fix(cli): handle non-tty stdin and EINVAL selector errors gracefully

[Bestbuybaby]

Problem

When hermes is launched via a pipe (e.g. the curl ... | bash installer, or any invocation where stdin is redirected), the CLI crashes with a user-facing traceback:

OSError: [Errno 22] Invalid argument
  File ".../prompt_toolkit/input/vt100.py", line 165, in _attached_input
    loop.add_reader(fd, callback_wrapper)
  File ".../asyncio/selector_events.py", line 271, in _add_reader
    self._selector.register(fd, selectors.EVENT_READ, ...)

Root cause

In this scenario fd 0 is a valid file descriptor (so the existing os.fstat(0) pre-flight guard passes) but is not a tty. prompt_toolkit then asks asyncio's kqueue selector to watch fd 0, which fails on macOS with OSError: [Errno 22] Invalid argument (EINVAL).

The existing fallback exception handler in HermesCLI.run() only matched error strings containing "is not registered" or "Bad file descriptor", so EINVAL slipped through and re-raised — producing the traceback after the "Goodbye! ⚕" exit message.

Fix

1. Pre-flight guard now also requires os.isatty(0), catching piped-stdin cases up front with a helpful message.
2. Fallback handler now also matches "Invalid argument" and errnos 9 (EBADF) and 22 (EINVAL).
3. Error messages now mention the curl | bash case and suggest opening a fresh terminal.

Verification

• New regression tests in tests/cli/test_cli_stdin_guard.py (7 tests, all pass):
  • Pre-flight tty guard exits cleanly with the helpful message
  • Pre-flight fstat failure exits cleanly
  • Selector matcher catches EINVAL / EBADF / "is not registered" KeyError
  • Selector matcher ignores unrelated OSError/KeyError
• Full tests/cli/ suite: 479 pass (2 pre-existing, unrelated failures in test_reasoning_command.py re: AIAgent._stream_callback — confirmed to fail on origin/main without this change)
• Live integration repro (hermes < /dev/null): now exits with code 0, prints the helpful message, no traceback

Co-authors

Co-Authored-By: Oz oz-agent@warp.dev

---

Generated with Warp: https://app.warp.dev/conversation/0127fff1-2474-43ae-ac87-6866375c2cd1

[Copilot]

Pull request overview

This PR prevents hermes interactive chat from crashing when stdin is redirected or non-TTY (e.g. curl ... | bash), and broadens the fallback handling for asyncio selector registration failures (including macOS EINVAL).

Changes:

• Add a pre-flight stdin validation that requires fd 0 to be a TTY (and improves the user-facing guidance message).
• Expand the selector/stdio error fallback matcher to include EINVAL (errno=22) in addition to existing broken-stdin symptoms.
• Add regression tests covering the non-TTY guard and selector error matching behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
cli.py	Adds non-TTY stdin guard and extends fallback handling for selector registration errors.
tests/cli/test_cli_stdin_guard.py	Introduces regression tests for the stdin guard and selector error matching.

Comments suppressed due to low confidence (1)

cli.py:10120

• The non‑TTY stdin guard returns early after starting spinner/process daemon threads (started above). Because _should_exit is never set to True on this early-return path, those background loops keep running for the remainder of the process (notably under pytest), repeatedly polling queues/config and importing modules. Set self._should_exit = True before returning (and/or move the stdin validation earlier so threads aren’t started when stdin is unusable).

            _run_cleanup()
            self._print_exit_summary()
            return

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

TestSelectorErrorMatcher._matcher() duplicates the production matching logic instead of exercising the real code path, so the tests can pass even if cli.py’s matcher regresses. Consider extracting the matcher into a small helper (e.g., _is_stdin_selector_error(exc)) and importing it here, or drive HermesCLI.run() through a mocked app.run() that raises the target OSError/KeyError and assert on the printed message.

[Copilot]

The fallback matcher treats any OSError whose message contains "Invalid argument" as a broken-stdin symptom. That string is generic and could hide unrelated EINVAL failures coming from inside app.run(). Prefer matching on errno == 22 (and keep the KeyError substring match for "is not registered") rather than the message text, which is both broad and locale-dependent.

Suggested change

	# - "Bad file descriptor" (EBADF / errno 9)
	# - "Invalid argument" (EINVAL / errno 22, raised by kqueue on
	# macOS when fd 0 is not a tty/pipe the kernel will watch)
	_err_str = str(_stdin_err)
	_errno = getattr(_stdin_err, "errno", None)
	if (
	"is not registered" in _err_str
	or "Bad file descriptor" in _err_str
	or "Invalid argument" in _err_str
	or _errno in (9, 22)
	# - errno 9 (EBADF / bad file descriptor)
	# - errno 22 (EINVAL, raised by kqueue on macOS when fd 0 is not
	# a tty/pipe the kernel will watch)
	_err_str = str(_stdin_err)
	_errno = getattr(_stdin_err, "errno", None)
	if (
	(isinstance(_stdin_err, KeyError) and "is not registered" in _err_str)
	or (isinstance(_stdin_err, OSError) and _errno in (9, 22))

[alt-glitch]

Multiple competing PRs for the same stdin EINVAL crash: #13251, #11253, #8796. Recommend maintainer pick one.

[teknium1]

Closing — superseded by #26077 (merged as commit d3d5916), which preventively probes kqueue at startup and falls back to SelectSelector when fd 0 cannot be registered. The widened except-clause matching EINVAL / EBADF / 'Invalid argument' — which most PRs in this cluster including yours added — is also included.

Thanks for the fix; closing as duplicate of the merged work.