ripperdoc 0.2.3__py3-none-any.whl → 0.2.5__py3-none-any.whl

ripperdoc/core/providers/gemini.py

@@ -1,10 +1,14 @@
-"""Gemini provider client."""
+"""Gemini provider client with function/tool calling support."""
 
 from __future__ import annotations
 
+import asyncio
+import copy
+import inspect
 import os
 import time
-from typing import Any, Dict, List, Optional
+from typing import Any, AsyncIterator, Dict, List, Optional, Tuple, cast
+from uuid import uuid4
 
 from ripperdoc.core.config import ModelProfile
 from ripperdoc.core.providers.base import (
@@ -12,43 +16,358 @@ from ripperdoc.core.providers.base import (
     ProviderClient,
     ProviderResponse,
     call_with_timeout_and_retries,
+    iter_with_timeout,
 )
+from ripperdoc.core.query_utils import _normalize_tool_args, build_tool_description
 from ripperdoc.core.tool import Tool
 from ripperdoc.utils.log import get_logger
+from ripperdoc.utils.session_usage import record_usage
+from ripperdoc.core.query_utils import estimate_cost_usd
 
 logger = get_logger()
 
+# Constants
+GEMINI_SDK_IMPORT_ERROR = (
+    "Gemini client requires the 'google-genai' package. Install it with: pip install google-genai"
+)
+GEMINI_MODELS_ENDPOINT_ERROR = "Gemini client is missing 'models' endpoint"
+GEMINI_GENERATE_CONTENT_ERROR = "Gemini client is missing generate_content() method"
+
+
+def _classify_gemini_error(exc: Exception) -> tuple[str, str]:
+    """Classify a Gemini exception into error code and user-friendly message."""
+    exc_type = type(exc).__name__
+    exc_msg = str(exc)
+
+    # Try to import Google's exception types for more specific handling
+    try:
+        from google.api_core import exceptions as google_exceptions  # type: ignore
+
+        if isinstance(exc, google_exceptions.Unauthenticated):
+            return "authentication_error", f"Authentication failed: {exc_msg}"
+        if isinstance(exc, google_exceptions.PermissionDenied):
+            return "permission_denied", f"Permission denied: {exc_msg}"
+        if isinstance(exc, google_exceptions.NotFound):
+            return "model_not_found", f"Model not found: {exc_msg}"
+        if isinstance(exc, google_exceptions.InvalidArgument):
+            if "context" in exc_msg.lower() or "token" in exc_msg.lower():
+                return "context_length_exceeded", f"Context length exceeded: {exc_msg}"
+            return "bad_request", f"Invalid request: {exc_msg}"
+        if isinstance(exc, google_exceptions.ResourceExhausted):
+            return "rate_limit", f"Rate limit exceeded: {exc_msg}"
+        if isinstance(exc, google_exceptions.ServiceUnavailable):
+            return "service_unavailable", f"Service unavailable: {exc_msg}"
+        if isinstance(exc, google_exceptions.GoogleAPICallError):
+            return "api_error", f"API error: {exc_msg}"
+    except ImportError:
+        pass
+
+    # Fallback for generic exceptions
+    if isinstance(exc, asyncio.TimeoutError):
+        return "timeout", f"Request timed out: {exc_msg}"
+    if isinstance(exc, ConnectionError):
+        return "connection_error", f"Connection error: {exc_msg}"
+    if "quota" in exc_msg.lower() or "limit" in exc_msg.lower():
+        return "rate_limit", f"Rate limit exceeded: {exc_msg}"
+    if "auth" in exc_msg.lower() or "key" in exc_msg.lower():
+        return "authentication_error", f"Authentication error: {exc_msg}"
+    if "not found" in exc_msg.lower():
+        return "model_not_found", f"Model not found: {exc_msg}"
+
+    return "unknown_error", f"Unexpected error ({exc_type}): {exc_msg}"
+
 
 def _extract_usage_metadata(payload: Any) -> Dict[str, int]:
     """Best-effort token extraction from Gemini responses."""
     usage = getattr(payload, "usage_metadata", None) or getattr(payload, "usageMetadata", None)
     if not usage:
         usage = getattr(payload, "usage", None)
-    get = lambda key: int(getattr(usage, key, 0) or 0) if usage else 0  # noqa: E731
+    if not usage and getattr(payload, "candidates", None):
+        usage = getattr(payload.candidates[0], "usage_metadata", None)
+
+    def safe_get_int(key: str) -> int:
+        """Safely extract integer value from usage metadata."""
+        if not usage:
+            return 0
+        value = getattr(usage, key, 0)
+        return int(value) if value else 0
+
+    thought_tokens = safe_get_int("thoughts_token_count")
+    candidate_tokens = safe_get_int("candidates_token_count")
+
     return {
-        "input_tokens": get("prompt_token_count") + get("cached_content_token_count"),
-        "output_tokens": get("candidates_token_count"),
-        "cache_read_input_tokens": get("cached_content_token_count"),
+        "input_tokens": safe_get_int("prompt_token_count")
+        + safe_get_int("cached_content_token_count"),
+        "output_tokens": candidate_tokens + thought_tokens,
+        "cache_read_input_tokens": safe_get_int("cached_content_token_count"),
         "cache_creation_input_tokens": 0,
     }
 
 
-def _collect_text_parts(candidate: Any) -> str:
-    parts = getattr(candidate, "content", None)
-    if not parts:
-        return ""
-    if isinstance(parts, list):
-        texts = []
-        for part in parts:
-            text_val = getattr(part, "text", None) or getattr(part, "content", None)
-            if isinstance(text_val, str):
-                texts.append(text_val)
-        return "".join(texts)
-    return str(parts)
+def _collect_parts(candidate: Any) -> List[Any]:
+    """Return a list of parts from a candidate regardless of SDK shape."""
+    content = getattr(candidate, "content", None)
+    if content is None:
+        return []
+    if hasattr(content, "parts"):
+        return list(getattr(content, "parts", []) or [])
+    if isinstance(content, list):
+        return content
+    return []
+
+
+def _collect_text_from_parts(parts: List[Any]) -> str:
+    texts: List[str] = []
+    for part in parts:
+        text_val = (
+            getattr(part, "text", None)
+            or getattr(part, "content", None)
+            or getattr(part, "raw_text", None)
+        )
+        if isinstance(text_val, str):
+            texts.append(text_val)
+    return "".join(texts)
+
+
+def _extract_function_calls(parts: List[Any]) -> List[Dict[str, Any]]:
+    calls: List[Dict[str, Any]] = []
+    for part in parts:
+        fn_call = getattr(part, "function_call", None) or getattr(part, "functionCall", None)
+        if not fn_call:
+            continue
+        name = getattr(fn_call, "name", None) or getattr(fn_call, "function_name", None)
+        args = getattr(fn_call, "args", None) or getattr(fn_call, "arguments", None) or {}
+        call_id = getattr(fn_call, "id", None) or getattr(fn_call, "call_id", None)
+        calls.append({"name": name, "args": _normalize_tool_args(args), "id": call_id})
+    return calls
+
+
+def _flatten_schema(schema: Dict[str, Any]) -> Dict[str, Any]:
+    """Inline $ref entries and drop $defs/$ref for Gemini Schema compatibility.
+
+    Gemini API doesn't support JSON Schema references, so this function
+    resolves all $ref pointers by inlining the referenced definitions.
+    """
+    definitions = copy.deepcopy(schema.get("$defs") or schema.get("definitions") or {})
+
+    def _resolve(node: Any) -> Any:
+        """Recursively resolve $ref pointers and remove unsupported fields."""
+        if isinstance(node, dict):
+            # Handle $ref resolution
+            ref = node.get("$ref")
+            if isinstance(ref, str) and ref.startswith("#/"):
+                ref_key = ref.split("/")[-1]
+                if ref_key in definitions:
+                    return _resolve(copy.deepcopy(definitions[ref_key]))
+
+            # Process remaining fields, excluding schema metadata
+            resolved: Dict[str, Any] = {}
+            for key, value in node.items():
+                if key in {"$ref", "$defs", "definitions"}:
+                    continue
+                resolved[key] = _resolve(value)
+            return resolved
+
+        if isinstance(node, list):
+            return [_resolve(item) for item in node]
+
+        return node
+
+    return cast(Dict[str, Any], _resolve(copy.deepcopy(schema)))
+
+
+def _supports_stream_arg(fn: Any) -> bool:
+    """Return True if the callable appears to accept a 'stream' kwarg."""
+    try:
+        sig = inspect.signature(fn)
+    except (TypeError, ValueError):
+        # If we cannot inspect, avoid passing stream to prevent TypeErrors.
+        return False
+
+    for param in sig.parameters.values():
+        if param.kind == param.VAR_KEYWORD:
+            return True
+        if param.name == "stream":
+            return True
+    return False
+
+
+def _build_thinking_config(max_thinking_tokens: int, model_name: str) -> Dict[str, Any]:
+    """Map max_thinking_tokens to Gemini thinking_config settings."""
+    if max_thinking_tokens <= 0:
+        return {}
+    name = (model_name or "").lower()
+    config: Dict[str, Any] = {"include_thoughts": True}
+    if "gemini-3" in name:
+        config["thinking_level"] = "low" if max_thinking_tokens <= 2048 else "high"
+    else:
+        config["thinking_budget"] = max_thinking_tokens
+    return config
+
+
+def _collect_thoughts_from_parts(parts: List[Any]) -> List[str]:
+    """Extract thought summaries from parts flagged as thoughts."""
+    snippets: List[str] = []
+    for part in parts:
+        is_thought = getattr(part, "thought", None)
+        if is_thought is None and isinstance(part, dict):
+            is_thought = part.get("thought")
+        if not is_thought:
+            continue
+        text_val = (
+            getattr(part, "text", None)
+            or getattr(part, "content", None)
+            or getattr(part, "raw_text", None)
+        )
+        if isinstance(text_val, str):
+            snippets.append(text_val)
+    return snippets
+
+
+async def _async_build_tool_declarations(tools: List[Tool[Any, Any]]) -> List[Dict[str, Any]]:
+    declarations: List[Dict[str, Any]] = []
+    try:
+        from google.genai import types as genai_types  # type: ignore
+    except (ImportError, ModuleNotFoundError):  # pragma: no cover - fallback when SDK not installed
+        genai_types = None  # type: ignore[assignment]
+
+    for tool in tools:
+        description = await build_tool_description(tool, include_examples=True, max_examples=2)
+        parameters_schema = _flatten_schema(tool.input_schema.model_json_schema())
+        if genai_types:
+            func_decl = genai_types.FunctionDeclaration(
+                name=tool.name,
+                description=description,
+                parameters_json_schema=parameters_schema,
+            )
+            declarations.append(
+                func_decl.model_dump(mode="json", exclude_none=True)
+            )
+        else:
+            declarations.append(
+                {
+                    "name": tool.name,
+                    "description": description,
+                    "parameters_json_schema": parameters_schema,
+                }
+            )
+    return declarations
+
+
+def _convert_messages_to_genai_contents(
+    normalized_messages: List[Dict[str, Any]],
+) -> Tuple[List[Any], Dict[str, str]]:
+    """Map normalized OpenAI-style messages to Gemini content payloads.
+
+    Returns:
+        contents: List of Content-like dicts/objects
+        tool_name_by_id: Map of tool_call_id -> function name (for pairing responses)
+    """
+    tool_name_by_id: Dict[str, str] = {}
+    contents: List[Any] = []
+
+    # Lazy import to avoid hard dependency in tests.
+    try:
+        from google.genai import types as genai_types  # type: ignore
+    except (ImportError, ModuleNotFoundError):  # pragma: no cover - fallback when SDK not installed
+        genai_types = None  # type: ignore[assignment]
+
+    def _mk_part_from_text(text: str) -> Any:
+        if genai_types:
+            return genai_types.Part(text=text)
+        return {"text": text}
+
+    def _mk_part_from_function_call(name: str, args: Dict[str, Any], call_id: Optional[str]) -> Any:
+        # Store mapping using actual call_id if available, otherwise generate one
+        actual_id = call_id or str(uuid4())
+        tool_name_by_id[actual_id] = name
+        if genai_types:
+            return genai_types.Part(function_call=genai_types.FunctionCall(name=name, args=args))
+        return {"function_call": {"name": name, "args": args, "id": actual_id}}
+
+    def _mk_part_from_function_response(
+        name: str, response: Dict[str, Any], call_id: Optional[str]
+    ) -> Any:
+        if call_id:
+            response = {**response, "call_id": call_id}
+        if genai_types:
+            return genai_types.Part.from_function_response(name=name, response=response)
+        payload = {"function_response": {"name": name, "response": response}}
+        if call_id:
+            payload["function_response"]["id"] = call_id
+        return payload
+
+    def _mk_content(role: str, parts: List[Any]) -> Any:
+        if genai_types:
+            return genai_types.Content(role=role, parts=parts)
+        return {"role": role, "parts": parts}
+
+    for message in normalized_messages:
+        role = message.get("role") or ""
+        msg_parts: List[Any] = []
+
+        # Assistant tool calls
+        for tool_call in message.get("tool_calls") or []:
+            func = tool_call.get("function") or {}
+            name = func.get("name") or ""
+            args = _normalize_tool_args(func.get("arguments") or {})
+            call_id = tool_call.get("id")
+            msg_parts.append(_mk_part_from_function_call(name, args, call_id))
+
+        content_value = message.get("content")
+        if isinstance(content_value, str) and content_value:
+            msg_parts.append(_mk_part_from_text(content_value))
+
+        if role == "tool":
+            call_id = message.get("tool_call_id") or ""
+            name = tool_name_by_id.get(call_id, call_id or "tool_response")
+            response = {"result": content_value}
+            msg_parts.append(_mk_part_from_function_response(name, response, call_id))
+            role = "user"  # Tool responses are treated as user-provided context
+
+        if not msg_parts:
+            continue
+
+        mapped_role = "user" if role == "user" else "model"
+        contents.append(_mk_content(mapped_role, msg_parts))
+
+    return contents, tool_name_by_id
 
 
 class GeminiClient(ProviderClient):
-    """Gemini client with streaming and basic text support."""
+    """Gemini client with streaming and function calling support."""
+
+    def __init__(self, client_factory: Optional[Any] = None) -> None:
+        self._client_factory = client_factory
+
+    async def _client(self, model_profile: ModelProfile) -> Any:
+        if self._client_factory is not None:
+            client = self._client_factory
+            if inspect.iscoroutinefunction(client):
+                return await client()
+            if inspect.isawaitable(client):
+                return await client  # type: ignore[return-value]
+            if callable(client):
+                result = client()
+                return await result if inspect.isawaitable(result) else result
+            return client
+
+        try:
+            from google import genai  # type: ignore
+        except (ImportError, ModuleNotFoundError) as exc:  # pragma: no cover - import guard
+            raise RuntimeError(GEMINI_SDK_IMPORT_ERROR) from exc
+
+        client_kwargs: Dict[str, Any] = {}
+        api_key = (
+            model_profile.api_key or os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY")
+        )
+        if api_key:
+            client_kwargs["api_key"] = api_key
+        if model_profile.api_base:
+            from google.genai import types as genai_types  # type: ignore
+
+            client_kwargs["http_options"] = genai_types.HttpOptions(base_url=model_profile.api_base)
+        return genai.Client(**client_kwargs)
 
     async def call(
         self,
@@ -62,96 +381,218 @@ class GeminiClient(ProviderClient):
         progress_callback: Optional[ProgressCallback],
         request_timeout: Optional[float],
         max_retries: int,
+        max_thinking_tokens: int,
     ) -> ProviderResponse:
+        start_time = time.time()
+
         try:
-            import google.generativeai as genai  # type: ignore
-        except Exception as exc:  # pragma: no cover - import guard
-            msg = (
-                "Gemini client requires the 'google-generativeai' package. "
-                "Install it to enable Gemini support."
+            client = await self._client(model_profile)
+        except asyncio.CancelledError:
+            raise  # Don't suppress task cancellation
+        except Exception as exc:
+            duration_ms = (time.time() - start_time) * 1000
+            error_code, error_message = _classify_gemini_error(exc)
+            logger.error(
+                "[gemini_client] Initialization failed",
+                extra={
+                    "model": model_profile.model,
+                    "error_code": error_code,
+                    "error_message": error_message,
+                    "duration_ms": round(duration_ms, 2),
+                },
             )
-            logger.warning(msg, extra={"error": str(exc)})
-            return ProviderResponse(
-                content_blocks=[{"type": "text", "text": msg}],
-                usage_tokens={},
-                cost_usd=0.0,
-                duration_ms=0.0,
+            return ProviderResponse.create_error(
+                error_code=error_code,
+                error_message=error_message,
+                duration_ms=duration_ms,
             )
 
+        declarations: List[Dict[str, Any]] = []
         if tools and tool_mode != "text":
-            msg = (
-                "Gemini client currently supports text-only responses; "
-                "tool/function calling is not yet implemented."
+            declarations = await _async_build_tool_declarations(tools)
+
+        contents, _ = _convert_messages_to_genai_contents(normalized_messages)
+
+        config: Dict[str, Any] = {"system_instruction": system_prompt}
+        if model_profile.max_tokens:
+            config["max_output_tokens"] = model_profile.max_tokens
+        thinking_config = _build_thinking_config(max_thinking_tokens, model_profile.model)
+        if thinking_config:
+            try:
+                from google.genai import types as genai_types  # type: ignore
+
+                config["thinking_config"] = genai_types.ThinkingConfig(**thinking_config)
+            except (ImportError, ModuleNotFoundError, TypeError, ValueError):  # pragma: no cover - fallback when SDK not installed
+                config["thinking_config"] = thinking_config
+        if declarations:
+            config["tools"] = [{"function_declarations": declarations}]
+
+        generate_kwargs: Dict[str, Any] = {
+            "model": model_profile.model,
+            "contents": contents,
+            "config": config,
+        }
+        usage_tokens: Dict[str, int] = {}
+        collected_text: List[str] = []
+        function_calls: List[Dict[str, Any]] = []
+        reasoning_parts: List[str] = []
+        response_metadata: Dict[str, Any] = {}
+
+        async def _call_generate(streaming: bool) -> Any:
+            models_api = getattr(client, "models", None) or getattr(
+                getattr(client, "aio", None), "models", None
             )
-            return ProviderResponse(
-                content_blocks=[{"type": "text", "text": msg}],
-                usage_tokens={},
-                cost_usd=0.0,
-                duration_ms=0.0,
+            if models_api is None:
+                raise RuntimeError(GEMINI_MODELS_ENDPOINT_ERROR)
+
+            generate_fn = getattr(models_api, "generate_content", None)
+            stream_fn = getattr(models_api, "generate_content_stream", None) or getattr(
+                models_api, "stream_generate_content", None
             )
 
-        api_key = (
-            model_profile.api_key or os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY")
-        )
-        genai.configure(api_key=api_key, client_options={"api_endpoint": model_profile.api_base})
+            if streaming:
+                if stream_fn:
+                    result = stream_fn(**generate_kwargs)
+                    if inspect.isawaitable(result):
+                        return await result
+                    return result
+
+                if generate_fn is None:
+                    raise RuntimeError(GEMINI_GENERATE_CONTENT_ERROR)
 
-        # Flatten normalized messages into a single text prompt (Gemini supports multi-turn, but keep it simple).
-        prompt_parts: List[str] = [system_prompt]
-        for msg in normalized_messages:  # type: ignore[assignment]
-            role: str = (
-                str(msg.get("role", "")) if isinstance(msg, dict) else str(getattr(msg, "role", ""))  # type: ignore[assignment]
+                if _supports_stream_arg(generate_fn):
+                    gen_kwargs: Dict[str, Any] = dict(generate_kwargs)
+                    gen_kwargs["stream"] = True
+                    result = generate_fn(**gen_kwargs)
+                    if inspect.isawaitable(result):
+                        return await result
+                    return result
+
+                # Fallback: non-streaming generate; wrap to keep downstream iterator usage
+                result = generate_fn(**generate_kwargs)
+                if inspect.isawaitable(result):
+                    result = await result
+
+                async def _single_chunk_stream() -> AsyncIterator[Any]:
+                    yield result
+
+                return _single_chunk_stream()
+
+            if generate_fn is None:
+                raise RuntimeError(GEMINI_GENERATE_CONTENT_ERROR)
+
+        try:
+            if stream:
+                stream_resp = await _call_generate(streaming=True)
+
+                # Normalize streams into an async iterator to avoid StopIteration surfacing through
+                # asyncio executors and to handle sync iterables.
+                def _to_async_iter(obj: Any) -> AsyncIterator[Any]:
+                    """Convert various iterable types to async generator."""
+                    if inspect.isasyncgen(obj) or hasattr(obj, "__aiter__"):
+
+                        async def _wrap_async() -> AsyncIterator[Any]:
+                            async for item in obj:
+                                yield item
+
+                        return _wrap_async()
+                    if hasattr(obj, "__iter__"):
+
+                        async def _wrap_sync() -> AsyncIterator[Any]:
+                            for item in obj:
+                                yield item
+
+                        return _wrap_sync()
+
+                    async def _single() -> AsyncIterator[Any]:
+                        yield obj
+
+                    return _single()
+
+                stream_iter = _to_async_iter(stream_resp)
+
+                async for chunk in iter_with_timeout(stream_iter, request_timeout):
+                    candidates = getattr(chunk, "candidates", None) or []
+                    for candidate in candidates:
+                        parts = _collect_parts(candidate)
+                        text_chunk = _collect_text_from_parts(parts)
+                        if progress_callback:
+                            if text_chunk:
+                                try:
+                                    await progress_callback(text_chunk)
+                                except (RuntimeError, ValueError, TypeError, OSError) as cb_exc:
+                                    logger.warning(
+                                        "[gemini_client] Stream callback failed: %s: %s",
+                                        type(cb_exc).__name__, cb_exc,
+                                    )
+                        if text_chunk:
+                            collected_text.append(text_chunk)
+                        reasoning_parts.extend(_collect_thoughts_from_parts(parts))
+                        function_calls.extend(_extract_function_calls(parts))
+                    usage_tokens = _extract_usage_metadata(chunk) or usage_tokens
+            else:
+                # Use retry logic for non-streaming calls
+                response = await call_with_timeout_and_retries(
+                    lambda: _call_generate(streaming=False),
+                    request_timeout,
+                    max_retries,
+                )
+                candidates = getattr(response, "candidates", None) or []
+                if candidates:
+                    parts = _collect_parts(candidates[0])
+                    collected_text.append(_collect_text_from_parts(parts))
+                    reasoning_parts.extend(_collect_thoughts_from_parts(parts))
+                    function_calls.extend(_extract_function_calls(parts))
+                else:
+                    # Fallback: try to read text directly
+                    collected_text.append(getattr(response, "text", "") or "")
+                usage_tokens = _extract_usage_metadata(response)
+        except asyncio.CancelledError:
+            raise  # Don't suppress task cancellation
+        except Exception as exc:
+            duration_ms = (time.time() - start_time) * 1000
+            error_code, error_message = _classify_gemini_error(exc)
+            logger.error(
+                "[gemini_client] API call failed",
+                extra={
+                    "model": model_profile.model,
+                    "error_code": error_code,
+                    "error_message": error_message,
+                    "duration_ms": round(duration_ms, 2),
+                },
+            )
+            return ProviderResponse.create_error(
+                error_code=error_code,
+                error_message=error_message,
+                duration_ms=duration_ms,
             )
-            content = msg.get("content") if isinstance(msg, dict) else getattr(msg, "content", "")
-            if isinstance(content, list):
-                for item in content:
-                    text_val = (
-                        getattr(item, "text", None)
-                        or item.get("text", "")  # type: ignore[union-attr]
-                        if isinstance(item, dict)
-                        else ""
-                    )
-                    if text_val:
-                        prompt_parts.append(f"{role}: {text_val}")
-            elif isinstance(content, str):
-                prompt_parts.append(f"{role}: {content}")
-        full_prompt = "\n".join(part for part in prompt_parts if part)
-
-        model = genai.GenerativeModel(model_profile.model)
-        collected_text: List[str] = []
-        start_time = time.time()
 
-        async def _stream_request() -> Dict[str, Dict[str, int]]:
-            stream_resp = model.generate_content(full_prompt, stream=True)
-            usage_tokens: Dict[str, int] = {}
-            for chunk in stream_resp:
-                text_delta = _collect_text_parts(chunk)
-                if text_delta:
-                    collected_text.append(text_delta)
-                    if progress_callback:
-                        try:
-                            await progress_callback(text_delta)
-                        except Exception:
-                            logger.exception("[gemini_client] Stream callback failed")
-                usage_tokens = _extract_usage_metadata(chunk) or usage_tokens
-            return {"usage": usage_tokens}
-
-        async def _non_stream_request() -> Any:
-            return model.generate_content(full_prompt)
-
-        response: Any = await call_with_timeout_and_retries(
-            _stream_request if stream and progress_callback else _non_stream_request,
-            request_timeout,
-            max_retries,
-        )
+        content_blocks: List[Dict[str, Any]] = []
+        combined_text = "".join(collected_text).strip()
+        if combined_text:
+            content_blocks.append({"type": "text", "text": combined_text})
+        if reasoning_parts:
+            response_metadata["reasoning_content"] = "".join(reasoning_parts)
 
-        duration_ms = (time.time() - start_time) * 1000
-        usage_tokens = _extract_usage_metadata(response)
-        cost_usd = 0.0  # Pricing unknown; leave as 0
+        for call in function_calls:
+            if not call.get("name"):
+                continue
+            content_blocks.append(
+                {
+                    "type": "tool_use",
+                    "tool_use_id": call.get("id") or str(uuid4()),
+                    "name": call["name"],
+                    "input": call.get("args") or {},
+                }
+            )
 
-        content_blocks = (
-            [{"type": "text", "text": "".join(collected_text)}]
-            if collected_text
-            else [{"type": "text", "text": _collect_text_parts(response)}]
+        duration_ms = (time.time() - start_time) * 1000
+        cost_usd = estimate_cost_usd(model_profile, usage_tokens) if usage_tokens else 0.0
+        record_usage(
+            model_profile.model,
+            duration_ms=duration_ms,
+            cost_usd=cost_usd,
+            **(usage_tokens or {}),
         )
 
         logger.info(
@@ -161,12 +602,14 @@ class GeminiClient(ProviderClient):
                 "duration_ms": round(duration_ms, 2),
                 "tool_mode": tool_mode,
                 "stream": stream,
+                "function_call_count": len(function_calls),
             },
         )
 
         return ProviderResponse(
-            content_blocks=content_blocks,
+            content_blocks=content_blocks or [{"type": "text", "text": ""}],
             usage_tokens=usage_tokens,
             cost_usd=cost_usd,
             duration_ms=duration_ms,
+            metadata=response_metadata,
         )