Endless Terminals

Description

Endless Terminals is an environment for training and evaluating terminal agents on procedurally generated command-line tasks. The benchmark provides diverse terminal-use tasks spanning file operations, log management, data processing, scripting, and database operations. Tasks are designed for reinforcement learning with binary episode-level rewards.

This OpenReward implementation is based on the Endless Terminals repository. The original benchmark contains 3,255 tasks; this implementation includes 2,490 tasks.

Capabilities

• File operations and management
• Log analysis and processing
• Data transformation and scripting
• Database operations
• System administration tasks
• Multi-step terminal command execution

Compute Requirements

Agents are given a sandbox with 1 CPU and 2GB RAM. Each task runs in an isolated Docker container with task-specific files and tooling.

License

Apache 2.0

Tasks

There is one split in this environment:

• train: 2,490 terminal-based tasks (subset of 3,255 in original benchmark)

Each task provides a containerized environment with specific files and objectives. Agents must execute terminal commands to transform the initial state into the expected final state.

Reward Structure

This is a sparse, verifiable reward environment. Rewards are computed when the agent submits their answer:

• 1.0: All verification tests pass (final state matches expected)
• 0.0: Any test fails

No LLM grader is used. Each task has pytest-based verification scripts (test_initial_state.py, test_final_state.py) that validate the container state.

Data

Task data is sourced from HuggingFace. Each task contains:

• instruction.md: Task description and requirements
• environment/Dockerfile: Task-specific container definition
• environment/image_sha.txt: Docker image digest
• tests/test_final_state.py: Pytest verification logic
• tests/test.sh: Test execution wrapper

Tools

Agents have access to 4 tools:

• bash: Execute bash commands in the container
• view: View file contents or directory listings
• str_replace: Replace unique strings in files
• create_file: Create new files with specified content

Grading uses a hidden @terminal tool: when the agent has finished modifying the container and replies with a plain message (no tool call), the environment runs the task's verification tests against the sandbox state and returns the reward. The assistant message text itself is not graded.

Time Horizon

Endless Terminals is a multi-turn environment where agents iteratively execute commands, explore the file system, and modify state before finishing with a plain-text reply.

[Statistics on average tool calls here]

Environment Difficulty

Results from the original paper (dev set performance):

Gains transfer to human-curated benchmarks like TerminalBench 2.0.

Safety

Endless Terminals tasks are run in isolated Docker containers. Agents interact only with pre-defined task environments and cannot affect external systems or the host machine.

Citations

This environment implements the Endless Terminals benchmark. If you use this environment, please cite the original paper:

@article{gandhi2026endless,
  title     = {Endless Terminals: Scaling RL Environments for Terminal Agents},
  author    = {Gandhi, Kanishk and Garg, Shivam and Goodman, Noah D. and Papailiopoulos, Dimitris},
  journal   = {arXiv preprint arXiv:2601.16443},
  year      = {2026}
}