llm-simple-router 0.5.0 → 0.5.2

package/dist/proxy/{enhancement-handler.js → enhancement/enhancement-handler.js}

@@ -1,9 +1,9 @@
 import { randomUUID } from "crypto";
-import { getSetting } from "../db/settings.js";
-import { getActiveProviderModels, resolveByProviderModel } from "../db/index.js";
-import { resolveMapping } from "./mapping-resolver.js";
-import { parseDirective } from "./directive-parser.js";
-import { modelState } from "./model-state.js";
+import { getSetting } from "../../db/settings.js";
+import { getActiveProviderModels, resolveByProviderModel } from "../../db/index.js";
+import { resolveMapping } from "../mapping-resolver.js";
+import { parseDirective, parseToolResult, TOOL_USE_ID_PREFIX } from "./directive-parser.js";
+import { modelState } from "../model-state.js";
 import { cleanRouterResponses } from "./response-cleaner.js";
 const MODEL_INFO_TAG_TYPE = "model-info";
 /**
@@ -17,6 +17,42 @@ function resolveProviderModel(db, providerSlashModel) {
     const resolved = resolveByProviderModel(db, match[1], match[2]);
     return resolved?.client_model ?? null;
 }
+/** 检查请求中是否包含 AskUserQuestion 工具（判断客户端是否为 Claude Code） */
+function hasAskUserQuestion(body) {
+    const tools = body.tools;
+    if (!tools)
+        return false;
+    return tools.some(t => t.name === "AskUserQuestion");
+}
+/**
+ * 获取去重后的 provider/backend_model 显示列表，按 allowed_models 过滤。
+ * 供 buildSelectModelResponse 和 AskUserQuestion 路径复用。
+ */
+function buildDisplayModels(db, allowedModelsRaw) {
+    const providerModels = getActiveProviderModels(db);
+    let allowedSet = null;
+    if (allowedModelsRaw) {
+        try {
+            const parsed = JSON.parse(allowedModelsRaw).filter((s) => s.trim() !== "");
+            if (parsed.length > 0)
+                allowedSet = new Set(parsed);
+        }
+        catch { /* 解析失败时不做过滤 */ }
+    }
+    const filtered = allowedSet
+        ? providerModels.filter(m => allowedSet.has(m.backend_model))
+        : providerModels;
+    const seen = new Set();
+    const displayModels = [];
+    for (const m of filtered) {
+        const key = `${m.provider_name}/${m.backend_model}`;
+        if (!seen.has(key)) {
+            seen.add(key);
+            displayModels.push(key);
+        }
+    }
+    return displayModels;
+}
 /**
  * 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。
@@ -34,6 +70,35 @@ export function applyEnhancement(db, request, clientModel, sessionId) {
     if (enhancement?.claude_code_enabled !== true) {
         return nullResult;
     }
+    // 检测 AskUserQuestion 的 tool_result 回调（用户在 UI 上选择了模型）
+    const toolResult = parseToolResult(request.body);
+    if (toolResult.isRouterToolResult) {
+        const routerKeyId = request.routerKey?.id ?? null;
+        if (toolResult.selectedModel) {
+            const resolvedClientModel = resolveProviderModel(db, toolResult.selectedModel);
+            if (resolvedClientModel) {
+                modelState.set(routerKeyId, toolResult.selectedModel, sessionId, clientModel, "command");
+                return {
+                    effectiveModel: toolResult.selectedModel,
+                    originalModel: null,
+                    interceptResponse: {
+                        ...buildTextResponse("model-selected", `已选择模型: ${toolResult.selectedModel}`),
+                        meta: { action: "模型选择", detail: toolResult.selectedModel },
+                    },
+                };
+            }
+            return {
+                effectiveModel: clientModel,
+                originalModel: null,
+                interceptResponse: {
+                    ...buildTextResponse("error", `未找到模型: ${toolResult.selectedModel}`),
+                    meta: { action: "模型选择失败", detail: toolResult.selectedModel },
+                },
+            };
+        }
+        // tool_result 匹配前缀但无法提取模型名，降级处理
+        return nullResult;
+    }
     // 清理历史消息中的 <router-response> 标签
     const cleaned = cleanRouterResponses(request.body);
     request.body.messages = cleaned.messages;
@@ -67,6 +132,27 @@ export function applyEnhancement(db, request, clientModel, sessionId) {
             };
         }
         // 无参数：返回模型列表
+        if (hasAskUserQuestion(request.body)) {
+            const displayModels = buildDisplayModels(db, request.routerKey?.allowed_models ?? null);
+            if (displayModels.length === 0) {
+                return {
+                    effectiveModel: clientModel,
+                    originalModel: null,
+                    interceptResponse: {
+                        ...buildTextResponse("model-list", "（无可用模型）"),
+                        meta: { action: "模型列表" },
+                    },
+                };
+            }
+            return {
+                effectiveModel: clientModel,
+                originalModel: null,
+                interceptResponse: {
+                    ...buildAskUserQuestionResponse(displayModels),
+                    meta: { action: "模型列表(AskUserQuestion)" },
+                },
+            };
+        }
         return {
             effectiveModel: clientModel,
             originalModel: null,
@@ -119,34 +205,9 @@ function buildTextResponse(type, inner) {
     };
     return { statusCode: 200, body };
 }
-/** 查询所有可用的 provider_model 并构造响应 */
+/** 查询所有可用的 provider_model 并构造文本列表响应 */
 function buildSelectModelResponse(db, allowedModelsRaw, selectedModel) {
-    const providerModels = getActiveProviderModels(db);
-    // 按 allowed_models 过滤（allowed_models 存储的是 client_model 列表）
-    let allowedSet = null;
-    if (allowedModelsRaw) {
-        try {
-            const parsed = JSON.parse(allowedModelsRaw).filter((s) => s.trim() !== "");
-            if (parsed.length > 0)
-                allowedSet = new Set(parsed);
-        }
-        catch {
-            return buildTextResponse("model-list", "（解析 allowed_models 失败）");
-        }
-    }
-    const filtered = allowedSet
-        ? providerModels.filter(m => allowedSet.has(m.backend_model))
-        : providerModels;
-    // 去重并格式化为 "provider_name/backend_model"
-    const seen = new Set();
-    const displayModels = [];
-    for (const m of filtered) {
-        const key = `${m.provider_name}/${m.backend_model}`;
-        if (!seen.has(key)) {
-            seen.add(key);
-            displayModels.push(key);
-        }
-    }
+    const displayModels = buildDisplayModels(db, allowedModelsRaw);
     let inner;
     let responseType;
     if (selectedModel) {
@@ -163,6 +224,65 @@ function buildSelectModelResponse(db, allowedModelsRaw, selectedModel) {
     }
     return buildTextResponse(responseType, inner);
 }
+/** 将模型列表分块为 AskUserQuestion 的 questions（每组 2-4 个选项，最多 4 组） */
+function buildModelQuestions(models) {
+    if (models.length <= 4) {
+        const options = models.map(m => {
+            const sep = m.indexOf("/");
+            const provider = sep > 0 ? m.substring(0, sep) : "";
+            return { label: m, description: provider || "模型" };
+        });
+        if (options.length === 1) {
+            options.push({ label: "保持当前", description: "不切换模型" });
+        }
+        return [{
+                question: "请选择要使用的模型",
+                header: "模型选择",
+                options,
+                multiSelect: false,
+            }];
+    }
+    // 均匀分配到最多 4 个 question，每个 2-4 个选项
+    const numChunks = Math.min(4, Math.ceil(models.length / 4));
+    const chunks = Array.from({ length: numChunks }, () => []);
+    for (let i = 0; i < models.length; i++) {
+        chunks[i % numChunks].push(models[i]);
+    }
+    return chunks.map((chunk, idx) => ({
+        question: "请选择要使用的模型",
+        header: idx === 0 ? "模型选择" : "更多模型",
+        options: chunk.map(m => {
+            const sep = m.indexOf("/");
+            const provider = sep > 0 ? m.substring(0, sep) : "";
+            return { label: m, description: provider || "模型" };
+        }),
+        multiSelect: false,
+    }));
+}
+/** 构造 AskUserQuestion synthetic tool_use 响应 */
+function buildAskUserQuestionResponse(displayModels) {
+    const capped = displayModels.slice(0, 16);
+    const questions = buildModelQuestions(capped);
+    const toolUseId = `${TOOL_USE_ID_PREFIX}${randomUUID()}`;
+    return {
+        statusCode: 200,
+        body: {
+            id: `msg-${randomUUID()}`,
+            type: "message",
+            role: "assistant",
+            content: [{
+                    type: "tool_use",
+                    id: toolUseId,
+                    name: "AskUserQuestion",
+                    input: { questions },
+                }],
+            model: "router",
+            stop_reason: "tool_use",
+            stop_sequence: null,
+            usage: { input_tokens: 0, output_tokens: 0 },
+        },
+    };
+}
 /** 生成注入到非流式响应中的模型信息标签 */
 export function buildModelInfoTag(effectiveModel) {
     return `<router-response type="${MODEL_INFO_TAG_TYPE}">当前模型: ${effectiveModel}</router-response>`;

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

@@ -0,0 +1,3 @@
+export { applyEnhancement, buildModelInfoTag } from "./enhancement-handler.js";
+export { parseDirective, parseToolResult, TOOL_USE_ID_PREFIX } from "./directive-parser.js";
+export { cleanRouterResponses } from "./response-cleaner.js";

package/dist/proxy/enhancement/index.js

@@ -0,0 +1,3 @@
+export { applyEnhancement, buildModelInfoTag } from "./enhancement-handler.js";
+export { parseDirective, parseToolResult, TOOL_USE_ID_PREFIX } from "./directive-parser.js";
+export { cleanRouterResponses } from "./response-cleaner.js";

package/dist/proxy/{response-cleaner.js → enhancement/response-cleaner.js}

@@ -1,3 +1,4 @@
+import { TOOL_USE_ID_PREFIX } from "./directive-parser.js";
 const RE_ROUTER_RESPONSE = /<router-response[^>]*>[\s\S]*?<\/router-response>/g;
 const RE_COMMAND = /\[router-command:/;
 /**
@@ -30,9 +31,22 @@ export function cleanRouterResponses(body) {
                         return false;
                 }
             }
+            // 清理 router synthetic AskUserQuestion 的 tool_result 回调
+            const toolResultBlocks = blocks.filter((b) => b && typeof b === "object" && b.type === "tool_result"
+                && typeof b.tool_use_id === "string"
+                && b.tool_use_id.startsWith(TOOL_USE_ID_PREFIX));
+            if (toolResultBlocks.length > 0 && toolResultBlocks.length === blocks.length)
+                return false;
         }
         if (msg.role === "assistant") {
             const blocks = Array.isArray(msg.content) ? msg.content : [msg.content];
+            // 清理 router synthetic AskUserQuestion 的 tool_use 消息
+            const toolUseBlocks = blocks.filter((b) => b && typeof b === "object" && b.type === "tool_use"
+                && b.name === "AskUserQuestion"
+                && typeof b.id === "string"
+                && b.id.startsWith(TOOL_USE_ID_PREFIX));
+            if (toolUseBlocks.length > 0 && toolUseBlocks.length === blocks.length)
+                return false;
             const texts = blocks
                 .filter((b) => b?.type === "text" && typeof b.text === "string")
                 .map((b) => b.text);

package/dist/proxy/log-helpers.d.ts

@@ -1,6 +1,6 @@
 import Database from "better-sqlite3";
 import type { Provider } from "../db/index.js";
-import type { RawHeaders } from "./proxy-core.js";
+import type { RawHeaders } from "./types.js";
 export interface FailoverContext {
     isFailoverIteration: boolean;
     rootLogId: string;

package/dist/proxy/mapping-resolver.js

@@ -3,7 +3,7 @@ import { ScheduledStrategy } from "./strategy/scheduled.js";
 import { RoundRobinStrategy } from "./strategy/round-robin.js";
 import { RandomStrategy } from "./strategy/random.js";
 import { FailoverStrategy } from "./strategy/failover.js";
-import { getMappingGroup } from "../db/index.js";
+import { getMappingGroup, getActiveProviderByName, getActiveProvidersWithModels } from "../db/index.js";
 // 策略注册表：key 为数据库中 mapping_groups.strategy 字段的值。
 // 新增策略时：
 // 1. 在 src/proxy/strategy/ 下创建实现文件
@@ -21,7 +21,7 @@ export function resolveMapping(db, clientModel, context) {
     if (slashMatch) {
         const providerName = slashMatch[1];
         const backendModel = slashMatch[2];
-        const provider = db.prepare("SELECT id, models FROM providers WHERE name = ? AND is_active = 1").get(providerName);
+        const provider = getActiveProviderByName(db, providerName);
         if (provider) {
             try {
                 const models = JSON.parse(provider.models);
@@ -39,7 +39,7 @@ export function resolveMapping(db, clientModel, context) {
     const group = getMappingGroup(db, clientModel);
     if (!group) {
         // Fallback: 没有 mapping group 时，直接查 provider 的 models 字段
-        const providers = db.prepare("SELECT id, models FROM providers WHERE is_active = 1").all();
+        const providers = getActiveProvidersWithModels(db);
         for (const p of providers) {
             try {
                 const models = JSON.parse(p.models);
@@ -48,7 +48,7 @@ export function resolveMapping(db, clientModel, context) {
                 }
             }
             catch {
-                break;
+                continue;
             }
         }
         return null;

package/dist/proxy/openai.d.ts

@@ -3,13 +3,14 @@ import Database from "better-sqlite3";
 import { RetryRuleMatcher } from "./retry-rules.js";
 import { ProviderSemaphoreManager } from "./semaphore.js";
 import type { RequestTracker } from "../monitor/request-tracker.js";
+import type { UsageWindowTracker } from "./usage-window-tracker.js";
 export interface OpenaiProxyOptions {
     db: Database.Database;
     streamTimeoutMs: number;
-    retryMaxAttempts: number;
     retryBaseDelayMs: number;
     matcher?: RetryRuleMatcher;
     semaphoreManager?: ProviderSemaphoreManager;
     tracker?: RequestTracker;
+    usageWindowTracker?: UsageWindowTracker;
 }
 export declare const openaiProxy: FastifyPluginCallback<OpenaiProxyOptions>;

package/dist/proxy/openai.js

@@ -19,15 +19,15 @@ const OPENAI_ERROR_META = {
 };
 const openaiErrors = createErrorFormatter((kind, message) => ({ error: { message, ...OPENAI_ERROR_META[kind] } }));
 function sendError(reply, e) {
-    return reply.status(e.statusCode).send(e.body);
+    return reply.code(e.statusCode).send(e.body);
 }
 const openaiProxyRaw = (app, opts, done) => {
-    const { db, streamTimeoutMs, retryMaxAttempts, retryBaseDelayMs, matcher, semaphoreManager, tracker } = opts;
+    const { db, streamTimeoutMs, retryBaseDelayMs, matcher, semaphoreManager, tracker, usageWindowTracker } = opts;
     const orchestrator = createOrchestrator(semaphoreManager, tracker);
     app.post(CHAT_COMPLETIONS_PATH, async (request, reply) => {
         if (!orchestrator)
             return sendError(reply, openaiErrors.providerUnavailable());
-        const deps = { db, streamTimeoutMs, retryMaxAttempts, retryBaseDelayMs, matcher, tracker, orchestrator };
+        const deps = { db, streamTimeoutMs, retryBaseDelayMs, matcher, tracker, orchestrator, usageWindowTracker };
         return handleProxyRequest(request, reply, "openai", CHAT_COMPLETIONS_PATH, openaiErrors, deps, {
             beforeSendProxy: (body, isStream) => {
                 if (isStream && !body.stream_options) {
@@ -50,7 +50,7 @@ const openaiProxyRaw = (app, opts, done) => {
             const result = await proxyGetRequest(provider, apiKey, cliHdrs, MODELS_PATH);
             for (const [k, v] of Object.entries(result.headers))
                 reply.header(k, v);
-            return reply.status(result.statusCode).send(result.body);
+            return reply.code(result.statusCode).send(result.body);
         }
         catch (err) {
             request.log.error({ err: err instanceof Error ? err.message : String(err) }, "Failed to reach OpenAI backend for /v1/models");

package/dist/proxy/orchestrator.d.ts

@@ -28,7 +28,6 @@ export interface OrchestratorConfig {
 }
 export interface HandleContext {
     streamTimeoutMs?: number;
-    retryMaxAttempts?: number;
     retryBaseDelayMs?: number;
     failoverThreshold?: number;
     isFailover?: boolean;

package/dist/proxy/orchestrator.js

@@ -1,7 +1,6 @@
 import { ResilienceLayer as ResilienceLayerClass } from "./resilience.js";
 import { SemaphoreScope as SemaphoreScopeClass } from "./scope.js";
 import { TrackerScope as TrackerScopeClass } from "./scope.js";
-const DEFAULT_MAX_RETRIES = 3;
 const DEFAULT_BASE_DELAY_MS = 1000;
 const DEFAULT_FAILOVER_THRESHOLD = 400;
 /**
@@ -66,7 +65,6 @@ export class ProxyOrchestrator {
         if (!ctx?.transportFn)
             throw new Error("HandleContext.transportFn is required");
         const resilienceConfig = {
-            maxRetries: ctx.retryMaxAttempts ?? DEFAULT_MAX_RETRIES,
             baseDelayMs: ctx.retryBaseDelayMs ?? DEFAULT_BASE_DELAY_MS,
             failoverThreshold: ctx.failoverThreshold ?? DEFAULT_FAILOVER_THRESHOLD,
             isFailover: ctx.isFailover ?? false,
@@ -87,7 +85,7 @@ export class ProxyOrchestrator {
                 reply.header(key, value);
             }
         }
-        reply.status(result.statusCode).send(result.body);
+        reply.code(result.statusCode).send(result.body);
     }
     extractTrackStatus(result) {
         const transport = result.result;

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

@@ -1,8 +1,6 @@
 import type { Provider } from "../db/index.js";
 import type { GetTransportResult } from "./transport.js";
 import type { RawHeaders } from "./types.js";
-export { UPSTREAM_SUCCESS } from "./types.js";
-export type { RawHeaders } from "./types.js";
 export interface ProxyErrorResponse {
     statusCode: number;
     body: unknown;
@@ -16,8 +14,6 @@ export interface ProxyErrorFormatter {
     concurrencyQueueFull(providerId: string): ProxyErrorResponse;
     concurrencyTimeout(providerId: string, timeoutMs: number): ProxyErrorResponse;
 }
-export type { ProxyResult, StreamProxyResult } from "./transport.js";
-export type { GetTransportResult as GetProxyResult } from "./transport.js";
 export type ErrorKind = "modelNotFound" | "modelNotAllowed" | "providerUnavailable" | "providerTypeMismatch" | "upstreamConnectionFailed" | "concurrencyQueueFull" | "concurrencyTimeout";
 /**
  * 工厂函数，消除 openai/anthropic 错误格式化的重复代码。

package/dist/proxy/proxy-core.js

@@ -1,6 +1,4 @@
 import { callGet as upstreamGet } from "./transport.js";
-// Re-export for external consumers (openai.ts, anthropic.ts, etc.)
-export { UPSTREAM_SUCCESS } from "./types.js";
 /**
  * 工厂函数，消除 openai/anthropic 错误格式化的重复代码。
  * statusCode 和 message 两个 provider 完全一致，仅 body 格式不同，

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

@@ -7,11 +7,11 @@ import type { ProxyErrorFormatter } from "./proxy-core.js";
 export interface RouteHandlerDeps {
     db: Database.Database;
     streamTimeoutMs: number;
-    retryMaxAttempts: number;
     retryBaseDelayMs: number;
     matcher?: RetryRuleMatcher;
     tracker?: RequestTracker;
     orchestrator: ProxyOrchestrator;
+    usageWindowTracker?: import("./usage-window-tracker.js").UsageWindowTracker;
 }
 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;