verifiers 0.3.1.dev38__tar.gz → 0.3.1.dev40__tar.gz

{verifiers-0.3.1.dev38 → verifiers-0.3.1.dev40}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: verifiers
-Version: 0.3.1.dev38
+Version: 0.3.1.dev40
 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-0.3.1.dev38 → verifiers-0.3.1.dev40}/tests/v1/fixtures/counter_tool_v1.py

@@ -44,9 +44,9 @@ class CounterTaskset(vf.Taskset[CounterTask, CounterConfig]):
                 vf.TaskData(
                     idx=0,
                     prompt=(
-                        f"Call the `counter_bump` tool {TARGET} times, one call per turn — wait "
-                        "for each result before the next. After the last result, reply with "
-                        "<answer>done</answer>."
+                        f"Call the `bump` tool from the `counter` MCP server {TARGET} times, "
+                        "one call per turn — wait for each result before the next. After the "
+                        "last result, reply with <answer>done</answer>."
                     ),
                 ),
                 self.config.task,

{verifiers-0.3.1.dev38 → verifiers-0.3.1.dev40}/tests/v1/fixtures/echo_acp_resume_v1.py

@@ -59,8 +59,9 @@ class ACPResumeEnv(vf.SingleAgentEnv):
             if not first.terminated:
                 segments.append(
                     await interaction.turn(
-                        "Call `resume_recall` with the codeword from my previous "
-                        "message, then reply with exactly the tool result."
+                        "Call the `recall` tool from the `resume` MCP server with the "
+                        "codeword from my previous message, then reply with exactly the "
+                        "tool result."
                     )
                 )
             interaction.trace.info["acp_segments"] = [

{verifiers-0.3.1.dev38 → verifiers-0.3.1.dev40}/tests/v1/fixtures/echo_tool_v1.py

@@ -18,7 +18,7 @@ ECHO_TOKEN = "ok-7f3"  # the tool stamps this; only a real tool call can surface
 
 
 class EchoToolset(vf.Toolset[vf.ToolsetConfig]):
-    TOOL_PREFIX = "echo"  # the model sees `echo_back` (matches the prompt)
+    TOOL_PREFIX = "echo"
 
     @vf.tool
     def back(self, message: str) -> str:
@@ -53,7 +53,8 @@ class EchoToolTaskset(vf.Taskset[EchoToolTask, EchoToolConfig]):
                 vf.TaskData(
                     idx=0,
                     prompt=(
-                        f'Call the `echo_back` tool with the message "{PHRASE}", then reply '
+                        f"Call the `back` tool from the `echo` MCP server with the message "
+                        f'"{PHRASE}", then reply '
                         "with exactly what it returns inside <answer></answer> tags."
                     ),
                 ),

{verifiers-0.3.1.dev38 → verifiers-0.3.1.dev40}/tests/v1/fixtures/tool_response_image_v1.py

@@ -13,7 +13,7 @@ SYSTEM = "Call the requested tool before answering."
 
 
 class VisionToolset(vf.Toolset[vf.ToolsetConfig]):
-    TOOL_PREFIX = "vision"  # the model sees `vision_snapshot`
+    TOOL_PREFIX = "vision"
 
     @vf.tool
     def snapshot(self) -> list:
@@ -60,8 +60,8 @@ class ToolResponseImageTaskset(
                 vf.TaskData(
                     idx=0,
                     prompt=(
-                        "Call the `vision_snapshot` tool exactly once. After it returns, "
-                        "reply with exactly `done`."
+                        "Call the `snapshot` tool from the `vision` MCP server exactly once. "
+                        "After it returns, reply with exactly `done`."
                     ),
                     system_prompt=SYSTEM,
                 ),

{verifiers-0.3.1.dev38 → verifiers-0.3.1.dev40}/tests/v1/test_e2e.py

@@ -288,10 +288,10 @@ async def test_tool(run_v1, harness_runtime, tool_runtime, tmp_path):
     assert trace.ok
     assert trace.num_turns >= 2  # tool call + answer
     assert trace.reward == 1.0
-    # The interception server captured the advertised tools onto the trace (for tool-use SFT):
-    # the null harness offered the task's MCP tool as `echo_back`, schema included.
+    # The interception server captured the harness's advertised name and schema. Harnesses
+    # may qualify the same raw MCP tool differently, so the test checks its stable suffix.
     assert trace.tools
-    (echo_tool,) = [t for t in trace.tools if t.name == "echo_back"]
+    (echo_tool,) = [tool for tool in trace.tools if tool.name.endswith("back")]
     assert "message" in echo_tool.parameters.get("properties", {})
 
 
@@ -599,9 +599,9 @@ async def test_env_id_user_sim_with_tools(run_v1, tmp_path):
     assert assistant.task.data.prompt is None  # the scenario stayed off the wire
     assert user.num_turns >= 1  # the modeled user actually drove the exchange
     assert assistant.rewards["echoed"].score == 1.0  # the tool ran, mid-conversation
-    # The tool was advertised to the masked chat exactly as to any run.
+    # The tool was advertised to the masked chat, regardless of harness qualification.
     assert assistant.tools
-    assert any(tool.name == "echo_back" for tool in assistant.tools)
+    assert any(tool.name.endswith("back") for tool in assistant.tools)
 
 
 @pytest.mark.e2e

{verifiers-0.3.1.dev38 → verifiers-0.3.1.dev40}/verifiers/v1/harnesses/rlm/harness.py

@@ -28,9 +28,7 @@ RLM_STATE_DIR = ".vf-rlm"
 
 
 class RLMHarnessConfig(HarnessConfig):
-    version: str = Field(
-        default="41739cf3c2eb9859a90dc630db0d4739df404d23", min_length=1
-    )
+    version: str = Field(default="main", min_length=1)
     """Git ref (branch, tag, or commit) of nano-rlm to install."""
     max_depth: int = 0
     """Recursion depth rlm may spawn sub-harnesses to (RLM_MAX_DEPTH)."""