mslxdff 0.1.134 → 0.1.135

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mslxdff",
-  "version": "0.1.134",
+  "version": "0.1.135",
   "description": "测试项目，请勿使用。",
   "type": "module",
   "bin": {

package/src/cli/commands/stats.js

@@ -0,0 +1,88 @@
+// `-stats` 模型用量报表：近 N 小时每模型 token 消耗 + 首字/总耗时/速度。
+// 数据来自 src/usage/（逐请求 JSONL），不是 state.json 的终生 EMA —— 见 .scratch/stats-report/SPEC.md
+import { usageReport } from "../../usage/report.js";
+import { usageEnabled, usageKeepDays } from "../../usage/record.js";
+
+const FLAGS = ["-stats", "--stats"];
+
+export function isStatsFlag(args) {
+  return FLAGS.some((f) => args.includes(f));
+}
+
+export function parseStatsArgs(args) {
+  const hoursIdx = args.findIndex((a) => a === "--hours" || a === "-hours");
+  const rawHours = hoursIdx >= 0 ? Number(args[hoursIdx + 1]) : NaN;
+  const hours = Number.isFinite(rawHours) && rawHours > 0 ? Math.min(168, Math.floor(rawHours)) : 24;
+  const modelIdx = args.findIndex((a) => a === "--model" || a === "-model");
+  const rawModel = modelIdx >= 0 && args[modelIdx + 1] && !String(args[modelIdx + 1]).startsWith("-") ? args[modelIdx + 1] : null;
+  return { hours, model: rawModel, json: args.includes("--json") };
+}
+
+function fmtTok(n) {
+  const v = Number(n) || 0;
+  if (v < 1000) return String(v);
+  if (v < 1_000_000) return `${(v / 1000).toFixed(1)}k`;
+  return `${(v / 1_000_000).toFixed(2)}M`;
+}
+
+function fmtMs(v) {
+  if (v == null || !Number.isFinite(v)) return "—";
+  return v < 1000 ? `${v}ms` : `${(v / 1000).toFixed(1)}s`;
+}
+
+function fmtTps(v) {
+  return v == null || !Number.isFinite(v) ? "—" : `${v} tok/s`;
+}
+
+// 中文字符按 2 列宽算，否则表格错位
+function width(s) {
+  let w = 0;
+  for (const ch of String(s)) w += /[\u1100-\u115F\u2E80-\uA4CF\uAC00-\uD7A3\uF900-\uFAFF\uFE30-\uFE4F\uFF00-\uFF60\uFFE0-\uFFE6]/.test(ch) ? 2 : 1;
+  return w;
+}
+
+function padW(s, w) {
+  const t = String(s);
+  return t + " ".repeat(Math.max(0, w - width(t)));
+}
+
+function rowText(id, r) {
+  return `  ${padW(id, 30)}  ${padW(r.requests, 6)}  ${padW(fmtTok(r.promptTokens), 8)}  ${padW(fmtTok(r.completionTokens), 10)}  ${padW(fmtTok(r.totalTokens), 8)}  ${padW(fmtMs(r.avgTtfbMs), 7)}  ${padW(fmtMs(r.avgTotalMs), 8)}  ${fmtTps(r.avgTps)}`;
+}
+
+export function renderStats(report, { hours = 24, model = null } = {}) {
+  const lines = [];
+  const { models, totals } = report;
+  if (!models.length) {
+    lines.push(`暂无用量记录（近 ${hours}h${model ? ` · 模型 ${model}` : ""}）— 经 8989 网关发一次请求后出现`);
+    lines.push("提示：mslxdff -status 看当前体检 · mslxdff -log 20 看最近事件");
+    lines.push("说明：-chat 直连 mimo/big-pickle 不经网关，不计入本表");
+    return lines.join("\n");
+  }
+  lines.push(`模型用量（近 ${hours}h · 成功请求 ${totals.requests} 次 · ${models.length} 个模型）`);
+  lines.push(`  ${padW("模型", 30)}  ${padW("请求", 6)}  ${padW("prompt", 8)}  ${padW("输出", 10)}  ${padW("合计", 8)}  ${padW("首字", 7)}  ${padW("总耗时", 8)}  速度`);
+  for (const m of models) lines.push(rowText(m.id, m));
+  lines.push(`  ${"-".repeat(76)}`);
+  lines.push(rowText("合计", totals));
+  if (totals.reasoningTokens > 0) lines.push(`  其中思考 tokens：${fmtTok(totals.reasoningTokens)}`);
+  lines.push("");
+  lines.push("说明：速度 = 输出 tokens ÷ 生成耗时（总耗时−首字），按窗口加权；只统计成功请求。");
+  lines.push(`      -chat 直连 mimo/big-pickle 不经 8989 网关，不计入。数据保留 ${usageKeepDays()} 天，mslxdff -stats --hours 1|--json|--model <id> 可调。`);
+  return lines.join("\n");
+}
+
+export async function handleStats(args) {
+  if (!isStatsFlag(args)) return false;
+  const { hours, model, json } = parseStatsArgs(args);
+  if (!usageEnabled()) {
+    console.log("用量记录已关闭（MSLXDFF_USAGE_LOG=0）—— 去掉该 env 后重启 daemon 即可恢复采集。");
+    return true;
+  }
+  const report = await usageReport({ hours, model, now: Date.now() });
+  if (json) {
+    console.log(JSON.stringify(report, null, 2));
+    return true;
+  }
+  console.log(renderStats(report, { hours, model }));
+  return true;
+}

package/src/cli/help.js

@@ -5,6 +5,8 @@ Usage:
   mslxdff                          start as a background daemon and exit (status + help if one is already running)
   mslxdff -d                       start as a background daemon
   mslxdff -status                  show current status (daemon/health/port/config, upstream providers, models + metrics/体检表, autostart/plugins, groups/peers, recent calls with ttfb/tps, last error)
+  mslxdff -stats [--hours N] [--json] [--model <id>]  per-model token usage + speed over the last N hours (default 24; success requests only, from the local usage JSONL)
+
   mslxdff -log [N]                 show last N events (default 10, e.g. -log 100)
   mslxdff -models                interactive picker: ↑/↓ select a model, Enter sets it as the default (non-TTY: plain list)
   mslxdff -model list            list the free models this proxy serves (cached)

package/src/cli/index.js

@@ -23,6 +23,8 @@ export async function run(args = process.argv.slice(2)) {
   if (await handlePlugins(args)) return;
   if (await handleChat(args)) return;
   if (await handleStatus(args, VERSION)) return;
+  const { handleStats } = await import("./commands/stats.js");
+  if (await handleStats(args)) return;
 
   const { handleModel } = await import("./commands/model.js");
   if (await handleModel(args)) return;

package/src/cli/status.js

@@ -243,25 +243,20 @@ export async function printStatus(VERSION) {
   }
 
   const { recentCalls, lastError } = await import("../logs.js");
-  console.log("\nrecent calls: (gateway 持久化，最近5条，含首字/tok/s)");
+  console.log("\nrecent calls: (gateway 持久化，最近5条；token/速度看 mslxdff -stats)");
   const calls = recentCalls(5);
   if (calls.length) {
-    let sumDur = 0, sumTps = 0, tpsN = 0;
+    let sumDur = 0;
     for (const c of calls) {
       if (Number.isFinite(c.durationMs)) sumDur += c.durationMs;
       else if (Number.isFinite(c.totalMs)) sumDur += c.totalMs;
-      if (Number.isFinite(c.tps)) { sumTps += c.tps; tpsN++; } else if (Number.isFinite(c.charsPerSec)) { sumTps += c.charsPerSec; tpsN++; }
     }
     const avgDur = calls.length ? Math.round(sumDur / calls.length) : null;
-    const avgTps = tpsN ? Math.round(sumTps / tpsN) : null;
-    console.log(`  avg ${avgDur ? avgDur + "ms" : "—"}${avgTps ? ` · ${avgTps} tok/s` : ""}  —  mslxdff -log 20 查看详情`);
+    console.log(`  avg ${avgDur ? avgDur + "ms" : "—"}  —  mslxdff -log 20 查看详情`);
     for (const c of calls) {
       const dur = c.totalMs ?? c.durationMs;
-      const ttfb = c.ttfbMs != null ? ` 首字${c.ttfbMs}ms` : "";
-      const tps = c.tps != null ? ` ${c.tps}tok/s` : (c.charsPerSec ? ` ${c.charsPerSec}ch/s` : "");
-      const tok = c.usage?.completion_tokens != null ? ` tok${c.usage.completion_tokens}` : (c.chars ? ` ch${c.chars}` : "");
       const tm = fmtTs(c.ts);
-      console.log(`  ${tm}  ${(c.model || "-").padEnd(28)}  ${String(c.status || "-").padEnd(4)}  ${dur ? dur + "ms" : ""}${ttfb}${tps}${tok}${c.auto ? "  auto" : ""}`);
+      console.log(`  ${tm}  ${(c.model || "-").padEnd(28)}  ${String(c.status || "-").padEnd(4)}  ${dur ? dur + "ms" : ""}${c.stream ? "  stream" : ""}${c.auto ? "  auto" : ""}`);
     }
   } else {
     console.log("  (none yet — 发一次请求后出现，mslxdff -chat hi)");

package/src/routes/chat/relay-pipeline.js

@@ -2,6 +2,7 @@ import { runHook } from "../../plugins.js";
 import { recordModelStats } from "../../state.js";
 import { normalizeFullId } from "../../providers/model-id.js";
 import { computeMetrics } from "../../metrics.js";
+import { recordChatUsage } from "../../usage/record.js"; // 窗口报表唯一写入点（canonical 名单记防双计）— 见 .agents/notes/implemented/feature/2026-09-19-usage-report-jsonl.md
 
 // 唯一/最后候选没有 failover 去向：首块闸门退化为纯"防连接泄漏"，放宽避免误杀慢模型
 // （参考 opencode：zen 通道不设超时；openai responses 硬编码 300s headerTimeout）
@@ -134,6 +135,15 @@ export function createRelayPipeline({
       }
       _evt("slow-model", { reqId, model: actual, elapsedMs: out.totalMs ?? (Date.now() - curStartedAt), threshold: C.STALL_TIMEOUT_MS, interrupted: true, detail: out.detail ?? null });
       try { _logCall(actual, 200); } catch {}
+      // interrupted 的 200 也是真实消耗（最贵的长生成）——照常落 usage 标 interrupted:1，口径与 5c 一致 — 见 .agents/notes/implemented/feature/2026-09-19-usage-report-jsonl.md
+      if (out.status === 200) {
+        try {
+          const u = out.detail?.usage || null;
+          const t1 = Number.isFinite(out.totalMs) && out.totalMs > 0 ? out.totalMs : (Date.now() - curStartedAt);
+          const t0 = Number.isFinite(out.ttfMs) && out.ttfMs > 0 ? out.ttfMs : null;
+          recordChatUsage({ model: normalizeFullId(actual), via, usage: u, interrupted: 1, ttfbMs: t0, totalMs: t1, tps: null }).catch(() => {});
+        } catch {}
+      }
       _evt("result", { reqId, model: actual, status: out.status, via, timing: upRes?._t ?? null, ttfMs: out.ttfMs, totalMs: out.totalMs, interrupted: true, detail: out.detail ?? null, fallback, requested, actual });
       _evt("client-response", { requested, actual, via, fallback, status: out.status, reqId, interrupted: true });
       if (plugins?.length) runHook(plugins, "request:completed", { reqId, requested, useAuto, hops, stream: Boolean(body?.stream), durationMs: Date.now() - curStartedAt, via, status: out.status, actual, interrupted: true, fallback }).catch(() => {});
@@ -183,6 +193,9 @@ export function createRelayPipeline({
         const fullId = normalizeFullId(actual);
         recordModelStats(fullId, { ttfbMs: ttfb, totalMs: total, tps, completionTokens: compTok });
         if (fullId !== actual) recordModelStats(actual, { ttfbMs: ttfb, totalMs: total, tps, completionTokens: compTok });
+        // 窗口报表：逐请求落 usage（行形状由 usage/record.js 拥有，含 prompt/total ——
+        // state 的 modelStats 只存 completion 的 EMA）。只按 canonical 名记一次，避免双计。
+        recordChatUsage({ model: fullId, via, usage, ttfbMs: ttfb, totalMs: total, tps: m.tps }).catch(() => {});
       } catch {}
     }
 

package/src/usage/record.js

@@ -0,0 +1,99 @@
+// 逐请求 usage 落盘：按日切片的 JSONL，供 -stats 做时间窗口聚合。
+// Note: 与 logs.js 的 calls/errors/events 不同，这里按日分片 + 保留期删旧文件，
+// 不走 1MB 环形截断 —— 环形会把 24h 窗口的数据裁到末 100 行。见 .scratch/stats-report/SPEC.md
+import { appendFile, mkdir, readdir, unlink } from "node:fs/promises";
+import { join } from "node:path";
+import { logDir } from "../logs.js";
+
+const DEFAULT_KEEP_DAYS = 2;
+const DAY_MS = 86_400_000;
+
+export function usageEnabled() {
+  return process.env.MSLXDFF_USAGE_LOG !== "0";
+}
+
+export function usageKeepDays() {
+  const n = Number(process.env.MSLXDFF_USAGE_KEEP_DAYS);
+  return Number.isInteger(n) && n > 0 ? n : DEFAULT_KEEP_DAYS;
+}
+
+export function usageDir({ dir } = {}) {
+  return join(dir || logDir(), "usage");
+}
+
+export function ymd(date) {
+  const y = date.getFullYear();
+  const m = String(date.getMonth() + 1).padStart(2, "0");
+  const d = String(date.getDate()).padStart(2, "0");
+  return `${y}-${m}-${d}`;
+}
+
+export function usageFileFor(date, { dir } = {}) {
+  return join(usageDir({ dir }), `${ymd(date)}.jsonl`);
+}
+
+// 删掉早于保留期的日文件。文件名是 YYYY-MM-DD，可直接字典序比较。
+export async function pruneUsage({ dir, keepDays, now = new Date() } = {}) {
+  const keep = Number.isInteger(keepDays) && keepDays > 0 ? keepDays : usageKeepDays();
+  const cutoff = ymd(new Date(now.getTime() - keep * DAY_MS));
+  const removed = [];
+  let files = [];
+  try {
+    files = await readdir(usageDir({ dir }));
+  } catch {
+    return removed; // 目录还不存在 = 无事可做
+  }
+  for (const f of files) {
+    if (!f.endsWith(".jsonl")) continue;
+    const day = f.slice(0, -".jsonl".length);
+    if (!/^\d{4}-\d{2}-\d{2}$/.test(day) || day >= cutoff) continue;
+    try {
+      await unlink(join(usageDir({ dir }), f));
+      removed.push(f);
+    } catch {}
+  }
+  return removed;
+}
+
+let lastPrunedDay = null;
+
+// 追加一行 usage。异步写，调用方可 fire-and-forget（不阻塞流式响应）。
+// 清理是惰性的且每天最多触发一次，避免每个请求都去 readdir。
+export async function recordUsage(entry, { dir, now = new Date() } = {}) {
+  if (!usageEnabled() || !entry || typeof entry !== "object") return null;
+  const row = { ts: now.getTime(), ...entry };
+  try {
+    await mkdir(usageDir({ dir }), { recursive: true });
+    await appendFile(usageFileFor(now, { dir }), JSON.stringify(row) + "\n");
+  } catch {
+    return null;
+  }
+  const today = ymd(now);
+  if (lastPrunedDay !== today) {
+    lastPrunedDay = today;
+    pruneUsage({ dir, now }).catch(() => {});
+  }
+  return row;
+}
+
+// 从 relay 结果组装一行 usage —— 行形状由本模块拥有，调用方只交原始字段。
+// usage 即 metrics.js 的 extractUsageFromJson 输出（prompt/completion/total/reasoning）。
+export function recordChatUsage({ model, via, usage, ttfbMs, totalMs, tps, interrupted } = {}) {
+  return recordUsage({
+    model,
+    via,
+    interrupted,
+    prompt_tokens: usage?.prompt_tokens ?? 0,
+    completion_tokens: usage?.completion_tokens ?? 0,
+    total_tokens: usage?.total_tokens ?? 0,
+    reasoning_tokens: usage?.reasoning_tokens ?? 0,
+    ttfbMs: Number.isFinite(ttfbMs) ? ttfbMs : null,
+    totalMs: Number.isFinite(totalMs) ? totalMs : null,
+    tps: Number.isFinite(tps) ? tps : null,
+  });
+}
+
+// 测试用：重置惰性清理标记，避免跨用例串味（置于文件末，业务函数在其上）
+export function _resetPruneMarker() {
+  lastPrunedDay = null;
+}

package/src/usage/report.js

@@ -0,0 +1,152 @@
+// 窗口用量聚合：JSONL 行数组 + 时间窗口 → 每模型 token/速度报表。
+// 纯函数（aggregateUsage）与薄 IO（readUsageRows / usageReport）分开，前者可离线单测。
+// Note: 速度用加权口径 Σcompletion / ΣcompletionMs，不用算术平均 —— 短回答会把算术均值拉飞。
+import { readFile, readdir } from "node:fs/promises";
+import { join } from "node:path";
+import { usageDir } from "./record.js";
+
+const HOUR_MS = 3_600_000;
+
+function tok(v) {
+  const n = Number(v);
+  return Number.isFinite(n) && n > 0 ? n : 0;
+}
+
+function ms(v) {
+  const n = Number(v);
+  return Number.isFinite(n) && n >= 0 ? n : null;
+}
+
+function blank(id) {
+  return {
+    id,
+    requests: 0,
+    promptTokens: 0,
+    completionTokens: 0,
+    totalTokens: 0,
+    reasoningTokens: 0,
+    ttfbSumMs: 0,
+    ttfbN: 0,
+    totalSumMs: 0,
+    totalN: 0,
+    completionMsSum: 0,
+    tpsTokSum: 0,
+  };
+}
+
+// 把累计量收敛成对外字段：平均首字/平均总耗时/加权速度。
+function finalize(a) {
+  return {
+    id: a.id,
+    requests: a.requests,
+    promptTokens: a.promptTokens,
+    completionTokens: a.completionTokens,
+    totalTokens: a.totalTokens,
+    reasoningTokens: a.reasoningTokens,
+    avgTtfbMs: a.ttfbN ? Math.round(a.ttfbSumMs / a.ttfbN) : null,
+    avgTotalMs: a.totalN ? Math.round(a.totalSumMs / a.totalN) : null,
+    avgTps: a.completionMsSum > 0 ? Number((a.tpsTokSum / (a.completionMsSum / 1000)).toFixed(1)) : null,
+  };
+}
+
+function fold(acc, r) {
+  acc.requests++;
+  const prompt = tok(r.prompt_tokens);
+  const comp = tok(r.completion_tokens);
+  acc.promptTokens += prompt;
+  acc.completionTokens += comp;
+  const total = tok(r.total_tokens);
+  acc.totalTokens += total || prompt + comp;
+  acc.reasoningTokens += tok(r.reasoning_tokens);
+
+  const ttfb = ms(r.ttfbMs ?? r.ttfb_ms);
+  if (ttfb != null) {
+    acc.ttfbSumMs += ttfb;
+    acc.ttfbN++;
+  }
+  const totalMs = ms(r.totalMs ?? r.total_ms);
+  if (totalMs != null && totalMs > 0) {
+    acc.totalSumMs += totalMs;
+    acc.totalN++;
+  }
+  // 生成阶段耗时 = 总耗时 - 首字（首字缺失按 0 计）
+  if (totalMs != null && totalMs > 0 && comp > 0) {
+    const compMs = Math.max(0, totalMs - (ttfb ?? 0));
+    if (compMs > 0) {
+      acc.completionMsSum += compMs;
+      acc.tpsTokSum += comp;
+    }
+  }
+  return acc;
+}
+
+// 纯函数：只做过滤 + 归并，不碰磁盘。
+export function aggregateUsage(rows, { hours = 24, now = Date.now(), model = null } = {}) {
+  const windowHours = Number.isFinite(Number(hours)) && Number(hours) > 0 ? Number(hours) : 24;
+  const untilN = Number(now);
+  const until = Number.isFinite(untilN) ? untilN : Date.now();
+  const since = until - windowHours * HOUR_MS;
+  const byModel = new Map();
+  const sum = blank(null);
+  let scanned = 0;
+
+  if (Array.isArray(rows)) {
+    for (const r of rows) {
+      if (!r || typeof r !== "object") continue;
+      const ts = Number(r.ts);
+      if (!Number.isFinite(ts) || ts < since || ts > until) continue;
+      const id = String(r.model || "").trim();
+      if (!id) continue;
+      // model 过滤同时接受 canonical 全称与裸 id（如 --model big-pickle 匹配 opencode/big-pickle 行）
+      if (model && id !== model && !id.endsWith("/" + model)) continue;
+      scanned++;
+      let acc = byModel.get(id);
+      if (!acc) { acc = blank(id); byModel.set(id, acc); }
+      fold(acc, r);
+      fold(sum, r);
+    }
+  }
+  const models = [...byModel.values()].map(finalize);
+  models.sort((a, b) => (b.totalTokens - a.totalTokens) || (b.requests - a.requests) || a.id.localeCompare(b.id));
+  return {
+    windowHours,
+    since,
+    until,
+    rows: scanned,
+    models,
+    totals: finalize(sum),
+  };
+}
+
+// 读保留期内的日文件并只留窗口内的行（保留期默认 2 天，文件数很少，全读即可）。
+export async function readUsageRows({ dir, hours = 24, now = Date.now() } = {}) {
+  const since = Number(now) - (Number(hours) > 0 ? Number(hours) : 24) * HOUR_MS;
+  let files = [];
+  try {
+    files = (await readdir(usageDir({ dir }))).filter((f) => f.endsWith(".jsonl")).sort();
+  } catch {
+    return [];
+  }
+  const rows = [];
+  for (const f of files) {
+    let text = "";
+    try {
+      text = await readFile(join(usageDir({ dir }), f), "utf8");
+    } catch {
+      continue;
+    }
+    for (const line of text.split("\n")) {
+      if (!line) continue;
+      try {
+        const r = JSON.parse(line);
+        if (Number(r?.ts) >= since) rows.push(r);
+      } catch {}
+    }
+  }
+  return rows;
+}
+
+export async function usageReport({ dir, hours = 24, now = Date.now(), model = null } = {}) {
+  const rows = await readUsageRows({ dir, hours, now });
+  return aggregateUsage(rows, { hours, now, model });
+}