llm-simple-router 0.6.4 → 0.6.5

package/config/recommended-retry-rules.json

@@ -6,6 +6,5 @@
   { "name": "ZAI 操作失败 (code 500)", "status_code": 400, "body_pattern": "\"type\"\\s*:\\s*\"error\".*\"code\"\\s*:\\s*\"500\"", "retry_strategy": "exponential", "retry_delay_ms": 5000, "max_retries": 10, "max_delay_ms": 60000 },
   { "name": "ZAI 速率限制 (HTTP 200, code 1302)", "status_code": 200, "body_pattern": "\"error\".*\"code\"\\s*:\\s*\"1302\"", "retry_strategy": "exponential", "retry_delay_ms": 5000, "max_retries": 10, "max_delay_ms": 60000 },
   { "name": "ZAI SSE 错误 (HTTP 200, code 500)", "status_code": 200, "body_pattern": "\"error\".*\"code\"\\s*:\\s*\"500\"", "retry_strategy": "exponential", "retry_delay_ms": 5000, "max_retries": 10, "max_delay_ms": 60000 },
-  { "name": "ZAI SSE 错误 (HTTP 200, code 1234)", "status_code": 200, "body_pattern": "\"error\".*\"code\"\\s*:\\s*\"1234\"", "retry_strategy": "exponential", "retry_delay_ms": 5000, "max_retries": 10, "max_delay_ms": 60000 },
-  { "name": "ZAI 过载限流 (HTTP 200, code 1305)", "status_code": 200, "body_pattern": "\"error\".*\"code\"\\s*:\\s*\"1305\"", "retry_strategy": "exponential", "retry_delay_ms": 5000, "max_retries": 10, "max_delay_ms": 60000 }
+  { "name": "ZAI SSE 错误 (HTTP 200, code 1234)", "status_code": 200, "body_pattern": "\"error\".*\"code\"\\s*:\\s*\"1234\"", "retry_strategy": "exponential", "retry_delay_ms": 5000, "max_retries": 10, "max_delay_ms": 60000 }
 ]

package/dist/admin/proxy-enhancement.js

@@ -3,6 +3,8 @@ import { setSetting } from "../db/settings.js";
 import { loadEnhancementConfig } from "../proxy/enhancement-config.js";
 const UpdateProxyEnhancementSchema = Type.Object({
     claude_code_enabled: Type.Boolean(),
+    tool_call_loop_enabled: Type.Boolean(),
+    stream_loop_enabled: Type.Boolean(),
 });
 const SessionParamsSchema = Type.Object({
     keyId: Type.String(),
@@ -16,12 +18,16 @@ export const adminProxyEnhancementRoutes = (app, options, done) => {
         const config = loadEnhancementConfig(db);
         return reply.send({
             claude_code_enabled: config.claude_code_enabled,
+            tool_call_loop_enabled: config.tool_call_loop_enabled,
+            stream_loop_enabled: config.stream_loop_enabled,
         });
     });
     app.put("/admin/api/proxy-enhancement", { schema: { body: UpdateProxyEnhancementSchema } }, async (request, reply) => {
         const body = request.body;
         const config = {
             claude_code_enabled: body.claude_code_enabled,
+            tool_call_loop_enabled: body.tool_call_loop_enabled,
+            stream_loop_enabled: body.stream_loop_enabled,
         };
         setSetting(db, "proxy_enhancement", JSON.stringify(config));
         return reply.send({ success: true });

package/dist/config.d.ts

@@ -1,4 +1,3 @@
-import type { LoopPreventionConfig } from "./proxy/loop-prevention/types.js";
 export interface Config {
     PORT: number;
     DB_PATH: string;
@@ -6,7 +5,6 @@ export interface Config {
     TZ: string;
     STREAM_TIMEOUT_MS: number;
     RETRY_BASE_DELAY_MS: number;
-    LOOP_PREVENTION: LoopPreventionConfig;
 }
 export declare function resetConfig(): void;
 export declare function getBaseConfig(): Config;

package/dist/config.js

@@ -1,6 +1,5 @@
 import { homedir } from "node:os";
 import { join } from "node:path";
-import { DEFAULT_LOOP_PREVENTION_CONFIG } from "./proxy/loop-prevention/types.js";
 let cachedConfig = null;
 function getDefaultDbPath() {
     if (process.env.DB_PATH)
@@ -20,10 +19,6 @@ export function getBaseConfig() {
         TZ: process.env.TZ || "Asia/Shanghai",
         STREAM_TIMEOUT_MS: parseInt(process.env.STREAM_TIMEOUT_MS || "3000000", 10),
         RETRY_BASE_DELAY_MS: parseInt(process.env.RETRY_BASE_DELAY_MS || "1000", 10),
-        LOOP_PREVENTION: {
-            ...DEFAULT_LOOP_PREVENTION_CONFIG,
-            ...(process.env.LOOP_PREVENTION ? JSON.parse(process.env.LOOP_PREVENTION) : {}),
-        },
     };
     return cachedConfig;
 }

package/dist/index.js

@@ -25,7 +25,6 @@ import { modelState } from "./proxy/model-state.js";
 import { UsageWindowTracker } from "./proxy/usage-window-tracker.js";
 import { SessionTracker } from "./proxy/loop-prevention/session-tracker.js";
 import { DEFAULT_LOOP_PREVENTION_CONFIG } from "./proxy/loop-prevention/types.js";
-import { setLoopPreventionConfig } from "./proxy/transport-fn.js";
 import { scheduleLogCleanup } from "./db/log-cleaner.js";
 import { scheduleDbSizeMonitor } from "./db/db-size-monitor.js";
 import { startUpgradeChecker, stopUpgradeChecker } from "./admin/upgrade.js";
@@ -159,14 +158,8 @@ export async function buildApp(options) {
     // 5h 用量窗口追踪器，启动时自动补齐缺失窗口
     const usageWindowTracker = new UsageWindowTracker(db);
     usageWindowTracker.reconcileOnStartup();
-    const loopConfig = config.LOOP_PREVENTION ?? DEFAULT_LOOP_PREVENTION_CONFIG;
-    const sessionTracker = new SessionTracker(loopConfig.sessionTracker);
-    // buildApp() 默认启用循环预防。
-    // 用户可通过环境变量 LOOP_PREVENTION='{"enabled":false}' 关闭。
-    // 直接注册插件的测试不使用 buildApp()，不受此影响。
-    setLoopPreventionConfig(process.env.LOOP_PREVENTION
-        ? loopConfig
-        : { ...loopConfig, enabled: true });
+    // Session tracker（工具调用循环检测用），始终创建但检测受 proxy_enhancement 配置控制
+    const sessionTracker = new SessionTracker(DEFAULT_LOOP_PREVENTION_CONFIG.sessionTracker);
     // 从 DB 读取已有 provider 的并发配置，初始化信号量管理器和 tracker
     const allProviders = getAllProviders(db);
     for (const p of allProviders) {

package/dist/metrics/sse-metrics-transform.d.ts

@@ -31,7 +31,7 @@ export declare class SSEMetricsTransform extends Transform {
     _transform(chunk: Buffer, _encoding: BufferEncoding, callback: TransformCallback): void;
     _flush(callback: TransformCallback): void;
     getExtractor(): MetricsExtractor;
-    /** 提取 SSE 事件中的内容文本，触发 onContentDelta 回调 */
+    /** 从 SSE 事件中提取内容文本，触发 onContentDelta 回调 */
     private emitContentDelta;
     /**
      * 从 SSE data 字段中提取实际内容文本（thinking / text / tool JSON delta）。

package/dist/metrics/sse-metrics-transform.js

@@ -47,7 +47,6 @@ export class SSEMetricsTransform extends Transform {
             this.extractor.processEvent(event);
             this.emitContentDelta(event);
         }
-        // flush 无条件推送最终状态，确保消费者能拿到完整指标
         if (this.onMetrics && !this.flushed) {
             this.flushed = true;
             this.lastCallbackTime = Date.now();
@@ -58,7 +57,7 @@ export class SSEMetricsTransform extends Transform {
     getExtractor() {
         return this.extractor;
     }
-    /** 提取 SSE 事件中的内容文本，触发 onContentDelta 回调 */
+    /** 从 SSE 事件中提取内容文本，触发 onContentDelta 回调 */
     emitContentDelta(event) {
         if (!this.onContentDelta || !event.data)
             return;

package/dist/proxy/enhancement/enhancement-handler.d.ts

@@ -1,5 +1,6 @@
 import type { FastifyRequest } from "fastify";
 import Database from "better-sqlite3";
+import { type EnhancementConfig } from "../enhancement-config.js";
 export interface InterceptResponse {
     statusCode: number;
     body: unknown;
@@ -18,6 +19,6 @@ export interface EnhancementResult {
  * 在代理转发前应用代理增强逻辑（指令解析 + 会话记忆 + 模型替换 + 命令拦截）。
  * 仅当 proxy_enhancement.claude_code_enabled 开启时生效。
  */
-export declare function applyEnhancement(db: Database.Database, request: FastifyRequest, clientModel: string, sessionId?: string): EnhancementResult;
+export declare function applyEnhancement(db: Database.Database, request: FastifyRequest, clientModel: string, sessionId?: string, enhancementConfig?: EnhancementConfig): EnhancementResult;
 /** 生成注入到非流式响应中的模型信息标签 */
 export declare function buildModelInfoTag(effectiveModel: string): string;

package/dist/proxy/enhancement/enhancement-handler.js

@@ -60,9 +60,9 @@ function buildDisplayModels(db, allowedModelsRaw) {
  * 在代理转发前应用代理增强逻辑（指令解析 + 会话记忆 + 模型替换 + 命令拦截）。
  * 仅当 proxy_enhancement.claude_code_enabled 开启时生效。
  */
-export function applyEnhancement(db, request, clientModel, sessionId) {
+export function applyEnhancement(db, request, clientModel, sessionId, enhancementConfig) {
     const nullResult = { effectiveModel: clientModel, originalModel: null, interceptResponse: null };
-    const enhancement = loadEnhancementConfig(db);
+    const enhancement = enhancementConfig ?? loadEnhancementConfig(db);
     if (!enhancement.claude_code_enabled) {
         return nullResult;
     }

package/dist/proxy/enhancement-config.d.ts

@@ -1,6 +1,8 @@
 import Database from "better-sqlite3";
 export interface EnhancementConfig {
     claude_code_enabled: boolean;
+    tool_call_loop_enabled: boolean;
+    stream_loop_enabled: boolean;
 }
 /** 集中加载 proxy_enhancement 配置，避免多处重复 getSetting + JSON.parse */
 export declare function loadEnhancementConfig(db: Database.Database): EnhancementConfig;

package/dist/proxy/enhancement-config.js

@@ -1,6 +1,8 @@
 import { getSetting } from "../db/settings.js";
 const DEFAULT_CONFIG = {
     claude_code_enabled: false,
+    tool_call_loop_enabled: false,
+    stream_loop_enabled: false,
 };
 /** 集中加载 proxy_enhancement 配置，避免多处重复 getSetting + JSON.parse */
 export function loadEnhancementConfig(db) {
@@ -11,6 +13,8 @@ export function loadEnhancementConfig(db) {
         const parsed = JSON.parse(raw);
         return {
             claude_code_enabled: parsed.claude_code_enabled ?? false,
+            tool_call_loop_enabled: parsed.tool_call_loop_enabled ?? false,
+            stream_loop_enabled: parsed.stream_loop_enabled ?? false,
         };
     }
     catch {

package/dist/proxy/patch/deepseek/index.d.ts

@@ -1,6 +1,6 @@
 /**
  * 按序执行所有 DeepSeek 特定补丁。
- * 非 DeepSeek tool 消息补丁先执行（转换 tool_use/tool_result），
+ * thinking 补丁先执行（影响 assistant 消息结构），
  * tool_result 配对修复后执行。
  */
 export declare function applyDeepSeekPatches(body: Record<string, unknown>): void;

package/dist/proxy/patch/deepseek/index.js

@@ -1,11 +1,11 @@
-import { patchNonDeepSeekToolMessages } from "./patch-thinking-blocks.js";
+import { patchMissingThinkingBlocks } from "./patch-thinking-blocks.js";
 import { patchOrphanToolResults } from "./patch-orphan-tool-results.js";
 /**
  * 按序执行所有 DeepSeek 特定补丁。
- * 非 DeepSeek tool 消息补丁先执行（转换 tool_use/tool_result），
+ * thinking 补丁先执行（影响 assistant 消息结构），
  * tool_result 配对修复后执行。
  */
 export function applyDeepSeekPatches(body) {
-    patchNonDeepSeekToolMessages(body);
+    patchMissingThinkingBlocks(body);
     patchOrphanToolResults(body);
 }

package/dist/proxy/patch/deepseek/patch-orphan-tool-results.d.ts

@@ -4,8 +4,9 @@
  *
  * 算法：
  * 1. 收集所有 assistant 消息中的 tool_use ID
- * 2. 将 tool_use_id 不在集合中的 tool_result 块转为 text（保留信息）
- * 3. 合并相邻的 user 消息（Anthropic API 不允许连续 user 消息）
- * 4. 合并相邻的 assistant 消息（同理）
+ * 2. 移除 tool_use_id 不在集合中的 tool_result 块
+ * 3. 移除清空后的空 user 消息
+ * 4. 合并相邻的 user 消息（Anthropic API 不允许连续 user 消息）
+ * 5. 合并相邻的 assistant 消息（同理）
  */
 export declare function patchOrphanToolResults(body: Record<string, unknown>): void;

package/dist/proxy/patch/deepseek/patch-orphan-tool-results.js

@@ -4,9 +4,10 @@
  *
  * 算法：
  * 1. 收集所有 assistant 消息中的 tool_use ID
- * 2. 将 tool_use_id 不在集合中的 tool_result 块转为 text（保留信息）
- * 3. 合并相邻的 user 消息（Anthropic API 不允许连续 user 消息）
- * 4. 合并相邻的 assistant 消息（同理）
+ * 2. 移除 tool_use_id 不在集合中的 tool_result 块
+ * 3. 移除清空后的空 user 消息
+ * 4. 合并相邻的 user 消息（Anthropic API 不允许连续 user 消息）
+ * 5. 合并相邻的 assistant 消息（同理）
  */
 export function patchOrphanToolResults(body) {
     if (!body.messages)
@@ -25,35 +26,41 @@ export function patchOrphanToolResults(body) {
             }
         }
     }
-    // Step 2: 将孤儿 tool_result 块转为 text，而非丢弃
-    let convertedAny = false;
+    // Step 2: 移除孤儿 tool_result 块
+    let removedAny = false;
     for (const msg of messages) {
         if (msg.role !== "user" || !Array.isArray(msg.content))
             continue;
         const blocks = msg.content;
-        const newBlocks = [];
-        let changed = false;
-        for (const block of blocks) {
-            if (block?.type === "tool_result" &&
-                typeof block.tool_use_id === "string" &&
-                !knownToolUseIds.has(block.tool_use_id)) {
-                newBlocks.push({ type: "text", text: JSON.stringify(block) });
-                changed = true;
+        const before = blocks.length;
+        const filtered = blocks.filter(block => {
+            if (block?.type === "tool_result" && typeof block.tool_use_id === "string") {
+                return knownToolUseIds.has(block.tool_use_id);
             }
-            else {
-                newBlocks.push(block);
-            }
-        }
-        if (changed) {
-            msg.content = newBlocks;
-            convertedAny = true;
+            return true;
+        });
+        if (filtered.length < before) {
+            msg.content = filtered;
+            removedAny = true;
         }
     }
-    if (!convertedAny)
+    if (!removedAny)
         return;
-    // Step 3: 合并相邻的 user 消息（转换后可能产生连续 user 消息）
+    // Step 3: 移除清空后的空 user 消息（向后遍历避免索引错乱）
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const msg = messages[i];
+        if (msg.role !== "user")
+            continue;
+        if (Array.isArray(msg.content) && msg.content.length === 0) {
+            messages.splice(i, 1);
+        }
+        else if (typeof msg.content === "string" && msg.content.trim() === "") {
+            messages.splice(i, 1);
+        }
+    }
+    // Step 4: 合并相邻的 user 消息
     mergeConsecutive(messages, "user");
-    // Step 4: 合并相邻的 assistant 消息
+    // Step 5: 合并相邻的 assistant 消息（删除空 user 消息后可能产生）
     mergeConsecutive(messages, "assistant");
 }
 function mergeConsecutive(messages, role) {

package/dist/proxy/patch/deepseek/patch-thinking-blocks.d.ts

@@ -1,10 +1,6 @@
 /**
- * 将非 DeepSeek 生成的消息中的 tool_use/tool_result 转为 text 块。
- *
- * 判断标准：assistant 消息含 tool_use 但（无 thinking 块或 thinking.signature 为空/缺失）。
- *
- * 背景：DeepSeek Anthropic API 开启 thinking 后要求含 tool_use 的 assistant 消息
- * 必须携带 thinking 块。跨模型切换（GLM → DeepSeek）时历史消息缺 thinking 会导致 400。
- * 转化为 text 规避格式校验，同时完整保留信息。
+ * DeepSeek thinking 协议实现不完整：开启 thinking 模式后部分轮次不返回 thinking block，
+ * 但后续请求要求历史 assistant 消息必须携带 thinking block。
+ * 在 content 数组开头补一个空 thinking block 以绕过上游校验。
  */
-export declare function patchNonDeepSeekToolMessages(body: Record<string, unknown>): void;
+export declare function patchMissingThinkingBlocks(body: Record<string, unknown>): void;

package/dist/proxy/patch/deepseek/patch-thinking-blocks.js

@@ -1,69 +1,24 @@
 /**
- * 将非 DeepSeek 生成的消息中的 tool_use/tool_result 转为 text 块。
- *
- * 判断标准：assistant 消息含 tool_use 但（无 thinking 块或 thinking.signature 为空/缺失）。
- *
- * 背景：DeepSeek Anthropic API 开启 thinking 后要求含 tool_use 的 assistant 消息
- * 必须携带 thinking 块。跨模型切换（GLM → DeepSeek）时历史消息缺 thinking 会导致 400。
- * 转化为 text 规避格式校验，同时完整保留信息。
+ * DeepSeek thinking 协议实现不完整：开启 thinking 模式后部分轮次不返回 thinking block，
+ * 但后续请求要求历史 assistant 消息必须携带 thinking block。
+ * 在 content 数组开头补一个空 thinking block 以绕过上游校验。
  */
-export function patchNonDeepSeekToolMessages(body) {
+export function patchMissingThinkingBlocks(body) {
     if (!body.messages)
         return;
     const messages = body.messages;
-    if (!Array.isArray(messages) || messages.length === 0)
+    // DeepSeek 可能在不传 thinking 参数时也启用 thinking 模式（从历史推断），
+    // 所以只要历史中存在任何 thinking block，就视为 thinking 模式激活。
+    const thinkingActive = !!body.thinking || messages.some((msg) => msg.role === "assistant" && Array.isArray(msg.content)
+        && msg.content.some((b) => b && typeof b === "object" && b.type === "thinking"));
+    if (!thinkingActive)
         return;
-    // Step 1: 识别非 DeepSeek 的 assistant 消息，转换 tool_use → text
-    const convertedIds = new Set();
     for (const msg of messages) {
         if (msg.role !== "assistant" || !Array.isArray(msg.content))
             continue;
-        const blocks = msg.content;
-        const hasToolUse = blocks.some((b) => b && typeof b === "object" && b.type === "tool_use");
-        if (!hasToolUse)
-            continue;
-        const thinkingBlock = blocks.find((b) => b && typeof b === "object" && b.type === "thinking");
-        const hasValidSignature = thinkingBlock &&
-            typeof thinkingBlock.signature === "string" &&
-            thinkingBlock.signature !== "";
-        // 有合法 signature → DeepSeek 原生，不动
-        if (hasValidSignature)
-            continue;
-        // 非 DeepSeek：将 tool_use 块替换为 text
-        const newBlocks = [];
-        for (const block of blocks) {
-            if (block && typeof block === "object" && block.type === "tool_use") {
-                newBlocks.push({ type: "text", text: JSON.stringify(block) });
-                if (typeof block.id === "string") {
-                    convertedIds.add(block.id);
-                }
-            }
-            else {
-                newBlocks.push(block);
-            }
-        }
-        msg.content = newBlocks;
-    }
-    if (convertedIds.size === 0)
-        return;
-    // Step 2: 转换对应 user 消息中的 tool_result → text
-    for (const msg of messages) {
-        if (msg.role !== "user" || !Array.isArray(msg.content))
-            continue;
-        const blocks = msg.content;
-        const newBlocks = [];
-        for (const block of blocks) {
-            if (block &&
-                typeof block === "object" &&
-                block.type === "tool_result" &&
-                typeof block.tool_use_id === "string" &&
-                convertedIds.has(block.tool_use_id)) {
-                newBlocks.push({ type: "text", text: JSON.stringify(block) });
-            }
-            else {
-                newBlocks.push(block);
-            }
+        const hasThinking = msg.content.some((b) => b && typeof b === "object" && b.type === "thinking");
+        if (!hasThinking) {
+            msg.content.unshift({ type: "thinking", thinking: "", signature: "" });
         }
-        msg.content = newBlocks;
     }
 }

package/dist/proxy/patch/index.d.ts

@@ -1,18 +1,9 @@
 interface ProviderInfo {
     base_url: string;
 }
-export interface ProviderPatchMeta {
-    types: string[];
-}
 /**
- * 通用消息补丁入口。
- * 返回新的 body（必要时深拷贝），不修改原始 body。
- * 执行顺序：
- * 1. 清理 router 合成的 tool_use/tool_result（通用，所有 provider）
- * 2. Provider-specific patches（如 DeepSeek thinking 校验）
+ * 根据 provider 信息分发到对应的补丁逻辑。
+ * 每个补丁直接修改 body，不返回新对象。
  */
-export declare function applyProviderPatches(body: Record<string, unknown>, provider: ProviderInfo): {
-    body: Record<string, unknown>;
-    meta: ProviderPatchMeta;
-};
+export declare function applyProviderPatches(body: Record<string, unknown>, provider: ProviderInfo): void;
 export {};

package/dist/proxy/patch/index.js

@@ -1,21 +1,12 @@
 import { applyDeepSeekPatches } from "./deepseek/index.js";
-import { patchRouterSyntheticToolCalls } from "./router-cleanup.js";
 /**
- * 通用消息补丁入口。
- * 返回新的 body（必要时深拷贝），不修改原始 body。
- * 执行顺序：
- * 1. 清理 router 合成的 tool_use/tool_result（通用，所有 provider）
- * 2. Provider-specific patches（如 DeepSeek thinking 校验）
+ * 根据 provider 信息分发到对应的补丁逻辑。
+ * 每个补丁直接修改 body，不返回新对象。
  */
 export function applyProviderPatches(body, provider) {
-    // router cleanup 始终执行，先克隆以避免修改原始 body
-    const cloned = JSON.parse(JSON.stringify(body));
-    patchRouterSyntheticToolCalls(cloned);
-    if (!needsDeepSeekPatch(body, provider)) {
-        return { body: cloned, meta: { types: [] } };
+    if (needsDeepSeekPatch(body, provider)) {
+        applyDeepSeekPatches(body);
     }
-    applyDeepSeekPatches(cloned);
-    return { body: cloned, meta: { types: ["deepseek_tool_use_to_text"] } };
 }
 /** DeepSeek patch 触发条件：直连 DeepSeek，或经代理转发且模型名含 deepseek */
 function needsDeepSeekPatch(body, provider) {

package/dist/proxy/proxy-handler.js

@@ -16,6 +16,7 @@ import { TOOL_USE_ID_PREFIX, TOOL_USE_ID_PROVIDER_PREFIX } from "./enhancement/d
 import { buildTransportFn } from "./transport-fn.js";
 import { applyOverflowRedirect } from "./overflow.js";
 import { applyProviderPatches } from "./patch/index.js";
+import { loadEnhancementConfig } from "./enhancement-config.js";
 const HTTP_ERROR_THRESHOLD = 400;
 const MAX_LOG_FIELD_LENGTH = 80;
 const UPSTREAM_ERROR_STATUS = 502;
@@ -73,9 +74,10 @@ export async function handleProxyRequest(request, reply, apiType, upstreamPath,
     });
     const clientModel = request.body.model || "unknown";
     const sessionId = request.headers["x-claude-code-session-id"];
-    const { effectiveModel, originalModel, interceptResponse } = applyEnhancement(deps.db, request, clientModel, sessionId);
-    // --- 工具调用循环检测 ---
-    if (deps.sessionTracker && sessionId) {
+    const enhancementConfig = loadEnhancementConfig(deps.db);
+    const { effectiveModel, originalModel, interceptResponse } = applyEnhancement(deps.db, request, clientModel, sessionId, enhancementConfig);
+    // --- 工具调用循环检测（受 proxy_enhancement 配置控制） ---
+    if (enhancementConfig.tool_call_loop_enabled && deps.sessionTracker && sessionId) {
         const routerKeyId = request.routerKey?.id ?? null;
         const sessionKey = routerKeyId ? `${routerKeyId}:${sessionId}` : sessionId;
         const lastToolUse = extractLastToolUse(request.body);
@@ -119,11 +121,12 @@ export async function handleProxyRequest(request, reply, apiType, upstreamPath,
         effectiveModel, originalModel,
         originalBody: JSON.parse(JSON.stringify(request.body)),
         sessionId,
+        streamLoopEnabled: enhancementConfig.stream_loop_enabled,
     });
 }
 // ---------- Failover loop ----------
 async function executeFailoverLoop(ctx) {
-    const { request, reply, apiType, upstreamPath, errors, deps, options, effectiveModel, originalModel, originalBody, sessionId } = ctx;
+    const { request, reply, apiType, upstreamPath, errors, deps, options, effectiveModel, originalModel, originalBody, sessionId, streamLoopEnabled } = ctx;
     const excludeTargets = [];
     let rootLogId = null;
     while (true) {
@@ -133,7 +136,7 @@ async function executeFailoverLoop(ctx) {
             rootLogId = logId;
         const isFailoverIteration = rootLogId !== logId;
         const routerKeyId = request.routerKey?.id ?? null;
-        let body = request.body;
+        const body = request.body;
         const isStream = body.stream === true;
         const cliHdrs = request.headers;
         const rCtx = {
@@ -185,8 +188,7 @@ async function executeFailoverLoop(ctx) {
                 body.model = overflowResult.backend_model;
             }
         }
-        const patchResult = applyProviderPatches(body, provider);
-        body = patchResult.body;
+        applyProviderPatches(body, provider);
         const apiKey = decrypt(provider.api_key, getSetting(deps.db, "encryption_key"));
         options?.beforeSendProxy?.(body, isStream);
         const reqBodyStr = JSON.stringify(body);
@@ -200,6 +202,7 @@ async function executeFailoverLoop(ctx) {
             provider, apiKey, body, cliHdrs, reply, upstreamPath, apiType,
             isStream, startTime, logId, effectiveModel, originalModel,
             streamTimeoutMs: deps.streamTimeoutMs, tracker: deps.tracker, matcher: deps.matcher, request,
+            streamLoopEnabled,
         });
         try {
             const resilienceResult = await deps.orchestrator.handle(request, reply, apiType, { resolved, provider, clientModel: effectiveModel, isStream, trackerId: logId, sessionId, clientRequest: clientReq, concurrencyOverride }, { retryBaseDelayMs: deps.retryBaseDelayMs, isFailover, ruleMatcher: deps.matcher, transportFn });
@@ -294,12 +297,10 @@ function extractLastToolUse(body) {
         if (msg.role === "assistant") {
             const content = msg.content;
             if (Array.isArray(content)) {
-                // 从后往前遍历，找到第一个非 router 合成的 tool_use
                 for (let j = content.length - 1; j >= 0; j--) {
                     const block = content[j];
                     if (block.type !== "tool_use")
                         continue;
-                    // 跳过 router 生成的 synthetic tool_use（如 AskUserQuestion 的模型选择）
                     const id = block.id;
                     if (id && (id.startsWith(TOOL_USE_ID_PREFIX) || id.startsWith(TOOL_USE_ID_PROVIDER_PREFIX))) {
                         continue;

package/dist/proxy/stream-proxy.js

@@ -213,8 +213,6 @@ class StreamProxy {
             }
         }
         this.pipeEntry.write(chunk);
-        // loopGuard 由 SSEMetricsTransform 的 onContentDelta 回调驱动，
-        // 此处仅检查是否已触发（触发后终止流）
         if (this.loopGuard?.isTriggered()) {
             this.terminal("stream_abort", { metrics: this.collectMetrics(false) });
             return;

package/dist/proxy/transport-fn.d.ts

@@ -1,11 +1,9 @@
 import type { FastifyReply, FastifyRequest } from "fastify";
 import { getProviderById } from "../db/index.js";
-import type { LoopPreventionConfig } from "./loop-prevention/types.js";
 import type { RawHeaders, TransportResult } from "./types.js";
 import type { Target } from "./strategy/types.js";
 import type { RequestTracker } from "../monitor/request-tracker.js";
 import type { RetryRuleMatcher } from "./retry-rules.js";
-export declare function setLoopPreventionConfig(config: LoopPreventionConfig): void;
 export interface TransportFnParams {
     provider: NonNullable<ReturnType<typeof getProviderById>>;
     apiKey: string;
@@ -23,5 +21,6 @@ export interface TransportFnParams {
     tracker?: RequestTracker;
     matcher?: RetryRuleMatcher;
     request: FastifyRequest;
+    streamLoopEnabled: boolean;
 }
 export declare function buildTransportFn(p: TransportFnParams): (target: Target) => Promise<TransportResult>;

package/dist/proxy/transport-fn.js

@@ -4,10 +4,12 @@ import { MetricsExtractor } from "../metrics/metrics-extractor.js";
 import { buildUpstreamHeaders } from "./proxy-core.js";
 import { StreamLoopGuard } from "./loop-prevention/stream-loop-guard.js";
 import { NGramLoopDetector } from "./loop-prevention/detectors/ngram-detector.js";
-import { DEFAULT_LOOP_PREVENTION_CONFIG } from "./loop-prevention/types.js";
 import { UPSTREAM_SUCCESS } from "./types.js";
 import { buildModelInfoTag } from "./enhancement/enhancement-handler.js";
 import { DEFAULT_MAX_RAW as STREAM_CONTENT_MAX_RAW, DEFAULT_MAX_TEXT as STREAM_CONTENT_MAX_TEXT } from "../monitor/stream-content-accumulator.js";
+const LOOP_DETECTOR_N = 6;
+const LOOP_DETECTOR_WINDOW_SIZE = 1000;
+const LOOP_DETECTOR_REPEAT_THRESHOLD = 10;
 function toStreamMetrics(m) {
     return {
         inputTokens: m.input_tokens,
@@ -29,22 +31,14 @@ function toStreamMetrics(m) {
         toolUseTokens: m.tool_use_tokens,
     };
 }
-let _loopConfig;
-export function setLoopPreventionConfig(config) {
-    _loopConfig = config;
-}
-function getLoopPreventionConfig() {
-    return _loopConfig ?? DEFAULT_LOOP_PREVENTION_CONFIG;
-}
 export function buildTransportFn(p) {
     const buildHeaders = (cliHdrs, key, bytes) => buildUpstreamHeaders(cliHdrs, key, bytes, p.apiType);
     // eslint-disable-next-line @typescript-eslint/no-unused-vars
     return async (_target) => {
         if (p.isStream) {
-            const loopPreventionConfig = getLoopPreventionConfig();
             let streamLoopGuard;
-            if (loopPreventionConfig.enabled && loopPreventionConfig.stream.enabled) {
-                streamLoopGuard = new StreamLoopGuard(loopPreventionConfig.stream, new NGramLoopDetector(loopPreventionConfig.stream.detectorConfig), (reason) => {
+            if (p.streamLoopEnabled) {
+                streamLoopGuard = new StreamLoopGuard({ enabled: true, detectorConfig: { n: LOOP_DETECTOR_N, windowSize: LOOP_DETECTOR_WINDOW_SIZE, repeatThreshold: LOOP_DETECTOR_REPEAT_THRESHOLD } }, new NGramLoopDetector({ n: LOOP_DETECTOR_N, windowSize: LOOP_DETECTOR_WINDOW_SIZE, repeatThreshold: LOOP_DETECTOR_REPEAT_THRESHOLD }), (reason) => {
                     p.request.log.warn({ logId: p.logId, reason }, "Stream loop detected, aborting");
                 });
             }

package/frontend-dist/assets/{CardContent-CpiBn1Oc.js → CardContent-DG1NiXMU.js}

@@ -1 +1 @@
-import{G as e,It as t,J as n,Pt as r,lt as i,ot as a,r as o}from"./button-q9xTxPJh.js";var s=[`data-size`],c=n({__name:`Card`,props:{class:{type:[Boolean,null,String,Object,Array]},size:{default:`default`}},setup(n){let c=n;return(l,u)=>(a(),e(`div`,{"data-slot":`card`,"data-size":n.size,class:t(r(o)(`ring-foreground/10 bg-card text-card-foreground gap-4 overflow-hidden rounded-lg py-4 text-sm ring-1 has-data-[slot=card-footer]:pb-0 has-[>img:first-child]:pt-0 data-[size=sm]:gap-3 data-[size=sm]:py-3 data-[size=sm]:has-data-[slot=card-footer]:pb-0 *:[img:first-child]:rounded-t-lg *:[img:last-child]:rounded-b-lg group/card flex flex-col`,c.class))},[i(l.$slots,`default`)],10,s))}}),l=n({__name:`CardContent`,props:{class:{type:[Boolean,null,String,Object,Array]}},setup(n){let s=n;return(n,c)=>(a(),e(`div`,{"data-slot":`card-content`,class:t(r(o)(`px-4 group-data-[size=sm]/card:px-3`,s.class))},[i(n.$slots,`default`)],2))}});export{c as n,l as t};
+import{G as e,It as t,J as n,Pt as r,lt as i,ot as a,r as o}from"./button-D4qBQ0nA.js";var s=[`data-size`],c=n({__name:`Card`,props:{class:{type:[Boolean,null,String,Object,Array]},size:{default:`default`}},setup(n){let c=n;return(l,u)=>(a(),e(`div`,{"data-slot":`card`,"data-size":n.size,class:t(r(o)(`ring-foreground/10 bg-card text-card-foreground gap-4 overflow-hidden rounded-lg py-4 text-sm ring-1 has-data-[slot=card-footer]:pb-0 has-[>img:first-child]:pt-0 data-[size=sm]:gap-3 data-[size=sm]:py-3 data-[size=sm]:has-data-[slot=card-footer]:pb-0 *:[img:first-child]:rounded-t-lg *:[img:last-child]:rounded-b-lg group/card flex flex-col`,c.class))},[i(l.$slots,`default`)],10,s))}}),l=n({__name:`CardContent`,props:{class:{type:[Boolean,null,String,Object,Array]}},setup(n){let s=n;return(n,c)=>(a(),e(`div`,{"data-slot":`card-content`,class:t(r(o)(`px-4 group-data-[size=sm]/card:px-3`,s.class))},[i(n.$slots,`default`)],2))}});export{c as n,l as t};

package/frontend-dist/assets/{CardTitle-dwtgd_nl.js → CardTitle-6JvqGNd9.js}

@@ -1 +1 @@
-import{G as e,It as t,J as n,Pt as r,lt as i,ot as a,r as o}from"./button-q9xTxPJh.js";var s=n({__name:`CardHeader`,props:{class:{type:[Boolean,null,String,Object,Array]}},setup(n){let s=n;return(n,c)=>(a(),e(`div`,{"data-slot":`card-header`,class:t(r(o)(`gap-1 rounded-t-xl px-4 group-data-[size=sm]/card:px-3 [.border-b]:pb-4 group-data-[size=sm]/card:[.border-b]:pb-3 group/card-header @container/card-header grid auto-rows-min items-start has-data-[slot=card-action]:grid-cols-[1fr_auto] has-data-[slot=card-description]:grid-rows-[auto_auto]`,s.class))},[i(n.$slots,`default`)],2))}}),c=n({__name:`CardTitle`,props:{class:{type:[Boolean,null,String,Object,Array]}},setup(n){let s=n;return(n,c)=>(a(),e(`div`,{"data-slot":`card-title`,class:t(r(o)(`text-base leading-snug font-medium group-data-[size=sm]/card:text-sm cn-font-heading`,s.class))},[i(n.$slots,`default`)],2))}});export{s as n,c as t};
+import{G as e,It as t,J as n,Pt as r,lt as i,ot as a,r as o}from"./button-D4qBQ0nA.js";var s=n({__name:`CardHeader`,props:{class:{type:[Boolean,null,String,Object,Array]}},setup(n){let s=n;return(n,c)=>(a(),e(`div`,{"data-slot":`card-header`,class:t(r(o)(`gap-1 rounded-t-xl px-4 group-data-[size=sm]/card:px-3 [.border-b]:pb-4 group-data-[size=sm]/card:[.border-b]:pb-3 group/card-header @container/card-header grid auto-rows-min items-start has-data-[slot=card-action]:grid-cols-[1fr_auto] has-data-[slot=card-description]:grid-rows-[auto_auto]`,s.class))},[i(n.$slots,`default`)],2))}}),c=n({__name:`CardTitle`,props:{class:{type:[Boolean,null,String,Object,Array]}},setup(n){let s=n;return(n,c)=>(a(),e(`div`,{"data-slot":`card-title`,class:t(r(o)(`text-base leading-snug font-medium group-data-[size=sm]/card:text-sm cn-font-heading`,s.class))},[i(n.$slots,`default`)],2))}});export{s as n,c as t};

package/frontend-dist/assets/{CascadingModelSelect-CxEXwaeM.js → CascadingModelSelect-BzDHmRLv.js}

@@ -1 +1 @@
-import{Dt as e,G as t,H as n,It as r,J as i,Pt as a,U as o,V as s,W as c,ct as l,ot as u,q as d,yt as f,z as p,zt as m}from"./button-q9xTxPJh.js";import{b as h}from"./Teleport-CRn-gy0B.js";import{a as g}from"./PopperContent-BoOYHCag.js";import{n as _,r as v,t as y}from"./PopoverTrigger-CIN3yOIw.js";var b=h(`chevron-right`,[[`path`,{d:`m9 18 6-6-6-6`,key:`mthhwq`}]]),x=[`onMouseenter`],S={class:`truncate max-w-40`},C=[`onMouseenter`],w=[`onClick`],T={class:`truncate`},E={key:0,class:`shrink-0 text-xs text-muted-foreground`},D={key:0,class:`px-2 py-1.5 text-sm text-muted-foreground`},O=i({__name:`CascadingSelect`,props:{groups:{},modelValue:{},placeholder:{default:`请选择...`}},emits:[`update:modelValue`],setup(i,{emit:h}){let O=i,k=h,A=e(!1),j=e(null),M=s(()=>{if(!O.modelValue)return``;let e=O.groups.find(e=>e.key===O.modelValue.groupKey);if(!e)return``;let t=e.options.find(e=>e.value===O.modelValue.value);return t?`${e.label} / ${t.label}`:``});function N(e,t){k(`update:modelValue`,{groupKey:e,value:t}),A.value=!1}function P(e){A.value=e,e||(j.value=null)}return(e,s)=>(u(),o(a(v),{open:A.value,"onUpdate:open":P},{default:f(()=>[d(a(y),{"as-child":``},{default:f(()=>[n(`div`,{class:r([`flex h-10 w-full items-center justify-between rounded-md border border-input bg-background px-3 py-2 text-sm ring-offset-background cursor-pointer hover:bg-accent hover:text-accent-foreground`,{"ring-2 ring-ring ring-offset-2":A.value}])},[n(`span`,{class:r([`truncate`,i.modelValue?`text-foreground`:`text-muted-foreground`])},m(M.value||i.placeholder),3),d(a(g),{class:`h-4 w-4 shrink-0 opacity-50`})],2)]),_:1}),d(a(_),{align:`start`,"side-offset":4,class:`z-[200] w-auto min-w-56 overflow-visible p-1`},{default:f(()=>[(u(!0),t(p,null,l(i.groups,e=>(u(),t(`div`,{key:e.key,class:r([`relative flex cursor-pointer items-center justify-between rounded-sm px-2 py-1.5 text-sm hover:bg-accent hover:text-accent-foreground`,{"bg-accent text-accent-foreground z-10":j.value===e.key}]),onMouseenter:t=>j.value=e.key},[n(`span`,S,m(e.label),1),d(a(b),{class:`ml-1 h-4 w-4 shrink-0 opacity-50`}),j.value===e.key&&e.options.length>0?(u(),t(`div`,{key:0,class:`absolute left-full top-0 ml-0.5 min-w-48 rounded-md border bg-popover p-1 text-popover-foreground shadow-md`,onMouseenter:t=>j.value=e.key},[(u(!0),t(p,null,l(e.options,a=>(u(),t(`div`,{key:a.value,class:r([`flex cursor-pointer items-center justify-between gap-2 rounded-sm px-2 py-1.5 text-sm hover:bg-accent hover:text-accent-foreground`,{"bg-accent text-accent-foreground":i.modelValue?.groupKey===e.key&&i.modelValue?.value===a.value}]),onClick:t=>N(e.key,a.value)},[n(`span`,T,m(a.label),1),a.tag?(u(),t(`span`,E,m(a.tag),1)):c(``,!0)],10,w))),128))],40,C)):c(``,!0)],42,x))),128)),i.groups.length===0?(u(),t(`div`,D,` 暂无选项 `)):c(``,!0)]),_:1})]),_:1},8,[`open`]))}}),k=i({__name:`CascadingModelSelect`,props:{providers:{},modelValue:{},placeholder:{default:`选择供应商 / 模型`}},emits:[`update:modelValue`],setup(e,{emit:t}){let n=e,r=t;function i(e){return e>=1e6?`${e/1e6}M`:`${e/1e3}K`}let a=s(()=>n.providers.map(e=>({key:e.provider.id,label:e.provider.name,options:e.models.map(e=>({value:e.name,label:e.name,tag:i(e.contextWindow)}))}))),c=s(()=>n.modelValue?{groupKey:n.modelValue.provider_id,value:n.modelValue.model}:void 0);function l(e){r(`update:modelValue`,{provider_id:e.groupKey,model:e.value})}return(t,n)=>(u(),o(O,{groups:a.value,"model-value":c.value,placeholder:e.placeholder,"onUpdate:modelValue":l},null,8,[`groups`,`model-value`,`placeholder`]))}});export{k as t};
+import{Dt as e,G as t,H as n,It as r,J as i,Pt as a,U as o,V as s,W as c,ct as l,ot as u,q as d,yt as f,z as p,zt as m}from"./button-D4qBQ0nA.js";import{b as h}from"./Teleport-BHTgdtZR.js";import{a as g}from"./PopperContent-BbE-uPX0.js";import{n as _,r as v,t as y}from"./PopoverTrigger-D0nT5UVQ.js";var b=h(`chevron-right`,[[`path`,{d:`m9 18 6-6-6-6`,key:`mthhwq`}]]),x=[`onMouseenter`],S={class:`truncate max-w-40`},C=[`onMouseenter`],w=[`onClick`],T={class:`truncate`},E={key:0,class:`shrink-0 text-xs text-muted-foreground`},D={key:0,class:`px-2 py-1.5 text-sm text-muted-foreground`},O=i({__name:`CascadingSelect`,props:{groups:{},modelValue:{},placeholder:{default:`请选择...`}},emits:[`update:modelValue`],setup(i,{emit:h}){let O=i,k=h,A=e(!1),j=e(null),M=s(()=>{if(!O.modelValue)return``;let e=O.groups.find(e=>e.key===O.modelValue.groupKey);if(!e)return``;let t=e.options.find(e=>e.value===O.modelValue.value);return t?`${e.label} / ${t.label}`:``});function N(e,t){k(`update:modelValue`,{groupKey:e,value:t}),A.value=!1}function P(e){A.value=e,e||(j.value=null)}return(e,s)=>(u(),o(a(v),{open:A.value,"onUpdate:open":P},{default:f(()=>[d(a(y),{"as-child":``},{default:f(()=>[n(`div`,{class:r([`flex h-10 w-full items-center justify-between rounded-md border border-input bg-background px-3 py-2 text-sm ring-offset-background cursor-pointer hover:bg-accent hover:text-accent-foreground`,{"ring-2 ring-ring ring-offset-2":A.value}])},[n(`span`,{class:r([`truncate`,i.modelValue?`text-foreground`:`text-muted-foreground`])},m(M.value||i.placeholder),3),d(a(g),{class:`h-4 w-4 shrink-0 opacity-50`})],2)]),_:1}),d(a(_),{align:`start`,"side-offset":4,class:`z-[200] w-auto min-w-56 overflow-visible p-1`},{default:f(()=>[(u(!0),t(p,null,l(i.groups,e=>(u(),t(`div`,{key:e.key,class:r([`relative flex cursor-pointer items-center justify-between rounded-sm px-2 py-1.5 text-sm hover:bg-accent hover:text-accent-foreground`,{"bg-accent text-accent-foreground z-10":j.value===e.key}]),onMouseenter:t=>j.value=e.key},[n(`span`,S,m(e.label),1),d(a(b),{class:`ml-1 h-4 w-4 shrink-0 opacity-50`}),j.value===e.key&&e.options.length>0?(u(),t(`div`,{key:0,class:`absolute left-full top-0 ml-0.5 min-w-48 rounded-md border bg-popover p-1 text-popover-foreground shadow-md`,onMouseenter:t=>j.value=e.key},[(u(!0),t(p,null,l(e.options,a=>(u(),t(`div`,{key:a.value,class:r([`flex cursor-pointer items-center justify-between gap-2 rounded-sm px-2 py-1.5 text-sm hover:bg-accent hover:text-accent-foreground`,{"bg-accent text-accent-foreground":i.modelValue?.groupKey===e.key&&i.modelValue?.value===a.value}]),onClick:t=>N(e.key,a.value)},[n(`span`,T,m(a.label),1),a.tag?(u(),t(`span`,E,m(a.tag),1)):c(``,!0)],10,w))),128))],40,C)):c(``,!0)],42,x))),128)),i.groups.length===0?(u(),t(`div`,D,` 暂无选项 `)):c(``,!0)]),_:1})]),_:1},8,[`open`]))}}),k=i({__name:`CascadingModelSelect`,props:{providers:{},modelValue:{},placeholder:{default:`选择供应商 / 模型`}},emits:[`update:modelValue`],setup(e,{emit:t}){let n=e,r=t;function i(e){return e>=1e6?`${e/1e6}M`:`${e/1e3}K`}let a=s(()=>n.providers.map(e=>({key:e.provider.id,label:e.provider.name,options:e.models.map(e=>({value:e.name,label:e.name,tag:i(e.contextWindow)}))}))),c=s(()=>n.modelValue?{groupKey:n.modelValue.provider_id,value:n.modelValue.model}:void 0);function l(e){r(`update:modelValue`,{provider_id:e.groupKey,model:e.value})}return(t,n)=>(u(),o(O,{groups:a.value,"model-value":c.value,placeholder:e.placeholder,"onUpdate:modelValue":l},null,8,[`groups`,`model-value`,`placeholder`]))}});export{k as t};