aliasrobotics
Enhance error handling for OutputGuardrailTripwireTriggered exception…
When CAI generates reports at the end of an assessment with CAI_STREAM=true, instead of displaying the report cleanly, the terminal output shows corrupted behavior:
- The entire report reprints multiple times
- Each reprint adds one new line to the previous output
- Results in exponentially growing output with duplicate content
- Terminal fills with repeated copies of the same report content
My assumption is that this is caused by output guardraisl detecting dangerous content (like exploit code or sqli-payloads) and raises an OutputGuardrailTripwireTriggered exception. But this exception is not handled properly by the streaming-system. Therefore the live display context remains active and is not cleaned up properly
I added this piece of code in cli.py at line 1511 as a workaround / solution
except OutputGuardrailTripwireTriggered as e:
# Handle guardrail exception specifically - MUST come before broad Exception handler
# Clean up streaming display before showing error
try:
from cai.util import cleanup_all_streaming_resources
cleanup_all_streaming_resources()
except Exception:
pass
# Clean up the async generator
if stream_iterator is not None:
try:
await stream_iterator.aclose()
except Exception:
pass
# Clean up the result object if it has cleanup methods
if result is not None and hasattr(result, '_cleanup_tasks'):
try:
result._cleanup_tasks()
except Exception:
pass
# Re-raise to be caught by outer handler which shows user-friendly message
raise
this stops the stream to break but leads to the output of the exception-message found in cli.py at around line 1598
except OutputGuardrailTripwireTriggered as e:
# Display a user-friendly warning instead of crashing (streaming mode)
guardrail_name = e.guardrail_result.guardrail.get_name()
reason = e.guardrail_result.output.output_info.get("reason", "Security policy violation")
# Use red color for the warning message
print(f"\n\033[91m🛡️ SECURITY GUARDRAIL TRIGGERED\033[0m")
print(f"\033[91mGuardrail: {guardrail_name}\033[0m")
print(f"\033[91mReason: {reason}\033[0m")
print(f"\033[93mThe agent's output was blocked for security reasons.\033[0m")
print(f"\033[96mYou can continue the conversation with a different request.\033[0m\n")
# Continue the conversation loop instead of crashing
continue
indicating that the guardrail is indeed triggered by the stream (but not blocking the stream)
