verifiers 0.4.2.dev30__tar.gz → 0.4.2.dev32__tar.gz

{verifiers-0.4.2.dev30 → verifiers-0.4.2.dev32}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.4.2.dev30
+Version: 0.4.2.dev32
 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.dev30 → verifiers-0.4.2.dev32}/verifiers/v1/agent.py

@@ -95,6 +95,16 @@ def resolve_rollout_timeouts(timeout: TimeoutConfig, task: Task) -> RolloutTimeo
     )
 
 
+def _check_artifact_runtime(runtime_config: RuntimeConfig) -> None:
+    """Artifact collection reads the task's paths out of the box, so the box needs a
+    filesystem of its own."""
+    if not backend_cls(runtime_config).SUPPORTS_TASK_IMAGES:
+        raise TypeError(
+            "artifact collection requires a runtime with its own filesystem; the "
+            f"{runtime_config.type} runtime cannot run images"
+        )
+
+
 def _check_borrowed_placement(
     task: Task, runtime: Runtime, base_config: RuntimeConfig
 ) -> None:
@@ -747,14 +757,8 @@ class Agent:
         attempt: int,
     ) -> Trace:
         params = self._rollout_params(task, runtime, dict(shared_tools or {}))
-        if (
-            collect_artifacts
-            and not backend_cls(params["runtime_config"]).SUPPORTS_TASK_IMAGES
-        ):
-            raise TypeError(
-                "artifact collection requires a runtime with its own filesystem; the "
-                f"{params['runtime_config'].type} runtime cannot run images"
-            )
+        if collect_artifacts:
+            _check_artifact_runtime(params["runtime_config"])
         run = Rollout(
             task=task,
             on_trace=on_trace,
@@ -786,6 +790,7 @@ class Agent:
         runtime: Runtime | None = None,
         tools: Mapping[str, SharedToolServer] | None = None,
         on_trace: Callable[[Trace], None] | None = None,
+        collect_artifacts: bool = False,
     ) -> AsyncIterator[Interaction]:
         """Interact with this agent turn-by-turn: a full rollout of `task` where
         the CALLER is the run's user — the one exchange surface. Yields an
@@ -801,7 +806,8 @@ class Agent:
         caller's to hide: hand the interaction a task whose `data.prompt` is None
         and keep the scenario on a scoring-side field (the user-sim env's contract).
 
-        `runtime` and `tools` borrow live resources from their owners, just as
+        `runtime` and `tools` borrow live resources from their owners, and
+        `collect_artifacts` captures the task's declared artifacts at close, just as
         they do for `run()`; an env supplies its taskset's shared tools
         automatically for tasks loaded from that taskset.
 
@@ -818,10 +824,13 @@ class Agent:
         trace_id = uuid.uuid4().hex
         with scope(trace=short_id(trace_id), agent=self._log_name):
             params = self._rollout_params(task, runtime, dict(tools or {}))
+            if collect_artifacts:
+                _check_artifact_runtime(params["runtime_config"])
             run = Rollout(
                 task=task,
                 has_user=True,
                 on_trace=on_trace,
+                collect_artifacts=collect_artifacts,
                 trace_id=trace_id,
                 **params,
             )
@@ -1030,6 +1039,7 @@ class _EpisodeAgent(Agent):
         runtime: Runtime | None = None,
         tools: Mapping[str, SharedToolServer] | None = None,
         on_trace: Callable[[Trace], None] | None = None,
+        collect_artifacts: bool = False,
     ) -> AsyncIterator[Interaction]:
         """The agent's `interaction`, with every trace stamped with its standing
         at mint and captured in `completed` at close — an interaction driven from
@@ -1051,6 +1061,7 @@ class _EpisodeAgent(Agent):
                 runtime=runtime,
                 tools=tools if tools is not None else self._shared_for(task),
                 on_trace=self._watch(remember),
+                collect_artifacts=collect_artifacts,
             ) as interaction:
                 yield interaction
         finally:

{verifiers-0.4.2.dev30 → verifiers-0.4.2.dev32}/verifiers/v1/dialects/chat.py

@@ -422,7 +422,7 @@ class ChatDialect(Dialect[ChatCompletion]):
             "extra_body",
         }
     )
-    routes = ("/v1/chat/completions",)
+    routes = ("/v1/chat/completions", "/chat/completions")
     upstream_path = "/chat/completions"
     response_type = ModdedChatCompletion
 

{verifiers-0.4.2.dev30 → verifiers-0.4.2.dev32}/verifiers/v1/dialects/responses.py

@@ -369,7 +369,7 @@ class ResponsesDialect(Dialect[OpenAIResponse]):
             "truncation",
         }
     )
-    routes = ("/v1/responses",)
+    routes = ("/v1/responses", "/responses")
     upstream_path = "/responses"
     response_type = OpenAIResponse
 

{verifiers-0.4.2.dev30 → verifiers-0.4.2.dev32}/verifiers/v1/harnesses/prime_agent/harness.py

@@ -27,7 +27,6 @@ PRIME_AGENT_COMMIT: Literal["ea9d9602ec0938d9e454e8963107cd85af1565c6"] = (
 PRIME_AGENT_VERSION = "0.10.2"
 PRIME_AGENT_DIR = "/var/tmp/vf-prime-agent"
 STATE_ROOT = "/tmp/vf-prime-agent-runs"
-PROVIDER = "intercept"
 LIFECYCLE_META_NAMESPACE = "ai.primeintellect.prime-agent"
 KEY_VAR = "PRIME_AGENT_INTERCEPT_KEY"
 ENV_AGENT_DIR = "PRIME_AGENT_CODING_AGENT_DIR"
@@ -82,8 +81,7 @@ class PrimeAgentHarnessConfig(HarnessConfig):
 
     autonomous: bool = False
     """Enable Prime Agent's autonomous continuation loop."""
-    # Over Responses it replays an earlier turn without its reasoning.
-    dialect: Literal["chat_completions", "anthropic_messages"] = "chat_completions"
+    dialect: Literal["auto"] = "auto"
 
 
 class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
@@ -208,22 +206,20 @@ class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
             None,
             "none",
         ) or ctx.model.rsplit("/", 1)[-1].startswith(("gpt-5", "o1", "o3", "o4"))
+        provider, separator, model = ctx.model.partition("/")
+        if not separator:
+            provider, model = "openai", ctx.model
+        elif provider in ("internal", "dev"):
+            provider, model = "prime-inference", ctx.model
         models = {
             "providers": {
-                PROVIDER: {
-                    # Its Anthropic provider appends `/v1/messages` to the base URL.
-                    "baseUrl": endpoint.removesuffix("/v1")
-                    if self.config.dialect == "anthropic_messages"
-                    else endpoint,
-                    "api": {
-                        "chat_completions": "openai-completions",
-                        "responses": "openai-responses",
-                        "anthropic_messages": "anthropic-messages",
-                    }[self.config.dialect],
+                provider: {
+                    # The native provider owns its API path, including any version prefix.
+                    "baseUrl": endpoint.removesuffix("/v1"),
                     "apiKey": KEY_VAR,
                     "models": [
                         {
-                            "id": ctx.model,
+                            "id": model,
                             "reasoning": reasoning,
                             "input": ["text", "image"],
                         }
@@ -246,9 +242,9 @@ class PrimeAgentHarness(ACPHarness[PrimeAgentHarnessConfig]):
             "acp",
             "--no-session",  # The ACP connection owns the session's lifetime.
             "--provider",
-            PROVIDER,
+            provider,
             "--model",
-            ctx.model,
+            model,
             "--daemon-socket",
             f"{root}/daemon.sock",
             "--offline",