mslxdff 0.1.203 → 0.1.207

package/src/model-trace.js

@@ -1,187 +1,187 @@
-// Per-model request trace formatting. Never include prompt, response body, headers or credentials.
-// Note: 同步 appendFileSync（异步 append 会在并发下打乱同一请求的阶段行序）— 见 .agents/notes/implemented/feature/2026-09-25-model-trace-log.md
-// Note: 事件面用黑名单（默认全可见，只排除噪声/敏感面）+ 决定类事件只渲染登记过的标量字段——起因是 empty-turn-retry 曾被阶段白名单静默吞掉，"关键决定不许再被漏登记" — 见 .agents/notes/implemented/feature/2026-09-27-qoder-per-request-sticky-account.md
-import { appendFileSync, existsSync, mkdirSync, readFileSync, statSync, writeFileSync } from "node:fs";
-import { join } from "node:path";
-import { logDir } from "./logs.js";
-import { fmtShanghaiYMDHMS } from "./time.js";
-
-const MAX_TRACE_BYTES = 1024 * 1024;
-// 裁剪后保留的比例（按**字节**回退）。与 logs.js 的 KEEP_RATIO 同口径：
-// 旧的 `slice(-100)` 是按**行数**回退，与按字节的阈值严重不匹配 —— 实测
-// cline-cline-pass-deepseek-v4.1-flash.log 到 1,006,482 字节 / 5346 行时，
-// 一次裁剪丢 5246 行（98.1% 历史），事后回溯基本不可用。按字节留 3/4 则每次只丢 ~1/4。
-const TRACE_KEEP_RATIO = 3 / 4;
-
-export function modelLogName(model) {
-  const s = String(model || "unknown").trim().toLowerCase();
-  const safe = s.replace(/[^a-z0-9._-]+/g, "-").replace(/^[-.]+|[-.]+$/g, "").slice(0, 180);
-  return `${safe || "unknown"}.log`;
-}
-
-export function modelLogFile(model) {
-  return join(logDir(), modelLogName(model));
-}
-
-
-function count(v) {
-  const n = Number(v);
-  return Number.isFinite(n) && n >= 0 ? n : null;
-}
-
-export function summarizeRequest(body = {}) {
-  const messages = Array.isArray(body.messages) ? body.messages : [];
-  const roles = {};
-  for (const m of messages) roles[String(m?.role || "unknown")] = (roles[String(m?.role || "unknown")] || 0) + 1;
-  return {
-    stream: body.stream === true,
-    messages: messages.length,
-    roles,
-    tools: Array.isArray(body.tools) ? body.tools.length : 0,
-    maxTokens: count(body.max_tokens ?? body.max_completion_tokens),
-    temperature: Number.isFinite(Number(body.temperature)) ? Number(body.temperature) : null,
-    topP: Number.isFinite(Number(body.top_p)) ? Number(body.top_p) : null,
-  };
-}
-
-function hostPort(value) {
-  try { const u = new URL(String(value)); return u.port ? `${u.hostname}:${u.port}` : u.hostname; } catch { return "-"; }
-}
-// 黑名单（而非白名单）：模型日志默认收下**所有**事件，只排除噪声与凭据面。
-// 起因：`empty-turn-retry`（"切号换 URL"的真实触发者）曾被白名单静默吞掉，排障时只见现象不见原因。
-// 关键决定不允许再被漏登记：新增事件默认可见，确属噪声/敏感才在此登记排除。
-const TRACE_DENY = new Set([
-  "peer-health",     // 组员心跳，秒级高频，无决策价值
-  "heartbeat",       // 运行心跳
-  "client-session",  // 客户端会话标识（路由用），非决定且属敏感面
-]);
-
-export function shouldTraceModel(type) {
-  const t = String(type || "");
-  if (!t) return false;
-  if (TRACE_DENY.has(t)) return false;
-  if (t.startsWith("upstream-probe")) return false; // 探针类（probe/probe-error/probe-skip）
-  return true;
-}
-
-function safeText(value, n = 140) {
-  return String(value ?? "")
-    .replace(/([?&](?:token|refreshToken|access[_-]?token|refresh[_-]?token|api[_-]?key|apikey|key|secret|password|cookie)=)[^&\s]+/gi, "$1[redacted]")
-    .replace(/(authorization|bearer|access[_-]?token|refresh[_-]?token|cookie|api[-_]?key|password|secret)\s*[:=]\s*[^,; ]+/gi, "$1=[redacted]")
-    .replace(/\s+/g, " ").trim().slice(0, n);
-}
-
-function kv(obj) {
-  return Object.entries(obj).filter(([, v]) => v != null && v !== "").map(([k, v]) => `${k}=${typeof v === "object" ? JSON.stringify(v) : v}`).join(" ");
-}
-
-// 决定类事件的可见字段白名单：只渲染这些标量；事件里的 payload/detail/stages/请求正文等一律不落模型日志。
-// 「加日志」不等于「倒数据」——新增决定字段要显式登记在这里才会出现。
-const DECISION_FIELDS = [
-  "status", "from", "to", "reason", "hook", "applied", "skipped", "plugin",
-  "provider", "before", "after", "models", "skippedFaulty", "limit", "tried",
-  "winPeer", "winTarget", "peer", "peers", "peerLabel", "routeBest", "at",
-  "latencyMs", "ttfMs", "hedged", "hops", "hasShare", "elapsedMs", "threshold",
-  "stallHits", "maxGapMs", "interrupted", "providers", "via",
-  "retry", "step", "max", "delayMs", "upstream", "account", "pick", "cooled", "queued", "nudged", "retries", "waitedMs", "raiseFrom", "raiseTo",
-  "heldOpen", "terminalForm", "requestWaitedMs", "budgetMs", "budgetOut",
-  // 终局判决（ADR-0058）：取证时第一眼要看的三件——判的哪种、命中哪条判据、是否因此没问组员
-  "kind", "marker", "peersSkipped",
-];
-
-function decisionKv(data = {}) {
-  const out = [];
-  for (const k of DECISION_FIELDS) {
-    const v = data[k];
-    if (v == null || v === "") continue;
-    out.push(`${k}=${Array.isArray(v) ? v.join("|") : v}`);
-  }
-  return out.join(" ");
-}
-
-// provider 回显头 → 日志字段（谁上的 / 打哪个站 / 为什么选它 / 是否被冷却 / 是否被判排队）：
-// 单一来源，防各处重复写法漂移；取不到时返回空对象（其它供应商零变化）。
-export function upstreamEcho(res) {
-  const h = res?.headers;
-  if (typeof h?.get !== "function") return {};
-  const pick = (name) => h.get(name) || null;
-  return {
-    upstream: pick("x-mslxdff-upstream"),
-    account: pick("x-mslxdff-workbuddy-uid") || pick("x-mslxdff-qoder-region"),
-    pick: pick("x-mslxdff-qoder-account"),
-    cooled: pick("x-mslxdff-qoder-cooldown"),
-    queued: pick("x-mslxdff-qoder-queued"),
-  };
-}
-
-export function formatModelTrace({ type, reqId, model, data = {}, request = null, totalMs = null } = {}) {
-  const base = { time: fmtShanghaiYMDHMS(new Date()), req: reqId || "-", model: model || "-", stage: type || "-" };
-  let detail = "";
-  // api= 打在 incoming 行最前：门是排查第一眼要的东西，grep "incoming api=" 直接分得开三门；无值时 kv 自然不落该段（不发散成 api=other）
-  if (type === "request") detail = `incoming ${kv({ api: data.api, ...(request || summarizeRequest(data.body)), hops: data.hops, useAuto: data.useAuto })}`;
-  else if (type === "ordered") detail = `route order=${(data.order || []).join(" | ")} hops=${data.hops ?? "-"} fallback=${data.canFallback ? 1 : 0}`;
-  else if (type === "model-try") detail = `target=${data.model || model || "-"} idx=${data.idx ?? "-"} remaining=${data.remaining ?? "-"}`;
-  else if (type === "alias") detail = `rawModel=${data.rawModel || "-"} requested=${data.requested || model || "-"}`;
-  else if (type === "exhausted-local" || type === "exhausted-all") detail = `last=${data.lastModel || model || "-"} status=${data.lastStatus ?? "-"} order=${(data.order || []).join(" | ")}`;
-  else if (type === "upstream-try") detail = `target=${data.model || model || "-"} attempt=${data.attempt ?? "-"} payload=${kv(data.payload || {})}`;
-  else if (type === "upstream-done" || type === "upstream-error") detail = `upstream status=${data.status ?? "-"} timing=${data.timing?.totalMs ?? "-"}ms${data.upstream ? ` via=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}${data.cooled ? ` cooled=${data.cooled}` : ""}${data.queued ? ` queued=1` : ""} ${safeText(data.message || data.error || "")}`;
-  else if (type === "empty-turn-retry") detail = `空转重试：retry ${data.retry ?? "-"}/${data.max ?? "-"}${data.step != null ? ` step=${data.step}` : ""} delay=${data.delayMs ?? "-"}ms${data.waitedMs ? ` waited=${data.waitedMs}ms` : ""}${data.requestWaitedMs != null ? ` 请求级累计=${data.requestWaitedMs}ms/预算${data.budgetMs ?? "-"}ms` : ""}${data.raiseTo ? ` max_tokens ${data.raiseFrom}→${data.raiseTo}` : ""}${data.upstream ? ` after=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}${data.nudged ? ` nudged=1` : ""} reason=empty turn${data.reason ? ` :: ${safeText(data.reason, 120)}` : ""}`;
-  else if (type === "empty-turn-recovered") detail = `空转重试成功：第 ${data.retries ?? "-"}/${data.max ?? "-"} 次重拉拿到真实输出，本轮共白等 ${data.waitedMs ?? "-"}ms${data.upstream ? ` via=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}`;
-  else if (type === "empty-turn-exhausted") detail = `空转重试已用尽：重试 ${data.retries ?? 0}/${data.max ?? "-"} 次仍零输出（累计等待 ${data.waitedMs ?? 0}ms${data.budgetOut ? `，请求级等待预算 ${data.budgetMs ?? "-"}ms 用尽` : ""}），交回换候选${data.heldOpen ? "（响应仍开着，换过去的正文还能送达）" : "（响应已封口，重拉只为省额度）"}${data.upstream ? ` after=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}${data.reason ? ` :: ${safeText(data.reason, 120)}` : ""}`;
-  else if (type === "peer-request" || type === "peer-forward" || type === "peer-attempt" || type === "peer-error") detail = `peer=${hostPort(data.peer)}${data.name ? ` name=${data.name}` : ""} status=${data.status ?? "-"} latency=${data.latencyMs ?? "-"}ms payload=${kv(data.payload || {})} ${safeText(data.message || data.error || "")}`;
-  else if (type === "relay-done") {
-    const d = data.detail || {};
-    const u = d.usage || {};
-    // 缺 usage 时显式标注来源，别与「真的用了 0 token」混成同一个 `-`（BUG-13：经组员中继的响应
-    // 少数轮次拿不到 usage，成本报表会静默偏低，读日志的人也会误判成"这轮没消耗"）。
-    const noUsage = u.prompt_tokens == null && u.completion_tokens == null;
-    const usageTxt = noUsage
-      ? `usage(none${data.via === "peer" ? ":peer 未回" : ""})`
-      : `usage(prompt=${u.prompt_tokens ?? "-"}/completion=${u.completion_tokens ?? "-"})`;
-    detail = `upstream_response status=${data.status ?? "-"} finish=${d.sawFinishReason || "-"} chunks=${d.receivedChunks ?? "-"} bytes=${d.receivedBytes ?? "-"} tools=${d.toolCalls ?? 0} chars=${d.chars ?? 0} ${usageTxt} elapsed=${data.totalMs ?? "-"}ms${data.upstream ? ` upstream=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}`;
-  } else if (type === "result" || type === "client-response") {
-    detail = `client status=${data.status ?? "-"} via=${data.via || "-"} actual=${data.actual || model || "-"} fallback=${data.fallback?.fallback ? 1 : 0} total=${totalMs ?? data.durationMs ?? "-"}ms${data.upstream ? ` upstream=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}`;
-  } else detail = safeText(`${decisionKv(data)} ${data.message || data.error || ""}`, 200);
-  return `[${base.time}] [req=${base.req}] [model=${base.model}] [stage=${base.stage}] ${detail}`;
-}
-
-/**
- * 按字节从尾部回退（不截半行）。行有长有短，固定行数 = 固定容量 ⇒ 一次裁剪就是断崖。
- * 与 logs.js 的 keepTailByBytes 同口径（那边不导出，这里留一份最小实现，避免为一行日志反向依赖）。
- */
-function keepTailByBytes(text, maxBytes) {
-  const limit = Math.max(1, Math.floor(maxBytes * TRACE_KEEP_RATIO));
-  if (Buffer.byteLength(text) <= limit) return text;
-  const lines = text.split("\n");
-  let bytes = 0;
-  let i = lines.length - 1;
-  for (; i >= 0; i--) {
-    const b = Buffer.byteLength(lines[i]) + 1;
-    if (bytes + b > limit) break;
-    bytes += b;
-  }
-  // 至少保留最后一条：单行超 limit 时否则返回空串 = 清空整个 trace。
-  // 判据是「结果为空」——文件尾有换行时 split 的最后一个元素是空串，`i === lines.length-1` 判不到内容行。
-  const kept = lines.slice(i + 1).join("\n");
-  if (!kept) {
-    const nonEmpty = lines.filter((l) => l !== "");
-    if (nonEmpty.length) return nonEmpty[nonEmpty.length - 1];
-  }
-  return kept;
-}
-
-export function appendModelTrace(model, entry, { file = modelLogFile(model) } = {}) {
-  const line = `${formatModelTrace(entry)}\n`;
-  try {
-    mkdirSync(logDir(), { recursive: true });
-    if (existsSync(file) && statSync(file).size > MAX_TRACE_BYTES) {
-      // 按字节留 3/4，不再固定砍到 100 行（那一次就丢 98% 历史，见 TRACE_KEEP_RATIO 注释）。
-      const kept = keepTailByBytes(readFileSync(file, "utf8"), MAX_TRACE_BYTES);
-      // 末尾补换行：下一次 appendFileSync 才不会把新行接到半截行尾巴上
-      writeFileSync(file, kept.endsWith("\n") ? kept : `${kept}\n`);
-    }
-    appendFileSync(file, line);
-  } catch {
-    // Logging must never affect the chat request.
-  }
-}
+// Per-model request trace formatting. Never include prompt, response body, headers or credentials.
+// Note: 同步 appendFileSync（异步 append 会在并发下打乱同一请求的阶段行序）— 见 .agents/notes/implemented/feature/2026-09-25-model-trace-log.md
+// Note: 事件面用黑名单（默认全可见，只排除噪声/敏感面）+ 决定类事件只渲染登记过的标量字段——起因是 empty-turn-retry 曾被阶段白名单静默吞掉，"关键决定不许再被漏登记" — 见 .agents/notes/implemented/feature/2026-09-27-qoder-per-request-sticky-account.md
+import { appendFileSync, existsSync, mkdirSync, readFileSync, statSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+import { logDir } from "./logs.js";
+import { fmtShanghaiYMDHMS } from "./time.js";
+
+const MAX_TRACE_BYTES = 1024 * 1024;
+// 裁剪后保留的比例（按**字节**回退）。与 logs.js 的 KEEP_RATIO 同口径：
+// 旧的 `slice(-100)` 是按**行数**回退，与按字节的阈值严重不匹配 —— 实测
+// cline-cline-pass-deepseek-v4.1-flash.log 到 1,006,482 字节 / 5346 行时，
+// 一次裁剪丢 5246 行（98.1% 历史），事后回溯基本不可用。按字节留 3/4 则每次只丢 ~1/4。
+const TRACE_KEEP_RATIO = 3 / 4;
+
+export function modelLogName(model) {
+  const s = String(model || "unknown").trim().toLowerCase();
+  const safe = s.replace(/[^a-z0-9._-]+/g, "-").replace(/^[-.]+|[-.]+$/g, "").slice(0, 180);
+  return `${safe || "unknown"}.log`;
+}
+
+export function modelLogFile(model) {
+  return join(logDir(), modelLogName(model));
+}
+
+
+function count(v) {
+  const n = Number(v);
+  return Number.isFinite(n) && n >= 0 ? n : null;
+}
+
+export function summarizeRequest(body = {}) {
+  const messages = Array.isArray(body.messages) ? body.messages : [];
+  const roles = {};
+  for (const m of messages) roles[String(m?.role || "unknown")] = (roles[String(m?.role || "unknown")] || 0) + 1;
+  return {
+    stream: body.stream === true,
+    messages: messages.length,
+    roles,
+    tools: Array.isArray(body.tools) ? body.tools.length : 0,
+    maxTokens: count(body.max_tokens ?? body.max_completion_tokens),
+    temperature: Number.isFinite(Number(body.temperature)) ? Number(body.temperature) : null,
+    topP: Number.isFinite(Number(body.top_p)) ? Number(body.top_p) : null,
+  };
+}
+
+function hostPort(value) {
+  try { const u = new URL(String(value)); return u.port ? `${u.hostname}:${u.port}` : u.hostname; } catch { return "-"; }
+}
+// 黑名单（而非白名单）：模型日志默认收下**所有**事件，只排除噪声与凭据面。
+// 起因：`empty-turn-retry`（"切号换 URL"的真实触发者）曾被白名单静默吞掉，排障时只见现象不见原因。
+// 关键决定不允许再被漏登记：新增事件默认可见，确属噪声/敏感才在此登记排除。
+const TRACE_DENY = new Set([
+  "peer-health",     // 组员心跳，秒级高频，无决策价值
+  "heartbeat",       // 运行心跳
+  "client-session",  // 客户端会话标识（路由用），非决定且属敏感面
+]);
+
+export function shouldTraceModel(type) {
+  const t = String(type || "");
+  if (!t) return false;
+  if (TRACE_DENY.has(t)) return false;
+  if (t.startsWith("upstream-probe")) return false; // 探针类（probe/probe-error/probe-skip）
+  return true;
+}
+
+function safeText(value, n = 140) {
+  return String(value ?? "")
+    .replace(/([?&](?:token|refreshToken|access[_-]?token|refresh[_-]?token|api[_-]?key|apikey|key|secret|password|cookie)=)[^&\s]+/gi, "$1[redacted]")
+    .replace(/(authorization|bearer|access[_-]?token|refresh[_-]?token|cookie|api[-_]?key|password|secret)\s*[:=]\s*[^,; ]+/gi, "$1=[redacted]")
+    .replace(/\s+/g, " ").trim().slice(0, n);
+}
+
+function kv(obj) {
+  return Object.entries(obj).filter(([, v]) => v != null && v !== "").map(([k, v]) => `${k}=${typeof v === "object" ? JSON.stringify(v) : v}`).join(" ");
+}
+
+// 决定类事件的可见字段白名单：只渲染这些标量；事件里的 payload/detail/stages/请求正文等一律不落模型日志。
+// 「加日志」不等于「倒数据」——新增决定字段要显式登记在这里才会出现。
+const DECISION_FIELDS = [
+  "status", "from", "to", "reason", "hook", "applied", "skipped", "plugin",
+  "provider", "before", "after", "models", "skippedFaulty", "limit", "tried",
+  "winPeer", "winTarget", "peer", "peers", "peerLabel", "routeBest", "at",
+  "latencyMs", "ttfMs", "hedged", "hops", "hasShare", "elapsedMs", "threshold",
+  "stallHits", "maxGapMs", "interrupted", "providers", "via",
+  "retry", "step", "max", "delayMs", "upstream", "account", "pick", "cooled", "queued", "nudged", "retries", "waitedMs", "raiseFrom", "raiseTo",
+  "heldOpen", "terminalForm", "requestWaitedMs", "budgetMs", "budgetOut",
+  // 终局判决（ADR-0058）：取证时第一眼要看的三件——判的哪种、命中哪条判据、是否因此没问组员
+  "kind", "marker", "peersSkipped",
+];
+
+function decisionKv(data = {}) {
+  const out = [];
+  for (const k of DECISION_FIELDS) {
+    const v = data[k];
+    if (v == null || v === "") continue;
+    out.push(`${k}=${Array.isArray(v) ? v.join("|") : v}`);
+  }
+  return out.join(" ");
+}
+
+// provider 回显头 → 日志字段（谁上的 / 打哪个站 / 为什么选它 / 是否被冷却 / 是否被判排队）：
+// 单一来源，防各处重复写法漂移；取不到时返回空对象（其它供应商零变化）。
+export function upstreamEcho(res) {
+  const h = res?.headers;
+  if (typeof h?.get !== "function") return {};
+  const pick = (name) => h.get(name) || null;
+  return {
+    upstream: pick("x-mslxdff-upstream"),
+    account: pick("x-mslxdff-workbuddy-uid") || pick("x-mslxdff-qoder-region"),
+    pick: pick("x-mslxdff-qoder-account"),
+    cooled: pick("x-mslxdff-qoder-cooldown"),
+    queued: pick("x-mslxdff-qoder-queued"),
+  };
+}
+
+export function formatModelTrace({ type, reqId, model, data = {}, request = null, totalMs = null } = {}) {
+  const base = { time: fmtShanghaiYMDHMS(new Date()), req: reqId || "-", model: model || "-", stage: type || "-" };
+  let detail = "";
+  // api= 打在 incoming 行最前：门是排查第一眼要的东西，grep "incoming api=" 直接分得开三门；无值时 kv 自然不落该段（不发散成 api=other）
+  if (type === "request") detail = `incoming ${kv({ api: data.api, ...(request || summarizeRequest(data.body)), hops: data.hops, useAuto: data.useAuto })}`;
+  else if (type === "ordered") detail = `route order=${(data.order || []).join(" | ")} hops=${data.hops ?? "-"} fallback=${data.canFallback ? 1 : 0}`;
+  else if (type === "model-try") detail = `target=${data.model || model || "-"} idx=${data.idx ?? "-"} remaining=${data.remaining ?? "-"}`;
+  else if (type === "alias") detail = `rawModel=${data.rawModel || "-"} requested=${data.requested || model || "-"}`;
+  else if (type === "exhausted-local" || type === "exhausted-all") detail = `last=${data.lastModel || model || "-"} status=${data.lastStatus ?? "-"} order=${(data.order || []).join(" | ")}`;
+  else if (type === "upstream-try") detail = `target=${data.model || model || "-"} attempt=${data.attempt ?? "-"} payload=${kv(data.payload || {})}`;
+  else if (type === "upstream-done" || type === "upstream-error") detail = `upstream status=${data.status ?? "-"} timing=${data.timing?.totalMs ?? "-"}ms${data.upstream ? ` via=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}${data.cooled ? ` cooled=${data.cooled}` : ""}${data.queued ? ` queued=1` : ""} ${safeText(data.message || data.error || "")}`;
+  else if (type === "empty-turn-retry") detail = `空转重试：retry ${data.retry ?? "-"}/${data.max ?? "-"}${data.step != null ? ` step=${data.step}` : ""} delay=${data.delayMs ?? "-"}ms${data.waitedMs ? ` waited=${data.waitedMs}ms` : ""}${data.requestWaitedMs != null ? ` 请求级累计=${data.requestWaitedMs}ms/预算${data.budgetMs ?? "-"}ms` : ""}${data.raiseTo ? ` max_tokens ${data.raiseFrom}→${data.raiseTo}` : ""}${data.upstream ? ` after=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}${data.nudged ? ` nudged=1` : ""} reason=empty turn${data.reason ? ` :: ${safeText(data.reason, 120)}` : ""}`;
+  else if (type === "empty-turn-recovered") detail = `空转重试成功：第 ${data.retries ?? "-"}/${data.max ?? "-"} 次重拉拿到真实输出，本轮共白等 ${data.waitedMs ?? "-"}ms${data.upstream ? ` via=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}`;
+  else if (type === "empty-turn-exhausted") detail = `空转重试已用尽：重试 ${data.retries ?? 0}/${data.max ?? "-"} 次仍零输出（累计等待 ${data.waitedMs ?? 0}ms${data.budgetOut ? `，请求级等待预算 ${data.budgetMs ?? "-"}ms 用尽` : ""}），交回换候选${data.heldOpen ? "（响应仍开着，换过去的正文还能送达）" : "（响应已封口，重拉只为省额度）"}${data.upstream ? ` after=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}${data.reason ? ` :: ${safeText(data.reason, 120)}` : ""}`;
+  else if (type === "peer-request" || type === "peer-forward" || type === "peer-attempt" || type === "peer-error") detail = `peer=${hostPort(data.peer)}${data.name ? ` name=${data.name}` : ""} status=${data.status ?? "-"} latency=${data.latencyMs ?? "-"}ms payload=${kv(data.payload || {})} ${safeText(data.message || data.error || "")}`;
+  else if (type === "relay-done") {
+    const d = data.detail || {};
+    const u = d.usage || {};
+    // 缺 usage 时显式标注来源，别与「真的用了 0 token」混成同一个 `-`（BUG-13：经组员中继的响应
+    // 少数轮次拿不到 usage，成本报表会静默偏低，读日志的人也会误判成"这轮没消耗"）。
+    const noUsage = u.prompt_tokens == null && u.completion_tokens == null;
+    const usageTxt = noUsage
+      ? `usage(none${data.via === "peer" ? ":peer 未回" : ""})`
+      : `usage(prompt=${u.prompt_tokens ?? "-"}/completion=${u.completion_tokens ?? "-"})`;
+    detail = `upstream_response status=${data.status ?? "-"} finish=${d.sawFinishReason || "-"} chunks=${d.receivedChunks ?? "-"} bytes=${d.receivedBytes ?? "-"} tools=${d.toolCalls ?? 0} chars=${d.chars ?? 0} ${usageTxt} elapsed=${data.totalMs ?? "-"}ms${data.upstream ? ` upstream=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}`;
+  } else if (type === "result" || type === "client-response") {
+    detail = `client status=${data.status ?? "-"} via=${data.via || "-"} actual=${data.actual || model || "-"} fallback=${data.fallback?.fallback ? 1 : 0} total=${totalMs ?? data.durationMs ?? "-"}ms${data.upstream ? ` upstream=${data.upstream}` : ""}${data.account ? ` account=${data.account}` : ""}${data.pick ? ` pick=${data.pick}` : ""}`;
+  } else detail = safeText(`${decisionKv(data)} ${data.message || data.error || ""}`, 200);
+  return `[${base.time}] [req=${base.req}] [model=${base.model}] [stage=${base.stage}] ${detail}`;
+}
+
+/**
+ * 按字节从尾部回退（不截半行）。行有长有短，固定行数 = 固定容量 ⇒ 一次裁剪就是断崖。
+ * 与 logs.js 的 keepTailByBytes 同口径（那边不导出，这里留一份最小实现，避免为一行日志反向依赖）。
+ */
+function keepTailByBytes(text, maxBytes) {
+  const limit = Math.max(1, Math.floor(maxBytes * TRACE_KEEP_RATIO));
+  if (Buffer.byteLength(text) <= limit) return text;
+  const lines = text.split("\n");
+  let bytes = 0;
+  let i = lines.length - 1;
+  for (; i >= 0; i--) {
+    const b = Buffer.byteLength(lines[i]) + 1;
+    if (bytes + b > limit) break;
+    bytes += b;
+  }
+  // 至少保留最后一条：单行超 limit 时否则返回空串 = 清空整个 trace。
+  // 判据是「结果为空」——文件尾有换行时 split 的最后一个元素是空串，`i === lines.length-1` 判不到内容行。
+  const kept = lines.slice(i + 1).join("\n");
+  if (!kept) {
+    const nonEmpty = lines.filter((l) => l !== "");
+    if (nonEmpty.length) return nonEmpty[nonEmpty.length - 1];
+  }
+  return kept;
+}
+
+export function appendModelTrace(model, entry, { file = modelLogFile(model) } = {}) {
+  const line = `${formatModelTrace(entry)}\n`;
+  try {
+    mkdirSync(logDir(), { recursive: true });
+    if (existsSync(file) && statSync(file).size > MAX_TRACE_BYTES) {
+      // 按字节留 3/4，不再固定砍到 100 行（那一次就丢 98% 历史，见 TRACE_KEEP_RATIO 注释）。
+      const kept = keepTailByBytes(readFileSync(file, "utf8"), MAX_TRACE_BYTES);
+      // 末尾补换行：下一次 appendFileSync 才不会把新行接到半截行尾巴上
+      writeFileSync(file, kept.endsWith("\n") ? kept : `${kept}\n`);
+    }
+    appendFileSync(file, line);
+  } catch {
+    // Logging must never affect the chat request.
+  }
+}

package/src/models.js

@@ -3,7 +3,7 @@ const CACHE_TTL_MS = 2 * 60 * 60 * 1000;
 const DEFAULT_REFRESH_MS = 2 * 60 * 60 * 1000;
 import { mkdirSync, writeFileSync } from "node:fs";
 import { dirname } from "node:path";
-import { compatFetch } from "./compat.js";
+import { compatFetch } from "./compat.js";
 
 export function isFreeModel(id) {
   return (typeof id === "string" && id.endsWith("-free")) ||

package/src/opencode-identity.js

@@ -1,72 +1,72 @@
-import crypto from "node:crypto";
-
-// opencode 客户端身份规格单一来源。真值 = 上游 packages/schema/src/identifier.ts：
-// id = <prefix>_ + 12 位 hex 时间段 + 14 位 base62；UA 必须 "opencode/<semver>"。
-// ⚠ 时间段是**倒序**的：value = ~(timestamp*4096 + counter) & 48bit（descending，最新排最前）。
-//   正序时钟不合规——zen 免费层自 2026-10-13 起把低 12 位**反解回 counter** 校验，只接受 [0,63]；
-//   正序 id 反解出 4094 → 403 FreeTierError（现网实测：counter=1/40 → 200，counter=4094 → 403）。
-//   时钟半段不校验：ms*4096 如今已需 53bit，截到 48bit 必回绕——真客户端自己反解出来是 1972 年。
-//   ⇒ 会话 id 可以继续由摘要确定性派生（跨重启稳定），只需把低 12 位强制成合法 counter。
-// 旧判据全部仍有效：形状正则（2026-09-17）、UA ≥ 1.18.0（2026-09-18，低版本 426 UpgradeRequired）、
-//   以及请求体那半道 agent 形状闸（stream + 核心五工具，见 free-lane.js）。
-const ID_CHARS = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-const MASK48 = 0xffffffffffffn;
-const MASK12 = 0xfffn;
-// 同毫秒内第 N 个 id：免费层只认反解后 ∈[0,63]，越界必 403 ⇒ 到 63 回头，不无限自增
-const COUNTER_MAX = 63;
-const DEFAULT_OPENCODE_UA = "opencode/1.18.31";
-let idLastTs = 0;
-let idCounter = 0;
-
-/** 合规时间段：倒序的 timestamp*4096+counter 截 48bit，恒 12 位小写 hex */
-export function identifierTimeHex(timestamp = Date.now(), counter = 1) {
-  return ((~(BigInt(timestamp) * 0x1000n + BigInt(counter))) & MASK48).toString(16).padStart(12, "0");
-}
-
-/**
- * 从 12 位 hex 时间段反解 per-ms 计数——与免费层闸门同一判据。
- * 值域 0..4095，闸门只放行 ≤63；形状不符返回 NaN（绝不抛，调用方用 Number.isInteger 自检）。
- */
-export function counterFromTimeHex(hex) {
-  if (typeof hex !== "string" || !/^[0-9a-f]{12}$/i.test(hex)) return Number.NaN;
-  return Number((~BigInt("0x" + hex)) & MASK12);
-}
-
-function base62From(bytes) {
-  let out = "";
-  for (const b of bytes) out += ID_CHARS[b % 62];
-  return out;
-}
-
-export function opencodeIdTail(timestamp = Date.now()) {
-  if (timestamp !== idLastTs) {
-    idLastTs = timestamp;
-    idCounter = 0;
-  }
-  idCounter = (idCounter % COUNTER_MAX) + 1;
-  return identifierTimeHex(timestamp, idCounter) + base62From(crypto.randomBytes(14));
-}
-
-// 摘要派生：同种子恒同 id（会话亲和——上游按 session 归组免费额度与 prompt cache，改成现取
-// 时钟会让同一会话每请求换号，绝不）。时钟半段照摘要填（不校验），低 12 位强制 counter=1。
-export function digestIdTail(digest) {
-  const time = ((BigInt("0x" + digest.subarray(0, 6).toString("hex")) & ~MASK12) | ((~1n) & MASK12)) & MASK48;
-  return time.toString(16).padStart(12, "0") + base62From(digest.subarray(6, 20));
-}
-
-export function genId(prefix) {
-  return `${prefix}${opencodeIdTail()}`;
-}
-
-export function opencodeUa() {
-  return process.env.MSLXDFF_OPENCODE_UA || DEFAULT_OPENCODE_UA;
-}
-
-// 旁路（-chat curl 工具 / bench 直连）共用同一身份，避免各自伪造头再被 403
-export function opencodeClientIdentity() {
-  return {
-    "User-Agent": opencodeUa(),
-    "x-opencode-session": genId("ses_"),
-    "x-opencode-request": genId("msg_"),
-  };
-}
+import crypto from "node:crypto";
+
+// opencode 客户端身份规格单一来源。真值 = 上游 packages/schema/src/identifier.ts：
+// id = <prefix>_ + 12 位 hex 时间段 + 14 位 base62；UA 必须 "opencode/<semver>"。
+// ⚠ 时间段是**倒序**的：value = ~(timestamp*4096 + counter) & 48bit（descending，最新排最前）。
+//   正序时钟不合规——zen 免费层自 2026-10-13 起把低 12 位**反解回 counter** 校验，只接受 [0,63]；
+//   正序 id 反解出 4094 → 403 FreeTierError（现网实测：counter=1/40 → 200，counter=4094 → 403）。
+//   时钟半段不校验：ms*4096 如今已需 53bit，截到 48bit 必回绕——真客户端自己反解出来是 1972 年。
+//   ⇒ 会话 id 可以继续由摘要确定性派生（跨重启稳定），只需把低 12 位强制成合法 counter。
+// 旧判据全部仍有效：形状正则（2026-09-17）、UA ≥ 1.18.0（2026-09-18，低版本 426 UpgradeRequired）、
+//   以及请求体那半道 agent 形状闸（stream + 核心五工具，见 free-lane.js）。
+const ID_CHARS = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+const MASK48 = 0xffffffffffffn;
+const MASK12 = 0xfffn;
+// 同毫秒内第 N 个 id：免费层只认反解后 ∈[0,63]，越界必 403 ⇒ 到 63 回头，不无限自增
+const COUNTER_MAX = 63;
+const DEFAULT_OPENCODE_UA = "opencode/1.18.31";
+let idLastTs = 0;
+let idCounter = 0;
+
+/** 合规时间段：倒序的 timestamp*4096+counter 截 48bit，恒 12 位小写 hex */
+export function identifierTimeHex(timestamp = Date.now(), counter = 1) {
+  return ((~(BigInt(timestamp) * 0x1000n + BigInt(counter))) & MASK48).toString(16).padStart(12, "0");
+}
+
+/**
+ * 从 12 位 hex 时间段反解 per-ms 计数——与免费层闸门同一判据。
+ * 值域 0..4095，闸门只放行 ≤63；形状不符返回 NaN（绝不抛，调用方用 Number.isInteger 自检）。
+ */
+export function counterFromTimeHex(hex) {
+  if (typeof hex !== "string" || !/^[0-9a-f]{12}$/i.test(hex)) return Number.NaN;
+  return Number((~BigInt("0x" + hex)) & MASK12);
+}
+
+function base62From(bytes) {
+  let out = "";
+  for (const b of bytes) out += ID_CHARS[b % 62];
+  return out;
+}
+
+export function opencodeIdTail(timestamp = Date.now()) {
+  if (timestamp !== idLastTs) {
+    idLastTs = timestamp;
+    idCounter = 0;
+  }
+  idCounter = (idCounter % COUNTER_MAX) + 1;
+  return identifierTimeHex(timestamp, idCounter) + base62From(crypto.randomBytes(14));
+}
+
+// 摘要派生：同种子恒同 id（会话亲和——上游按 session 归组免费额度与 prompt cache，改成现取
+// 时钟会让同一会话每请求换号，绝不）。时钟半段照摘要填（不校验），低 12 位强制 counter=1。
+export function digestIdTail(digest) {
+  const time = ((BigInt("0x" + digest.subarray(0, 6).toString("hex")) & ~MASK12) | ((~1n) & MASK12)) & MASK48;
+  return time.toString(16).padStart(12, "0") + base62From(digest.subarray(6, 20));
+}
+
+export function genId(prefix) {
+  return `${prefix}${opencodeIdTail()}`;
+}
+
+export function opencodeUa() {
+  return process.env.MSLXDFF_OPENCODE_UA || DEFAULT_OPENCODE_UA;
+}
+
+// 旁路（-chat curl 工具 / bench 直连）共用同一身份，避免各自伪造头再被 403
+export function opencodeClientIdentity() {
+  return {
+    "User-Agent": opencodeUa(),
+    "x-opencode-session": genId("ses_"),
+    "x-opencode-request": genId("msg_"),
+  };
+}