llm-simple-router 0.6.3 → 0.6.4

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

@@ -4,9 +4,8 @@
  *
  * 算法：
  * 1. 收集所有 assistant 消息中的 tool_use ID
- * 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）
+ * 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）
  */
 export declare function patchOrphanToolResults(body: Record<string, unknown>): void;

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

@@ -4,10 +4,9 @@
  *
  * 算法：
  * 1. 收集所有 assistant 消息中的 tool_use ID
- * 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）
+ * 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）
  */
 export function patchOrphanToolResults(body) {
     if (!body.messages)
@@ -26,41 +25,35 @@ export function patchOrphanToolResults(body) {
             }
         }
     }
-    // Step 2: 移除孤儿 tool_result 块
-    let removedAny = false;
+    // Step 2: 将孤儿 tool_result 块转为 text，而非丢弃
+    let convertedAny = false;
     for (const msg of messages) {
         if (msg.role !== "user" || !Array.isArray(msg.content))
             continue;
         const blocks = msg.content;
-        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);
+        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;
+            }
+            else {
+                newBlocks.push(block);
             }
-            return true;
-        });
-        if (filtered.length < before) {
-            msg.content = filtered;
-            removedAny = true;
-        }
-    }
-    if (!removedAny)
-        return;
-    // 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);
+        if (changed) {
+            msg.content = newBlocks;
+            convertedAny = true;
         }
     }
-    // Step 4: 合并相邻的 user 消息
+    if (!convertedAny)
+        return;
+    // Step 3: 合并相邻的 user 消息（转换后可能产生连续 user 消息）
     mergeConsecutive(messages, "user");
-    // Step 5: 合并相邻的 assistant 消息（删除空 user 消息后可能产生）
+    // Step 4: 合并相邻的 assistant 消息
     mergeConsecutive(messages, "assistant");
 }
 function mergeConsecutive(messages, role) {

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

@@ -1,6 +1,10 @@
 /**
- * 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。
+ * 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。
  */
-export declare function patchMissingThinkingBlocks(body: Record<string, unknown>): void;
+export declare function patchNonDeepSeekToolMessages(body: Record<string, unknown>): void;

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

@@ -1,24 +1,69 @@
 /**
- * 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。
+ * 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。
  */
-export function patchMissingThinkingBlocks(body) {
+export function patchNonDeepSeekToolMessages(body) {
     if (!body.messages)
         return;
     const messages = body.messages;
-    // 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)
+    if (!Array.isArray(messages) || messages.length === 0)
         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 hasThinking = msg.content.some((b) => b && typeof b === "object" && b.type === "thinking");
-        if (!hasThinking) {
-            msg.content.unshift({ type: "thinking", thinking: "", signature: "" });
+        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);
+            }
         }
+        msg.content = newBlocks;
     }
 }

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

@@ -1,9 +1,18 @@
 interface ProviderInfo {
     base_url: string;
 }
+export interface ProviderPatchMeta {
+    types: string[];
+}
 /**
- * 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。
+ * 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）
  */
-export declare function applyProviderPatches(body: Record<string, unknown>, provider: ProviderInfo): void;
+export declare function applyProviderPatches(body: Record<string, unknown>, provider: ProviderInfo): {
+    body: Record<string, unknown>;
+    meta: ProviderPatchMeta;
+};
 export {};

package/dist/proxy/patch/index.js

@@ -1,12 +1,21 @@
 import { applyDeepSeekPatches } from "./deepseek/index.js";
+import { patchRouterSyntheticToolCalls } from "./router-cleanup.js";
 /**
- * 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。
+ * 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）
  */
 export function applyProviderPatches(body, provider) {
-    if (needsDeepSeekPatch(body, provider)) {
-        applyDeepSeekPatches(body);
+    // 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: [] } };
     }
+    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/patch/router-cleanup.d.ts

@@ -0,0 +1,13 @@
+/**
+ * 从消息中移除 router 注入的合成 tool_use/tool_result。
+ *
+ * Router 通过 AskUserQuestion 注入的 tool_use（ID 以 toolu_router_ 开头）是给客户端
+ * UI 用的交互机制，不是任何 LLM 生成的内容，不应出现在发送给上游 provider 的上下文中。
+ *
+ * 处理步骤：
+ * 1. 移除 assistant 消息中的 router 合成 tool_use 块
+ * 2. 移除 user 消息中对应的 tool_result 块
+ * 3. 移除内容为空的消息
+ * 4. 合并连续的同角色消息
+ */
+export declare function patchRouterSyntheticToolCalls(body: Record<string, unknown>): void;

package/dist/proxy/patch/router-cleanup.js

@@ -0,0 +1,87 @@
+import { TOOL_USE_ID_PREFIX } from "../enhancement/directive-parser.js";
+/**
+ * 从消息中移除 router 注入的合成 tool_use/tool_result。
+ *
+ * Router 通过 AskUserQuestion 注入的 tool_use（ID 以 toolu_router_ 开头）是给客户端
+ * UI 用的交互机制，不是任何 LLM 生成的内容，不应出现在发送给上游 provider 的上下文中。
+ *
+ * 处理步骤：
+ * 1. 移除 assistant 消息中的 router 合成 tool_use 块
+ * 2. 移除 user 消息中对应的 tool_result 块
+ * 3. 移除内容为空的消息
+ * 4. 合并连续的同角色消息
+ */
+export function patchRouterSyntheticToolCalls(body) {
+    if (!body.messages)
+        return;
+    const messages = body.messages;
+    if (!Array.isArray(messages) || messages.length === 0)
+        return;
+    // Step 1: 收集 router 合成的 tool_use ID，移除这些块
+    const removedIds = new Set();
+    for (const msg of messages) {
+        if (msg.role !== "assistant" || !Array.isArray(msg.content))
+            continue;
+        const blocks = msg.content;
+        const filtered = blocks.filter((block) => {
+            if (block?.type === "tool_use" &&
+                typeof block.id === "string" &&
+                block.id.startsWith(TOOL_USE_ID_PREFIX)) {
+                removedIds.add(block.id);
+                return false;
+            }
+            return true;
+        });
+        if (filtered.length !== blocks.length)
+            msg.content = filtered;
+    }
+    if (removedIds.size === 0)
+        return;
+    // Step 2: 移除对应的 tool_result
+    for (const msg of messages) {
+        if (msg.role !== "user" || !Array.isArray(msg.content))
+            continue;
+        const blocks = msg.content;
+        const filtered = blocks.filter((block) => !(block?.type === "tool_result" &&
+            typeof block.tool_use_id === "string" &&
+            removedIds.has(block.tool_use_id)));
+        if (filtered.length !== blocks.length)
+            msg.content = filtered;
+    }
+    // Step 3: 移除内容为空的消息
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const content = messages[i].content;
+        const empty = content == null ||
+            content === "" ||
+            (Array.isArray(content) && content.length === 0);
+        if (empty)
+            messages.splice(i, 1);
+    }
+    // Step 4: 合并连续的同角色消息
+    mergeConsecutive(messages, "user");
+    mergeConsecutive(messages, "assistant");
+}
+function mergeConsecutive(messages, role) {
+    let i = 1;
+    while (i < messages.length) {
+        if (messages[i].role === role && messages[i - 1].role === role) {
+            const prev = messages[i - 1];
+            const curr = messages[i];
+            prev.content = [
+                ...normalizeToArray(prev.content),
+                ...normalizeToArray(curr.content),
+            ];
+            messages.splice(i, 1);
+        }
+        else {
+            i++;
+        }
+    }
+}
+function normalizeToArray(content) {
+    if (Array.isArray(content))
+        return content;
+    if (typeof content === "string")
+        return [{ type: "text", text: content }];
+    return [{ type: "text", text: String(content ?? "") }];
+}

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

@@ -12,6 +12,7 @@ export interface RouteHandlerDeps {
     tracker?: RequestTracker;
     orchestrator: ProxyOrchestrator;
     usageWindowTracker?: import("./usage-window-tracker.js").UsageWindowTracker;
+    sessionTracker?: import("./loop-prevention/session-tracker.js").SessionTracker;
 }
 export declare function handleProxyRequest(request: FastifyRequest, reply: FastifyReply, apiType: "openai" | "anthropic", upstreamPath: string, errors: ProxyErrorFormatter, deps: RouteHandlerDeps, options?: {
     beforeSendProxy?: (body: Record<string, unknown>, isStream: boolean) => void;

package/dist/proxy/proxy-handler.js

@@ -1,4 +1,5 @@
 import { randomUUID } from "crypto";
+import { HTTP_UNPROCESSABLE_ENTITY } from "../constants.js";
 import { getProviderById, insertRequestLog } from "../db/index.js";
 import { decrypt } from "../utils/crypto.js";
 import { getSetting } from "../db/settings.js";
@@ -10,12 +11,15 @@ import { buildUpstreamHeaders, buildUpstreamUrl } from "./proxy-core.js";
 import { ProviderSwitchNeeded } from "./types.js";
 import { updateLogStreamContent, updateLogClientStatus } from "../db/index.js";
 import { insertRejectedLog } from "./log-helpers.js";
+import { ToolLoopGuard } from "./loop-prevention/tool-loop-guard.js";
+import { TOOL_USE_ID_PREFIX, TOOL_USE_ID_PROVIDER_PREFIX } from "./enhancement/directive-parser.js";
 import { buildTransportFn } from "./transport-fn.js";
 import { applyOverflowRedirect } from "./overflow.js";
 import { applyProviderPatches } from "./patch/index.js";
 const HTTP_ERROR_THRESHOLD = 400;
 const MAX_LOG_FIELD_LENGTH = 80;
 const UPSTREAM_ERROR_STATUS = 502;
+const TIER2_LOOP_THRESHOLD = 2;
 /** 从 TransportResult 中提取最终 HTTP status code */
 function getTransportStatusCode(result) {
     if (result.kind === "success" || result.kind === "error" || result.kind === "stream_error")
@@ -70,6 +74,44 @@ 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 routerKeyId = request.routerKey?.id ?? null;
+        const sessionKey = routerKeyId ? `${routerKeyId}:${sessionId}` : sessionId;
+        const lastToolUse = extractLastToolUse(request.body);
+        if (lastToolUse) {
+            const toolGuard = new ToolLoopGuard(deps.sessionTracker, {
+                enabled: true,
+                minConsecutiveCount: 3,
+                detectorConfig: { n: 6, windowSize: 500, repeatThreshold: 5 },
+            });
+            const checkResult = toolGuard.check(sessionKey, lastToolUse);
+            if (checkResult.detected) {
+                const loopCount = deps.sessionTracker.getLoopCount(sessionKey);
+                if (loopCount === 1) {
+                    // 层级 1：透明重试 — 注入中断提示词
+                    toolGuard.injectLoopBreakPrompt(request.body, apiType, lastToolUse.toolName);
+                    request.log.warn({ sessionId, toolName: lastToolUse.toolName, loopCount }, "Tool call loop detected, injecting break prompt");
+                }
+                else if (loopCount === TIER2_LOOP_THRESHOLD) {
+                    // 层级 2：优雅中断
+                    return reply.code(HTTP_UNPROCESSABLE_ENTITY).send({
+                        error: {
+                            type: "tool_call_loop_detected",
+                            message: `检测到工具调用循环（连续重复调用 "${lastToolUse.toolName}"）。请求已中断。`,
+                            suggestion: "请回顾对话历史，停止重复调用工具，直接告知用户当前的进展和遇到的问题。",
+                        },
+                    });
+                }
+                else {
+                    // 层级 3：直接断开
+                    request.log.warn({ sessionId, toolName: lastToolUse.toolName, loopCount }, "Tool call loop detected, hard disconnecting");
+                    reply.raw.destroy();
+                    return reply;
+                }
+            }
+        }
+    }
     if (interceptResponse)
         return handleIntercept(deps.db, apiType, request, reply, interceptResponse, clientModel, sessionId);
     return executeFailoverLoop({
@@ -91,7 +133,7 @@ async function executeFailoverLoop(ctx) {
             rootLogId = logId;
         const isFailoverIteration = rootLogId !== logId;
         const routerKeyId = request.routerKey?.id ?? null;
-        const body = request.body;
+        let body = request.body;
         const isStream = body.stream === true;
         const cliHdrs = request.headers;
         const rCtx = {
@@ -143,7 +185,8 @@ async function executeFailoverLoop(ctx) {
                 body.model = overflowResult.backend_model;
             }
         }
-        applyProviderPatches(body, provider);
+        const patchResult = applyProviderPatches(body, provider);
+        body = patchResult.body;
         const apiKey = decrypt(provider.api_key, getSetting(deps.db, "encryption_key"));
         options?.beforeSendProxy?.(body, isStream);
         const reqBodyStr = JSON.stringify(body);
@@ -242,3 +285,46 @@ async function executeFailoverLoop(ctx) {
         }
     }
 }
+function extractLastToolUse(body) {
+    const messages = body.messages;
+    if (!messages)
+        return null;
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const msg = messages[i];
+        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;
+                    }
+                    const inputStr = JSON.stringify(block.input ?? {});
+                    return {
+                        toolName: block.name,
+                        toolUseId: typeof block.id === "string" ? block.id : undefined,
+                        inputHash: simpleHash(inputStr),
+                        inputText: inputStr,
+                        timestamp: Date.now(),
+                    };
+                }
+            }
+        }
+    }
+    return null;
+}
+function simpleHash(s) {
+    const HASH_SHIFT = 5;
+    let hash = 0;
+    for (let i = 0; i < s.length; i++) {
+        const char = s.charCodeAt(i);
+        hash = ((hash << HASH_SHIFT) - hash) + char;
+        hash |= 0;
+    }
+    return String(hash);
+}

package/dist/proxy/stream-proxy.d.ts

@@ -1,7 +1,8 @@
 import type { FastifyReply } from "fastify";
 import type { RawHeaders, TransportResult } from "./types.js";
 import type { SSEMetricsTransform } from "../metrics/sse-metrics-transform.js";
+import type { StreamLoopGuard } from "./loop-prevention/stream-loop-guard.js";
 import { type BuildHeadersFn } from "./transport.js";
 export declare function callStream(backend: {
     base_url: string;
-}, apiKey: string, body: Record<string, unknown>, clientHeaders: RawHeaders, reply: FastifyReply, timeoutMs: number, upstreamPath: string, buildHeaders: BuildHeadersFn, metricsTransform?: SSEMetricsTransform, checkEarlyError?: (bufferedData: string) => boolean, compatResolve?: (result: TransportResult) => void): Promise<TransportResult>;
+}, apiKey: string, body: Record<string, unknown>, clientHeaders: RawHeaders, reply: FastifyReply, timeoutMs: number, upstreamPath: string, buildHeaders: BuildHeadersFn, metricsTransform?: SSEMetricsTransform, checkEarlyError?: (bufferedData: string) => boolean, compatResolve?: (result: TransportResult) => void, loopGuard?: StreamLoopGuard): Promise<TransportResult>;

package/dist/proxy/stream-proxy.js

@@ -11,6 +11,7 @@ class StreamProxy {
     metricsTransform;
     checkEarlyError;
     timeoutMs;
+    loopGuard;
     state = "BUFFERING";
     resolved = false;
     resolveFn = null;
@@ -26,13 +27,14 @@ class StreamProxy {
     // 流式阶段 SSE error 扫描缓冲（跨 chunk 边界匹配）
     sseScanBuffer = "";
     static SSE_SCAN_MAX = 8192;
-    constructor(statusCode, rawUpstreamHeaders, sentUpstreamHeaders, reply, metricsTransform, checkEarlyError, timeoutMs) {
+    constructor(statusCode, rawUpstreamHeaders, sentUpstreamHeaders, reply, metricsTransform, checkEarlyError, timeoutMs, loopGuard) {
         this.statusCode = statusCode;
         this.sentUpstreamHeaders = sentUpstreamHeaders;
         this.reply = reply;
         this.metricsTransform = metricsTransform;
         this.checkEarlyError = checkEarlyError;
         this.timeoutMs = timeoutMs;
+        this.loopGuard = loopGuard;
         this.sseHeaders = filterHeaders(rawUpstreamHeaders);
         this.sseHeaders["Content-Type"] = "text/event-stream";
         this.sseHeaders["Cache-Control"] = "no-cache";
@@ -211,6 +213,12 @@ 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;
+        }
     }
     onEnd() {
         if (this.resolved)
@@ -259,7 +267,7 @@ class StreamProxy {
     }
 }
 // ---------- callStream ----------
-export function callStream(backend, apiKey, body, clientHeaders, reply, timeoutMs, upstreamPath, buildHeaders, metricsTransform, checkEarlyError, compatResolve) {
+export function callStream(backend, apiKey, body, clientHeaders, reply, timeoutMs, upstreamPath, buildHeaders, metricsTransform, checkEarlyError, compatResolve, loopGuard) {
     return new Promise((resolve) => {
         const effectiveResolve = compatResolve ?? resolve;
         const url = new URL(buildUpstreamUrl(backend.base_url, upstreamPath));
@@ -283,7 +291,7 @@ export function callStream(backend, apiKey, body, clientHeaders, reply, timeoutM
                 });
                 return;
             }
-            const proxy = new StreamProxy(statusCode, upstreamRes.headers, upstreamHeaders, reply, metricsTransform, checkEarlyError, timeoutMs);
+            const proxy = new StreamProxy(statusCode, upstreamRes.headers, upstreamHeaders, reply, metricsTransform, checkEarlyError, timeoutMs, loopGuard);
             proxy.bindResolve(effectiveResolve);
             proxy.registerCloseHandler();
             // 无 early error checker 时直接开始流式传输

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

@@ -1,9 +1,11 @@
 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;

package/dist/proxy/transport-fn.js

@@ -2,6 +2,9 @@ import { callNonStream, callStream } from "./transport.js";
 import { SSEMetricsTransform } from "../metrics/sse-metrics-transform.js";
 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";
@@ -26,17 +29,32 @@ 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) => {
+                    p.request.log.warn({ logId: p.logId, reason }, "Stream loop detected, aborting");
+                });
+            }
             const metricsTransform = new SSEMetricsTransform(p.apiType, p.startTime, {
                 onMetrics: (m) => { p.tracker?.update(p.logId, { streamMetrics: toStreamMetrics(m) }); },
                 onChunk: (rawLine) => { p.tracker?.appendStreamChunk(p.logId, rawLine, p.apiType, STREAM_CONTENT_MAX_RAW, STREAM_CONTENT_MAX_TEXT); },
+                onContentDelta: streamLoopGuard ? (text) => streamLoopGuard.feed(text) : undefined,
             });
             const checkEarlyError = p.matcher ? (data) => p.matcher.test(UPSTREAM_SUCCESS, data) : undefined;
-            const streamResult = await callStream(p.provider, p.apiKey, p.body, p.cliHdrs, p.reply, p.streamTimeoutMs, p.upstreamPath, buildHeaders, metricsTransform, checkEarlyError);
+            const streamResult = await callStream(p.provider, p.apiKey, p.body, p.cliHdrs, p.reply, p.streamTimeoutMs, p.upstreamPath, buildHeaders, metricsTransform, checkEarlyError, undefined, streamLoopGuard);
             const m = (streamResult.kind === "stream_success" || streamResult.kind === "stream_abort")
                 ? streamResult.metrics : undefined;
             if (m)

package/frontend-dist/assets/{CardContent-BtAcFNMy.js → CardContent-CpiBn1Oc.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-BLX8zWc1.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-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};

package/frontend-dist/assets/{CardTitle-Bmwf1S5Y.js → CardTitle-dwtgd_nl.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-BLX8zWc1.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-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};

package/frontend-dist/assets/{CascadingModelSelect-CicfrqcY.js → CascadingModelSelect-CxEXwaeM.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-BLX8zWc1.js";import{b as h}from"./Teleport-BGbwtNTD.js";import{a as g}from"./PopperContent-B23SzU9H.js";import{n as _,r as v,t as y}from"./PopoverTrigger-BVsxIE2L.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`:`${e/1e3}K`}let a=s(()=>n.providers.map(e=>({key:e.provider.id,label:e.provider.name,options:e.models.map(e=>({value:e.name,label:e.name,tag:i(e.contextWindow)}))}))),c=s(()=>n.modelValue?{groupKey:n.modelValue.provider_id,value:n.modelValue.model}:void 0);function l(e){r(`update:modelValue`,{provider_id:e.groupKey,model:e.value})}return(t,n)=>(u(),o(O,{groups:a.value,"model-value":c.value,placeholder:e.placeholder,"onUpdate:modelValue":l},null,8,[`groups`,`model-value`,`placeholder`]))}});export{k as t};
+import{Dt as e,G as t,H as n,It as r,J as i,Pt as a,U as o,V as s,W as c,ct as l,ot as u,q as d,yt as f,z as p,zt as m}from"./button-q9xTxPJh.js";import{b as h}from"./Teleport-CRn-gy0B.js";import{a as g}from"./PopperContent-BoOYHCag.js";import{n as _,r as v,t as y}from"./PopoverTrigger-CIN3yOIw.js";var b=h(`chevron-right`,[[`path`,{d:`m9 18 6-6-6-6`,key:`mthhwq`}]]),x=[`onMouseenter`],S={class:`truncate max-w-40`},C=[`onMouseenter`],w=[`onClick`],T={class:`truncate`},E={key:0,class:`shrink-0 text-xs text-muted-foreground`},D={key:0,class:`px-2 py-1.5 text-sm text-muted-foreground`},O=i({__name:`CascadingSelect`,props:{groups:{},modelValue:{},placeholder:{default:`请选择...`}},emits:[`update:modelValue`],setup(i,{emit:h}){let O=i,k=h,A=e(!1),j=e(null),M=s(()=>{if(!O.modelValue)return``;let e=O.groups.find(e=>e.key===O.modelValue.groupKey);if(!e)return``;let t=e.options.find(e=>e.value===O.modelValue.value);return t?`${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};