co-mono/packages/ai/src/providers/anthropic.ts
Mario Zechner e5aedfed29 feat(ai): Implement unified AI API with Anthropic provider
- Define clean API with complete() method and callbacks for streaming
- Add comprehensive type system for messages, tools, and usage
- Implement AnthropicAI provider with full feature support:
  - Thinking/reasoning with signatures
  - Tool calling with parallel execution
  - Streaming via callbacks (onText, onThinking)
  - Proper error handling and stop reasons
  - Cache tracking for input/output tokens
- Add working test/example demonstrating tool execution flow
- Support for system prompts, temperature, max tokens
- Proper message role types: user, assistant, toolResult
2025-08-17 23:30:20 +02:00

246 lines
6 KiB
TypeScript

import Anthropic from "@anthropic-ai/sdk";
import type {
ContentBlockParam,
MessageCreateParamsStreaming,
MessageParam,
Tool,
} from "@anthropic-ai/sdk/resources/messages.js";
import type { AI, AssistantMessage, Event, Message, Request, StopReason, TokenUsage, ToolCall } from "../types.js";
export interface AnthropicOptions {
thinking?: {
enabled: boolean;
budgetTokens?: number;
};
toolChoice?: "auto" | "any" | "none" | { type: "tool"; name: string };
}
export class AnthropicAI implements AI<AnthropicOptions> {
private client: Anthropic;
private model: string;
constructor(model: string, apiKey?: string, baseUrl?: string) {
if (!apiKey) {
if (!process.env.ANTHROPIC_API_KEY) {
throw new Error(
"Anthropic API key is required. Set ANTHROPIC_API_KEY environment variable or pass it as an argument.",
);
}
apiKey = process.env.ANTHROPIC_API_KEY;
}
this.client = new Anthropic({ apiKey, baseURL: baseUrl });
this.model = model;
}
async complete(request: Request, options?: AnthropicOptions): Promise<AssistantMessage> {
try {
const messages = this.convertMessages(request.messages);
const params: MessageCreateParamsStreaming = {
model: this.model,
messages,
max_tokens: request.maxTokens || 4096,
stream: true,
};
if (request.systemPrompt) {
params.system = request.systemPrompt;
}
if (request.temperature !== undefined) {
params.temperature = request.temperature;
}
if (request.tools) {
params.tools = this.convertTools(request.tools);
}
if (options?.thinking?.enabled) {
params.thinking = {
type: "enabled",
budget_tokens: options.thinking.budgetTokens || 1024,
};
}
if (options?.toolChoice) {
if (typeof options.toolChoice === "string") {
params.tool_choice = { type: options.toolChoice };
} else {
params.tool_choice = options.toolChoice;
}
}
const stream = this.client.messages.stream(
{
...params,
stream: true,
},
{
signal: request.signal,
},
);
for await (const event of stream) {
if (event.type === "content_block_delta") {
if (event.delta.type === "text_delta") {
request.onText?.(event.delta.text);
}
if (event.delta.type === "thinking_delta") {
request.onThinking?.(event.delta.thinking);
}
}
}
const msg = await stream.finalMessage();
const thinking = msg.content.some((block) => block.type === "thinking")
? msg.content
.filter((block) => block.type === "thinking")
.map((block) => block.thinking)
.join("\n")
: undefined;
// This is kinda wrong if there is more than one thinking block. We do not use interleaved thinking though, so we should
// always have a single thinking block.
const thinkingSignature = msg.content.some((block) => block.type === "thinking")
? msg.content
.filter((block) => block.type === "thinking")
.map((block) => block.signature)
.join("\n")
: undefined;
const content = msg.content.some((block) => block.type === "text")
? msg.content
.filter((block) => block.type === "text")
.map((block) => block.text)
.join("\n")
: undefined;
const toolCalls: ToolCall[] = msg.content
.filter((block) => block.type === "tool_use")
.map((block) => ({
id: block.id,
name: block.name,
arguments: block.input as Record<string, any>,
}));
const usage: TokenUsage = {
input: msg.usage.input_tokens,
output: msg.usage.output_tokens,
cacheRead: msg.usage.cache_read_input_tokens || 0,
cacheWrite: msg.usage.cache_creation_input_tokens || 0,
// TODO add cost
};
return {
role: "assistant",
content,
thinking,
thinkingSignature,
toolCalls,
model: this.model,
usage,
stopResaon: this.mapStopReason(msg.stop_reason),
};
} catch (error) {
return {
role: "assistant",
model: this.model,
usage: {
input: 0,
output: 0,
cacheRead: 0,
cacheWrite: 0,
},
stopResaon: "error",
error: error instanceof Error ? error.message : String(error),
};
}
}
private convertMessages(messages: Message[]): MessageParam[] {
const params: MessageParam[] = [];
for (const msg of messages) {
if (msg.role === "user") {
params.push({
role: "user",
content: msg.content,
});
} else if (msg.role === "assistant") {
const blocks: ContentBlockParam[] = [];
if (msg.thinking && msg.thinkingSignature) {
blocks.push({
type: "thinking",
thinking: msg.thinking,
signature: msg.thinkingSignature,
});
}
if (msg.content) {
blocks.push({
type: "text",
text: msg.content,
});
}
if (msg.toolCalls) {
for (const toolCall of msg.toolCalls) {
blocks.push({
type: "tool_use",
id: toolCall.id,
name: toolCall.name,
input: toolCall.arguments,
});
}
}
params.push({
role: "assistant",
content: blocks,
});
} else if (msg.role === "toolResult") {
params.push({
role: "user",
content: [
{
type: "tool_result",
tool_use_id: msg.toolCallId,
content: msg.content,
is_error: msg.isError,
},
],
});
}
}
return params;
}
private convertTools(tools: Request["tools"]): Tool[] {
if (!tools) return [];
return tools.map((tool) => ({
name: tool.name,
description: tool.description,
input_schema: {
type: "object" as const,
properties: tool.parameters.properties || {},
required: tool.parameters.required || [],
},
}));
}
private mapStopReason(reason: Anthropic.Messages.StopReason | null): StopReason {
switch (reason) {
case "end_turn":
return "stop";
case "max_tokens":
return "length";
case "tool_use":
return "toolUse";
case "refusal":
return "safety";
case "pause_turn": // Stop is good enough -> resubmit
return "stop";
case "stop_sequence":
return "stop"; // We don't supply stop sequences, so this should never happen
default:
return "stop";
}
}
}