verifiers 0.4.2.dev44__tar.gz → 0.4.2.dev45__tar.gz

{verifiers-0.4.2.dev44 → verifiers-0.4.2.dev45}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.4.2.dev44
+Version: 0.4.2.dev45
 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-0.4.2.dev44 → verifiers-0.4.2.dev45}/tests/unit/harnesses/test_rlm.py

@@ -57,12 +57,28 @@ def test_rlm_stop_reason_survives_completion_and_serialization(reason, condition
     assert restored.is_format_violation is (condition == "format_violation")
 
 
-def _sent_policy(config: RLMHarnessConfig, runtime_config) -> dict:
+def _sent_contract(config: RLMHarnessConfig, runtime_config) -> dict:
     runtime = SimpleNamespace(config=runtime_config, env={})
     metadata = RLMHarness(config)._runtime_metadata(
         SimpleNamespace(model="m"), _trace(), runtime, "http://endpoint", "s", None
     )
-    return metadata[RLM_RUNTIME_METADATA_KEY]["policy"]
+    return metadata[RLM_RUNTIME_METADATA_KEY]
+
+
+def _sent_policy(config: RLMHarnessConfig, runtime_config) -> dict:
+    return _sent_contract(config, runtime_config)["policy"]
+
+
+@pytest.mark.parametrize("max_retries", [0, 30])
+def test_rlm_sends_provider_max_retries_when_set(max_retries):
+    config = RLMHarnessConfig(id="rlm", provider_max_retries=max_retries)
+    provider = _sent_contract(config, DockerConfig())["provider"]
+    assert provider["max_retries"] == max_retries
+
+
+def test_rlm_leaves_provider_max_retries_to_nano_rlm_by_default():
+    provider = _sent_contract(RLMHarnessConfig(id="rlm"), DockerConfig())["provider"]
+    assert "max_retries" not in provider
 
 
 @pytest.mark.parametrize(

{verifiers-0.4.2.dev44 → verifiers-0.4.2.dev45}/verifiers/v1/harnesses/rlm/harness.py

@@ -101,6 +101,9 @@ class RLMHarnessConfig(HarnessConfig):
     exec_timeout: PositiveInt | None = None
     """IPython/native tool execution timeout in seconds; `None` uses nano-rlm's
     default (300). Separate from `tool_timeout`, which controls MCP calls."""
+    provider_max_retries: NonNegativeInt | None = None
+    """SDK retries per model call (429, 5xx, connection errors); `None` uses nano-rlm's
+    default (5). Raise it so long rollouts ride out extended rate-limit bursts."""
     allow_git: bool | None = None
     """Allow unrestricted Git history queries (e.g. `git log --all`); `None` uses
     nano-rlm's default (false). Ordinary Git commands remain allowed either way;
@@ -225,6 +228,11 @@ class RLMHarness(ACPHarness[RLMHarnessConfig]):
             "provider": {
                 "base_url": endpoint,
                 "api_key": secret,
+                **(
+                    {"max_retries": self.config.provider_max_retries}
+                    if self.config.provider_max_retries is not None
+                    else {}
+                ),
             },
             # None = passthrough: the key stays off the wire and nano-rlm's own
             # default applies.