fix: allow both append_messages and generate_tool_call_response in tool runner loop

[AliSharjeell]

Summary

When users call both append_messages() AND generate_tool_call_response() in the same loop iteration (e.g., to log the tool result), the tool result was never added to history because _messages_modified = True bypassed the auto-append. This caused infinite loops as the model re-called the same tool with no history of the result.

Root Cause

The original code:

if not self._messages_modified:
    self.append_messages(message, response)

When append_messages() was called first, it set _messages_modified = True, so the auto-append was skipped. But since generate_tool_call_response() was also called (and cached), the tool result was never added.

Fix

Changed the logic to:

if self._messages_modified:
    self._cached_tool_call_response = None
    self._messages_modified = False

self.append_messages(message, response)

Now the auto-append always happens, clearing the cache and reset flag if needed. Users can still access the tool result via generate_tool_call_response() before the loop continues.

Testing

The fix was verified against the minimal repro from the issue:

for message in runner:
    tool_response = runner.generate_tool_call_response()
    if tool_response:
        print(f"Tool result: {tool_response}")

    runner.append_messages({"role": "user", "content": "be concise in your response"})

Now terminates correctly instead of infinite looping.

Related Issue

Closes #1536 — append_messages() inside tool runner loop causes infinite loop