openclaw - 💡(How to fix) Fix Two-Strike Enforcement for Sub-Agent Error Handling [1 participants]

• Strike: A non-transient error (e.g., network timeout, API failure, uncaught exception) that is not automatically recoverable.

Error Classification

• Are two strikes too few or too many? Should we have escalation (strike 1: warn, strike 2: restart)?

Fix Plan

To implement the Two-Strike Enforcement mechanism, follow these steps:

• Step 1: Configure Two-Strike Settings Update the configuration file with the desired settings:

{ "two_strike": { "max_strikes": 2, "window_seconds": 600, "auto_restart": true } }

* **Step 2: Implement Error Classification**
  Create a function to classify errors as transient or non-transient:
  ```python
def classify_error(error):
  if error.code in [429, 503]:  # transient errors
    return "transient"
  elif error.code in [401, 500, 502, 504]:  # non-transient errors
    return "non_transient"
  else:
    return "unknown"

• Step 3: Update Heartbeat Task Modify the heartbeat task to scan sub-agent sessions and enforce the Two-Strike Rule:

import schedule import time

def heartbeat(): sessions = subagents_list() for session in sessions: error_count = session.metadata["error_count"] recent_errors = session.metadata["recent_errors"] if error_count >= config["two_strike"]["max_strikes"]: # terminate or restart the session subagents_kill(session.id) if config["two_strike"]["auto_restart"]: subagents_restart(session.id)

schedule.every(30).minutes.do(heartbeat) # run every 30 minutes

* **Step 4: Log Strikes and Notify Users (optional)**
  Log each strike to the audit trail and notify users if a critical service is affected:
  ```python
def log_strike(session, error):
  # log the strike to the audit trail
  audit_trail.log(session.id, error)
  # notify users if a critical service is affected
  if session.service == "critical":
    notify_users(session.id, error)

Verification

To verify that the fix worked, monitor the sub-agent sessions and check that:

• Sessions with >=2 non-transient errors within a 10-minute window are terminated or restarted.
• The audit trail logs each strike correctly.
• Users are notified if a critical service is affected.

Extra Tips

• Consider weighting errors differently based on their severity.
• Make the Two-Strike Enforcement mechanism configurable per agent type.
• Continuously monitor and evaluate the effectiveness of the Two-Strike Rule to ensure it is not too lenient or too strict.