[Bug] ScenarioResult.messages returns judge's internal messages instead of full conversation in 0.7.15

[rajdeepmahal24]

Description

In langwatch-scenario==0.7.15, ScenarioResult.messages returns the judge's internal messages (system prompt + transcript text) instead of the full conversation history. This is a breaking change from 0.7.14.

Root Cause

In v0.7.15, the JudgeAgent.call() method was refactored to use a transcript-based approach for the judge's context:

v0.7.14 built messages by spreading the actual conversation:

messages = [
    {"role": "system", "content": self.system_prompt or ...},
    *input.messages,  # Actual conversation messages spread here
]

v0.7.15 converts messages to a text transcript instead:

transcript = JudgeUtils.build_transcript_from_messages(input.messages)
content_for_judge = f"""
<transcript>
{transcript}
</transcript>
...
"""

messages = [
    {"role": "system", "content": self.system_prompt or ...},
    {"role": "user", "content": content_for_judge},  # Just text, not the actual messages
]

Both versions then return messages=messages in the ScenarioResult, but in 0.7.15 this no longer contains the actual message objects.

Expected Behavior

ScenarioResult.messages should contain the full conversation history between the user simulator and the agent under test, including:

• User messages
• Assistant messages
• Tool calls
• Tool results

Actual Behavior

ScenarioResult.messages contains only:

• Judge's system prompt
• A single user message containing the transcript as text

The actual structured message objects (with tool calls, etc.) are lost.

Impact

This breaks any code that relies on result.messages to:

• Extract tool calls from the conversation
• Extract assistant responses
• Log the full conversation for debugging/reporting
• Generate HTML reports showing the conversation
• Iterate over individual messages

Workaround

Downgrade to langwatch-scenario==0.7.14 where result.messages contains the actual conversation messages.

Suggested Fix

Change the JudgeAgent to return input.messages instead of the local messages variable:

return ScenarioResult(
    success=verdict == "success" and len(failed_criteria) == 0,
    messages=input.messages,  # Use the actual conversation, not judge's internal messages
    reasoning=reasoning,
    passed_criteria=passed_criteria,
    failed_criteria=failed_criteria,
)

Environment

• Python: 3.13.4
• langwatch-scenario: 0.7.15
• OS: macOS