fixed.** The Windows console delivers legacy key
+  codes to raw byte reads (Backspace → `0x08`), whereas a Unix PTY expects the
+  xterm convention `0x7f` (DEL) to match the default `stty erase`. The client
+  now enables virtual-terminal console modes after raw mode
+  (`ENABLE_VIRTUAL_TERMINAL_INPUT` on stdin, `ENABLE_VIRTUAL_TERMINAL_PROCESSING`
+  on stdout) via a new Windows-only `windows-sys` dependency, so it emits the
+  xterm byte sequences the remote expects and renders the remote's ANSI output.
+  No-op on Unix. Verified live (Windows `etr` → Unix `etrs`): Backspace now
+  erases correctly.
+- Bumped `anyhow` 1.0.102→1.0.103 to clear RUSTSEC-2026-0190 (an unsoundness
+  advisory against a pre-existing transitive dep). (`crossbeam-epoch` was
+  already bumped to 0.9.20 in v0.6.3 for RUSTSEC-2026-0204.)
+- Test count: 110 (unchanged).
+
+### Known issue — Windows: first line of input not echoed until Enter
+
+When connecting from a Windows `etr` client to a Unix host, the shell prompt
+renders correctly, but the **first line** the user types is not echoed until
+Enter is pressed (after which the whole line appears and runs, and the session
+behaves normally thereafter). This does **not** occur linux→linux and is not
+shell-specific (reproduced with `zsh`+`zellij`+`starship` and with
+`bash --norc`).
+
+Diagnosis: `-vvv` logs show the keystrokes reach the server, but the Windows
+client sends the first line as a single batched chunk (with the trailing Enter)
+rather than per-keystroke, so the remote PTY echoes the whole line only on
+submission. The fault is in the Windows client input path
+(`std::io::stdin().read()`), which does not deliver bytes per-keystroke at
+session start; a large contributor is the shell's startup terminal-probing,
+which — with VT-input mode on — the Windows console auto-answers with a ~7 KB
+burst (256-color palette, size, mode) plus mouse-motion events that
+back-pressure/batch the reader.
+
+Two fixes were attempted and **reverted** (neither is in this release): (1)
+deferring the server-side PTY reader until the first client connection — no
+effect, since the input path (not output) is at fault; (2) reading discrete key
+events via `crossterm` and translating to bytes on the client — eliminated the
+batching but caused the console's terminal-query auto-responses to be echoed
+back as visible garbage (`]4;N;rgb:…`). Tracked in
+[GitHub issue #54](https://github.com/l1a/etr/issues/54).
+
+## Previous: v0.6.3 — project logo
 
 New in v0.6.3:
 - Added `assets/logos/etr-logo.svg` (source) plus rendered `etr-logo-256.png`,
diff --git a/src/bin/etr.rs b/src/bin/etr.rs
index cdc0df8..58d0946 100644
--- a/src/bin/etr.rs
+++ b/src/bin/etr.rs
@@ -28,6 +28,47 @@ static IN_RAW_MODE: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBoo
 /// prevents false triggers from `~` in shell paths or git refs.
 const ESCAPE_CHAR: u8 = b'~';
 
+/// Put the console into virtual-terminal mode.
+///
+/// On Windows the console, by default, hands raw byte reads the legacy key
+/// codes (Backspace → `0x08`, no ESC sequences for arrows/function keys) and
+/// does not interpret ANSI output.  A Unix PTY on the far end expects the
+/// xterm conventions instead — notably Backspace → `0x7f` (DEL) — so without
+/// this the remote `stty erase` (DEL) never matches what we send and Backspace
+/// misbehaves.  Enabling `ENABLE_VIRTUAL_TERMINAL_INPUT` makes the console
+/// translate keys into the same VT byte sequences a real terminal emits, and
+/// `ENABLE_VIRTUAL_TERMINAL_PROCESSING` makes it render the remote's ANSI
+/// output.  Call this after raw mode is enabled, on every (re)connect.
+///
+/// No-op on Unix, where the terminal already speaks VT natively.
+#[cfg(windows)]
+fn enable_vt_console() {
+    use windows_sys::Win32::System::Console::{
+        CONSOLE_MODE, ENABLE_VIRTUAL_TERMINAL_INPUT, ENABLE_VIRTUAL_TERMINAL_PROCESSING,
+        GetConsoleMode, GetStdHandle, STD_INPUT_HANDLE, STD_OUTPUT_HANDLE, SetConsoleMode,
+    };
+    // SAFETY: standard Win32 console calls.  `GetStdHandle` returns a process
+    // std handle; `GetConsoleMode` fails (returns 0) for non-console handles
+    // (e.g. redirected stdio), so we only call `SetConsoleMode` on a handle it
+    // confirmed is a console.  All pointers point to locals that outlive the call.
+    unsafe {
+        let h_in = GetStdHandle(STD_INPUT_HANDLE);
+        let mut mode: CONSOLE_MODE = 0;
+        if GetConsoleMode(h_in, &mut mode) != 0 {
+            SetConsoleMode(h_in, mode | ENABLE_VIRTUAL_TERMINAL_INPUT);
+        }
+        let h_out = GetStdHandle(STD_OUTPUT_HANDLE);
+        let mut out_mode: CONSOLE_MODE = 0;
+        if GetConsoleMode(h_out, &mut out_mode) != 0 {
+            SetConsoleMode(h_out, out_mode | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
+        }
+    }
+}
+
+/// No-op on Unix: the terminal already delivers VT key sequences and renders ANSI.
+#[cfg(not(windows))]
+fn enable_vt_console() {}
+
 macro_rules! vlog {
     ($verbose:expr, $level:expr, $($arg:tt)*) => {
         if $verbose >= $level {
@@ -760,6 +801,10 @@ async fn run_connection_loop(
         vlog!(verbose, 2, "[etr] {}", quic::tls_info());
 
         enable_raw_mode().unwrap();
+        // On Windows, raw mode alone still leaves the console emitting legacy key
+        // codes (Backspace → 0x08) and not rendering ANSI; switch it to VT mode so
+        // we speak the same conventions as the remote Unix PTY.  No-op on Unix.
+        enable_vt_console();
         IN_RAW_MODE.store(true, std::sync::atomic::Ordering::Relaxed);
         in_raw = true;
         let result = tokio::select! {