llm-simple-router 0.10.6 → 0.10.7

package/dist/config/model-context.d.ts

@@ -16,5 +16,7 @@ export declare const OVERFLOW_THRESHOLD = 1000000;
 export declare function lookupContextWindow(modelName: string): number;
 /** 标准化 patch 名称：连字符 → 下划线 */
 export declare function normalizePatchName(name: string): string;
+/** 清除缓存（仅供测试使用） */
+export declare function clearModelsCache(): void;
 export declare function parseModels(raw: string): ModelEntry[];
 export declare function buildModelInfoList(modelEntries: ModelEntry[], overrides: Map<string, number>): ModelInfo[];

package/dist/config/model-context.js

@@ -104,14 +104,23 @@ const PATCH_ID_MIGRATION = {
     non_ds_tools: "thinking_consistency",
     cache_control: "thinking_consistency",
 };
+// parseModels 缓存，key 为 raw 字符串引用
+const modelsCache = new Map();
+/** 清除缓存（仅供测试使用） */
+export function clearModelsCache() {
+    modelsCache.clear();
+}
 export function parseModels(raw) {
     if (!raw)
         return [];
+    const cached = modelsCache.get(raw);
+    if (cached)
+        return cached;
     try {
         const parsed = JSON.parse(raw);
         if (!Array.isArray(parsed))
             return [];
-        return parsed.map((item) => {
+        const result = parsed.map((item) => {
             if (typeof item === 'string') {
                 return item ? { name: item, patches: [] } : null;
             }
@@ -124,14 +133,16 @@ export function parseModels(raw) {
             const rawPatches = (obj.patches ?? []).map(normalizePatchName);
             const migrated = rawPatches.map(p => PATCH_ID_MIGRATION[p] ?? p);
             const patches = [...new Set(migrated)];
-            const result = {
+            const entry = {
                 name: modelName,
                 patches,
             };
             if (obj.stream_timeout_ms != null)
-                result.stream_timeout_ms = obj.stream_timeout_ms;
-            return result;
+                entry.stream_timeout_ms = obj.stream_timeout_ms;
+            return entry;
         }).filter((e) => e !== null);
+        modelsCache.set(raw, result);
+        return result;
     }
     catch {
         return [];

package/dist/core/monitor/request-tracker.d.ts

@@ -23,6 +23,7 @@ export declare class RequestTracker {
     private providerConfigCache;
     private pushTimer;
     private tickCount;
+    private requestUpdateDirty;
     private streamAccumulators;
     private streamContentPending;
     private streamContentTimer;

package/dist/core/monitor/request-tracker.js

@@ -15,6 +15,7 @@ export class RequestTracker {
     providerConfigCache = new Map();
     pushTimer = null;
     tickCount = 0;
+    requestUpdateDirty = true;
     streamAccumulators = new Map();
     streamContentPending = new Set();
     streamContentTimer = null;
@@ -38,6 +39,7 @@ export class RequestTracker {
     // --- Core methods ---
     start(req) {
         this.activeMap.set(req.id, { ...req });
+        this.requestUpdateDirty = true;
         this.logger?.debug?.({ reqId: req.id, model: req.model, providerId: req.providerId, activeCount: this.activeMap.size }, "Tracker: start");
         this.broadcast("request_start", req);
     }
@@ -133,6 +135,7 @@ export class RequestTracker {
             this.recentCompleted.length = RECENT_COMPLETED_MAX;
         }
         this.logger?.debug?.({ reqId: id, status: result.status, statusCode, latency, activeCount: this.activeMap.size }, "Tracker: complete");
+        this.requestUpdateDirty = true;
         this.broadcast("request_complete", completed);
     }
     /** Update stream metrics for a completed request (e.g., after cache estimation) */
@@ -278,7 +281,10 @@ export class RequestTracker {
             this.tickCount++;
             this.cleanupRecent();
             this.cleanupStaleActive();
-            this.broadcast("request_update", this.getActive());
+            if (this.requestUpdateDirty) {
+                this.broadcast("request_update", this.getActive());
+                this.requestUpdateDirty = false;
+            }
             this.broadcast("concurrency_update", this.getConcurrency());
             this.broadcast("stats_update", this.getStats());
             // Every 10s (every 2nd tick)

package/dist/core/types.d.ts

@@ -26,6 +26,8 @@ export interface ResolveResult {
     concurrency_override?: ConcurrencyOverride;
     /** 活跃规则（schedule 或 base）中的 target 总数，用于 failover 判断 */
     targetCount: number;
+    /** 排除前的完整 target 列表，用于请求级缓存（BP-H2） */
+    allTargets?: Target[];
 }
 export interface MetricsResult {
     input_tokens: number | null;

package/dist/db/helpers.d.ts

@@ -1,4 +1,5 @@
 import Database from "better-sqlite3";
+export declare function getCachedStmt(db: Database.Database, sql: string): Database.Statement;
 /**
  * 通用 UPDATE 构建器。
  * 用白名单过滤安全字段，拼接 SET 子句。

package/dist/db/helpers.js

@@ -1,3 +1,18 @@
+/** WeakMap 按 db 实例缓存 prepared statements，避免重复 prepare() */
+const stmtCache = new WeakMap();
+export function getCachedStmt(db, sql) {
+    let cache = stmtCache.get(db);
+    if (!cache) {
+        cache = new Map();
+        stmtCache.set(db, cache);
+    }
+    let stmt = cache.get(sql);
+    if (!stmt) {
+        stmt = db.prepare(sql);
+        cache.set(sql, stmt);
+    }
+    return stmt;
+}
 /**
  * 通用 UPDATE 构建器。
  * 用白名单过滤安全字段，拼接 SET 子句。

package/dist/db/index.js

@@ -28,6 +28,12 @@ export function initDatabase(dbPath) {
     db.pragma("journal_mode = WAL");
     db.pragma("auto_vacuum = INCREMENTAL");
     db.pragma("foreign_keys = ON");
+    db.pragma("synchronous = NORMAL");
+    db.pragma("cache_size = -16000");
+    db.pragma("busy_timeout = 5000");
+    db.pragma("temp_store = MEMORY");
+    db.pragma("mmap_size = 67108864");
+    db.pragma("journal_size_limit = 67108864");
     db.exec(`
     CREATE TABLE IF NOT EXISTS migrations (
       name TEXT PRIMARY KEY,

package/dist/db/logs.d.ts

@@ -74,7 +74,7 @@ export declare function updateLogStreamContent(db: Database.Database, logId: str
 /** 当 router 返回给客户端的 status code 与上游不同时，记录实际发送的 status */
 export declare function updateLogClientStatus(db: Database.Database, logId: string, clientStatusCode: number): void;
 export declare function deleteLogsBefore(db: Database.Database, beforeDate: string): number;
-/** 估算 request_logs 表占用字节数 */
+/** 采样估算 request_logs 表占用字节数（避免全表 SUM 扫描） */
 export declare function estimateLogTableSize(db: Database.Database): number;
 /** 删除最旧的日志，保留 keepCount 条，返回实际删除条数。分批删除避免长时间锁表 */
 export declare function deleteOldestLogs(db: Database.Database, keepCount: number): number;

package/dist/db/logs.js

@@ -1,4 +1,5 @@
 import { shouldPreserveDetail } from "../proxy/log-detail-policy.js";
+import { getCachedStmt } from "./helpers.js";
 // --- request_logs ---
 const LOG_LIST_SELECT = `rl.id, rl.api_type, rl.model, rl.provider_id, rl.status_code, rl.client_status_code, rl.latency_ms,
             rl.is_stream, rl.error_message, rl.created_at, rl.is_retry, rl.is_failover, rl.original_request_id, rl.original_model,
@@ -26,7 +27,7 @@ export function insertRequestLog(db, log, writeContext) {
     }
     // 详情保留判定
     const preserveDetail = shouldPreserveDetail(log.status_code, writeContext?.responseBody ?? null, writeContext?.matcher ?? null, !!writeContext?.logFileWriter);
-    db.prepare(`INSERT INTO request_logs (id, api_type, model, provider_id, status_code, client_status_code, latency_ms,
+    getCachedStmt(db, `INSERT INTO request_logs (id, api_type, model, provider_id, status_code, client_status_code, latency_ms,
       is_stream, error_message, created_at, client_request, upstream_request, upstream_response,
       is_retry, is_failover, original_request_id, router_key_id, original_model, session_id, pipeline_snapshot)
      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`).run(log.id, log.api_type, log.model, log.provider_id, log.status_code, log.client_status_code ?? null, log.latency_ms, log.is_stream, log.error_message, log.created_at, preserveDetail ? (log.client_request ?? null) : null, preserveDetail ? (log.upstream_request ?? null) : null, preserveDetail ? (log.upstream_response ?? null) : null, log.is_retry ?? 0, log.is_failover ?? 0, log.original_request_id ?? null, log.router_key_id ?? null, log.original_model ?? null, log.session_id ?? null, log.pipeline_snapshot ?? null);
@@ -92,11 +93,11 @@ export function getRequestLogById(db, id) {
 }
 /** 流式请求完成后，将 tracker 中累积的文本内容写入 request_logs */
 export function updateLogStreamContent(db, logId, textContent) {
-    db.prepare("UPDATE request_logs SET stream_text_content = ? WHERE id = ?").run(textContent, logId);
+    getCachedStmt(db, "UPDATE request_logs SET stream_text_content = ? WHERE id = ?").run(textContent, logId);
 }
 /** 当 router 返回给客户端的 status code 与上游不同时，记录实际发送的 status */
 export function updateLogClientStatus(db, logId, clientStatusCode) {
-    db.prepare("UPDATE request_logs SET client_status_code = ? WHERE id = ?").run(clientStatusCode, logId);
+    getCachedStmt(db, "UPDATE request_logs SET client_status_code = ? WHERE id = ?").run(clientStatusCode, logId);
 }
 export function deleteLogsBefore(db, beforeDate) {
     const changes = db.prepare("DELETE FROM request_logs WHERE created_at < ?").run(beforeDate).changes;
@@ -107,17 +108,20 @@ export function deleteLogsBefore(db, beforeDate) {
 }
 /** 每行元数据（数字列+索引）的估算字节数 */
 const ROW_METADATA_BYTES = 500;
-/** 估算 request_logs 表占用字节数 */
+/** 采样估算 request_logs 表占用字节数（避免全表 SUM 扫描） */
 export function estimateLogTableSize(db) {
-    const row = db.prepare(`
-    SELECT COALESCE(SUM(
-      COALESCE(length(client_request), 0) + COALESCE(length(upstream_request), 0) +
-      COALESCE(length(upstream_response), 0) + COALESCE(length(stream_text_content), 0) +
-      COALESCE(length(error_message), 0) + COALESCE(length(pipeline_snapshot), 0) + ?
-    ), 0) as size
-    FROM request_logs
-  `).get(ROW_METADATA_BYTES);
-    return row.size;
+    const countRow = db.prepare("SELECT COUNT(*) as cnt FROM request_logs").get();
+    if (countRow.cnt === 0)
+        return 0;
+    // 采样最近 100 行，计算平均行大小
+    const samples = db.prepare(`
+    SELECT COALESCE(length(client_request), 0) + COALESCE(length(upstream_request), 0) +
+           COALESCE(length(upstream_response), 0) + COALESCE(length(stream_text_content), 0) +
+           COALESCE(length(error_message), 0) + COALESCE(length(pipeline_snapshot), 0) + ? AS row_size
+    FROM request_logs ORDER BY created_at DESC LIMIT 100
+  `).all(ROW_METADATA_BYTES);
+    const avgRowSize = samples.reduce((s, r) => s + r.row_size, 0) / samples.length;
+    return Math.round(avgRowSize * countRow.cnt);
 }
 const DELETE_BATCH_SIZE = 1000;
 /** 删除最旧的日志，保留 keepCount 条，返回实际删除条数。分批删除避免长时间锁表 */
@@ -164,15 +168,28 @@ export function getRequestLogsGrouped(db, options) {
     const total = db.prepare(`SELECT COUNT(*) as count FROM request_logs rl WHERE ${where}`).get(...params).count;
     const offset = (options.page - 1) * options.limit;
     const data = db
-        .prepare(`SELECT ${LOG_LIST_SELECT},
-              (SELECT COUNT(*) FROM request_logs c WHERE c.original_request_id = rl.id) AS child_count
-       FROM request_logs rl
+        .prepare(`WITH page_ids AS (
+         SELECT rl.id FROM request_logs rl
+         ${LOG_LIST_JOIN}
+         WHERE ${where}
+         ORDER BY rl.created_at DESC LIMIT ? OFFSET ?
+       )
+       SELECT ${LOG_LIST_SELECT},
+              COALESCE(child.cnt, 0) AS child_count
+       FROM page_ids pg
+       JOIN request_logs rl ON rl.id = pg.id
        ${LOG_LIST_JOIN}
-       WHERE ${where} ORDER BY rl.created_at DESC LIMIT ? OFFSET ?`)
+       LEFT JOIN (
+         SELECT original_request_id, COUNT(*) AS cnt
+         FROM request_logs
+         WHERE original_request_id IN (SELECT id FROM page_ids)
+         GROUP BY original_request_id
+       ) child ON child.original_request_id = rl.id
+       ORDER BY rl.created_at DESC`)
         .all(...params, options.limit, offset);
     return { data, total };
 }
 /** 后续 pipeline 阶段完成后，回写 snapshot 到已有日志 */
 export function updateLogPipelineSnapshot(db, logId, snapshot) {
-    db.prepare("UPDATE request_logs SET pipeline_snapshot = ? WHERE id = ?").run(snapshot, logId);
+    getCachedStmt(db, "UPDATE request_logs SET pipeline_snapshot = ? WHERE id = ?").run(snapshot, logId);
 }

package/dist/db/metrics.js

@@ -1,8 +1,9 @@
 import { randomUUID } from "crypto";
 import { MS_PER_SECOND } from "../core/constants.js";
+import { getCachedStmt } from "./helpers.js";
 export function insertMetrics(db, m) {
     const id = randomUUID();
-    db.prepare(`INSERT INTO request_metrics (id, request_log_id, provider_id, backend_model, api_type, router_key_id, status_code,
+    getCachedStmt(db, `INSERT INTO request_metrics (id, request_log_id, provider_id, backend_model, api_type, router_key_id, status_code,
        input_tokens, output_tokens, cache_creation_tokens, cache_read_tokens, ttft_ms, total_duration_ms, tokens_per_second, stop_reason, is_complete, input_tokens_estimated,
        client_type, cache_read_tokens_estimated,
        thinking_tokens, text_tokens, tool_use_tokens, thinking_duration_ms,

package/dist/db/migrations/044_add_performance_indexes.sql

@@ -0,0 +1,8 @@
+-- 覆盖 provider_id 过滤 + 时间范围分页
+CREATE INDEX IF NOT EXISTS idx_request_logs_provider_id ON request_logs(provider_id);
+CREATE INDEX IF NOT EXISTS idx_request_logs_created_at_provider ON request_logs(created_at DESC, provider_id);
+CREATE INDEX IF NOT EXISTS idx_request_logs_created_at_router_key ON request_logs(created_at DESC, router_key_id);
+
+-- 覆盖按密钥过滤的聚合查询
+CREATE INDEX IF NOT EXISTS idx_metrics_router_key ON request_metrics(router_key_id);
+CREATE INDEX IF NOT EXISTS idx_metrics_created_at_router_key ON request_metrics(created_at, router_key_id);

package/dist/db/settings.js

@@ -1,9 +1,26 @@
+import { getCachedStmt } from "./helpers.js";
+// TTL 缓存：WeakMap 按 db 实例隔离，确保测试中 :memory: db 互不干扰
+const settingsCache = new WeakMap();
+const CACHE_TTL_MS = 30_000;
 export function getSetting(db, key) {
-    const row = db.prepare("SELECT value FROM settings WHERE key = ?").get(key);
-    return row?.value ?? null;
+    let cache = settingsCache.get(db);
+    if (!cache) {
+        cache = new Map();
+        settingsCache.set(db, cache);
+    }
+    const cached = cache.get(key);
+    if (cached && Date.now() < cached.expiresAt)
+        return cached.value;
+    const row = getCachedStmt(db, "SELECT value FROM settings WHERE key = ?").get(key);
+    const value = row?.value ?? null;
+    cache.set(key, { value, expiresAt: Date.now() + CACHE_TTL_MS });
+    return value;
 }
 export function setSetting(db, key, value) {
-    db.prepare("INSERT OR REPLACE INTO settings (key, value) VALUES (?, ?)").run(key, value);
+    getCachedStmt(db, "INSERT OR REPLACE INTO settings (key, value) VALUES (?, ?)").run(key, value);
+    const cache = settingsCache.get(db);
+    if (cache)
+        cache.delete(key);
 }
 export function isInitialized(db) {
     return getSetting(db, "initialized") === "true";
@@ -40,7 +57,7 @@ export function setConfigSyncSource(db, source) {
     setSetting(db, "config_sync_source", source);
 }
 export function getDetailLogEnabled(db) {
-    const row = db.prepare("SELECT value FROM settings WHERE key = ?").get("detail_log_enabled");
+    const row = getCachedStmt(db, "SELECT value FROM settings WHERE key = ?").get("detail_log_enabled");
     return row ? row.value !== "0" : true;
 }
 export function getTokenEstimationEnabled(db) {
@@ -52,7 +69,7 @@ export function setTokenEstimationEnabled(db, enabled) {
 }
 const DEFAULT_LOG_FILE_RETENTION_DAYS = 3;
 export function getLogFileRetentionDays(db) {
-    const row = db.prepare("SELECT value FROM settings WHERE key = ?").get("log_file_retention_days");
+    const row = getCachedStmt(db, "SELECT value FROM settings WHERE key = ?").get("log_file_retention_days");
     return row ? parseInt(row.value, 10) : DEFAULT_LOG_FILE_RETENTION_DAYS;
 }
 const DEFAULT_CLIENT_SESSION_HEADERS = [

package/dist/index.js

@@ -330,6 +330,8 @@ export async function buildApp(options) {
         proxyAgentFactory.invalidateAll();
         const sessionTracker = container.resolve(SERVICE_KEYS.sessionTracker);
         sessionTracker.stop();
+        // Flush LogFileWriter 的 WriteStream 缓冲数据到磁盘
+        await logFileWriter?.stop();
         // 等待活跃代理请求自然完成，超时后强制关闭所有连接。
         // 先调用 app.close() 停止接受新连接并等待现有连接结束，
         // 如果 2 秒内未完成则调用 closeAllConnections() 强制断开，防止 SSE 长连接导致无限等待。

package/dist/metrics/metrics-extractor.d.ts

@@ -13,12 +13,17 @@ export declare class MetricsExtractor {
     private stopReason;
     private firstContentReceived;
     private complete;
-    private thinkingContentBuffer;
+    /** Buffer 容量上限，超过后停止 push 避免内存膨胀 */
+    private static readonly MAX_BUFFER_SIZE;
+    private thinkingChunks;
+    private thinkingTotalLength;
     private thinkingStreamStartTime;
     private thinkingStreamEndTime;
-    private textContentBuffer;
+    private textChunks;
+    private textTotalLength;
     private textStreamStartTime;
-    private toolUseContentBuffer;
+    private toolUseChunks;
+    private toolUseTotalLength;
     private toolUseStreamStartTime;
     constructor(apiType: "openai" | "openai-responses" | "anthropic", requestStartTime: number);
     processEvent(event: SSEEvent): void;

package/dist/metrics/metrics-extractor.js

@@ -1,6 +1,6 @@
 // TODO: 当文件超过 400 行时拆分为 metrics-streaming.ts（流式事件处理 + TPS 计算）和 metrics-extractor.ts（非流式 + 类型）
 import { MS_PER_SECOND } from "../core/constants.js";
-import { encode } from "gpt-tokenizer";
+import { countTokens } from "../utils/token-counter.js";
 export class MetricsExtractor {
     apiType;
     requestStartTime;
@@ -14,13 +14,18 @@ export class MetricsExtractor {
     stopReason = null;
     firstContentReceived = false;
     complete = false;
-    // --- Phase content buffers + timing ---
-    thinkingContentBuffer = "";
+    // --- Phase content buffers (array-based) + timing ---
+    /** Buffer 容量上限，超过后停止 push 避免内存膨胀 */
+    static MAX_BUFFER_SIZE = 500_000; // eslint-disable-line no-magic-numbers
+    thinkingChunks = [];
+    thinkingTotalLength = 0;
     thinkingStreamStartTime = null;
     thinkingStreamEndTime = null;
-    textContentBuffer = "";
+    textChunks = [];
+    textTotalLength = 0;
     textStreamStartTime = null;
-    toolUseContentBuffer = "";
+    toolUseChunks = [];
+    toolUseTotalLength = 0;
     toolUseStreamStartTime = null;
     constructor(apiType, requestStartTime) {
         this.apiType = apiType;
@@ -49,7 +54,8 @@ export class MetricsExtractor {
         let thinkingDurationMs = null;
         let textTokens = null;
         let toolUseTokens = null;
-        const hasThinking = this.thinkingContentBuffer.length > 0;
+        const thinkingContent = this.thinkingChunks.join("");
+        const hasThinking = thinkingContent.length > 0;
         if (this.streamEndTime !== null &&
             this.outputTokens !== null) {
             // total_duration: T6 - T0 (proxy end-to-end, not just stream window)
@@ -58,7 +64,7 @@ export class MetricsExtractor {
                 totalTps = this.outputTokens / (totalDurationMs / MS_PER_SECOND);
             }
             if (hasThinking) {
-                thinkingTokens = encode(this.thinkingContentBuffer).length;
+                thinkingTokens = countTokens(thinkingContent);
                 // thinking_duration: T3 - T0 (includes network RTT + generation)
                 if (this.thinkingStreamEndTime !== null) {
                     thinkingDurationMs = this.thinkingStreamEndTime - this.requestStartTime;
@@ -81,11 +87,13 @@ export class MetricsExtractor {
                 }
             }
             // content token counts (for analysis only)
-            if (this.textContentBuffer.length > 0) {
-                textTokens = encode(this.textContentBuffer).length;
+            const textContent = this.textChunks.join("");
+            if (textContent.length > 0) {
+                textTokens = countTokens(textContent);
             }
-            if (this.toolUseContentBuffer.length > 0) {
-                toolUseTokens = encode(this.toolUseContentBuffer).length;
+            const toolUseContent = this.toolUseChunks.join("");
+            if (toolUseContent.length > 0) {
+                toolUseTokens = countTokens(toolUseContent);
             }
         }
         return {
@@ -140,7 +148,8 @@ export class MetricsExtractor {
                 this.firstContentReceived = true;
                 this.ttftMs = Date.now() - this.requestStartTime;
             }
-            this.textContentBuffer += delta;
+            this.textChunks.push(delta);
+            this.textTotalLength += delta.length;
         }
         // Track completion
         if (type === "response.completed" || type === "response.incomplete") {
@@ -193,8 +202,9 @@ export class MetricsExtractor {
                     this.thinkingStreamStartTime = Date.now();
                 }
                 const thinking = delta.thinking ?? "";
-                if (thinking) {
-                    this.thinkingContentBuffer += thinking;
+                if (thinking && this.thinkingTotalLength < MetricsExtractor.MAX_BUFFER_SIZE) {
+                    this.thinkingChunks.push(thinking);
+                    this.thinkingTotalLength += thinking.length;
                     this.thinkingStreamEndTime = Date.now();
                 }
             }
@@ -202,16 +212,18 @@ export class MetricsExtractor {
                 if (this.textStreamStartTime === null) {
                     this.textStreamStartTime = Date.now();
                 }
-                if (delta.text) {
-                    this.textContentBuffer += delta.text;
+                if (delta.text && this.textTotalLength < MetricsExtractor.MAX_BUFFER_SIZE) {
+                    this.textChunks.push(delta.text);
+                    this.textTotalLength += delta.text.length;
                 }
             }
             else if (delta?.type === "input_json_delta") {
                 if (this.toolUseStreamStartTime === null) {
                     this.toolUseStreamStartTime = Date.now();
                 }
-                if (delta.partial_json) {
-                    this.toolUseContentBuffer += delta.partial_json;
+                if (delta.partial_json && this.toolUseTotalLength < MetricsExtractor.MAX_BUFFER_SIZE) {
+                    this.toolUseChunks.push(delta.partial_json);
+                    this.toolUseTotalLength += delta.partial_json.length;
                 }
             }
         }
@@ -252,8 +264,9 @@ export class MetricsExtractor {
                 this.ttftMs = Date.now() - this.requestStartTime;
                 this.textStreamStartTime = Date.now();
             }
-            if (delta?.content) {
-                this.textContentBuffer += delta.content;
+            if (delta?.content && this.textTotalLength < MetricsExtractor.MAX_BUFFER_SIZE) {
+                this.textChunks.push(delta.content);
+                this.textTotalLength += delta.content.length;
             }
             if (choice.finish_reason) {
                 this.stopReason = choice.finish_reason;

package/dist/metrics/sse-parser.js

@@ -9,8 +9,11 @@ export class SSEParser {
         if (this.isDone)
             return [];
         this.buffer += chunk;
-        // SSE 规范允许 \r\n 行尾，统一为 \n
-        this.buffer = this.buffer.replace(/\r\n/g, "\n");
+        // SSE 规范允许 \\r\\n 行尾，统一为 \\n
+        // 绝大多数 chunk 不含 \\r，跳过全局 replace
+        if (this.buffer.includes('\r')) {
+            this.buffer = this.buffer.replace(/\r\n/g, "\n");
+        }
         return this.drainEvents();
     }
     flush() {

package/dist/middleware/auth.d.ts

@@ -5,7 +5,7 @@ declare module "fastify" {
         routerKey?: {
             id: string;
             name: string;
-            allowed_models: string | null;
+            allowed_models: string[] | null;
         };
     }
 }

package/dist/middleware/auth.js

@@ -1,7 +1,6 @@
-import { createHash, randomUUID } from "crypto";
+import { createHash } from "crypto";
 import fp from "fastify-plugin";
 import { isInitialized } from "../db/settings.js";
-import { insertRequestLog } from "../db/logs.js";
 import { getProxyApiType, HTTP_SERVICE_UNAVAILABLE } from "../core/constants.js";
 const SKIP_PATHS = ["/health", "/admin"];
 const HTTP_UNAUTHORIZED = 401;
@@ -19,20 +18,6 @@ function unauthorizedReply(reply) {
         },
     });
 }
-function logRejectedAuth(db, apiType, statusCode, errorMessage, request) {
-    insertRequestLog(db, {
-        id: randomUUID(),
-        api_type: apiType,
-        model: null,
-        provider_id: null,
-        status_code: statusCode,
-        latency_ms: 0,
-        is_stream: 0,
-        error_message: errorMessage,
-        created_at: new Date().toISOString(),
-        client_request: JSON.stringify({ method: request.method, ip: request.ip, headers: request.headers }),
-    });
-}
 const authMiddlewareRaw = (app, options, done) => {
     const stmt = options.db.prepare("SELECT id, name, allowed_models FROM router_keys WHERE key_hash = ? AND is_active = 1");
     app.addHook("onRequest", async (request, reply) => {
@@ -47,7 +32,7 @@ const authMiddlewareRaw = (app, options, done) => {
         // 未初始化时代理层不可用
         if (!isInitialized(options.db)) {
             if (proxyApiType) {
-                logRejectedAuth(options.db, proxyApiType, HTTP_SERVICE_UNAVAILABLE, "Service not initialized", request);
+                request.log.info({ method: request.method, url: request.url, ip: request.ip }, `Rejected: service not initialized [${proxyApiType}]`);
             }
             reply.code(HTTP_SERVICE_UNAVAILABLE).send({ error: { message: "Service not initialized" } });
             return reply;
@@ -66,7 +51,7 @@ const authMiddlewareRaw = (app, options, done) => {
         }
         if (!token) {
             if (proxyApiType) {
-                logRejectedAuth(options.db, proxyApiType, HTTP_UNAUTHORIZED, "Invalid API key", request);
+                request.log.info({ method: request.method, url: request.url, ip: request.ip }, `Rejected: no API key [${proxyApiType}]`);
             }
             unauthorizedReply(reply);
             return reply;
@@ -75,12 +60,19 @@ const authMiddlewareRaw = (app, options, done) => {
         const row = stmt.get(hash);
         if (!row) {
             if (proxyApiType) {
-                logRejectedAuth(options.db, proxyApiType, HTTP_UNAUTHORIZED, "Invalid API key", request);
+                request.log.info({ method: request.method, url: request.url, ip: request.ip }, `Rejected: invalid API key [${proxyApiType}]`);
             }
             unauthorizedReply(reply);
             return reply;
         }
-        request.routerKey = { id: row.id, name: row.name, allowed_models: row.allowed_models };
+        let parsedAllowedModels = null;
+        if (row.allowed_models) {
+            try {
+                parsedAllowedModels = JSON.parse(row.allowed_models);
+            }
+            catch { /* JSON 解析失败时 allowed_models 保持为 null，允许所有模型 */ } // eslint-disable-line taste/no-silent-catch
+        }
+        request.routerKey = { id: row.id, name: row.name, allowed_models: parsedAllowedModels };
     });
     done();
 };

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

@@ -11,6 +11,7 @@
 import { randomUUID } from "crypto";
 import fp from "fastify-plugin";
 import { insertRequestLog, getAllProviders } from "../../db/index.js";
+import { parseModels } from "../../config/model-context.js";
 import { createErrorFormatter } from "../proxy-core.js";
 import { createOrchestrator } from "../orchestration/orchestrator.js";
 import { HTTP_OK, HTTP_BAD_GATEWAY, HTTP_CLIENT_CLOSED, MS_PER_SECOND } from "../../core/constants.js";
@@ -24,7 +25,6 @@ import { HTTP_UNPROCESSABLE_ENTITY } from "../../core/constants.js";
 import { PipelineAbort } from "../pipeline/types.js";
 import { applyToolRoundLimit } from "../patch/tool-round-limiter.js";
 import { extractLastToolUse } from "./proxy-handler-utils.js";
-import { parseModels } from "../../config/model-context.js";
 // ---------- Models handler (shared across openai/anthropic) ----------
 const ANTHROPIC_DEFAULT_PAGE_SIZE = 20;
 const ANTHROPIC_MAX_PAGE_SIZE = 1000;
@@ -44,7 +44,11 @@ function handleModelsRequest(db) {
                 continue;
             }
         }
-        const sortedIds = [...modelMeta.keys()].sort();
+        // 如果请求的 key 配置了 allowed_models 白名单，则过滤
+        const allowedModels = request.routerKey?.allowed_models;
+        const sortedIds = allowedModels
+            ? [...modelMeta.keys()].filter(id => allowedModels.includes(id)).sort()
+            : [...modelMeta.keys()].sort();
         const isAnthropicFormat = !!request.headers["anthropic-version"];
         if (isAnthropicFormat) {
             const query = request.query;