Alt+] wedges keyboard input in legacy terminals: unterminated OSC in pinned crossterm

Environment

Codex CLI 0.154.0 on Linux, GNOME Terminal 3.56.2 / VTE 0.80.1, X11. Codex pins openai-oss-forks/crossterm at 45fecb9508105988f42fe6ff0441783ed3717f92.

Reproduction

1. Open Codex in a terminal that sends ESC ] (1b 5d) for Alt+].
2. Press Alt+], then Ctrl+C, Esc, or Ctrl+D.
3. Subsequent keyboard input no longer reaches the application.

Initially observed after Alt+Up opened an asynchronous question, but also reproduced in the ordinary empty composer. CODEX_TUI_DISABLE_KEYBOARD_ENHANCEMENT=1 does not prevent it.

Expected: the key is handled or ignored and subsequent input remains available.

Actual: subsequent key bytes remain buffered inside an unterminated OSC sequence.

Evidence

A real PTY driving the exact pinned Crossterm EventStream reproduces:

1b 5d       # Alt+]
03          # Ctrl+C after 100 ms: no KeyEvent
1b 04 03    # Esc, Ctrl+D, Ctrl+C: still no KeyEvents
07 03       # BEL then Ctrl+C: input recovers

A Tokio timer continues running during the stall; this is input parsing/buffering failure, not evidence that all runtime threads are deadlocked.

Both fork Unix backends (mio/use-dev-tty) fail. The installed CLI was independently reproduced using isolated PTYs and a localhost Responses fixture that emits an async question. The test verifies the expanded answer UI before injecting the trigger, and checks enhancement mode requests in the default/disabled comparison. No real model calls are needed.

System VTE 0.80.1, in a minimal GTK host driven by XTest on an isolated Xvfb display, emits 1b 5d both before and after Codex's enhancement request. This is a terminal-library test, not a capture of the original session's physical keypress.

Root cause and proposed fix

The fork adds b']' => parse_osc(buffer). parse_osc returns Ok(None) until BEL/ST arrives. The reader keeps appending input without an OSC deadline. The existing lone-ESC timeout does not cover this state.

Consequently, no Alt+] KeyEvent reaches the question UI; ignoring it in async_questions/input.rs cannot fix the problem.

A dependency-layer fix should bound unfinished OSC parsing, preserve Ctrl+C even if it arrives before timeout, and retain legitimate fragmented OSC/paste behavior in both Unix backends. A tested implementation and real PTY regression are available in https://github.com/openai-oss-forks/crossterm/pull/5, targeting the dependency branch that Codex pins. Codex will also need a dependency/lockfile update after that fix becomes available upstream.

A separate incomplete-CSI issue also reproduces in original Crossterm; it is outside the OSC fix.

Patched Codex integration validation

Built Codex commit 02a8f038b87ad34d4a1dc5058eda26972ed7aa6c with dependency commit 995d98a using a test-only local Cargo path override. The installed CLI and daily configuration were not changed.

• All 12 real CLI PTY cases passed: ordinary composer and expanded asynchronous-question UI, keyboard enhancement default/disabled, and control/legacy Alt+]/CSI-u input. Ctrl+C exited normally without a BEL recovery byte. Question cases verified the expanded UI before injecting input; fixtures used localhost only.
• The normal TUI suite initially ran 4,613 tests: 4,603 passed, 9 failed, and 1 timed out. The 10 unsuccessful cases all passed in a focused 14-test rerun with umask 077 and NO_COLOR unset. The original failures came from group-writable test socket directories and disabled color output; the IPC timeout waited on a server thread after its client connection was rejected. No source or snapshot changes were needed. Previously passing tests were not rerun.
• The 6 default-ignored entries were separately attempted: one tmux resize smoke passed; three stopped at the GPT-5.4 model-migration prompt before reaching resize assertions. The Windows-only AltGr test is inapplicable on Linux and fails if forced. The ignored startup PTY child returned successfully but performs no standalone validation without its parent-provided environment; its real parent test passed in the normal suite. These results do not establish full manual resize coverage.

Read the pinned Crossterm parser in src/event/sys/unix/parse.rs and the Unix mio reader in src/event/source/unix/mio.rs, then review the related Codex handling in async_questions/input.rs. Use the PTY regression cases described in the issue as the starting test; done means the dependency and lockfile use the available fix and legacy Alt+] no longer swallows subsequent control input.