What

Ports the claude-loop PTY proxy (#269) to Windows. The Unix proxy is POSIX-only (pty/termios/AF_UNIX); Windows needs ConPTY + a named pipe. This is strategy B — a standalone nested-ConPTY proxy that ships the feature now; strategy A (psmux emits the signal natively) is the planned follow-up (see the design doc).

Same feature as the Unix side: tell a human typing apart from claude's output and from the loop's own wake injection — live, busy included.

New crate: windows/cl-pty-proxy/ (Rust)

• GNU toolchain (x86_64-pc-windows-gnu) — no MSVC / VS Build Tools needed.
• Nested ConPTY for claude via portable-pty (same layer psmux is built on); bridges three channels: human stdin → claude, claude output → stdout, named-pipe injection → claude (marker untouched).
• win32-input-mode parsing — the key Windows gotcha: under psmux/ConPTY keystrokes arrive as ESC[Vk;Sc;Uc;Kd;Cs;Rc_, not raw VT, so every keystroke starts with ESC. A naive "first byte printable" test never fires. The parser extracts the unicode + key-down to detect text keystrokes. 7 unit tests on real captured sequences.
• Markers (human-typing, PID-stamped proxy-alive), @cl_human painting (#274 parity), resize + exit-code propagation, fail-safe direct-run if ConPTY init fails (pane never bricked).

Wiring

• cli.ts: on win32, launch cl-pty-proxy.exe -- claude … when the binary exists; the Python proxy branch is gated to non-Windows (its POSIX APIs would crash the pane). claude-loop check reports proxy active/inactive.
• state.ts: injectWakePhrase → named pipe \.\pipe\cl-inject-<name> via Node net (named pipes are first-class, no native dep), gated on proxyIsAlive (a pipe can't be stat-ed). New injectPipeName() helper.

Validation (under real psmux)

• ✅ ConPTY bridging (cmd renders, DSR cursor handshake completes)
• ✅ human typing → human-typing marker + @cl_human=stop, reverts to loop after TTL
• ✅ named-pipe injection reaches claude without touching the marker (channel separation)
• ✅ graceful cleanup; kill -9 leaves a stale marker handled by the PID-liveness check (#278 contract)
• ✅ end-to-end via claude-loop start
• ✅ cargo test (7), cargo clippy, TS typecheck + eslint all clean

Build (not committed — platform-specific artifact, target/ gitignored)

cargo build --release --manifest-path windows/cl-pty-proxy/Cargo.toml

Without it, claude-loop still works — falls back to idle-only pane-diff detection.

Known limitation → motivates strategy A

claude runs under a second ConPTY nested inside psmux's ConPTY (double conhost). Basic rendering / resize work, but complex TUI cases can glitch. This is inherent to any nested-proxy approach. Strategy A (psmux touches the marker natively in forward_key_to_active — no nested PTY, no double translation, no native dep) is documented in docs/PTY-PROXY-WINDOWS.md as the follow-up PR; strategy B validates the markers/named-pipe/status wiring that A would reuse.

Docs

• docs/PTY-PROXY-WINDOWS.md — full design + strategy A proposal.
• docs/WIN-INSTALL.md — optional build step.