@acosmi/sdk-ts 2.19.3 → 2.19.5

package/dist/node/index.cjs

@@ -150,6 +150,14 @@ var init_types = __esm({
 });
 
 // src/shared/errors.ts
+function finiteNonNegative(value) {
+  return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : void 0;
+}
+function isStreamError(err) {
+  if (err == null || typeof err !== "object") return false;
+  const e = err;
+  return e.name === "StreamError" && typeof e.errorCode === "string";
+}
 function nullableID(value) {
   return value === null || typeof value === "string";
 }
@@ -359,8 +367,23 @@ var init_errors = __esm({
       userMessage;
       /** gateway 原始 error 字符串 */
       rawError;
-      /** 客户端是否值得重试 */
+      /** 客户端是否值得重试 (= 服务端 `retryable` **且** `requestDisposition === 'not_accepted'`) */
       retryable;
+      /**
+       * 服务端在帧上声明的 `retryable` **原值** (2.19.5 起) —— 不与 requestDisposition 合取,
+       * 缺席即 undefined。
+       *
+       * 为什么要并列存一份: 上面的 `retryable` 是两个判据的合取, 网关的流内错误帧并不总是
+       * 下发 `requestDisposition`, 于是"服务端明说可重试"会被合取成 false —— 两端各自正确,
+       * 合起来恒否。消费方要区分**「服务端明说不可重试」**(false, 一票否决) 与**「服务端没说」**
+       * (undefined, 回落到自己的本地分类表) 时, 只能读这个原值; 读 `retryable` 两种情形不可分。
+       */
+      serverRetryable;
+      /**
+       * 服务端建议的等待秒数 (2.19.5 起)。**有限非负数**才收, 其余 (null / 负数 /
+       * NaN / Infinity / 非数字) 一律 undefined —— 宁可不给, 绝不编造等待时间。
+       */
+      retryAfterSecs;
       errorContractVersion;
       faultDomain;
       requestDisposition;
@@ -383,6 +406,8 @@ var init_errors = __esm({
         this.userMessage = userMessage;
         this.rawError = rawError;
         this.retryable = retryable;
+        this.serverRetryable = opts.retryable;
+        this.retryAfterSecs = finiteNonNegative(opts.retryAfterSecs);
         this.errorContractVersion = opts.errorContractVersion;
         this.faultDomain = opts.faultDomain;
         this.requestDisposition = opts.requestDisposition;
@@ -407,272 +432,15 @@ var init_errors = __esm({
   }
 });
 
-// src/models/betas.ts
-function buildBetas(caps, req) {
-  const betas = [];
-  if (caps.supports_isp) {
-    betas.push(betaInterleavedThinking);
-    betas.push(betaContextManagement);
-  }
-  if (caps.supports_redact_thinking && req.thinking && req.thinking.display === "summary") {
-    betas.push(betaRedactThinking);
-  }
-  if (caps.supports_1m_context) {
-    betas.push(betaContext1M);
-  }
-  const hasStructuredOutput = caps.supports_structured_output && req.outputConfig != null;
-  if (hasStructuredOutput) {
-    betas.push(betaStructuredOutputs);
-  } else if (caps.supports_token_efficient) {
-    betas.push(betaTokenEfficientTools);
-  }
-  if (caps.supports_tool_search) {
-    betas.push(betaAdvancedToolUse);
-  }
-  const needsEffort = req.effort != null || req.thinking != null && req.thinking.level && req.thinking.level !== exports.ThinkingOff;
-  if (caps.supports_effort && needsEffort) {
-    betas.push(betaEffort);
-  }
-  if (caps.supports_fast_mode && req.speed === "fast") {
-    betas.push(betaFastMode);
-  }
-  if (caps.supports_prompt_cache) {
-    betas.push(betaPromptCachingScope);
-  }
-  return uniqueMerge(betas, req.betas ?? []);
-}
-function uniqueMerge(base, extra) {
-  if (extra.length === 0) return base;
-  const seen = new Set(base);
-  for (const s of extra) {
-    if (!seen.has(s)) {
-      base.push(s);
-      seen.add(s);
-    }
-  }
-  return base;
-}
-var betaInterleavedThinking, betaContext1M, betaContextManagement, betaStructuredOutputs, betaAdvancedToolUse, betaEffort, betaPromptCachingScope, betaFastMode, betaRedactThinking, betaTokenEfficientTools;
-var init_betas = __esm({
-  "src/models/betas.ts"() {
-    init_types();
-    betaInterleavedThinking = "interleaved-thinking-2025-05-14";
-    betaContext1M = "context-1m-2025-08-07";
-    betaContextManagement = "context-management-2025-06-27";
-    betaStructuredOutputs = "structured-outputs-2025-11-13";
-    betaAdvancedToolUse = "advanced-tool-use-2025-11-20";
-    betaEffort = "effort-2025-11-24";
-    betaPromptCachingScope = "prompt-caching-scope-2026-01-05";
-    betaFastMode = "fast-mode-2026-02-01";
-    betaRedactThinking = "redact-thinking-2026-02-12";
-    betaTokenEfficientTools = "token-efficient-tools-2025-02-19";
-  }
-});
-
-// src/models/adapters/anthropic.ts
-function resolveThinkingLevel(body, req, caps) {
-  const level = req.thinking?.level ?? "";
-  if (level === exports.ThinkingOff) {
-    body["thinking"] = { type: "disabled" };
-    return;
-  }
-  if (!caps.supports_adaptive_thinking && !caps.supports_thinking) {
-    return;
-  }
-  let maxTokens = req.max_tokens ?? 0;
-  if (maxTokens <= 0) {
-    maxTokens = exports.ThinkingHighMinMaxTokens;
-  }
-  if (level === exports.ThinkingMax) {
-    let modelMax = caps.max_output_tokens;
-    if (modelMax <= 0) modelMax = exports.ThinkingMaxFallbackMaxTokens;
-    if (maxTokens < modelMax) maxTokens = modelMax;
-  } else {
-    if (maxTokens < exports.ThinkingHighMinMaxTokens) maxTokens = exports.ThinkingHighMinMaxTokens;
-  }
-  body["max_tokens"] = maxTokens;
-  if (caps.supports_adaptive_thinking) {
-    const thinking = { type: "adaptive" };
-    if (req.thinking?.display) thinking["display"] = req.thinking.display;
-    body["thinking"] = thinking;
-  } else if (caps.supports_thinking) {
-    let budget = maxTokens - 1;
-    if (budget < 1024) budget = 1024;
-    const thinking = {
-      type: "enabled",
-      budget_tokens: budget
-    };
-    if (req.thinking?.display) thinking["display"] = req.thinking.display;
-    body["thinking"] = thinking;
-  }
-  if (caps.supports_effort) {
-    let effortLevel = "high";
-    if (level === "low") effortLevel = "low";
-    if (level === "medium" || level === "xhigh") effortLevel = level;
-    if (level === exports.ThinkingMax && caps.supports_max_effort) {
-      effortLevel = "max";
-    }
-    body["effort"] = { level: effortLevel };
-  }
-  delete body["temperature"];
-}
-var ANTHROPIC_SDK_MANAGED_BODY_KEYS; exports.AnthropicAdapter = void 0;
-var init_anthropic = __esm({
-  "src/models/adapters/anthropic.ts"() {
-    init_types();
-    init_errors();
-    init_betas();
-    init_adapters();
-    ANTHROPIC_SDK_MANAGED_BODY_KEYS = /* @__PURE__ */ new Set([
-      "thinking",
-      "effort",
-      "max_tokens",
-      "temperature",
-      "betas"
-    ]);
-    exports.AnthropicAdapter = class {
-      format() {
-        return 0 /* Anthropic */;
-      }
-      endpointSuffix() {
-        return "/anthropic";
-      }
-      /**
-       * 构建 Anthropic 格式请求体
-       * 逻辑等同于原 buildChatRequest, 包含完整的 betas/tools/serverTools/extraBody 处理
-       */
-      buildRequestBody(caps, req) {
-        const body = {};
-        if (req.rawMessages != null) {
-          body["messages"] = req.rawMessages;
-        } else if ((req.messages?.length ?? 0) > 0) {
-          body["messages"] = req.messages;
-        }
-        body["stream"] = req.stream === true;
-        if (req.max_tokens && req.max_tokens > 0) {
-          body["max_tokens"] = req.max_tokens;
-        }
-        if (req.system != null) {
-          body["system"] = req.system;
-        }
-        if (req.temperature != null) {
-          body["temperature"] = req.temperature;
-        }
-        if (req.thinking && req.thinking.level && req.thinking.level !== "") {
-          resolveThinkingLevel(body, req, caps);
-        } else {
-          if (req.thinking) {
-            body["thinking"] = req.thinking;
-          }
-          if (req.effort) {
-            body["effort"] = req.effort;
-          }
-        }
-        if (req.metadata || req.endUserId && req.endUserId !== "") {
-          const meta = {};
-          if (req.metadata) {
-            for (const [k, v] of Object.entries(req.metadata)) {
-              meta[k] = v;
-            }
-          }
-          if (req.endUserId && req.endUserId !== "" && meta["user_id"] === void 0) {
-            meta["user_id"] = req.endUserId;
-          }
-          body["metadata"] = meta;
-        }
-        const allTools = [];
-        if (req.tools != null) {
-          try {
-            const parsed = JSON.parse(JSON.stringify(req.tools));
-            if (Array.isArray(parsed)) {
-              for (const t of parsed) allTools.push(t);
-            }
-          } catch {
-          }
-        }
-        for (const st of req.serverTools ?? []) {
-          const schema = {
-            type: st.type,
-            name: st.name
-          };
-          if (st.config) {
-            for (const [k, v] of Object.entries(st.config)) {
-              schema[k] = v;
-            }
-          }
-          allTools.push(schema);
-        }
-        if (allTools.length > 0) {
-          body["tools"] = allTools;
-        }
-        if (req.speed && req.speed !== "") {
-          body["speed"] = req.speed;
-        }
-        if (req.outputConfig) {
-          body["output_config"] = req.outputConfig;
-        }
-        const betas = buildBetas(caps, req);
-        if (betas.length > 0) {
-          body["betas"] = betas;
-        }
-        if (req.extraBody) {
-          for (const [k, v] of Object.entries(req.extraBody)) {
-            if (ANTHROPIC_SDK_MANAGED_BODY_KEYS.has(k)) {
-              console.warn(
-                `acosmi-sdk: extraBody key "${k}" is SDK-managed and was ignored (use the dedicated request field instead).`
-              );
-              continue;
-            }
-            body[k] = v;
-          }
-        }
-        return body;
-      }
-      /**
-       * 解析 Anthropic 格式同步响应
-       * 兼容 APIResponse 包装 {"code":0,"data":{...}} 和裸 Anthropic JSON 两种格式
-       */
-      parseResponse(bodyInput) {
-        const bodyStr = typeof bodyInput === "string" ? bodyInput : new TextDecoder().decode(bodyInput);
-        let raw = bodyStr;
-        try {
-          const wrapper = JSON.parse(bodyStr);
-          if (wrapper.data != null && wrapper.data !== null) {
-            if ((wrapper.code ?? 0) !== 0) {
-              throw new exports.BusinessError(wrapper.code ?? 0, wrapper.message ?? "");
-            }
-            raw = JSON.stringify(wrapper.data);
-          }
-        } catch (e) {
-          if (e instanceof exports.BusinessError) throw e;
-        }
-        let resp;
-        try {
-          resp = JSON.parse(raw);
-        } catch (e) {
-          throw new Error(
-            `decode anthropic response: ${e instanceof Error ? e.message : String(e)}`
-          );
-        }
-        resp.tokenRemaining = -1;
-        resp.callRemaining = -1;
-        resp.modelTokenRemaining = -1;
-        resp.modelTokenRemainingETU = -1;
-        return resp;
-      }
-      /**
-       * 解析 Anthropic SSE 行
-       *
-       * Anthropic 原生协议无 [DONE] (message_stop 后上游关闭连接),
-       * 但 Nexus Gateway 的 ChatStream 路径会追加 [DONE] 哨兵, 此处一并处理。
-       */
-      parseStreamLine(eventType, data) {
-        if (data === "[DONE]") {
-          return { event: { event: "", data: "" }, done: true };
-        }
-        return { event: { event: eventType, data }, done: false };
-      }
-    };
+// src/models/adapters/format.ts
+exports.ProviderFormat = void 0;
+var init_format = __esm({
+  "src/models/adapters/format.ts"() {
+    exports.ProviderFormat = /* @__PURE__ */ ((ProviderFormat2) => {
+      ProviderFormat2[ProviderFormat2["Anthropic"] = 0] = "Anthropic";
+      ProviderFormat2[ProviderFormat2["OpenAI"] = 1] = "OpenAI";
+      return ProviderFormat2;
+    })(exports.ProviderFormat || {});
   }
 });
 
@@ -745,16 +513,19 @@ function convertOpenAIToChatResponse(oai) {
     role: "assistant",
     content: [],
     stop_reason: "",
+    // [W-SDK-OPENAI-PARITY] 缺 usage 的响应按 0 计数, 而不是让 `oai.usage.prompt_tokens`
+    // 抛 TypeError 把整条响应打死 —— 上游不计量是常见形态, 正文照样是有效的。
+    // 与 Go 侧同名函数逐字同语义 (那里 Usage 是值类型, 缺席即零值)。
     usage: {
-      input_tokens: oai.usage.prompt_tokens,
-      output_tokens: oai.usage.completion_tokens
+      input_tokens: oai.usage?.prompt_tokens ?? 0,
+      output_tokens: oai.usage?.completion_tokens ?? 0
     },
     tokenRemaining: -1,
     callRemaining: -1,
     modelTokenRemaining: -1,
     modelTokenRemainingETU: -1
   };
-  if (oai.choices.length > 0) {
+  if (Array.isArray(oai.choices) && oai.choices.length > 0) {
     const choice = oai.choices[0];
     switch (choice.finish_reason) {
       case "stop":
@@ -828,12 +599,14 @@ function parseOpenAIResponseToAnthropic(raw) {
     content: [],
     model: oaiResp.model,
     stop_reason: "",
+    // [W-SDK-OPENAI-PARITY] 同 convertOpenAIToChatResponse: 缺 usage 按 0 计数, 不抛。
+    // 两处都要改 —— 只改一处等于「非流式路径修好了一半」, 而两条路径是同一族上游响应。
     usage: {
-      input_tokens: oaiResp.usage.prompt_tokens,
-      output_tokens: oaiResp.usage.completion_tokens
+      input_tokens: oaiResp.usage?.prompt_tokens ?? 0,
+      output_tokens: oaiResp.usage?.completion_tokens ?? 0
     }
   };
-  if (oaiResp.choices.length > 0) {
+  if (Array.isArray(oaiResp.choices) && oaiResp.choices.length > 0) {
     const choice = oaiResp.choices[0];
     switch (choice.finish_reason) {
       case "stop":
@@ -871,6 +644,21 @@ function parseOpenAIResponseToAnthropic(raw) {
   }
   return resp;
 }
+function readStreamUsage(chunk) {
+  const raw = chunk.usage;
+  if (raw === null || typeof raw !== "object" || Array.isArray(raw)) {
+    return null;
+  }
+  const { prompt_tokens: promptTokens, completion_tokens: completionTokens } = raw;
+  const usage = {};
+  if (typeof promptTokens === "number" && Number.isFinite(promptTokens)) {
+    usage.input_tokens = promptTokens;
+  }
+  if (typeof completionTokens === "number" && Number.isFinite(completionTokens)) {
+    usage.output_tokens = completionTokens;
+  }
+  return usage;
+}
 function newOpenAIStreamConverter() {
   return new OpenAIStreamConverter();
 }
@@ -879,7 +667,7 @@ var init_openai = __esm({
   "src/models/adapters/openai.ts"() {
     init_types();
     init_errors();
-    init_adapters();
+    init_format();
     exports.OpenAIAdapter = class {
       format() {
         return 1 /* OpenAI */;
@@ -1004,8 +792,18 @@ var init_openai = __esm({
       toolArgsAccum = /* @__PURE__ */ new Map();
       /** 上一次解析出的 tool 键，供缺 index 且缺 id 的后续增量沿用。 */
       lastToolKey = null;
-      /** 已发出 message_delta/message_stop，避免 finish_reason 与 `[DONE]` 各收一次。 */
+      /** 已发出 message_delta/message_stop，避免 finish_reason、usage 尾帧、`[DONE]` 与
+       *  {@link flush} 重复收口 —— 整条流恰好一个 message_stop。 */
       messageClosed = false;
+      /** 已为仍打开的块发出 content_block_stop。与 {@link messageClosed} 分开记：块在
+       *  finish_reason 帧上就关，message_delta/message_stop 却可能推迟到之后的帧。 */
+      blocksClosed = false;
+      /** [W-QUAD-CHAIN-20260914] finish_reason 已到、但 message_delta/message_stop 因等待
+       *  usage 尾帧而推迟时记下的 stop_reason；没有推迟中的收尾时为 null。 */
+      pendingStopReason = null;
+      /** [W-QUAD-CHAIN-20260914] 最近一次带 usage 对象的帧搬出的 usage（后到覆盖先到）；
+       *  整条流从未出现 usage 对象时为 null，收尾 message_delta 据此不写 usage 键。 */
+      usage = null;
       /**
        * 解析一条 tool_call delta 归属的块键。
        *
@@ -1030,15 +828,18 @@ var init_openai = __esm({
         return 0;
       }
       /**
-       * 关闭仍打开的 text / thinking / tool 块，并收口 message。
+       * 关闭仍打开的 text / thinking / tool 块（整条流只关一次）。
        *
        * 由 `finish_reason` 分支与 `[DONE]` 分支共用：上游断流或只发 `[DONE]` 而不发
        * `finish_reason` 时，此前一个 `content_block_stop` 都不会发，下游拿到的是一个
        * 永不闭合的 tool_use 块。
+       *
+       * [W-QUAD-CHAIN-20260914] 此前关块与 message_delta/message_stop 是同一个方法里同一时刻
+       * 的事；usage 尾帧要求收尾推迟，于是拆成本方法与 {@link emitMessageEnd} 两段，各自防重。
        */
-      closeOpenBlocks(events, stopReason) {
-        if (this.messageClosed) return;
-        this.messageClosed = true;
+      closeContentBlocks(events) {
+        if (this.blocksClosed) return;
+        this.blocksClosed = true;
         if (this.textStarted) {
           events.push({
             event: "content_block_stop",
@@ -1058,10 +859,23 @@ var init_openai = __esm({
             data: JSON.stringify({ type: "content_block_stop", index: idx })
           });
         }
-        events.push({
-          event: "message_delta",
-          data: JSON.stringify({ type: "message_delta", delta: { stop_reason: stopReason } })
-        });
+      }
+      /**
+       * 发出 message_delta + message_stop，整条流只发一次；同时清掉推迟中的收尾。
+       *
+       * 见过 usage 对象就把它放进 message_delta：usage 必须出现在唯一的 message_stop 之前，
+       * message_stop 之后下游已无处安放用量。整条流从未出现 usage 对象时不写 usage 键 ——
+       * 缺席表示「上游没给」，不是 0。
+       */
+      emitMessageEnd(events, stopReason) {
+        if (this.messageClosed) return;
+        this.messageClosed = true;
+        this.pendingStopReason = null;
+        const messageDelta = { type: "message_delta", delta: { stop_reason: stopReason } };
+        if (this.usage !== null) {
+          messageDelta.usage = this.usage;
+        }
+        events.push({ event: "message_delta", data: JSON.stringify(messageDelta) });
         events.push({
           event: "message_stop",
           data: JSON.stringify({ type: "message_stop" })
@@ -1074,8 +888,11 @@ var init_openai = __esm({
       convert(data) {
         if (data === "[DONE]") {
           const events2 = [];
-          if (this.messageStarted) {
-            this.closeOpenBlocks(events2, "end_turn");
+          if (this.pendingStopReason !== null) {
+            this.emitMessageEnd(events2, this.pendingStopReason);
+          } else if (this.messageStarted) {
+            this.closeContentBlocks(events2);
+            this.emitMessageEnd(events2, "end_turn");
           }
           return { events: events2, done: true };
         }
@@ -1086,10 +903,18 @@ var init_openai = __esm({
           throw new Error(`parse openai stream chunk: ${e instanceof Error ? e.message : String(e)}`);
         }
         const events = [];
+        const frameUsage = readStreamUsage(chunk);
+        if (frameUsage !== null) {
+          this.usage = frameUsage;
+        }
         if (!Array.isArray(chunk.choices) || chunk.choices.length === 0) {
+          if (frameUsage !== null && this.pendingStopReason !== null) {
+            this.emitMessageEnd(events, this.pendingStopReason);
+          }
           return { events, done: false };
         }
         const choice = chunk.choices[0];
+        const delta = choice.delta ?? {};
         if (!this.messageStarted) {
           this.messageStarted = true;
           const msgJSON = JSON.stringify({
@@ -1104,8 +929,17 @@ var init_openai = __esm({
           });
           events.push({ event: "message_start", data: msgJSON });
         }
-        if (choice.delta.reasoning_content && choice.delta.reasoning_content !== "") {
+        if (delta.reasoning_content && delta.reasoning_content !== "") {
           if (!this.thinkingStarted) {
+            if (this.textStarted) {
+              const stopJSON = JSON.stringify({
+                type: "content_block_stop",
+                index: this.blockIndex
+              });
+              events.push({ event: "content_block_stop", data: stopJSON });
+              this.blockIndex++;
+              this.textStarted = false;
+            }
             this.thinkingStarted = true;
             this.thinkingBlockIndex = this.blockIndex;
             const blockJSON = JSON.stringify({
@@ -1118,11 +952,11 @@ var init_openai = __esm({
           const deltaJSON = JSON.stringify({
             type: "content_block_delta",
             index: this.thinkingBlockIndex,
-            delta: { type: "thinking_delta", thinking: choice.delta.reasoning_content }
+            delta: { type: "thinking_delta", thinking: delta.reasoning_content }
           });
           events.push({ event: "content_block_delta", data: deltaJSON });
         }
-        if (choice.delta.content && choice.delta.content !== "") {
+        if (delta.content && delta.content !== "") {
           if (this.thinkingStarted && !this.thinkingStopped) {
             this.thinkingStopped = true;
             const stopJSON = JSON.stringify({
@@ -1144,11 +978,11 @@ var init_openai = __esm({
           const deltaJSON = JSON.stringify({
             type: "content_block_delta",
             index: this.blockIndex,
-            delta: { type: "text_delta", text: choice.delta.content }
+            delta: { type: "text_delta", text: delta.content }
           });
           events.push({ event: "content_block_delta", data: deltaJSON });
         }
-        for (const tc of choice.delta.tool_calls ?? []) {
+        for (const tc of delta.tool_calls ?? []) {
           const fn = tc.function;
           const toolKey = this.resolveToolKey(tc);
           if (!this.toolBlockIndex.has(toolKey)) {
@@ -1223,67 +1057,39 @@ var init_openai = __esm({
             default:
               stopReason = choice.finish_reason;
           }
-          this.closeOpenBlocks(events, stopReason);
+          if (!this.messageClosed && this.pendingStopReason === null) {
+            this.closeContentBlocks(events);
+            if (this.usage !== null) {
+              this.emitMessageEnd(events, stopReason);
+            } else {
+              this.pendingStopReason = stopReason;
+            }
+          }
+        }
+        if (frameUsage !== null && this.pendingStopReason !== null) {
+          this.emitMessageEnd(events, this.pendingStopReason);
         }
         return { events, done: false };
       }
+      /**
+       * 流在**没有** `[DONE]` 的情况下结束 (EOF) 时, 由驱动方在读循环结束后调用一次。
+       *
+       * 只补发「finish_reason 已到、仅因等待 usage 尾帧而推迟」的 message_delta + message_stop。
+       * 从未收到 finish_reason 的流是被截断的流, 这里刻意**不**替它伪造正常结束 —— 一个
+       * `end_turn` 的 message_stop 会把被截断的回答当成完整回答交给下游。
+       * 已经收口 (usage 帧 / `[DONE]` / 上一次 flush) 后再调用返回空数组, 不会发出第二个 message_stop。
+       */
+      flush() {
+        const events = [];
+        if (this.pendingStopReason !== null) {
+          this.emitMessageEnd(events, this.pendingStopReason);
+        }
+        return events;
+      }
     };
   }
 });
 
-// src/models/adapters/index.ts
-function getAdapter(provider) {
-  const a = adapterRegistry[provider.toLowerCase()];
-  if (a) return a;
-  return defaultOpenAIAdapter;
-}
-function getAdapterForModel(m) {
-  let hasAnthropic = false;
-  let hasOpenAI = false;
-  for (const f of m.supported_formats ?? []) {
-    switch (f.trim().toLowerCase()) {
-      case "anthropic":
-        hasAnthropic = true;
-        break;
-      case "openai":
-        hasOpenAI = true;
-        break;
-    }
-  }
-  const declared = hasAnthropic || hasOpenAI;
-  const pref = (m.preferred_format ?? "").trim().toLowerCase();
-  switch (pref) {
-    case "anthropic":
-      if (!declared || hasAnthropic) return new exports.AnthropicAdapter();
-      break;
-    case "openai":
-      if (!declared || hasOpenAI) return new exports.OpenAIAdapter();
-      break;
-  }
-  if (hasAnthropic) return new exports.AnthropicAdapter();
-  if (hasOpenAI) return new exports.OpenAIAdapter();
-  return getAdapter((m.provider ?? "").toLowerCase());
-}
-exports.ProviderFormat = void 0; var adapterRegistry, defaultOpenAIAdapter;
-var init_adapters = __esm({
-  "src/models/adapters/index.ts"() {
-    init_anthropic();
-    init_openai();
-    init_openai();
-    exports.ProviderFormat = /* @__PURE__ */ ((ProviderFormat2) => {
-      ProviderFormat2[ProviderFormat2["Anthropic"] = 0] = "Anthropic";
-      ProviderFormat2[ProviderFormat2["OpenAI"] = 1] = "OpenAI";
-      return ProviderFormat2;
-    })(exports.ProviderFormat || {});
-    adapterRegistry = {
-      anthropic: new exports.AnthropicAdapter(),
-      /** Acosmi 自有模型走 Anthropic 格式 */
-      acosmi: new exports.AnthropicAdapter()
-    };
-    defaultOpenAIAdapter = new exports.OpenAIAdapter();
-  }
-});
-
 // src/core/credentials.ts
 var uuid = () => globalThis.crypto.randomUUID();
 var pause = () => new Promise((resolve) => setTimeout(resolve, 25));
@@ -2416,7 +2222,10 @@ var FileTokenStore = class {
   /**
    * 跨进程临界区. 用 sidecar `<path>.lock` 文件 + O_EXCL 创建语义实现互斥:
    *   - 创建成功 = 持有锁; 写入 pid+timestamp 便于诊断
-   *   - 创建失败 (EEXIST) = 别的进程持锁, backoff 重试
+   *   - 创建失败 (EEXIST) = 别的进程持锁, stale 检测后 backoff 重试
+   *   - 创建失败 (Windows 的 EPERM / EBUSY) = 前一持有者的 unlink 处于 delete-pending,
+   *     同属争用; 跳过 stale 检测直接 backoff 重试 (见 catch 内注释)。EACCES 不在其中 —
+   *     它是真权限错误, 立即抛
    *   - 锁文件 mtime > staleMs = 进程崩溃残留, unlink 后重试
    *   - acquireTimeoutMs 超时 = 抛错 (caller 应作为 transient error 处理, 上层 retry)
    *
@@ -2450,20 +2259,22 @@ ${Date.now()}
         };
         break;
       } catch (e) {
-        if (!isAlreadyExistsError(e)) throw e;
-        let stale = false;
-        try {
-          const st = await fs.stat(lockPath);
-          if (Date.now() - st.mtimeMs > fileLockDefaults.staleMs) stale = true;
-        } catch {
-          continue;
-        }
-        if (stale) {
+        if (!isLockContendedError(e)) throw e;
+        if (!isWindowsLockBusyError(e)) {
+          let stale = false;
           try {
-            await fs.unlink(lockPath);
+            const st = await fs.stat(lockPath);
+            if (Date.now() - st.mtimeMs > fileLockDefaults.staleMs) stale = true;
           } catch {
+            continue;
+          }
+          if (stale) {
+            try {
+              await fs.unlink(lockPath);
+            } catch {
+            }
+            continue;
           }
-          continue;
         }
         if (Date.now() - startMs > fileLockDefaults.acquireTimeoutMs) {
           throw new Error(
@@ -2569,6 +2380,15 @@ function isNotExistError(e) {
   }
   return false;
 }
+function isWindowsLockBusyError(e) {
+  if (typeof process === "undefined" || process.platform !== "win32") return false;
+  if (typeof e !== "object" || e === null || !("code" in e)) return false;
+  const code = e.code;
+  return code === "EPERM" || code === "EBUSY";
+}
+function isLockContendedError(e) {
+  return isAlreadyExistsError(e) || isWindowsLockBusyError(e);
+}
 function isAlreadyExistsError(e) {
   if (typeof e === "object" && e !== null && "code" in e) {
     return e.code === "EEXIST";
@@ -2674,8 +2494,308 @@ function computeBackoff(p, attempt, err) {
   return d;
 }
 
-// src/core/client.ts
-init_adapters();
+// src/models/adapters/anthropic.ts
+init_types();
+init_errors();
+
+// src/models/betas.ts
+init_types();
+var betaInterleavedThinking = "interleaved-thinking-2025-05-14";
+var betaContext1M = "context-1m-2025-08-07";
+var betaContextManagement = "context-management-2025-06-27";
+var betaStructuredOutputs = "structured-outputs-2025-11-13";
+var betaAdvancedToolUse = "advanced-tool-use-2025-11-20";
+var betaEffort = "effort-2025-11-24";
+var betaPromptCachingScope = "prompt-caching-scope-2026-01-05";
+var betaFastMode = "fast-mode-2026-02-01";
+var betaRedactThinking = "redact-thinking-2026-02-12";
+var betaTokenEfficientTools = "token-efficient-tools-2025-02-19";
+function buildBetas(caps, req) {
+  const betas = [];
+  if (caps.supports_isp) {
+    betas.push(betaInterleavedThinking);
+    betas.push(betaContextManagement);
+  }
+  if (caps.supports_redact_thinking && req.thinking && req.thinking.display === "summary") {
+    betas.push(betaRedactThinking);
+  }
+  if (caps.supports_1m_context) {
+    betas.push(betaContext1M);
+  }
+  const hasStructuredOutput = caps.supports_structured_output && req.outputConfig != null;
+  if (hasStructuredOutput) {
+    betas.push(betaStructuredOutputs);
+  } else if (caps.supports_token_efficient) {
+    betas.push(betaTokenEfficientTools);
+  }
+  if (caps.supports_tool_search) {
+    betas.push(betaAdvancedToolUse);
+  }
+  const needsEffort = req.effort != null || req.thinking != null && req.thinking.level && req.thinking.level !== exports.ThinkingOff;
+  if (caps.supports_effort && needsEffort) {
+    betas.push(betaEffort);
+  }
+  if (caps.supports_fast_mode && req.speed === "fast") {
+    betas.push(betaFastMode);
+  }
+  if (caps.supports_prompt_cache) {
+    betas.push(betaPromptCachingScope);
+  }
+  return uniqueMerge(betas, req.betas ?? []);
+}
+function uniqueMerge(base, extra) {
+  if (extra.length === 0) return base;
+  const seen = new Set(base);
+  for (const s of extra) {
+    if (!seen.has(s)) {
+      base.push(s);
+      seen.add(s);
+    }
+  }
+  return base;
+}
+
+// src/models/adapters/anthropic.ts
+init_format();
+var ANTHROPIC_SDK_MANAGED_BODY_KEYS = /* @__PURE__ */ new Set([
+  "thinking",
+  "effort",
+  "max_tokens",
+  "temperature",
+  "betas"
+]);
+var AnthropicAdapter = class {
+  format() {
+    return 0 /* Anthropic */;
+  }
+  endpointSuffix() {
+    return "/anthropic";
+  }
+  /**
+   * 构建 Anthropic 格式请求体
+   * 逻辑等同于原 buildChatRequest, 包含完整的 betas/tools/serverTools/extraBody 处理
+   */
+  buildRequestBody(caps, req) {
+    const body = {};
+    if (req.rawMessages != null) {
+      body["messages"] = req.rawMessages;
+    } else if ((req.messages?.length ?? 0) > 0) {
+      body["messages"] = req.messages;
+    }
+    body["stream"] = req.stream === true;
+    if (req.max_tokens && req.max_tokens > 0) {
+      body["max_tokens"] = req.max_tokens;
+    }
+    if (req.system != null) {
+      body["system"] = req.system;
+    }
+    if (req.temperature != null) {
+      body["temperature"] = req.temperature;
+    }
+    if (req.thinking && req.thinking.level && req.thinking.level !== "") {
+      resolveThinkingLevel(body, req, caps);
+    } else {
+      if (req.thinking) {
+        body["thinking"] = req.thinking;
+      }
+      if (req.effort) {
+        body["effort"] = req.effort;
+      }
+    }
+    if (req.metadata || req.endUserId && req.endUserId !== "") {
+      const meta = {};
+      if (req.metadata) {
+        for (const [k, v] of Object.entries(req.metadata)) {
+          meta[k] = v;
+        }
+      }
+      if (req.endUserId && req.endUserId !== "" && meta["user_id"] === void 0) {
+        meta["user_id"] = req.endUserId;
+      }
+      body["metadata"] = meta;
+    }
+    const allTools = [];
+    if (req.tools != null) {
+      try {
+        const parsed = JSON.parse(JSON.stringify(req.tools));
+        if (Array.isArray(parsed)) {
+          for (const t of parsed) allTools.push(t);
+        }
+      } catch {
+      }
+    }
+    for (const st of req.serverTools ?? []) {
+      const schema = {
+        type: st.type,
+        name: st.name
+      };
+      if (st.config) {
+        for (const [k, v] of Object.entries(st.config)) {
+          schema[k] = v;
+        }
+      }
+      allTools.push(schema);
+    }
+    if (allTools.length > 0) {
+      body["tools"] = allTools;
+    }
+    if (req.speed && req.speed !== "") {
+      body["speed"] = req.speed;
+    }
+    if (req.outputConfig) {
+      body["output_config"] = req.outputConfig;
+    }
+    const betas = buildBetas(caps, req);
+    if (betas.length > 0) {
+      body["betas"] = betas;
+    }
+    if (req.extraBody) {
+      for (const [k, v] of Object.entries(req.extraBody)) {
+        if (ANTHROPIC_SDK_MANAGED_BODY_KEYS.has(k)) {
+          console.warn(
+            `acosmi-sdk: extraBody key "${k}" is SDK-managed and was ignored (use the dedicated request field instead).`
+          );
+          continue;
+        }
+        body[k] = v;
+      }
+    }
+    return body;
+  }
+  /**
+   * 解析 Anthropic 格式同步响应
+   * 兼容 APIResponse 包装 {"code":0,"data":{...}} 和裸 Anthropic JSON 两种格式
+   */
+  parseResponse(bodyInput) {
+    const bodyStr = typeof bodyInput === "string" ? bodyInput : new TextDecoder().decode(bodyInput);
+    let raw = bodyStr;
+    try {
+      const wrapper = JSON.parse(bodyStr);
+      if (wrapper.data != null && wrapper.data !== null) {
+        if ((wrapper.code ?? 0) !== 0) {
+          throw new exports.BusinessError(wrapper.code ?? 0, wrapper.message ?? "");
+        }
+        raw = JSON.stringify(wrapper.data);
+      }
+    } catch (e) {
+      if (e instanceof exports.BusinessError) throw e;
+    }
+    let resp;
+    try {
+      resp = JSON.parse(raw);
+    } catch (e) {
+      throw new Error(
+        `decode anthropic response: ${e instanceof Error ? e.message : String(e)}`
+      );
+    }
+    resp.tokenRemaining = -1;
+    resp.callRemaining = -1;
+    resp.modelTokenRemaining = -1;
+    resp.modelTokenRemainingETU = -1;
+    return resp;
+  }
+  /**
+   * 解析 Anthropic SSE 行
+   *
+   * Anthropic 原生协议无 [DONE] (message_stop 后上游关闭连接),
+   * 但 Nexus Gateway 的 ChatStream 路径会追加 [DONE] 哨兵, 此处一并处理。
+   */
+  parseStreamLine(eventType, data) {
+    if (data === "[DONE]") {
+      return { event: { event: "", data: "" }, done: true };
+    }
+    return { event: { event: eventType, data }, done: false };
+  }
+};
+function resolveThinkingLevel(body, req, caps) {
+  const level = req.thinking?.level ?? "";
+  if (level === exports.ThinkingOff) {
+    body["thinking"] = { type: "disabled" };
+    return;
+  }
+  if (!caps.supports_adaptive_thinking && !caps.supports_thinking) {
+    return;
+  }
+  let maxTokens = req.max_tokens ?? 0;
+  if (maxTokens <= 0) {
+    maxTokens = exports.ThinkingHighMinMaxTokens;
+  }
+  if (level === exports.ThinkingMax) {
+    let modelMax = caps.max_output_tokens;
+    if (modelMax <= 0) modelMax = exports.ThinkingMaxFallbackMaxTokens;
+    if (maxTokens < modelMax) maxTokens = modelMax;
+  } else {
+    if (maxTokens < exports.ThinkingHighMinMaxTokens) maxTokens = exports.ThinkingHighMinMaxTokens;
+  }
+  body["max_tokens"] = maxTokens;
+  if (caps.supports_adaptive_thinking) {
+    const thinking = { type: "adaptive" };
+    if (req.thinking?.display) thinking["display"] = req.thinking.display;
+    body["thinking"] = thinking;
+  } else if (caps.supports_thinking) {
+    let budget = maxTokens - 1;
+    if (budget < 1024) budget = 1024;
+    const thinking = {
+      type: "enabled",
+      budget_tokens: budget
+    };
+    if (req.thinking?.display) thinking["display"] = req.thinking.display;
+    body["thinking"] = thinking;
+  }
+  if (caps.supports_effort) {
+    let effortLevel = "high";
+    if (level === "low") effortLevel = "low";
+    if (level === "medium" || level === "xhigh") effortLevel = level;
+    if (level === exports.ThinkingMax && caps.supports_max_effort) {
+      effortLevel = "max";
+    }
+    body["effort"] = { level: effortLevel };
+  }
+  delete body["temperature"];
+}
+
+// src/models/adapters/index.ts
+init_openai();
+init_format();
+init_openai();
+var adapterRegistry = {
+  anthropic: new AnthropicAdapter(),
+  /** Acosmi 自有模型走 Anthropic 格式 */
+  acosmi: new AnthropicAdapter()
+};
+var defaultOpenAIAdapter = new exports.OpenAIAdapter();
+function getAdapter(provider) {
+  const a = adapterRegistry[provider.toLowerCase()];
+  if (a) return a;
+  return defaultOpenAIAdapter;
+}
+function getAdapterForModel(m) {
+  let hasAnthropic = false;
+  let hasOpenAI = false;
+  for (const f of m.supported_formats ?? []) {
+    switch (f.trim().toLowerCase()) {
+      case "anthropic":
+        hasAnthropic = true;
+        break;
+      case "openai":
+        hasOpenAI = true;
+        break;
+    }
+  }
+  const declared = hasAnthropic || hasOpenAI;
+  const pref = (m.preferred_format ?? "").trim().toLowerCase();
+  switch (pref) {
+    case "anthropic":
+      if (!declared || hasAnthropic) return new AnthropicAdapter();
+      break;
+    case "openai":
+      if (!declared || hasOpenAI) return new exports.OpenAIAdapter();
+      break;
+  }
+  if (hasAnthropic) return new AnthropicAdapter();
+  if (hasOpenAI) return new exports.OpenAIAdapter();
+  return getAdapter((m.provider ?? "").toLowerCase());
+}
 
 // src/models/stream-meta.ts
 function extractAnthropicBlockMeta(eventType, data, blockTypeMap) {
@@ -2862,7 +2982,7 @@ function parseStreamError(data) {
   let code = payload.errorCode ?? "";
   const stage = payload.stage ?? "";
   let message = payload.message ?? "";
-  const retryable = payload.retryable ?? false;
+  const retryable = payload.retryable;
   let rawError = "";
   if (payload.error != null) {
     if (typeof payload.error === "string") {
@@ -2880,6 +3000,7 @@ function parseStreamError(data) {
     message,
     rawError,
     retryable,
+    retryAfterSecs: payload.retryAfterSecs,
     errorContractVersion: payload.errorContractVersion,
     faultDomain: payload.faultDomain,
     requestDisposition: payload.requestDisposition,
@@ -3090,6 +3211,25 @@ function notifyGatewayRequestID(cb, headers) {
   } catch {
   }
 }
+function notifyResponseHeaders(cb, headers) {
+  if (!cb) return;
+  try {
+    cb(headers);
+  } catch {
+  }
+}
+function armStreamRequestTimeout(opts, parent) {
+  const ms = opts?.requestTimeoutMs;
+  if (typeof ms !== "number" || !Number.isFinite(ms) || ms <= 0) return void 0;
+  return withRequestTimeout(ms, parent);
+}
+async function* disposeTimeoutWhenSettled(inner, ctl) {
+  try {
+    yield* inner;
+  } finally {
+    ctl.dispose();
+  }
+}
 var DEFAULT_API_TIMEOUT_MS = 6e4;
 function newDeferred() {
   let resolve;
@@ -4223,10 +4363,23 @@ var Client = class _Client {
    *
    * @param onUpstreamActivity 见 {@link UpstreamActivityCallback}
    * @param onGatewayRequestID 见 {@link GatewayRequestIDCallback}
+   * @param opts 见 {@link ChatStreamOptions} (2.19.5 起)
    */
-  chatStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID) {
+  chatStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID, opts) {
     return {
-      [Symbol.asyncIterator]: () => this.chatStreamGen(modelID, req, signal, false, onUpstreamActivity, onGatewayRequestID)
+      [Symbol.asyncIterator]: () => {
+        const ctl = armStreamRequestTimeout(opts, signal);
+        const inner = this.chatStreamGen(
+          modelID,
+          req,
+          ctl?.signal ?? signal,
+          false,
+          onUpstreamActivity,
+          onGatewayRequestID,
+          opts?.onResponseHeaders
+        );
+        return ctl ? disposeTimeoutWhenSettled(inner, ctl) : inner;
+      }
     };
   }
   /**
@@ -4236,13 +4389,26 @@ var Client = class _Client {
    *
    * @param onUpstreamActivity 见 {@link UpstreamActivityCallback}
    * @param onGatewayRequestID 见 {@link GatewayRequestIDCallback}
+   * @param opts 见 {@link ChatStreamOptions} (2.19.5 起)
    */
-  chatMessagesStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID) {
+  chatMessagesStream(modelID, req, signal, onUpstreamActivity, onGatewayRequestID, opts) {
     return {
-      [Symbol.asyncIterator]: () => this.chatMessagesStreamGen(modelID, req, signal, false, onUpstreamActivity, onGatewayRequestID)
+      [Symbol.asyncIterator]: () => {
+        const ctl = armStreamRequestTimeout(opts, signal);
+        const inner = this.chatMessagesStreamGen(
+          modelID,
+          req,
+          ctl?.signal ?? signal,
+          false,
+          onUpstreamActivity,
+          onGatewayRequestID,
+          opts?.onResponseHeaders
+        );
+        return ctl ? disposeTimeoutWhenSettled(inner, ctl) : inner;
+      }
     };
   }
-  async *chatStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID) {
+  async *chatStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID, onResponseHeaders) {
     const r = { ...req, stream: true };
     const { body, adapter } = await this.buildChatRequest(modelID, r, signal);
     const token = await this.ensureToken(signal);
@@ -4274,10 +4440,11 @@ var Client = class _Client {
           `stream: unauthorized and refresh failed: ${refreshErr instanceof Error ? refreshErr.message : String(refreshErr)}`
         );
       }
-      yield* this.chatStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID);
+      yield* this.chatStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID, onResponseHeaders);
       return;
     }
     notifyGatewayRequestID(onGatewayRequestID, resp.headers);
+    notifyResponseHeaders(onResponseHeaders, resp.headers);
     if (!resp.ok) {
       const bodyBytes = await readLimited(resp.body, maxErrorBodySize);
       throw parseHTTPErrorWithHeader(resp.status, bodyBytes, resp.headers);
@@ -4312,7 +4479,7 @@ var Client = class _Client {
       }
     }
   }
-  async *chatMessagesStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID) {
+  async *chatMessagesStreamGen(modelID, req, signal, retried, onUpstreamActivity, onGatewayRequestID, onResponseHeaders) {
     const r = { ...req, stream: true };
     const { body, adapter } = await this.buildChatRequest(modelID, r, signal);
     const token = await this.ensureToken(signal);
@@ -4344,10 +4511,11 @@ var Client = class _Client {
           `messages stream: unauthorized and refresh failed: ${refreshErr instanceof Error ? refreshErr.message : String(refreshErr)}`
         );
       }
-      yield* this.chatMessagesStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID);
+      yield* this.chatMessagesStreamGen(modelID, req, signal, true, onUpstreamActivity, onGatewayRequestID, onResponseHeaders);
       return;
     }
     notifyGatewayRequestID(onGatewayRequestID, resp.headers);
+    notifyResponseHeaders(onResponseHeaders, resp.headers);
     if (!resp.ok) {
       const bodyBytes = await readLimited(resp.body, maxErrorBodySize);
       throw parseHTTPErrorWithHeader(resp.status, bodyBytes, resp.headers);
@@ -4373,6 +4541,7 @@ var Client = class _Client {
           if (done) return;
         }
       }
+      for (const ev of converter.flush()) yield ev;
     } else {
       const blockTypeMap = /* @__PURE__ */ new Map();
       let currentEvent = "";
@@ -5290,10 +5459,6 @@ function findDesktopVisualUnderstandingModel(models) {
   return candidates[0];
 }
 
-// src/models/index.ts
-init_adapters();
-init_betas();
-
 // src/billing/entitlements.ts
 Client.prototype.getBalance = async function(signal) {
   const resp = await this.doJSON(
@@ -8893,6 +9058,7 @@ exports.ALL_PLATFORMS = ALL_PLATFORMS;
 exports.ALL_REGIONS = ALL_REGIONS;
 exports.AgentRunStreamError = AgentRunStreamError;
 exports.AgentRunsClient = AgentRunsClient;
+exports.AnthropicAdapter = AnthropicAdapter;
 exports.AudienceEnum = AudienceEnum;
 exports.BillingModeEnum = BillingModeEnum;
 exports.CHAT_REQUEST_TIMEOUT_MS = CHAT_REQUEST_TIMEOUT_MS;
@@ -9035,6 +9201,7 @@ exports.isPlatform = isPlatform;
 exports.isRegion = isRegion;
 exports.isSSECommentLine = isSSECommentLine;
 exports.isSSLError = isSSLError;
+exports.isStreamError = isStreamError;
 exports.isTerminalRemoteEvent = isTerminalRemoteEvent;
 exports.isUserAccessTokenRejected = isUserAccessTokenRejected;
 exports.isValidTokenSet = isValidTokenSet;