Get the path where trajectory is being saved.

Get working directory input from user (for interactive mode). Returns: Working directory path

Async factory to create and initialize TraeAgent.

Check if the folder is a git repository.

Check if the LLM indicates that the task is completed. Override for custom logic.

Log an agent step to the execution log.

Main entry point for benchmark evaluation script. Parses command-line arguments and runs patch generation and/or evaluation.

Main entry point for the CLI.

A powerful CLI wrapper for the TextEditorTool that supports sub-commands.

Truncate the output if it's too long.

Execute tool calls in parallel

Parse a ToolCall into a Gemini FunctionCall Part for history.

Parse the tool call from the LLM response.

Parse the tool call from the LLM response.

Parse the tool call from the LLM response.

Parse the tool call result from the LLM response.

Parse the tool call result from the LLM response to FunctionCallOutput format.

Parse the tool call result from the LLM response.

Prepare experiment Docker container for a given instance. The container mounts the results directory for this instance, so al

Build Trae Agent and UV inside a base Ubuntu container. Save built artifacts to workspace for later use in experiment containers.

Print initial task configuration details.

Pull missing Docker images required for all instances.

Search for a class in the database. Args: identifier: the identifier of the class to search for Returns:

Record an agent execution step. Args: step_number: Step number in the execution state: Current state of the agent

Reflect on tool execution result. Override for custom reflection logic.

Run a shell command asynchronously with a timeout.

Run a shell command asynchronously.

Run patch generation for all instances in the dataset, with parallelism controlled by max_workers.

Run all instances concurrently using ThreadPoolExecutor. Args: max_workers: Maximum number of worker threads. If None, defaults t

Run evaluation using the benchmark harness. Evaluation results and predictions.json are stored in the task results directory.

Execute tool calls in sequential

Executes tool calls sequentially, routing to Docker if necessary.

Set the chat history.

Set the chat history.

Set the initial task for RUN mode.

Set the trajectory recorder for this agent.

Determine whether to use the max_completion_tokens parameter.Primarily used for Azure OpenAI's newer models (e.g., gpt-5).

Start recording a new trajectory. Args: task: The task being executed provider: LLM provider being used m

Stops the pexpect shell and cleans up the container if managed by this instance.

Implement the str_replace command, which replaces old_str with new_str in the file content

Implement the str_replace command, which replaces old_str with new_str in the file content

Return a message indicating that the task is incomplete. Override for custom logic.

Update the console with agent status. Args: agent_step: Current agent step information agent_execution: Complete agen

Update token counts from agent execution.

Validate the path for the str_replace_editor tool.

Initialize the BenchmarkEvaluation class. Args: benchmark: Benchmark name. working_dir: Path for workspace (

Initializes the DockerToolExecutor.

Initialize trajectory recorder. Args: trajectory_path: Path to save trajectory file. If None, generates default path.