SWE-agent takes a GitHub issue and tries to automatically fix it, using your LM of choice. It can also be employed for offensive cybersecurity or competitive coding challenges. [NeurIPS 2024]
Resample: Better handling of format errors
Kilian Lieretf8e64ce1f306
diff --git a/sweagent/agent/best_response_picker.py b/sweagent/agent/best_response_picker.pyindex 7b269129..472beb9d 100644--- a/sweagent/agent/best_response_picker.py+++ b/sweagent/agent/best_response_picker.py@@ -173,7 +173,17 @@ class BinaryTrajectoryComparison(AbstractBestActionPicker):history: list[dict[str, Any]],completions: list[dict[str, Any]],) -> GetActionOutput:- parsed_completions = [self._tools.parse_actions(a) for a in completions]+ parsed_completions = []+ for completion in completions:+ try:+ thought, action = self._tools.parse_actions(completion)+ except FormatError:+ logger.warning("Could not parse completion %s, skipping.", completion)+ continue+ parsed_completions.append((thought, action))+ if len(parsed_completions) == 0:+ msg = "No completions could be parsed."+ raise FormatError(msg)thoughts: list[str] = [h[0] for h in parsed_completions]actions: list[str] = [h[1] for h in parsed_completions]assert len(thoughts) == len(actions)