remove unused scripts in gaia example

examples/gaia_agent/eval.py

@@ -21,8 +21,7 @@
 from tapeagents.tools.web_search import SearchAction
 
 from .scorer import question_scorer
-from .steps import GaiaAnswer, GaiaQuestion, ImageObservation
-from .tape import GaiaMetadata, GaiaTape
+from .steps import GaiaAnswer, GaiaMetadata, GaiaQuestion, GaiaTape, ImageObservation
 
 logger = logging.getLogger(__name__)
 

examples/gaia_agent/scripts/browsergym_ui.py

@@ -14,8 +14,7 @@
 from tapeagents.steps import ReasoningThought
 from tapeagents.tools.container_executor import init_code_sandbox
 
-from ..steps import GaiaQuestion
-from ..tape import GaiaTape
+from ..steps import GaiaQuestion, GaiaTape
 
 logger = logging.getLogger(__name__)
 

examples/gaia_agent/scripts/collect_results.py

@@ -2,10 +2,10 @@
 import os
 from collections import defaultdict
 
-from examples.gaia_agent.tape import GaiaTape
 from tapeagents.io import load_tapes
 
 from ..eval import ensemble_files, majority_vote, tape_correct
+from ..steps import GaiaTape
 
 logging.basicConfig(level=logging.INFO)
 

examples/gaia_agent/scripts/debug.py

@@ -10,7 +10,7 @@
 from tapeagents.orchestrator import get_agent_and_env_from_config, main_loop
 
 from ..eval import load_dataset, task_to_observations
-from ..tape import GaiaMetadata, GaiaTape
+from ..steps import GaiaMetadata, GaiaTape
 
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)

examples/gaia_agent/scripts/prepare_submission.py

@@ -6,7 +6,7 @@
 from tapeagents.io import load_tapes
 
 from ..eval import get_exp_config_dict, load_dataset
-from ..tape import GaiaTape
+from ..steps import GaiaTape
 
 logging.basicConfig(level=logging.INFO)
 

examples/gaia_agent/scripts/studio.py

@@ -11,8 +11,7 @@
 from tapeagents.studio import Studio
 from tapeagents.tools.container_executor import init_code_sandbox
 
-from ..steps import GaiaQuestion
-from ..tape import GaiaTape
+from ..steps import GaiaQuestion, GaiaTape
 
 logging.basicConfig(
     level=logging.INFO,

examples/gaia_agent/scripts/tape_browser.py

@@ -12,8 +12,7 @@
 from tapeagents.tape_browser import TapeBrowser
 
 from ..eval import calculate_accuracy, get_exp_config_dict, tape_correct
-from ..steps import GaiaStep
-from ..tape import GaiaTape
+from ..steps import GaiaStep, GaiaTape
 
 logging.basicConfig(level=logging.INFO)
 

examples/gaia_agent/scripts/tape_diff.py

@@ -10,7 +10,7 @@
 from tapeagents.utils import diff_strings
 
 from ..eval import calculate_accuracy, tape_correct
-from ..tape import GaiaTape
+from ..steps import GaiaTape
 
 
 class TapeDiffGUI:

examples/gaia_agent/steps.py

@@ -4,8 +4,16 @@
 
 from pydantic import Field
 
-from tapeagents.core import LLMOutputParsingFailureAction, Observation, SetNextNode, StopStep, Thought
-from tapeagents.dialog_tape import UserStep
+from tapeagents.core import (
+    LLMOutputParsingFailureAction,
+    Observation,
+    SetNextNode,
+    StopStep,
+    Tape,
+    TapeMetadata,
+    Thought,
+)
+from tapeagents.dialog_tape import DialogContext, UserStep
 from tapeagents.environment import CodeExecutionResult, ExecuteCode
 from tapeagents.steps import (
     ActionExecutionFailure,
@@ -156,3 +164,17 @@ class GaiaAnswer(StopStep):
     SetNextNode,
     UserStep,
 ]
+
+
+class GaiaMetadata(TapeMetadata):
+    task: dict = Field(default_factory=dict)
+    result: Any = None
+    terminated: bool = False
+    attempt_number: int = 0
+    level: int = 0
+    other: dict = Field(default_factory=dict)
+
+
+class GaiaTape(Tape[DialogContext, GaiaStep]):
+    metadata: GaiaMetadata = Field(default_factory=GaiaMetadata)
+    context: DialogContext | None = DialogContext(tools=[])

tests/examples/test_examples.py

@@ -40,7 +40,7 @@
     load_user_input_tapes,
     load_user_reference_tapes,
 )
-from examples.gaia_agent.tape import GaiaTape
+from examples.gaia_agent.steps import GaiaTape
 from examples.llama_agent import LLAMAChatBot
 from examples.optimize.optimize import make_agentic_rag_agent, make_env
 from examples.tape_improver import tape_improver