llm-simple-router 0.10.6 → 0.10.8

package/dist/admin/providers.js

@@ -138,6 +138,7 @@ export const adminProviderRoutes = (app, options, done) => {
                 name: s.name,
                 api_type: s.api_type,
                 base_url: s.base_url,
+                upstream_path: s.upstream_path,
                 api_key: s.api_key ? decrypt(s.api_key, encryptionKey) : "",
                 models: buildModelInfoList(modelEntries, overrides),
                 is_active: s.is_active,

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/container.d.ts

@@ -10,6 +10,7 @@ export declare const SERVICE_KEYS: {
     readonly pluginRegistry: "pluginRegistry";
     readonly formatRegistry: "formatRegistry";
     readonly logFileWriter: "logFileWriter";
+    readonly logWriteBuffer: "logWriteBuffer";
     readonly proxyAgentFactory: "proxyAgentFactory";
 };
 export type ServiceKey = (typeof SERVICE_KEYS)[keyof typeof SERVICE_KEYS];

package/dist/core/container.js

@@ -10,6 +10,7 @@ export const SERVICE_KEYS = {
     pluginRegistry: "pluginRegistry",
     formatRegistry: "formatRegistry",
     logFileWriter: "logFileWriter",
+    logWriteBuffer: "logWriteBuffer",
     proxyAgentFactory: "proxyAgentFactory",
 };
 /**

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/log-write-buffer.d.ts

@@ -0,0 +1,43 @@
+import type Database from "better-sqlite3";
+import type { RequestLogInsert, LogWriteContext } from "./logs.js";
+import type { MetricsInsert } from "./metrics.js";
+export interface LogWriteBufferOptions {
+    flushIntervalMs?: number;
+    maxBufferSize?: number;
+}
+/**
+ * 批量缓冲 DB 日志写入，减少 SQLite 事务频率。
+ *
+ * 透明使用：logs.ts / metrics.ts 内部判断 buffer 是否存在，
+ * 存在则 push 到缓冲，否则走原始同步路径。
+ *
+ * flush 策略：定时（100ms）+ 阈值（50 条）+ stop() 时同步 flush。
+ */
+export declare class LogWriteBuffer {
+    private readonly db;
+    private readonly rawInsertLog;
+    private readonly rawInsertMetrics;
+    private buffer;
+    private flushTimer;
+    private readonly flushIntervalMs;
+    private readonly maxBufferSize;
+    private stopped;
+    constructor(db: Database.Database, rawInsertLog: (db: Database.Database, data: RequestLogInsert, ctx?: LogWriteContext) => void, rawInsertMetrics: (db: Database.Database, data: MetricsInsert & {
+        id: string;
+    }) => void, options?: LogWriteBufferOptions);
+    /** 推入日志条目到缓冲，达到阈值时立即 flush */
+    pushLog(data: RequestLogInsert, context?: LogWriteContext): void;
+    /**
+     * 推入 metrics 条目到缓冲。
+     * 调用方需预生成 UUID 并在 data.id 中传入。
+     */
+    pushMetrics(data: MetricsInsert & {
+        id: string;
+    }): void;
+    /** 同步批量写入所有缓冲条目（db.transaction 包裹） */
+    flush(): void;
+    /** 停止定时器 + 同步 flush 剩余数据。stop 后的 push 直接走同步写入。 */
+    stop(): void;
+    /** 当前缓冲区中的条目数（测试用） */
+    get pendingCount(): number;
+}

package/dist/db/log-write-buffer.js

@@ -0,0 +1,91 @@
+const DEFAULT_FLUSH_INTERVAL_MS = 100;
+const DEFAULT_MAX_BUFFER_SIZE = 50;
+/**
+ * 批量缓冲 DB 日志写入，减少 SQLite 事务频率。
+ *
+ * 透明使用：logs.ts / metrics.ts 内部判断 buffer 是否存在，
+ * 存在则 push 到缓冲，否则走原始同步路径。
+ *
+ * flush 策略：定时（100ms）+ 阈值（50 条）+ stop() 时同步 flush。
+ */
+export class LogWriteBuffer {
+    db;
+    rawInsertLog;
+    rawInsertMetrics;
+    buffer = [];
+    flushTimer = null;
+    flushIntervalMs;
+    maxBufferSize;
+    stopped = false;
+    constructor(db, rawInsertLog, rawInsertMetrics, options) {
+        this.db = db;
+        this.rawInsertLog = rawInsertLog;
+        this.rawInsertMetrics = rawInsertMetrics;
+        this.flushIntervalMs = options?.flushIntervalMs ?? DEFAULT_FLUSH_INTERVAL_MS;
+        this.maxBufferSize = options?.maxBufferSize ?? DEFAULT_MAX_BUFFER_SIZE;
+        this.flushTimer = setInterval(() => {
+            this.flush();
+        }, this.flushIntervalMs);
+        // 不阻止进程退出
+        if (this.flushTimer && typeof this.flushTimer === "object" && "unref" in this.flushTimer) {
+            this.flushTimer.unref();
+        }
+    }
+    /** 推入日志条目到缓冲，达到阈值时立即 flush */
+    pushLog(data, context) {
+        if (this.stopped) {
+            this.rawInsertLog(this.db, data, context);
+            return;
+        }
+        this.buffer.push({ type: "log", data, context });
+        if (this.buffer.length >= this.maxBufferSize) {
+            this.flush();
+        }
+    }
+    /**
+     * 推入 metrics 条目到缓冲。
+     * 调用方需预生成 UUID 并在 data.id 中传入。
+     */
+    pushMetrics(data) {
+        if (this.stopped) {
+            this.rawInsertMetrics(this.db, data);
+            return;
+        }
+        this.buffer.push({ type: "metrics", data });
+        if (this.buffer.length >= this.maxBufferSize) {
+            this.flush();
+        }
+    }
+    /** 同步批量写入所有缓冲条目（db.transaction 包裹） */
+    flush() {
+        if (this.buffer.length === 0)
+            return;
+        // 取出当前缓冲，重置数组，避免 flush 期间新 push 的条目丢失
+        const entries = this.buffer;
+        this.buffer = [];
+        const transaction = this.db.transaction(() => {
+            for (const entry of entries) {
+                if (entry.type === "log") {
+                    this.rawInsertLog(this.db, entry.data, entry.context);
+                }
+                else {
+                    this.rawInsertMetrics(this.db, entry.data);
+                }
+            }
+        });
+        transaction();
+    }
+    /** 停止定时器 + 同步 flush 剩余数据。stop 后的 push 直接走同步写入。 */
+    stop() {
+        this.stopped = true;
+        if (this.flushTimer) {
+            clearInterval(this.flushTimer);
+            this.flushTimer = null;
+        }
+        this.flush();
+    }
+    /** 当前缓冲区中的条目数（测试用） */
+    get pendingCount() {
+        return this.buffer.length;
+    }
+}

package/dist/db/logs.d.ts

@@ -1,6 +1,7 @@
 import Database from "better-sqlite3";
 import type { LogFileWriter } from "../storage/log-file-writer.js";
 import { type RetryMatcher } from "../proxy/log-detail-policy.js";
+import type { LogWriteBuffer } from "./log-write-buffer.js";
 export interface RequestLog {
     id: string;
     api_type: string;
@@ -54,6 +55,13 @@ export interface LogWriteContext {
     logFileWriter?: LogFileWriter | null;
     responseBody?: string | null;
 }
+/** 初始化日志缓冲（buildApp 时调用） */
+export declare function initLogBuffer(buffer: LogWriteBuffer): void;
+/** 停止日志缓冲，同步 flush 剩余数据（close 时调用） */
+export declare function stopLogBuffer(): void;
+/** 原始 DB INSERT 逻辑（无缓冲） */
+declare function rawInsertRequestLog(db: Database.Database, log: RequestLogInsert, writeContext?: LogWriteContext): void;
+export { rawInsertRequestLog };
 export declare function insertRequestLog(db: Database.Database, log: RequestLogInsert, writeContext?: LogWriteContext): void;
 export declare function getRequestLogs(db: Database.Database, options: {
     page: number;
@@ -74,7 +82,7 @@ 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,
@@ -9,8 +10,32 @@ const LOG_LIST_SELECT = `rl.id, rl.api_type, rl.model, rl.provider_id, rl.status
             rm.input_tokens_estimated, rm.client_type, rm.cache_read_tokens_estimated,
             COALESCE(p.name, rl.provider_id) AS provider_name`;
 const LOG_LIST_JOIN = `LEFT JOIN providers p ON p.id = rl.provider_id LEFT JOIN request_metrics rm ON rm.request_log_id = rl.id`;
+/** 模块级缓冲实例，由 initLogBuffer 设置 */
+let logBuffer = null;
+/** 初始化日志缓冲（buildApp 时调用） */
+export function initLogBuffer(buffer) {
+    logBuffer = buffer;
+}
+/** 停止日志缓冲，同步 flush 剩余数据（close 时调用） */
+export function stopLogBuffer() {
+    if (logBuffer) {
+        logBuffer.stop();
+        logBuffer = null;
+    }
+}
+/** 原始 DB INSERT 逻辑（无缓冲） */
+function rawInsertRequestLog(db, log, writeContext) {
+    // 详情保留判定
+    const preserveDetail = shouldPreserveDetail(log.status_code, writeContext?.responseBody ?? null, writeContext?.matcher ?? null, !!writeContext?.logFileWriter);
+    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);
+}
+// 导出给 LogWriteBuffer 的原始插入函数引用
+export { rawInsertRequestLog };
 export function insertRequestLog(db, log, writeContext) {
-    // 文件写入：始终写入全文
+    // 文件写入：始终同步调用（WriteStream 内部异步，不阻塞事件循环）
     if (writeContext?.logFileWriter) {
         writeContext.logFileWriter.write({
             id: log.id,
@@ -24,12 +49,13 @@ export function insertRequestLog(db, log, writeContext) {
             pipeline_snapshot: log.pipeline_snapshot ?? null,
         });
     }
-    // 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,
-      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);
+    // DB INSERT：缓冲已初始化则走缓冲，否则走同步
+    if (logBuffer) {
+        logBuffer.pushLog(log, writeContext);
+    }
+    else {
+        rawInsertRequestLog(db, log, writeContext);
+    }
 }
 function buildLogWhereClause(options, baseCondition) {
     let where = baseCondition;
@@ -92,11 +118,11 @@ 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 +133,20 @@ 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 +193,28 @@ 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.d.ts

@@ -1,4 +1,5 @@
 import Database from "better-sqlite3";
+import type { LogWriteBuffer } from "./log-write-buffer.js";
 export type MetricsPeriod = "1h" | "5h" | "6h" | "24h" | "7d" | "30d";
 export type MetricsMetric = "ttft" | "tps" | "text_tps" | "thinking_tps" | "tool_use_tps" | "non_thinking_tps" | "total_tps" | "tokens" | "cache_rate" | "request_count" | "input_tokens" | "output_tokens" | "cache_hit_tokens";
 export interface MetricsRow {
@@ -48,6 +49,15 @@ export type MetricsInsert = {
     non_thinking_tps?: number | null;
     total_tps?: number | null;
 };
+/** 设置缓冲实例（由 index.ts buildApp 调用，传入与 logs.ts 共享的 buffer） */
+export declare function setLogBuffer(buffer: LogWriteBuffer): void;
+/** 清除缓冲引用（由 stopLogBuffer 调用） */
+export declare function clearLogBuffer(): void;
+/** 原始 DB INSERT 逻辑（无缓冲） */
+declare function rawInsertMetrics(db: Database.Database, m: MetricsInsert & {
+    id: string;
+}): void;
+export { rawInsertMetrics };
 export declare function insertMetrics(db: Database.Database, m: MetricsInsert): string;
 export interface MetricsSummaryRow {
     provider_id: string;

package/dist/db/metrics.js

@@ -1,13 +1,35 @@
 import { randomUUID } from "crypto";
 import { MS_PER_SECOND } from "../core/constants.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,
+import { getCachedStmt } from "./helpers.js";
+/** 模块级缓冲实例，与 logs.ts 共享同一个 LogWriteBuffer */
+let logBuffer = null;
+/** 设置缓冲实例（由 index.ts buildApp 调用，传入与 logs.ts 共享的 buffer） */
+export function setLogBuffer(buffer) {
+    logBuffer = buffer;
+}
+/** 清除缓冲引用（由 stopLogBuffer 调用） */
+export function clearLogBuffer() {
+    logBuffer = null;
+}
+/** 原始 DB INSERT 逻辑（无缓冲） */
+function rawInsertMetrics(db, m) {
+    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,
        thinking_tps, total_tps, non_thinking_duration_ms, non_thinking_tps)
-     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`).run(id, m.request_log_id, m.provider_id, m.backend_model, m.api_type, m.router_key_id ?? null, m.status_code ?? null, m.input_tokens ?? null, m.output_tokens ?? null, m.cache_creation_tokens ?? null, m.cache_read_tokens ?? null, m.ttft_ms ?? null, m.total_duration_ms ?? null, m.tokens_per_second ?? null, m.stop_reason ?? null, m.is_complete ?? 1, m.input_tokens_estimated ?? 0, m.client_type ?? 'unknown', m.cache_read_tokens_estimated ?? 0, m.thinking_tokens ?? null, m.text_tokens ?? null, m.tool_use_tokens ?? null, m.thinking_duration_ms ?? null, m.thinking_tps ?? null, m.total_tps ?? null, m.non_thinking_duration_ms ?? null, m.non_thinking_tps ?? null);
+     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`).run(m.id, m.request_log_id, m.provider_id, m.backend_model, m.api_type, m.router_key_id ?? null, m.status_code ?? null, m.input_tokens ?? null, m.output_tokens ?? null, m.cache_creation_tokens ?? null, m.cache_read_tokens ?? null, m.ttft_ms ?? null, m.total_duration_ms ?? null, m.tokens_per_second ?? null, m.stop_reason ?? null, m.is_complete ?? 1, m.input_tokens_estimated ?? 0, m.client_type ?? 'unknown', m.cache_read_tokens_estimated ?? 0, m.thinking_tokens ?? null, m.text_tokens ?? null, m.tool_use_tokens ?? null, m.thinking_duration_ms ?? null, m.thinking_tps ?? null, m.total_tps ?? null, m.non_thinking_duration_ms ?? null, m.non_thinking_tps ?? null);
+}
+// 导出给 LogWriteBuffer 的原始插入函数引用
+export { rawInsertMetrics };
+export function insertMetrics(db, m) {
+    const id = randomUUID();
+    if (logBuffer) {
+        logBuffer.pushMetrics({ ...m, id });
+    }
+    else {
+        rawInsertMetrics(db, { ...m, id });
+    }
     return id;
 }
 const PERIOD_OFFSET = {

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/migrations/045_add_metrics_composite_indexes.sql

@@ -0,0 +1,8 @@
+-- request_metrics 核心聚合索引（Dashboard getMetricsSummary + getMetricsTimeseries）
+CREATE INDEX IF NOT EXISTS idx_metrics_agg
+  ON request_metrics(is_complete, created_at DESC, provider_id, backend_model);
+
+-- request_logs 子请求+时间排序复合索引
+-- 注：018 已有 idx_request_logs_original_request_id(original_request_id)，本索引额外覆盖时间排序
+CREATE INDEX IF NOT EXISTS idx_logs_original_time
+  ON request_logs(original_request_id, created_at DESC);

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 = [