Skip to content

Command Configuration

In this document, we describe how to implement your own commands for the SWE-agent ACI. To see examples of command implementations, open the .sh and .py files in the config/commands folder.

Scaffolding

Every command subscribes to the following skeleton code.

# @yaml
# signature: [command] [argument(s)]
# docstring: [Brief description of what your command does.]
# arguments:
#   [argument 1 name]:
#       type: [type (i.e. integer, string)]
#       description: [Brief description of this argument]
#       required: [true|false]
#   [argument 2 name]:
#       ...
[command]() {
    # Implementation here
}
  • If a command takes in arguments, reference them via positional parameters notation (i.e. $1).
  • If there are no arguments, omit the arguments section.
  • The implementation for your command is unconstrained. There are no limitations on the form of the underlying command code.
  • The minimal documentation requirements are signature and docstring.
  • If you'd like multiple commands to make modifications to a similar body of functions, we recommend using global variables.
    • For instance, in config/commands/default.sh, you'll see we define the CURRENT_LINE variable for the file viewer. This variable is modified across multiple commands, including open, goto, scroll_up, scroll_down, and edit.
    • You can also third-party libraries (check out how we enable linting in config/commands/edit_linting.sh).
  • To show effects of the command, print to standard output (i.e. echo). SWE-agent is implemented such that it does not look for a return value from these commands.
  • The following environment variables are used to persist information between commands:
    • CURRENT_FILE: File that is currently open
    • CURRENT_LINE: First line of the window that is currently being shown/edited
    • WINDOW (start line to end line): Part of the file that is currently shown/edited
    • START_CURSOR, END_CURSOR: Only used for the cursors_* commands.
  • You can also use environment variables to set parameters for your commands. For this, edit the env_vars section of the config. For example, the WINDOW setting controls the number of context lines shown when editing a file.

Displaying the Command to SWE-agent

After you define a command, there are a small set of additional steps to making it available for the agent to use.

First, within your config file...

  • Add config/commands/<file name>.sh file to the command_files field.
  • Set the parse_command field to ParseCommandBash or ParseCommandDetailed. This key points to the functionality that generates how command documentation is shown to the agent.
  • Decide which template(s) you want to show the {command_docs} in.
    • We strongly recommend including {command_docs} in the system_template, which is the first message shown to the agent for every task instance episode.
    • You might also consider adding {command_docs} to the format_error_template, which is shown if the response provided by a model is malformed.
  • (Optional) Including a demonstration that uses a command is helpful to showcase proper use + increases the frequency with which the agent uses the command. If you'd like to add a demonstration...
    • Create a demonstration manually (i.e. python run.py --model human_thought ...) or automatically (i.e. python run_replay --traj_path ...)
    • Add/Update the demonstration to the demonstrations argument.
    • Update demonstration_template to control how the demonstration is displayed to the agent.

Config files

If you're not familiar with how SWE-agent configuration files work, we recommend checking out the config documentation.

Next, run your configuration and see how your agent uses the commands! 

python run.py --config_file config/[your config].yaml ...