Every agent response via Claude models (MyClaw provider) triggers an Unexpected event order, got message_start before receiving "message_stop" error that is displayed as a visible error message in the Telegram chat. This makes the chat essentially unusable — every reply from the agent is followed by a red error notification. {"subsystem":"agent/embedded", "event":"embedded_run_agent_end", "tags":["error_handling","lifecycle","agent_end","assistant_error"], "isError":true, "error":"Unexpected event order, got message_start before receiving "message_stop"", "model":"claude-opus-4.6", "provider":"myclaw"} {"subsystem":"agent/embedded", "event":"embedded_run_agent_end", "isError":true, "error":"Unexpected event order, got message_start before receiving "message_stop"", "model":"claude-sonnet-4.6", "provider":"myclaw"} This error repeats for every single response from the agent, regardless of model. 3. Agent responds correctly BUT an error message appears in the Telegram chat: Unexpected event order, got message_start before receiving "message_stop" 4. Repeat for every message — error appears every time No error message should appear in the Telegram chat. Even if the streaming event order is unexpected, the error should be handled gracefully and silently (or logged internally only). An error notification is sent to the Telegram chat after every agent response, making the chat unusable.

• The error likely originates from MyClaw sending a malformed SSE stream (sending message_start before the previous stream is closed with message_stop)
• OpenClaw should handle this case gracefully without exposing the error to the end user in the chat

Fix Plan

To resolve the Unexpected event order, got message_start before receiving "message_stop" error, we need to modify the OpenClaw code to handle the malformed SSE stream from MyClaw gracefully. Here are the steps:

• Modify the embedded_run_agent_end function to catch and log the error internally instead of sending it to the Telegram chat.
• Implement a retry mechanism to handle the case where message_start is received before message_stop.
• Update the error handling to silently ignore the error and continue processing the agent response.

Example code changes:

// Catch and log the error internally
try {
    // Existing code to handle agent response
} catch (error) {
    if (error.message === 'Unexpected event order, got message_start before receiving "message_stop"') {
        console.error('Error handling agent response:', error);
    } else {
        throw error;
    }
}

// Implement a retry mechanism
const maxRetries = 3;
let retries = 0;
while (retries < maxRetries) {
    try {
        // Existing code to handle agent response
        break;
    } catch (error) {
        if (error.message === 'Unexpected event order, got message_start before receiving "message_stop"') {
            retries++;
            console.error('Retry ', retries, ' failed. Retrying...');
        } else {
            throw error;
        }
    }
}

// Update error handling to silently ignore the error
if (error.message === 'Unexpected event order, got message_start before receiving "message_stop"') {
    console.error('Error handling agent response:', error);
    // Continue processing the agent response
} else {
    throw error;
}

Verification

To verify that the fix worked, follow these steps:

1. Apply the code changes and restart the OpenClaw service.
2. Test the agent response via Telegram.
3. Verify that the error message no longer appears in the Telegram chat.
4. Check the logs to ensure that the error is being caught and logged internally.

Extra Tips

• Monitor the logs to ensure that the error is not occurring frequently and adjust the retry mechanism as needed.
• Consider implementing a more robust error handling mechanism to handle other potential errors that may occur.
• Test the fix with non-Claude models to determine if the issue is specific to Claude models or a general MyClaw gateway issue.