fixes

Author: ollmer

src/agentlab/backends/browser/env.py

@@ -29,7 +29,7 @@ def reset(self, seed: int):
         self.backend.goto(self.task.url)
         setup_js = self.task.get_setup_js()
         if setup_js:
-            self.goal = self.backend.run_js(setup_js)
+            self.goal = self.task.parse_setup_result(self.backend.run_js(setup_js))
             logger.info(f"Task goal: {self.goal}")
         html = self.backend.page_html()
         screenshot = self.backend.page_screenshot()

src/agentlab/benchmarks/miniwob/task.py

@@ -107,6 +107,12 @@ def get_setup_js(self) -> str:
     """
         return f"async () => {{{js}}}"
 
+    def parse_setup_result(self, setup_result: str | dict | list) -> str:
+        if isinstance(setup_result, dict):
+            return setup_result["utterance"]
+        else:
+            return setup_result
+
     def get_teardown_js(self) -> str:
         return ""
 

src/agentlab/benchmarks/web_task.py

@@ -19,6 +19,9 @@ def filter_actions(cls, actions: list[ToolSpec]) -> list[str]:
     def get_setup_js(self) -> str:
         raise NotImplementedError
 
+    def parse_setup_result(self, setup_result: str | dict | list) -> str:
+        raise NotImplementedError
+
     def get_teardown_js(self) -> str:
         raise NotImplementedError