mslxdff 0.1.160 → 0.1.162

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (49) hide show
  1. package/package.json +1 -1
  2. package/src/autostart.js +3 -0
  3. package/src/chat-pipeline/empty-turn.js +119 -0
  4. package/src/chat-pipeline/index.js +4 -1
  5. package/src/chat-pipeline/serial-trial.js +53 -19
  6. package/src/cli/commands/daemon.js +4 -4
  7. package/src/cli/commands/model/list-providers.js +1 -1
  8. package/src/cli/commands/provider/globalqwenwork-login.js +126 -0
  9. package/src/cli/commands/provider/index.js +2 -0
  10. package/src/cli/commands/provider/models.js +3 -0
  11. package/src/cli/commands/system.js +2 -2
  12. package/src/cli/format.js +7 -0
  13. package/src/cli/policy.js +2 -2
  14. package/src/cli/provider-row.js +2 -2
  15. package/src/daemon.js +7 -2
  16. package/src/logs.js +14 -1
  17. package/src/model-trace.js +8 -4
  18. package/src/providers/classify.js +1 -1
  19. package/src/providers/dispatcher.js +26 -14
  20. package/src/providers/globalqwenwork/account-store.js +141 -0
  21. package/src/providers/globalqwenwork/constants.js +73 -0
  22. package/src/providers/globalqwenwork/cosy.js +123 -0
  23. package/src/providers/globalqwenwork/crypto.js +220 -0
  24. package/src/providers/globalqwenwork/http.js +22 -0
  25. package/src/providers/globalqwenwork/index.js +331 -0
  26. package/src/providers/globalqwenwork/payload.js +145 -0
  27. package/src/providers/globalqwenwork/rsa.js +56 -0
  28. package/src/providers/globalqwenwork/sse.js +270 -0
  29. package/src/providers/globalqwenwork/stream.js +132 -0
  30. package/src/providers/globalqwenwork/upstream.js +122 -0
  31. package/src/providers/globalqwenwork.js +1 -0
  32. package/src/providers/qoder/chat.js +16 -6
  33. package/src/providers/qoder/index.js +13 -4
  34. package/src/providers/qoder/sse.js +39 -1
  35. package/src/providers/registry.js +8 -0
  36. package/src/providers/zcode/sse.js +19 -2
  37. package/src/routes/chat/exhausted-handler.js +23 -8
  38. package/src/routes/chat/local-handler.js +3 -1
  39. package/src/routes/chat/relay-pipeline.js +22 -10
  40. package/src/routes/helpers.js +20 -4
  41. package/src/routes/stream-hold.js +138 -0
  42. package/src/routes/stream-scan.js +192 -0
  43. package/src/routes/stream.js +97 -104
  44. package/src/runtime/lifecycle-forensics.js +116 -0
  45. package/src/runtime/lifecycle-log.js +23 -0
  46. package/src/runtime/provider-gate.js +4 -3
  47. package/src/runtime/server-lifecycle.js +5 -6
  48. package/src/talk-log.js +226 -0
  49. package/src/timeline.js +5 -2
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "mslxdff",
3
- "version": "0.1.160",
3
+ "version": "0.1.162",
4
4
  "description": "测试项目,请勿使用。",
5
5
  "type": "module",
6
6
  "bin": {
package/src/autostart.js CHANGED
@@ -3,6 +3,8 @@ import { existsSync, mkdirSync, writeFileSync, readFileSync, rmSync } from "node
3
3
  import { homedir, userInfo } from "node:os";
4
4
  import { join, dirname } from "node:path";
5
5
  import { fileURLToPath } from "node:url";
6
+ import { daemonDir, readPid } from "./daemon.js";
7
+ import { writeExitMarker } from "./runtime/lifecycle-forensics.js";
6
8
 
7
9
  const TASK_NAME = "mslxdff";
8
10
  const SERVICE_NAME = "mslxdff";
@@ -203,6 +205,7 @@ async function linuxEnable() {
203
205
  }
204
206
  for (const p of pids) {
205
207
  if (p === process.pid) continue;
208
+ if (p === readPid()) { try { writeExitMarker(daemonDir(), { reason: "autostart-cleanup", prevPid: p, prevVersion: null, byPid: process.pid }); } catch {} }
206
209
  try { process.kill(p, "SIGTERM"); killed++; } catch {}
207
210
  }
208
211
  if (killed) await new Promise((r2) => setTimeout(r2, 400));
@@ -0,0 +1,119 @@
1
+ // 空转(EMPTY_MODEL_RESPONSE:上游 200 但零正文零工具调用)重试策略:阶梯等待 + 抬额度 + 末次追问。
2
+ // 本模块只做「配置解析 + 载荷改写」的纯决策,不碰 HTTP 与下游响应;执行编排在 serial-trial.js。
3
+ // 拆出原因:serial-trial.js 承载串行换人/hedge/peer/broadband 编排,再加政策面即破 20KB 硬门。
4
+ import { extractRetryAfterSec } from "../routes/stream-scan.js"; // 限流窗口解析:与 relay 的「错误包络暂扣」同源,不复制第二份正则
5
+
6
+ // 空转重试阶梯(用户定:网络质量差 → 2s→8s→30s;末档对齐实测 retryAfterSeconds=30 的限流窗口)。
7
+ // MSLXDFF_EMPTY_TURN_RETRY_STEPS: JSON 数组(ms)。未设 = 用默认阶梯 [2000,8000,30000];显式 "[]" = 逃生阀,回旧式固定延迟。
8
+ // MSLXDFF_EMPTY_TURN_RETRIES: 重试上限。未设 = 阶梯档数(默认 3);0 = 关闭重试(空转直接换候选/终结)。
9
+ // 非法值(语法错/非正数)→ 每进程告警一次并按默认阶梯跑:绝不因配置手滑把重试整体关掉。
10
+ export const DEFAULT_EMPTY_TURN_STEPS = Object.freeze([2000, 8000, 30000]);
11
+
12
+ let stepsWarned = false;
13
+ function warnBadSteps(raw) {
14
+ if (stepsWarned) return;
15
+ stepsWarned = true;
16
+ try { console.warn(`[mslxdff] MSLXDFF_EMPTY_TURN_RETRY_STEPS 非法(${String(raw).slice(0, 60)}) → 改用默认阶梯 [${DEFAULT_EMPTY_TURN_STEPS.join(",")}]`); } catch {}
17
+ }
18
+
19
+ export function emptyRetryCfg() {
20
+ // 空串/纯空白 = 未设(.env 里 `MSLXDFF_EMPTY_TURN_RETRIES=` 是常见手滑,Number("")===0 会把整条恢复静默清零)
21
+ const rawMax = String(process.env.MSLXDFF_EMPTY_TURN_RETRIES ?? "").trim();
22
+ const r = rawMax === "" ? NaN : Number(rawMax);
23
+ const resolveMax = (fallback) => (Number.isInteger(r) && r >= 0 ? r : fallback);
24
+ const raw = process.env.MSLXDFF_EMPTY_TURN_RETRY_STEPS;
25
+ const s = raw == null ? "" : String(raw).trim();
26
+ // 逃生阀:显式空数组 → 旧式固定延迟(MSLXDFF_EMPTY_TURN_RETRY_DELAY_MS,默认 2s、最多 2 次)
27
+ if (s === "[]") {
28
+ const d = Number(process.env.MSLXDFF_EMPTY_TURN_RETRY_DELAY_MS);
29
+ return { max: resolveMax(2), delayMs: Number.isFinite(d) && d >= 0 ? d : 2000, steps: null };
30
+ }
31
+ let steps = DEFAULT_EMPTY_TURN_STEPS;
32
+ if (s !== "") {
33
+ let parsed = null;
34
+ try { parsed = JSON.parse(s); } catch { warnBadSteps(s); }
35
+ if (Array.isArray(parsed) && parsed.length && parsed.every((v) => Number.isFinite(Number(v)) && Number(v) > 0)) {
36
+ steps = parsed.map(Number);
37
+ } else if (parsed !== null) {
38
+ warnBadSteps(s); // 语法过了但形状不对(非数组/空数组元素/含非正数)
39
+ }
40
+ }
41
+ return { max: resolveMax(steps.length), steps };
42
+ }
43
+
44
+ // retryAfterSeconds 来自上游可控文本,不设顶就会被拿捏:`retryAfterSeconds: 3600` → 单档挂 1 小时;
45
+ // 且 ms 数超过 setTimeout 上限时 Node 会钳成 1ms 并告警 → 对刚限流的上游连开三枪,正是要防的事。
46
+ // 顶只约束「上游能 imposed 多少」;用户自己配的更大阶梯档照办(那是显式意愿),仅保留 setTimeout 硬安全顶。
47
+ const SET_TIMEOUT_CEILING_MS = 2_147_483_647;
48
+ export function emptyTurnMaxWaitMs() {
49
+ const v = Number(process.env.MSLXDFF_EMPTY_TURN_MAX_WAIT_MS);
50
+ return Number.isFinite(v) && v > 0 ? Math.min(v, SET_TIMEOUT_CEILING_MS) : 60_000;
51
+ }
52
+
53
+ // 本次重试实际该等多久:按档取阶梯(超出档数循环复用),再按上游 retryAfterSeconds 抬到不低于该窗口。
54
+ // 限流没到点就重拉等于白烧一发(实测 2s 重拉撞 retryAfterSeconds=30 → 连 2/2 全灭,正是本次改动起因)。
55
+ export function computeNextDelay(stepIndex, steps, lastErr) {
56
+ const base = Number(process.env.MSLXDFF_EMPTY_TURN_RETRY_DELAY_MS);
57
+ let delayMs = Number.isFinite(base) && base >= 0 ? base : 2000;
58
+ if (Array.isArray(steps) && steps.length && Number.isInteger(stepIndex) && stepIndex >= 0) {
59
+ delayMs = steps[stepIndex % steps.length];
60
+ }
61
+ const after = extractRetryAfterSec(lastErr?.message || "");
62
+ if (after != null) delayMs = Math.max(delayMs, Math.min(after * 1000, emptyTurnMaxWaitMs()));
63
+ return Math.min(Math.round(delayMs), SET_TIMEOUT_CEILING_MS);
64
+ }
65
+
66
+ // 空转重试抬额度:思考型模型常把 max_tokens 全花在 reasoning 上(实测 8192 刷满 → 零正文
67
+ // finish_reason=length),同参重拉必然复现 → 重试前把额度翻倍抬到顶,给模型"想完还能说话"的空间。
68
+ // 只加不减、有顶;客户端没设额度时由调用方决定是否兜底发明一次(floor 参数,见 emptyTurnMinRaiseTo;floor=0 回旧口径不发明)。
69
+ // MSLXDFF_EMPTY_TURN_RAISE_TOKENS=0 关闭抬额;默认顶 16384。
70
+ export function emptyRaiseCap() {
71
+ const c = Number(process.env.MSLXDFF_EMPTY_TURN_RAISE_TOKENS);
72
+ return Number.isInteger(c) && c >= 0 ? c : 16384;
73
+ }
74
+
75
+ // 空转重试抬额度(含兜底):现网三类空轮里最多的是「思考吃满 max_tokens、零正文」
76
+ // (events.log:reasoningChars 10290/22652、wroteFrames 1972/2573、finish=length)。客户端设过额度 → 翻倍抬到 cap;
77
+ // 客户端根本没设额度 → 同参重拉必然复现,故用 MIN_RAISE_TO 兜底发明一次上限(=0 关,回旧口径「不替它发明」)。
78
+ export function emptyTurnMinRaiseTo() {
79
+ const v = Number(process.env.MSLXDFF_EMPTY_TURN_MIN_RAISE_TO);
80
+ if (v === 0 || (Number.isFinite(v) && v < 0)) return 0;
81
+ return Number.isInteger(v) && v > 0 ? v : 16384;
82
+ }
83
+
84
+ export function withRaisedMaxTokens(payload, cap, floor = 0) {
85
+ if (!payload || typeof payload !== "object") return payload;
86
+ for (const key of ["max_tokens", "max_completion_tokens"]) {
87
+ const cur = Number(payload[key]);
88
+ if (Number.isFinite(cur) && cur > 0) {
89
+ if (!cap || cur >= cap) return payload; // 已到顶:只剩同参重拉(赌偶发空轮)
90
+ return { ...payload, [key]: Math.min(cap, cur * 2) };
91
+ }
92
+ }
93
+ if (floor) return { ...payload, max_tokens: Math.min(floor, cap || floor) }; // 客户端没设额度:兜底发明一次
94
+ return payload;
95
+ }
96
+ // 空转追问:最后一次空转重试前追加一句 user 兜底(content 空轮大概率是模型“收尾犹豫”,
97
+ // 追问把它逼出一句“任务完成了”即可终结;前面重试仍原样重拉,防改写正常对话)。
98
+ // MSLXDFF_EMPTY_NUDGE=0 关闭;MSLXDFF_EMPTY_NUDGE_TEXT 改话术。
99
+ export function emptyNudgeCfg() {
100
+ const off = process.env.MSLXDFF_EMPTY_NUDGE;
101
+ if (off === "0" || String(off || "").toLowerCase() === "off" || String(off || "").toLowerCase() === "false") return { enabled: false, text: "" };
102
+ const custom = String(process.env.MSLXDFF_EMPTY_NUDGE_TEXT || "").trim();
103
+ return { enabled: true, text: custom || "如果你完成任务了,请简短的说:任务完成了。" };
104
+ }
105
+ export function withEmptyNudge(payload, text) {
106
+ const t = String(text || "").trim();
107
+ if (!t || !payload || !Array.isArray(payload.messages) || !payload.messages.length) return payload;
108
+ const last = payload.messages[payload.messages.length - 1];
109
+ if (last && last.role === "user" && String(last.content || "").includes(t)) return payload;
110
+ return { ...payload, messages: [...payload.messages, { role: "user", content: t }] };
111
+ }
112
+
113
+ // 请求级空轮等待预算:阶梯是「每发候选」独立算的,多候选串起来最坏 = N×(2+8+30),
114
+ // 客户端只是在等一个永远空白的回复。预算封顶「这一次请求最多白等多久」,超限直接收场。
115
+ // MSLXDFF_EMPTY_TURN_BUDGET_MS 覆盖(<=0 或未设 = 默认 45s:够装下一整个阶梯 + 一发换候选)。
116
+ export function emptyTurnBudgetMs() {
117
+ const v = Number(process.env.MSLXDFF_EMPTY_TURN_BUDGET_MS);
118
+ return Number.isFinite(v) && v > 0 ? v : 45_000;
119
+ }
@@ -27,7 +27,7 @@ export function createChatPipeline({ upstream, auto, logs, peers, groups, bus, t
27
27
  mark("parsed");
28
28
  if (aliasInfo) { try { res?.setHeader?.("x-mslxdff-alias", aliasInfo); } catch {} }
29
29
  // mslxdff/ 前缀或 alias 命中时,把 body.model 改写为还原后的模型(与原 gateway 语义一致)
30
- const timeline = { direct: [], peers: [], retries: 0, result: null };
30
+ const timeline = { direct: [], peers: [], retries: 0, retryResult: null, result: null };
31
31
  if (aliasInfo && req?.body && req.body.model !== requested) {
32
32
  req.body = { ...req.body, model: requested };
33
33
  }
@@ -53,6 +53,9 @@ export function createChatPipeline({ upstream, auto, logs, peers, groups, bus, t
53
53
  else timeline.peers.push({ peer: entry.peer, ok: false, status: entry.status, message: entry.message || entry.error || "" });
54
54
  }
55
55
  if (type === "empty-turn-retry") timeline.retries += 1;
56
+ // 空转重试必须交代结局:只记次数的话,"重试后救回来"和"重试到死"在日志里长得一模一样
57
+ if (type === "empty-turn-recovered") timeline.retryResult = "ok";
58
+ if (type === "empty-turn-exhausted") timeline.retryResult = timeline.retryResult || "lost";
56
59
  if (type === "result" || (type === "client-response" && !timeline.result)) {
57
60
  timeline.result = { status: entry.status, detail: entry.detail || null };
58
61
  try { logs?.appendTimeline?.(formatTimeline({ reqId, model: entry.model || requested, ...timeline, totalMs: Date.now() - startedAt })); } catch {}
@@ -12,20 +12,10 @@ import { handleExhaustedLocal, handleExhaustedAll } from "../routes/chat/exhaust
12
12
  import { shouldUseGroupForModel, isHardLocalOnly, isKeyProviderDirectOnly } from "../state/schemas/use-group.js";
13
13
  import { summarizeRequest, upstreamEcho } from "../model-trace.js";
14
14
  import { isEmptyTurnError } from "../routes/chat/relay-pipeline.js";
15
+ import { emptyRetryCfg, emptyRaiseCap, withRaisedMaxTokens, emptyNudgeCfg, withEmptyNudge, computeNextDelay, emptyTurnBudgetMs, emptyTurnMinRaiseTo } from "./empty-turn.js";
15
16
 
16
17
  const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
17
18
 
18
- // 空转重试档位(每次请求读取 env,单测可覆盖):默认同模型最多重试 2 次、间隔 1s;
19
- // MSLXDFF_EMPTY_TURN_RETRIES=0 关闭(回旧行为:空转直接换候选/终结)。
20
- function emptyRetryCfg() {
21
- const r = Number(process.env.MSLXDFF_EMPTY_TURN_RETRIES);
22
- const d = Number(process.env.MSLXDFF_EMPTY_TURN_RETRY_DELAY_MS);
23
- return {
24
- max: Number.isInteger(r) && r >= 0 ? r : 2,
25
- delayMs: Number.isFinite(d) && d >= 0 ? d : 1000,
26
- };
27
- }
28
-
29
19
  function groupSkipReason(model) {
30
20
  if (isHardLocalOnly(model)) return "provider local-only(禁组员,仅本机直连)";
31
21
  if (isKeyProviderDirectOnly(model)) return "key provider default direct(仅本机直连,MSLXDFF_USE_GROUP_KEYS=1 可开组员)";
@@ -69,6 +59,8 @@ export async function runSerialTrial(ctx, deps = {}) {
69
59
  }
70
60
 
71
61
  let lastErr = viaRouteLastErr;
62
+ // 请求级空轮等待累计(跨候选):阶梯是「每发候选」各算的,不设顶就会 N×(2+8+30) 把客户端吊在门外
63
+ let requestWaitedMs = 0;
72
64
  candidate: for (let idx = 0; idx < order.length; idx++) {
73
65
  const model = order[idx];
74
66
  handlerCtx.model = model;
@@ -94,10 +86,13 @@ export async function runSerialTrial(ctx, deps = {}) {
94
86
  // reqId = 本次客户端请求的身份:供应商据此"同请求粘号"(重试不换号,只有 401/403/429/5xx 冷却才换)
95
87
  if (reqId) chatOpts.reqId = reqId;
96
88
  const chatOptsArg = Object.keys(chatOpts).length ? chatOpts : undefined;
97
- // 空转 200(模型无输出)同模型暂停重试:默认 2 次、间隔 1s;仅 EMPTY_MODEL_RESPONSE,
89
+ // 空转 200(模型无输出)同模型暂停重试:默认 3 次、阶梯 [2s,8s,30s];仅 EMPTY_MODEL_RESPONSE,
98
90
  // 429/403/500 与 fetch 异常走原有切号/failover(防烧额度)。MSLXDFF_EMPTY_TURN_RETRIES=0 关闭。
99
91
  const emptyCfg = emptyRetryCfg();
92
+ const raiseCap = emptyRaiseCap();
93
+ const nudgeCfg = emptyNudgeCfg();
100
94
  let emptyRetried = 0;
95
+ let emptyWaitedMs = 0;
101
96
  for (;;) {
102
97
  const tUp = performance.now();
103
98
  evt("upstream-try", { reqId, model, attempt: idx + 1, emptyRetry: emptyRetried, payload: summarizeRequest(forwarded) });
@@ -169,13 +164,52 @@ export async function runSerialTrial(ctx, deps = {}) {
169
164
  }
170
165
  if (upRes) {
171
166
  const lr = await localRelay({ upRes, model, body, order, idx, lastErr, requested, useAuto, lockModel, auto, handlerCtx, evt, logCall, logError, mark, perf0, attemptStartMs: tUp, stages, startedAt, plugins, res });
172
- if (lr.handled) return { done: true };
173
- if (lr.lastErr && isEmptyTurnError(lr.lastErr) && emptyRetried < emptyCfg.max) {
174
- emptyRetried++;
175
- // 带上"刚空转的是哪个号/哪个站":切号是重试驱动的,日志必须能自证
176
- evt("empty-turn-retry", { reqId, model, retry: emptyRetried, max: emptyCfg.max, delayMs: emptyCfg.delayMs, ...upstreamEcho(upRes) });
177
- await sleep(emptyCfg.delayMs);
178
- continue;
167
+ if (lr.handled) {
168
+ // 空转重试后真拿到输出 = 降级但成功(WARN 语义):必须交代"第几次救回来的、白等了多久",
169
+ // 否则用户只看到"这一发变慢了",无从判断是重试在兜底还是上游真的死了。
170
+ // 「救回」必须有送达证据:handled:true 也可能来自下游已断开/零输出路径(见 ADR-0043),
171
+ // 只认 relay 报上来的 wrotePayload —— 没有正文到下游就不许记成功,否则日志在骗排障的人。
172
+ if (emptyRetried > 0 && lr.wrotePayload === true) {
173
+ evt("empty-turn-recovered", { reqId, model, retries: emptyRetried, max: emptyCfg.max, waitedMs: emptyWaitedMs, ...upstreamEcho(upRes) });
174
+ }
175
+ return { done: true };
176
+ }
177
+ if (lr.lastErr && isEmptyTurnError(lr.lastErr)) {
178
+ // 空转判据原文(finish_reason / 零正文 / 上游错误摘要)随事件落盘,排障不必再翻第二个文件
179
+ const _why = String(lr.lastErr.message || "").replace(/\s+/g, " ").trim().slice(0, 220);
180
+ const budgetMs = emptyTurnBudgetMs();
181
+ const budgetLeftMs = budgetMs - requestWaitedMs;
182
+ if (emptyRetried < emptyCfg.max && budgetLeftMs > 0) {
183
+ // 口径(用户定):正文为空就重试,最多 emptyCfg.max 次——不设"能不能送达"的前提,
184
+ // 大不了两次都空,反正不是无限重试。抬额度只加不减且有顶(默认 16384):
185
+ // 思考刷满 max_tokens 是零正文的主因,同参重拉必然复现,抬一次才算换了打法。
186
+ emptyRetried++;
187
+ const _before = Number(forwarded.max_tokens ?? forwarded.max_completion_tokens) || null;
188
+ // 客户端没设额度也兜底发明一次:现网主流空轮就是「思考吃满 max_tokens、正文为零」,同参重拉必复现
189
+ const _raised = withRaisedMaxTokens(forwarded, raiseCap, emptyTurnMinRaiseTo());
190
+ let _after = null;
191
+ if (_raised !== forwarded) { forwarded = _raised; _after = Number(forwarded.max_tokens ?? forwarded.max_completion_tokens); }
192
+ const _isLast = emptyRetried >= emptyCfg.max;
193
+ let _nudged = false;
194
+ if (nudgeCfg.enabled && _isLast) {
195
+ const _next = withEmptyNudge(forwarded, nudgeCfg.text);
196
+ if (_next !== forwarded) { forwarded = _next; _nudged = true; }
197
+ }
198
+ // 事件在 sleep **之前**发:对着日志能立刻看到"正在暂停 Nms 重拉",而不是等结果
199
+ // 带上"刚空转的是哪个号/哪个站":切号是重试驱动的,日志必须能自证
200
+ const stepIdx = emptyCfg.steps ? (emptyRetried - 1) % emptyCfg.steps.length : 0;
201
+ const rawDelayMs = computeNextDelay(emptyRetried - 1, emptyCfg.steps, lr.lastErr);
202
+ const delayMs = Math.min(rawDelayMs, budgetLeftMs); // 末次等待不越过请求级预算
203
+ emptyWaitedMs += delayMs;
204
+ requestWaitedMs += delayMs;
205
+ evt("empty-turn-retry", { reqId, model, retry: emptyRetried, step: stepIdx, max: emptyCfg.max, delayMs, waitedMs: emptyWaitedMs, requestWaitedMs, budgetMs, nudged: _nudged ? 1 : undefined, raiseFrom: _after != null ? _before : undefined, raiseTo: _after ?? undefined, reason: _why, ...upstreamEcho(upRes) });
206
+ await sleep(delayMs);
207
+ continue;
208
+ }
209
+ // 次数用尽(或被 MSLXDFF_EMPTY_TURN_RETRIES=0 关掉)→ 记一行"不再重试"再交回 failover
210
+ const _budgetOut = emptyRetried < emptyCfg.max && emptyTurnBudgetMs() - requestWaitedMs <= 0;
211
+ evt("empty-turn-exhausted", { reqId, model, retries: emptyRetried, max: emptyCfg.max, waitedMs: emptyWaitedMs, requestWaitedMs, budgetMs: emptyTurnBudgetMs(), budgetOut: _budgetOut ? 1 : undefined, reason: _why, ...upstreamEcho(upRes) });
212
+ try { logError(model, 502, `空转重试 ${emptyRetried}/${emptyCfg.max} 后仍无输出${emptyCfg.max === 0 ? "(MSLXDFF_EMPTY_TURN_RETRIES=0 已关闭重试)" : _budgetOut ? `(请求级等待预算 ${emptyTurnBudgetMs()}ms 用尽)` : ""}:${_why}`); } catch {}
179
213
  }
180
214
  if (lr.lastErr) { lastErr = lr.lastErr; continue candidate; }
181
215
  return { done: true };
@@ -23,11 +23,11 @@ export async function handleRestart(args, VERSION) {
23
23
  const alive = pid ? isPidAlive(pid) : false;
24
24
  if (alive) {
25
25
  console.log(`restarting daemon (pid ${pid})...`);
26
- stopDaemon();
26
+ stopDaemon({ reason: "restart" });
27
27
  await new Promise((r) => setTimeout(r, 300));
28
28
  } else if (pid) {
29
29
  console.log(`daemon pid ${pid} is stale (not running) — starting fresh...`);
30
- try { stopDaemon(); } catch {}
30
+ try { stopDaemon({ reason: "restart" }); } catch {}
31
31
  } else {
32
32
  console.log(`daemon not running — starting...`);
33
33
  }
@@ -58,7 +58,7 @@ export async function handlePort(args) {
58
58
  setPort(port);
59
59
  const daemon = readPid();
60
60
  if (daemon) {
61
- stopDaemon();
61
+ stopDaemon({ reason: "restart" });
62
62
  startDaemon(["-port", String(port)]);
63
63
  await waitForHealth(port, 4000);
64
64
  console.log(`mslxdff restarted on port ${port} (pid ${readPid()})`);
@@ -108,7 +108,7 @@ export async function handleDebug(args) {
108
108
  }
109
109
  } catch {}
110
110
  }
111
- const { stopped, pid } = stopDaemon();
111
+ const { stopped, pid } = stopDaemon({ reason: "debug-takeover" });
112
112
  if (stopped) console.log(`[debug] stopped background daemon (pid ${pid})`);
113
113
  // 等旧 daemon 真正退出再抢端口(Windows 端口释放有延迟,否则 EADDRINUSE 会让 debug 立即崩)
114
114
  if (stopped && pid) {
@@ -10,7 +10,7 @@ export async function renderOtherProviders({ pickedIds, ids, fullAliases }) {
10
10
  try { _la2(); } catch {}
11
11
  const configs = loadProviderConfigs();
12
12
  const otherIds = Object.keys(configs).filter((k) => String(k).toLowerCase() !== "opencode");
13
- const order2 = ["workbuddy", "cline", "openrouter", "bai", "qoder", "qwenwork", "traework", "zcode", "codearts"];
13
+ const order2 = ["workbuddy", "cline", "openrouter", "bai", "qoder", "qwenwork", "globalqwenwork", "traework", "zcode", "codearts"];
14
14
  otherIds.sort((a, b) => {
15
15
  const ia = order2.indexOf(a), ib = order2.indexOf(b);
16
16
  if (ia !== -1 || ib !== -1) {
@@ -0,0 +1,126 @@
1
+ // mslxdff -provider globalqwenwork login — 千问办公**国际站**设备授权(PKCE + poll),对齐 qwenwork-login。
2
+ // 与 cn 站的三处实质差异(2026-10-01 现网取证):
3
+ // ① auth 域名 gateway.qwenwork.ai ② client_id cc65e5fc…(用 cn 的会整串 query 被判 not_allowed)
4
+ // ③ 深链 redirect_uri=qwenwork://;三处全部来自 constants.js,本文件不重复硬编码。
5
+ // 授权后落盘 auths/globalqwenwork-<uid>.json + state 双写,默认 allowAny=false 并种 2 个实测模型(ADR-0041)。
6
+ import { compatFetch } from "../../../compat.js";
7
+ import { normalizeProviderId } from "../../../providers/model-id.js";
8
+ import { newDeviceFlow, pollGrant, userInfo, accountContext, expiryUnix } from "../../../providers/globalqwenwork/upstream.js";
9
+
10
+ const POLL_MS = 3000;
11
+ const POLL_TIMEOUT_MS = 10 * 60 * 1000;
12
+ // 国际站实测 qwork 切片恒为这 2 个(价因子 0,仍是额度池供应商 → 保持 allowAny=false 防 auto 烧分)
13
+ const SEED_MODELS = ["qwork-auto", "qwork-advanced"];
14
+
15
+ export async function handleGlobalQwenworkLogin(id, sub, rest = [], deps = {}) {
16
+ if (normalizeProviderId(id) !== "globalqwenwork") return false;
17
+ if (sub !== "login" && sub !== "auth" && sub !== "oauth") return false;
18
+ const fetchImpl = deps.fetchImpl || compatFetch;
19
+ const log = deps.log || ((m) => console.log(m));
20
+
21
+ const flow = await newDeviceFlow();
22
+ log("=".repeat(60));
23
+ log(" 千问办公 · 国际站(QwenWork Global)登录");
24
+ log("=".repeat(60));
25
+ log("");
26
+ log(" 1. 在浏览器打开下面链接,用已登录**国际版**千问办公的账号确认授权");
27
+ log(" (cn 站账号无效:实测 cn token 打 .ai userinfo → 401 INVALID_TOKEN)");
28
+ log(" 2. 授权成功后会自动完成,不要关闭本终端");
29
+ log("");
30
+ log(` ${flow.url}`);
31
+ log("");
32
+ log("🔄 等待你授权(自动轮询,每 3 秒,最多 10 分钟)...");
33
+
34
+ const deadline = Date.now() + POLL_TIMEOUT_MS;
35
+ let grant = null;
36
+ for (;;) {
37
+ if (Date.now() > deadline) {
38
+ console.error("\n❌ 授权超时(10 分钟),请重新运行 mslxdff -provider globalqwenwork login");
39
+ process.exit(1);
40
+ }
41
+ let outcome;
42
+ try {
43
+ outcome = await pollGrant(flow.nonce, flow.verifier, fetchImpl);
44
+ } catch (e) {
45
+ console.error(`\n❌ 轮询失败: ${e.message}`);
46
+ process.exit(1);
47
+ }
48
+ if (!outcome.pending) { grant = outcome.grant; break; }
49
+ try { process.stdout.write("."); } catch {}
50
+ await new Promise((r) => setTimeout(r, POLL_MS));
51
+ }
52
+ const accessToken = grant.token || grant.device_token;
53
+ const refreshToken = grant.refresh_token || "";
54
+ if (!accessToken) {
55
+ console.error("\n❌ 未获取到 token");
56
+ process.exit(1);
57
+ }
58
+ log(`\n✅ 授权成功!token=${String(accessToken).slice(0, 12)}...`);
59
+ if (refreshToken) log(` refreshToken=${String(refreshToken).slice(0, 12)}...`);
60
+
61
+ let uid = String(grant.user_id || "");
62
+ let name = String(grant.user_name || "");
63
+ let email = "";
64
+ try {
65
+ const info = await userInfo(accessToken, fetchImpl);
66
+ if (info?.id) uid = String(info.id);
67
+ if (info?.name) name = String(info.name);
68
+ if (info?.email) email = String(info.email);
69
+ log(` UID: ${uid} 昵称: ${name || "(空)"} 邮箱: ${email || "(空)"}`);
70
+ } catch (e) {
71
+ log(` ⚠️ userinfo 失败: ${e.message}(用 grant 兜底)`);
72
+ if (!uid) uid = String(accessToken).slice(0, 8);
73
+ }
74
+ if (!uid) {
75
+ console.error("❌ 无法获取用户标识");
76
+ process.exit(1);
77
+ }
78
+
79
+ const { saveGlobalQwenworkAccount } = await import("../../../providers/globalqwenwork/account-store.js");
80
+ const save = deps.saveAccount || saveGlobalQwenworkAccount;
81
+ const saved = await save({
82
+ uid,
83
+ accessToken,
84
+ refreshToken,
85
+ expiresAt: expiryUnix(grant),
86
+ name,
87
+ email,
88
+ });
89
+
90
+ log("");
91
+ log("=".repeat(60));
92
+ log("✅ 登录完成!");
93
+ log(` UID: ${uid}`);
94
+ log(` 账号数: ${saved.accounts}${saved.updated ? "(旧号凭证已更新)" : ""}`);
95
+ log(` 数据文件: ${saved.file}`);
96
+ log("=".repeat(60));
97
+ // 安全默认:allowAny 关 + 只放 2 个实测模型(额度池供应商,与 qwenwork/qoder 惯例一致,见 ADR-0041)
98
+ try {
99
+ const { saveProviderAllowAnyModels, saveProviderAllowedModels } = await import("../../../state.js");
100
+ saveProviderAllowAnyModels("globalqwenwork", false);
101
+ saveProviderAllowedModels("globalqwenwork", SEED_MODELS);
102
+ log(` allowAnyModels=off + allowlist 已种 ${SEED_MODELS.join("/")}(额度池供应商,防 auto 烧分)`);
103
+ } catch {}
104
+ // 额度展示:国际站免费档 pid 实测为 subscription-sgp-free(新加坡池,与 cn 积分池独立)
105
+ try {
106
+ const ctx = await accountContext(accessToken, fetchImpl);
107
+ const plan = ctx?.plan || {};
108
+ const quota = ctx?.quota || {};
109
+ if (plan?.name || quota?.remaining !== undefined) {
110
+ log(` 套餐: ${plan?.name || "-"} (${plan?.pid || "-"}) 积分剩余: ${quota?.remaining ?? "-"}`);
111
+ }
112
+ } catch {
113
+ log(" 额度查询失败,不影响使用");
114
+ }
115
+
116
+ log("下一步:");
117
+ log(" mslxdff -provider globalqwenwork models 列模型列表");
118
+ log(" mslxdff -restart(用户终端跑) 重启网关使新账号生效");
119
+ log(` 然后 curl http://localhost:8989/v1/chat/completions -d '{"model":"globalqwenwork/qwork-auto","messages":[{"role":"user","content":"hi"}]}'`);
120
+ log("");
121
+ log("多账号:重复 `mslxdff -provider globalqwenwork login` 追加,每号独立 uid,对话自动轮换");
122
+ log("(与 cn 站 qwenwork 凭证不互通,两站可在同一 state.json 共存,账号文件前缀互不吞)");
123
+
124
+ if (!deps.noExit) process.exit(0);
125
+ return true;
126
+ }
@@ -83,6 +83,8 @@ const { handleQoderLogin } = await import("./qoder-login.js");
83
83
  if (await handleQoderCheckin(id, sub, rest)) return true;
84
84
  const { handleQwenworkLogin } = await import("./qwenwork-login.js");
85
85
  if (await handleQwenworkLogin(id, sub, rest)) return true;
86
+ const { handleGlobalQwenworkLogin } = await import("./globalqwenwork-login.js");
87
+ if (await handleGlobalQwenworkLogin(id, sub, rest)) return true;
86
88
  const { handleZcodeLogin } = await import("./zcode-login.js");
87
89
  if (await handleZcodeLogin(id, sub, rest)) return true;
88
90
  const { handleZcodeQuota } = await import("./zcode-quota.js");
@@ -53,6 +53,9 @@ export async function handleProviderModels(id, sub, args, rest) {
53
53
  } else if (id === "qwenwork") {
54
54
  const { createQwenworkProvider } = await import("../../../providers/qwenwork/index.js");
55
55
  provider = createQwenworkProvider({ id, apiKeys: keys, file: defaultStateFile() });
56
+ } else if (id === "globalqwenwork") {
57
+ const { createGlobalQwenworkProvider } = await import("../../../providers/globalqwenwork/index.js");
58
+ provider = createGlobalQwenworkProvider({ id, apiKeys: keys, file: defaultStateFile() });
56
59
  } else if (id === "zcode") {
57
60
  const { createZcodeProvider } = await import("../../../providers/zcode/index.js");
58
61
  provider = createZcodeProvider({ id, apiKeys: keys, file: defaultStateFile() });
@@ -58,7 +58,7 @@ export async function handleUpdate(args, VERSION) {
58
58
  const daemon = readPid();
59
59
  if (daemon) {
60
60
  console.log("restarting daemon on the new version…");
61
- stopDaemon();
61
+ stopDaemon({ reason: "restart" });
62
62
  startDaemon([]);
63
63
  await waitForHealth(resolvePort(), 4000);
64
64
  console.log(`restarted (pid ${readPid()})`);
@@ -82,7 +82,7 @@ export async function handleShowToken(args) {
82
82
 
83
83
  export async function handleUninstall(args) {
84
84
  if (!(args.includes("-uninstall") || args.includes("--uninstall"))) return false;
85
- const { stopped, pid } = stopDaemon();
85
+ const { stopped, pid } = stopDaemon({ reason: "uninstall" });
86
86
  if (stopped) console.log(`mslxdff daemon stopped (pid ${pid})`);
87
87
  else console.log("mslxdff daemon not running");
88
88
  const stateFile = defaultStateFile();
package/src/cli/format.js CHANGED
@@ -46,8 +46,15 @@ export function fmtEvent(e) {
46
46
  return `${head} upstream ok ${m(e.model)} HTTP ${e.status}${e.timing ? ` total=${e.timing.totalMs}ms` : ""}`;
47
47
  case "upstream-error":
48
48
  return `${head} upstream err ${m(e.model)} ${e.status ? `HTTP ${e.status}` : "network"}: ${m(e.message)}`;
49
+ case "empty-turn-retry":
50
+ return `${head} 空转重试 模型零输出 model=${m(e.model)} 第 ${e.retry || "?"}/${e.max || "?"} 次,暂停 ${e.delayMs || 0}ms 后同模型重拉${e.raiseTo ? `(max_tokens ${e.raiseFrom}→${e.raiseTo})` : ""}${e.nudged ? "(末次已追问)" : ""} reqId=${e.reqId || ""}`;
51
+ case "empty-turn-recovered":
52
+ return `${head} 空转救回 重试成功 model=${m(e.model)} 第 ${e.retries || "?"}/${e.max || "?"} 次拿到输出,本轮多花 ${e.waitedMs || 0}ms reqId=${e.reqId || ""}`;
53
+ case "empty-turn-exhausted":
54
+ return `${head} 空转耗尽 重试无用 model=${m(e.model)} 已重试 ${e.retries || 0}/${e.max || "?"} 次仍零输出(等 ${e.waitedMs || 0}ms)→ 换候选/报错${e.reason ? ` ${String(e.reason).slice(0, 90)}` : ""} reqId=${e.reqId || ""}`;
49
55
  case "upstream-preheat":
50
56
  if (e.skipped) return `${head} preheat 跳过 (MSLXDFF_PREHEAT disabled)`;
57
+ if (Array.isArray(e.results)) return `${head} preheat 定制 ${e.okCount}/${e.total} ok${(e.results || []).filter((x) => !x.ok).map((x) => x.id).join(",") ? ` fail:${(e.results || []).filter((x) => !x.ok).map((x) => x.id).join(",")}` : ""}`;
51
58
  return `${head} preheat 预热 opencode models ${e.ok ? "ok" : "fail"} ${e.status ? `HTTP ${e.status}` : e.error || ""} ${e.ms ? `${e.ms}ms` : ""}`;
52
59
  case "peer-race-start":
53
60
  return `${head} peer race 开始并发给组员 model=${m(e.model)} peers=${e.peers}`;
package/src/cli/policy.js CHANGED
@@ -60,7 +60,7 @@ export function stopDaemonIfOutdated(VERSION) {
60
60
  const runningVersion = readPidVersion();
61
61
  if (!runningVersion) {
62
62
  console.log(`daemon version unknown — restarting with v${VERSION}...`);
63
- stopDaemon();
63
+ stopDaemon({ reason: "upgrade" });
64
64
  return;
65
65
  }
66
66
  if (runningVersion === VERSION) return;
@@ -70,7 +70,7 @@ export function stopDaemonIfOutdated(VERSION) {
70
70
  return;
71
71
  }
72
72
  console.log(`daemon running v${runningVersion} — upgrading to v${VERSION}, restarting...`);
73
- stopDaemon();
73
+ stopDaemon({ reason: "upgrade" });
74
74
  }
75
75
 
76
76
  export function refreshIntervalMs() {
@@ -40,12 +40,12 @@ export function buildProviderRows({ stateFile, env = process.env } = {}) {
40
40
  let allowed = [];
41
41
  let allowAny = false;
42
42
  try { keys = loadProviderKeys(gid, stateFile ? { file: stateFile } : undefined); } catch {}
43
- try { baseUrl = loadProviderBaseUrl(gid, stateFile ? { file: stateFile } : undefined) || cfg?.baseUrl || (gid === "openrouter" ? "https://openrouter.ai/api/v1" : gid === "workbuddy" ? "https://copilot.tencent.com" : gid === "traework" ? "https://trae-api-cn.mchost.guru" : gid === "qoder" ? "qoder://native" : gid === "qwenwork" ? "qwenwork://native" : gid === "zcode" ? "zcode://native" : ""); } catch { baseUrl = cfg?.baseUrl || ""; }
43
+ try { baseUrl = loadProviderBaseUrl(gid, stateFile ? { file: stateFile } : undefined) || cfg?.baseUrl || (gid === "openrouter" ? "https://openrouter.ai/api/v1" : gid === "workbuddy" ? "https://copilot.tencent.com" : gid === "traework" ? "https://trae-api-cn.mchost.guru" : gid === "qoder" ? "qoder://native" : gid === "qwenwork" ? "qwenwork://native" : gid === "globalqwenwork" ? "globalqwenwork://native" : gid === "zcode" ? "zcode://native" : ""); } catch { baseUrl = cfg?.baseUrl || ""; }
44
44
  try { allowed = loadProviderAllowedModels(gid, stateFile ? { file: stateFile } : undefined); } catch {}
45
45
  try { allowAny = loadProviderAllowAnyModels(gid, stateFile ? { file: stateFile } : undefined); } catch {}
46
46
  // openrouter 特殊:opencode 例外默认 allowAny true,其余默认 false
47
47
  if (gid === "opencode") allowAny = true;
48
- let enabled = Boolean(baseUrl && keys.length) || (gid === "openrouter" && keys.length > 0) || ((gid === "qoder" || gid === "qwenwork" || gid === "traework" || gid === "zcode") && keys.length > 0);
48
+ let enabled = Boolean(baseUrl && keys.length) || (gid === "openrouter" && keys.length > 0) || ((gid === "qoder" || gid === "qwenwork" || gid === "globalqwenwork" || gid === "traework" || gid === "zcode") && keys.length > 0);
49
49
  const auths = gid === "workbuddy" ? (() => { try { return loadProviderAuths(gid, stateFile ? { file: stateFile } : undefined) || []; } catch { return []; } })() : [];
50
50
  let note = "";
51
51
  const isWorkbuddyStub = gid === "workbuddy" && (keys.includes("k-new") || String(baseUrl).includes("127.0.0.1") || (keys.length === 1 && keys[0].length < 20));
package/src/daemon.js CHANGED
@@ -3,6 +3,7 @@ import { appendFileSync, existsSync, mkdirSync, openSync, readFileSync, writeFil
3
3
  import { join, dirname } from "node:path";
4
4
  import os from "node:os";
5
5
  import { fileURLToPath } from "node:url";
6
+ import { writeExitMarker } from "./runtime/lifecycle-forensics.js";
6
7
 
7
8
  export function daemonDir() {
8
9
  return process.env.MSLXDFF_DAEMON_DIR || join(os.homedir(), ".config", "mslxdff");
@@ -73,15 +74,19 @@ export function isPidAlive(pid) {
73
74
  }
74
75
  }
75
76
 
76
- export function stopDaemon() {
77
+ export function stopDaemon({ reason = "stop" } = {}) {
77
78
  const pid = readPid();
78
79
  if (!pid) return { stopped: false, reason: "no pid file" };
80
+ // 杀者牌(验尸官的另一半证据):动手前写、自带死者身份。只给在世者留牌——给尸体留牌
81
+ // 会让下次验尸把"早已静默死亡"误判成"被本次杀死"。
82
+ const victimAlive = isPidAlive(pid);
79
83
  // 杀者留痕(写 daemon.log):Windows 的 SIGTERM 是 TerminateProcess 强杀,被杀的 daemon
80
84
  // 在 JS 层收不到任何事件(无法自记),故由"杀者"记录调用方 pid——
81
85
  // 没有这行 = 非我方所杀(外部 taskkill/任务管理器/崩溃),配合最后一条 heartbeat 定位死亡时刻。
82
86
  try {
83
87
  appendFileSync(logFile(), `[lifecycle] stopDaemon called by pid=${process.pid} — killing daemon pid=${pid}\n`);
84
88
  } catch {}
89
+ if (victimAlive) { try { writeExitMarker(daemonDir(), { reason, prevPid: pid, prevVersion: readPidVersion(), byPid: process.pid }); } catch {} }
85
90
  try {
86
91
  process.kill(pid, "SIGTERM");
87
92
  } catch (err) {
@@ -92,5 +97,5 @@ export function stopDaemon() {
92
97
  } catch {
93
98
  // already gone
94
99
  }
95
- return { stopped: true, pid };
100
+ return { stopped: true, pid, stale: !victimAlive };
96
101
  }
package/src/logs.js CHANGED
@@ -108,7 +108,20 @@ export function appendError(entry, { file = errorsFile() } = {}) {
108
108
  // Structured debug event stream: one JSON line per request/error/forward
109
109
  // step, consumed live by `mslxdff -debug`.
110
110
  export function appendEvent(entry, { file = eventsFile() } = {}) {
111
- appendLine(file, entry);
111
+ appendLine(file, dropTalk(entry));
112
+ }
113
+
114
+ // events.log 是 debug 流(1MB 环形),对话正文只准进 talk/*.log。
115
+ // 单点剥掉 detail.talk:任何 relay 路径(含绕过 relay-pipeline 的 exhausted 收尾)都不会
116
+ // 把整段问答灌进 debug 行;留 talkChars 计数,debug 时仍看得出这轮有没有出字。
117
+ function dropTalk(entry) {
118
+ const t = entry?.detail?.talk;
119
+ if (!t) return entry;
120
+ const detail = { ...entry.detail };
121
+ delete detail.talk;
122
+ detail.talkChars = (t.content || []).join("").length;
123
+ detail.talkReasoningChars = (t.reasoning || []).join("").length;
124
+ return { ...entry, detail };
112
125
  }
113
126
 
114
127
  export function recentEvents(n = 100, { file = eventsFile() } = {}) {