codepp 0.0.437__py3-none-any.whl

code_puppy/gemini_model.py

@@ -0,0 +1,754 @@
+"""Standalone Gemini Model for pydantic_ai - no google-genai dependency.
+
+This module provides a custom Model implementation that uses Google's
+Generative Language API directly via httpx, without the bloated google-genai
+SDK dependency.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import logging
+import uuid
+from collections.abc import AsyncIterator
+from contextlib import asynccontextmanager
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Any
+
+import httpx
+from pydantic_ai._run_context import RunContext
+from pydantic_ai.messages import (
+    ModelMessage,
+    ModelRequest,
+    ModelResponse,
+    ModelResponsePart,
+    ModelResponseStreamEvent,
+    RetryPromptPart,
+    SystemPromptPart,
+    TextPart,
+    ThinkingPart,
+    ToolCallPart,
+    ToolReturnPart,
+    UserPromptPart,
+)
+from pydantic_ai.models import Model, ModelRequestParameters, StreamedResponse
+from pydantic_ai.settings import ModelSettings
+from pydantic_ai.tools import ToolDefinition
+from pydantic_ai.usage import RequestUsage
+
+logger = logging.getLogger(__name__)
+
+# Bypass thought signature for Gemini when no pending signature is available.
+# This allows function calls to work with thinking models.
+BYPASS_THOUGHT_SIGNATURE = "context_engineering_is_the_way_to_go"
+
+
+def generate_tool_call_id() -> str:
+    """Generate a unique tool call ID."""
+    return str(uuid.uuid4())
+
+
+def _flatten_union_to_object_gemini(union_items: list, defs: dict, resolve_fn) -> dict:
+    """Flatten a union of object types into a single object with all properties.
+
+    For discriminated unions like EditFilePayload, we merge all object types
+    into one with all properties (Gemini doesn't support anyOf/oneOf).
+    """
+    import copy as copy_module
+
+    merged_properties = {}
+    has_string_type = False
+
+    for item in union_items:
+        if not isinstance(item, dict):
+            continue
+
+        # Resolve $ref first
+        if "$ref" in item:
+            ref_path = item["$ref"]
+            ref_name = None
+            if ref_path.startswith("#/$defs/"):
+                ref_name = ref_path[8:]
+            elif ref_path.startswith("#/definitions/"):
+                ref_name = ref_path[14:]
+            if ref_name and ref_name in defs:
+                item = copy_module.deepcopy(defs[ref_name])
+            else:
+                continue
+
+        if item.get("type") == "string":
+            has_string_type = True
+            continue
+
+        if item.get("type") == "null":
+            continue
+
+        if item.get("type") == "object" or "properties" in item:
+            props = item.get("properties", {})
+            for prop_name, prop_schema in props.items():
+                if prop_name not in merged_properties:
+                    merged_properties[prop_name] = resolve_fn(
+                        copy_module.deepcopy(prop_schema)
+                    )
+
+    if not merged_properties:
+        return {"type": "string"} if has_string_type else {"type": "object"}
+
+    return {
+        "type": "object",
+        "properties": merged_properties,
+    }
+
+
+def _sanitize_schema_for_gemini(schema: dict) -> dict:
+    """Sanitize JSON schema for Gemini API compatibility.
+
+    Removes/transforms fields that Gemini doesn't support:
+    - $defs, definitions, $schema, $id
+    - additionalProperties
+    - $ref (inlined)
+    - anyOf/oneOf/allOf (flattened - Gemini doesn't support unions!)
+      - For unions of objects: merges into single object with all properties
+      - For simple unions (string | null): picks first non-null type
+    """
+    import copy
+
+    if not isinstance(schema, dict):
+        return schema
+
+    # Make a deep copy to avoid modifying original
+    schema = copy.deepcopy(schema)
+
+    # Extract $defs for reference resolution
+    defs = schema.pop("$defs", schema.pop("definitions", {}))
+
+    def resolve_refs(obj):
+        """Recursively resolve $ref references and clean schema."""
+        if isinstance(obj, dict):
+            # Handle anyOf/oneOf unions
+            for union_key in ["anyOf", "oneOf"]:
+                if union_key in obj:
+                    union = obj[union_key]
+                    if isinstance(union, list):
+                        # Check if this is a complex union of objects
+                        object_count = 0
+                        has_refs = False
+                        for item in union:
+                            if isinstance(item, dict):
+                                if "$ref" in item:
+                                    has_refs = True
+                                    object_count += 1
+                                elif (
+                                    item.get("type") == "object" or "properties" in item
+                                ):
+                                    object_count += 1
+
+                        # If multiple objects or has refs, flatten to single object
+                        if object_count > 1 or has_refs:
+                            flattened = _flatten_union_to_object_gemini(
+                                union, defs, resolve_refs
+                            )
+                            if "description" in obj:
+                                flattened["description"] = obj["description"]
+                            return flattened
+
+                        # Simple union - pick first non-null type
+                        for item in union:
+                            if isinstance(item, dict) and item.get("type") != "null":
+                                result = dict(item)
+                                if "description" in obj:
+                                    result["description"] = obj["description"]
+                                return resolve_refs(result)
+
+            # Handle allOf by merging all schemas
+            if "allOf" in obj:
+                all_of = obj["allOf"]
+                if isinstance(all_of, list):
+                    merged = {}
+                    merged_properties = {}
+                    for item in all_of:
+                        if isinstance(item, dict):
+                            resolved_item = resolve_refs(item)
+                            if "properties" in resolved_item:
+                                merged_properties.update(
+                                    resolved_item.pop("properties")
+                                )
+                            merged.update(resolved_item)
+                    if merged_properties:
+                        merged["properties"] = merged_properties
+                    for k, v in obj.items():
+                        if k != "allOf":
+                            merged[k] = v
+                    return resolve_refs(merged)
+
+            # Check for $ref
+            if "$ref" in obj:
+                ref_path = obj["$ref"]
+                ref_name = None
+
+                # Parse ref like "#/$defs/SomeType" or "#/definitions/SomeType"
+                if ref_path.startswith("#/$defs/"):
+                    ref_name = ref_path[8:]
+                elif ref_path.startswith("#/definitions/"):
+                    ref_name = ref_path[14:]
+
+                if ref_name and ref_name in defs:
+                    resolved = resolve_refs(copy.deepcopy(defs[ref_name]))
+                    other_props = {k: v for k, v in obj.items() if k != "$ref"}
+                    if other_props:
+                        resolved.update(resolve_refs(other_props))
+                    return resolved
+                else:
+                    return {"type": "object"}
+
+            # Recursively process and transform
+            result = {}
+            for key, value in obj.items():
+                # Skip unsupported fields
+                if key in (
+                    "$defs",
+                    "definitions",
+                    "$schema",
+                    "$id",
+                    "additionalProperties",
+                    "default",
+                    "examples",
+                    "const",
+                    "anyOf",  # Skip any remaining union types
+                    "oneOf",
+                    "allOf",
+                ):
+                    continue
+
+                result[key] = resolve_refs(value)
+            return result
+        elif isinstance(obj, list):
+            return [resolve_refs(item) for item in obj]
+        else:
+            return obj
+
+    return resolve_refs(schema)
+
+
+class GeminiModel(Model):
+    """Standalone Model implementation for Google's Generative Language API.
+
+    Uses httpx directly instead of google-genai SDK.
+    """
+
+    def __init__(
+        self,
+        model_name: str,
+        api_key: str,
+        base_url: str = "https://generativelanguage.googleapis.com/v1beta",
+        http_client: httpx.AsyncClient | None = None,
+    ):
+        self._model_name = model_name
+        self.api_key = api_key
+        self._base_url = base_url.rstrip("/")
+        self._http_client = http_client
+        self._owns_client = http_client is None
+
+    @property
+    def model_name(self) -> str:
+        """Return the model name."""
+        return self._model_name
+
+    @property
+    def base_url(self) -> str:
+        """Return the base URL for the API."""
+        return self._base_url
+
+    @property
+    def system(self) -> str:
+        """Return the provider system identifier."""
+        return "google"
+
+    def _get_instructions(
+        self,
+        messages: list,
+        model_request_parameters,
+    ) -> str | None:
+        """Get additional instructions to prepend to system prompt.
+
+        This is a compatibility method for pydantic-ai interface.
+        Override in subclasses to inject custom instructions.
+        """
+        return None
+
+    def prepare_request(
+        self,
+        model_settings: ModelSettings | None,
+        model_request_parameters,
+    ) -> tuple:
+        """Prepare request by normalizing settings.
+
+        This is a compatibility method for pydantic-ai interface.
+        """
+        return model_settings, model_request_parameters
+
+    async def _get_client(self) -> httpx.AsyncClient:
+        """Get or create HTTP client."""
+        if self._http_client is None:
+            self._http_client = httpx.AsyncClient(timeout=180)
+        return self._http_client
+
+    async def _close_client(self) -> None:
+        """Close HTTP client if we own it."""
+        if self._owns_client and self._http_client is not None:
+            await self._http_client.aclose()
+            self._http_client = None
+
+    def _get_headers(self) -> dict[str, str]:
+        """Get HTTP headers for the request."""
+        return {
+            "Content-Type": "application/json",
+            "Accept": "application/json",
+            "x-goog-api-key": self.api_key,
+        }
+
+    async def _map_user_prompt(self, part: UserPromptPart) -> list[dict[str, Any]]:
+        """Map a user prompt part to Gemini format."""
+        parts = []
+
+        if isinstance(part.content, str):
+            parts.append({"text": part.content})
+        elif isinstance(part.content, list):
+            for item in part.content:
+                if isinstance(item, str):
+                    parts.append({"text": item})
+                elif hasattr(item, "media_type") and hasattr(item, "data"):
+                    # Handle file/image content
+                    data = item.data
+                    if isinstance(data, bytes):
+                        data = base64.b64encode(data).decode("utf-8")
+                    parts.append(
+                        {
+                            "inline_data": {
+                                "mime_type": item.media_type,
+                                "data": data,
+                            }
+                        }
+                    )
+                else:
+                    parts.append({"text": str(item)})
+        else:
+            parts.append({"text": str(part.content)})
+
+        return parts
+
+    async def _map_messages(
+        self,
+        messages: list[ModelMessage],
+        model_request_parameters: ModelRequestParameters,
+    ) -> tuple[dict[str, Any] | None, list[dict[str, Any]]]:
+        """Map pydantic-ai messages to Gemini API format."""
+        contents: list[dict[str, Any]] = []
+        system_parts: list[dict[str, Any]] = []
+
+        for m in messages:
+            if isinstance(m, ModelRequest):
+                message_parts: list[dict[str, Any]] = []
+
+                for part in m.parts:
+                    if isinstance(part, SystemPromptPart):
+                        system_parts.append({"text": part.content})
+                    elif isinstance(part, UserPromptPart):
+                        mapped_parts = await self._map_user_prompt(part)
+                        message_parts.extend(mapped_parts)
+                    elif isinstance(part, ToolReturnPart):
+                        message_parts.append(
+                            {
+                                "function_response": {
+                                    "name": part.tool_name,
+                                    "response": part.model_response_object(),
+                                    "id": part.tool_call_id,
+                                }
+                            }
+                        )
+                    elif isinstance(part, RetryPromptPart):
+                        if part.tool_name is None:
+                            message_parts.append({"text": part.model_response()})
+                        else:
+                            message_parts.append(
+                                {
+                                    "function_response": {
+                                        "name": part.tool_name,
+                                        "response": {"error": part.model_response()},
+                                        "id": part.tool_call_id,
+                                    }
+                                }
+                            )
+
+                if message_parts:
+                    # Merge with previous user message if exists
+                    if contents and contents[-1].get("role") == "user":
+                        contents[-1]["parts"].extend(message_parts)
+                    else:
+                        contents.append({"role": "user", "parts": message_parts})
+
+            elif isinstance(m, ModelResponse):
+                model_parts = self._map_model_response(m)
+                if model_parts:
+                    # Merge with previous model message if exists
+                    if contents and contents[-1].get("role") == "model":
+                        contents[-1]["parts"].extend(model_parts["parts"])
+                    else:
+                        contents.append(model_parts)
+
+        # Ensure at least one content
+        if not contents:
+            contents = [{"role": "user", "parts": [{"text": ""}]}]
+
+        # Get any injected instructions
+        instructions = self._get_instructions(messages, model_request_parameters)
+        if instructions:
+            system_parts.insert(0, {"text": instructions})
+
+        # Build system instruction
+        system_instruction = None
+        if system_parts:
+            system_instruction = {"role": "user", "parts": system_parts}
+
+        return system_instruction, contents
+
+    def _map_model_response(self, m: ModelResponse) -> dict[str, Any] | None:
+        """Map a ModelResponse to Gemini content format.
+
+        For Gemini thinking models, we need to track thought signatures from
+        ThinkingParts and apply them to subsequent function_call parts.
+        """
+        parts: list[dict[str, Any]] = []
+        pending_signature: str | None = None
+
+        for item in m.parts:
+            if isinstance(item, ToolCallPart):
+                part_dict: dict[str, Any] = {
+                    "function_call": {
+                        "name": item.tool_name,
+                        "args": item.args_as_dict(),
+                        "id": item.tool_call_id,
+                    }
+                }
+                # Gemini thinking models REQUIRE thoughtSignature on function calls
+                # Use pending signature from thinking or bypass signature
+                part_dict["thoughtSignature"] = (
+                    pending_signature
+                    if pending_signature is not None
+                    else BYPASS_THOUGHT_SIGNATURE
+                )
+                parts.append(part_dict)
+            elif isinstance(item, TextPart):
+                part_dict = {"text": item.content}
+                # Apply pending signature to text parts too if present
+                if pending_signature is not None:
+                    part_dict["thoughtSignature"] = pending_signature
+                    pending_signature = None
+                parts.append(part_dict)
+            elif isinstance(item, ThinkingPart):
+                if item.content:
+                    part_dict = {"text": item.content, "thought": True}
+                    if item.signature:
+                        part_dict["thoughtSignature"] = item.signature
+                        # Store signature for subsequent parts
+                        pending_signature = item.signature
+                    else:
+                        # No signature on thinking part, use bypass
+                        pending_signature = BYPASS_THOUGHT_SIGNATURE
+                    parts.append(part_dict)
+
+        if not parts:
+            return None
+        return {"role": "model", "parts": parts}
+
+    def _build_tools(self, tools: list[ToolDefinition]) -> list[dict[str, Any]]:
+        """Build tool definitions for the API."""
+        function_declarations = []
+
+        for tool in tools:
+            func_decl: dict[str, Any] = {
+                "name": tool.name,
+                "description": tool.description or "",
+            }
+            if tool.parameters_json_schema:
+                # Sanitize schema for Gemini compatibility
+                func_decl["parameters"] = _sanitize_schema_for_gemini(
+                    tool.parameters_json_schema
+                )
+            function_declarations.append(func_decl)
+
+        return [{"functionDeclarations": function_declarations}]
+
+    def _build_generation_config(
+        self, model_settings: ModelSettings | None
+    ) -> dict[str, Any]:
+        """Build generation config from model settings."""
+        config: dict[str, Any] = {}
+
+        if model_settings:
+            # ModelSettings is a TypedDict, so use .get() for all access
+            temperature = model_settings.get("temperature")
+            if temperature is not None:
+                config["temperature"] = temperature
+
+            top_p = model_settings.get("top_p")
+            if top_p is not None:
+                config["topP"] = top_p
+
+            max_tokens = model_settings.get("max_tokens")
+            if max_tokens is not None:
+                config["maxOutputTokens"] = max_tokens
+
+            # Handle Gemini 3 Pro thinking settings
+            thinking_enabled = model_settings.get("thinking_enabled")
+            thinking_level = model_settings.get("thinking_level")
+
+            # Build thinkingConfig if thinking settings are present
+            if thinking_enabled is False:
+                # Disable thinking by not including thinkingConfig
+                pass
+            elif thinking_level is not None:
+                # Gemini 3 Pro uses thinkingLevel with values "low" or "high"
+                # includeThoughts=True is required to surface the thinking in the response
+                config["thinkingConfig"] = {
+                    "thinkingLevel": thinking_level,
+                    "includeThoughts": True,
+                }
+
+        return config
+
+    async def request(
+        self,
+        messages: list[ModelMessage],
+        model_settings: ModelSettings | None,
+        model_request_parameters: ModelRequestParameters,
+    ) -> ModelResponse:
+        """Make a non-streaming request to the Gemini API."""
+        system_instruction, contents = await self._map_messages(
+            messages, model_request_parameters
+        )
+
+        # Build request body
+        body: dict[str, Any] = {"contents": contents}
+
+        gen_config = self._build_generation_config(model_settings)
+        if gen_config:
+            body["generationConfig"] = gen_config
+        if system_instruction:
+            body["systemInstruction"] = system_instruction
+
+        # Add tools
+        if model_request_parameters.function_tools:
+            body["tools"] = self._build_tools(model_request_parameters.function_tools)
+
+        # Make request
+        client = await self._get_client()
+        url = f"{self._base_url}/models/{self._model_name}:generateContent"
+        headers = self._get_headers()
+
+        response = await client.post(url, json=body, headers=headers)
+
+        if response.status_code != 200:
+            raise RuntimeError(
+                f"Gemini API error {response.status_code}: {response.text}"
+            )
+
+        data = response.json()
+        return self._parse_response(data)
+
+    def _parse_response(self, data: dict[str, Any]) -> ModelResponse:
+        """Parse the Gemini API response."""
+        candidates = data.get("candidates", [])
+        if not candidates:
+            return ModelResponse(
+                parts=[TextPart(content="")],
+                model_name=self._model_name,
+                usage=RequestUsage(),
+            )
+
+        candidate = candidates[0]
+        content = candidate.get("content", {})
+        parts = content.get("parts", [])
+
+        response_parts: list[ModelResponsePart] = []
+
+        for part in parts:
+            if part.get("thought") and part.get("text") is not None:
+                # Thinking part
+                signature = part.get("thoughtSignature")
+                response_parts.append(
+                    ThinkingPart(content=part["text"], signature=signature)
+                )
+            elif "text" in part:
+                response_parts.append(TextPart(content=part["text"]))
+            elif "functionCall" in part:
+                fc = part["functionCall"]
+                response_parts.append(
+                    ToolCallPart(
+                        tool_name=fc["name"],
+                        args=fc.get("args", {}),
+                        tool_call_id=fc.get("id") or generate_tool_call_id(),
+                    )
+                )
+
+        # Extract usage
+        usage_meta = data.get("usageMetadata", {})
+        usage = RequestUsage(
+            input_tokens=usage_meta.get("promptTokenCount", 0),
+            output_tokens=usage_meta.get("candidatesTokenCount", 0),
+        )
+
+        return ModelResponse(
+            parts=response_parts,
+            model_name=self._model_name,
+            usage=usage,
+            provider_response_id=data.get("requestId"),
+            provider_name=self.system,
+        )
+
+    @asynccontextmanager
+    async def request_stream(
+        self,
+        messages: list[ModelMessage],
+        model_settings: ModelSettings | None,
+        model_request_parameters: ModelRequestParameters,
+        run_context: RunContext[Any] | None = None,
+    ) -> AsyncIterator[StreamedResponse]:
+        """Make a streaming request to the Gemini API."""
+        system_instruction, contents = await self._map_messages(
+            messages, model_request_parameters
+        )
+
+        # Build request body
+        body: dict[str, Any] = {"contents": contents}
+
+        gen_config = self._build_generation_config(model_settings)
+        if gen_config:
+            body["generationConfig"] = gen_config
+        if system_instruction:
+            body["systemInstruction"] = system_instruction
+
+        # Add tools
+        if model_request_parameters.function_tools:
+            body["tools"] = self._build_tools(model_request_parameters.function_tools)
+
+        # Make streaming request
+        client = await self._get_client()
+        url = (
+            f"{self._base_url}/models/{self._model_name}:streamGenerateContent?alt=sse"
+        )
+        headers = self._get_headers()
+
+        async def stream_chunks() -> AsyncIterator[dict[str, Any]]:
+            async with client.stream(
+                "POST", url, json=body, headers=headers
+            ) as response:
+                if response.status_code != 200:
+                    text = await response.aread()
+                    raise RuntimeError(
+                        f"Gemini API error {response.status_code}: {text.decode()}"
+                    )
+
+                async for line in response.aiter_lines():
+                    line = line.strip()
+                    if not line:
+                        continue
+                    if line.startswith("data: "):
+                        json_str = line[6:]
+                        if json_str:
+                            try:
+                                yield json.loads(json_str)
+                            except json.JSONDecodeError:
+                                continue
+
+        yield GeminiStreamingResponse(
+            model_request_parameters=model_request_parameters,
+            _chunks=stream_chunks(),
+            _model_name_str=self._model_name,
+            _provider_name_str=self.system,
+            _provider_url_str=self._base_url,
+        )
+
+
+@dataclass
+class GeminiStreamingResponse(StreamedResponse):
+    """Streaming response handler for Gemini API."""
+
+    _chunks: AsyncIterator[dict[str, Any]]
+    _model_name_str: str
+    _provider_name_str: str = "google"
+    _provider_url_str: str | None = None
+    _timestamp_val: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
+
+    async def _get_event_iterator(self) -> AsyncIterator[ModelResponseStreamEvent]:
+        """Process streaming chunks and yield events."""
+        async for chunk in self._chunks:
+            # Extract usage
+            usage_meta = chunk.get("usageMetadata", {})
+            if usage_meta:
+                self._usage = RequestUsage(
+                    input_tokens=usage_meta.get("promptTokenCount", 0),
+                    output_tokens=usage_meta.get("candidatesTokenCount", 0),
+                )
+
+            # Extract response ID
+            if chunk.get("responseId"):
+                self.provider_response_id = chunk["responseId"]
+
+            candidates = chunk.get("candidates", [])
+            if not candidates:
+                continue
+
+            candidate = candidates[0]
+            content = candidate.get("content", {})
+            parts = content.get("parts", [])
+
+            for part in parts:
+                # Handle thinking part
+                if part.get("thought") and part.get("text") is not None:
+                    for event in self._parts_manager.handle_thinking_delta(
+                        vendor_part_id=None,
+                        content=part["text"],
+                    ):
+                        yield event
+
+                # Handle regular text
+                elif part.get("text") is not None and not part.get("thought"):
+                    text = part["text"]
+                    if len(text) == 0:
+                        continue
+                    for event in self._parts_manager.handle_text_delta(
+                        vendor_part_id=None,
+                        content=text,
+                    ):
+                        yield event
+
+                # Handle function call
+                elif part.get("functionCall"):
+                    fc = part["functionCall"]
+                    event = self._parts_manager.handle_tool_call_delta(
+                        vendor_part_id=uuid.uuid4(),
+                        tool_name=fc.get("name"),
+                        args=fc.get("args"),
+                        tool_call_id=fc.get("id") or generate_tool_call_id(),
+                    )
+                    if event is not None:
+                        yield event
+
+    @property
+    def model_name(self) -> str:
+        return self._model_name_str
+
+    @property
+    def provider_name(self) -> str | None:
+        return self._provider_name_str
+
+    @property
+    def provider_url(self) -> str | None:
+        return self._provider_url_str
+
+    @property
+    def timestamp(self) -> datetime:
+        return self._timestamp_val