Issue-Fixing Agent (SWE-Bench Pro)¶

Source: agent_library/agents/swebenchpro_agent

What This Agent Does¶

swebenchpro_agent is the OpenSage team's submission agent for SWE-Bench Pro. It fixes a bug described in a natural-language issue by exploring a real repository, reproducing the failure, editing source files, running the project's test suite, and finally emitting a prediction.patch file ready for evaluation.

The thing that makes this agent different from a vanilla coding agent:

• Two-phase workflow (explore -> solve). The file exports two agents: mk_agent (the solver) and mk_explore_agent (a read-only explorer). The explorer produces a structured "Exploration Summary" that the solver consumes as its starting context. The explorer never modifies code.

Key Design¶

• Explore-then-solve. Running an explorer first is essentially a pre-computed reconnaissance pass that seeds the solver with the critical files and a bug hypothesis.
• Structured handoff. The explorer's final output follows a strict template (### 1. Critical Files (MUST READ), ### 2. Bug Location Hypothesis, ### 3. NOT FOUND, and similar) that the solver's prompt references by section name.
• Multi-file-change checklist. The solver's prompt has an explicit "Before finish_task" checklist that enumerates ways multi-file refactors go wrong (missed callers, imports, helpers).
• Tool output persistence. Long tool outputs are written to /workspace/.tool_outputs/ and /workspace/.memory_observer_outputs/; the agent knows to read those if a summarized tool response is not enough.
• Multiple sandbox variants. The directory ships config.toml, docker_config.toml, nitrobox_config.toml, nitrobox_overlayfs_config.toml, docker_no_neo4j_config.toml: different sandbox backends for different eval rigs.

Solver Source (Abbreviated)¶

from opensage.agents.opensage_agent import OpenSageAgent
from opensage.toolbox.finish_task.finish_task import finish_task
from opensage.toolbox.general.agent_tools import (
    complain,
)
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
from opensage.toolbox.general.orchestration_tools import (
    call_subagent, create_subagent, get_available_models, list_subagents,
)

def mk_agent(opensage_session_id, model=None, planner=None):
    if model is None:
        model = LiteLlm(
            model="openai/gpt-5.5",
            api_key=os.environ.get("LITELLM_API_KEY"),
            base_url=os.environ.get("LITELLM_BASE_URL"),
            cache_control_injection_points=[
                {"location": "message", "role": "system"},
                {"location": "message", "index": -2},
                {"location": "message", "index": -1},
            ],
        )

    return OpenSageAgent(
        name="benchmark_agent",
        model=model,
        planner=planner,
        instruction=SOLVER_PROMPT,   # see repo for the full prompt
        tools=[
            finish_task,
            view_file, str_replace_edit,
            get_available_models,
            create_subagent, list_subagents, call_subagent,
            list_background_tasks, get_background_task_output,
            run_terminal_command,
        ],
    )

mk_explore_agent is a read-only variant with a shorter tool list (view_file, terminal commands only; no edits). See the source for both prompts.

Run It¶

For standard SWE-Bench Pro evaluation the repo is mounted at /app and the agent writes prediction.patch to /shared:

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

The directory ships several sandbox variants: config.toml, docker_config.toml, docker_no_neo4j_config.toml, nitrobox_config.toml, nitrobox_overlayfs_config.toml. Use docker_no_neo4j_config.toml for a plain Docker rig without the Neo4j sidecar.

Prerequisites¶

• Neo4j sandbox (auto-launched from [sandbox.sandboxes.neo4j] unless you use the no_neo4j config).
• A model that supports prompt caching if you want the 3-point cache_control_injection_points hint to reduce cost on long sessions.