Terminal Bench Integration into rLLM

[JasonWei05]

This PR integrates Terminal Bench’s Terminus 1 agent into rLLM for reproducible terminal task evaluation. Requirements: Python ≥ 3.12, pip install terminal-bench, and OPENAI_API_KEY set. Run the example with:

python examples/terminal/run_terminal.py

Edited files (key changes)

• rllm/rllm/environments/terminal/terminal_terminus.py

  • Adds TerminalTerminusEnv that bridges rLLM and TB.
  • Spins up Docker containers via TB Terminal and DockerComposeManager, and creates a TmuxSession for the agent.
  • Builds the initial prompt from TB’s Terminus prompt template; executes command batches; triggers unit tests on completion or episode cap; parses results via TB ParserFactory.
  • Records asciinema markers when logging_dir is provided; cleans up containers on close.

• rllm/rllm/agents/terminal_terminus_agent.py

  • TerminalTerminusAgent: thin, state‑holding agent that maintains alternating user/assistant messages and mirrors raw model responses into Action objects.
  • Terminal‑specific logic is kept out of the agent and handled by the environment.

• rllm/examples/terminal/prepare_terminal_data.py

  • Converts TB datasets into minimal rLLM task dicts: task_path, task_id, instruction.
  • Defers parser/timeouts/shell settings to TB’s task config (loaded by the env).

• rllm/examples/terminal/run_terminal.py

  • Runnable example that wires TerminalTerminusAgent, TerminalTerminusEnv, and TerminalLiteLLMEngine through AgentWorkflowEngine and TerminalWorkflow.
  • Loads tasks from TB registry and prints final accuracy.

• rllm/rllm/workflows/terminal_workflow.py

  • Workflow that runs multi‑step interactions with Terminus; evaluates tests on completion or error and surfaces a consistent termination.

• rllm/workflows/workflow.py

  • Minor edit to hand uid to the environment's reset function.

• rllm/rllm/integrations/terminal_terminus_1.py

  • RLLMTerminus: thin wrapper subclass over TB’s Terminus exposing private functions and variables.
  • TerminalLiteLLMEngine: minimal rollout engine delegating completions to TB’s Chat + LiteLLM and Terminus’ _handle_llm_interaction wrapper.
  • Returns ModelOutput with assistant JSON text and token counts.

[jeffreysijuntan]

Since this engine is for the sole purpose of TerminalBench, can you move it to /integrations/terminal_terminus_1.py.

[JasonWei05]

Done!