Use real parallel tool calls for mini-swe-agent (#1505)

Author: xeophon

packages/harnesses/harnesses/mini_swe_agent.py

@@ -18,8 +18,8 @@
 MINI_SWE_AGENT_DEFAULT_PACKAGE_SHA256 = (
     "694df4de1337e665e3cd82e99f93374f573bf52b8e7c362ac5d8045ad9f7c37c"
 )
-MINI_SWE_AGENT_DEFAULT_CONFIG_SPEC = "mini_textbased"
-MINI_SWE_AGENT_DEFAULT_MODEL_CLASS = "litellm_textbased"
+MINI_SWE_AGENT_DEFAULT_CONFIG_SPEC = "mini"
+MINI_SWE_AGENT_DEFAULT_MODEL_CLASS = "litellm"
 MINI_SWE_AGENT_DEFAULT_ENVIRONMENT_TIMEOUT = 120
 
 
@@ -71,6 +71,7 @@ class MiniSWEAgentProgramConfig(vf.ProgramConfig):
     config_spec: str = MINI_SWE_AGENT_DEFAULT_CONFIG_SPEC
     model_class: str = MINI_SWE_AGENT_DEFAULT_MODEL_CLASS
     environment_timeout: int = MINI_SWE_AGENT_DEFAULT_ENVIRONMENT_TIMEOUT
+    parallel_tool_calls: bool = True
     extra_config_specs: list[str] | None = None
     sandbox: vf.SandboxConfig | None = vf.SandboxConfig()
 
@@ -117,6 +118,8 @@ def resolve(self) -> vf.ProgramConfig:
             "model.cost_tracking=ignore_errors",
             "-c",
             "model.model_kwargs.custom_llm_provider=openai",
+            "-c",
+            f"model.model_kwargs.parallel_tool_calls={str(self.parallel_tool_calls).lower()}",
         ]
         for spec in self.extra_config_specs or []:
             config_args.extend(["-c", shlex.quote(spec)])

tests/test_v1_mini_swe_agent.py

@@ -69,10 +69,15 @@ def test_mini_swe_agent_builds_sandbox_program():
         )
     )
     program = cast(dict[str, Any], harness.config.program.data())
+    command = cast(list[str], program["command"])
+    script = command[-1]
 
     assert isinstance(harness, vf.Harness)
     assert program["sandbox"] is not False
     assert "OPENAI_MODEL" in cast(dict[str, object], program["env"])
+    assert "-c mini " in script
+    assert "model.model_class=litellm" in script
+    assert "model.model_kwargs.parallel_tool_calls=true" in script
     assert "apt-get -o Acquire::Retries=3 update" in cast(str, program["setup"])
     assert "apt-get -o Acquire::Retries=3 install" in cast(str, program["setup"])
     assert "/mini-swe-agent/prompt.txt" in cast(dict[str, object], program["files"])

verifiers/envs/experimental/composable/harnesses/mini_swe_agent.py

@@ -24,8 +24,8 @@
 DEFAULT_LOG_PATH = f"{DEFAULT_LOG_DIR}/mini-swe-agent.log"
 DEFAULT_TRAJECTORY_PATH = f"{DEFAULT_LOG_DIR}/mini-swe-agent.traj.json"
 DEFAULT_AGENT_WORKDIR = "${AGENT_WORKDIR:-/app}"
-DEFAULT_CONFIG_SPEC = "mini_textbased"
-DEFAULT_MODEL_CLASS = "litellm_textbased"
+DEFAULT_CONFIG_SPEC = "mini"
+DEFAULT_MODEL_CLASS = "litellm"
 DEFAULT_ENVIRONMENT_TIMEOUT = 120
 
 
@@ -52,6 +52,7 @@ def build_mini_swe_agent_run_command(
     config_spec: str = DEFAULT_CONFIG_SPEC,
     model_class: str = DEFAULT_MODEL_CLASS,
     environment_timeout: int = DEFAULT_ENVIRONMENT_TIMEOUT,
+    parallel_tool_calls: bool = True,
     extra_config_specs: list[str] | None = None,
 ) -> str:
     """Build the shell command that configures and runs mini-SWE-agent.
@@ -80,6 +81,8 @@ def build_mini_swe_agent_run_command(
         "model.cost_tracking=ignore_errors",
         "-c",
         "model.model_kwargs.custom_llm_provider=openai",
+        "-c",
+        f"model.model_kwargs.parallel_tool_calls={str(parallel_tool_calls).lower()}",
     ]
     # Config specs are the mini CLI's native override format; use them for cwd,
     # timeout, model class, and optional system prompt wiring.
@@ -141,6 +144,7 @@ def mini_swe_agent_harness(
     config_spec: str = DEFAULT_CONFIG_SPEC,
     model_class: str = DEFAULT_MODEL_CLASS,
     environment_timeout: int = DEFAULT_ENVIRONMENT_TIMEOUT,
+    parallel_tool_calls: bool = True,
     extra_config_specs: list[str] | None = None,
 ):
     """Create a Harness configured for mini-SWE-agent."""
@@ -168,6 +172,7 @@ def mini_swe_agent_harness(
             config_spec=config_spec,
             model_class=model_class,
             environment_timeout=environment_timeout,
+            parallel_tool_calls=parallel_tool_calls,
             extra_config_specs=extra_config_specs,
         ),
         system_prompt=system_prompt,