zero-agent-core 0.2.0__py3-none-any.whl

agent_core/model/providers/anthropic.py

@@ -0,0 +1,294 @@
+"""适配器 2：Anthropic Claude。"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import AsyncIterator
+from typing import Any
+
+from agent_core.errors import ModelInvocationError
+from agent_core.model.core import ContextUsage, StreamChunk
+from agent_core.model.providers.common import known_context_window, logger
+from agent_core.protocol.messages import Message, assistant_message
+
+
+class AnthropicProvider:
+    """Anthropic Claude 提供商。
+
+    需要安装：uv add anthropic
+    """
+
+    def __init__(
+        self,
+        *,
+        api_key: str,
+        model: str = "claude-3-5-sonnet-20241022",
+        temperature: float = 0.0,
+        max_tokens: int = 4096,
+        context_window_tokens: int | None = None,
+    ) -> None:
+        try:
+            import anthropic
+
+            self._client = anthropic.AsyncAnthropic(api_key=api_key)
+        except ImportError as exc:
+            raise ImportError("AnthropicProvider 需要 anthropic 包，请执行：uv add anthropic") from exc
+
+        self.model = model
+        self._temperature = temperature
+        self._max_tokens = max_tokens
+        self._context_window_tokens = context_window_tokens
+        self.provider_name = "anthropic"
+
+    # ──────────────────────────────────────────────
+    # 上下文使用
+    # ──────────────────────────────────────────────
+    async def context_usage(
+        self,
+        messages: list[Message],
+        *,
+        tools: list[dict[str, Any]] | None = None,
+    ) -> ContextUsage:
+        system_prompt, anthropic_msgs = self._to_anthropic_messages(messages)
+        try:
+            count_tokens = self._client.messages.count_tokens
+            kwargs: dict[str, Any] = {
+                "model": self.model,
+                "messages": anthropic_msgs,
+            }
+            if system_prompt:
+                kwargs["system"] = system_prompt
+            if tools:
+                kwargs["tools"] = self._to_anthropic_tools(tools)
+            response = await count_tokens(**kwargs)
+            input_tokens = getattr(response, "input_tokens", None)
+            if isinstance(input_tokens, int):
+                return ContextUsage(
+                    input_tokens=input_tokens,
+                    context_window_tokens=(self._context_window_tokens or known_context_window(self.model)),
+                    exact=True,
+                    source="provider_tokenizer",
+                )
+        except Exception as exc:
+            logger.debug("Anthropic tokenizer 不可用：%s", exc)
+        return ContextUsage(
+            input_tokens=None,
+            context_window_tokens=self._context_window_tokens or known_context_window(self.model),
+            exact=False,
+            source="unavailable",
+        )
+
+    # ──────────────────────────────────────────────
+    # 转换为 Anthropic API 格式
+    # ──────────────────────────────────────────────
+    def _to_anthropic_messages(self, messages: list[Message]) -> tuple[str | None, list[dict[str, Any]]]:
+        """将 Message 列表转为 Anthropic API 格式。
+
+        Returns:
+            (system_prompt, messages_list)
+        """
+        system_parts: list[str] = []
+        anthropic_msgs: list[dict[str, Any]] = []
+
+        def append_user_content(content: str | list[dict[str, Any]]) -> None:
+            """合并连续 user 消息，满足 Anthropic 的消息交替约束。"""
+            if anthropic_msgs and anthropic_msgs[-1]["role"] == "user":
+                existing = anthropic_msgs[-1]["content"]
+                if isinstance(existing, str):
+                    existing = [{"type": "text", "text": existing}]
+                    anthropic_msgs[-1]["content"] = existing
+                if isinstance(content, str):
+                    existing.append({"type": "text", "text": content})
+                else:
+                    existing.extend(content)
+                return
+            anthropic_msgs.append({"role": "user", "content": content})
+
+        for msg in messages:
+            if msg.role == "system":
+                if msg.content:
+                    system_parts.append(msg.content)
+            elif msg.role == "user":
+                append_user_content(msg.content)
+            elif msg.role == "assistant":
+                content: list[dict[str, Any]] = []
+                if msg.content:
+                    content.append({"type": "text", "text": msg.content})
+                if msg.tool_calls:
+                    for tc in msg.tool_calls:
+                        content.append(
+                            {
+                                "type": "tool_use",
+                                "id": tc["id"],
+                                "name": tc["name"],
+                                "input": tc.get("args", {}),
+                            }
+                        )
+                if content:
+                    anthropic_msgs.append({"role": "assistant", "content": content})
+            elif msg.role == "tool":
+                append_user_content(
+                    [
+                        {
+                            "type": "tool_result",
+                            "tool_use_id": msg.tool_call_id,
+                            "content": msg.content,
+                        }
+                    ]
+                )
+
+        return "\n\n".join(system_parts) or None, anthropic_msgs
+
+    # ──────────────────────────────────────────────
+    # 转换为 Anthropic API 格式工具定义
+    # ──────────────────────────────────────────────
+    def _to_anthropic_tools(self, tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """将 OpenAI 格式工具定义转为 Anthropic 格式。"""
+        result = []
+        for t in tools:
+            fn = t.get("function", t)
+            result.append(
+                {
+                    "name": fn["name"],
+                    "description": fn.get("description", ""),
+                    "input_schema": fn.get("parameters", {"type": "object", "properties": {}}),
+                }
+            )
+        return result
+
+    # ──────────────────────────────────────────────
+    # 调用 Anthropic API 进行 chat completion
+    # ──────────────────────────────────────────────
+    async def chat(
+        self,
+        messages: list[Message],
+        *,
+        tools: list[dict[str, Any]] | None = None,
+        temperature: float | None = None,
+        max_tokens: int | None = None,
+    ) -> Message:
+        system_prompt, anthropic_msgs = self._to_anthropic_messages(messages)
+
+        kwargs: dict[str, Any] = {
+            "model": self.model,
+            "max_tokens": max_tokens or self._max_tokens,
+            "temperature": temperature if temperature is not None else self._temperature,
+            "messages": anthropic_msgs,
+        }
+        if system_prompt:
+            kwargs["system"] = system_prompt
+        if tools:
+            kwargs["tools"] = self._to_anthropic_tools(tools)
+
+        try:
+            response = await self._client.messages.create(**kwargs)
+        except Exception as exc:
+            raise ModelInvocationError(f"Anthropic API call failed: {exc}") from exc
+
+        # 解析响应
+        content_text = ""
+        tool_calls: list[dict[str, Any]] = []
+
+        for block in response.content:
+            if block.type == "text":
+                content_text += block.text
+            elif block.type == "tool_use":
+                tool_calls.append(
+                    {
+                        "id": block.id,
+                        "type": "function",
+                        "name": block.name,
+                        "args": block.input,
+                    }
+                )
+
+        logger.debug(
+            "Anthropic response: content_len=%d, tool_calls=%d, stop=%s",
+            len(content_text),
+            len(tool_calls),
+            response.stop_reason,
+        )
+        return assistant_message(content_text, tool_calls=tool_calls)
+
+    # ──────────────────────────────────────────────
+    # 调用 Anthropic API 进行流式 chat completion
+    # ──────────────────────────────────────────────
+    async def stream_chat(
+        self,
+        messages: list[Message],
+        *,
+        tools: list[dict[str, Any]] | None = None,
+        temperature: float | None = None,
+        max_tokens: int | None = None,
+    ) -> AsyncIterator[StreamChunk]:
+        system_prompt, anthropic_msgs = self._to_anthropic_messages(messages)
+
+        kwargs: dict[str, Any] = {
+            "model": self.model,
+            "max_tokens": max_tokens or self._max_tokens,
+            "temperature": temperature if temperature is not None else self._temperature,
+            "messages": anthropic_msgs,
+        }
+        if system_prompt:
+            kwargs["system"] = system_prompt
+        if tools:
+            kwargs["tools"] = self._to_anthropic_tools(tools)
+
+        tool_call_accum: dict[str, dict[str, Any]] = {}
+
+        try:
+            async with self._client.messages.stream(**kwargs) as stream:
+                async for event in stream:
+                    # 不依赖 SDK 私有事件类名，兼容 anthropic 不同版本的事件对象。
+                    if hasattr(event, "content_block"):
+                        block = event.content_block
+                        if getattr(block, "type", None) == "tool_use":
+                            block_id = str(getattr(event, "index", len(tool_call_accum)))
+                            tool_call_accum[block_id] = {
+                                "id": getattr(block, "id", ""),
+                                "name": getattr(block, "name", ""),
+                                "args_json": "",
+                            }
+
+                    if hasattr(event, "delta"):
+                        delta = event.delta
+                        # 取值经 getattr 而非属性直访：delta 在不同 SDK 版本下的
+                        # 类型联合不同（mypy 按 union-attr 检查），鸭子类型取值
+                        # 既保持跨版本兼容，也让类型检查可静态通过。
+                        delta_text = getattr(delta, "text", None)
+                        partial_json = getattr(delta, "partial_json", None)
+                        if delta_text:
+                            yield StreamChunk(text=delta_text)
+                        elif partial_json is not None:
+                            # 工具调用参数增量
+                            block_id = str(getattr(event, "index", len(tool_call_accum)))
+                            if block_id not in tool_call_accum:
+                                tool_call_accum[block_id] = {"args_json": ""}
+                            tool_call_accum[block_id]["args_json"] += partial_json
+
+        except Exception as exc:
+            raise ModelInvocationError(f"Anthropic stream failed: {exc}") from exc
+
+        # 流结束后发送工具调用
+        if tool_call_accum:
+            tool_calls = []
+            for acc in tool_call_accum.values():
+                try:
+                    args = json.loads(acc.get("args_json", "{}") or "{}")
+                except json.JSONDecodeError as exc:
+                    raise ModelInvocationError(
+                        f"Anthropic 流式工具 {acc.get('name')!r} 参数不是有效 JSON：{exc}"
+                    ) from exc
+                if not isinstance(args, dict):
+                    raise ModelInvocationError(f"Anthropic 流式工具 {acc.get('name')!r} 参数必须是 JSON 对象")
+                tool_calls.append(
+                    {
+                        "id": acc.get("id", ""),
+                        "type": "function",
+                        "name": acc.get("name", ""),
+                        "args": args,
+                    }
+                )
+            yield StreamChunk(tool_calls=tool_calls, is_tool_call=True)
+        else:
+            yield StreamChunk()

agent_core/model/providers/common.py

@@ -0,0 +1,28 @@
+"""Provider 共享辅助：已公开模型规格与统一 logger。
+
+具体厂商适配器见 ``openai_compatible``、``anthropic``、``gemini``、``ollama``；
+工厂与注册逻辑见 ``factory``。
+"""
+
+from __future__ import annotations
+
+import logging
+
+# 保留拆分前单文件模块的 logger 名，日志过滤行为保持不变。
+logger = logging.getLogger("agent_core.model.providers")
+
+
+# ── token / 已公开模型规格辅助函数 ────────────────────────────────
+def known_context_window(model: str) -> int | None:
+    """已公开模型规格；未知模型必须由 provider metadata 或配置提供。"""
+    normalized = model.lower()
+    known_windows = {
+        "gpt-4o": 128_000,
+        "gpt-4o-mini": 128_000,
+        "gpt-4.1": 1_047_576,
+        "gpt-4.1-mini": 1_047_576,
+        "gpt-4.1-nano": 1_047_576,
+        "qwen3.7-plus": 1_000_000,
+        "qwen3-max": 262_144,
+    }
+    return known_windows.get(normalized)

agent_core/model/providers/factory.py

@@ -0,0 +1,107 @@
+"""提供商构造器与注册：统一模型工厂入口。"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agent_core.model.core import (
+    ModelAdapter,
+    create_model_provider,
+    default_model_provider_registry,
+)
+from agent_core.model.providers.anthropic import AnthropicProvider
+from agent_core.model.providers.common import logger
+from agent_core.model.providers.gemini import GeminiProvider
+from agent_core.model.providers.ollama import OllamaProvider
+from agent_core.model.providers.openai_compatible import OpenAIProvider
+
+# 模型提供商注册表和具体适配器都属于 agent-core；第三方 SDK 只在实例化对应
+# 提供商时延迟导入，未使用的模型不会影响 Core 的基础导入。
+MODEL_PROVIDER_REGISTRY = default_model_provider_registry
+
+
+def _model_name(config: Any, model: str | None) -> str:
+    """读取统一模型名称，并保留旧配置对象的访问方式。"""
+    return model or str(getattr(config, "model_name", ""))
+
+
+def _build_openai(config: Any, *, model: str | None = None, **_: Any) -> ModelAdapter:
+    """构造 OpenAI 及其兼容接口适配器。"""
+    api_key = getattr(config, "openai_api_key", None)
+    if not api_key:
+        raise ValueError("使用 OpenAI 兼容接口需要配置 AGENT_OPENAI_API_KEY")
+    model_name = _model_name(config, model)
+    logger.info("LLM Provider: OpenAI-compatible (%s)", model_name)
+    return OpenAIProvider(
+        api_key=api_key,
+        base_url=getattr(config, "openai_base_url", "https://api.openai.com/v1"),
+        model=model_name,
+        temperature=getattr(config, "temperature", 0.0),
+        max_tokens=getattr(config, "max_tokens", 4096),
+        context_window_tokens=getattr(config, "context_window_tokens", None),
+    )
+
+
+def _build_anthropic(config: Any, *, model: str | None = None, **_: Any) -> ModelAdapter:
+    """构造 Anthropic 适配器。"""
+    api_key = getattr(config, "anthropic_api_key", None)
+    if not api_key:
+        raise ValueError("使用 Anthropic 需要配置 AGENT_ANTHROPIC_API_KEY")
+    model_name = _model_name(config, model)
+    logger.info("LLM Provider: Anthropic (%s)", model_name)
+    return AnthropicProvider(
+        api_key=api_key,
+        model=model_name,
+        temperature=getattr(config, "temperature", 0.0),
+        max_tokens=getattr(config, "max_tokens", 4096),
+        context_window_tokens=getattr(config, "context_window_tokens", None),
+    )
+
+
+def _build_ollama(config: Any, *, model: str | None = None, **_: Any) -> ModelAdapter:
+    """构造 Ollama 本地模型适配器。"""
+    model_name = _model_name(config, model)
+    base_url = getattr(config, "ollama_base_url", "http://localhost:11434/v1")
+    logger.info("LLM Provider: Ollama (%s @ %s)", model_name, base_url)
+    return OllamaProvider(
+        model=model_name,
+        base_url=base_url,
+        temperature=getattr(config, "temperature", 0.0),
+        max_tokens=getattr(config, "max_tokens", 4096),
+        context_window_tokens=getattr(config, "context_window_tokens", None),
+    )
+
+
+def _build_gemini(config: Any, *, model: str | None = None, **_: Any) -> ModelAdapter:
+    """构造 Google Gemini 适配器。"""
+    api_key = getattr(config, "gemini_api_key", None)
+    if not api_key:
+        raise ValueError("使用 Gemini 需要配置 AGENT_GEMINI_API_KEY")
+    model_name = _model_name(config, model)
+    logger.info("LLM Provider: Gemini (%s)", model_name)
+    return GeminiProvider(
+        api_key=api_key,
+        model=model_name,
+        temperature=getattr(config, "temperature", 0.0),
+        max_tokens=getattr(config, "max_tokens", 4096),
+        context_window_tokens=getattr(config, "context_window_tokens", None),
+    )
+
+
+MODEL_PROVIDER_REGISTRY.register("openai", _build_openai)
+MODEL_PROVIDER_REGISTRY.register("anthropic", _build_anthropic)
+MODEL_PROVIDER_REGISTRY.register("ollama", _build_ollama)
+MODEL_PROVIDER_REGISTRY.register("gemini", _build_gemini)
+
+
+def create_llm_provider(config: Any, *, model: str | None = None) -> ModelAdapter:
+    """兼容旧调用方的模型工厂入口。
+
+    选择、注册和具体 SDK 适配器都由 ``agent_core`` 提供。
+    旧代码仍可继续调用 ``create_llm_provider(config, model=...)``。
+    """
+    return create_model_provider(
+        config,
+        provider=getattr(config, "llm_provider", "openai"),
+        model=model,
+    )