Add workplace-claude demo notebook and config

fern/versions/latest/pages/evaluation-tutorials/index.mdx

@@ -12,6 +12,12 @@ Here are the hands-on walkthroughs for running benchmarks, collecting rollouts,
 Run the EvalPlus coding benchmark and inspect rollout and aggregate metric outputs.
 </Card>
 
+<Card title="Workplace Assistant with Claude Code" href="https://github.com/NVIDIA-NeMo/Gym/blob/main/resources_servers/workplace_assistant/notebooks/workplace-claude-demo.ipynb">
+Run an agentic tool-use benchmark end to end with the Claude Code agent harness — config, rollouts, and BLADE analysis.
+
+<Badge minimal outlined>notebook</Badge>
+</Card>
+
 <Card title="Environment List" href="/evaluation/environment-list">
 Browse the built-in benchmark and training environments.
 </Card>

resources_servers/workplace_assistant/app.py

@@ -61,6 +61,13 @@ def setup_webserver(self) -> FastAPI:
         app.post("/{path}")(self.route_to_python_function)
         return app
 
+    # Register all 27 workplace tools as MCP tools via the catch-all route when expose_tools_over_mcp is enabled.
+    def mcp_tools(self, harvested, catchall):
+        specs = get_tools(["email", "calendar", "analytics", "project_management", "customer_relationship_manager"])[
+            "schemas"
+        ]
+        return harvested + [catchall.tool(s["name"], s["parameters"], s.get("description")) for s in specs]
+
     async def seed_session(self, request: Request, body: BaseSeedSessionRequest) -> BaseSeedSessionResponse:
         # init session once for each sample.
         session_id = request.session[SESSION_ID_KEY]

resources_servers/workplace_assistant/dataset_preprocess.py

@@ -16,7 +16,6 @@
 # Run as `python resources_servers/workplace_assistant/dataset_preprocess.py --split test``
 import argparse
 import json
-from copy import deepcopy
 from dataclasses import dataclass
 from typing import Any
 
@@ -84,26 +83,15 @@ def __init__(self, prompt_to_use: str = "SYS_PROMPT"):
         )
 
     def get_samples(self, split):
-        dataset = load_dataset("Nexusflow/250319-workplace_assistant-fulleval", split=split)
+        dataset = load_dataset("nvidia/Nemotron-RL-agent-workplace_assistant", split=split)
 
         processed_samples = []
 
         for d in dataset:
-            # convert into create params
-            create_params = deepcopy(self.base_create_params)
-            create_params["input"].append(
-                {
-                    "role": "user",
-                    "content": d["problem"],
-                }
-            )
-
-            ground_truth = json.loads(d["solution"])  # json loads ground truths/solutions
-
             processed_samples.append(
                 self.WorkbenchSample(
-                    create_params=create_params,
-                    reference=ground_truth,
+                    create_params=d["responses_create_params"],
+                    reference=d["ground_truth"],
                     category=d["category"],
                     environment_name=d["environment_name"],
                 )