agent

Agents

The agent folder contains the logic for handling model inference and facilitating their interaction with SWEEnv. The following documentation describing the purpose and classes of each file.

agents.py

This file defines the Agent class, which facilitates the interaction between an agent and the environment. The AgentConfig and AgentArguments data classes compile all arguments into a single file.

• Agent: Main class for handling model behavior + interaction with environment
  • __init__: Sets up model, assistant, configurations, and arguments
  • state_command: Getter for bash command for extracting env. state
  • setup: Resets cost stats, initializes system message (+ demonstrations), and returns full list of bash commands to define within environment.
  • forward: Main inference call to model.
  • forward_model: Determines appropriate observation template, then makes inference call to model
  • forward_with_format_check: Invokes forward_model, with retry calls to handle blocked or malformed actions.
  • forward_with_error_check: Wraps forward_with_format_check with exception handling.

commands.py

This file defines the abstraction for custom commands (non-native functions that are implemented in bash) that agents can invoke in swe-agent environment. On top of the abstraction, helper functions to extract commands' documentation and compile .sh files into separate Command objects are provided. There are also fields for establishing the input/output of each action and control flow of actions via templates.

• AssistantMetadata: Defines templates for formatting input/output to sub-assistant calls
• Command: Defines fields of a custom command
• ControlMetadata (WIP): Defines template fields that format the observations for the next agent forward inference call
• generate_command_docs: Extracts docstrings from each command to form comprehensive documentation.
• parse_command_file: Converts bash file content to separate Command objects

models.py

This file defines the abstraction for running inference on API models. In addition, the BaseModel abstraction also defines a set of cost-related fields for tracking instance-level and total expenses accumulated across a single model run.

• AnthropicModel: Handles inference + cost logging for Anthropic Models
• BedrockModel: handles inference + cost logging for Amazon Bedrock-provided models (Anthropic Claude only)
• APIStats: Cost tracking fields that are updated per model inference
• BaseModel: Abstract class that defines the common logic for updating cost stats
• get_model: Returns initialized [Anthropic|Bedrock|Human|OpenAI]Model based on given arguments + commands
• HumanModel: Handles inference for human task worker
• ModelArguments: Model name, hyperparameter, and cost limit arguments
• OpenAIModel: Handles inference + cost logging for OpenAI models

parsing.py

This file defines the abstraction for parsing the output of the model inference. The Parsing class is used to extract the relevant information from the model's output and format it into a response that can be used by the Agent class.

• Parsing: Abstract class that defines the common logic for parsing model output

history_processors.py

This file defines the abstraction for processing the history of the environment. The HistoryProcessor class is used to extract the relevant information from the history of the environment and format it into a response that can be used by the Agent class.

• HistoryProcessor: Abstract class that defines the common logic for processing the history of the environment
• DefaultHistoryProcessor: Default implementation of HistoryProcessor that processes the history of the environment

Environment Usage

• To skip over a task instance, use the skip keyword
• To submit for evaluation, use the submit keyword
• To exit the SWEEnv environment, perform a keyboard interrupt (^ c)