klaus-agent 0.1.0 → 0.2.0

package/dist/providers/shared.js

@@ -0,0 +1,49 @@
+// Shared streaming utilities for LLM providers
+// Retryable error patterns common across providers
+const COMMON_RETRYABLE = ["429", "500", "503", "ECONNRESET"];
+export const RETRYABLE_PATTERNS = {
+    anthropic: [...COMMON_RETRYABLE, "rate_limit", "overloaded", "529"],
+    openai: [...COMMON_RETRYABLE, "rate_limit"],
+    google: [...COMMON_RETRYABLE],
+};
+export function isRetryableError(error, patterns) {
+    return patterns.some((p) => error.message.includes(p));
+}
+/**
+ * Wraps a streaming generator with exponential backoff retry logic.
+ * Only retries on connection-level failures before streaming starts.
+ */
+export async function* withRetry(streamOnce, retryablePatterns, maxRetries = 3) {
+    let lastError = null;
+    for (let attempt = 0; attempt <= maxRetries; attempt++) {
+        if (attempt > 0) {
+            const delay = Math.min(1000 * Math.pow(2, attempt - 1), 4000);
+            await new Promise((r) => setTimeout(r, delay));
+        }
+        try {
+            yield* streamOnce();
+            return;
+        }
+        catch (err) {
+            lastError = err instanceof Error ? err : new Error(String(err));
+            if (!isRetryableError(lastError, retryablePatterns) || attempt === maxRetries) {
+                yield { type: "error", error: lastError };
+                return;
+            }
+        }
+    }
+}
+/** Maps ThinkingLevel to token budget (shared by Anthropic and Gemini). */
+export function mapThinkingBudget(level) {
+    if (!level || level === "off")
+        return undefined;
+    const budgets = {
+        minimal: 1024,
+        low: 4096,
+        medium: 10240,
+        high: 20480,
+        xhigh: 40960,
+    };
+    return budgets[level];
+}
+//# sourceMappingURL=shared.js.map

package/dist/providers/shared.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"shared.js","sourceRoot":"","sources":["../../src/providers/shared.ts"],"names":[],"mappings":"AAAA,+CAA+C;AAI/C,mDAAmD;AACnD,MAAM,gBAAgB,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,EAAE,YAAY,CAAC,CAAC;AAE7D,MAAM,CAAC,MAAM,kBAAkB,GAA6B;IAC1D,SAAS,EAAE,CAAC,GAAG,gBAAgB,EAAE,YAAY,EAAE,YAAY,EAAE,KAAK,CAAC;IACnE,MAAM,EAAE,CAAC,GAAG,gBAAgB,EAAE,YAAY,CAAC;IAC3C,MAAM,EAAE,CAAC,GAAG,gBAAgB,CAAC;CAC9B,CAAC;AAEF,MAAM,UAAU,gBAAgB,CAAC,KAAY,EAAE,QAAkB;IAC/D,OAAO,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;AACzD,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,SAAS,CAAC,CAAC,SAAS,CAC9B,UAAsD,EACtD,iBAA2B,EAC3B,UAAU,GAAG,CAAC;IAEd,IAAI,SAAS,GAAiB,IAAI,CAAC;IAEnC,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,UAAU,EAAE,OAAO,EAAE,EAAE,CAAC;QACvD,IAAI,OAAO,GAAG,CAAC,EAAE,CAAC;YAChB,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;YAC9D,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC;QACjD,CAAC;QAED,IAAI,CAAC;YACH,KAAK,CAAC,CAAC,UAAU,EAAE,CAAC;YACpB,OAAO;QACT,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,SAAS,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;YAEhE,IAAI,CAAC,gBAAgB,CAAC,SAAS,EAAE,iBAAiB,CAAC,IAAI,OAAO,KAAK,UAAU,EAAE,CAAC;gBAC9E,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,CAAC;gBAC1C,OAAO;YACT,CAAC;QACH,CAAC;IACH,CAAC;AACH,CAAC;AAED,2EAA2E;AAC3E,MAAM,UAAU,iBAAiB,CAAC,KAAqB;IACrD,IAAI,CAAC,KAAK,IAAI,KAAK,KAAK,KAAK;QAAE,OAAO,SAAS,CAAC;IAChD,MAAM,OAAO,GAA2B;QACtC,OAAO,EAAE,IAAI;QACb,GAAG,EAAE,IAAI;QACT,MAAM,EAAE,KAAK;QACb,IAAI,EAAE,KAAK;QACX,KAAK,EAAE,KAAK;KACb,CAAC;IACF,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC;AACxB,CAAC"}

package/dist/providers/volcengine.d.ts

@@ -0,0 +1,4 @@
+import { OpenAIProvider } from "./openai.js";
+export declare class VolcengineProvider extends OpenAIProvider {
+    constructor(apiKey?: string, baseUrl?: string);
+}

package/dist/providers/volcengine.js

@@ -0,0 +1,8 @@
+// Volcengine (Doubao) LLM provider (OpenAI Chat Completions API compatible)
+import { OpenAIProvider } from "./openai.js";
+export class VolcengineProvider extends OpenAIProvider {
+    constructor(apiKey, baseUrl) {
+        super(apiKey ?? process.env.VOLCENGINE_API_KEY, baseUrl ?? "https://ark.cn-beijing.volces.com/api/v3");
+    }
+}
+//# sourceMappingURL=volcengine.js.map

package/dist/providers/volcengine.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"volcengine.js","sourceRoot":"","sources":["../../src/providers/volcengine.ts"],"names":[],"mappings":"AAAA,4EAA4E;AAE5E,OAAO,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAE7C,MAAM,OAAO,kBAAmB,SAAQ,cAAc;IACpD,YAAY,MAAe,EAAE,OAAgB;QAC3C,KAAK,CACH,MAAM,IAAI,OAAO,CAAC,GAAG,CAAC,kBAAkB,EACxC,OAAO,IAAI,0CAA0C,CACtD,CAAC;IACJ,CAAC;CACF"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "klaus-agent",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Universal agent framework",
   "type": "module",
   "main": "dist/index.js",
@@ -17,10 +17,11 @@
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.39.0",
-    "@sinclair/typebox": "^0.34.0"
+    "@google/generative-ai": "^0.21.0",
+    "@sinclair/typebox": "^0.34.0",
+    "openai": "^4.104.0"
   },
   "devDependencies": {
-    "@types/node": "^25.5.0",
     "typescript": "^5.7.0"
   }
 }

package/src/core/agent.ts

@@ -289,6 +289,20 @@ export class Agent {
     if (this._config.extensions?.length) {
       this._extensionRunner = new ExtensionRunner();
       await this._extensionRunner.loadExtensions(this._config.extensions);
+
+      // Re-resolve provider if extensions registered one for the configured provider
+      const extProviders = this._extensionRunner.getRegisteredProviders();
+      const extFactory = extProviders.get(this._config.model.provider);
+      if (extFactory) {
+        try {
+          this._provider = extFactory({
+            apiKey: this._config.model.apiKey,
+            baseUrl: this._config.model.baseUrl,
+          });
+        } catch {
+          // Extension provider factory failed, keep the original provider
+        }
+      }
     }
 
     // Notify extensions of session start

package/src/extensions/runner.ts

@@ -22,12 +22,14 @@ import type {
 } from "./types.js";
 import type { AgentTool } from "../tools/types.js";
 import type { AgentMessage } from "../types.js";
+import type { LLMProviderFactory } from "../llm/types.js";
 import type { TSchema } from "@sinclair/typebox";
 
 interface RegisteredExtension {
   handlers: Map<ExtensionEventType, ExtensionHandler<any>[]>;
   tools: AgentTool[];
   commands: Map<string, CommandHandler>;
+  providers: Map<string, LLMProviderFactory>;
 }
 
 export class ExtensionRunner {
@@ -40,6 +42,7 @@ export class ExtensionRunner {
         handlers: new Map(),
         tools: [],
         commands: new Map(),
+        providers: new Map(),
       };
 
       const api: ExtensionAPI = {
@@ -53,6 +56,9 @@ export class ExtensionRunner {
         registerCommand: (name: string, handler: CommandHandler) => {
           ext.commands.set(name, handler);
         },
+        registerProvider: (name: string, factory: LLMProviderFactory) => {
+          ext.providers.set(name, factory);
+        },
         sendMessage: (message: AgentMessage) => {
           this._pendingMessages.push(message);
         },
@@ -75,6 +81,16 @@ export class ExtensionRunner {
     return undefined;
   }
 
+  getRegisteredProviders(): Map<string, LLMProviderFactory> {
+    const merged = new Map<string, LLMProviderFactory>();
+    for (const ext of this._extensions) {
+      for (const [name, factory] of ext.providers) {
+        merged.set(name, factory);
+      }
+    }
+    return merged;
+  }
+
   drainPendingMessages(): AgentMessage[] {
     const msgs = [...this._pendingMessages];
     this._pendingMessages = [];

package/src/extensions/types.ts

@@ -3,7 +3,7 @@
 import type { TSchema } from "@sinclair/typebox";
 import type { AgentTool, AgentToolResult } from "../tools/types.js";
 import type { AgentMessage, ThinkingLevel } from "../types.js";
-import type { Message, ToolDefinition } from "../llm/types.js";
+import type { Message, ToolDefinition, LLMProviderFactory } from "../llm/types.js";
 
 // --- Extension factory ---
 
@@ -15,6 +15,7 @@ export interface ExtensionAPI {
   on<E extends ExtensionEventType>(event: E, handler: ExtensionHandler<E>): void;
   registerTool<TParams extends TSchema>(tool: AgentTool<TParams>): void;
   registerCommand(name: string, handler: CommandHandler): void;
+  registerProvider(name: string, factory: LLMProviderFactory): void;
   sendMessage(message: AgentMessage): void;
 }
 

package/src/index.ts

@@ -78,6 +78,12 @@ export function createAgent(config: CreateAgentConfig): Agent {
 export { Agent } from "./core/agent.js";
 export { ApprovalImpl } from "./approval/approval.js";
 export { registerProvider, resolveProvider } from "./llm/provider.js";
+export { AnthropicProvider } from "./providers/anthropic.js";
+export { OpenAIProvider } from "./providers/openai.js";
+export { GeminiProvider } from "./providers/google.js";
+export { MiniMaxProvider } from "./providers/minimax.js";
+export { KimiProvider } from "./providers/kimi.js";
+export { VolcengineProvider } from "./providers/volcengine.js";
 export { executeToolCalls } from "./tools/executor.js";
 export { SessionManager } from "./session/session-manager.js";
 export { buildSessionContext } from "./session/session-context-builder.js";
@@ -130,6 +136,7 @@ export type {
 // LLM types
 export type {
   LLMProvider,
+  LLMProviderFactory,
   LLMRequestOptions,
   AssistantMessage,
   AssistantMessageEvent,

package/src/llm/provider.ts

@@ -1,247 +1,29 @@
-// Anthropic LLM provider
-
-import Anthropic from "@anthropic-ai/sdk";
-import type {
-  LLMProvider,
-  LLMRequestOptions,
-  AssistantMessageEvent,
-  AssistantMessage,
-  AssistantContentBlock,
-  ToolDefinition,
-  TokenUsage,
-  ThinkingLevel,
-} from "./types.js";
-
-function mapThinkingLevel(level?: ThinkingLevel): number | undefined {
-  if (!level || level === "off") return undefined;
-  const budgets: Record<string, number> = {
-    minimal: 1024,
-    low: 4096,
-    medium: 10240,
-    high: 20480,
-    xhigh: 40960,
-  };
-  return budgets[level];
-}
-
-export class AnthropicProvider implements LLMProvider {
-  private client: Anthropic;
-
-  constructor(apiKey?: string, baseUrl?: string) {
-    this.client = new Anthropic({
-      apiKey: apiKey || process.env.ANTHROPIC_API_KEY,
-      ...(baseUrl ? { baseURL: baseUrl } : {}),
-    });
-  }
-
-  async *stream(options: LLMRequestOptions): AsyncIterable<AssistantMessageEvent> {
-    const maxRetries = 3;
-    let lastError: Error | null = null;
-
-    for (let attempt = 0; attempt <= maxRetries; attempt++) {
-      if (attempt > 0) {
-        const delay = Math.min(1000 * Math.pow(2, attempt - 1), 4000);
-        await new Promise((r) => setTimeout(r, delay));
-      }
-
-      try {
-        // On last attempt or first attempt, stream directly for real-time output.
-        // On intermediate retries, errors will be caught and retried.
-        yield* this._streamOnce(options);
-        return;
-      } catch (err) {
-        lastError = err instanceof Error ? err : new Error(String(err));
-
-        // Only retry on transient errors, and only if we haven't yielded
-        // any events yet (connection-level failures before streaming starts)
-        const isRetryable = lastError.message.includes("rate_limit")
-          || lastError.message.includes("overloaded")
-          || lastError.message.includes("529")
-          || lastError.message.includes("500")
-          || lastError.message.includes("ECONNRESET");
-
-        if (!isRetryable || attempt === maxRetries) {
-          yield { type: "error", error: lastError };
-          return;
-        }
-        // Note: if _streamOnce yielded partial events before throwing,
-        // those events are already consumed by the caller. Retries only
-        // help with pre-stream connection failures, not mid-stream breaks.
-      }
-    }
-  }
-
-  private async *_streamOnce(options: LLMRequestOptions): AsyncIterable<AssistantMessageEvent> {
-    const { model, systemPrompt, messages, tools, thinkingLevel, maxTokens, signal } = options;
-
-    const thinkingBudget = mapThinkingLevel(thinkingLevel);
-
-    const params: Anthropic.MessageCreateParamsStreaming = {
-      model,
-      max_tokens: maxTokens ?? 8192,
-      system: systemPrompt,
-      messages: messages.map((m) => {
-        if (m.role === "user") {
-          return { role: "user" as const, content: typeof m.content === "string" ? m.content : m.content.map(mapContentBlock) };
-        }
-        if (m.role === "assistant") {
-          return { role: "assistant" as const, content: m.content.map(mapAssistantBlock) };
-        }
-        // tool_result
-        return {
-          role: "user" as const,
-          content: [{
-            type: "tool_result" as const,
-            tool_use_id: m.toolCallId,
-            content: typeof m.content === "string" ? m.content : m.content.map(mapToolResultContent),
-            ...(m.isError ? { is_error: true as const } : {}),
-          }],
-        };
-      }),
-      stream: true,
-      ...(tools?.length ? {
-        tools: tools.map((t) => ({
-          name: t.name,
-          description: t.description,
-          input_schema: t.inputSchema as Anthropic.Tool.InputSchema,
-        })),
-      } : {}),
-      ...(thinkingBudget ? {
-        thinking: { type: "enabled" as const, budget_tokens: thinkingBudget },
-      } : {}),
-    };
-
-    const contentBlocks: AssistantContentBlock[] = [];
-    const toolInputBuffers = new Map<number, string>();
-    let usage: TokenUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
-
-    const stream = this.client.messages.stream(params, { signal });
-
-    try {
-      for await (const event of stream) {
-        if (event.type === "content_block_start") {
-          const block = event.content_block;
-          if (block.type === "text") {
-            contentBlocks.push({ type: "text", text: "" });
-          } else if (block.type === "tool_use") {
-            contentBlocks.push({ type: "tool_call", id: block.id, name: block.name, input: {} });
-            toolInputBuffers.set(event.index, "");
-            yield { type: "tool_call_start", id: block.id, name: block.name };
-          } else if (block.type === "thinking") {
-            contentBlocks.push({ type: "thinking", thinking: "" });
-          }
-        } else if (event.type === "content_block_delta") {
-          const delta = event.delta;
-          if (delta.type === "text_delta") {
-            const block = contentBlocks[event.index];
-            if (block && block.type === "text") {
-              block.text += delta.text;
-            }
-            yield { type: "text", text: delta.text };
-          } else if (delta.type === "input_json_delta") {
-            const buf = (toolInputBuffers.get(event.index) ?? "") + delta.partial_json;
-            toolInputBuffers.set(event.index, buf);
-            const block = contentBlocks[event.index];
-            if (block && block.type === "tool_call") {
-              yield { type: "tool_call_delta", id: block.id, input: delta.partial_json };
-            }
-          } else if (delta.type === "thinking_delta") {
-            const block = contentBlocks[event.index];
-            if (block && block.type === "thinking") {
-              block.thinking += delta.thinking;
-            }
-            yield { type: "thinking", thinking: delta.thinking };
-          }
-        } else if (event.type === "content_block_stop") {
-          // Parse accumulated tool input JSON
-          const block = contentBlocks[event.index];
-          if (block && block.type === "tool_call") {
-            const buf = toolInputBuffers.get(event.index) ?? "{}";
-            try {
-              block.input = JSON.parse(buf || "{}");
-            } catch {
-              block.input = {};
-            }
-            toolInputBuffers.delete(event.index);
-          }
-        } else if (event.type === "message_delta") {
-          if (event.usage) {
-            usage = {
-              inputTokens: usage.inputTokens,
-              outputTokens: event.usage.output_tokens,
-              totalTokens: usage.inputTokens + event.usage.output_tokens,
-            };
-          }
-        } else if (event.type === "message_start") {
-          if (event.message.usage) {
-            usage = {
-              inputTokens: event.message.usage.input_tokens,
-              outputTokens: event.message.usage.output_tokens,
-              totalTokens: event.message.usage.input_tokens + event.message.usage.output_tokens,
-              cacheReadTokens: (event.message.usage as any).cache_read_input_tokens,
-              cacheWriteTokens: (event.message.usage as any).cache_creation_input_tokens,
-            };
-          }
-        }
-      }
-    } catch (err) {
-      throw err; // Let retry wrapper handle it
-    }
-
-    const message: AssistantMessage = { role: "assistant", content: contentBlocks };
-    yield { type: "done", message, usage };
-  }
-}
-
-function mapToolResultContent(block: { type: string; [key: string]: any }): Anthropic.TextBlockParam | Anthropic.ImageBlockParam {
-  if (block.type === "image" && block.source?.type === "base64") {
-    return {
-      type: "image",
-      source: { type: "base64", media_type: block.source.mediaType, data: block.source.data },
-    };
-  }
-  if (block.type === "image" && block.source?.type === "url") {
-    return { type: "image", source: { type: "url", url: block.source.url } };
-  }
-  return { type: "text", text: block.type === "text" ? block.text : JSON.stringify(block) };
-}
-
-function mapContentBlock(block: { type: string; [key: string]: any }): Anthropic.ContentBlockParam {
-  if (block.type === "text") {
-    return { type: "text", text: block.text };
-  }
-  if (block.type === "image") {
-    if (block.source.type === "base64") {
-      return {
-        type: "image",
-        source: { type: "base64", media_type: block.source.mediaType, data: block.source.data },
-      };
-    }
-    return { type: "image", source: { type: "url", url: block.source.url } };
-  }
-  return { type: "text", text: JSON.stringify(block) };
-}
-
-function mapAssistantBlock(block: AssistantContentBlock): Anthropic.ContentBlockParam {
-  if (block.type === "text") {
-    return { type: "text", text: block.text };
-  }
-  if (block.type === "tool_call") {
-    return { type: "tool_use", id: block.id, name: block.name, input: block.input };
-  }
-  if (block.type === "thinking") {
-    return { type: "thinking", thinking: block.thinking, signature: "" };
-  }
-  return { type: "text", text: JSON.stringify(block) };
-}
-
-// --- Provider registry ---
-
-const providers = new Map<string, (config: { apiKey?: string; baseUrl?: string }) => LLMProvider>();
-
-providers.set("anthropic", (config) => new AnthropicProvider(config.apiKey, config.baseUrl));
-
-export function registerProvider(name: string, factory: (config: { apiKey?: string; baseUrl?: string }) => LLMProvider): void {
+// LLM provider registry
+
+import { AnthropicProvider } from "../providers/anthropic.js";
+import { OpenAIProvider } from "../providers/openai.js";
+import { GeminiProvider } from "../providers/google.js";
+import { MiniMaxProvider } from "../providers/minimax.js";
+import { KimiProvider } from "../providers/kimi.js";
+import { VolcengineProvider } from "../providers/volcengine.js";
+import type { LLMProvider, LLMProviderFactory } from "./types.js";
+
+const providers = new Map<string, LLMProviderFactory>();
+
+// Built-in providers
+providers.set("anthropic", (c) => new AnthropicProvider(c.apiKey, c.baseUrl));
+providers.set("openai", (c) => new OpenAIProvider(c.apiKey, c.baseUrl));
+providers.set("google", (c) => new GeminiProvider(c.apiKey, c.baseUrl));
+providers.set("minimax", (c) => new MiniMaxProvider(c.apiKey, c.baseUrl));
+providers.set("kimi", (c) => new KimiProvider(c.apiKey, c.baseUrl));
+providers.set("volcengine", (c) => new VolcengineProvider(c.apiKey, c.baseUrl));
+
+// Protocol-compatible proxies — user provides baseUrl to connect any compatible service
+providers.set("openai-compatible", (c) => new OpenAIProvider(c.apiKey, c.baseUrl));
+providers.set("anthropic-compatible", (c) => new AnthropicProvider(c.apiKey, c.baseUrl));
+providers.set("gemini-compatible", (c) => new GeminiProvider(c.apiKey, c.baseUrl));
+
+export function registerProvider(name: string, factory: LLMProviderFactory): void {
   providers.set(name, factory);
 }
 

package/src/llm/types.ts

@@ -144,3 +144,5 @@ export interface LLMRequestOptions {
   maxTokens?: number;
   signal?: AbortSignal;
 }
+
+export type LLMProviderFactory = (config: { apiKey?: string; baseUrl?: string }) => LLMProvider;

package/src/providers/anthropic.ts

@@ -0,0 +1,185 @@
+// Anthropic LLM provider
+
+import Anthropic from "@anthropic-ai/sdk";
+import type {
+  LLMProvider,
+  LLMRequestOptions,
+  AssistantMessageEvent,
+  AssistantMessage,
+  AssistantContentBlock,
+  TokenUsage,
+} from "../llm/types.js";
+import { withRetry, RETRYABLE_PATTERNS, mapThinkingBudget } from "./shared.js";
+
+export class AnthropicProvider implements LLMProvider {
+  private client: Anthropic;
+
+  constructor(apiKey?: string, baseUrl?: string) {
+    this.client = new Anthropic({
+      apiKey: apiKey || process.env.ANTHROPIC_API_KEY,
+      ...(baseUrl ? { baseURL: baseUrl } : {}),
+    });
+  }
+
+  async *stream(options: LLMRequestOptions): AsyncIterable<AssistantMessageEvent> {
+    yield* withRetry(() => this._streamOnce(options), RETRYABLE_PATTERNS.anthropic);
+  }
+
+  private async *_streamOnce(options: LLMRequestOptions): AsyncIterable<AssistantMessageEvent> {
+    const { model, systemPrompt, messages, tools, thinkingLevel, maxTokens, signal } = options;
+
+    const thinkingBudget = mapThinkingBudget(thinkingLevel);
+
+    const params: Anthropic.MessageCreateParamsStreaming = {
+      model,
+      max_tokens: maxTokens ?? 8192,
+      system: systemPrompt,
+      messages: messages.map((m) => {
+        if (m.role === "user") {
+          return { role: "user" as const, content: typeof m.content === "string" ? m.content : m.content.map(mapContentBlock) };
+        }
+        if (m.role === "assistant") {
+          return { role: "assistant" as const, content: m.content.map(mapAssistantBlock) };
+        }
+        // tool_result
+        return {
+          role: "user" as const,
+          content: [{
+            type: "tool_result" as const,
+            tool_use_id: m.toolCallId,
+            content: typeof m.content === "string" ? m.content : m.content.map(mapToolResultContent),
+            ...(m.isError ? { is_error: true as const } : {}),
+          }],
+        };
+      }),
+      stream: true,
+      ...(tools?.length ? {
+        tools: tools.map((t) => ({
+          name: t.name,
+          description: t.description,
+          input_schema: t.inputSchema as Anthropic.Tool.InputSchema,
+        })),
+      } : {}),
+      ...(thinkingBudget ? {
+        thinking: { type: "enabled" as const, budget_tokens: thinkingBudget },
+      } : {}),
+    };
+
+    const contentBlocks: AssistantContentBlock[] = [];
+    const toolInputBuffers = new Map<number, string>();
+    let usage: TokenUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
+
+    const stream = this.client.messages.stream(params, { signal });
+
+    for await (const event of stream) {
+        if (event.type === "content_block_start") {
+          const block = event.content_block;
+          if (block.type === "text") {
+            contentBlocks.push({ type: "text", text: "" });
+          } else if (block.type === "tool_use") {
+            contentBlocks.push({ type: "tool_call", id: block.id, name: block.name, input: {} });
+            toolInputBuffers.set(event.index, "");
+            yield { type: "tool_call_start", id: block.id, name: block.name };
+          } else if (block.type === "thinking") {
+            contentBlocks.push({ type: "thinking", thinking: "" });
+          }
+        } else if (event.type === "content_block_delta") {
+          const delta = event.delta;
+          if (delta.type === "text_delta") {
+            const block = contentBlocks[event.index];
+            if (block && block.type === "text") {
+              block.text += delta.text;
+            }
+            yield { type: "text", text: delta.text };
+          } else if (delta.type === "input_json_delta") {
+            const buf = (toolInputBuffers.get(event.index) ?? "") + delta.partial_json;
+            toolInputBuffers.set(event.index, buf);
+            const block = contentBlocks[event.index];
+            if (block && block.type === "tool_call") {
+              yield { type: "tool_call_delta", id: block.id, input: delta.partial_json };
+            }
+          } else if (delta.type === "thinking_delta") {
+            const block = contentBlocks[event.index];
+            if (block && block.type === "thinking") {
+              block.thinking += delta.thinking;
+            }
+            yield { type: "thinking", thinking: delta.thinking };
+          }
+        } else if (event.type === "content_block_stop") {
+          const block = contentBlocks[event.index];
+          if (block && block.type === "tool_call") {
+            const buf = toolInputBuffers.get(event.index) ?? "{}";
+            try {
+              block.input = JSON.parse(buf || "{}");
+            } catch {
+              block.input = {};
+            }
+            toolInputBuffers.delete(event.index);
+          }
+        } else if (event.type === "message_delta") {
+          if (event.usage) {
+            usage = {
+              inputTokens: usage.inputTokens,
+              outputTokens: event.usage.output_tokens,
+              totalTokens: usage.inputTokens + event.usage.output_tokens,
+            };
+          }
+        } else if (event.type === "message_start") {
+          if (event.message.usage) {
+            usage = {
+              inputTokens: event.message.usage.input_tokens,
+              outputTokens: event.message.usage.output_tokens,
+              totalTokens: event.message.usage.input_tokens + event.message.usage.output_tokens,
+              cacheReadTokens: (event.message.usage as any).cache_read_input_tokens,
+              cacheWriteTokens: (event.message.usage as any).cache_creation_input_tokens,
+            };
+          }
+        }
+      }
+
+    const message: AssistantMessage = { role: "assistant", content: contentBlocks };
+    yield { type: "done", message, usage };
+  }
+}
+
+function mapToolResultContent(block: { type: string; [key: string]: any }): Anthropic.TextBlockParam | Anthropic.ImageBlockParam {
+  if (block.type === "image" && block.source?.type === "base64") {
+    return {
+      type: "image",
+      source: { type: "base64", media_type: block.source.mediaType, data: block.source.data },
+    };
+  }
+  if (block.type === "image" && block.source?.type === "url") {
+    return { type: "image", source: { type: "url", url: block.source.url } };
+  }
+  return { type: "text", text: block.type === "text" ? block.text : JSON.stringify(block) };
+}
+
+function mapContentBlock(block: { type: string; [key: string]: any }): Anthropic.ContentBlockParam {
+  if (block.type === "text") {
+    return { type: "text", text: block.text };
+  }
+  if (block.type === "image") {
+    if (block.source.type === "base64") {
+      return {
+        type: "image",
+        source: { type: "base64", media_type: block.source.mediaType, data: block.source.data },
+      };
+    }
+    return { type: "image", source: { type: "url", url: block.source.url } };
+  }
+  return { type: "text", text: JSON.stringify(block) };
+}
+
+function mapAssistantBlock(block: AssistantContentBlock): Anthropic.ContentBlockParam {
+  if (block.type === "text") {
+    return { type: "text", text: block.text };
+  }
+  if (block.type === "tool_call") {
+    return { type: "tool_use", id: block.id, name: block.name, input: block.input };
+  }
+  if (block.type === "thinking") {
+    return { type: "thinking", thinking: block.thinking, signature: "" };
+  }
+  return { type: "text", text: JSON.stringify(block) };
+}