tracia 0.0.1__py3-none-any.whl → 0.1.0__py3-none-any.whl

tracia/_llm.py

@@ -0,0 +1,898 @@
+"""LiteLLM wrapper for unified LLM access."""
+
+from __future__ import annotations
+
+import json
+import os
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, AsyncIterator, Iterator
+
+from ._constants import ENV_VAR_MAP
+from ._errors import TraciaError, TraciaErrorCode, sanitize_error_message
+from ._types import (
+    ContentPart,
+    FinishReason,
+    LLMProvider,
+    LocalPromptMessage,
+    TextPart,
+    ToolCall,
+    ToolCallPart,
+    ToolChoice,
+    ToolDefinition,
+)
+
+if TYPE_CHECKING:
+    from litellm import ModelResponse
+
+
+@dataclass
+class CompletionResult:
+    """Result from an LLM completion."""
+
+    text: str
+    input_tokens: int
+    output_tokens: int
+    total_tokens: int
+    tool_calls: list[ToolCall]
+    finish_reason: FinishReason
+    provider: LLMProvider
+
+
+# Model to provider mapping for common models
+_MODEL_PROVIDER_MAP: dict[str, LLMProvider] = {
+    # OpenAI
+    "gpt-3.5-turbo": LLMProvider.OPENAI,
+    "gpt-4": LLMProvider.OPENAI,
+    "gpt-4-turbo": LLMProvider.OPENAI,
+    "gpt-4o": LLMProvider.OPENAI,
+    "gpt-4o-mini": LLMProvider.OPENAI,
+    "gpt-4.1": LLMProvider.OPENAI,
+    "gpt-4.1-mini": LLMProvider.OPENAI,
+    "gpt-4.1-nano": LLMProvider.OPENAI,
+    "gpt-4.5-preview": LLMProvider.OPENAI,
+    "gpt-5": LLMProvider.OPENAI,
+    "o1": LLMProvider.OPENAI,
+    "o1-mini": LLMProvider.OPENAI,
+    "o1-preview": LLMProvider.OPENAI,
+    "o3": LLMProvider.OPENAI,
+    "o3-mini": LLMProvider.OPENAI,
+    "o4-mini": LLMProvider.OPENAI,
+    # Anthropic
+    "claude-3-haiku-20240307": LLMProvider.ANTHROPIC,
+    "claude-3-sonnet-20240229": LLMProvider.ANTHROPIC,
+    "claude-3-opus-20240229": LLMProvider.ANTHROPIC,
+    "claude-3-5-haiku-20241022": LLMProvider.ANTHROPIC,
+    "claude-3-5-sonnet-20241022": LLMProvider.ANTHROPIC,
+    "claude-sonnet-4-20250514": LLMProvider.ANTHROPIC,
+    "claude-opus-4-20250514": LLMProvider.ANTHROPIC,
+    # Google
+    "gemini-2.0-flash": LLMProvider.GOOGLE,
+    "gemini-2.0-flash-lite": LLMProvider.GOOGLE,
+    "gemini-2.5-pro": LLMProvider.GOOGLE,
+    "gemini-2.5-flash": LLMProvider.GOOGLE,
+}
+
+
+def resolve_provider(model: str, explicit_provider: LLMProvider | None) -> LLMProvider:
+    """Resolve the provider for a model.
+
+    Args:
+        model: The model name.
+        explicit_provider: Explicitly specified provider.
+
+    Returns:
+        The resolved provider.
+
+    Raises:
+        TraciaError: If the provider cannot be determined.
+    """
+    if explicit_provider is not None:
+        return explicit_provider
+
+    # Check the model registry
+    if model in _MODEL_PROVIDER_MAP:
+        return _MODEL_PROVIDER_MAP[model]
+
+    # Try prefix-based detection
+    if model.startswith("gpt-") or model.startswith("o1") or model.startswith("o3") or model.startswith("o4"):
+        return LLMProvider.OPENAI
+    if model.startswith("claude-"):
+        return LLMProvider.ANTHROPIC
+    if model.startswith("gemini-"):
+        return LLMProvider.GOOGLE
+
+    raise TraciaError(
+        code=TraciaErrorCode.UNSUPPORTED_MODEL,
+        message=f"Cannot determine provider for model '{model}'. Please specify the provider explicitly.",
+    )
+
+
+def get_litellm_model(model: str, provider: LLMProvider) -> str:
+    """Get the litellm-compatible model name.
+
+    LiteLLM requires a ``gemini/`` prefix to route Google AI Studio models
+    correctly. Without it, litellm defaults to the Vertex AI path which
+    requires Application Default Credentials instead of an API key.
+
+    Args:
+        model: The user-facing model name (e.g. ``gemini-2.0-flash``).
+        provider: The resolved provider.
+
+    Returns:
+        The model string suitable for ``litellm.completion()``.
+    """
+    if provider == LLMProvider.GOOGLE and not model.startswith("gemini/"):
+        return f"gemini/{model}"
+    return model
+
+
+def get_provider_api_key(
+    provider: LLMProvider,
+    provider_api_key: str | None = None,
+) -> str:
+    """Get the API key for a provider.
+
+    Args:
+        provider: The LLM provider.
+        provider_api_key: Explicitly provided API key.
+
+    Returns:
+        The API key.
+
+    Raises:
+        TraciaError: If no API key is found.
+    """
+    if provider_api_key:
+        return provider_api_key
+
+    env_var = ENV_VAR_MAP.get(provider.value)
+    if env_var:
+        key = os.environ.get(env_var)
+        if key:
+            return key
+
+    raise TraciaError(
+        code=TraciaErrorCode.MISSING_PROVIDER_API_KEY,
+        message=f"No API key found for provider '{provider.value}'. "
+        f"Set the {ENV_VAR_MAP.get(provider.value, 'PROVIDER_API_KEY')} environment variable "
+        "or pass provider_api_key parameter.",
+    )
+
+
+def convert_messages(
+    messages: list[LocalPromptMessage],
+) -> list[dict[str, Any]]:
+    """Convert Tracia messages to LiteLLM/OpenAI format.
+
+    Args:
+        messages: The Tracia messages.
+
+    Returns:
+        Messages in LiteLLM format.
+    """
+    result: list[dict[str, Any]] = []
+
+    for msg in messages:
+        # Handle tool role
+        if msg.role == "tool":
+            content = msg.content if isinstance(msg.content, str) else str(msg.content)
+            result.append({
+                "role": "tool",
+                "tool_call_id": msg.tool_call_id,
+                "content": content,
+            })
+            continue
+
+        # Handle developer role (map to system)
+        role = "system" if msg.role == "developer" else msg.role
+
+        # Handle string content
+        if isinstance(msg.content, str):
+            result.append({"role": role, "content": msg.content})
+            continue
+
+        # Handle list content (text parts and tool calls)
+        content_parts: list[dict[str, Any]] = []
+        tool_calls: list[dict[str, Any]] = []
+
+        for part in msg.content:
+            if isinstance(part, TextPart) or (isinstance(part, dict) and part.get("type") == "text"):
+                text = part.text if isinstance(part, TextPart) else part.get("text", "")
+                content_parts.append({"type": "text", "text": text})
+            elif isinstance(part, ToolCallPart) or (isinstance(part, dict) and part.get("type") == "tool_call"):
+                if isinstance(part, ToolCallPart):
+                    tc_id = part.id
+                    tc_name = part.name
+                    tc_args = part.arguments
+                else:
+                    tc_id = part.get("id", "")
+                    tc_name = part.get("name", "")
+                    tc_args = part.get("arguments", {})
+
+                tool_calls.append({
+                    "id": tc_id,
+                    "type": "function",
+                    "function": {
+                        "name": tc_name,
+                        "arguments": json.dumps(tc_args) if isinstance(tc_args, dict) else tc_args,
+                    },
+                })
+
+        # Build the message
+        msg_dict: dict[str, Any] = {"role": role}
+
+        if content_parts:
+            # If only text parts, we can simplify
+            if len(content_parts) == 1 and not tool_calls:
+                msg_dict["content"] = content_parts[0]["text"]
+            else:
+                msg_dict["content"] = content_parts
+        elif not tool_calls:
+            msg_dict["content"] = ""
+
+        if tool_calls:
+            msg_dict["tool_calls"] = tool_calls
+
+        result.append(msg_dict)
+
+    return result
+
+
+def convert_tools(tools: list[ToolDefinition] | None) -> list[dict[str, Any]] | None:
+    """Convert tool definitions to LiteLLM format.
+
+    Args:
+        tools: The tool definitions.
+
+    Returns:
+        Tools in LiteLLM format.
+    """
+    if not tools:
+        return None
+
+    result = []
+    for tool in tools:
+        result.append({
+            "type": "function",
+            "function": {
+                "name": tool.name,
+                "description": tool.description,
+                "parameters": tool.parameters.model_dump(exclude_none=True),
+            },
+        })
+    return result
+
+
+def convert_tool_choice(tool_choice: ToolChoice | None) -> str | dict[str, Any] | None:
+    """Convert tool choice to LiteLLM format.
+
+    Args:
+        tool_choice: The tool choice.
+
+    Returns:
+        Tool choice in LiteLLM format.
+    """
+    if tool_choice is None:
+        return None
+
+    if isinstance(tool_choice, str):
+        return tool_choice
+
+    if isinstance(tool_choice, dict) and "tool" in tool_choice:
+        return {"type": "function", "function": {"name": tool_choice["tool"]}}
+
+    return None
+
+
+def parse_finish_reason(reason: str | None) -> FinishReason:
+    """Parse the finish reason from LiteLLM response.
+
+    Args:
+        reason: The raw finish reason.
+
+    Returns:
+        The normalized finish reason.
+    """
+    if reason == "tool_calls":
+        return "tool_calls"
+    if reason == "length":
+        return "max_tokens"
+    return "stop"
+
+
+def extract_tool_calls(response: "ModelResponse") -> list[ToolCall]:
+    """Extract tool calls from a LiteLLM response.
+
+    Args:
+        response: The LiteLLM response.
+
+    Returns:
+        The extracted tool calls.
+    """
+    tool_calls: list[ToolCall] = []
+
+    choices = getattr(response, "choices", [])
+    if not choices:
+        return tool_calls
+
+    message = getattr(choices[0], "message", None)
+    if not message:
+        return tool_calls
+
+    raw_tool_calls = getattr(message, "tool_calls", None)
+    if not raw_tool_calls:
+        return tool_calls
+
+    for tc in raw_tool_calls:
+        func = getattr(tc, "function", None)
+        if func:
+            try:
+                args = json.loads(func.arguments) if isinstance(func.arguments, str) else func.arguments
+            except json.JSONDecodeError:
+                args = {}
+
+            tool_calls.append(ToolCall(
+                id=tc.id,
+                name=func.name,
+                arguments=args,
+            ))
+
+    return tool_calls
+
+
+class LLMClient:
+    """Client for making LLM calls via LiteLLM."""
+
+    def complete(
+        self,
+        model: str,
+        messages: list[LocalPromptMessage],
+        *,
+        provider: LLMProvider | None = None,
+        temperature: float | None = None,
+        max_tokens: int | None = None,
+        top_p: float | None = None,
+        stop: list[str] | None = None,
+        tools: list[ToolDefinition] | None = None,
+        tool_choice: ToolChoice | None = None,
+        api_key: str | None = None,
+        timeout: float | None = None,
+    ) -> CompletionResult:
+        """Make a synchronous completion request.
+
+        Args:
+            model: The model name.
+            messages: The messages to send.
+            provider: The LLM provider.
+            temperature: Sampling temperature.
+            max_tokens: Maximum output tokens.
+            top_p: Top-p sampling.
+            stop: Stop sequences.
+            tools: Tool definitions.
+            tool_choice: Tool choice setting.
+            api_key: Provider API key.
+            timeout: Request timeout in seconds.
+
+        Returns:
+            The completion result.
+
+        Raises:
+            TraciaError: If the request fails.
+        """
+        try:
+            import litellm
+        except ImportError as e:
+            raise TraciaError(
+                code=TraciaErrorCode.MISSING_PROVIDER_SDK,
+                message="litellm is not installed. Install it with: pip install litellm",
+            ) from e
+
+        resolved_provider = resolve_provider(model, provider)
+        resolved_api_key = get_provider_api_key(resolved_provider, api_key)
+
+        # Build the request
+        litellm_messages = convert_messages(messages)
+        litellm_tools = convert_tools(tools)
+        litellm_tool_choice = convert_tool_choice(tool_choice)
+
+        request_kwargs: dict[str, Any] = {
+            "model": get_litellm_model(model, resolved_provider),
+            "messages": litellm_messages,
+            "api_key": resolved_api_key,
+        }
+
+        if temperature is not None:
+            request_kwargs["temperature"] = temperature
+        if max_tokens is not None:
+            request_kwargs["max_tokens"] = max_tokens
+        if top_p is not None:
+            request_kwargs["top_p"] = top_p
+        if stop is not None:
+            request_kwargs["stop"] = stop
+        if litellm_tools is not None:
+            request_kwargs["tools"] = litellm_tools
+        if litellm_tool_choice is not None:
+            request_kwargs["tool_choice"] = litellm_tool_choice
+        if timeout is not None:
+            request_kwargs["timeout"] = timeout
+
+        try:
+            response = litellm.completion(**request_kwargs)
+        except Exception as e:
+            error_msg = sanitize_error_message(str(e))
+            raise TraciaError(
+                code=TraciaErrorCode.PROVIDER_ERROR,
+                message=f"LLM provider error: {error_msg}",
+            ) from e
+
+        # Extract result
+        usage = getattr(response, "usage", None)
+        choices = getattr(response, "choices", [])
+        message = choices[0].message if choices else None
+        content = getattr(message, "content", "") or ""
+        finish_reason = choices[0].finish_reason if choices else "stop"
+
+        return CompletionResult(
+            text=content,
+            input_tokens=getattr(usage, "prompt_tokens", 0) if usage else 0,
+            output_tokens=getattr(usage, "completion_tokens", 0) if usage else 0,
+            total_tokens=getattr(usage, "total_tokens", 0) if usage else 0,
+            tool_calls=extract_tool_calls(response),
+            finish_reason=parse_finish_reason(finish_reason),
+            provider=resolved_provider,
+        )
+
+    async def acomplete(
+        self,
+        model: str,
+        messages: list[LocalPromptMessage],
+        *,
+        provider: LLMProvider | None = None,
+        temperature: float | None = None,
+        max_tokens: int | None = None,
+        top_p: float | None = None,
+        stop: list[str] | None = None,
+        tools: list[ToolDefinition] | None = None,
+        tool_choice: ToolChoice | None = None,
+        api_key: str | None = None,
+        timeout: float | None = None,
+    ) -> CompletionResult:
+        """Make an asynchronous completion request.
+
+        Args:
+            model: The model name.
+            messages: The messages to send.
+            provider: The LLM provider.
+            temperature: Sampling temperature.
+            max_tokens: Maximum output tokens.
+            top_p: Top-p sampling.
+            stop: Stop sequences.
+            tools: Tool definitions.
+            tool_choice: Tool choice setting.
+            api_key: Provider API key.
+            timeout: Request timeout in seconds.
+
+        Returns:
+            The completion result.
+
+        Raises:
+            TraciaError: If the request fails.
+        """
+        try:
+            import litellm
+        except ImportError as e:
+            raise TraciaError(
+                code=TraciaErrorCode.MISSING_PROVIDER_SDK,
+                message="litellm is not installed. Install it with: pip install litellm",
+            ) from e
+
+        resolved_provider = resolve_provider(model, provider)
+        resolved_api_key = get_provider_api_key(resolved_provider, api_key)
+
+        # Build the request
+        litellm_messages = convert_messages(messages)
+        litellm_tools = convert_tools(tools)
+        litellm_tool_choice = convert_tool_choice(tool_choice)
+
+        request_kwargs: dict[str, Any] = {
+            "model": get_litellm_model(model, resolved_provider),
+            "messages": litellm_messages,
+            "api_key": resolved_api_key,
+        }
+
+        if temperature is not None:
+            request_kwargs["temperature"] = temperature
+        if max_tokens is not None:
+            request_kwargs["max_tokens"] = max_tokens
+        if top_p is not None:
+            request_kwargs["top_p"] = top_p
+        if stop is not None:
+            request_kwargs["stop"] = stop
+        if litellm_tools is not None:
+            request_kwargs["tools"] = litellm_tools
+        if litellm_tool_choice is not None:
+            request_kwargs["tool_choice"] = litellm_tool_choice
+        if timeout is not None:
+            request_kwargs["timeout"] = timeout
+
+        try:
+            response = await litellm.acompletion(**request_kwargs)
+        except Exception as e:
+            error_msg = sanitize_error_message(str(e))
+            raise TraciaError(
+                code=TraciaErrorCode.PROVIDER_ERROR,
+                message=f"LLM provider error: {error_msg}",
+            ) from e
+
+        # Extract result
+        usage = getattr(response, "usage", None)
+        choices = getattr(response, "choices", [])
+        message = choices[0].message if choices else None
+        content = getattr(message, "content", "") or ""
+        finish_reason = choices[0].finish_reason if choices else "stop"
+
+        return CompletionResult(
+            text=content,
+            input_tokens=getattr(usage, "prompt_tokens", 0) if usage else 0,
+            output_tokens=getattr(usage, "completion_tokens", 0) if usage else 0,
+            total_tokens=getattr(usage, "total_tokens", 0) if usage else 0,
+            tool_calls=extract_tool_calls(response),
+            finish_reason=parse_finish_reason(finish_reason),
+            provider=resolved_provider,
+        )
+
+    def stream(
+        self,
+        model: str,
+        messages: list[LocalPromptMessage],
+        *,
+        provider: LLMProvider | None = None,
+        temperature: float | None = None,
+        max_tokens: int | None = None,
+        top_p: float | None = None,
+        stop: list[str] | None = None,
+        tools: list[ToolDefinition] | None = None,
+        tool_choice: ToolChoice | None = None,
+        api_key: str | None = None,
+        timeout: float | None = None,
+    ) -> tuple[Iterator[str], list[CompletionResult], LLMProvider]:
+        """Make a streaming completion request.
+
+        Args:
+            model: The model name.
+            messages: The messages to send.
+            provider: The LLM provider.
+            temperature: Sampling temperature.
+            max_tokens: Maximum output tokens.
+            top_p: Top-p sampling.
+            stop: Stop sequences.
+            tools: Tool definitions.
+            tool_choice: Tool choice setting.
+            api_key: Provider API key.
+            timeout: Request timeout in seconds.
+
+        Returns:
+            A tuple of (chunk iterator, result holder list, provider).
+
+        Raises:
+            TraciaError: If the request fails.
+        """
+        try:
+            import litellm
+        except ImportError as e:
+            raise TraciaError(
+                code=TraciaErrorCode.MISSING_PROVIDER_SDK,
+                message="litellm is not installed. Install it with: pip install litellm",
+            ) from e
+
+        resolved_provider = resolve_provider(model, provider)
+        resolved_api_key = get_provider_api_key(resolved_provider, api_key)
+
+        # Build the request
+        litellm_messages = convert_messages(messages)
+        litellm_tools = convert_tools(tools)
+        litellm_tool_choice = convert_tool_choice(tool_choice)
+
+        request_kwargs: dict[str, Any] = {
+            "model": get_litellm_model(model, resolved_provider),
+            "messages": litellm_messages,
+            "api_key": resolved_api_key,
+            "stream": True,
+        }
+
+        if temperature is not None:
+            request_kwargs["temperature"] = temperature
+        if max_tokens is not None:
+            request_kwargs["max_tokens"] = max_tokens
+        if top_p is not None:
+            request_kwargs["top_p"] = top_p
+        if stop is not None:
+            request_kwargs["stop"] = stop
+        if litellm_tools is not None:
+            request_kwargs["tools"] = litellm_tools
+        if litellm_tool_choice is not None:
+            request_kwargs["tool_choice"] = litellm_tool_choice
+        if timeout is not None:
+            request_kwargs["timeout"] = timeout
+
+        result_holder: list[CompletionResult] = []
+
+        def generate_chunks() -> Iterator[str]:
+            full_text = ""
+            input_tokens = 0
+            output_tokens = 0
+            total_tokens = 0
+            tool_calls: list[ToolCall] = []
+            finish_reason: FinishReason = "stop"
+            tool_call_chunks: dict[int, dict[str, Any]] = {}
+
+            try:
+                response = litellm.completion(**request_kwargs)
+
+                for chunk in response:
+                    choices = getattr(chunk, "choices", [])
+                    if not choices:
+                        continue
+
+                    delta = getattr(choices[0], "delta", None)
+                    if delta:
+                        content = getattr(delta, "content", None)
+                        if content:
+                            full_text += content
+                            yield content
+
+                        # Handle streaming tool calls
+                        delta_tool_calls = getattr(delta, "tool_calls", None)
+                        if delta_tool_calls:
+                            for tc in delta_tool_calls:
+                                idx = tc.index
+                                if idx not in tool_call_chunks:
+                                    tool_call_chunks[idx] = {
+                                        "id": "",
+                                        "name": "",
+                                        "arguments": "",
+                                    }
+                                if tc.id:
+                                    tool_call_chunks[idx]["id"] = tc.id
+                                if tc.function:
+                                    if tc.function.name:
+                                        tool_call_chunks[idx]["name"] = tc.function.name
+                                    if tc.function.arguments:
+                                        tool_call_chunks[idx]["arguments"] += tc.function.arguments
+
+                    chunk_finish = getattr(choices[0], "finish_reason", None)
+                    if chunk_finish:
+                        finish_reason = parse_finish_reason(chunk_finish)
+
+                    # Extract usage from final chunk
+                    usage = getattr(chunk, "usage", None)
+                    if usage:
+                        input_tokens = getattr(usage, "prompt_tokens", 0)
+                        output_tokens = getattr(usage, "completion_tokens", 0)
+                        total_tokens = getattr(usage, "total_tokens", 0)
+
+                # Convert accumulated tool calls
+                for idx in sorted(tool_call_chunks.keys()):
+                    tc_data = tool_call_chunks[idx]
+                    try:
+                        args = json.loads(tc_data["arguments"]) if tc_data["arguments"] else {}
+                    except json.JSONDecodeError:
+                        args = {}
+                    tool_calls.append(ToolCall(
+                        id=tc_data["id"],
+                        name=tc_data["name"],
+                        arguments=args,
+                    ))
+
+                result_holder.append(CompletionResult(
+                    text=full_text,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    total_tokens=total_tokens,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    provider=resolved_provider,
+                ))
+
+            except Exception as e:
+                error_msg = sanitize_error_message(str(e))
+                raise TraciaError(
+                    code=TraciaErrorCode.PROVIDER_ERROR,
+                    message=f"LLM provider error: {error_msg}",
+                ) from e
+
+        return generate_chunks(), result_holder, resolved_provider
+
+    async def astream(
+        self,
+        model: str,
+        messages: list[LocalPromptMessage],
+        *,
+        provider: LLMProvider | None = None,
+        temperature: float | None = None,
+        max_tokens: int | None = None,
+        top_p: float | None = None,
+        stop: list[str] | None = None,
+        tools: list[ToolDefinition] | None = None,
+        tool_choice: ToolChoice | None = None,
+        api_key: str | None = None,
+        timeout: float | None = None,
+    ) -> tuple[AsyncIterator[str], list[CompletionResult], LLMProvider]:
+        """Make an async streaming completion request.
+
+        Args:
+            model: The model name.
+            messages: The messages to send.
+            provider: The LLM provider.
+            temperature: Sampling temperature.
+            max_tokens: Maximum output tokens.
+            top_p: Top-p sampling.
+            stop: Stop sequences.
+            tools: Tool definitions.
+            tool_choice: Tool choice setting.
+            api_key: Provider API key.
+            timeout: Request timeout in seconds.
+
+        Returns:
+            A tuple of (async chunk iterator, result holder list, provider).
+
+        Raises:
+            TraciaError: If the request fails.
+        """
+        try:
+            import litellm
+        except ImportError as e:
+            raise TraciaError(
+                code=TraciaErrorCode.MISSING_PROVIDER_SDK,
+                message="litellm is not installed. Install it with: pip install litellm",
+            ) from e
+
+        resolved_provider = resolve_provider(model, provider)
+        resolved_api_key = get_provider_api_key(resolved_provider, api_key)
+
+        # Build the request
+        litellm_messages = convert_messages(messages)
+        litellm_tools = convert_tools(tools)
+        litellm_tool_choice = convert_tool_choice(tool_choice)
+
+        request_kwargs: dict[str, Any] = {
+            "model": get_litellm_model(model, resolved_provider),
+            "messages": litellm_messages,
+            "api_key": resolved_api_key,
+            "stream": True,
+        }
+
+        if temperature is not None:
+            request_kwargs["temperature"] = temperature
+        if max_tokens is not None:
+            request_kwargs["max_tokens"] = max_tokens
+        if top_p is not None:
+            request_kwargs["top_p"] = top_p
+        if stop is not None:
+            request_kwargs["stop"] = stop
+        if litellm_tools is not None:
+            request_kwargs["tools"] = litellm_tools
+        if litellm_tool_choice is not None:
+            request_kwargs["tool_choice"] = litellm_tool_choice
+        if timeout is not None:
+            request_kwargs["timeout"] = timeout
+
+        result_holder: list[CompletionResult] = []
+
+        async def generate_chunks() -> AsyncIterator[str]:
+            full_text = ""
+            input_tokens = 0
+            output_tokens = 0
+            total_tokens = 0
+            tool_calls: list[ToolCall] = []
+            finish_reason: FinishReason = "stop"
+            tool_call_chunks: dict[int, dict[str, Any]] = {}
+
+            try:
+                response = await litellm.acompletion(**request_kwargs)
+
+                async for chunk in response:
+                    choices = getattr(chunk, "choices", [])
+                    if not choices:
+                        continue
+
+                    delta = getattr(choices[0], "delta", None)
+                    if delta:
+                        content = getattr(delta, "content", None)
+                        if content:
+                            full_text += content
+                            yield content
+
+                        # Handle streaming tool calls
+                        delta_tool_calls = getattr(delta, "tool_calls", None)
+                        if delta_tool_calls:
+                            for tc in delta_tool_calls:
+                                idx = tc.index
+                                if idx not in tool_call_chunks:
+                                    tool_call_chunks[idx] = {
+                                        "id": "",
+                                        "name": "",
+                                        "arguments": "",
+                                    }
+                                if tc.id:
+                                    tool_call_chunks[idx]["id"] = tc.id
+                                if tc.function:
+                                    if tc.function.name:
+                                        tool_call_chunks[idx]["name"] = tc.function.name
+                                    if tc.function.arguments:
+                                        tool_call_chunks[idx]["arguments"] += tc.function.arguments
+
+                    chunk_finish = getattr(choices[0], "finish_reason", None)
+                    if chunk_finish:
+                        finish_reason = parse_finish_reason(chunk_finish)
+
+                    # Extract usage from final chunk
+                    usage = getattr(chunk, "usage", None)
+                    if usage:
+                        input_tokens = getattr(usage, "prompt_tokens", 0)
+                        output_tokens = getattr(usage, "completion_tokens", 0)
+                        total_tokens = getattr(usage, "total_tokens", 0)
+
+                # Convert accumulated tool calls
+                for idx in sorted(tool_call_chunks.keys()):
+                    tc_data = tool_call_chunks[idx]
+                    try:
+                        args = json.loads(tc_data["arguments"]) if tc_data["arguments"] else {}
+                    except json.JSONDecodeError:
+                        args = {}
+                    tool_calls.append(ToolCall(
+                        id=tc_data["id"],
+                        name=tc_data["name"],
+                        arguments=args,
+                    ))
+
+                result_holder.append(CompletionResult(
+                    text=full_text,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    total_tokens=total_tokens,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    provider=resolved_provider,
+                ))
+
+            except Exception as e:
+                error_msg = sanitize_error_message(str(e))
+                raise TraciaError(
+                    code=TraciaErrorCode.PROVIDER_ERROR,
+                    message=f"LLM provider error: {error_msg}",
+                ) from e
+
+        return generate_chunks(), result_holder, resolved_provider
+
+
+def build_assistant_message(
+    text: str,
+    tool_calls: list[ToolCall],
+) -> LocalPromptMessage:
+    """Build an assistant message from completion result.
+
+    Args:
+        text: The text content.
+        tool_calls: Any tool calls made.
+
+    Returns:
+        The assistant message.
+    """
+    if not tool_calls:
+        return LocalPromptMessage(role="assistant", content=text)
+
+    content: list[ContentPart] = []
+
+    if text:
+        content.append(TextPart(type="text", text=text))
+
+    for tc in tool_calls:
+        content.append(ToolCallPart(
+            type="tool_call",
+            id=tc.id,
+            name=tc.name,
+            arguments=tc.arguments,
+        ))
+
+    return LocalPromptMessage(role="assistant", content=content)