mingdao-harness 0.1.54

package/src/batch.js

@@ -0,0 +1,182 @@
+// Batch API 半价通道（四报告共识 A1/OfficeACE P0：批量任务省 50%）：
+// 单轮批量任务（无工具、无流式）走 OpenAI 兼容批处理协议：
+//   POST {base}/files（multipart，purpose=batch）→ input_file_id
+//   POST {base}/batches {input_file_id, endpoint, completion_window:'24h'} → batchId
+//   轮询 GET {base}/batches/{id} → completed → 下载结果（DeepSeek 风格 /files/result，回退 OpenAI /files/{id}/content）
+// 端点不可用（404/405）→ 明确报错告知网关不支持批处理，绝不静默假装成功。
+// 计费：闲时全未命中 × BATCH_DISCOUNT（0.5），结果记入 cache-stats（batch:true）供 /cost 汇总。
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { resolveProviderConfig } from './providers/index.js';
+import { estimateBatchCost, BATCH_DISCOUNT } from './pricing.js';
+import { recordCacheStats } from './cachestats.js';
+import { buildSystemPrompt } from './prompts.js';
+
+const DEFAULT_WINDOW = '24h';
+const DEFAULT_ENDPOINT = '/v1/chat/completions';
+
+// 批处理端点基址：config.batchBaseUrl 优先；否则取当前服务商 baseUrl 去掉 /v1 后缀
+function batchBase(cfg, model) {
+  const explicit = String(cfg?.batchBaseUrl || '').trim().replace(/\/+$/, '');
+  if (explicit) return explicit;
+  const pc = resolveProviderConfig(cfg, model);
+  // 审计 P2-10：DeepSeek 的批处理端点在根路径（/files /batches），其余 OpenAI 兼容网关在 /v1 下
+  const base = String(pc.baseUrl || '').replace(/\/+$/, '');
+  return pc.name === 'deepseek' ? base.replace(/\/v1\/?$/, '') : base;
+}
+
+async function api(base, apiKey, methodPath, payload, httpMethod = 'POST') {
+  const res = await fetch(base + methodPath, {
+    method: httpMethod,
+    headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${apiKey}` },
+    body: payload === undefined ? undefined : JSON.stringify(payload),
+  });
+  const j = await res.json().catch(() => ({}));
+  if (!res.ok) {
+    const e = new Error(j?.error?.message || j?.message || `HTTP ${res.status}`);
+    e.status = res.status;
+    throw e;
+  }
+  return j;
+}
+
+async function uploadFile(base, apiKey, jsonl) {
+  const form = new FormData();
+  form.append('file', new Blob([jsonl], { type: 'application/jsonl' }), 'mingdao-batch.jsonl');
+  form.append('purpose', 'batch');
+  const res = await fetch(base + '/files', {
+    method: 'POST',
+    headers: { Authorization: `Bearer ${apiKey}` },
+    body: form,
+  });
+  const j = await res.json().catch(() => ({}));
+  if (!res.ok) {
+    const e = new Error(j?.error?.message || j?.message || `上传失败 HTTP ${res.status}`);
+    e.status = res.status;
+    throw e;
+  }
+  return j.id;
+}
+
+async function downloadResults(base, apiKey, batch) {
+  // DeepSeek 风格：直接取结果文件；回退 OpenAI 风格：按 output_file_id 取内容
+  const attempts = [
+    `/batches/${batch.id}/files/result`,
+    ...(batch.output_file_id ? [`/files/${batch.output_file_id}/content`] : []),
+  ];
+  for (const m of attempts) {
+    const res = await fetch(base + m, { headers: { Authorization: `Bearer ${apiKey}` } });
+    if (!res.ok) continue;
+    const text = await res.text();
+    return text
+      .split('\n')
+      .filter(Boolean)
+      .map((l) => {
+        try {
+          return JSON.parse(l);
+        } catch {
+          return null;
+        }
+      })
+      .filter(Boolean);
+  }
+  throw new Error('批处理结果文件不可用（端点不支持或文件已过期）');
+}
+
+// 执行一次批处理。questions: string[]。返回 { ok, outputFile, results, usage, cost, batchId }
+export async function runBatch({ cfg, model, questions, workingDir = process.cwd(), maxTokens = 4096, temperature, signal, onStatus }) {
+  const list = (questions || []).map((q) => String(q).trim()).filter(Boolean);
+  if (!list.length) return { error: '没有可批处理的问题（每行一个问题）' };
+  const pc = resolveProviderConfig(cfg, model);
+  if (!pc.apiKey) return { error: `模型 ${model} 没有可用 API Key（mingdao key set ${pc.name}）` };
+  const base = batchBase(cfg, model);
+  const apiKey = pc.apiKey;
+  const systemPrompt = buildSystemPrompt({ workingDir });
+  const bodyTemplate = {
+    model,
+    messages: null, // 逐行填充
+    max_tokens: maxTokens,
+    temperature: temperature ?? cfg?.temperature ?? 0.6,
+    stream: false,
+  };
+  const jsonl =
+    list
+      .map((q, i) =>
+        JSON.stringify({
+          custom_id: `md-${i}`,
+          method: 'POST',
+          url: cfg?.batchEndpoint || DEFAULT_ENDPOINT,
+          body: { ...bodyTemplate, messages: [{ role: 'system', content: systemPrompt }, { role: 'user', content: q }] },
+        })
+      )
+      .join('\n') + '\n';
+
+  try {
+    onStatus?.('上传输入文件…');
+    const fileId = await uploadFile(base, apiKey, jsonl);
+    onStatus?.('创建批处理任务…');
+    const batch = await api(base, apiKey, '/batches', {
+      input_file_id: fileId,
+      endpoint: cfg?.batchEndpoint || DEFAULT_ENDPOINT,
+      completion_window: cfg?.batchWindow || DEFAULT_WINDOW,
+    });
+    onStatus?.(`任务已创建：${batch.id}`);
+    // 轮询（间隔可用 MINGDAO_BATCH_POLL_MS 覆盖，测试用）；连续失败 10 次 → 报错，绝不无限重试
+    const interval = Math.max(500, Number(process.env.MINGDAO_BATCH_POLL_MS) || 5000);
+    const t0 = Date.now();
+    let st = batch.status;
+    let failures = 0;
+    for (;;) {
+      if (signal?.aborted) return { error: '已取消轮询（任务仍在服务端运行）', batchId: batch.id };
+      if (Date.now() - t0 > 24 * 3600 * 1000) return { error: '批处理超过 24h 窗口', batchId: batch.id };
+      let j = null;
+      try {
+        j = await api(base, apiKey, `/batches/${batch.id}`, undefined, 'GET');
+        failures = 0;
+      } catch (err) {
+        failures += 1;
+        if (failures >= 10) return { error: `轮询失败：${err?.message || err}（任务仍在服务端，ID ${batch.id}）`, batchId: batch.id };
+        await new Promise((r) => setTimeout(r, interval));
+        continue;
+      }
+      st = j.status;
+      if (st === 'completed') {
+        batch.output_file_id = j.output_file_id;
+        break;
+      }
+      if (['failed', 'expired', 'cancelled', 'canceled'].includes(st)) {
+        const detail = j?.errors?.data?.[0]?.message || j?.errors?.message || '';
+        return { error: `批处理失败：${st}${detail ? '（' + detail + '）' : ''}`, batchId: batch.id };
+      }
+      if (st !== 'in_progress' && st !== 'validating' && st !== 'finalizing') {
+        // 未知状态：继续等待但报告
+        onStatus?.(`状态：${st}`);
+      }
+      await new Promise((r) => setTimeout(r, interval));
+    }
+    onStatus?.('下载结果…');
+    const results = await downloadResults(base, apiKey, batch);
+    // 汇总 usage 与费用（batch 半价）
+    let prompt = 0;
+    let completion = 0;
+    const outputs = [];
+    for (const r of results) {
+      const body = r?.response?.body || {};
+      prompt += body.usage?.prompt_tokens || 0;
+      completion += body.usage?.completion_tokens || 0;
+      const content = body.choices?.[0]?.message?.content ?? body.choices?.[0]?.text ?? (r?.response?.status_code !== 200 ? `（错误 ${r?.response?.status_code}）` : '');
+      outputs.push({ id: r.custom_id, content: String(content || '').trim() });
+    }
+    const usage = { prompt_tokens: prompt, completion_tokens: completion };
+    const cost = estimateBatchCost(model, prompt, completion);
+    const outFile = path.join(workingDir, `mingdao-batch-result-${Date.now()}.jsonl`);
+    fs.writeFileSync(outFile, outputs.map((o) => JSON.stringify(o)).join('\n') + '\n');
+    recordCacheStats({ model, prompt, completion, hit: null, miss: null, cost, saved: null, batch: true });
+    onStatus?.(`完成：${outputs.length} 条结果`);
+    return { ok: true, batchId: batch.id, outputFile: outFile, results: outputs, usage, cost, discount: BATCH_DISCOUNT };
+  } catch (err) {
+    // 批处理端点不支持（404/405 等）→ 明确告知，不静默
+    return { error: `批处理不可用：${err?.message || err}（该服务商可能不支持 Batch API，可用 config.batchBaseUrl 指定支持的网关）` };
+  }
+}

package/src/cachestats.js

@@ -0,0 +1,215 @@
+// 缓存命中统计：记录每次用量的缓存命中/未命中，支撑「缓存命中率仪表盘」。
+// 落盘 <home>/cache-stats.jsonl，每行：{at, model, prompt, completion, hit, miss, cost, saved}
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { mingdaoHome, ensureHome } from './config.js';
+import { estimateCost, cacheSplit, beijingDayStart, beijingParts } from './pricing.js';
+import { modelPreset } from './models.js';
+
+export function cacheStatsFile() {
+  return path.join(mingdaoHome(), 'cache-stats.jsonl');
+}
+
+// 内存计数 + 低频轮转（评估 P3-3：此前无限增长且每次全量解析）
+const MAX_LINES = 20000;
+const KEEP_LINES = 10000;
+let cacheStatsCount = 0;
+
+export function recordCacheStats(entry) {
+  try {
+    ensureHome(); // 审计 B10：与 audit/journal 一致，目录缺失不静默丢统计
+    const line = JSON.stringify({
+      at: Date.now(),
+      model: entry.model || '',
+      prompt: entry.prompt || 0,
+      completion: entry.completion || 0,
+      hit: entry.hit ?? null,
+      miss: entry.miss ?? null,
+      cost: entry.cost ?? null,
+      saved: entry.saved ?? null,
+      batch: entry.batch === true ? true : undefined, // Batch API 半价任务标记（/cost 分账展示）
+      steps: entry.steps ?? undefined, // 本回合步数（状态栏 步数统计）
+      llmMs: entry.llmMs ?? undefined, // 模型调用总耗时（状态栏 LLM 时长/tok-s）
+      toolMs: entry.toolMs ?? undefined, // 工具执行总耗时（状态栏 工具调用时长）
+      firstTokenMs: entry.firstTokenMs ?? undefined, // 首 token 延迟（状态栏平均首 token）
+    });
+    fs.appendFileSync(cacheStatsFile(), line + '\n');
+    cacheStatsCount += 1;
+  } catch {}
+  if (cacheStatsCount > MAX_LINES && cacheStatsCount % 200 === 0) {
+    try {
+      const raw = fs.readFileSync(cacheStatsFile(), 'utf8');
+      const lines = raw.split('\n').filter(Boolean);
+      if (lines.length > MAX_LINES) {
+        fs.writeFileSync(cacheStatsFile(), lines.slice(-KEEP_LINES).join('\n') + '\n');
+        cacheStatsCount = KEEP_LINES;
+      }
+    } catch {}
+  }
+}
+
+export function listCacheStats(limit = 2000) {
+  try {
+    const raw = fs.readFileSync(cacheStatsFile(), 'utf8');
+    const out = [];
+    for (const l of raw.split('\n')) {
+      if (!l.trim()) continue;
+      try {
+        out.push(JSON.parse(l));
+      } catch {}
+    }
+    return out.slice(-limit);
+  } catch {
+    return [];
+  }
+}
+
+export function summarizeCacheStats(entries) {
+  const sum = { turns: entries.length, prompt: 0, completion: 0, hit: 0, miss: 0, cost: 0, saved: 0, steps: 0, llmMs: 0, toolMs: 0, firstTokenCount: 0, firstTokenSumMs: 0 };
+  for (const e of entries) {
+    sum.prompt += e.prompt || 0;
+    sum.completion += e.completion || 0;
+    sum.hit += e.hit || 0;
+    sum.miss += e.miss || 0;
+    sum.cost += e.cost || 0;
+    sum.saved += e.saved || 0;
+    sum.steps += e.steps || 0;
+    sum.llmMs += e.llmMs || 0;
+    sum.toolMs += e.toolMs || 0;
+    if (e.firstTokenMs != null) {
+      sum.firstTokenCount += 1;
+      sum.firstTokenSumMs += e.firstTokenMs;
+    }
+  }
+  sum.rate = sum.hit + sum.miss > 0 ? sum.hit / (sum.hit + sum.miss) : null;
+  sum.firstTokenAvgMs = sum.firstTokenCount > 0 ? sum.firstTokenSumMs / sum.firstTokenCount : null;
+  sum.tokensPerSec = sum.llmMs > 0 ? (sum.completion / (sum.llmMs / 1000)) : null;
+  return sum;
+}
+
+// 记录一次用量（agent 侧调用）：自动计算命中拆分与节省额
+// 记录一次用量（agent 侧调用）：自动计算命中拆分与节省额；perf 为回合性能指标（状态栏）
+export function recordUsage(modelName, usage, perf = null) {
+  const split = cacheSplit(usage);
+  const prompt = usage?.prompt_tokens || 0;
+  const completion = usage?.completion_tokens || 0;
+  let cost = null;
+  let saved = null;
+  if (split) {
+    cost = estimateCost(modelName, prompt, completion, split);
+    saved = estimateCost(modelName, prompt, completion, null) - cost;
+  } else if (modelPreset(modelName)?.pricing) {
+    // 审计 B10：无缓存字段时按全未命中估算（不再计 0 元，费用护栏口径更真实）
+    cost = estimateCost(modelName, prompt, completion, null);
+  }
+  recordCacheStats({
+    model: modelName,
+    prompt,
+    completion,
+    hit: split?.hit ?? null,
+    miss: split?.miss ?? null,
+    cost,
+    saved,
+    steps: perf?.steps ?? undefined,
+    llmMs: perf?.llmMs ?? undefined,
+    toolMs: perf?.toolMs ?? undefined,
+    firstTokenMs: perf?.firstTokenMs ?? undefined,
+  });
+}
+
+export function formatCacheSummary(sum) {
+  const fmt = (n) => (n >= 1000 ? (n / 1000).toFixed(1) + 'k' : String(n));
+  const lines = [
+    `轮次  ${sum.turns} · ↑${fmt(sum.prompt)} ↓${fmt(sum.completion)} tokens`,
+    `缓存命中率  ${sum.rate != null ? (sum.rate * 100).toFixed(0) + '%' : '暂无缓存数据'}`,
+    `实际费用  ≈¥${sum.cost.toFixed(5)} · 相比全未命中节省 ≈¥${sum.saved.toFixed(5)}`,
+  ];
+  return lines;
+}
+
+// 分账统计（评估 /cost 升级）：按模型分账、今日费用、batch 半价任务、节省归因
+export function costBreakdown() {
+  const entries = listCacheStats(100000);
+  const byModel = new Map();
+  const start = beijingDayStart().getTime();
+  let totalCost = 0;
+  let totalSaved = 0;
+  let today = 0;
+  let batchCost = 0;
+  let hit = 0;
+  let miss = 0;
+  for (const e of entries) {
+    totalCost += e.cost || 0;
+    totalSaved += e.saved || 0;
+    hit += e.hit || 0;
+    miss += e.miss || 0;
+    if (e.batch) batchCost += e.cost || 0;
+    if (e.at >= start) today += e.cost || 0;
+    const m = byModel.get(e.model) || { prompt: 0, completion: 0, cost: 0, saved: 0, turns: 0, batchTurns: 0 };
+    m.prompt += e.prompt || 0;
+    m.completion += e.completion || 0;
+    m.cost += e.cost || 0;
+    m.saved += e.saved || 0;
+    m.turns += 1;
+    if (e.batch) m.batchTurns += 1;
+    byModel.set(e.model, m);
+  }
+  return {
+    totalCost,
+    totalSaved,
+    today,
+    batchCost,
+    hit,
+    miss,
+    rate: hit + miss > 0 ? hit / (hit + miss) : null,
+    byModel: [...byModel.entries()].map(([model, m]) => ({ model, ...m })).sort((a, b) => b.cost - a.cost),
+  };
+}
+
+// 月度费用报告（/cost 导出）：按北京时间月份聚合，month 形如 'YYYY-MM'，缺省返回全部月份
+export function costMonthlyReport(month) {
+  const entries = listCacheStats(100000);
+  const want = String(month || '').trim();
+  const months = new Map();
+  for (const e of entries) {
+    const p = beijingParts(new Date(e.at));
+    const key = `${p.year}-${String(p.month).padStart(2, '0')}`;
+    if (want && key !== want) continue;
+    let m = months.get(key);
+    if (!m) {
+      m = { month: key, cost: 0, saved: 0, prompt: 0, completion: 0, hit: 0, miss: 0, batchCost: 0, days: new Map(), models: new Map() };
+      months.set(key, m);
+    }
+    m.cost += e.cost || 0;
+    m.saved += e.saved || 0;
+    m.prompt += e.prompt || 0;
+    m.completion += e.completion || 0;
+    m.hit += e.hit || 0;
+    m.miss += e.miss || 0;
+    if (e.batch) m.batchCost += e.cost || 0;
+    const day = String(p.day).padStart(2, '0');
+    m.days.set(day, (m.days.get(day) || 0) + (e.cost || 0));
+    const mm = m.models.get(e.model) || { prompt: 0, completion: 0, cost: 0, turns: 0 };
+    mm.prompt += e.prompt || 0;
+    mm.completion += e.completion || 0;
+    mm.cost += e.cost || 0;
+    mm.turns += 1;
+    m.models.set(e.model, mm);
+  }
+  return [...months.values()]
+    .sort((a, b) => a.month.localeCompare(b.month))
+    .map((m) => ({
+      month: m.month,
+      cost: m.cost,
+      saved: m.saved,
+      prompt: m.prompt,
+      completion: m.completion,
+      hit: m.hit,
+      miss: m.miss,
+      rate: m.hit + m.miss > 0 ? m.hit / (m.hit + m.miss) : null,
+      batchCost: m.batchCost,
+      days: [...m.days.entries()].sort((a, b) => a[0].localeCompare(b[0])).map(([day, cost]) => ({ day, cost })),
+      models: [...m.models.entries()].map(([model, x]) => ({ model, ...x })).sort((a, b) => b.cost - a.cost),
+    }));
+}