Terminal Coding Agent (Harbor)¶

Source: agent_library/agents/harbor_agent

What This Agent Does¶

harbor_agent is the minimal general-purpose terminal coding agent that the OpenSage team uses as the baseline for Harbor / T-Bench runs. It operates entirely inside a sandboxed Linux container with a small, disciplined toolset: terminal commands, file view/edit, and nothing else. No sub-agents, no ensembles, no MCP services. This is the "plain vanilla" production agent, good to copy when you want to start from a clean slate.

The prompt leans heavily on two principles: verify your work by running the code and re-read the task before finishing. Together, they are what keep a single-agent, single-model setup honest on coding tasks.

Key Design¶

• No multi-agent orchestration. One root agent, one model, a short tool list.
• File ops as first-class tools. view_file and str_replace_edit are given directly to the LLM instead of asking it to assemble sed commands.
• Background-task awareness. The agent can start long-running commands, check their status, and fetch output later without blocking its own turn.
• Skills disabled. enabled_skills=None: no bash-tool scripts; the agent works from raw terminal primitives only.

Agent Source¶

from opensage.agents.opensage_agent import OpenSageAgent
from opensage.toolbox.finish_task.finish_task import finish_task
from opensage.toolbox.general.bash_tools_interface import (
    get_background_task_output,
    list_background_tasks,
    run_terminal_command,
)
from opensage.toolbox.general.fileop import str_replace_edit, view_file

def mk_agent(opensage_session_id, model=None):
    if model is None:
        model = LiteLlm(model="openai/gpt-4o")

    return OpenSageAgent(
        name="harbor_agent",
        model=model,
        instruction=SYSTEM_PROMPT,
        tools=[
            finish_task,
            view_file, str_replace_edit,
            run_terminal_command,
            list_background_tasks, get_background_task_output,
        ],
        enabled_skills=None,
    )

The full SYSTEM_PROMPT (Role / Environment / Verify / Review / Best Practices) lives in the source file.

Run It¶

uv run opensage web \
  --agent agent_library/agents/harbor_agent \
  --config agent_library/agents/harbor_agent/config.toml \
  --port 8000

The bundled config.toml uses openai/gpt-4o with history compaction at 240k characters. For T-Bench runs, the main sandbox image name is derived from ${TASK_NAME}; override it per task.