Terminus-2

terminus2 runs the Terminus-2 terminal agent for Terminal-Bench 2, Verified, and 2.1 tasks. It accepts the model credentials supplied by --model-* and supports openai-chat and openai-responses model APIs.

How it works

• Prepare a terminal session. The harness starts Terminus-2 in the task environment and operates a tmux terminal pane in the workspace prepared by the benchmark and environment recipe.
• Run the agent loop. Model output is decoded with the selected response parser. The agent sends shell keystrokes, observes the terminal, and continues until it calls task completion or reaches a configured limit.
• Control context and return results. Context summarization is enabled by default for long terminal sessions. The harness returns the normalized trajectory, final answer, status, and execution metadata as a RunResult.

Parameters

Pass harness options with --harness-params '{...}'. The table intentionally exposes only behavior that normally needs tuning; terminal dimensions, service aliases, recording, and other integration options use their built-in defaults.

Parameter reference

Parameter	Type	Default	Choices / values	Description
parser_name	string	json	json / xml / tool_call	Response format.
max_turns	int	300	integer ≥ 1	Maximum agent turns for one task.
timeout	int	14400	integer ≥ 1	Harness wall-clock timeout in seconds.
enable_summarize	bool	true	true / false	Enable context summarization for long trajectories.
modality	string	llm	llm / vlm	Agent modality.
interleaved_thinking	bool	false	true / false	Enable interleaved thinking for compatible models.

Model API and response format

Supply the model endpoint and credentials through --model-base-url and --model-api-key; openai-chat and openai-responses are supported through --model-api-protocol. Use parser_name only when the model’s response format differs from the default JSON tool-call format.

Run examples

terminus2 is the second positional argument to agentcompass run <benchmark> terminus2 <model>.

Default configuration

Run Terminus-2 with its built-in Terminal-Bench defaults.

agentcompass run \
  terminal_bench_2 \
  terminus2 \
  "$MODEL_NAME" \
  --env docker \
  --model-base-url "$MODEL_BASE_URL" \
  --model-api-key "$MODEL_API_KEY"

Custom parameters

Adjust the response parser, turn limit, and harness timeout for a model or workload.

agentcompass run \
  terminal_bench_2 \
  terminus2 \
  "$MODEL_NAME" \
  --env docker \
  --harness-params '{
    "parser_name": "json",
    "max_turns": 120,
    "timeout": 7200
  }' \
  --model-base-url "$MODEL_BASE_URL" \
  --model-api-key "$MODEL_API_KEY"

Output

For each task, the harness returns a RunResult containing the normalized trajectory, final answer, execution status, and diagnostic metadata. The benchmark writes these task details and aggregate metrics under results/<benchmark>/<model>/<run>/; see Results.