Function process_single_instance

src/benchmarks/swe_bench_verified/run.py:205–241

Process a single instance. Used by ThreadPoolExecutor.

(instance_data: tuple)

    output_dir.mkdir(parents=True, exist_ok=True)

    def process_single_instance(instance_data: tuple) -> InstanceResult:
        """Process a single instance. Used by ThreadPoolExecutor."""
        idx, total, instance = instance_data
        instance_id = instance["instance_id"]

        print(f"\n{'='*80}")
        print(f"[{idx}/{total}] Starting: {instance_id}")
        print(f"{'='*80}")

        docker_client = None

        try:
            image = DockerExecClient.resolve_image_name(instance_id)
            print(f"[{instance_id}] Starting SWE-bench container: {image}")
            docker_client = DockerExecClient(image=image)
            agent = AgentSPEX(mcp_client=docker_client)

            result = run_instance(
                instance,
                agent,
                agent.mcp_client,
                args,
                output_dir,
                save_logs=args.save_logs,
            )
            print(f"[{instance_id}] Completed successfully")
            return result
        except Exception as e:
            print(f"[{instance_id}] Error: {e}")
            return InstanceResult(
                instance_id=instance_id,
                success=False,
                error=str(e),
            )
        finally:
            if docker_client is not None:
                docker_client.close()

    # Prepare instance data with index info
    instance_data_list = [