Description
When the user sends an interrupt message while Hermes CLI is executing a long-running task, the process crashes with OSError: [Errno 5] Input/output error during prompt_toolkit shutdown.

Reproduction

1. Start a long-running task in Hermes CLI
2. Send an interrupt message (press Enter with busy_input_mode=interrupt)
3. Observe: process crashes and restarts

Error Log

OSError: [Errno 5] Input/output error

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File ".../prompt_toolkit/renderer.py", line 429, in reset
    self.output.flush()
  File ".../prompt_toolkit/output/vt100.py", line 706, in flush
    flush_stdout(self.stdout, data)
  File ".../prompt_toolkit/output/flush_stdout.py", line 37, in flush_stdout
    stdout.flush()
OSError: [Errno 5] Input/output error

Root Cause Analysis

Two issues compound:

1. Signal handler raises KeyboardInterrupt at arbitrary C stack point (cli.py:9682): The _signal_handler raises KeyboardInterrupt inside a Unix signal handler. This causes prompt_toolkit to attempt emergency shutdown, which tries to flush stdout that is in a broken/devnull state.

2. OSError with errno 5 not suppressed (cli.py:9699-9706): The _suppress_closed_loop_errors exception handler only suppresses:

   • RuntimeError: "Event loop is closed"
   • KeyError: "is not registered"

   But OSError: [Errno 5] Input/output error from stdout flush falls through to the default handler, which then crashes the process.

Additional Context from interrupt_debug.log

05:31:21 interrupt fired: msg=..., agent_running=True
05:45:43 ... OSError: [Errno 5] Input/output error (crash + restart)
05:55:42 interrupt fired: msg=..., agent_running=True
06:00:09 ... OSError: [Errno 5] Input/output error (crash + restart)

Environment

• macOS Darwin 25.4.0 (Apple Silicon M3)
• Python 3.9.6 via uv
• Hermes Agent latest (NousResearch/hermes-agent)
• busy_input_mode: interrupt

Expected Behavior
Hermes CLI should gracefully handle interrupts without crashing.

Suggested Fix

1. Add OSError with errno 5 (Input/output error) to _suppress_closed_loop_errors
2. Alternatively, change signal handler to use a flag-based approach instead of raising an exception