mslxdff 0.1.245 → 0.1.247

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 (82) hide show
  1. package/package.json +1 -1
  2. package/src/autostart-health.js +65 -0
  3. package/src/autostart.js +57 -1
  4. package/src/chat-pipeline/terminal-verdict.js +4 -4
  5. package/src/cli/commands/daemon.js +9 -1
  6. package/src/cli/commands/provider/index.js +6 -1
  7. package/src/cli/commands/provider/models.js +3 -0
  8. package/src/cli/commands/provider/todaycn-login.js +121 -0
  9. package/src/cli/commands/provider/todaycn-persona.js +117 -0
  10. package/src/cli/commands/system.js +21 -5
  11. package/src/cli/help.js +4 -4
  12. package/src/cli/index.js +0 -5
  13. package/src/cli/status.js +11 -4
  14. package/src/daemon.js +3 -3
  15. package/src/providers/classify.js +1 -1
  16. package/src/providers/codearts/stream.js +17 -0
  17. package/src/providers/dispatcher.js +1 -1
  18. package/src/providers/railway-endpoints.js +192 -0
  19. package/src/providers/railway.js +111 -45
  20. package/src/providers/registry.js +8 -2
  21. package/src/providers/share-keys.js +2 -1
  22. package/src/providers/todaycn/account-store.js +169 -0
  23. package/src/providers/todaycn/chat.js +168 -0
  24. package/src/providers/todaycn/const.js +59 -0
  25. package/src/providers/todaycn/index.js +202 -0
  26. package/src/providers/todaycn/models.js +93 -0
  27. package/src/providers/todaycn/persona.js +103 -0
  28. package/src/providers/todaycn/prompt.js +194 -0
  29. package/src/providers/todaycn/session.js +107 -0
  30. package/src/providers/todaycn/sse.js +86 -0
  31. package/src/providers/todaycn/token.js +66 -0
  32. package/src/providers/todaycn.js +2 -0
  33. package/src/runtime/autostart-repair.js +108 -0
  34. package/src/runtime/bootstrap.js +1 -1
  35. package/src/runtime/proc-info.js +2 -2
  36. package/src/runtime/provider-gate.js +2 -2
  37. package/src/runtime/providers-setup.js +2 -4
  38. package/src/runtime/server-lifecycle.js +21 -48
  39. package/src/server.js +3 -3
  40. package/src/state/facade.js +1 -0
  41. package/src/state/merge.js +2 -0
  42. package/src/state/schemas/autostart.js +79 -0
  43. package/src/upstream-engine/index.js +2 -17
  44. package/src/upstream.js +1 -61
  45. package/src/cli/commands/multi-ip-pool-fetch.js +0 -173
  46. package/src/cli/commands/multi-ip-pool-test.js +0 -136
  47. package/src/cli/commands/multi-ip-pool.js +0 -83
  48. package/src/cli/commands/v6-pool.js +0 -123
  49. package/src/multi-ip-pool/config.js +0 -157
  50. package/src/multi-ip-pool/diagnose.js +0 -159
  51. package/src/multi-ip-pool/dial.js +0 -179
  52. package/src/multi-ip-pool/dispatch.js +0 -244
  53. package/src/multi-ip-pool/index.js +0 -174
  54. package/src/multi-ip-pool/nodes-fetch.js +0 -160
  55. package/src/multi-ip-pool/nodes.js +0 -205
  56. package/src/multi-ip-pool/report.js +0 -162
  57. package/src/multi-ip-pool/selftest.js +0 -168
  58. package/src/multi-ip-pool/standby.js +0 -157
  59. package/src/multi-ip-pool/tunnel.js +0 -291
  60. package/src/v6pool/claim.js +0 -157
  61. package/src/v6pool/config.js +0 -65
  62. package/src/v6pool/detect.js +0 -130
  63. package/src/v6pool/index.js +0 -157
  64. package/src/v6pool/prefix.js +0 -140
  65. package/src/v6pool/probe.js +0 -40
  66. package/src/v6pool/report.js +0 -104
  67. package/src/v6pool/rotate.js +0 -263
  68. package/src/v6pool/selftest.js +0 -178
  69. package/src/warp-pool/config.js +0 -74
  70. package/src/warp-pool/health.js +0 -106
  71. package/src/warp-pool/identity.js +0 -141
  72. package/src/warp-pool/index.js +0 -358
  73. package/src/warp-pool/keys.js +0 -29
  74. package/src/warp-pool/nodes.js +0 -95
  75. package/src/warp-pool/ports.js +0 -160
  76. package/src/warp-pool/register.js +0 -113
  77. package/src/warp-pool/report.js +0 -232
  78. package/src/warp-pool/revive.js +0 -168
  79. package/src/warp-pool/rotate.js +0 -214
  80. package/src/warp-pool/strategy.js +0 -43
  81. package/src/warp-pool/verdict.js +0 -69
  82. package/src/warp-pool/wireproxy.js +0 -162
@@ -98,6 +98,20 @@ const chunkLine = (id, model, delta, finish) => {
98
98
  return `data: ${JSON.stringify({ id, object: "chat.completion.chunk", created: Math.floor(Date.now() / 1000), model, choices: [choice] })}\n\n`;
99
99
  };
100
100
 
101
+ // 流式 usage 收尾:OpenAI 兼容 chunk 把 usage 放顶层(choices 之外),与 aggregateToCompletion 同形状。
102
+ // 若上游从未给 usage(state.usage 为 null)则不伪造 —— 上游没报 → 留 null,报表层渲染 —/0。
103
+ function usageChunk(id, model, usage) {
104
+ if (!usage || typeof usage !== "object") return "";
105
+ const pt = Number(usage.prompt_tokens) || 0;
106
+ const ct = Number(usage.completion_tokens) || 0;
107
+ const tt = Number(usage.total_tokens);
108
+ const total = Number.isFinite(tt) ? tt : pt + ct;
109
+ return `data: ${JSON.stringify({
110
+ id, object: "chat.completion.chunk", created: Math.floor(Date.now() / 1000), model,
111
+ choices: [], usage: { prompt_tokens: pt, completion_tokens: ct, total_tokens: total },
112
+ })}\n\n`;
113
+ }
114
+
101
115
  /** 上游 SSE → OpenAI SSE 流式转换(快照替换/增量混合语义,保证至少一个 finish chunk + [DONE])。 */
102
116
  export function sseToOpenAIResponse(resp, { model, id }) {
103
117
  const dec = new TextDecoder();
@@ -122,6 +136,7 @@ export function sseToOpenAIResponse(resp, { model, id }) {
122
136
  if (!ev) continue;
123
137
  if (ev.data === "[DONE]") {
124
138
  finishChunk();
139
+ emit(usageChunk(id, model, state.usage));
125
140
  emit("data: [DONE]\n\n");
126
141
  controller.close();
127
142
  return;
@@ -145,6 +160,7 @@ export function sseToOpenAIResponse(resp, { model, id }) {
145
160
  if (fwd.finish) finishChunk(); // 上游 finish_reason 出现即发(含 tool_calls;空 calls 自动降 stop)
146
161
  if (state.done) {
147
162
  finishChunk();
163
+ emit(usageChunk(id, model, state.usage));
148
164
  emit("data: [DONE]\n\n");
149
165
  controller.close();
150
166
  return;
@@ -152,6 +168,7 @@ export function sseToOpenAIResponse(resp, { model, id }) {
152
168
  }
153
169
  if (done) {
154
170
  finishChunk();
171
+ emit(usageChunk(id, model, state.usage));
155
172
  emit("data: [DONE]\n\n");
156
173
  controller.close();
157
174
  return;
@@ -3,7 +3,7 @@
3
3
  import { isModelAllowed as stateIsAllowed, loadProviderAllowedModels as stateLoadAllowed, loadProviderAllowAnyModels as stateLoadAllowAny } from "../state.js";
4
4
  // 定制预热名单:registry customProviders + 默认 opencode;openrouter/自配 generic 不预热(按需自拉)
5
5
  // agentrouter 已于 2026-10-21 随 registry 一并摘除(见 src/providers/registry.js 注释)。
6
- const CUSTOM_PREHEAT_IDS = new Set(["opencode", "workbuddy", "traework", "cline", "codearts", "qoder", "globalqwenwork", "qwenwork", "zcode", "raccoon"]);
6
+ const CUSTOM_PREHEAT_IDS = new Set(["opencode", "workbuddy", "traework", "cline", "codearts", "qoder", "globalqwenwork", "qwenwork", "zcode", "raccoon", "todaycn"]);
7
7
  function isPreheatDisabled() {
8
8
  const raw = process.env.MSLXDFF_PREHEAT;
9
9
  if (raw === undefined || raw === null || raw === "") return false;
@@ -0,0 +1,192 @@
1
+ // railway 的多出口选路:同一套 zen 契约、不同出口 IP(免费额度按出口 IP 计 ⇒ 多份额度)。
2
+ // 缺省顺序固定为「新隧道(instabox)→ 原 railway 反代」——新出口先用,用完/挂掉即下沉到下一个。
3
+ //
4
+ // 失败分档(2026-10-25 用户裁决):
5
+ // 429(额度被上游收走) → 该出口长冷却 1 小时;本请求内出局,换另一个出口**不等待**
6
+ // 5xx / 连不上 / 超时 → 该出口短冷却 3s;本请求内**等 3s 再换另一个出口**(两侧轮流重试)
7
+ // 其它 4xx(400/401/403) → 原样交回(换出口没用;FreeTierError / 模型不存在的终局判决靠它)
8
+ // 每出口最多试 N 次(缺省 2)→ 最多 4 发、最多等 3×3s = 9s,然后交回最后那条错误
9
+ //
10
+ // env:
11
+ // MSLXDFF_RAILWAY_BASES —— 逗号分隔,**完整替换**出口列表(顺序=优先级)
12
+ // MSLXDFF_RAILWAY_ENDPOINT_COOLDOWN_MS —— 非 429 失败的出口短冷却(缺省 3000,0=不冷却)
13
+ // MSLXDFF_RAILWAY_ENDPOINT_QUOTA_COOLDOWN_MS —— 429 的出口长冷却(缺省 3600000 = 1h)
14
+ // MSLXDFF_RAILWAY_ENDPOINT_RETRY_DELAY_MS —— 换出口前的等待(缺省 3000,0=不等待)
15
+ // MSLXDFF_RAILWAY_ENDPOINT_ROUNDS —— 每出口最多试几次(缺省 2)
16
+ //
17
+ // 纯逻辑模块:无 fs / 无 HTTP,便于单测(sleep 可注入)。
18
+ import { envInt } from "./base.js";
19
+
20
+ /** 新隧道出口(用户 2026-10-25 提供,缺省第一优先)。 */
21
+ export const INSTABOX_RAILWAY_BASE = "https://3000-agqxlhk7jgtegt4n5acnqyrndi.instabox.tech/ai-api";
22
+ /** 原 railway 反代出口(兜底)。 */
23
+ export const UPSTREAM_RAILWAY_BASE = "https://mytestip-production.up.railway.app/ai-api";
24
+ /** 缺省出口列表面(顺序=优先级)。 */
25
+ export const DEFAULT_RAILWAY_BASES = [INSTABOX_RAILWAY_BASE, UPSTREAM_RAILWAY_BASE];
26
+
27
+ export function normalizeBase(url) {
28
+ return String(url ?? "").trim().replace(/\/+$/, "");
29
+ }
30
+
31
+ function dedupe(list) {
32
+ const out = [];
33
+ for (const item of list) {
34
+ const u = normalizeBase(item);
35
+ if (u && !out.includes(u)) out.push(u);
36
+ }
37
+ return out;
38
+ }
39
+
40
+ /** 各档时长/轮数的单一真相(显式入参 > env > 缺省)。 */
41
+ export function endpointTuning(overrides = {}) {
42
+ const nums = {
43
+ shortCooldownMs: overrides.shortCooldownMs ?? envInt("MSLXDFF_RAILWAY_ENDPOINT_COOLDOWN_MS", 3_000),
44
+ quotaCooldownMs: overrides.quotaCooldownMs ?? envInt("MSLXDFF_RAILWAY_ENDPOINT_QUOTA_COOLDOWN_MS", 3_600_000),
45
+ retryDelayMs: overrides.retryDelayMs ?? envInt("MSLXDFF_RAILWAY_ENDPOINT_RETRY_DELAY_MS", 3_000),
46
+ rounds: Math.max(1, overrides.rounds ?? envInt("MSLXDFF_RAILWAY_ENDPOINT_ROUNDS", 2)),
47
+ };
48
+ return nums;
49
+ }
50
+
51
+ /**
52
+ * 解析该供应商的出口列表(顺序 = 优先级)。
53
+ * ① env `MSLXDFF_RAILWAY_BASES` —— 完整替换(运维逃生阀)
54
+ * ② 显式 `baseUrl` 为数组 —— 完整替换;为字符串且**不在**缺省列表里 —— 该地址排第一 + 缺省列表兜底
55
+ * ③ 字符串等于缺省之一(state 里的播种值、或用户显式指到缺省)—— 用缺省顺序(保证新出口仍优先)
56
+ * ④ 都没有 —— 缺省顺序
57
+ */
58
+ export function resolveRailwayBases({ baseUrl, env = process.env } = {}) {
59
+ const fromEnv = dedupe(String(env?.MSLXDFF_RAILWAY_BASES ?? "").split(","));
60
+ if (fromEnv.length) return fromEnv;
61
+ if (Array.isArray(baseUrl)) {
62
+ const list = dedupe(baseUrl);
63
+ if (list.length) return list;
64
+ }
65
+ const one = normalizeBase(baseUrl);
66
+ if (one) {
67
+ if (DEFAULT_RAILWAY_BASES.includes(one)) return [...DEFAULT_RAILWAY_BASES];
68
+ return dedupe([one, ...DEFAULT_RAILWAY_BASES]);
69
+ }
70
+ return [...DEFAULT_RAILWAY_BASES];
71
+ }
72
+
73
+ /** 额度耗尽(429):换出口=换额度池,值得换、且该出口要歇很久。 */
74
+ export function isQuotaStatus(status) {
75
+ return Number(status) === 429;
76
+ }
77
+
78
+ /**
79
+ * 出口侧故障(5xx):值得换出口。
80
+ * 其余 4xx —— 400 参数错、401/403 身份与形状错 —— 换出口只会把同一堵墙撞两次,交回原样。
81
+ */
82
+ export function isSwapWorthyStatus(status) {
83
+ const s = Number(status);
84
+ if (!Number.isFinite(s)) return false;
85
+ return s >= 500 && s < 600;
86
+ }
87
+
88
+ /**
89
+ * 出口冷却账:`order()` 把未冷却的按原相对顺序排前、冷却中的排后(全冷却时就是原顺序)。
90
+ * 短冷却(非 429)与长冷却(429)共用一张截止时刻表,只是时长不同。
91
+ */
92
+ export function createEndpointRing(
93
+ bases,
94
+ {
95
+ cooldownMs = endpointTuning().shortCooldownMs,
96
+ quotaCooldownMs = endpointTuning().quotaCooldownMs,
97
+ clock = Date.now,
98
+ } = {},
99
+ ) {
100
+ const list = dedupe(bases);
101
+ const cooling = new Map(); // base -> 冷却截止时刻
102
+ const isCooling = (base, now) => (cooling.get(normalizeBase(base)) || 0) > now;
103
+ return {
104
+ bases: [...list],
105
+ cooldownMs,
106
+ quotaCooldownMs,
107
+ /** 当前尝试顺序(顺序=优先级,冷却者沉底但仍在列)。 */
108
+ order() {
109
+ const now = clock();
110
+ const ready = [];
111
+ const cooled = [];
112
+ for (const b of list) (isCooling(b, now) ? cooled : ready).push(b);
113
+ return [...ready, ...cooled];
114
+ },
115
+ cooldownUntil(base) { return cooling.get(normalizeBase(base)) || 0; },
116
+ /** 出口侧故障(5xx/连不上):短冷却,轮转几发后又会优先试它。 */
117
+ noteExhausted(base) { cooling.set(normalizeBase(base), clock() + cooldownMs); },
118
+ /** 额度耗尽(429):长冷却(缺省 1 小时),期间不再排前。 */
119
+ noteQuotaExhausted(base) { cooling.set(normalizeBase(base), clock() + quotaCooldownMs); },
120
+ /** 出口可用:清冷却(额度回血后立刻恢复优先级)。 */
121
+ noteOk(base) { cooling.delete(normalizeBase(base)); },
122
+ };
123
+ }
124
+
125
+ const defaultSleep = (ms) => (ms > 0 ? new Promise((r) => setTimeout(r, ms)) : Promise.resolve());
126
+
127
+ /**
128
+ * 出口轮流重试编排(实际发请求由 `attempt` 注入)。
129
+ *
130
+ * 顺序 = A→B→A→B(每个出口最多 `rounds` 次机会),直到成功或用完机会:
131
+ * - `{kind:"ok", result}` —— 成功,立刻返回(不再试别的出口)
132
+ * - `{kind:"quota", result}` —— 429:该出口本请求内出局 + `onQuota`(长冷却),换出口**不等待**
133
+ * - `{kind:"swap", ...}` —— 5xx/连不上/超时:`onSwap`(短冷却),还有机会就**等 retryDelayMs** 换下一个
134
+ * - `{kind:"fatal", result}` —— 其它 4xx:立刻交回原响应,不冷却、不换出口
135
+ * - attempt 抛错视作 `swap`(连不上)
136
+ *
137
+ * 只有「确定放弃这一发、还要继续试」时才取消它的 body;**最后交回的那一发 body 保持完好**,
138
+ * 上层(终局判决/错误文案)才能读到上游原文。
139
+ */
140
+ export async function runEndpointRotation({
141
+ bases,
142
+ rounds = endpointTuning().rounds,
143
+ retryDelayMs = endpointTuning().retryDelayMs,
144
+ sleep = defaultSleep,
145
+ attempt,
146
+ onOk,
147
+ onSwap,
148
+ onQuota,
149
+ } = {}) {
150
+ const list = Array.isArray(bases) ? bases.filter(Boolean) : [];
151
+ if (!list.length || typeof attempt !== "function") return { kind: "none", err: new Error("railway: no endpoint available") };
152
+ const totalRounds = Math.max(1, Number(rounds) || 1);
153
+ const left = new Map(list.map((b) => [b, totalRounds]));
154
+ let last = { kind: "none", err: new Error("railway: no endpoint available") };
155
+ for (let round = 0; round < totalRounds; round++) {
156
+ let progressed = false;
157
+ for (const base of list) {
158
+ if ((left.get(base) || 0) <= 0) continue;
159
+ progressed = true;
160
+ const remaining = [...left.values()].reduce((a, b) => a + b, 0);
161
+ const isFinal = remaining <= 1; // 这是最后一次机会 → 结果必须原样带回
162
+ let outcome;
163
+ try {
164
+ outcome = await attempt(base);
165
+ } catch (err) {
166
+ outcome = { kind: "swap", err };
167
+ }
168
+ const kind = outcome?.kind || "swap";
169
+ if (kind === "ok") {
170
+ onOk?.(base);
171
+ return { kind: "ok", base, result: outcome.result };
172
+ }
173
+ if (kind === "fatal") return { kind: "fatal", base, result: outcome.result, err: outcome.err };
174
+ // 失败(quota=429 / swap=5xx·连不上·超时):先记账冷却(最后一发同样要记),再决定交回还是继续
175
+ if (kind === "quota") onQuota?.(base);
176
+ else onSwap?.(base);
177
+ const record = { kind, base, result: outcome.result, err: outcome.err };
178
+ if (isFinal) return record; // 最后一次机会:原样交回(body 不能动)
179
+ // 还有后续机会:释放这一发的 body,再决定等不等
180
+ try { await outcome.result?.body?.cancel?.(); } catch {}
181
+ last = record;
182
+ if (kind === "quota") {
183
+ left.set(base, 0); // 额度耗尽:本请求内不再试它
184
+ continue; // 换出口不等待
185
+ }
186
+ left.set(base, (left.get(base) || 0) - 1);
187
+ await sleep(retryDelayMs); // 其它失败:等一会再换另一个 URL
188
+ }
189
+ if (!progressed) break;
190
+ }
191
+ return last;
192
+ }
@@ -1,12 +1,18 @@
1
- // Railway 定制供应商:opencode zen free 层的**另一个 IP 出口**(Railway 边缘反代 opencode.ai)。
2
- // 本质:同一套 zen 契约,只是出口 IP 不同 ⇒ 免费额度按出口 IP 计,等于多一份额度。
1
+ // Railway 定制供应商:opencode zen free 层的**多个 IP 出口**(第三方边缘反代 opencode.ai)。
2
+ // 本质:同一套 zen 契约,只是出口 IP 不同 ⇒ 免费额度按出口 IP 计,等于多份额度。
3
+ //
4
+ // 多出口轮流重试(顺序=优先级,缺省「新隧道 → 原 railway 反代」,判据/冷却全在 railway-endpoints.js):
5
+ // 429(额度被上游收走)→ 该出口歇 1 小时;本请求内出局,换另一个出口不等待
6
+ // 5xx / 连不上 / 超时 → 该出口歇 3s;本请求内等 3s 换另一个 URL(两侧轮流),每出口最多 2 次
7
+ // 400 / 401 / 403 → 原样交回(换出口无用;FreeTierError、模型不存在的终局判决靠它)
8
+ // 上层(dispatcher / chat-pipeline)零感知——仍是同一个 `railway/<model>`。
3
9
  //
4
10
  // 与内置 opencode 的区别:
5
11
  // ① 只暴露 free 模型(上游 /models 会把付费模型一起列出来,这里过滤掉);
6
12
  // ② 每个请求强制补 free-lane 形状(stream:true + 五个核心工具名),否则 403 FreeTierError;
7
13
  // ③ 带 opencode 身份头(UA + client/session/request/project),否则同样 403。
8
14
  //
9
- // 端点:<base>/zen/v1/models、<base>/zen/v1/chat/completions(base 缺省为 Railway 反代根)。
15
+ // 端点:<base>/zen/v1/models、<base>/zen/v1/chat/completions(base 取自出口列表)。
10
16
  import crypto from "node:crypto";
11
17
  import { createKeyRing } from "./keyring.js";
12
18
  import { collectApiKeysGeneric, envInt, createAgent, getUndici, joinUrl } from "./base.js";
@@ -15,10 +21,19 @@ import { loadProviderBaseUrl, loadProviderKeys, saveProviderConfig } from "../st
15
21
  import { ensureFreeLaneShape, aggregateChatSse } from "../free-lane.js";
16
22
  import { genId, opencodeUa, digestIdTail } from "../opencode-identity.js";
17
23
  import { joinModelId } from "./model-id.js";
24
+ import {
25
+ UPSTREAM_RAILWAY_BASE,
26
+ resolveRailwayBases,
27
+ isSwapWorthyStatus,
28
+ createEndpointRing,
29
+ endpointTuning,
30
+ runEndpointRotation,
31
+ } from "./railway-endpoints.js";
18
32
 
19
33
  const { UndiciFetch } = getUndici();
20
34
 
21
- export const DEFAULT_RAILWAY_BASE = "https://mytestip-production.up.railway.app/ai-api";
35
+ /** 兼容旧常量:老代码(含测试)引用的「原 railway 反代根」。 */
36
+ export const DEFAULT_RAILWAY_BASE = UPSTREAM_RAILWAY_BASE;
22
37
  export const MODELS_PATH = "/zen/v1/models";
23
38
  export const CHAT_PATH = "/zen/v1/chat/completions";
24
39
 
@@ -37,12 +52,6 @@ export function filterFreeModels(list) {
37
52
  return (Array.isArray(list) ? list : []).filter((m) => m && isFreeModelId(m.id));
38
53
  }
39
54
 
40
- function resolveBase(baseUrl) {
41
- if (baseUrl) return String(baseUrl).trim().replace(/\/+$/, "");
42
- const env = loadProviderBaseUrl("railway");
43
- return env || DEFAULT_RAILWAY_BASE;
44
- }
45
-
46
55
  export function createRailwayProvider({
47
56
  id = "railway",
48
57
  baseUrl,
@@ -50,12 +59,25 @@ export function createRailwayProvider({
50
59
  apiKey,
51
60
  connectTimeoutMs = envInt("MSLXDFF_RAILWAY_TIMEOUT_MS", 60_000),
52
61
  cooldownMs = envInt("MSLXDFF_RAILWAY_COOLDOWN_MS", 30_000),
62
+ endpointCooldownMs,
63
+ endpointQuotaCooldownMs,
64
+ endpointRetryDelayMs,
65
+ endpointRounds,
53
66
  fetchImpl,
54
67
  file,
55
68
  clock = Date.now,
56
69
  } = {}) {
57
70
  if (id !== "railway") throw new Error(`railway provider id mismatch: ${id}`);
58
- const resolvedBase = resolveBase(baseUrl);
71
+ const tuning = endpointTuning({
72
+ shortCooldownMs: endpointCooldownMs,
73
+ quotaCooldownMs: endpointQuotaCooldownMs,
74
+ retryDelayMs: endpointRetryDelayMs,
75
+ rounds: endpointRounds,
76
+ });
77
+ const endpoints = createEndpointRing(resolveRailwayBases({ baseUrl }), {
78
+ cooldownMs: tuning.shortCooldownMs,
79
+ quotaCooldownMs: tuning.quotaCooldownMs,
80
+ });
59
81
  if (!fetchImpl) fetchImpl = UndiciFetch || compatFetch;
60
82
 
61
83
  const rawKeys = collectApiKeysGeneric(id, apiKeys, apiKey, (pid) => loadProviderKeys(pid, file ? { file } : {}));
@@ -101,70 +123,110 @@ export function createRailwayProvider({
101
123
  };
102
124
  }
103
125
 
104
- async function chat(body, opts = {}) {
105
- const wantStream = body?.stream === true;
106
- // free-lane 强制 stream:true:非流式请求在这里补形状,聚合回 JSON 后再返回。
107
- const outBody = { ...body, stream: true };
108
- ensureFreeLaneShape(outBody);
126
+ /** 单出口一发(不含判据,成败交由编排裁决)。 */
127
+ async function postTo(base, outBody, opts, key) {
109
128
  const headers = buildHeaders(opts);
110
- const key = ring.size ? ring.next() : null;
111
129
  if (key) headers.Authorization = `Bearer ${key}`;
112
-
113
- const url = joinUrl(resolvedBase, CHAT_PATH);
114
130
  const ac = new AbortController();
115
131
  const timer = setTimeout(() => ac.abort(new Error(`railway timed out after ${connectTimeoutMs}ms`)), connectTimeoutMs);
116
- let res;
117
132
  try {
118
- res = await fetchImpl(url, {
133
+ return await fetchImpl(joinUrl(base, CHAT_PATH), {
119
134
  method: "POST",
120
135
  headers,
121
136
  body: JSON.stringify(outBody),
122
137
  signal: ac.signal,
123
138
  ...(dispatcher ? { dispatcher } : {}),
124
139
  });
125
- } catch (err) {
140
+ } finally {
126
141
  clearTimeout(timer);
127
- if (key) ring.onError(key);
128
- throw err;
129
142
  }
130
- clearTimeout(timer);
143
+ }
131
144
 
132
- if (res.status >= 400) {
133
- if (key) ring.onError(key);
134
- return res;
135
- }
136
- if (!wantStream) {
137
- // 客户端要非流式:把上游 SSE 聚合回 chat.completion JSON
138
- return aggregateChatSse(res);
145
+ /** 一发上游的分类:ok / quota(429) / swap(5xx 或连不上) / fatal(其它 4xx)。 */
146
+ function classifyResponse(res, key) {
147
+ if (res.status < 400) return { kind: "ok", result: res };
148
+ if (key) ring.onError(key);
149
+ if (res.status === 429) return { kind: "quota", result: res };
150
+ if (isSwapWorthyStatus(res.status)) return { kind: "swap", result: res };
151
+ return { kind: "fatal", result: res };
152
+ }
153
+
154
+ /** 出口编排放账:成功清冷却、5xx 短冷却、429 长冷却。 */
155
+ const endpointHooks = {
156
+ onOk: (base) => endpoints.noteOk(base),
157
+ onSwap: (base) => endpoints.noteExhausted(base),
158
+ onQuota: (base) => endpoints.noteQuotaExhausted(base),
159
+ };
160
+
161
+ async function chat(body, opts = {}) {
162
+ const wantStream = body?.stream === true;
163
+ // free-lane 强制 stream:true:非流式请求在这里补形状,聚合回 JSON 后再返回。
164
+ const outBody = { ...body, stream: true };
165
+ ensureFreeLaneShape(outBody);
166
+ const outcome = await runEndpointRotation({
167
+ bases: endpoints.order(),
168
+ rounds: tuning.rounds,
169
+ retryDelayMs: tuning.retryDelayMs,
170
+ ...endpointHooks,
171
+ attempt: async (base) => {
172
+ const key = ring.size ? ring.next() : null;
173
+ try {
174
+ return classifyResponse(await postTo(base, outBody, opts, key), key);
175
+ } catch (err) {
176
+ if (key) ring.onError(key);
177
+ return { kind: "swap", err }; // 连不上 / 超时:等 3s 换另一个 URL
178
+ }
179
+ },
180
+ });
181
+ if (outcome.kind === "ok") {
182
+ const res = outcome.result;
183
+ return wantStream ? res : aggregateChatSse(res); // 客户端要非流式:SSE 聚合回 JSON
139
184
  }
140
- return res;
185
+ if (outcome.result) return outcome.result; // 交回上游原响应(429 / 5xx / 其它 4xx),body 完好
186
+ throw outcome.err || new Error("railway: no endpoint available");
141
187
  }
142
188
 
143
189
  let modelsCache = { at: 0, list: [] };
144
190
  const MODELS_TTL_MS = 10 * 60 * 1000;
145
191
 
146
- async function fetchModels() {
192
+ async function fetchModels(base) {
147
193
  const key = ring.size ? ring.next() : null;
148
194
  const headers = { "Accept": "application/json", "User-Agent": opencodeUa() };
149
195
  if (key) headers.Authorization = `Bearer ${key}`;
150
- const res = await fetchImpl(joinUrl(resolvedBase, MODELS_PATH), {
196
+ const res = await fetchImpl(joinUrl(base, MODELS_PATH), {
151
197
  headers,
152
198
  ...(dispatcher ? { dispatcher } : {}),
153
199
  });
154
- if (!res.ok) throw new Error(`railway models: http ${res.status}`);
200
+ if (!res.ok) {
201
+ const err = new Error(`railway models: http ${res.status}`);
202
+ err.status = res.status;
203
+ try { await res.body?.cancel?.(); } catch {}
204
+ throw err;
205
+ }
155
206
  const j = await res.json();
156
207
  return filterFreeModels(j?.data);
157
208
  }
158
209
 
159
210
  async function listModels() {
160
211
  if (clock() - modelsCache.at < MODELS_TTL_MS && modelsCache.list.length) return modelsCache.list;
161
- let raw = [];
162
- try {
163
- raw = await fetchModels();
164
- } catch {
165
- return modelsCache.list; // 拉取失败:给上一次的缓存(首次则空表),不阻断启动
166
- }
167
- const list = raw.map((m) => ({
212
+ const outcome = await runEndpointRotation({
213
+ bases: endpoints.order(),
214
+ rounds: tuning.rounds,
215
+ retryDelayMs: tuning.retryDelayMs,
216
+ ...endpointHooks,
217
+ attempt: async (base) => {
218
+ try {
219
+ return { kind: "ok", result: await fetchModels(base) };
220
+ } catch (err) {
221
+ const status = Number(err?.status) || 0;
222
+ if (status === 429) return { kind: "quota", err };
223
+ if (status >= 400 && status < 500) return { kind: "fatal", err }; // 401/403/404:换出口无益
224
+ return { kind: "swap", err };
225
+ }
226
+ },
227
+ });
228
+ if (outcome.kind !== "ok" || !outcome.result) return modelsCache.list; // 全失败:给上一次缓存(首次空表)
229
+ const list = outcome.result.map((m) => ({
168
230
  id: joinModelId(id, m.id),
169
231
  object: "model",
170
232
  name: m.name || m.id,
@@ -194,7 +256,11 @@ export function createRailwayProvider({
194
256
  close,
195
257
  agent,
196
258
  keyRing: ring,
197
- baseUrl: resolvedBase,
259
+ bases: [...endpoints.bases],
260
+ orderBases: () => endpoints.order(),
261
+ endpointCooldownUntil: (base) => endpoints.cooldownUntil(base),
262
+ tuning: { ...tuning },
263
+ baseUrl: endpoints.bases[0] || DEFAULT_RAILWAY_BASE,
198
264
  _isFreeModelId: isFreeModelId,
199
265
  _filterFreeModels: filterFreeModels,
200
266
  };
@@ -210,4 +276,4 @@ export function saveRailwayConfig({ baseUrl, keys } = {}, { file } = {}) {
210
276
  baseUrl: baseUrl ?? cur.baseUrl,
211
277
  keys: keys ?? cur.keys,
212
278
  }, file ? { file } : {});
213
- }
279
+ }
@@ -51,8 +51,8 @@ export const customProviders = [
51
51
  {
52
52
  id: "railway",
53
53
  // opencode zen free 层的另一个 IP 出口(Railway 反代),只暴露 free 模型。
54
- // 匹配必须排在 generic 之前,且不能与其它条目撞(railway.app 域名足够独特)。
55
- match: (id, baseUrl) => id === "railway" || String(baseUrl).includes("railway.app"),
54
+ // 匹配必须排在 generic 之前,且不能与其它条目撞(railway.app / instabox.tech 域名足够独特)。
55
+ match: (id, baseUrl) => id === "railway" || /railway\.app|instabox\.tech/.test(String(baseUrl)),
56
56
  load: () => import("./railway.js").then((m) => m.createRailwayProvider),
57
57
  },
58
58
  {
@@ -60,6 +60,12 @@ export const customProviders = [
60
60
  match: (id, baseUrl) => id === "raccoon" || String(baseUrl).includes("xiaohuanxiong.com"),
61
61
  load: () => import("./raccoon/index.js").then((m) => m.createRaccoonProvider),
62
62
  },
63
+ {
64
+ id: "todaycn",
65
+ // Today AI(今日工作助手):本机桌面登录态导入的多号轮换 + 文本契约模拟工具调用。
66
+ match: (id, baseUrl) => id === "todaycn" || String(baseUrl).includes("todayai.cn"),
67
+ load: () => import("./todaycn.js").then((m) => m.createTodaycnProvider),
68
+ },
63
69
  // agentrouter 定制工厂已于 2026-10-21 摘除:现网成功率仅 ~48%(502/400/422/500 高发),
64
70
  // 从注册表 + 预热名单移除后网关不再解析 agentrouter/* 模型。工厂文件 src/providers/agentrouter.js
65
71
  // 与其测试保留,日后要重新启用只需在此加回一条 { id:"agentrouter", ... } 并同步 dispatcher 预热名单。
@@ -10,7 +10,8 @@ export const SHARE_KEYS_HEADER = "x-mslxdff-share-keys";
10
10
  // 硬排除:refresh-token 型凭据(对端刷新会轮换 token,与本机互相踢下线)——正确性问题。
11
11
  // 二层保护:① 按 id 硬排除(cline 恒 local-only);② 按凭据形状兜底(见 shareableProviderIds)。
12
12
  // traework/workbuddy/zcode 同为本机账号绑定型(auths/*.json + uid;zcode 还带 per-account deviceMid),恒不借出。
13
- const NEVER_SHARE_IDS = new Set(["cline", "codearts", "traework", "qoder", "zcode", "raccoon"]);
13
+ // todaycn 同理(refreshCredential 由本机桌面登录态 DPAPI 解出、与 uid 绑定;形状不是 JWT,兜底闸认不出,必须靠本名单)。
14
+ const NEVER_SHARE_IDS = new Set(["cline", "codearts", "traework", "qoder", "zcode", "raccoon", "todaycn"]);
14
15
  // 本节点应 cast key 到出站转发的供应商 id 集合 = 所有「本机有 key」的供应商,
15
16
  // 减去:opencode(无 key 且恒排除)、local-only(workbuddy/cline系,本就不走组员)、刷新型凭据。
16
17
  export function shareableProviderIds({ file } = {}) {