Calculate the score of the current state.

Initialize session and return the search results page

Creates a log file and logging object for the corresponding session ID

Summarize history steps using the OpenAI Chat Completions API. Args: history_steps: List of formatted history strings. api_key: O

executes actions in parallel and waits for all env to finish

Evaluate if the episodes are successful or not. (Default) implementation is to check info['won'] of the last step.

Convert to dictionary for serialization.

Collects trajectories through parallel agent-environment agent_loop. Parameters: gen_batch (DataProto): Initial batch wit

Constructor for simulated server serving WebShop application Arguments: filter_goals (`func`) -- Select specific goa

Build AlfWorld environment

Build GAIA/Tool Use environment

Build WebShop environment

Convert the list of dictionaries (each containing trajectory, step_rewards, env_score) from the internal rollout loop into a DataProt

Create followup message for subsequent turns.

Create HTML visualization of a trajectory.

Create initial prompt using the template for first turn.

Compare model answer with ground truth

Extract action from structured response.

Extract final answer from action text

Extract answer text from action string

Fetch history using summarization strategy.

Generate simple action.

Generate a mock staged response.

Generate staged response for environment interaction.

Check if action contains a final answer

Check if action contains a tool call

Initialize and return a list of specific AgentGym environment clients based on the ports provided in the config.

Check if action indicates task completion

Judge a single answer against ground truth using LLM. Returns (analysis, is_correct).

Load recent entries from memory.md into cache.

Load a specific tool

Load tools specified in tool_names

Parse tool call from action text. Returns: Tuple of (parsed_action, is_valid)

Process a single action

Query the LLM API using chat history.

Auto-register discovered tools.

Extract and save a specific patch from the image with 10% margins.

Configure API based on environment variables.

Create output directory if needed.

Simple packaging when parent methods not available.

Convert DataProto batch to list of dicts.

Convert a trajectory to formatted text.

Generate action based on current observation using chat conversation. Returns: Tuple of (raw_response, action)

Print data analysis

Build environment based on type

Group observations by index and then cluster identical observations within each index group. Assigns a unique step_group_uid (UUID) to each c

Builds and returns the EasyOCR reader model. Parameters: languages (list): A list of language codes for the OCR model.

Context manager to capture the standard output. Yields: StringIO: The captured output.

Handle special case when Faucet is toggled on. In this case, we need to execute a `CleanAction` in the simulator on every object in t