feat: conversational agent support

[mike-deem-uipath]

Depends on UiPath/uipath-runtime-python#65

Adds support for conversational agents:

• Adds some optional overrides for local testing with CAS (configured using env vars)
• Log conversation error events received from CAS over the websocket bridge
• Adds an is_conversational getter to AgentDefinition
• Adds --keep-state-file cli option that causes the state.db to be preserved. This allows conversational agent runs to accumulate conversation history.

Also

• Fixed a minor typo in the autonomous agent system prompt.

[cristipufu]

let's replace this with:

@click.option(
    "--state-file",
    required=False,
    type=click.Path(exists=True),
    help="Full path to the state file (takes priority over default location)"
)

Trying to shoot two birds with one stone, if we have this specified, we just use it as an override and don't delete it. Updated the runtime PR as well https://github.com/UiPath/uipath-runtime-python/pull/65/changes

We need to update the pyproject.toml with the new runtime version range + increment the minor here as well

[mike-deem-uipath]

Now if I want to keep the file, I have to know and provide the path to it? This doesn't make sense to me... the file you want to use and if you want to keep it are separate concepts.

[mike-deem-uipath]

Here is where the keep-state-file flag was used: https://github.com/UiPath/uipath-langchain-python/pull/395/changes#diff-7533e122e0f409dfcb61aa90e55e348bb528eac57d711599e88b198f38a467adR70. So this becomes self.context.state_file_path is not None. Why would one expect just setting that property to have the side effect of not deleting the file? And what if I want the file to be deleted AND I want to override the path?

Having an explicit keep-state-file flag avoids these unexpected side effects.

[cristipufu]

We need the state file pointer for something else, was thinking we can re-use the concept for this as well, and avoid adding another option.

If you insist, you can add the keep-state-file option back

[cristipufu]

The idea behind my thoughts is that you either:

• provide your own sqlite database, which we use to store the state
• or we create a temporary one for the local run and clean-up afterwards

[cristipufu]

state_file_path=state_file

[cristipufu]

let's change this method's signature to only use what it needs from LowCodeAgentDefinition (agent_name str, model str, system_message). we need to re-use it for the conversion flow

[cristipufu]

for reference, this is how a converted autonomous low-code agent looks today:

class AgentInput(BaseModel):
    pass

class AgentOutput(BaseModel):
    content: str | None = Field(None, description="Output content")

# Agent Messages Function
def create_messages(state: AgentInput) -> Sequence[SystemMessage | HumanMessage]:
    # Apply system prompt template
    current_date = datetime.now(timezone.utc).strftime('%Y-%m-%d')
    system_prompt_content = f"""You are an agentic assistant."""
    enhanced_system_prompt = (
        AGENT_SYSTEM_PROMPT_TEMPLATE
        .replace('{{systemPrompt}}', system_prompt_content)
        .replace('{{currentDate}}', current_date)
        .replace('{{agentName}}', 'Mr Assistant')
    )

    return [
        SystemMessage(content=enhanced_system_prompt),
        HumanMessage(content=f"""What is the current date?"""),
    ]

# Create agent graph
graph = create_agent(model=llm, messages=create_messages, tools=all_tools, input_schema=AgentInput, output_schema=AgentOutput)

[cristipufu]

I would also rename the method to get_chat_system_prompt to be consistent with current public namespaces

[cristipufu]

we use buit-in types (dict not Dict, x | None instead of Optional etc)