fix(cli): degrade _cprint to plain output when stdout has no console (fixes #65558)

[hansai-art]

What & why

On Windows, hermes chat -q "<prompt>" with piped/redirected stdout — the documented non-interactive path, and exactly what any orchestrator does to capture the response (stdio: pipe) — crashed with prompt_toolkit.output.win32.NoConsoleScreenBufferError. This made the -q flag unusable from any Windows automation/CI context.

Fixes #65558.

Root cause (confirmed on Windows): when no prompt_toolkit Application is running, _cprint calls _pt_print, which builds a fresh pt output backend via create_output(). On Windows that's Win32Output, which calls GetConsoleScreenBufferInfo on the process's stdout handle — a pipe, not a console — and raises. I reproduced it directly: _pt_print(ANSI(...)) with a non-tty stdout → NoConsoleScreenBufferError. (The POSIX branch of create_output() has an isatty() fallback to PlainTextOutput; the win32 branch has none.)

Fix

Suggested option 1 from the issue (in-tree, no upstream dependency): detect a non-console stdout up front and emit plain text instead of the pt renderer — but only in the no-app path, where a fresh backend would otherwise be constructed. This mirrors the isatty() gate prompt_toolkit uses on POSIX (and that this file's own _b/_d helpers already use). When an Application is running, its output backend already exists and works, so those paths (bg-thread run_in_terminal scheduling, etc.) are left completely unchanged.

The plain path also resolves the two related symptoms the reporter flagged, same "assume a real console" root cause:

• ANSI escapes leaking into captured output → stripped via a CSI regex.
• UnicodeEncodeError on box-drawing / braille characters under a legacy Windows code page (cp1252) → re-encode through stdout's encoding with replacement, so they degrade to ? instead of raising (previously silently dropped by a bare except).

New helpers: _stdout_is_console(), _plain_output(), and _emit_without_running_app() (shared by the no-app and the pt-import-failure branches). The old inline try: _pt_print except: print(text) fallback — which leaked raw ANSI and could drop unicode — is replaced by this.

Scope / compatibility

• Only the no-app and pt-import-failure branches of _cprint change. The app-running branches are byte-for-byte unchanged.
• When stdout is a console, behavior is identical to before (pt renderer, colors intact).

Testing

tests/cli/test_cli_cprint_no_console.py (8 new): console detection (incl. isatty() raising), ANSI + cursor/erase stripping, the cp1252 encoding-safe fallback (via a fake cp1252 stream), the end-to-end no-console _cprint regression (pytest's capsys stdout is already non-tty — the exact broken path — so it asserts no crash + ANSI stripped), and that a real console still routes through prompt_toolkit.

Updated 3 existing tests in test_cprint_bg_thread.py that assert the pt-routing path, pinning _stdout_is_console -> True (the console presence they implicitly assumed); the no-console behavior is covered by the new file.

• python -m pytest tests/cli/test_cli_cprint_no_console.py tests/cli/test_cprint_bg_thread.py -q → 21 passed.
• python -m pytest tests/cli -k "cprint or output or print or terminal or replay or history" -q → 161 passed.
• Authored and reproduced on Windows.

🤖 Generated with Claude Code

https://claude.ai/code/session_013b1XyXitAxV7phGmKWigJX

[tonydwb]

Code Review Summary

Verdict: Approved

Degrades _cprint to plain output when stdout has no console. Clean fix for SIGWINCH issue. No security concerns.

---

Reviewed by Hermes Agent