vision-agent 0.2.202__py3-none-any.whl → 0.2.204__py3-none-any.whl

vision_agent/agent/README.md

@@ -0,0 +1,89 @@
+## V2 Agents
+
+This gives an overview of all the V2 agents, how they communicate, and how human-in-the-loop works.
+
+- `vision_agent_v2` - This is the conversation agent. It can take exactly one action and one response. The actions are fixed JSON actions (meaning it does not execute code but instead returns a JSON and we execute the code on it's behalf). This is so we can control the action arguments as well as pass around the notebook code interpreter.
+- `vision_agent_planner_v2` - This agent is responsible for planning. It can run for N turns, each turn it can take some code action (it can execute it's own python code and has access to `planner_tools`) to test a new potential step in the plan.
+- `vision_agent_coder_v2` - This agent is responsible for the final code. It can call the planner on it's own or it can take in the final `PlanContext` returned by `vision_agent_planner_v2` and use that to write the final code and test it.
+
+### Communication
+The agents communicate through `AgentMessage`'s and return `PlanContext`'s and `CodeContext`'s for the planner and coder agent respectively.
+```
+_______________
+|VisionAgentV2|
+---------------
+       |                       ____________________
+       -----(AgentMessage)---> |VisionAgentCoderV2|
+                               --------------------
+                                         |                        ______________________
+                                         -----(AgentMessage)----> |VisionAgentPlannerV2|
+                                                                  ----------------------
+                               ____________________                         |
+                               |VisionAgentCoderV2| <----(PlanContext)-------
+                               --------------------
+_______________                          |
+|VisionAgentV2|<-----(CodeContext)--------
+---------------                    
+```
+
+#### AgentMessage and Contexts
+`AgentMessage` is a basic chat message but with extended roles. The roles can be your typical `user` and `assistant` but can also be `conversation`, `planner` or `coder` where the come from `VisionAgentV2`, `VisionAgentPlannerV2` and `VisionAgentCoderV2` respectively. `conversation`, `planner` and `coder` are all types of `assistant`. `observation`'s come from responses from executing python code internally by the planner.
+
+The `VisionAgentPlannerV2` returns `PlanContext` which contains the a finalized version of the plan, including instructions and code snippets used during planning. `VisionAgentCoderV2` will then take in that `PlanContext` and return a `CodeContext` which contains the final code and any additional information.
+
+
+#### Callbacks
+If you want to recieve intermediate messages you can use the `update_callback` argument in all the `V2` constructors. This will asynchronously send `AgentMessage`'s to the callback function you provide. You can see an example of how to run this in `examples/chat/app.py`
+
+### Human-in-the-loop
+Human-in-the-loop is a feature that allows the user to interact with the agents at certain points in the conversation. This is handled by using the `interaction` and `interaction_response` roles in the `AgentMessage`. You can enable this feature by passing `hil=True` to the `VisionAgentV2`, currently you can only use human-in-the-loop if you are also using the `update_callback` to collect the messages and pass them back to `VisionAgentV2`.
+
+When the planner agent wants to interact with a human, it will return `InteractionContext` which will propogate back up to `VisionAgentV2` and then to the user. This exits the planner so it can ask for human input. If you collect the messages from `update_callback`, you will see the last `AgentMessage` has a role of `interaction` and the content will include a JSON string surrounded by `<interaction>` tags:
+
+```
+AgentMessage(
+    role="interaction",
+    content="<interaction>{\"prompt\": \"Should I use owl_v2_image or countgd_counting?\"}</interaction>",
+    media=None,
+)
+```
+
+The user can then add an additional `AgentMessage` with the role `interaction_response` and the response they want to give:
+
+```
+AgentMessage(
+    role="interaction_response",
+    content="{\"function_name\": \"owl_v2_image\"}",
+    media=None,
+)
+```
+
+You can see an example of how this works in `examples/chat/chat-app/src/components/ChatSection.tsx` under the `handleSubmit` function.
+
+
+#### Human-in-the-loop Caveats
+One issue with this approach is the planner is running code on a notebook and has access to all the previous executions. This means that for the internal notebook that the agents use, usually named `code_interpreter` from `CodeInterpreterFactor.new_instance`, you cannot close it or restart it. This is an issue because the notebook will close once you exit from the `VisionAgentV2` call.
+
+To fix this you must construct a notebook outside of the chat and ensure it has `non_exiting=True`:
+
+```python
+agent = VisionAgentV2(
+    verbose=True,
+    update_callback=update_callback,
+    hil=True,
+)
+
+code_interpreter = CodeInterpreterFactory.new_instance(non_exiting=True)
+agent.chat(
+    [
+        AgentMessage(
+            role="user",
+            content="Hello",
+            media=None
+        )
+    ],
+    code_interpreter=code_interpreter
+)
+```
+
+An example of this can be seen in `examples/chat/app.py`. Here the `code_intepreter` is constructed outside of the chat and passed in. This is so the notebook does not close when the chat ends or returns when asking for human feedback.

vision_agent/agent/agent.py

@@ -2,7 +2,12 @@ from abc import ABC, abstractmethod
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
 
-from vision_agent.agent.types import AgentMessage, CodeContext, PlanContext
+from vision_agent.agent.types import (
+    AgentMessage,
+    CodeContext,
+    InteractionContext,
+    PlanContext,
+)
 from vision_agent.lmm.types import Message
 from vision_agent.utils.execute import CodeInterpreter
 
@@ -31,7 +36,7 @@ class AgentCoder(Agent):
         chat: List[AgentMessage],
         max_steps: Optional[int] = None,
         code_interpreter: Optional[CodeInterpreter] = None,
-    ) -> CodeContext:
+    ) -> Union[CodeContext, InteractionContext]:
         pass
 
     @abstractmethod
@@ -51,5 +56,5 @@ class AgentPlanner(Agent):
         chat: List[AgentMessage],
         max_steps: Optional[int] = None,
         code_interpreter: Optional[CodeInterpreter] = None,
-    ) -> PlanContext:
+    ) -> Union[PlanContext, InteractionContext]:
         pass

vision_agent/agent/types.py

@@ -17,12 +17,12 @@ class AgentMessage(BaseModel):
     interaction: An interaction between the user and the assistant. For example if the
         assistant wants to ask the user for help on a task, it could send an
         interaction message.
+    interaction_response: The user's response to an interaction message.
     conversation: Messages coming from the conversation agent, this is a type of
         assistant messages.
     planner: Messages coming from the planner agent, this is a type of assistant
         messages.
     coder: Messages coming from the coder agent, this is a type of assistant messages.
-
     """
 
     role: Union[
@@ -30,6 +30,7 @@ class AgentMessage(BaseModel):
         Literal["assistant"],  # planner, coder and conversation are of type assistant
         Literal["observation"],
         Literal["interaction"],
+        Literal["interaction_response"],
         Literal["conversation"],
         Literal["planner"],
         Literal["coder"],
@@ -39,13 +40,37 @@ class AgentMessage(BaseModel):
 
 
 class PlanContext(BaseModel):
+    """PlanContext is a data model that represents the context of a plan.
+
+    plan: A description of the overall plan.
+    instructions: A list of step-by-step instructions.
+    code: Code snippets that were used during planning.
+    """
+
     plan: str
     instructions: List[str]
     code: str
 
 
 class CodeContext(BaseModel):
+    """CodeContext is a data model that represents final code and test cases.
+
+    code: The final code that was written.
+    test: The test cases that were written.
+    success: A boolean value indicating whether the code passed the test cases.
+    test_result: The result of running the test cases.
+    """
+
     code: str
     test: str
     success: bool
     test_result: Execution
+
+
+class InteractionContext(BaseModel):
+    """InteractionContext is a data model that represents the context of an interaction.
+
+    chat: A list of messages exchanged between the user and the assistant.
+    """
+
+    chat: List[AgentMessage]

vision_agent/agent/vision_agent_coder_v2.py

@@ -18,7 +18,12 @@ from vision_agent.agent.agent_utils import (
     print_code,
     strip_function_calls,
 )
-from vision_agent.agent.types import AgentMessage, CodeContext, PlanContext
+from vision_agent.agent.types import (
+    AgentMessage,
+    CodeContext,
+    InteractionContext,
+    PlanContext,
+)
 from vision_agent.agent.vision_agent_coder_prompts_v2 import CODE, FIX_BUG, TEST
 from vision_agent.agent.vision_agent_planner_v2 import VisionAgentPlannerV2
 from vision_agent.lmm import LMM, AnthropicLMM
@@ -257,6 +262,7 @@ class VisionAgentCoderV2(AgentCoder):
         tester: Optional[LMM] = None,
         debugger: Optional[LMM] = None,
         tool_recommender: Optional[Union[str, Sim]] = None,
+        hil: bool = False,
         verbose: bool = False,
         code_sandbox_runtime: Optional[str] = None,
         update_callback: Callable[[Dict[str, Any]], None] = lambda _: None,
@@ -272,6 +278,7 @@ class VisionAgentCoderV2(AgentCoder):
                 None, a default AnthropicLMM will be used.
             debugger (Optional[LMM]): The language model to use for the debugger agent.
             tool_recommender (Optional[Union[str, Sim]]): The tool recommender to use.
+            hil (bool): Whether to use human-in-the-loop mode.
             verbose (bool): Whether to print out debug information.
             code_sandbox_runtime (Optional[str]): The code sandbox runtime to use, can
                 be one of: None, "local" or "e2b". If None, it will read from the
@@ -283,8 +290,11 @@ class VisionAgentCoderV2(AgentCoder):
         self.planner = (
             planner
             if planner is not None
-            else VisionAgentPlannerV2(verbose=verbose, update_callback=update_callback)
+            else VisionAgentPlannerV2(
+                verbose=verbose, update_callback=update_callback, hil=hil
+            )
         )
+
         self.coder = (
             coder
             if coder is not None
@@ -311,6 +321,8 @@ class VisionAgentCoderV2(AgentCoder):
         self.verbose = verbose
         self.code_sandbox_runtime = code_sandbox_runtime
         self.update_callback = update_callback
+        if hasattr(self.planner, "update_callback"):
+            self.planner.update_callback = update_callback
 
     def __call__(
         self,
@@ -331,14 +343,17 @@ class VisionAgentCoderV2(AgentCoder):
         """
 
         input_msg = convert_message_to_agentmessage(input, media)
-        return self.generate_code(input_msg).code
+        code_or_interaction = self.generate_code(input_msg)
+        if isinstance(code_or_interaction, InteractionContext):
+            return code_or_interaction.chat[-1].content
+        return code_or_interaction.code
 
     def generate_code(
         self,
         chat: List[AgentMessage],
         max_steps: Optional[int] = None,
         code_interpreter: Optional[CodeInterpreter] = None,
-    ) -> CodeContext:
+    ) -> Union[CodeContext, InteractionContext]:
         """Generate vision code from a conversation.
 
         Parameters:
@@ -353,6 +368,11 @@ class VisionAgentCoderV2(AgentCoder):
         """
 
         chat = copy.deepcopy(chat)
+        if not chat or chat[-1].role not in {"user", "interaction_response"}:
+            raise ValueError(
+                f"Last chat message must be from the user or interaction_response, got {chat[-1].role}."
+            )
+
         with (
             CodeInterpreterFactory.new_instance(self.code_sandbox_runtime)
             if code_interpreter is None
@@ -362,6 +382,10 @@ class VisionAgentCoderV2(AgentCoder):
             plan_context = self.planner.generate_plan(
                 int_chat, max_steps=max_steps, code_interpreter=code_interpreter
             )
+            # the planner needs an interaction, so return before generating code
+            if isinstance(plan_context, InteractionContext):
+                return plan_context
+
             code_context = self.generate_code_from_plan(
                 orig_chat,
                 plan_context,
@@ -391,6 +415,19 @@ class VisionAgentCoderV2(AgentCoder):
         """
 
         chat = copy.deepcopy(chat)
+        if not chat or chat[-1].role not in {"user", "interaction_response"}:
+            raise ValueError(
+                f"Last chat message must be from the user or interaction_response, got {chat[-1].role}."
+            )
+
+        # we don't need the user_interaction response for generating code since it's
+        # already in the plan context
+        while chat[-1].role != "user":
+            chat.pop()
+
+        if not chat:
+            raise ValueError("Chat must have at least one user message.")
+
         with (
             CodeInterpreterFactory.new_instance(self.code_sandbox_runtime)
             if code_interpreter is None

vision_agent/agent/vision_agent_planner_v2.py

@@ -1,4 +1,5 @@
 import copy
+import json
 import logging
 import time
 from concurrent.futures import ThreadPoolExecutor, as_completed
@@ -21,7 +22,7 @@ from vision_agent.agent.agent_utils import (
     print_code,
     print_table,
 )
-from vision_agent.agent.types import AgentMessage, PlanContext
+from vision_agent.agent.types import AgentMessage, InteractionContext, PlanContext
 from vision_agent.agent.vision_agent_planner_prompts_v2 import (
     CRITIQUE_PLAN,
     EXAMPLE_PLAN1,
@@ -32,6 +33,7 @@ from vision_agent.agent.vision_agent_planner_prompts_v2 import (
     PLAN,
 )
 from vision_agent.lmm import LMM, AnthropicLMM, Message
+from vision_agent.tools.planner_tools import check_function_call, get_tool_documentation
 from vision_agent.utils.execute import (
     CodeInterpreter,
     CodeInterpreterFactory,
@@ -210,7 +212,7 @@ def execute_code_action(
     obs = execution.text(include_results=False).strip()
     if verbose:
         _CONSOLE.print(
-            f"[bold cyan]Code Execution Output ({end - start:.2f} sec):[/bold cyan] [yellow]{escape(obs)}[/yellow]"
+            f"[bold cyan]Code Execution Output ({end - start:.2f}s):[/bold cyan] [yellow]{escape(obs)}[/yellow]"
         )
 
     count = 1
@@ -247,6 +249,31 @@ def find_and_replace_code(response: str, code: str) -> str:
     return response[:code_start] + code + response[code_end:]
 
 
+def create_hil_response(
+    execution: Execution,
+) -> AgentMessage:
+    content = []
+    for result in execution.results:
+        for format in result.formats():
+            if format == "json":
+                data = result.json
+                try:
+                    data_dict: Dict[str, Any] = dict(data)  # type: ignore
+                    if (
+                        "request" in data_dict
+                        and "function_name" in data_dict["request"]
+                    ):
+                        content.append(data_dict)
+                except Exception:
+                    continue
+
+    return AgentMessage(
+        role="interaction",
+        content="<interaction>" + json.dumps(content) + "</interaction>",
+        media=None,
+    )
+
+
 def maybe_run_code(
     code: Optional[str],
     response: str,
@@ -254,6 +281,7 @@ def maybe_run_code(
     media_list: List[Union[str, Path]],
     model: LMM,
     code_interpreter: CodeInterpreter,
+    hil: bool = False,
     verbose: bool = False,
 ) -> List[AgentMessage]:
     return_chat: List[AgentMessage] = []
@@ -270,6 +298,11 @@ def maybe_run_code(
             AgentMessage(role="planner", content=fixed_response, media=None)
         )
 
+        # if we are running human-in-the-loop mode, send back an interaction message
+        # make sure we return code from planner and the hil response
+        if check_function_call(code, "get_tool_for_task") and hil:
+            return return_chat + [create_hil_response(execution)]
+
         media_data = capture_media_from_exec(execution)
         int_chat_elt = AgentMessage(role="observation", content=obs, media=None)
         if media_list:
@@ -285,6 +318,10 @@ def create_finalize_plan(
     model: LMM,
     verbose: bool = False,
 ) -> Tuple[List[AgentMessage], PlanContext]:
+    # if we're in the middle of an interaction, don't finalize the plan
+    if chat[-1].role == "interaction":
+        return [], PlanContext(plan="", instructions=[], code="")
+
     prompt = FINALIZE_PLAN.format(
         planning=get_planning(chat),
         excluded_tools=str([t.__name__ for t in pt.PLANNER_TOOLS]),
@@ -318,6 +355,27 @@ def get_steps(chat: List[AgentMessage], max_steps: int) -> int:
     return max_steps
 
 
+def replace_interaction_with_obs(chat: List[AgentMessage]) -> List[AgentMessage]:
+    chat = copy.deepcopy(chat)
+    new_chat = []
+
+    for i, chat_i in enumerate(chat):
+        if chat_i.role == "interaction" and (
+            i < len(chat) and chat[i + 1].role == "interaction_response"
+        ):
+            try:
+                response = json.loads(chat[i + 1].content)
+                function_name = response["function_name"]
+                tool_doc = get_tool_documentation(function_name)
+                new_chat.append(AgentMessage(role="observation", content=tool_doc))
+            except json.JSONDecodeError:
+                raise ValueError(f"Invalid JSON in interaction response: {chat_i}")
+        else:
+            new_chat.append(chat_i)
+
+    return new_chat
+
+
 class VisionAgentPlannerV2(AgentPlanner):
     """VisionAgentPlannerV2 is a class that generates a plan to solve a vision task."""
 
@@ -328,6 +386,7 @@ class VisionAgentPlannerV2(AgentPlanner):
         max_steps: int = 10,
         use_multi_trial_planning: bool = False,
         critique_steps: int = 11,
+        hil: bool = False,
         verbose: bool = False,
         code_sandbox_runtime: Optional[str] = None,
         update_callback: Callable[[Dict[str, Any]], None] = lambda _: None,
@@ -343,6 +402,7 @@ class VisionAgentPlannerV2(AgentPlanner):
             use_multi_trial_planning (bool): Whether to use multi-trial planning.
             critique_steps (int): The number of steps between critiques. If critic steps
                 is larger than max_steps no critiques will be made.
+            hil (bool): Whether to use human-in-the-loop mode.
             verbose (bool): Whether to print out debug information.
             code_sandbox_runtime (Optional[str]): The code sandbox runtime to use, can
                 be one of: None, "local" or "e2b". If None, it will read from the
@@ -365,6 +425,11 @@ class VisionAgentPlannerV2(AgentPlanner):
         self.use_multi_trial_planning = use_multi_trial_planning
         self.critique_steps = critique_steps
 
+        self.hil = hil
+        if self.hil:
+            DefaultPlanningImports.imports.append(
+                "from vision_agent.tools.planner_tools import get_tool_for_task_human_reviewer as get_tool_for_task"
+            )
         self.verbose = verbose
         self.code_sandbox_runtime = code_sandbox_runtime
         self.update_callback = update_callback
@@ -388,15 +453,17 @@ class VisionAgentPlannerV2(AgentPlanner):
         """
 
         input_msg = convert_message_to_agentmessage(input, media)
-        plan = self.generate_plan(input_msg)
-        return plan.plan
+        plan_or_interaction = self.generate_plan(input_msg)
+        if isinstance(plan_or_interaction, InteractionContext):
+            return plan_or_interaction.chat[-1].content
+        return plan_or_interaction.plan
 
     def generate_plan(
         self,
         chat: List[AgentMessage],
         max_steps: Optional[int] = None,
         code_interpreter: Optional[CodeInterpreter] = None,
-    ) -> PlanContext:
+    ) -> Union[PlanContext, InteractionContext]:
         """Generate a plan to solve a vision task.
 
         Parameters:
@@ -409,24 +476,30 @@ class VisionAgentPlannerV2(AgentPlanner):
                 needed to solve the task.
         """
 
-        if not chat:
-            raise ValueError("Chat cannot be empty")
+        if not chat or chat[-1].role not in {"user", "interaction_response"}:
+            raise ValueError(
+                f"Last chat message must be from the user or interaction_response, got {chat[-1].role}."
+            )
 
         chat = copy.deepcopy(chat)
-        code_interpreter = code_interpreter or CodeInterpreterFactory.new_instance(
-            self.code_sandbox_runtime
-        )
         max_steps = max_steps or self.max_steps
 
-        with code_interpreter:
+        with (
+            CodeInterpreterFactory.new_instance(self.code_sandbox_runtime)
+            if code_interpreter is None
+            else code_interpreter
+        ) as code_interpreter:
             critque_steps = 1
             finished = False
+            interaction = False
             int_chat, _, media_list = add_media_to_chat(chat, code_interpreter)
+            int_chat = replace_interaction_with_obs(int_chat)
 
             step = get_steps(int_chat, max_steps)
             if "<count>" not in int_chat[-1].content and step == max_steps:
                 int_chat[-1].content += f"\n<count>{step}</count>\n"
-            while step > 0 and not finished:
+
+            while step > 0 and not finished and not interaction:
                 if self.use_multi_trial_planning:
                     response = run_multi_trial_planning(
                         int_chat, media_list, self.planner
@@ -456,8 +529,10 @@ class VisionAgentPlannerV2(AgentPlanner):
                     media_list,
                     self.planner,
                     code_interpreter,
-                    self.verbose,
+                    hil=self.hil,
+                    verbose=self.verbose,
                 )
+                interaction = updated_chat[-1].role == "interaction"
 
                 if critque_steps % self.critique_steps == 0:
                     critique = run_critic(int_chat, media_list, self.critic)
@@ -478,14 +553,18 @@ class VisionAgentPlannerV2(AgentPlanner):
                 for chat_elt in updated_chat:
                     self.update_callback(chat_elt.model_dump())
 
-            updated_chat, plan_context = create_finalize_plan(
-                int_chat, self.planner, self.verbose
-            )
-            int_chat.extend(updated_chat)
-            for chat_elt in updated_chat:
-                self.update_callback(chat_elt.model_dump())
+            context: Union[PlanContext, InteractionContext]
+            if interaction:
+                context = InteractionContext(chat=int_chat)
+            else:
+                updated_chat, context = create_finalize_plan(
+                    int_chat, self.planner, self.verbose
+                )
+                int_chat.extend(updated_chat)
+                for chat_elt in updated_chat:
+                    self.update_callback(chat_elt.model_dump())
 
-        return plan_context
+        return context
 
     def log_progress(self, data: Dict[str, Any]) -> None:
         pass

vision_agent/agent/vision_agent_prompts.py

@@ -280,7 +280,7 @@ def main():
 if __name__ == "__main__":
     main()
 '''
-edit_code_artifact(artifacts, 'streamlit_app.py', CODE, 0, 0)
+edit_code_artifact(artifacts, 'streamlit_app.py', 0, 0, CODE)
 </execute_python>
 
 OBSERVATION:

vision_agent/agent/vision_agent_v2.py

@@ -1,4 +1,5 @@
 import copy
+import json
 from pathlib import Path
 from typing import Any, Callable, Dict, List, Optional, Union, cast
 
@@ -8,7 +9,12 @@ from vision_agent.agent.agent_utils import (
     convert_message_to_agentmessage,
     extract_tag,
 )
-from vision_agent.agent.types import AgentMessage, PlanContext
+from vision_agent.agent.types import (
+    AgentMessage,
+    CodeContext,
+    InteractionContext,
+    PlanContext,
+)
 from vision_agent.agent.vision_agent_coder_v2 import format_code_context
 from vision_agent.agent.vision_agent_prompts_v2 import CONVERSATION
 from vision_agent.lmm import LMM, AnthropicLMM
@@ -39,10 +45,24 @@ def run_conversation(agent: LMM, chat: List[AgentMessage]) -> str:
     return cast(str, response)
 
 
+def check_for_interaction(chat: List[AgentMessage]) -> bool:
+    return (
+        len(chat) > 2
+        and chat[-2].role == "interaction"
+        and chat[-1].role == "interaction_response"
+    )
+
+
 def extract_conversation_for_generate_code(
     chat: List[AgentMessage],
 ) -> List[AgentMessage]:
     chat = copy.deepcopy(chat)
+
+    # if we are in the middle of an interaction, return all the intermediate planning
+    # steps
+    if check_for_interaction(chat):
+        return chat
+
     extracted_chat = []
     for chat_i in chat:
         if chat_i.role == "user":
@@ -66,13 +86,20 @@ def maybe_run_action(
         # to the outside user via it's update_callback, but we don't necessarily have
         # access to that update_callback here, so we re-create the message using
         # format_code_context.
-        code_context = coder.generate_code(
-            extracted_chat, code_interpreter=code_interpreter
-        )
-        return [
-            AgentMessage(role="coder", content=format_code_context(code_context)),
-            AgentMessage(role="observation", content=code_context.test_result.text()),
-        ]
+        context = coder.generate_code(extracted_chat, code_interpreter=code_interpreter)
+
+        if isinstance(context, CodeContext):
+            return [
+                AgentMessage(role="coder", content=format_code_context(context)),
+                AgentMessage(role="observation", content=context.test_result.text()),
+            ]
+        elif isinstance(context, InteractionContext):
+            return [
+                AgentMessage(
+                    role="interaction",
+                    content=json.dumps([elt.model_dump() for elt in context.chat]),
+                )
+            ]
     elif action == "edit_code":
         extracted_chat = extract_conversation_for_generate_code(chat)
         plan_context = PlanContext(
@@ -80,12 +107,12 @@ def maybe_run_action(
             instructions=[],
             code="",
         )
-        code_context = coder.generate_code_from_plan(
+        context = coder.generate_code_from_plan(
             extracted_chat, plan_context, code_interpreter=code_interpreter
         )
         return [
-            AgentMessage(role="coder", content=format_code_context(code_context)),
-            AgentMessage(role="observation", content=code_context.test_result.text()),
+            AgentMessage(role="coder", content=format_code_context(context)),
+            AgentMessage(role="observation", content=context.test_result.text()),
         ]
     elif action == "view_image":
         pass
@@ -102,6 +129,7 @@ class VisionAgentV2(Agent):
         self,
         agent: Optional[LMM] = None,
         coder: Optional[AgentCoder] = None,
+        hil: bool = False,
         verbose: bool = False,
         code_sandbox_runtime: Optional[str] = None,
         update_callback: Callable[[Dict[str, Any]], None] = lambda x: None,
@@ -113,6 +141,7 @@ class VisionAgentV2(Agent):
                 default AnthropicLMM will be used.
             coder (Optional[AgentCoder]): The coder agent to use for generating vision
                 code. If None, a default VisionAgentCoderV2 will be used.
+            hil (bool): Whether to use human-in-the-loop mode.
             verbose (bool): Whether to print out debug information.
             code_sandbox_runtime (Optional[str]): The code sandbox runtime to use, can
                 be one of: None, "local" or "e2b". If None, it will read from the
@@ -132,7 +161,9 @@ class VisionAgentV2(Agent):
         self.coder = (
             coder
             if coder is not None
-            else VisionAgentCoderV2(verbose=verbose, update_callback=update_callback)
+            else VisionAgentCoderV2(
+                verbose=verbose, update_callback=update_callback, hil=hil
+            )
         )
 
         self.verbose = verbose
@@ -169,6 +200,7 @@ class VisionAgentV2(Agent):
     def chat(
         self,
         chat: List[AgentMessage],
+        code_interpreter: Optional[CodeInterpreter] = None,
     ) -> List[AgentMessage]:
         """Conversational interface to the agent. This is the main method to use to
         interact with the agent. It takes in a list of messages and returns the agent's
@@ -177,28 +209,47 @@ class VisionAgentV2(Agent):
         Parameters:
             chat (List[AgentMessage]): The input to the agent. This should be a list of
                 AgentMessage objects.
+            code_interpreter (Optional[CodeInterpreter]): The code interpreter to use.
 
         Returns:
             List[AgentMessage]: The agent's response as a list of AgentMessage objects.
         """
 
+        chat = copy.deepcopy(chat)
+        if not chat or chat[-1].role not in {"user", "interaction_response"}:
+            raise ValueError(
+                f"Last chat message must be from the user or interaction_response, got {chat[-1].role}."
+            )
+
         return_chat = []
-        with CodeInterpreterFactory.new_instance(
-            self.code_sandbox_runtime
+        with (
+            CodeInterpreterFactory.new_instance(self.code_sandbox_runtime)
+            if code_interpreter is None
+            else code_interpreter
         ) as code_interpreter:
             int_chat, _, _ = add_media_to_chat(chat, code_interpreter)
-            response_context = run_conversation(self.agent, int_chat)
-            return_chat.append(
-                AgentMessage(role="conversation", content=response_context)
-            )
-            self.update_callback(return_chat[-1].model_dump())
 
-            action = extract_tag(response_context, "action")
+            # if we had an interaction and then recieved an observation from the user
+            # go back into the same action to finish it.
+            action = None
+            if check_for_interaction(int_chat):
+                action = "generate_or_edit_vision_code"
+            else:
+                response_context = run_conversation(self.agent, int_chat)
+                return_chat.append(
+                    AgentMessage(role="conversation", content=response_context)
+                )
+                self.update_callback(return_chat[-1].model_dump())
+                action = extract_tag(response_context, "action")
 
             updated_chat = maybe_run_action(
                 self.coder, action, int_chat, code_interpreter=code_interpreter
             )
-            if updated_chat is not None:
+
+            # return an interaction early to get users feedback
+            if updated_chat is not None and updated_chat[-1].role == "interaction":
+                return_chat.extend(updated_chat)
+            elif updated_chat is not None and updated_chat[-1].role != "interaction":
                 # do not append updated_chat to return_chat becuase the observation
                 # from running the action will have already been added via the callbacks
                 obs_response_context = run_conversation(

vision_agent/tools/planner_tools.py

@@ -3,7 +3,9 @@ import shutil
 import tempfile
 from typing import Any, Callable, Dict, List, Optional, Tuple, cast
 
+import libcst as cst
 import numpy as np
+from IPython.display import display
 from PIL import Image
 
 import vision_agent.tools as T
@@ -21,8 +23,13 @@ from vision_agent.agent.vision_agent_planner_prompts_v2 import (
     TEST_TOOLS_EXAMPLE1,
     TEST_TOOLS_EXAMPLE2,
 )
-from vision_agent.lmm import AnthropicLMM
-from vision_agent.utils.execute import CodeInterpreterFactory
+from vision_agent.lmm import LMM, AnthropicLMM
+from vision_agent.utils.execute import (
+    CodeInterpreter,
+    CodeInterpreterFactory,
+    Execution,
+    MimeType,
+)
 from vision_agent.utils.image_utils import convert_to_b64
 from vision_agent.utils.sim import load_cached_sim
 
@@ -33,6 +40,16 @@ _LOGGER = logging.getLogger(__name__)
 EXAMPLES = f"\n{TEST_TOOLS_EXAMPLE1}\n{TEST_TOOLS_EXAMPLE2}\n"
 
 
+def format_tool_output(tool_thoughts: str, tool_docstring: str) -> str:
+    return_str = "[get_tool_for_task output]\n"
+    if tool_thoughts.strip() != "":
+        return_str += f"{tool_thoughts}\n\n"
+    return_str += (
+        f"Tool Documentation:\n{tool_docstring}\n[end of get_tool_for_task output]\n"
+    )
+    return return_str
+
+
 def extract_tool_info(
     tool_choice_context: Dict[str, Any]
 ) -> Tuple[Optional[Callable], str, str, str]:
@@ -46,6 +63,70 @@ def extract_tool_info(
     return tool, tool_thoughts, tool_docstring, ""
 
 
+def run_tool_testing(
+    task: str,
+    image_paths: List[str],
+    lmm: LMM,
+    exclude_tools: Optional[List[str]],
+    code_interpreter: CodeInterpreter,
+) -> tuple[str, str, Execution]:
+    """Helper function to generate and run tool testing code."""
+    query = lmm.generate(CATEGORIZE_TOOL_REQUEST.format(task=task))
+    category = extract_tag(query, "category")  # type: ignore
+    if category is None:
+        category = task
+    else:
+        category = (
+            f"I need models from the {category.strip()} category of tools. {task}"
+        )
+
+    tool_docs = TOOL_RECOMMENDER.top_k(category, k=10, thresh=0.2)
+    if exclude_tools is not None and len(exclude_tools) > 0:
+        cleaned_tool_docs = []
+        for tool_doc in tool_docs:
+            if not tool_doc["name"] in exclude_tools:
+                cleaned_tool_docs.append(tool_doc)
+        tool_docs = cleaned_tool_docs
+    tool_docs_str = "\n".join([e["doc"] for e in tool_docs])
+
+    prompt = TEST_TOOLS.format(
+        tool_docs=tool_docs_str,
+        previous_attempts="",
+        user_request=task,
+        examples=EXAMPLES,
+        media=str(image_paths),
+    )
+
+    response = lmm.generate(prompt, media=image_paths)
+    code = extract_tag(response, "code")  # type: ignore
+    if code is None:
+        raise ValueError(f"Could not extract code from response: {response}")
+    tool_output = code_interpreter.exec_isolation(DefaultImports.prepend_imports(code))
+    tool_output_str = tool_output.text(include_results=False).strip()
+
+    count = 1
+    while (
+        not tool_output.success
+        or (len(tool_output.logs.stdout) == 0 and len(tool_output.logs.stderr) == 0)
+    ) and count <= 3:
+        if tool_output_str.strip() == "":
+            tool_output_str = "EMPTY"
+        prompt = TEST_TOOLS.format(
+            tool_docs=tool_docs_str,
+            previous_attempts=f"<code>\n{code}\n</code>\nTOOL OUTPUT\n{tool_output_str}",
+            user_request=task,
+            examples=EXAMPLES,
+            media=str(image_paths),
+        )
+        code = extract_code(lmm.generate(prompt, media=image_paths))  # type: ignore
+        tool_output = code_interpreter.exec_isolation(
+            DefaultImports.prepend_imports(code)
+        )
+        tool_output_str = tool_output.text(include_results=False).strip()
+
+    return code, tool_docs_str, tool_output
+
+
 def get_tool_for_task(
     task: str, images: List[np.ndarray], exclude_tools: Optional[List[str]] = None
 ) -> None:
@@ -90,61 +171,11 @@ def get_tool_for_task(
             Image.fromarray(image).save(image_path)
             image_paths.append(image_path)
 
-        query = lmm.generate(CATEGORIZE_TOOL_REQUEST.format(task=task))
-        category = extract_tag(query, "category")  # type: ignore
-        if category is None:
-            category = task
-        else:
-            category = (
-                f"I need models from the {category.strip()} category of tools. {task}"
-            )
-
-        tool_docs = TOOL_RECOMMENDER.top_k(category, k=10, thresh=0.2)
-        if exclude_tools is not None and len(exclude_tools) > 0:
-            cleaned_tool_docs = []
-            for tool_doc in tool_docs:
-                if not tool_doc["name"] in exclude_tools:
-                    cleaned_tool_docs.append(tool_doc)
-            tool_docs = cleaned_tool_docs
-        tool_docs_str = "\n".join([e["doc"] for e in tool_docs])
-
-        prompt = TEST_TOOLS.format(
-            tool_docs=tool_docs_str,
-            previous_attempts="",
-            user_request=task,
-            examples=EXAMPLES,
-            media=str(image_paths),
-        )
-
-        response = lmm.generate(prompt, media=image_paths)
-        code = extract_tag(response, "code")  # type: ignore
-        if code is None:
-            raise ValueError(f"Could not extract code from response: {response}")
-        tool_output = code_interpreter.exec_isolation(
-            DefaultImports.prepend_imports(code)
+        code, tool_docs_str, tool_output = run_tool_testing(
+            task, image_paths, lmm, exclude_tools, code_interpreter
         )
         tool_output_str = tool_output.text(include_results=False).strip()
 
-        count = 1
-        while (
-            not tool_output.success
-            or (len(tool_output.logs.stdout) == 0 and len(tool_output.logs.stderr) == 0)
-        ) and count <= 3:
-            if tool_output_str.strip() == "":
-                tool_output_str = "EMPTY"
-            prompt = TEST_TOOLS.format(
-                tool_docs=tool_docs_str,
-                previous_attempts=f"<code>\n{code}\n</code>\nTOOL OUTPUT\n{tool_output_str}",
-                user_request=task,
-                examples=EXAMPLES,
-                media=str(image_paths),
-            )
-            code = extract_code(lmm.generate(prompt, media=image_paths))  # type: ignore
-            tool_output = code_interpreter.exec_isolation(
-                DefaultImports.prepend_imports(code)
-            )
-            tool_output_str = tool_output.text(include_results=False).strip()
-
         error_message = ""
         prompt = PICK_TOOL.format(
             tool_docs=tool_docs_str,
@@ -178,9 +209,87 @@ def get_tool_for_task(
         except Exception as e:
             _LOGGER.error(f"Error removing temp directory: {e}")
 
-    print(
-        f"[get_tool_for_task output]\n{tool_thoughts}\n\nTool Documentation:\n{tool_docstring}\n[end of get_tool_for_task output]\n"
-    )
+    print(format_tool_output(tool_thoughts, tool_docstring))
+
+
+def get_tool_documentation(tool_name: str) -> str:
+    # use same format as get_tool_for_task
+    tool_doc = T.TOOLS_DF[T.TOOLS_DF["name"] == tool_name]["doc"].values[0]
+    return format_tool_output("", tool_doc)
+
+
+def get_tool_for_task_human_reviewer(
+    task: str, images: List[np.ndarray], exclude_tools: Optional[List[str]] = None
+) -> None:
+    # NOTE: this should be the same documentation as get_tool_for_task
+    """Given a task and one or more images this function will find a tool to accomplish
+    the jobs. It prints the tool documentation and thoughts on why it chose the tool.
+
+    It can produce tools for the following types of tasks:
+        - Object detection and counting
+        - Classification
+        - Segmentation
+        - OCR
+        - VQA
+        - Depth and pose estimation
+        - Video object tracking
+
+    Wait until the documentation is printed to use the function so you know what the
+    input and output signatures are.
+
+    Parameters:
+        task: str: The task to accomplish.
+        images: List[np.ndarray]: The images to use for the task.
+        exclude_tools: Optional[List[str]]: A list of tool names to exclude from the
+            recommendations. This is helpful if you are calling get_tool_for_task twice
+            and do not want the same tool recommended.
+
+    Returns:
+        The tool to use for the task is printed to stdout
+
+    Examples
+    --------
+        >>> get_tool_for_task("Give me an OCR model that can find 'hot chocolate' in the image", [image])
+    """
+    lmm = AnthropicLMM()
+
+    with (
+        tempfile.TemporaryDirectory() as tmpdirname,
+        CodeInterpreterFactory.new_instance() as code_interpreter,
+    ):
+        image_paths = []
+        for i, image in enumerate(images[:3]):
+            image_path = f"{tmpdirname}/image_{i}.png"
+            Image.fromarray(image).save(image_path)
+            image_paths.append(image_path)
+
+        _, _, tool_output = run_tool_testing(
+            task, image_paths, lmm, exclude_tools, code_interpreter
+        )
+
+        # need to re-display results for the outer notebook to see them
+        for result in tool_output.results:
+            if "json" in result.formats():
+                display({MimeType.APPLICATION_JSON: result.json}, raw=True)
+
+
+def check_function_call(code: str, function_name: str) -> bool:
+    class FunctionCallVisitor(cst.CSTVisitor):
+        def __init__(self) -> None:
+            self.function_name = function_name
+            self.function_called = False
+
+        def visit_Call(self, node: cst.Call) -> None:
+            if (
+                isinstance(node.func, cst.Name)
+                and node.func.value == self.function_name
+            ):
+                self.function_called = True
+
+    tree = cst.parse_module(code)
+    visitor = FunctionCallVisitor()
+    tree.visit(visitor)
+    return visitor.function_called
 
 
 def finalize_plan(user_request: str, chain_of_thoughts: str) -> str:

vision_agent/tools/tool_utils.py

@@ -213,6 +213,8 @@ def _call_post(
     files_in_b64 = None
     if files:
         files_in_b64 = [(file[0], b64encode(file[1]).decode("utf-8")) for file in files]
+
+    tool_call_trace = None
     try:
         if files is not None:
             response = session.post(url, data=payload, files=files)
@@ -250,9 +252,10 @@ def _call_post(
         tool_call_trace.response = result
         return result
     finally:
-        trace = tool_call_trace.model_dump()
-        trace["type"] = "tool_call"
-        display({MimeType.APPLICATION_JSON: trace}, raw=True)
+        if tool_call_trace is not None:
+            trace = tool_call_trace.model_dump()
+            trace["type"] = "tool_call"
+            display({MimeType.APPLICATION_JSON: trace}, raw=True)
 
 
 def filter_bboxes_by_threshold(

vision_agent/utils/execute.py

@@ -398,17 +398,20 @@ class CodeInterpreter(abc.ABC):
         self,
         timeout: int,
         remote_path: Optional[Union[str, Path]] = None,
+        non_exiting: bool = False,
         *args: Any,
         **kwargs: Any,
     ) -> None:
         self.timeout = timeout
         self.remote_path = Path(remote_path if remote_path is not None else WORKSPACE)
+        self.non_exiting = non_exiting
 
     def __enter__(self) -> Self:
         return self
 
     def __exit__(self, *exc_info: Any) -> None:
-        self.close()
+        if not self.non_exiting:
+            self.close()
 
     def close(self, *args: Any, **kwargs: Any) -> None:
         raise NotImplementedError()
@@ -571,8 +574,9 @@ class LocalCodeInterpreter(CodeInterpreter):
         self,
         timeout: int = _SESSION_TIMEOUT,
         remote_path: Optional[Union[str, Path]] = None,
+        non_exiting: bool = False,
     ) -> None:
-        super().__init__(timeout=timeout)
+        super().__init__(timeout=timeout, non_exiting=non_exiting)
         self.nb = nbformat.v4.new_notebook()
         # Set the notebook execution path to the remote path
         self.remote_path = Path(remote_path if remote_path is not None else WORKSPACE)
@@ -692,6 +696,7 @@ class CodeInterpreterFactory:
     def new_instance(
         code_sandbox_runtime: Optional[str] = None,
         remote_path: Optional[Union[str, Path]] = None,
+        non_exiting: bool = False,
     ) -> CodeInterpreter:
         if not code_sandbox_runtime:
             code_sandbox_runtime = os.getenv("CODE_SANDBOX_RUNTIME", "local")
@@ -702,7 +707,9 @@ class CodeInterpreterFactory:
             )
         elif code_sandbox_runtime == "local":
             instance = LocalCodeInterpreter(
-                timeout=_SESSION_TIMEOUT, remote_path=remote_path
+                timeout=_SESSION_TIMEOUT,
+                remote_path=remote_path,
+                non_exiting=non_exiting,
             )
         else:
             raise ValueError(

{vision_agent-0.2.202.dist-info → vision_agent-0.2.204.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: vision-agent
-Version: 0.2.202
+Version: 0.2.204
 Summary: Toolset for Vision Agent
 Author: Landing AI
 Author-email: dev@landing.ai

{vision_agent-0.2.202.dist-info → vision_agent-0.2.204.dist-info}/RECORD

@@ -1,22 +1,23 @@
 vision_agent/.sim_tools/df.csv,sha256=0fmLwTDjnRTiqYwamTOdCPjruE6wZz0AVrONIPTHxZY,34086
 vision_agent/.sim_tools/embs.npy,sha256=xF8Cg7Xd09QCTySj831aL1O2_0kRNaaH8XRJIRjgWzQ,356480
 vision_agent/__init__.py,sha256=EAb4-f9iyuEYkBrX4ag1syM8Syx8118_t0R6_C34M9w,57
+vision_agent/agent/README.md,sha256=Q4w7FWw38qaWosQYAZ7NqWx8Q5XzuWrlv7nLhjUd1-8,5527
 vision_agent/agent/__init__.py,sha256=M8CffavdIh8Zh-skznLHIaQkYGCGK7vk4dq1FaVkbs4,617
-vision_agent/agent/agent.py,sha256=sf8JcA3LNy_4GaS_gQb2Q-PXkl4dBuGh-7raI9KAtZo,1470
+vision_agent/agent/agent.py,sha256=_1tHWAs7Jm5tqDzEcPfCRvJV3uRRveyh4n9_9pd6I1w,1565
 vision_agent/agent/agent_utils.py,sha256=NmrqjhSb6fpnrB8XGWtaywZjr9n89otusOZpcbWLf9k,13534
-vision_agent/agent/types.py,sha256=aAd_ez1-NQh04k27cmywyOV2uA_vWWYE-Ok7zq_JoAk,1532
+vision_agent/agent/types.py,sha256=DkFm3VMMrKlhYyfxEmZx4keppD72Ov3wmLCbM2J2o10,2437
 vision_agent/agent/vision_agent.py,sha256=I75bEU-os9Lf9OSICKfvQ_H_ftg-zOwgTwWnu41oIdo,23555
 vision_agent/agent/vision_agent_coder.py,sha256=ANwUuCO4JpTYJs4s6ynSRFcdjZFUVuSoSfcqp8ZQDDQ,27451
 vision_agent/agent/vision_agent_coder_prompts.py,sha256=gPLVXQMNSzYnQYpNm0wlH_5FPkOTaFDV24bqzK3jQ40,12221
 vision_agent/agent/vision_agent_coder_prompts_v2.py,sha256=9v5HwbNidSzYUEFl6ZMniWWOmyLITM_moWLtKVaTen8,4845
-vision_agent/agent/vision_agent_coder_v2.py,sha256=SVIJC0N5TBgq9z-F99UebLimRuQuAe_HHvTFupBzVfo,14715
+vision_agent/agent/vision_agent_coder_v2.py,sha256=nXbMsCLpKxTEi075ZE932227tW-lEJ6a_4PnNS9fFY0,16112
 vision_agent/agent/vision_agent_planner.py,sha256=KWMA7XemcSmc_jn-MwdWz9wnKDtj-sYQ9tINi70_OoU,18583
 vision_agent/agent/vision_agent_planner_prompts.py,sha256=Y3jz9HRf8fz9NLUseN7cTgZqewP0RazxR7vw1sPhcn0,6691
 vision_agent/agent/vision_agent_planner_prompts_v2.py,sha256=Tzon3h5iZdHJglesk8GVS-2myNf5-fhf7HUbkpZWHQk,33143
-vision_agent/agent/vision_agent_planner_v2.py,sha256=mxQxD_B8sKYharh8e7W0uc1tN11YCztyLowc83seScc,17023
-vision_agent/agent/vision_agent_prompts.py,sha256=PENFd8VM_vHKxeZPiotVM1RBVW9NrXimKbpvI1UteKI,13772
+vision_agent/agent/vision_agent_planner_v2.py,sha256=3sXW4A-GZ5Bg2rGheuIYspAu_N2e00Sii1f_1HJS934,20255
+vision_agent/agent/vision_agent_prompts.py,sha256=NtGdCfzzilCRtscKALC9FK55d1h4CBpMnbhLzg0PYlc,13772
 vision_agent/agent/vision_agent_prompts_v2.py,sha256=-vCWat-ARlCOOOeIDIFhg-kcwRRwjTXYEwsvvqPeaCs,1972
-vision_agent/agent/vision_agent_v2.py,sha256=Cudp_ZZBI9rDwMjIYlvY4jzh_srsulYgfRWZLo4_2TQ,8366
+vision_agent/agent/vision_agent_v2.py,sha256=6gGVV3FlL4NLzHRpjMqMz-fEP6f_JhwwOjUKczZ3TPA,10231
 vision_agent/clients/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 vision_agent/clients/http.py,sha256=k883i6M_4nl7zwwHSI-yP5sAgQZIDPM1nrKD6YFJ3Xs,2009
 vision_agent/clients/landing_public_api.py,sha256=lU2ev6E8NICmR8DMUljuGcVFy5VNJQ4WQkWC8WnnJEc,1503
@@ -27,19 +28,19 @@ vision_agent/lmm/lmm.py,sha256=x_nIyDNDZwq4-pfjnJTmcyyJZ2_B7TjkA5jZp88YVO8,17103
 vision_agent/lmm/types.py,sha256=ZEXR_ptBL0ZwDMTDYkgxUCmSZFmBYPQd2jreNzr_8UY,221
 vision_agent/tools/__init__.py,sha256=xuNt5e4syQH28Vr6EdjLmO9ni9i00yav9yqcPMUx1oo,2878
 vision_agent/tools/meta_tools.py,sha256=TPeS7QWnc_PmmU_ndiDT03dXbQ5yDSP33E7U8cSj7Ls,28660
-vision_agent/tools/planner_tools.py,sha256=FROahw_6Taqvytv6pOjCHUEypOfjsi_f8Vo1c5vz6Mw,8823
+vision_agent/tools/planner_tools.py,sha256=amaycM_REQ4cwZCaKSyIWr-6ExqlHGEVs3PuIXjf-9M,12373
 vision_agent/tools/prompts.py,sha256=V1z4YJLXZuUl_iZ5rY0M5hHc_2tmMEUKr0WocXKGt4E,1430
-vision_agent/tools/tool_utils.py,sha256=GDGOmBCo4UfYz-DJ-olREJHPsqs5mzHu0YXiAnpNE8E,10179
+vision_agent/tools/tool_utils.py,sha256=AT7rMcpKwZgIErfgfSvHS0gmtvd8KMHJoHnu5aMlgO0,10259
 vision_agent/tools/tools.py,sha256=vavzmDuIBHI-g13RMDnr9NALfWpiIvJWkXhD0pnhCuk,87576
 vision_agent/tools/tools_types.py,sha256=8hYf2OZhI58gvf65KGaeGkt4EQ56nwLFqIQDPHioOBc,2339
 vision_agent/utils/__init__.py,sha256=7fMgbZiEwbNS0fBOS_hJI5PuEYBblw36zLi_UjUzvj4,244
 vision_agent/utils/exceptions.py,sha256=booSPSuoULF7OXRr_YbC4dtKt6gM_HyiFQHBuaW86C4,2052
-vision_agent/utils/execute.py,sha256=b3AA1G16Ixwlgd-kke13brKclxh5nJXQTrk25oj1W3o,28027
+vision_agent/utils/execute.py,sha256=Qr-2coChvv0oa0ObutuO1j0zo2xyq0_YlsSsSn50Cnw,28291
 vision_agent/utils/image_utils.py,sha256=rRWcxKggPXIRXIY_XT9rZt30ECDRq8zq7FDeXRDqQWs,11679
 vision_agent/utils/sim.py,sha256=NZc9QGD6BTY5O29NVbHH7oxDePL_QMnylT1lYcDUn1Y,7437
 vision_agent/utils/type_defs.py,sha256=BE12s3JNQy36QvauXHjwyeffVh5enfcvd4vTzSwvEZI,1384
 vision_agent/utils/video.py,sha256=tRcGp4vEnaDycigL1hBO9k0FBPtDH35fCQciVr9GqYI,6013
-vision_agent-0.2.202.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-vision_agent-0.2.202.dist-info/METADATA,sha256=czcpEPxPUmbKBV1E5pOb-GpG_rkHYHjknLoOA3bekbk,19026
-vision_agent-0.2.202.dist-info/WHEEL,sha256=7Z8_27uaHI_UZAc4Uox4PpBhQ9Y5_modZXWMxtUi4NU,88
-vision_agent-0.2.202.dist-info/RECORD,,
+vision_agent-0.2.204.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+vision_agent-0.2.204.dist-info/METADATA,sha256=cuwR0b_QsTgq_dle_aATNpcNC-XGl78sLY11dS9OGbg,19026
+vision_agent-0.2.204.dist-info/WHEEL,sha256=7Z8_27uaHI_UZAc4Uox4PpBhQ9Y5_modZXWMxtUi4NU,88
+vision_agent-0.2.204.dist-info/RECORD,,