Move LLMBENCH_WORKSPACE handling to server level

The server now checks for LLMBENCH_WORKSPACE environment variable
and uses it when calling setup_problem and grade functions.
If not set, it falls back to the configured working_dir.

This centralizes the environment variable handling in the server,
making the benchmark modules simpler and more predictable.

Author: Keno

src/tools/grade_problem.jl

@@ -53,10 +53,13 @@ function ClaudeMCPTools.execute(tool::GradeProblemTool, params::Dict)
         # Push testset to capture test results (Test module won't print when inside a testset)
         Test.push_testset(ts)
         try
+            # Use LLMBENCH_WORKSPACE if set, otherwise use working_dir
+            workspace = get(ENV, "LLMBENCH_WORKSPACE", tool.working_dir)
+            
             # Call the grade function with all arguments
             # Use invokelatest to handle world age issues when loading modules dynamically
             # Always pass all three parameters - the function has a default value for problem_id
-            result = Base.invokelatest(tool.grade_fn, tool.working_dir, transcript, problem_id)
+            result = Base.invokelatest(tool.grade_fn, workspace, transcript, problem_id)
         finally
             Test.pop_testset()
             # Restore TESTSET_PRINT_ENABLE

src/tools/setup_problem.jl

@@ -32,11 +32,14 @@ function ClaudeMCPTools.execute(tool::SetupProblemTool, params::Dict)
     problem_id = get(params, "problem_id", "")
 
     try
-        # Call the setup function with the working directory and problem_id
+        # Use LLMBENCH_WORKSPACE if set, otherwise use working_dir
+        workspace = get(ENV, "LLMBENCH_WORKSPACE", tool.working_dir)
+        
+        # Call the setup function with the workspace directory and problem_id
         # Use invokelatest to handle world age issues when loading modules dynamically
         # Always pass both parameters if we have a problem_id
         # The function can have a default value for problem_id
-        result = Base.invokelatest(tool.setup_fn, tool.working_dir, problem_id)
+        result = Base.invokelatest(tool.setup_fn, workspace, problem_id)
 
         # The setup function should return a problem description
         # Format it as a proper MCP response