agent-runtime-core 0.6.0__py3-none-any.whl → 0.7.1__py3-none-any.whl

agent_runtime_core/interfaces.py

@@ -15,6 +15,18 @@ from typing import Any, Callable, Optional, Protocol, TypedDict, AsyncIterator
 from uuid import UUID
 
 
+class EventVisibility(str, Enum):
+    """
+    Visibility levels for events.
+
+    Controls which events are shown to users in the UI.
+    """
+
+    INTERNAL = "internal"  # Never shown to UI (checkpoints, heartbeats)
+    DEBUG = "debug"  # Shown only in debug mode (tool calls, tool results)
+    USER = "user"  # Always shown to users (assistant messages, errors)
+
+
 class EventType(str, Enum):
     """
     Standard event types emitted by agent runtimes.
@@ -41,6 +53,9 @@ class EventType(str, Enum):
     # State events
     STATE_CHECKPOINT = "state.checkpoint"
 
+    # Error events (distinct from run.failed - for runtime errors shown to users)
+    ERROR = "error"
+
     # Step execution events (for long-running multi-step agents)
     STEP_STARTED = "step.started"
     STEP_COMPLETED = "step.completed"
@@ -138,6 +153,30 @@ class RunContext(Protocol):
         """
         ...
 
+    async def emit_user_message(self, content: str) -> None:
+        """
+        Emit a message that will always be shown to the user.
+
+        This is a convenience method for emitting assistant messages.
+
+        Args:
+            content: The message content to display
+        """
+        ...
+
+    async def emit_error(self, error: str, details: dict = None) -> None:
+        """
+        Emit an error that will be shown to the user.
+
+        This is for runtime errors that should be displayed to users,
+        distinct from run.failed which is the final failure event.
+
+        Args:
+            error: The error message
+            details: Optional additional error details
+        """
+        ...
+
     async def checkpoint(self, state: dict) -> None:
         """
         Save a state checkpoint for recovery.
@@ -403,6 +442,56 @@ class LLMClient(ABC):
         ...
 
 
+class LLMToolCall:
+    """
+    Wrapper for tool call data from LLM responses to provide attribute access.
+
+    This provides a consistent interface for accessing tool call data
+    regardless of the underlying format (OpenAI, Anthropic, etc.).
+
+    Note: This is different from persistence.ToolCall which is a dataclass
+    for storing tool calls in conversations.
+    """
+
+    def __init__(self, data: dict):
+        self._data = data
+
+    @property
+    def id(self) -> str:
+        return self._data.get("id", "")
+
+    @property
+    def name(self) -> str:
+        func = self._data.get("function", {})
+        return func.get("name", "")
+
+    @property
+    def arguments(self) -> dict:
+        import json
+        import ast
+        func = self._data.get("function", {})
+        args = func.get("arguments", "{}")
+        if isinstance(args, str):
+            # First try standard JSON parsing
+            try:
+                return json.loads(args)
+            except json.JSONDecodeError:
+                pass
+
+            # Some models (e.g., Claude via certain providers) return Python dict syntax
+            # with single quotes instead of JSON double quotes. Try ast.literal_eval.
+            try:
+                result = ast.literal_eval(args)
+                if isinstance(result, dict):
+                    return result
+            except (ValueError, SyntaxError):
+                pass
+
+            # Last resort: return empty dict
+            return {}
+        return args
+
+
 @dataclass
 class LLMResponse:
     """Response from an LLM generation."""
@@ -413,6 +502,23 @@ class LLMResponse:
     finish_reason: str = ""
     raw_response: Optional[Any] = None
 
+    @property
+    def tool_calls(self) -> Optional[list["LLMToolCall"]]:
+        """Extract tool_calls from the message for convenience."""
+        if isinstance(self.message, dict):
+            calls = self.message.get("tool_calls")
+            if calls:
+                # Convert to LLMToolCall objects with name, arguments, id attributes
+                return [LLMToolCall(tc) for tc in calls]
+        return None
+
+    @property
+    def content(self) -> str:
+        """Extract content from the message for convenience."""
+        if isinstance(self.message, dict):
+            return self.message.get("content", "")
+        return ""
+
 
 @dataclass
 class LLMStreamChunk:

agent_runtime_core/json_runtime.py

@@ -0,0 +1,509 @@
+"""
+JsonAgentRuntime - A runtime that loads agent configuration from JSON.
+
+This allows running agents defined in the portable AgentConfig format,
+either from a JSON file or from a Django revision snapshot.
+
+Example:
+    # Load from file
+    config = AgentConfig.from_file("my_agent.json")
+    runtime = JsonAgentRuntime(config)
+
+    # Or load directly
+    runtime = JsonAgentRuntime.from_file("my_agent.json")
+
+    # Run the agent
+    result = await runtime.run(ctx)
+
+Multi-Agent Support:
+    The runtime supports sub-agent tools defined in the config. Sub-agents
+    can be embedded (agent_config) or referenced by slug (agent_slug).
+
+    # Config with embedded sub-agent
+    config = AgentConfig(
+        slug="triage-agent",
+        sub_agent_tools=[
+            SubAgentToolConfig(
+                name="billing_specialist",
+                description="Handle billing questions",
+                agent_config=AgentConfig(slug="billing-agent", ...),
+            )
+        ]
+    )
+"""
+
+import importlib
+import logging
+from typing import Any, Callable, Optional
+
+from agent_runtime_core.interfaces import (
+    AgentRuntime,
+    RunContext,
+    RunResult,
+    Tool,
+    ToolDefinition,
+)
+from agent_runtime_core.agentic_loop import run_agentic_loop
+from agent_runtime_core.config_schema import AgentConfig, ToolConfig, SubAgentToolConfig
+
+logger = logging.getLogger(__name__)
+
+
+def resolve_function(function_path: str) -> Callable:
+    """
+    Resolve a function path like 'myapp.services.orders.lookup_order' to the actual callable.
+    
+    Args:
+        function_path: Dotted path to the function
+        
+    Returns:
+        The callable function
+        
+    Raises:
+        ImportError: If the module cannot be imported
+        AttributeError: If the function doesn't exist in the module
+    """
+    parts = function_path.rsplit(".", 1)
+    if len(parts) != 2:
+        raise ValueError(f"Invalid function path: {function_path}. Expected 'module.function' format.")
+    
+    module_path, function_name = parts
+    module = importlib.import_module(module_path)
+    return getattr(module, function_name)
+
+
+class ConfiguredTool(Tool):
+    """A tool created from ToolConfig that resolves function_path at runtime."""
+    
+    def __init__(self, config: ToolConfig):
+        self.config = config
+        self._function: Optional[Callable] = None
+    
+    @property
+    def definition(self) -> ToolDefinition:
+        return ToolDefinition(
+            name=self.config.name,
+            description=self.config.description,
+            parameters=self.config.parameters,
+        )
+    
+    def _get_function(self) -> Callable:
+        """Lazily resolve the function."""
+        if self._function is None:
+            self._function = resolve_function(self.config.function_path)
+        return self._function
+    
+    async def execute(self, args: dict, ctx: RunContext) -> Any:
+        """Execute the tool by calling the resolved function."""
+        func = self._get_function()
+
+        # Check if function is async
+        if hasattr(func, "__call__"):
+            import asyncio
+            if asyncio.iscoroutinefunction(func):
+                return await func(**args)
+            else:
+                # Run sync function in thread pool
+                loop = asyncio.get_event_loop()
+                return await loop.run_in_executor(None, lambda: func(**args))
+
+        return func(**args)
+
+
+class SubAgentTool(Tool):
+    """
+    A tool that delegates to another agent (sub-agent).
+
+    This implements the "agent-as-tool" pattern where one agent can
+    invoke another agent as if it were a tool.
+    """
+
+    def __init__(
+        self,
+        config: SubAgentToolConfig,
+        agent_registry: Optional[dict[str, "JsonAgentRuntime"]] = None,
+    ):
+        """
+        Initialize a sub-agent tool.
+
+        Args:
+            config: The sub-agent tool configuration
+            agent_registry: Optional registry of pre-loaded agent runtimes
+                           (for embedded configs or external lookup)
+        """
+        self.config = config
+        self.agent_registry = agent_registry or {}
+        self._runtime: Optional["JsonAgentRuntime"] = None
+
+    @property
+    def definition(self) -> ToolDefinition:
+        """Get the tool definition for this sub-agent."""
+        # Sub-agent tools take a single 'message' parameter
+        # Note: The handler here is a placeholder - actual execution goes through
+        # execute() which receives the RunContext. The runtime handles this specially.
+        async def _placeholder_handler(**kwargs):
+            raise RuntimeError("SubAgentTool.execute() should be called directly with context")
+
+        return ToolDefinition(
+            name=self.config.name,
+            description=self.config.description,
+            parameters={
+                "type": "object",
+                "properties": {
+                    "message": {
+                        "type": "string",
+                        "description": "The message or task to send to the sub-agent",
+                    },
+                    "context": {
+                        "type": "string",
+                        "description": "Optional additional context for the sub-agent",
+                    },
+                },
+                "required": ["message"],
+            },
+            handler=_placeholder_handler,
+        )
+
+    def _get_runtime(self) -> "JsonAgentRuntime":
+        """Get or create the sub-agent runtime."""
+        if self._runtime is not None:
+            return self._runtime
+
+        # Try embedded config first
+        if self.config.agent_config:
+            self._runtime = JsonAgentRuntime(self.config.agent_config)
+            return self._runtime
+
+        # Try registry lookup by slug
+        if self.config.agent_slug:
+            if self.config.agent_slug in self.agent_registry:
+                self._runtime = self.agent_registry[self.config.agent_slug]
+                return self._runtime
+            raise ValueError(
+                f"Sub-agent '{self.config.agent_slug}' not found in registry. "
+                f"Available: {list(self.agent_registry.keys())}"
+            )
+
+        raise ValueError(
+            f"Sub-agent tool '{self.config.name}' has no agent_config or agent_slug"
+        )
+
+    async def execute(self, args: dict, ctx: RunContext) -> Any:
+        """
+        Execute the sub-agent with the given message.
+
+        Args:
+            args: Tool arguments (message, optional context)
+            ctx: The parent run context
+
+        Returns:
+            The sub-agent's response
+        """
+        message = args.get("message", "")
+        additional_context = args.get("context", "")
+
+        runtime = self._get_runtime()
+
+        # Build messages for sub-agent based on context_mode
+        sub_messages = self._build_sub_agent_messages(message, additional_context, ctx)
+
+        # Create a sub-context for the sub-agent
+        # Note: In a full implementation, you'd want to track parent-child relationships
+        from agent_runtime_core.interfaces import RunContext as RC
+        sub_ctx = RC(
+            run_id=f"{ctx.run_id}-sub-{self.config.name}",
+            conversation_id=ctx.conversation_id,
+            input_messages=sub_messages,
+            params=ctx.params,
+        )
+
+        # Run the sub-agent
+        logger.info(f"Invoking sub-agent '{self.config.name}' ({runtime.key})")
+        result = await runtime.run(sub_ctx)
+
+        # Extract the response
+        response = result.final_output.get("response", "")
+        if not response and result.final_messages:
+            # Try to get from last assistant message
+            for msg in reversed(result.final_messages):
+                if msg.get("role") == "assistant" and msg.get("content"):
+                    response = msg["content"]
+                    break
+
+        return response
+
+    def _build_sub_agent_messages(
+        self,
+        message: str,
+        additional_context: str,
+        ctx: RunContext,
+    ) -> list[dict]:
+        """
+        Build the message list for the sub-agent based on context_mode.
+
+        Args:
+            message: The message to send
+            additional_context: Optional additional context
+            ctx: Parent run context
+
+        Returns:
+            List of messages for the sub-agent
+        """
+        context_mode = self.config.context_mode
+
+        if context_mode == "full":
+            # Pass full conversation history
+            messages = list(ctx.input_messages)
+            # Add the delegation message
+            if additional_context:
+                messages.append({
+                    "role": "user",
+                    "content": f"{message}\n\nAdditional context: {additional_context}",
+                })
+            else:
+                messages.append({"role": "user", "content": message})
+            return messages
+
+        elif context_mode == "summary":
+            # Summarize context (simplified - in production you'd use LLM)
+            summary = f"Previous conversation context: {len(ctx.input_messages)} messages exchanged."
+            return [
+                {"role": "system", "content": summary},
+                {"role": "user", "content": message},
+            ]
+
+        else:  # "message_only" or default
+            # Just the message, no context
+            content = message
+            if additional_context:
+                content = f"{message}\n\nContext: {additional_context}"
+            return [{"role": "user", "content": content}]
+
+
+class JsonAgentRuntime(AgentRuntime):
+    """
+    An agent runtime that loads its configuration from AgentConfig.
+
+    This provides a portable way to define and run agents using JSON configuration,
+    without requiring Django or any specific framework.
+
+    Supports multi-agent systems through sub_agent_tools, which allow this agent
+    to delegate to other agents as if they were tools.
+    """
+
+    def __init__(
+        self,
+        config: AgentConfig,
+        agent_registry: Optional[dict[str, "JsonAgentRuntime"]] = None,
+    ):
+        """
+        Initialize the runtime with an AgentConfig.
+
+        Args:
+            config: The agent configuration
+            agent_registry: Optional registry of pre-loaded agent runtimes
+                           for sub-agent lookup by slug
+        """
+        self.config = config
+        self.agent_registry = agent_registry or {}
+        self._tools: list[Tool] = []
+        self._tools_loaded = False
+
+        # Build registry from embedded sub-agent configs
+        self._build_embedded_agent_registry()
+
+    def _build_embedded_agent_registry(self) -> None:
+        """Build registry entries for embedded sub-agent configs."""
+        for sub_tool in self.config.sub_agent_tools:
+            if sub_tool.agent_config:
+                slug = sub_tool.agent_config.slug
+                if slug not in self.agent_registry:
+                    # Create runtime for embedded config (recursively)
+                    self.agent_registry[slug] = JsonAgentRuntime(
+                        sub_tool.agent_config,
+                        agent_registry=self.agent_registry,
+                    )
+
+    @property
+    def key(self) -> str:
+        return self.config.slug
+
+    @classmethod
+    def from_file(cls, path: str) -> "JsonAgentRuntime":
+        """Load a runtime from a JSON file."""
+        config = AgentConfig.from_file(path)
+        return cls(config)
+
+    @classmethod
+    def from_json(cls, json_str: str) -> "JsonAgentRuntime":
+        """Load a runtime from a JSON string."""
+        config = AgentConfig.from_json(json_str)
+        return cls(config)
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "JsonAgentRuntime":
+        """Load a runtime from a dictionary."""
+        config = AgentConfig.from_dict(data)
+        return cls(config)
+
+    @classmethod
+    def from_system_export(cls, data: dict) -> "JsonAgentRuntime":
+        """
+        Load a runtime from an exported multi-agent system.
+
+        This handles the system export format which has entry_agent at the top level.
+
+        Args:
+            data: Exported system config (from AgentSystemVersion.export_config)
+
+        Returns:
+            JsonAgentRuntime for the entry agent with all sub-agents wired up
+        """
+        entry_agent_data = data.get("entry_agent")
+        if not entry_agent_data:
+            raise ValueError("System export must have 'entry_agent' key")
+
+        config = AgentConfig.from_dict(entry_agent_data)
+        return cls(config)
+
+    def _load_tools(self) -> list[Tool]:
+        """Load and resolve all tools from config, including sub-agent tools."""
+        if self._tools_loaded:
+            return self._tools
+
+        self._tools = []
+
+        # Load regular function tools
+        for tool_config in self.config.tools:
+            try:
+                tool = ConfiguredTool(tool_config)
+                # Validate that the function can be resolved
+                tool._get_function()
+                self._tools.append(tool)
+                logger.debug(f"Loaded tool: {tool_config.name}")
+            except Exception as e:
+                logger.error(f"Failed to load tool {tool_config.name}: {e}")
+                raise
+
+        # Load sub-agent tools
+        for sub_tool_config in self.config.sub_agent_tools:
+            try:
+                tool = SubAgentTool(sub_tool_config, agent_registry=self.agent_registry)
+                self._tools.append(tool)
+                logger.debug(f"Loaded sub-agent tool: {sub_tool_config.name}")
+            except Exception as e:
+                logger.error(f"Failed to load sub-agent tool {sub_tool_config.name}: {e}")
+                raise
+
+        self._tools_loaded = True
+        return self._tools
+
+    def _build_system_prompt(self) -> str:
+        """Build the full system prompt including knowledge."""
+        parts = []
+
+        # Add base system prompt
+        if self.config.system_prompt:
+            parts.append(self.config.system_prompt)
+
+        # Add always-included knowledge
+        for knowledge in self.config.knowledge:
+            if knowledge.inclusion_mode == "always" and knowledge.content:
+                parts.append(f"\n\n## {knowledge.name}\n{knowledge.content}")
+
+        return "\n".join(parts)
+
+    async def run(self, ctx: RunContext) -> RunResult:
+        """
+        Run the agent using the agentic loop.
+
+        Args:
+            ctx: The run context with conversation history and state
+
+        Returns:
+            RunResult with the agent's response
+        """
+        from agent_runtime_core.llm import get_llm_client
+
+        # Load tools
+        tools = self._load_tools()
+
+        # Build system prompt with knowledge
+        system_prompt = self._build_system_prompt()
+
+        # Get model settings
+        model = self.config.model
+        model_settings = self.config.model_settings or {}
+
+        # Get LLM client
+        llm = get_llm_client(model=model)
+
+        # Build messages list with system prompt
+        messages = []
+        if system_prompt:
+            messages.append({"role": "system", "content": system_prompt})
+        messages.extend(ctx.input_messages)
+
+        # Convert tools to OpenAI format
+        tool_schemas = None
+        tool_map = {}
+        if tools:
+            tool_schemas = []
+            for tool in tools:
+                tool_schemas.append({
+                    "type": "function",
+                    "function": {
+                        "name": tool.definition.name,
+                        "description": tool.definition.description,
+                        "parameters": tool.definition.parameters,
+                    }
+                })
+                tool_map[tool.definition.name] = tool
+
+        # Create tool executor
+        async def execute_tool(name: str, args: dict) -> Any:
+            if name not in tool_map:
+                raise ValueError(f"Unknown tool: {name}")
+            return await tool_map[name].execute(args, ctx)
+
+        # Run the agentic loop
+        result = await run_agentic_loop(
+            llm=llm,
+            messages=messages,
+            tools=tool_schemas,
+            execute_tool=execute_tool,
+            ctx=ctx,
+            model=model,
+            **model_settings,
+        )
+
+        return RunResult(
+            final_output={"response": result.final_content},
+            final_messages=result.messages,
+            usage=result.usage,
+        )
+
+    def get_tools(self) -> list[Tool]:
+        """Get the list of configured tools (including sub-agent tools)."""
+        return self._load_tools()
+
+    def get_system_prompt(self) -> str:
+        """Get the full system prompt including knowledge."""
+        return self._build_system_prompt()
+
+    def get_sub_agent_tools(self) -> list[SubAgentTool]:
+        """Get only the sub-agent tools."""
+        return [t for t in self._load_tools() if isinstance(t, SubAgentTool)]
+
+    def get_sub_agent_runtimes(self) -> dict[str, "JsonAgentRuntime"]:
+        """
+        Get all sub-agent runtimes available to this agent.
+
+        Returns:
+            Dict mapping agent slug to JsonAgentRuntime
+        """
+        return dict(self.agent_registry)
+
+    def has_sub_agents(self) -> bool:
+        """Check if this agent has any sub-agent tools."""
+        return len(self.config.sub_agent_tools) > 0