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]
Ref: Renaming best response picker -> action sampler
Kilian Lieret2b3774f0de3f
diff --git a/sweagent/agent/best_response_picker.py b/sweagent/agent/action_sampler.pysimilarity index 95%rename from sweagent/agent/best_response_picker.pyrename to sweagent/agent/action_sampler.pyindex e47d41ab..91f12ecc 100644--- a/sweagent/agent/best_response_picker.py+++ b/sweagent/agent/action_sampler.py@@ -15,14 +15,14 @@ from sweagent.utils.log import get_loggerlogger = get_logger(__name__)-class GetActionOutput(BaseModel):+class ActionSamplerOutput(BaseModel):completion: dict[str, Any]messages: list[dict[str, Any]] = []trajectory_items: list[dict[str, Any]] = []extra_info: dict[str, Any] = {}-class AbstractBestActionPicker(BaseModel):+class AbstractActionSampler(BaseModel):def setup(self, model: AbstractModel, tools: ToolHandler):self._model = modelself._tools = tools@@ -34,12 +34,12 @@ class AbstractBestActionPicker(BaseModel):trajectory: Trajectory,history: list[dict[str, Any]],completions: list[dict[str, Any]],- ) -> GetActionOutput:+ ) -> ActionSamplerOutput:"""Returns action with tool calls"""pass-class AskColleagues(AbstractBestActionPicker):+class AskColleagues(AbstractActionSampler):type: Literal["ask_colleagues"] = "ask_colleagues"def get_colleague_discussion(self, completions: list[dict[str, Any]]) -> str:@@ -70,7 +70,7 @@ class AskColleagues(AbstractBestActionPicker):trajectory: Trajectory,history: list[dict[str, Any]],completions: list[dict[str, Any]],- ) -> GetActionOutput:+ ) -> ActionSamplerOutput:"""Returns action with tool calls"""discussion = self.get_colleague_discussion(completions)logger.info(f"COLLEAGUE DISCUSSION:\n{discussion}")@@ -78,13 +78,13 @@ class AskColleagues(AbstractBestActionPicker):{"role": "user", "content": discussion},]final_completion = self._model.query(history + new_messages) # type: ignore- return GetActionOutput(+ return ActionSamplerOutput(completion=final_completion,extra_info={"colleagues": discussion},)-class BinaryTrajectoryComparison(AbstractBestActionPicker):+class BinaryTrajectoryComparison(AbstractActionSampler):type: Literal["binary_trajectory_comparison"] = "binary_trajectory_comparison"system_template: str = """<setting>You are an expert software engineer overseeing junior developers. They suggest actions to take to solve a problem. You must choose the best action to take. </setting>"""@@ -221,7 +221,7 @@ class BinaryTrajectoryComparison(AbstractBestActionPicker):trajectory: Trajectory,history: list[dict[str, Any]],completions: list[dict[str, Any]],- ) -> GetActionOutput:+ ) -> ActionSamplerOutput:parsed_completions = self.parse_completions(completions)parsed_completions = self.filter_duplicates(parsed_completions)if len(parsed_completions) == 1:@@ -251,7 +251,7 @@ class BinaryTrajectoryComparison(AbstractBestActionPicker):)best_idx = i if idx == 1 else best_idx- return GetActionOutput(+ return ActionSamplerOutput(completion=completions[best_idx],extra_info={"comparison_log": comparison_log},)@@ -267,4 +267,4 @@ class BinaryTrajectoryComparison(AbstractBestActionPicker):return 0-BestResponsePicker = BinaryTrajectoryComparison | AskColleagues+ActionSampler = BinaryTrajectoryComparison | AskColleaguesdiff --git a/sweagent/agent/agents.py b/sweagent/agent/agents.pyindex 21916540..42df518e 100644--- a/sweagent/agent/agents.py+++ b/sweagent/agent/agents.py@@ -16,7 +16,7 @@ from tenacity import RetryErrorfrom typing_extensions import Selffrom sweagent import __version__, get_agent_commit_hash, get_rex_commit_hash, get_rex_version-from sweagent.agent.best_response_picker import BestResponsePicker+from sweagent.agent.action_sampler import ActionSamplerfrom sweagent.agent.history_processors import DefaultHistoryProcessor, HistoryProcessorfrom sweagent.agent.hooks.abstract import AbstractAgentHook, CombinedAgentHookfrom sweagent.agent.models import (@@ -132,19 +132,11 @@ class AgentConfig(BaseModel):formatting error, a blocked action, or a bash syntax error."""review_loop: ReviewLoopConfig | None = None- best_response_picker: BestResponsePicker | None = None- n_samples: int = 1+ best_response_picker: ActionSampler | None = None# pydantic configmodel_config = ConfigDict(extra="forbid")- @model_validator(mode="after")- def validate_n_samples(self) -> Self:- if self.n_samples is not None and self.n_samples > 1 and self.best_response_picker is None:- msg = "n_samples must be 1 when best_response_picker is not set"- raise ValueError(msg)- return self-class _BlockedActionError(Exception):"""Raised when the agent's action is blocked"""@@ -173,7 +165,7 @@ class Agent:_catch_errors: bool = True,_always_require_zero_exit_code: bool = False,review_loop_config: ReviewLoopConfig | None = None,- best_response_picker: BestResponsePicker | None = None,+ best_response_picker: ActionSampler | None = None,n_samples: int = 1,):"""The agent handles the behaviour of the model and how it interacts with the environment.