@deepwhale/llm 1.0.10 → 1.0.12

package/LICENSE

@@ -1,21 +1,21 @@
-MIT License
-
-Copyright (c) 2026 yysf1949
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
+MIT License
+
+Copyright (c) 2026 yysf1949
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@deepwhale/llm",
-  "version": "1.0.10",
+  "version": "1.0.12",
   "description": "LLM client factory: DeepSeek + Anthropic providers with token pricing",
   "license": "MIT",
   "type": "module",

package/dist/anthropic-client.d.ts

@@ -1,147 +0,0 @@
-/**
- * @deepwhale/llm — AnthropicClient
- *
- * Sprint 1b.5 Step 2: Anthropic provider, 走 /anthropic endpoint 薄包装方案 (D1 拍板).
- *
- * 关键设计 (R-D 拍板 2026-06-03):
- * - 用官方 @anthropic-ai/sdk 实例发请求, 不手写 fetch
- * - SDK opts.fetch 注入是设计意图内的 escape hatch (Cloudflare Workers / Deno
- *   proxy), 我们用同 pattern 注入 mock fetch 用于测试, 真实部署走 SDK 实际 fetch
- * - baseURL 落 https://api.deepseek.com/anthropic, authToken 复用 DEEPSEEK_API_KEY
- *   (DeepSeek shim 接 /anthropic 路径, 同 key 验证) — **不**直连 api.anthropic.com
- * - 响应是 Anthropic-shape Message, 写 parseAnthropicMessage 翻译成 ChatResult
- * - SSE 走 SDK MessageStream, 写 parseAnthropicSseEvent 翻译 RawMessageStreamEvent
- *   → ChatChunk. 不复用 parseOai* (那是 OAI shape, 协议不同)
- * - X3 拍板: Step 2 不接真 API, 测试用 mock fetch, 不碰 key
- *
- * Cache 字段 (B1 拍板): Anthropic 有 cache_creation_input_tokens (新建) +
- * cache_read_input_tokens (命中) 两个**独立**字段. 我们合并到 cached_tokens:
- *   cached_tokens = (cache_creation ?? 0) + (cache_read ?? 0)
- * cache_creation 详细拆解留 sprint 2 改进 (跟 cache_hit_rate / cost_turn 一起).
- *
- * StopReason 翻译 (Anthropic → OAI-shape finish_reason):
- *   'end_turn'       → 'stop'
- *   'stop_sequence'  → 'stop'
- *   'max_tokens'     → 'length'
- *   'tool_use'       → 'tool_calls'
- *   null (in-flight) → undefined
- */
-import type { Message as AnthropicMessage, MessageDeltaUsage, RawMessageStreamEvent, Usage as AnthropicUsage } from '@anthropic-ai/sdk/resources/messages/messages.js';
-import type { ChatChunk, ChatMessage, ChatResult, LLMClient, LLMToolSchema, ModelId, Usage } from './types.js';
-import type { PricingConfig } from './pricing-config.js';
-/** DeepSeek shim 提供的 /anthropic 兼容端点 (相对 OAI v1 端点). */
-export declare const DEEPSEEK_ANTHROPIC_BASE_URL = "https://api.deepseek.com/anthropic";
-/** Anthropic 默认 model (走 DeepSeek shim 时, 服务端映射到 Claude Sonnet 4.5). */
-export declare const ANTHROPIC_DEFAULT_MODEL = "claude-sonnet-4-5";
-export interface AnthropicClientOptions {
-    /** API key。优先于 process.env.ANTHROPIC_AUTH_TOKEN / DEEPSEEK_API_KEY。 */
-    apiKey?: string;
-    /** 模型 ID，默认 claude-sonnet-4-5。 */
-    model?: string;
-    /**
-     * Base URL。
-     *
-     * 拍板 1C (2026-06-04): **DeepSeek-first + Anthropic SDK 协议兼容 + 多 Provider Adapter**。
-     * 默认 `https://api.deepseek.com/anthropic` (DeepSeek 提供的 /anthropic 兼容端点, 单 key 走两家),
-     * 但 caller 可显式指定其他 Anthropic-兼容 provider:
-     *   - `https://api.anthropic.com` — 真 Anthropic API (需 ANTHROPIC_AUTH_TOKEN 真 key, 不走 DEEPSEEK_API_KEY 退路)
-     *   - `https://openrouter.ai/api/v1/anthropic` — OpenRouter Anthropic Route
-     *   - 任何自定义 proxy / 兼容层
-     *
-     * Sprint 1d.5-B 揭示: DeepSeek /anthropic 端点**实际** routing 兜底到 OAI flash (server.model=deepseek-v4-flash),
-     * 行为稳定但 server 协议声明 mis-labeled. 1C 拍板**不**解决这个 routing 问题 — 现状保留, caller
-     * 自选 baseUrl 决定走哪条路.
-     */
-    baseUrl?: string;
-    /** 自定义 fetch (注入 mock 用于测试). 默认走 SDK 内部 fetch. */
-    fetchImpl?: typeof fetch;
-    /** 单次 HTTP 调用的超时毫秒，默认 60s。 */
-    timeoutMs?: number;
-    /**
-     * pricing config override. 不传 → 走 ship-in default.toml.
-     * Sprint 1b.5 pricing 抽象层 (per-model currency). 详见 pricing-config.ts.
-     */
-    pricing?: PricingConfig;
-}
-/**
- * AnthropicClient implements LLMClient.
- *
- * 内部包一个 @anthropic-ai/sdk 实例, .messages.create() / .messages.stream()
- * 走 SDK 真实 HTTP 路径. 我们负责:
- * - ChatMessage → Anthropic.MessageParam 转换 (Sprint 1c 再加 tool_use schema)
- * - Anthropic.Message → ChatResult 转换 (parseAnthropicMessage)
- * - RawMessageStreamEvent → ChatChunk 转换 (parseAnthropicSseEvent, 含 usage 翻译)
- *
- * 不**在**这里写 HTTP / SSE 解析 (SDK 负责), 不**在**这里做重试 (SDK 自带 maxRetries=2,
- * 我们接受默认). Sprint 1c 集成测真接 shim 时可调整 SDK timeout / maxRetries.
- */
-export declare class AnthropicClient implements LLMClient {
-    readonly model: ModelId;
-    private readonly apiKey;
-    private readonly baseUrl;
-    private readonly timeoutMs;
-    private readonly pricing;
-    private readonly sdk;
-    constructor(options?: AnthropicClientOptions);
-    chat(messages: ChatMessage[], options?: {
-        signal?: AbortSignal;
-        tools?: ReadonlyArray<LLMToolSchema>;
-        tool_choice?: 'auto' | 'none' | 'required';
-    }): Promise<ChatResult>;
-    stream(messages: ChatMessage[], options: {
-        signal?: AbortSignal;
-        tools?: ReadonlyArray<LLMToolSchema>;
-        tool_choice?: 'auto' | 'none' | 'required';
-        onChunk: (chunk: ChatChunk) => void;
-    }): Promise<ChatResult>;
-}
-/**
- * 把 Anthropic SDK 的 Message 翻译成 ChatResult.
- *
- * - content: 拼 text block, 跳过 tool_use (1c 实施) / thinking (1b.5 不暴露)
- * - stop_reason → finish_reason 翻译
- * - usage: cache_creation + cache_read → cached_tokens (B1 拍板). 算 cache_hit_rate
- *   + cost_turn + tokens_uncached 跟 OAI 路径一致.
- */
-export declare function parseAnthropicMessage(message: AnthropicMessage, fallbackModel: ModelId, pricing?: PricingConfig): ChatResult;
-/**
- * 把 Anthropic Usage 翻译成标准化 Usage 结构 (带 cache/cost 字段).
- *
- * B1 拍板 (Sprint 1b.5 Step 2): cached_tokens = (cache_creation ?? 0) + (cache_read ?? 0).
- *
- * ⚠️ 重要语义修正 (F4 拍板 2026-06-03, review 找到):
- * Anthropic 官方 prompt caching 文档: total input tokens = input_tokens + cache_creation_input_tokens
- * + cache_read_input_tokens. 之前 Step 2 写时把 input_tokens 当"总 prompt" 是错的, 漏算 cache
- * 字段对应的 token 总量. 修法:
- *
- * - total_prompt = input_tokens + cache_creation + cache_read
- * - cached_tokens = cache_creation + cache_read (跟 Step 2 一致)
- * - tokens_uncached = total_prompt - cached_tokens = input_tokens (不变量)
- * - cache_hit_rate = cached_tokens / total_prompt (cache 命中率, 包括 write + read)
- *
- * **Cost 1b.5 保守策略** (F4 拍板): cache_creation 跟 cache_read 价格不同 (Sonnet 1h TTL write
- * \$3.75/M, read \$0.30/M, 比例 12.5×), 1b.5 pricing 模型**不**拆 cache_write vs cache_read 字段.
- * 为避免假装知道 cache_creation 价格, **保守**: cache_creation OR cache_read 任一非零 →
- * cost_turn/cost_currency 字段 absent. 留 sprint 2 加 `cache_write_per_m` 字段.
- * - 注意: tokens_uncached 仍**可**算 (是 input_tokens, 跟 cache 字段无关), 不受 cost 限制
- *
- * 接受 Usage | MessageDeltaUsage (后者无 cache_creation / cache_read, 视为 0).
- * Sprint 1b.5 Step 2: 流末尾 message_delta event.usage 是 MessageDeltaUsage, 走 delta 路径
- * 不算 cost, 真实生产靠 stream.finalMessage() 拿完整 Usage.
- */
-export declare function parseAnthropicUsage(usage: AnthropicUsage | MessageDeltaUsage, fallbackModel: ModelId, pricing?: PricingConfig): Usage | undefined;
-/**
- * 把单个 RawMessageStreamEvent 翻译成 ChatChunk.
- * 返 null 表示跳过 (heartbeat / ping / 不该透出的 event).
- *
- * Event 类型 (来自 SDK type):
- * - message_start: 携带 message metadata (model, id, usage.input_tokens=0) — 不透出
- * - content_block_start: 新的 text/tool_use/thinking block 起点 — 不透出 (避免 0 token 块)
- * - content_block_delta: text 增量 (delta.text) 或 input_json_delta (tool_use) — 透出 content
- * - content_block_stop: block 结束 — 不透出
- * - message_delta: 顶层 delta (stop_reason + usage 更新) — 透出 finish_reason + 最终 usage
- * - message_stop: 流结束 — 不透出 (caller 走 stream.finalMessage 拿 final Message)
- * - ping: heartbeat — 不透出
- */
-export declare function parseAnthropicSseEvent(event: RawMessageStreamEvent, fallbackModel: ModelId, pricing?: PricingConfig): ChatChunk | null;
-//# sourceMappingURL=anthropic-client.d.ts.map

package/dist/anthropic-client.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"anthropic-client.d.ts","sourceRoot":"","sources":["../src/anthropic-client.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;GA2BG;AAGH,OAAO,KAAK,EACV,OAAO,IAAI,gBAAgB,EAE3B,iBAAiB,EACjB,qBAAqB,EACrB,KAAK,IAAI,cAAc,EACxB,MAAM,kDAAkD,CAAC;AAM1D,OAAO,KAAK,EACV,SAAS,EACT,WAAW,EACX,UAAU,EACV,SAAS,EACT,aAAa,EACb,OAAO,EAEP,KAAK,EACN,MAAM,YAAY,CAAC;AAEpB,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAEzD,wDAAwD;AACxD,eAAO,MAAM,2BAA2B,uCAAuC,CAAC;AAEhF,wEAAwE;AACxE,eAAO,MAAM,uBAAuB,sBAAsB,CAAC;AAI3D,MAAM,WAAW,sBAAsB;IACrC,uEAAuE;IACvE,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,kCAAkC;IAClC,KAAK,CAAC,EAAE,MAAM,CAAC;IACf;;;;;;;;;;;;;OAaG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,kDAAkD;IAClD,SAAS,CAAC,EAAE,OAAO,KAAK,CAAC;IACzB,8BAA8B;IAC9B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;OAGG;IACH,OAAO,CAAC,EAAE,aAAa,CAAC;CACzB;AAED;;;;;;;;;;;GAWG;AACH,qBAAa,eAAgB,YAAW,SAAS;IAC/C,QAAQ,CAAC,KAAK,EAAE,OAAO,CAAC;IACxB,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAS;IAChC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAS;IACjC,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAS;IACnC,OAAO,CAAC,QAAQ,CAAC,OAAO,CAA4B;IACpD,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAY;gBAEpB,OAAO,GAAE,sBAA2B;IAwB1C,IAAI,CACR,QAAQ,EAAE,WAAW,EAAE,EACvB,OAAO,CAAC,EAAE;QACR,MAAM,CAAC,EAAE,WAAW,CAAC;QACrB,KAAK,CAAC,EAAE,aAAa,CAAC,aAAa,CAAC,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,CAAC;KAC5C,GACA,OAAO,CAAC,UAAU,CAAC;IA0BhB,MAAM,CACV,QAAQ,EAAE,WAAW,EAAE,EACvB,OAAO,EAAE;QACP,MAAM,CAAC,EAAE,WAAW,CAAC;QACrB,KAAK,CAAC,EAAE,aAAa,CAAC,aAAa,CAAC,CAAC;QACrC,WAAW,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,CAAC;QAC3C,OAAO,EAAE,CAAC,KAAK,EAAE,SAAS,KAAK,IAAI,CAAC;KACrC,GACA,OAAO,CAAC,UAAU,CAAC;CAsCvB;AAqID;;;;;;;GAOG;AACH,wBAAgB,qBAAqB,CACnC,OAAO,EAAE,gBAAgB,EACzB,aAAa,EAAE,OAAO,EACtB,OAAO,CAAC,EAAE,aAAa,GACtB,UAAU,CA8BZ;AAiBD;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,wBAAgB,mBAAmB,CACjC,KAAK,EAAE,cAAc,GAAG,iBAAiB,EACzC,aAAa,EAAE,OAAO,EACtB,OAAO,CAAC,EAAE,aAAa,GACtB,KAAK,GAAG,SAAS,CA6CnB;AAED;;;;;;;;;;;;GAYG;AACH,wBAAgB,sBAAsB,CACpC,KAAK,EAAE,qBAAqB,EAC5B,aAAa,EAAE,OAAO,EACtB,OAAO,CAAC,EAAE,aAAa,GACtB,SAAS,GAAG,IAAI,CA2BlB"}

package/dist/anthropic-client.js

@@ -1,443 +0,0 @@
-/**
- * @deepwhale/llm — AnthropicClient
- *
- * Sprint 1b.5 Step 2: Anthropic provider, 走 /anthropic endpoint 薄包装方案 (D1 拍板).
- *
- * 关键设计 (R-D 拍板 2026-06-03):
- * - 用官方 @anthropic-ai/sdk 实例发请求, 不手写 fetch
- * - SDK opts.fetch 注入是设计意图内的 escape hatch (Cloudflare Workers / Deno
- *   proxy), 我们用同 pattern 注入 mock fetch 用于测试, 真实部署走 SDK 实际 fetch
- * - baseURL 落 https://api.deepseek.com/anthropic, authToken 复用 DEEPSEEK_API_KEY
- *   (DeepSeek shim 接 /anthropic 路径, 同 key 验证) — **不**直连 api.anthropic.com
- * - 响应是 Anthropic-shape Message, 写 parseAnthropicMessage 翻译成 ChatResult
- * - SSE 走 SDK MessageStream, 写 parseAnthropicSseEvent 翻译 RawMessageStreamEvent
- *   → ChatChunk. 不复用 parseOai* (那是 OAI shape, 协议不同)
- * - X3 拍板: Step 2 不接真 API, 测试用 mock fetch, 不碰 key
- *
- * Cache 字段 (B1 拍板): Anthropic 有 cache_creation_input_tokens (新建) +
- * cache_read_input_tokens (命中) 两个**独立**字段. 我们合并到 cached_tokens:
- *   cached_tokens = (cache_creation ?? 0) + (cache_read ?? 0)
- * cache_creation 详细拆解留 sprint 2 改进 (跟 cache_hit_rate / cost_turn 一起).
- *
- * StopReason 翻译 (Anthropic → OAI-shape finish_reason):
- *   'end_turn'       → 'stop'
- *   'stop_sequence'  → 'stop'
- *   'max_tokens'     → 'length'
- *   'tool_use'       → 'tool_calls'
- *   null (in-flight) → undefined
- */
-import Anthropic from '@anthropic-ai/sdk';
-import { readFileSync } from 'node:fs';
-import { fileURLToPath } from 'node:url';
-import { dirname, resolve } from 'node:path';
-import process from 'node:process';
-import { APIKeyMissingError, LLMUnknownError } from './types.js';
-import { computeCost, parsePricingConfig } from './pricing-config.js';
-/** DeepSeek shim 提供的 /anthropic 兼容端点 (相对 OAI v1 端点). */
-export const DEEPSEEK_ANTHROPIC_BASE_URL = 'https://api.deepseek.com/anthropic';
-/** Anthropic 默认 model (走 DeepSeek shim 时, 服务端映射到 Claude Sonnet 4.5). */
-export const ANTHROPIC_DEFAULT_MODEL = 'claude-sonnet-4-5';
-const DEFAULT_TIMEOUT_MS = 60_000;
-/**
- * AnthropicClient implements LLMClient.
- *
- * 内部包一个 @anthropic-ai/sdk 实例, .messages.create() / .messages.stream()
- * 走 SDK 真实 HTTP 路径. 我们负责:
- * - ChatMessage → Anthropic.MessageParam 转换 (Sprint 1c 再加 tool_use schema)
- * - Anthropic.Message → ChatResult 转换 (parseAnthropicMessage)
- * - RawMessageStreamEvent → ChatChunk 转换 (parseAnthropicSseEvent, 含 usage 翻译)
- *
- * 不**在**这里写 HTTP / SSE 解析 (SDK 负责), 不**在**这里做重试 (SDK 自带 maxRetries=2,
- * 我们接受默认). Sprint 1c 集成测真接 shim 时可调整 SDK timeout / maxRetries.
- */
-export class AnthropicClient {
-    model;
-    apiKey;
-    baseUrl;
-    timeoutMs;
-    pricing;
-    sdk;
-    constructor(options = {}) {
-        this.apiKey = options.apiKey ?? resolveApiKey();
-        if (this.apiKey === '') {
-            throw new APIKeyMissingError('Anthropic API key not set. Set ANTHROPIC_AUTH_TOKEN or DEEPSEEK_API_KEY env var, ' +
-                'or pass apiKey option.');
-        }
-        this.model = (options.model ?? ANTHROPIC_DEFAULT_MODEL);
-        this.baseUrl = options.baseUrl ?? DEEPSEEK_ANTHROPIC_BASE_URL;
-        this.timeoutMs = options.timeoutMs ?? DEFAULT_TIMEOUT_MS;
-        this.pricing = options.pricing ?? loadDefaultPricing();
-        // SDK opts.fetch 注入是设计意图内的 escape hatch (测试 mock + Cloudflare
-        // Workers / Deno proxy). 真实生产不传, SDK 走全局 fetch.
-        const sdkOptions = {
-            authToken: this.apiKey,
-            baseURL: this.baseUrl,
-            timeout: this.timeoutMs,
-        };
-        if (options.fetchImpl !== undefined)
-            sdkOptions.fetch = options.fetchImpl;
-        this.sdk = new Anthropic(sdkOptions);
-    }
-    async chat(messages, options) {
-        // Sprint 1c.5 拍板 (1c-revive-2-B-1, 2026-06-04): tool schema 转换 (OAI {parameters} → Anthropic
-        // {input_schema}), 跟 DeepSeekClient 同 LLMClient 契约 (5-7 行 production 改).
-        // 跟 pi-agent 4-layer 模式: model layer (AnthropicClient) 不知道 tool registry 细节, 只做协议转换.
-        const body = toAnthropicMessages(messages, options?.tools);
-        const createParams = {
-            model: this.model,
-            messages: body.messages,
-            max_tokens: 4096, // Anthropic API 必填, 4096 是合理默认 (Sprint 1c 让 caller 传)
-            ...(body.system !== undefined ? { system: body.system } : {}),
-            ...(body.tools !== undefined ? { tools: body.tools } : {}),
-            ...(options?.tool_choice !== undefined ? { tool_choice: mapToolChoice(options.tool_choice) } : {}),
-        };
-        const sdkOptions = {
-            ...(options?.signal !== undefined ? { signal: options.signal } : {}),
-        };
-        let response;
-        try {
-            response = await this.sdk.messages.create(createParams, sdkOptions);
-        }
-        catch (e) {
-            throw mapSdkError(e);
-        }
-        return parseAnthropicMessage(response, this.model, this.pricing);
-    }
-    async stream(messages, options) {
-        const body = toAnthropicMessages(messages, options.tools);
-        const streamParams = {
-            model: this.model,
-            messages: body.messages,
-            max_tokens: 4096,
-            ...(body.system !== undefined ? { system: body.system } : {}),
-            ...(body.tools !== undefined ? { tools: body.tools } : {}),
-            ...(options.tool_choice !== undefined ? { tool_choice: mapToolChoice(options.tool_choice) } : {}),
-        };
-        const sdkOptions = {
-            ...(options.signal !== undefined ? { signal: options.signal } : {}),
-        };
-        // SDK 的 messages.stream() 返 MessageStream (异步可迭代 + EventEmitter).
-        // 跟 for await 一起用, 跟 OAI SSE 处理模式不同 — 这里是 SDK 自家 stream,
-        // 不是 SSE 字节流. parseAnthropicSseEvent 负责把每个 RawMessageStreamEvent
-        // 翻译成 ChatChunk.
-        const stream = this.sdk.messages.stream(streamParams, sdkOptions);
-        let finalMessage;
-        for await (const event of stream) {
-            const chunk = parseAnthropicSseEvent(event, this.model, this.pricing);
-            if (chunk !== null) {
-                options.onChunk(chunk);
-                // 收到 message_stop 时, SDK 也提供 finalMessage(), 但我们靠 event 流本身
-                // 拼出 stop_reason + 完整 usage. 留 1c 集成测时优化 (现 8 tests 覆盖 main path).
-            }
-            // SDK 的 message_stop event 不携带 final message, 需调 stream.finalMessage()
-            // 才能拿到. 简化: 收完流后从 stream 实例读 final message.
-            if (event.type === 'message_stop') {
-                finalMessage = await stream.finalMessage();
-            }
-        }
-        if (finalMessage === undefined) {
-            throw new LLMUnknownError('Anthropic stream ended without message_stop event');
-        }
-        return parseAnthropicMessage(finalMessage, this.model, this.pricing);
-    }
-}
-// ---- 私有 helper ----
-function resolveApiKey() {
-    // 优先 ANTHROPIC_AUTH_TOKEN (Anthropic SDK 标准), 退到 DEEPSEEK_API_KEY (Sprint 1b.5 shim).
-    const anthropic = process.env['ANTHROPIC_AUTH_TOKEN'];
-    if (anthropic !== undefined && anthropic !== '')
-        return anthropic;
-    const deepseek = process.env['DEEPSEEK_API_KEY'];
-    if (deepseek !== undefined && deepseek !== '')
-        return deepseek;
-    return '';
-}
-function loadDefaultPricing() {
-    try {
-        // 跟 DeepSeekClient 走同 pattern: readFileSync pricing.default.toml relative to dist/.
-        // Sprint 1c 集成测时检查这条路径在 ESM 打包后是否仍 OK. Step 2 单测用 mock fetch,
-        // 不走 loadDefaultPricing (test fixture 显式传 pricing).
-        const here = dirname(fileURLToPath(import.meta.url));
-        const defaultPath = resolve(here, 'pricing.default.toml');
-        const tomlText = readFileSync(defaultPath, 'utf-8');
-        return parsePricingConfig(tomlText);
-    }
-    catch {
-        return undefined;
-    }
-}
-function mapSdkError(e) {
-    // SDK 错误分类: APIConnectionError / APIConnectionTimeoutError / RateLimitError /
-    // AuthenticationError / BadRequestError / InternalServerError 等. 简化: 透传
-    // SDK Error name + message 到 LLMUnknownError. Sprint 1c 集成测时细化 1:1 映射
-    // 到 LLMRateLimitError / LLMAuthError / LLMNetworkError (跟 DeepSeekClient 一致).
-    if (e instanceof Error)
-        return new LLMUnknownError(`Anthropic SDK: ${e.message}`, { cause: e });
-    return new LLMUnknownError(`Anthropic SDK: ${String(e)}`);
-}
-/** 拆分 system / 非 system messages (Anthropic 协议 system 是顶层字段). */
-function toAnthropicMessages(messages, tools) {
-    const out = [];
-    let system;
-    // Sprint 1c-revive-2-D-4-1 (P38, 2026-06-04): 合并连续 tool 消息到 1 个 user 消息
-    // (Anthropic 协议要求 N 个 tool_use 紧跟 1 个 user 消息含 N 个 tool_result blocks).
-    // 1c.5 拍板时 1-tool-call 路径碰巧合法 (N=1 时独立 user 消息仍可), 多 tool_calls 揭示.
-    let pendingToolResults;
-    const flushToolResults = () => {
-        if (pendingToolResults !== undefined && pendingToolResults.length > 0) {
-            out.push({
-                role: 'user',
-                content: pendingToolResults,
-            });
-        }
-        pendingToolResults = undefined;
-    };
-    for (const m of messages) {
-        if (m.role === 'system') {
-            // 多条 system 合并 (Anthropic system 是单 string, 重复 system 罕见)
-            system = system === undefined ? m.content : `${system}\n\n${m.content}`;
-            continue;
-        }
-        if (m.role === 'tool') {
-            // 拍板 (D-4-1): 累积 tool_result 进 pendingToolResults (跟下一个 tool 消息合并).
-            // flush 时机: 1) 遇到非 tool 角色, 2) loop 结束.
-            if (pendingToolResults === undefined) {
-                pendingToolResults = [];
-            }
-            pendingToolResults.push({
-                type: 'tool_result',
-                tool_use_id: m.tool_call_id ?? '',
-                content: m.content,
-            });
-            continue;
-        }
-        // 非 tool 角色: 先 flush pending tool_results (如果有)
-        flushToolResults();
-        if (m.role === 'user') {
-            out.push({ role: 'user', content: m.content });
-            continue;
-        }
-        // assistant: OAI tool_calls → Anthropic content blocks (text + tool_use)
-        if (m.role === 'assistant') {
-            if (m.tool_calls !== undefined && m.tool_calls.length > 0) {
-                const blocks = m.tool_calls.map((tc) => ({
-                    type: 'tool_use',
-                    id: tc.id,
-                    name: tc.name,
-                    input: tc.args,
-                }));
-                out.push({
-                    role: 'assistant',
-                    // ToolUseBlockParam[] 在 SDK 类型上是 ContentBlockParam[] 的子集, 但 TS 4.x 推断不到
-                    // (SDK 用 union 反推, 编译期会失配). 显式 cast: 真实运行时 server 接受.
-                    content: blocks,
-                });
-                continue;
-            }
-            out.push({ role: 'assistant', content: m.content });
-            continue;
-        }
-    }
-    // loop 结束: flush 最后一批 tool_results (避免末尾独立 tool 消息丢失)
-    flushToolResults();
-    // tool schema: OAI {name, description, parameters} → Anthropic {name, description, input_schema}
-    // 1c.5 拍板: 走 Tool 类型 (跟 SDK 对齐), 不拆 ToolUnion (Bash20250124 等 built-in 工具暂不用).
-    let anthropicTools;
-    if (tools !== undefined && tools.length > 0) {
-        anthropicTools = tools.map((t) => ({
-            name: t.name,
-            description: t.description,
-            input_schema: t.parameters,
-        }));
-    }
-    const out2 = { system, messages: out };
-    if (anthropicTools !== undefined)
-        out2.tools = anthropicTools;
-    return out2;
-}
-/** Map LLMClient 通用 tool_choice (OAI 风格) → Anthropic ToolChoice. */
-function mapToolChoice(choice) {
-    switch (choice) {
-        case 'auto':
-            return { type: 'auto' };
-        case 'none':
-            return { type: 'none' };
-        case 'required':
-            return { type: 'any' }; // Anthropic 协议 'any' 强制至少调 1 个, 跟 OAI 'required' 等价
-    }
-}
-// ---- 解析层: Anthropic.Message / RawMessageStreamEvent → ChatResult / ChatChunk ----
-/**
- * 把 Anthropic SDK 的 Message 翻译成 ChatResult.
- *
- * - content: 拼 text block, 跳过 tool_use (1c 实施) / thinking (1b.5 不暴露)
- * - stop_reason → finish_reason 翻译
- * - usage: cache_creation + cache_read → cached_tokens (B1 拍板). 算 cache_hit_rate
- *   + cost_turn + tokens_uncached 跟 OAI 路径一致.
- */
-export function parseAnthropicMessage(message, fallbackModel, pricing) {
-    // 拼 text + 提取 tool_use (1c.5 实施, 1b.5 留空)
-    let content = '';
-    const toolCalls = [];
-    for (const block of message.content) {
-        if (block.type === 'text') {
-            content = content === '' ? block.text : `${content}${block.text}`;
-        }
-        else if (block.type === 'tool_use') {
-            // Sprint 1c.5 (1c-revive-2-B-1): tool_use block → OAI-style ToolCall (跟 DeepSeek shape 对齐)
-            // Anthropic SDK 给 input: unknown, 我们假设是 parsed object (runToolLoop 给 args object)
-            const input = block.input;
-            const args = typeof input === 'object' && input !== null
-                ? input
-                : {};
-            toolCalls.push({ id: block.id, name: block.name, args });
-        }
-        // thinking / redacted_thinking 跳过 (跟 1b.5 范围一致)
-    }
-    const finishReason = mapStopReason(message.stop_reason);
-    const usage = parseAnthropicUsage(message.usage, fallbackModel, pricing);
-    const model = message.model ?? fallbackModel;
-    const result = { model, content };
-    if (toolCalls.length > 0)
-        result.tool_calls = toolCalls;
-    if (usage !== undefined)
-        result.usage = usage;
-    if (finishReason !== undefined)
-        result.finish_reason = finishReason;
-    return result;
-}
-/** 翻译 Anthropic stop_reason → ChatResult['finish_reason']. */
-function mapStopReason(stopReason) {
-    if (stopReason === null)
-        return undefined;
-    switch (stopReason) {
-        case 'end_turn':
-            return 'stop';
-        case 'stop_sequence':
-            return 'stop';
-        case 'max_tokens':
-            return 'length';
-        case 'tool_use':
-            return 'tool_calls';
-    }
-}
-/**
- * 把 Anthropic Usage 翻译成标准化 Usage 结构 (带 cache/cost 字段).
- *
- * B1 拍板 (Sprint 1b.5 Step 2): cached_tokens = (cache_creation ?? 0) + (cache_read ?? 0).
- *
- * ⚠️ 重要语义修正 (F4 拍板 2026-06-03, review 找到):
- * Anthropic 官方 prompt caching 文档: total input tokens = input_tokens + cache_creation_input_tokens
- * + cache_read_input_tokens. 之前 Step 2 写时把 input_tokens 当"总 prompt" 是错的, 漏算 cache
- * 字段对应的 token 总量. 修法:
- *
- * - total_prompt = input_tokens + cache_creation + cache_read
- * - cached_tokens = cache_creation + cache_read (跟 Step 2 一致)
- * - tokens_uncached = total_prompt - cached_tokens = input_tokens (不变量)
- * - cache_hit_rate = cached_tokens / total_prompt (cache 命中率, 包括 write + read)
- *
- * **Cost 1b.5 保守策略** (F4 拍板): cache_creation 跟 cache_read 价格不同 (Sonnet 1h TTL write
- * \$3.75/M, read \$0.30/M, 比例 12.5×), 1b.5 pricing 模型**不**拆 cache_write vs cache_read 字段.
- * 为避免假装知道 cache_creation 价格, **保守**: cache_creation OR cache_read 任一非零 →
- * cost_turn/cost_currency 字段 absent. 留 sprint 2 加 `cache_write_per_m` 字段.
- * - 注意: tokens_uncached 仍**可**算 (是 input_tokens, 跟 cache 字段无关), 不受 cost 限制
- *
- * 接受 Usage | MessageDeltaUsage (后者无 cache_creation / cache_read, 视为 0).
- * Sprint 1b.5 Step 2: 流末尾 message_delta event.usage 是 MessageDeltaUsage, 走 delta 路径
- * 不算 cost, 真实生产靠 stream.finalMessage() 拿完整 Usage.
- */
-export function parseAnthropicUsage(usage, fallbackModel, pricing) {
-    // MessageDeltaUsage (流末尾 message_delta 事件): 只有 output_tokens, 拿不到 input/cache.
-    //   → 视为 delta 增量, 不算完整 cost. 真实生产靠 stream.finalMessage() 拿完整 Usage.
-    if (!('input_tokens' in usage)) {
-        return {
-            prompt_tokens: 0,
-            completion_tokens: usage.output_tokens,
-            total_tokens: usage.output_tokens,
-        };
-    }
-    // 此后 usage 一定是 AnthropicUsage (有 input_tokens + cache_creation + cache_read)
-    const full = usage;
-    // F4 修正: total_prompt = input + cache_creation + cache_read (官方文档)
-    const cacheCreation = full.cache_creation_input_tokens ?? 0;
-    const cacheRead = full.cache_read_input_tokens ?? 0;
-    const cached = cacheCreation + cacheRead;
-    const totalPrompt = full.input_tokens + cached;
-    const completion = full.output_tokens;
-    const total = totalPrompt + completion;
-    const out = {
-        prompt_tokens: totalPrompt, // 跟 DeepSeek OAI shape 字段对齐 (prompt = 全部输入, 含 cache)
-        completion_tokens: completion,
-        total_tokens: total,
-    };
-    if (cached > 0)
-        out.cached_tokens = cached;
-    // tokens_uncached 仍算 (跟 cache 字段无关, 跟 computeCost 内部算的 uncached 一致)
-    if (cached > 0)
-        out.tokens_uncached = full.input_tokens; // = totalPrompt - cached
-    // F4 保守: cache_creation OR cache_read 非零 → cost_turn 字段 absent. 留 sprint 2 加 cache_write_per_m 字段.
-    // 1b.5 pricing 模型只有 cache_miss / cache_hit / completion, 不能区分 cache_creation 跟
-    // cache_read 的不同价. 假装按 cache_hit 价算 cache_creation 会**低估** Sonnet 12.5×.
-    if (cached === 0) {
-        // 跟 parseOaiSseUsageField 一致, 透传 pricing + model 给 computeCost
-        // cached=0 (LLM 显式说 0 cache hit) → 走完整 4 字段路径, 跟 OAI shape 对齐
-        const breakdown = computeCost(pricing, fallbackModel, totalPrompt, completion, 0);
-        if (breakdown !== undefined) {
-            out.cache_hit_rate = breakdown.cache_hit_rate;
-            if (breakdown.cost_turn !== undefined)
-                out.cost_turn = breakdown.cost_turn;
-            if (breakdown.cost_currency !== undefined)
-                out.cost_currency = breakdown.cost_currency;
-            out.tokens_uncached = breakdown.tokens_uncached;
-        }
-    }
-    else {
-        // cache_creation OR cache_read 非零: cost 字段 absent, 但 cache_hit_rate 仍算 (观测)
-        out.cache_hit_rate = totalPrompt > 0 ? cached / totalPrompt : 0;
-    }
-    return out;
-}
-/**
- * 把单个 RawMessageStreamEvent 翻译成 ChatChunk.
- * 返 null 表示跳过 (heartbeat / ping / 不该透出的 event).
- *
- * Event 类型 (来自 SDK type):
- * - message_start: 携带 message metadata (model, id, usage.input_tokens=0) — 不透出
- * - content_block_start: 新的 text/tool_use/thinking block 起点 — 不透出 (避免 0 token 块)
- * - content_block_delta: text 增量 (delta.text) 或 input_json_delta (tool_use) — 透出 content
- * - content_block_stop: block 结束 — 不透出
- * - message_delta: 顶层 delta (stop_reason + usage 更新) — 透出 finish_reason + 最终 usage
- * - message_stop: 流结束 — 不透出 (caller 走 stream.finalMessage 拿 final Message)
- * - ping: heartbeat — 不透出
- */
-export function parseAnthropicSseEvent(event, fallbackModel, pricing) {
-    switch (event.type) {
-        case 'content_block_delta': {
-            const delta = event.delta;
-            if (delta.type === 'text_delta') {
-                return { delta: { content: delta.text } };
-            }
-            // input_json_delta (tool_use) / thinking_delta / signature_delta: 1b.5 跳过
-            return null;
-        }
-        case 'message_delta': {
-            // 顶层 delta: stop_reason + usage. 这是流末尾的最终 usage 更新.
-            const finishReason = mapStopReason(event.delta.stop_reason);
-            const usage = parseAnthropicUsage(event.usage, fallbackModel, pricing);
-            const chunk = { delta: {} };
-            if (finishReason !== undefined)
-                chunk.finish_reason = finishReason;
-            if (usage !== undefined)
-                chunk.usage = usage;
-            // 1b.5 简化: 即使 chunk 是空 delta, 也透出 (caller 看 finish_reason 收尾)
-            return chunk;
-        }
-        // 其他 event 类型 (start/stop) 不透出. ping event 不在 union (SDK 内部 filter)
-        case 'message_start':
-        case 'content_block_start':
-        case 'content_block_stop':
-        case 'message_stop':
-            return null;
-    }
-}
-//# sourceMappingURL=anthropic-client.js.map