task_info

Author: TLSDC

src/agentlab/analyze/error_analysis/summarizer.py

@@ -91,8 +91,8 @@ def make_prompt(self, exp_results: ExpResult, summaries: list[str]): ...
     def __call__(self, exp_results: ExpResult) -> EpisodeAnalysis:
         """Run Change Summarizer for every step in the episode or extract a pre-computed one."""
 
-        # if exp_results.steps_info[-1].reward == 1:
-        #     return {"analysis": "Success", "summaries": {}}
+        if exp_results.steps_info[-1].reward == 1:
+            return {"analysis": "Success", "summaries": {}}
 
         with set_tracker("summary") as summaries_tracker:
             summaries = self.make_change_summaries(exp_results)
@@ -119,9 +119,7 @@ def make_change_summaries(self, exp_result: ExpResult) -> list[str]:
         return summaries
 
     def parse(self, raw_output: str) -> dict:
-        parsed_result = parse_html_tags(
-            raw_output, keys=["explanation", "success", "errorCategory"]
-        )[0]
+        parsed_result = parse_html_tags(raw_output, keys=["explanation", "errorCategory"])[0]
         return parsed_result
 
 
@@ -153,8 +151,12 @@ def format_summary(summary):
                 for action, action_error in zip(actions, action_errors)
             ]
         )
+
+        extra_info = exp_results.steps_info[-1].task_info
+
         return ERROR_CLASSIFICATION_PROMPT.format(
             goal=goal,
             historical_summaries=txt_summaries,
             action_history=txt_actions,
+            extra_info=extra_info,
         )

src/agentlab/analyze/error_analysis/summarizer_prompts.py

@@ -132,30 +132,24 @@
    - The current HTML or AX Tree observation
    - The user goal
 
-2. In case you think the task was unsuccessful, decide the category, or a combination thereof, under which the reason for failure lies.
-   If the task is successful, you can keep the error category as blank.
+2. Decide the error category, or a combination thereof, under which the reason for failure lies.
 
 3. Provide a brief explanation justifying your classification, referencing specific steps if helpful.
 
-Output format example for an unsuccessful interaction:
+Output format example for an interaction:
 
 <explanation>The agent opened the wrong GitLab page and never recovered...</explanation>
-<success>False</success>
 <errorCategory>["Navigation & Planning"]</errorCategory>
 
-Output format example for a successful interaction:
-
-<explanation>The agent opened the correct GitLab page and ...</explanation>
-<success>True</success>
-<errorCategory>[]</errorCategory>
-  
 Please follow this structure at every step. Keep your responses concise and clear. 
 
-Below are the details for the interaction.
+Below are the details for the interaction. Extra information yields additional information from the environment. It might not always be present or relevant.
 
 Overall goal: {goal}
 
 Historical change summaries: {historical_summaries}
 
 Action history: {action_history}
+
+Extra information: {extra_info}
 """