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]
Fix(reviewer): Handle context window issues in reviewer
Kilian Liereta88c0871dda2
diff --git a/sweagent/agent/reviewer.py b/sweagent/agent/reviewer.pyindex 3e17f2c4..8a6ed2a1 100644--- a/sweagent/agent/reviewer.py+++ b/sweagent/agent/reviewer.py@@ -236,7 +236,11 @@ class Reviewer(AbstractReviewer):answers = []accepts = []for _ in range(self._config.n_sample):- answer = self._model.query(messages)["message"]+ try:+ answer = self._model.query(messages)["message"]+ except Exception as e:+ self.logger.warning(f"Query failed: {e}", exc_info=True)+ continuetry:score = self.interpret(answer)except ValueError as e:@@ -244,6 +248,9 @@ class Reviewer(AbstractReviewer):continueanswers.append(answer)accepts.append(score)+ if not accepts:+ answers = ["No valid scores found, failing submission"]+ accepts = [-100.0]accept = sum(accepts) / len(accepts) - penaltystd = np.std(accepts).item()if self._config.reduce_by_std > 0: