verifiers 0.1.15.dev17__tar.gz → 0.1.15.dev18__tar.gz

{verifiers-0.1.15.dev17 → verifiers-0.1.15.dev18}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.1.15.dev17
+Version: 0.1.15.dev18
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.1.15.dev17 → verifiers-0.1.15.dev18}/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-0.1.15.dev17 → verifiers-0.1.15.dev18}/verifiers/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.1.15.dev17"
+__version__ = "0.1.15.dev18"
 
 import importlib
 import os

{verifiers-0.1.15.dev17 → verifiers-0.1.15.dev18}/verifiers/envs/experimental/composable/harnesses/mini_swe_agent.py

@@ -24,8 +24,8 @@ DEFAULT_LOG_DIR = "/logs/agent"
 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,

{verifiers-0.1.15.dev17 → verifiers-0.1.15.dev18}/verifiers/utils/interception_utils.py

@@ -37,7 +37,7 @@ logger = logging.getLogger(__name__)
 KEEPALIVE_INTERVAL_SECONDS = float(
     os.environ.get("INTERCEPTION_SERVER_KEEPALIVE_INTERVAL_SECONDS", "3.0")
 )
-DEFAULT_CLIENT_MAX_SIZE_BYTES = 16 * 1024 * 1024
+DEFAULT_CLIENT_MAX_SIZE_BYTES = 24 * 1024 * 1024
 
 
 class StreamInterrupted(InfraError):

{verifiers-0.1.15.dev17 → verifiers-0.1.15.dev18}/verifiers/v1/sandbox.py

@@ -25,10 +25,11 @@ class SandboxConfig(Config):
     install_timeout: int = 300
     setup_commands: list[str] = []
     setup_timeout: int = 300
+    labels: list[str] = []
     scope: Literal["rollout", "group", "global"] = "rollout"
     prefer: Literal["program"] | None = None
 
-    @field_validator("packages", "setup_commands", mode="before")
+    @field_validator("packages", "setup_commands", "labels", mode="before")
     @classmethod
     def validate_string_list(cls, value: object) -> object:
         if isinstance(value, str):