Mock player for testing arena validation and result parsing.

Metadata for a single simulation, storing team assignments explicitly.

Metadata for a single simulation, storing team assignments explicitly.

Result of execute_round, used to communicate status to get_results.

Result of execute_round, used to communicate status to get_results.

Normalized, arena-agnostic playback payload. The player shell only relies on ``w``, ``h``, ``frames`` (each frame carrying a ``turn``), ``win

Parses game log files into structured data

Mock environment that simulates container file system and command execution.

A class that gives agent access to a partial view of the game state. NOTE: Instead of passing `game` directly as a reference to the agent,

A single playable game located within a tournament folder. Either it lives inside a per-round tar archive (``archive`` + ``member``) or it is a

A wrapper around metadata dictionary with convenient access methods

Everything the replay layer needs about a tournament folder.

A dummy player that does nothing. Mainly for testing purposes.

Create a triple plot showing grounding, validation, and testing/analysis metrics.

Class for aggregating LLM-as-judge evaluation results.

Exception raised when a function times out

Monitor AWS Batch jobs

Information about a single agent

BattleCode23 arena implementation. Lifecycle: 1. validate_code() - Source-level structural checks only (in agent container) 2. execute_ro

BattleCode24 arena implementation. Lifecycle: 1. validate_code() - Source-level structural checks only (in agent container) 2. execute_ro

Cache entry with data and timestamp

`DefaultAgent` from mini-SWE-agent (https://github.com/SWE-agent/mini-swe-agent) with per-player debug logging.

DockerEnvironment that also accepts a plain command string. mini-swe-agent v2's `execute` takes an action dict (`{"command": ...}`), but CodeClas

Per-token costs in USD.

Metadata about a game session

Custom formatter that uses Rich for colorized output.

Simple in-memory cache with timeout and request coalescing Prevents cache stampede by using per-key locks to ensure only one request computes

Information about a single trajectory

Schema for structured output of the model.