mslxdff 0.1.161 → 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 (93) hide show
  1. package/bin/mslxdff.js +4 -4
  2. package/package.json +1 -1
  3. package/src/auto.js +254 -254
  4. package/src/bench/cline-bench.js +42 -42
  5. package/src/bench/probe.js +70 -70
  6. package/src/bench/report.js +162 -162
  7. package/src/bench/runner.js +77 -77
  8. package/src/bench/via-probe.js +124 -124
  9. package/src/bench/via-routes.js +87 -87
  10. package/src/bench/workbuddy-bench.js +54 -54
  11. package/src/chat/engine.js +160 -160
  12. package/src/chat/gateway.js +163 -163
  13. package/src/chat/orchestrator.js +234 -234
  14. package/src/chat/prompt.js +70 -70
  15. package/src/chat/repl.js +88 -88
  16. package/src/chat/terminal.js +135 -135
  17. package/src/chat/tools.js +306 -306
  18. package/src/chat-pipeline/index.js +126 -126
  19. package/src/chat-pipeline/policy.js +76 -76
  20. package/src/cli/commands/group.js +249 -249
  21. package/src/cli/commands/model/picks.js +50 -50
  22. package/src/cli/commands/provider/bench-via.js +247 -247
  23. package/src/cli/commands/provider/bench.js +141 -141
  24. package/src/cli/commands/provider/index.js +126 -126
  25. package/src/cli/commands/provider/models.js +142 -142
  26. package/src/cli/commands/provider/qwenwork-login.js +119 -119
  27. package/src/cli/commands/sync.js +232 -232
  28. package/src/cli/format.js +1 -0
  29. package/src/cli/status.js +279 -279
  30. package/src/daemon.js +101 -101
  31. package/src/model-capabilities/enrich.js +86 -86
  32. package/src/model-capabilities/index.js +183 -183
  33. package/src/model-capabilities/parse.js +70 -70
  34. package/src/model-trace.js +4 -3
  35. package/src/models.js +225 -225
  36. package/src/providers/cline/auth.js +228 -228
  37. package/src/providers/cline/chat.js +307 -307
  38. package/src/providers/cline.js +2 -2
  39. package/src/providers/dispatcher.js +26 -14
  40. package/src/providers/globalqwenwork/cosy.js +123 -123
  41. package/src/providers/globalqwenwork/crypto.js +218 -218
  42. package/src/providers/globalqwenwork/http.js +20 -20
  43. package/src/providers/globalqwenwork/rsa.js +54 -54
  44. package/src/providers/globalqwenwork/sse.js +268 -268
  45. package/src/providers/globalqwenwork/stream.js +130 -130
  46. package/src/providers/globalqwenwork/upstream.js +122 -122
  47. package/src/providers/keyring.js +60 -60
  48. package/src/providers/qoder/chat.js +193 -183
  49. package/src/providers/qoder/index.js +239 -230
  50. package/src/providers/qoder/sse.js +141 -103
  51. package/src/providers/qwenwork/account-store.js +133 -133
  52. package/src/providers/qwenwork/constants.js +67 -67
  53. package/src/providers/qwenwork/cosy.js +120 -120
  54. package/src/providers/qwenwork/crypto.js +218 -218
  55. package/src/providers/qwenwork/http.js +20 -20
  56. package/src/providers/qwenwork/index.js +327 -327
  57. package/src/providers/qwenwork/payload.js +142 -142
  58. package/src/providers/qwenwork/rsa.js +54 -54
  59. package/src/providers/qwenwork/sse.js +268 -268
  60. package/src/providers/qwenwork/stream.js +130 -130
  61. package/src/providers/qwenwork/upstream.js +120 -120
  62. package/src/providers/qwenwork.js +1 -1
  63. package/src/providers/registry.js +74 -74
  64. package/src/providers/workbuddy/chat.js +248 -248
  65. package/src/providers/workbuddy/reshape.js +152 -152
  66. package/src/providers/workbuddy.js +2 -2
  67. package/src/reasoning.js +32 -32
  68. package/src/routes/chat/gateway.js +46 -46
  69. package/src/routes/chat/relay-pipeline.js +276 -276
  70. package/src/routes/chat/via-route-handler.js +146 -146
  71. package/src/routes/hedge.js +255 -255
  72. package/src/routes/models-route.js +167 -167
  73. package/src/routes/peers.js +273 -273
  74. package/src/routes/stream.js +386 -386
  75. package/src/runtime/bootstrap.js +45 -45
  76. package/src/runtime/provider-gate.js +34 -34
  77. package/src/runtime/providers-setup.js +165 -165
  78. package/src/runtime/server-lifecycle.js +5 -6
  79. package/src/server.js +64 -64
  80. package/src/state/schemas/allowlist.js +92 -92
  81. package/src/sync-opencode.js +280 -280
  82. package/src/transport/index.js +244 -244
  83. package/src/transport/pool.js +56 -56
  84. package/src/transport/retry.js +24 -24
  85. package/src/transport/sse.js +93 -93
  86. package/src/upstream-probe/display.js +52 -52
  87. package/src/upstream-probe/probe.js +49 -49
  88. package/src/upstream-probe/rotate.js +110 -110
  89. package/src/upstream-probe/start.js +45 -45
  90. package/src/upstream.js +289 -289
  91. package/docs/cli_help_mini.md +0 -135
  92. package/src/providers/AGENTS.md +0 -46
  93. package/src/routes/AGENTS.md +0 -37
package/bin/mslxdff.js CHANGED
@@ -1,4 +1,4 @@
1
- #!/usr/bin/env node
2
- // thin adapter — all logic lives in src/cli (deep module, single inlet run)
3
- import { run } from "../src/cli/index.js";
4
- await run(process.argv.slice(2));
1
+ #!/usr/bin/env node
2
+ // thin adapter — all logic lives in src/cli (deep module, single inlet run)
3
+ import { run } from "../src/cli/index.js";
4
+ await run(process.argv.slice(2));
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "mslxdff",
3
- "version": "0.1.161",
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/auto.js CHANGED
@@ -1,254 +1,254 @@
1
- import { statSync } from "node:fs";
2
- import { loadModelErrors, saveModelErrors, loadModelLatencies, saveModelLatencies, loadPreferredModel, loadModelPicks, saveModelPicks, defaultStateFile } from "./state.js";
3
-
4
- // 出厂默认首选模型(state.json 的 preferredModel / env MSLXDFF_PREFERRED_MODEL 可覆盖)
5
- export const DEFAULT_PREFERRED_MODEL = "big-pickle";
6
- // 兼容旧导出名:语义为"出厂默认",当前生效值请用 getPreferredModel()
7
- export const PREFERRED_MODEL = DEFAULT_PREFERRED_MODEL;
8
-
9
- // 当前生效的首选模型:state.json > env > 出厂默认;mtime 缓存保证 daemon 热生效
10
- const _prefCache = { mtimeMs: -1, file: null, value: null };
11
- export function getPreferredModel({ file = defaultStateFile() } = {}) {
12
- try {
13
- const st = statSync(file);
14
- if (_prefCache.file !== file || st.mtimeMs !== _prefCache.mtimeMs) {
15
- _prefCache.file = file;
16
- _prefCache.mtimeMs = st.mtimeMs;
17
- _prefCache.value = loadPreferredModel({ file });
18
- }
19
- } catch {
20
- _prefCache.file = file;
21
- _prefCache.mtimeMs = -1;
22
- _prefCache.value = null;
23
- }
24
- const env = (process.env.MSLXDFF_PREFERRED_MODEL || "").trim();
25
- return _prefCache.value || env || DEFAULT_PREFERRED_MODEL;
26
- }
27
-
28
- export const DEFAULT_AUTO_MODELS = [
29
- PREFERRED_MODEL,
30
- "mimo-v2.5-free",
31
- "deepseek-v4-flash-free",
32
- "ling-3.0-flash-fin-free",
33
- "nemotron-3-ultra-free",
34
- "nemotron-3.5-lightning-free",
35
- "muse-spark-1.3-contributor-free",
36
- ].filter((id, i, arr) => id && arr.indexOf(id) === i);
37
-
38
- export function isAutoModel(model) {
39
- return !model || model === "auto";
40
- }
41
-
42
- export const MODEL_STATUS = Object.freeze({
43
- NORMAL: "normal",
44
- LIMIT: "limit",
45
- ERROR: "error",
46
- });
47
-
48
- // Legacy modelErrors entries are bare timestamps ({id: ts}); newer ones are
49
- // objects ({id: {status, at, code}}). Normalize both to an entry object.
50
- // `slow` flags a model whose last request was slow (over the wall-clock
51
- // threshold) — those get a longer cooldown so they lie low until they recover.
52
- function normEntry(e) {
53
- if (typeof e === "number") return { status: MODEL_STATUS.ERROR, at: e, code: null, slow: false };
54
- if (e && typeof e === "object") {
55
- return {
56
- status: e.status || MODEL_STATUS.ERROR,
57
- at: typeof e.at === "number" ? e.at : 0,
58
- code: e.code ?? null,
59
- slow: Boolean(e.slow),
60
- };
61
- }
62
- return null;
63
- }
64
-
65
- export function classifyErrorEvent(evt = {}) {
66
- if (evt.slow) return MODEL_STATUS.ERROR;
67
- const code = Number(evt.status);
68
- if (code === 429) return MODEL_STATUS.LIMIT;
69
- const msg = String(evt.message || evt.note || "").toLowerCase();
70
- if (msg.includes("rate limit") || msg.includes("limit exceeded") || msg.includes("429")) {
71
- return MODEL_STATUS.LIMIT;
72
- }
73
- return MODEL_STATUS.ERROR;
74
- }
75
-
76
- export const DEFAULT_COOLDOWN_MS = 60_000;
77
- export const DEFAULT_SLOW_COOLDOWN_MS = 5 * 60_000;
78
- export const DEFAULT_LATENCY_ALPHA = 0.3;
79
-
80
- function effectiveCooldown(entry, slowCooldownMs, cooldownMs) {
81
- if (entry && entry.slow) return slowCooldownMs || 0;
82
- return cooldownMs || 0;
83
- }
84
-
85
- function inCooldown(id, errors, now, cooldownMs, slowCooldownMs) {
86
- const e = normEntry(errors[id]);
87
- if (!e || !(e.at > 0)) return false;
88
- if (e.status === MODEL_STATUS.NORMAL) return false;
89
- const cd = effectiveCooldown(e, slowCooldownMs, cooldownMs);
90
- return cd > 0 && now - e.at < cd;
91
- }
92
-
93
- // Latency EMA helpers
94
- function normLatency(e) {
95
- if (!e || typeof e !== "object") return null;
96
- const ema = Number(e.emaMs);
97
- return Number.isFinite(ema) && ema > 0 ? ema : null;
98
- }
99
-
100
- export function rankModels(ids, errors = {}, { now = Date.now(), cooldownMs = 0, slowCooldownMs = 0, latencies = {}, preferred } = {}) {
101
- const pref = preferred ?? getPreferredModel();
102
- // 最近一次成功的模型(NORMAL 且非慢且 at 最大),用于“上次成功优先”——慢模型即使刚成功也不应钉死
103
- let lastSuccessId = null;
104
- let lastSuccessAt = 0;
105
- for (const [id, e] of Object.entries(errors)) {
106
- const ne = normEntry(e);
107
- if (ne && ne.status === MODEL_STATUS.NORMAL && !ne.slow && ne.at > lastSuccessAt) {
108
- lastSuccessAt = ne.at;
109
- lastSuccessId = id;
110
- }
111
- }
112
- return [...new Set(ids)]
113
- .filter(Boolean)
114
- .map((id) => ({
115
- id,
116
- e: normEntry(errors[id]),
117
- err: normEntry(errors[id])?.at ?? 0,
118
- isPreferred: id === pref,
119
- isLastSuccess: id === lastSuccessId,
120
- cooling: inCooldown(id, errors, now, cooldownMs, slowCooldownMs),
121
- latency: normLatency(latencies[id]) ?? Number.MAX_SAFE_INTEGER,
122
- }))
123
- .sort(
124
- (a, b) =>
125
- (a.cooling ? 1 : 0) - (b.cooling ? 1 : 0) ||
126
- (b.isLastSuccess ? 1 : 0) - (a.isLastSuccess ? 1 : 0) ||
127
- (b.isPreferred ? 1 : 0) - (a.isPreferred ? 1 : 0) ||
128
- a.latency - b.latency ||
129
- a.err - b.err
130
- )
131
- .map((x) => x.id);
132
- }
133
-
134
- export function createAutoSelector({
135
- loadCandidates,
136
- file,
137
- now = () => Date.now(),
138
- cooldownMs = DEFAULT_COOLDOWN_MS,
139
- slowCooldownMs = DEFAULT_SLOW_COOLDOWN_MS,
140
- latencyAlpha = DEFAULT_LATENCY_ALPHA,
141
- errors: seedErrors,
142
- latencies: seedLatencies,
143
- persist = (errors, f = file) => saveModelErrors(errors, f ? { file: f } : {}),
144
- persistLatencies = (latencies, f = file) => saveModelLatencies(latencies, f ? { file: f } : {}),
145
- loadPicks = () => (file ? loadModelPicks({ file }) : []),
146
- persistPicks = (picks) => (file ? saveModelPicks(picks, { file }) : picks),
147
- } = {}) {
148
- const lastErrorAt = { ...(seedErrors ?? loadModelErrors(file ? { file } : {})) };
149
- const latencies = { ...(seedLatencies ?? loadModelLatencies(file ? { file } : {})) };
150
-
151
- async function loadList() {
152
- let list;
153
- try {
154
- const loaded = await loadCandidates?.();
155
- list = Array.isArray(loaded) && loaded.length ? loaded : DEFAULT_AUTO_MODELS;
156
- } catch {
157
- list = DEFAULT_AUTO_MODELS;
158
- }
159
- return [...new Set(list)].filter(Boolean);
160
- }
161
-
162
- // 勾选集 = auto 候选池白名单:只在勾选的模型里择优;空勾选或勾选中无可用模型时回退全量
163
- async function pickedPool(list) {
164
- const picks = loadPicks();
165
- if (!picks.length) return list;
166
- const pickedSet = new Set(picks);
167
- const filtered = list.filter((id) => pickedSet.has(id));
168
- return filtered.length ? filtered : list;
169
- }
170
-
171
- async function candidates() {
172
- const list = await loadList();
173
- const pool = await pickedPool(list);
174
- return rankModels(pool, lastErrorAt, { now: now(), cooldownMs, slowCooldownMs, latencies, preferred: getPreferredModel({ file: file ?? undefined }) });
175
- }
176
-
177
- async function candidatesFor(requested) {
178
- if (!requested) return candidates();
179
- const list = await loadList();
180
- // 显式指定某模型 = 认可它,自动加入勾选集(仅当它是真实上游 free 模型时,避免垃圾 id 污染)
181
- const picks = loadPicks();
182
- if (list.includes(requested) && !picks.includes(requested) && persistPicks) {
183
- await persistPicks([...picks, requested]);
184
- }
185
- const all = list.includes(requested) ? list : [requested, ...list];
186
- const others = rankModels(all.filter((id) => id !== requested), lastErrorAt, {
187
- now: now(),
188
- cooldownMs,
189
- slowCooldownMs,
190
- latencies,
191
- preferred: getPreferredModel({ file: file ?? undefined }),
192
- });
193
- // 显式指定模型:严格优先,永不因冷却被挤到最后(原设计:A deepseek 失败 → B/D deepseek 并发 → 都失败才 fallback)
194
- // 冷却仅影响 auto 的择优,不影响指定模型的“很难被更改”语义
195
- return [requested, ...others];
196
- }
197
-
198
- function isCooling(id) {
199
- return inCooldown(id, lastErrorAt, now(), cooldownMs, slowCooldownMs);
200
- }
201
-
202
- async function recordError(id, evt = {}) {
203
- if (!id) return;
204
- lastErrorAt[id] = {
205
- status: classifyErrorEvent(evt),
206
- at: now(),
207
- code: Number.isInteger(Number(evt.status)) ? Number(evt.status) : null,
208
- slow: Boolean(evt.slow),
209
- };
210
- await persist({ ...lastErrorAt });
211
- }
212
-
213
- async function recordOk(id, evt = {}) {
214
- if (!id) return;
215
- lastErrorAt[id] = { status: MODEL_STATUS.NORMAL, at: now(), code: 200, slow: false };
216
- await persist({ ...lastErrorAt });
217
- const ms = Number(evt.latencyMs ?? evt.totalMs ?? evt.elapsedMs);
218
- if (Number.isFinite(ms) && ms > 0) {
219
- const prev = latencies[id]?.emaMs;
220
- const ema = prev ? Math.round(prev * (1 - latencyAlpha) + ms * latencyAlpha) : Math.round(ms);
221
- latencies[id] = { emaMs: ema, lastMs: Math.round(ms), at: now(), count: (latencies[id]?.count ?? 0) + 1 };
222
- await persistLatencies({ ...latencies });
223
- }
224
- }
225
-
226
- async function recordLatency(id, ms) {
227
- if (!id || !Number.isFinite(ms) || ms <= 0) return;
228
- const prev = latencies[id]?.emaMs;
229
- const ema = prev ? Math.round(prev * (1 - latencyAlpha) + ms * latencyAlpha) : Math.round(ms);
230
- latencies[id] = { emaMs: ema, lastMs: Math.round(ms), at: now(), count: (latencies[id]?.count ?? 0) + 1 };
231
- await persistLatencies({ ...latencies });
232
- }
233
-
234
- function statuses() {
235
- return { ...lastErrorAt };
236
- }
237
-
238
- function latencyStatuses() {
239
- return { ...latencies };
240
- }
241
-
242
- return {
243
- candidates,
244
- candidatesFor,
245
- recordError,
246
- recordOk,
247
- recordLatency,
248
- statuses,
249
- latencyStatuses,
250
- isCooling,
251
- errors: () => ({ ...lastErrorAt }),
252
- latencies: () => ({ ...latencies }),
253
- };
254
- }
1
+ import { statSync } from "node:fs";
2
+ import { loadModelErrors, saveModelErrors, loadModelLatencies, saveModelLatencies, loadPreferredModel, loadModelPicks, saveModelPicks, defaultStateFile } from "./state.js";
3
+
4
+ // 出厂默认首选模型(state.json 的 preferredModel / env MSLXDFF_PREFERRED_MODEL 可覆盖)
5
+ export const DEFAULT_PREFERRED_MODEL = "big-pickle";
6
+ // 兼容旧导出名:语义为"出厂默认",当前生效值请用 getPreferredModel()
7
+ export const PREFERRED_MODEL = DEFAULT_PREFERRED_MODEL;
8
+
9
+ // 当前生效的首选模型:state.json > env > 出厂默认;mtime 缓存保证 daemon 热生效
10
+ const _prefCache = { mtimeMs: -1, file: null, value: null };
11
+ export function getPreferredModel({ file = defaultStateFile() } = {}) {
12
+ try {
13
+ const st = statSync(file);
14
+ if (_prefCache.file !== file || st.mtimeMs !== _prefCache.mtimeMs) {
15
+ _prefCache.file = file;
16
+ _prefCache.mtimeMs = st.mtimeMs;
17
+ _prefCache.value = loadPreferredModel({ file });
18
+ }
19
+ } catch {
20
+ _prefCache.file = file;
21
+ _prefCache.mtimeMs = -1;
22
+ _prefCache.value = null;
23
+ }
24
+ const env = (process.env.MSLXDFF_PREFERRED_MODEL || "").trim();
25
+ return _prefCache.value || env || DEFAULT_PREFERRED_MODEL;
26
+ }
27
+
28
+ export const DEFAULT_AUTO_MODELS = [
29
+ PREFERRED_MODEL,
30
+ "mimo-v2.5-free",
31
+ "deepseek-v4-flash-free",
32
+ "ling-3.0-flash-fin-free",
33
+ "nemotron-3-ultra-free",
34
+ "nemotron-3.5-lightning-free",
35
+ "muse-spark-1.3-contributor-free",
36
+ ].filter((id, i, arr) => id && arr.indexOf(id) === i);
37
+
38
+ export function isAutoModel(model) {
39
+ return !model || model === "auto";
40
+ }
41
+
42
+ export const MODEL_STATUS = Object.freeze({
43
+ NORMAL: "normal",
44
+ LIMIT: "limit",
45
+ ERROR: "error",
46
+ });
47
+
48
+ // Legacy modelErrors entries are bare timestamps ({id: ts}); newer ones are
49
+ // objects ({id: {status, at, code}}). Normalize both to an entry object.
50
+ // `slow` flags a model whose last request was slow (over the wall-clock
51
+ // threshold) — those get a longer cooldown so they lie low until they recover.
52
+ function normEntry(e) {
53
+ if (typeof e === "number") return { status: MODEL_STATUS.ERROR, at: e, code: null, slow: false };
54
+ if (e && typeof e === "object") {
55
+ return {
56
+ status: e.status || MODEL_STATUS.ERROR,
57
+ at: typeof e.at === "number" ? e.at : 0,
58
+ code: e.code ?? null,
59
+ slow: Boolean(e.slow),
60
+ };
61
+ }
62
+ return null;
63
+ }
64
+
65
+ export function classifyErrorEvent(evt = {}) {
66
+ if (evt.slow) return MODEL_STATUS.ERROR;
67
+ const code = Number(evt.status);
68
+ if (code === 429) return MODEL_STATUS.LIMIT;
69
+ const msg = String(evt.message || evt.note || "").toLowerCase();
70
+ if (msg.includes("rate limit") || msg.includes("limit exceeded") || msg.includes("429")) {
71
+ return MODEL_STATUS.LIMIT;
72
+ }
73
+ return MODEL_STATUS.ERROR;
74
+ }
75
+
76
+ export const DEFAULT_COOLDOWN_MS = 60_000;
77
+ export const DEFAULT_SLOW_COOLDOWN_MS = 5 * 60_000;
78
+ export const DEFAULT_LATENCY_ALPHA = 0.3;
79
+
80
+ function effectiveCooldown(entry, slowCooldownMs, cooldownMs) {
81
+ if (entry && entry.slow) return slowCooldownMs || 0;
82
+ return cooldownMs || 0;
83
+ }
84
+
85
+ function inCooldown(id, errors, now, cooldownMs, slowCooldownMs) {
86
+ const e = normEntry(errors[id]);
87
+ if (!e || !(e.at > 0)) return false;
88
+ if (e.status === MODEL_STATUS.NORMAL) return false;
89
+ const cd = effectiveCooldown(e, slowCooldownMs, cooldownMs);
90
+ return cd > 0 && now - e.at < cd;
91
+ }
92
+
93
+ // Latency EMA helpers
94
+ function normLatency(e) {
95
+ if (!e || typeof e !== "object") return null;
96
+ const ema = Number(e.emaMs);
97
+ return Number.isFinite(ema) && ema > 0 ? ema : null;
98
+ }
99
+
100
+ export function rankModels(ids, errors = {}, { now = Date.now(), cooldownMs = 0, slowCooldownMs = 0, latencies = {}, preferred } = {}) {
101
+ const pref = preferred ?? getPreferredModel();
102
+ // 最近一次成功的模型(NORMAL 且非慢且 at 最大),用于“上次成功优先”——慢模型即使刚成功也不应钉死
103
+ let lastSuccessId = null;
104
+ let lastSuccessAt = 0;
105
+ for (const [id, e] of Object.entries(errors)) {
106
+ const ne = normEntry(e);
107
+ if (ne && ne.status === MODEL_STATUS.NORMAL && !ne.slow && ne.at > lastSuccessAt) {
108
+ lastSuccessAt = ne.at;
109
+ lastSuccessId = id;
110
+ }
111
+ }
112
+ return [...new Set(ids)]
113
+ .filter(Boolean)
114
+ .map((id) => ({
115
+ id,
116
+ e: normEntry(errors[id]),
117
+ err: normEntry(errors[id])?.at ?? 0,
118
+ isPreferred: id === pref,
119
+ isLastSuccess: id === lastSuccessId,
120
+ cooling: inCooldown(id, errors, now, cooldownMs, slowCooldownMs),
121
+ latency: normLatency(latencies[id]) ?? Number.MAX_SAFE_INTEGER,
122
+ }))
123
+ .sort(
124
+ (a, b) =>
125
+ (a.cooling ? 1 : 0) - (b.cooling ? 1 : 0) ||
126
+ (b.isLastSuccess ? 1 : 0) - (a.isLastSuccess ? 1 : 0) ||
127
+ (b.isPreferred ? 1 : 0) - (a.isPreferred ? 1 : 0) ||
128
+ a.latency - b.latency ||
129
+ a.err - b.err
130
+ )
131
+ .map((x) => x.id);
132
+ }
133
+
134
+ export function createAutoSelector({
135
+ loadCandidates,
136
+ file,
137
+ now = () => Date.now(),
138
+ cooldownMs = DEFAULT_COOLDOWN_MS,
139
+ slowCooldownMs = DEFAULT_SLOW_COOLDOWN_MS,
140
+ latencyAlpha = DEFAULT_LATENCY_ALPHA,
141
+ errors: seedErrors,
142
+ latencies: seedLatencies,
143
+ persist = (errors, f = file) => saveModelErrors(errors, f ? { file: f } : {}),
144
+ persistLatencies = (latencies, f = file) => saveModelLatencies(latencies, f ? { file: f } : {}),
145
+ loadPicks = () => (file ? loadModelPicks({ file }) : []),
146
+ persistPicks = (picks) => (file ? saveModelPicks(picks, { file }) : picks),
147
+ } = {}) {
148
+ const lastErrorAt = { ...(seedErrors ?? loadModelErrors(file ? { file } : {})) };
149
+ const latencies = { ...(seedLatencies ?? loadModelLatencies(file ? { file } : {})) };
150
+
151
+ async function loadList() {
152
+ let list;
153
+ try {
154
+ const loaded = await loadCandidates?.();
155
+ list = Array.isArray(loaded) && loaded.length ? loaded : DEFAULT_AUTO_MODELS;
156
+ } catch {
157
+ list = DEFAULT_AUTO_MODELS;
158
+ }
159
+ return [...new Set(list)].filter(Boolean);
160
+ }
161
+
162
+ // 勾选集 = auto 候选池白名单:只在勾选的模型里择优;空勾选或勾选中无可用模型时回退全量
163
+ async function pickedPool(list) {
164
+ const picks = loadPicks();
165
+ if (!picks.length) return list;
166
+ const pickedSet = new Set(picks);
167
+ const filtered = list.filter((id) => pickedSet.has(id));
168
+ return filtered.length ? filtered : list;
169
+ }
170
+
171
+ async function candidates() {
172
+ const list = await loadList();
173
+ const pool = await pickedPool(list);
174
+ return rankModels(pool, lastErrorAt, { now: now(), cooldownMs, slowCooldownMs, latencies, preferred: getPreferredModel({ file: file ?? undefined }) });
175
+ }
176
+
177
+ async function candidatesFor(requested) {
178
+ if (!requested) return candidates();
179
+ const list = await loadList();
180
+ // 显式指定某模型 = 认可它,自动加入勾选集(仅当它是真实上游 free 模型时,避免垃圾 id 污染)
181
+ const picks = loadPicks();
182
+ if (list.includes(requested) && !picks.includes(requested) && persistPicks) {
183
+ await persistPicks([...picks, requested]);
184
+ }
185
+ const all = list.includes(requested) ? list : [requested, ...list];
186
+ const others = rankModels(all.filter((id) => id !== requested), lastErrorAt, {
187
+ now: now(),
188
+ cooldownMs,
189
+ slowCooldownMs,
190
+ latencies,
191
+ preferred: getPreferredModel({ file: file ?? undefined }),
192
+ });
193
+ // 显式指定模型:严格优先,永不因冷却被挤到最后(原设计:A deepseek 失败 → B/D deepseek 并发 → 都失败才 fallback)
194
+ // 冷却仅影响 auto 的择优,不影响指定模型的“很难被更改”语义
195
+ return [requested, ...others];
196
+ }
197
+
198
+ function isCooling(id) {
199
+ return inCooldown(id, lastErrorAt, now(), cooldownMs, slowCooldownMs);
200
+ }
201
+
202
+ async function recordError(id, evt = {}) {
203
+ if (!id) return;
204
+ lastErrorAt[id] = {
205
+ status: classifyErrorEvent(evt),
206
+ at: now(),
207
+ code: Number.isInteger(Number(evt.status)) ? Number(evt.status) : null,
208
+ slow: Boolean(evt.slow),
209
+ };
210
+ await persist({ ...lastErrorAt });
211
+ }
212
+
213
+ async function recordOk(id, evt = {}) {
214
+ if (!id) return;
215
+ lastErrorAt[id] = { status: MODEL_STATUS.NORMAL, at: now(), code: 200, slow: false };
216
+ await persist({ ...lastErrorAt });
217
+ const ms = Number(evt.latencyMs ?? evt.totalMs ?? evt.elapsedMs);
218
+ if (Number.isFinite(ms) && ms > 0) {
219
+ const prev = latencies[id]?.emaMs;
220
+ const ema = prev ? Math.round(prev * (1 - latencyAlpha) + ms * latencyAlpha) : Math.round(ms);
221
+ latencies[id] = { emaMs: ema, lastMs: Math.round(ms), at: now(), count: (latencies[id]?.count ?? 0) + 1 };
222
+ await persistLatencies({ ...latencies });
223
+ }
224
+ }
225
+
226
+ async function recordLatency(id, ms) {
227
+ if (!id || !Number.isFinite(ms) || ms <= 0) return;
228
+ const prev = latencies[id]?.emaMs;
229
+ const ema = prev ? Math.round(prev * (1 - latencyAlpha) + ms * latencyAlpha) : Math.round(ms);
230
+ latencies[id] = { emaMs: ema, lastMs: Math.round(ms), at: now(), count: (latencies[id]?.count ?? 0) + 1 };
231
+ await persistLatencies({ ...latencies });
232
+ }
233
+
234
+ function statuses() {
235
+ return { ...lastErrorAt };
236
+ }
237
+
238
+ function latencyStatuses() {
239
+ return { ...latencies };
240
+ }
241
+
242
+ return {
243
+ candidates,
244
+ candidatesFor,
245
+ recordError,
246
+ recordOk,
247
+ recordLatency,
248
+ statuses,
249
+ latencyStatuses,
250
+ isCooling,
251
+ errors: () => ({ ...lastErrorAt }),
252
+ latencies: () => ({ ...latencies }),
253
+ };
254
+ }
@@ -1,42 +1,42 @@
1
- import { clineHeaders } from "../providers/cline/headers.js";
2
- import { computeMetrics } from "../metrics.js";
3
- import { createTransport } from "../transport/index.js";
4
-
5
- function sseContent(obj) {
6
- const c = obj?.choices?.[0]?.delta?.content || obj?.choices?.[0]?.message?.content || "";
7
- return typeof c === "string" ? c : "";
8
- }
9
-
10
- export async function clineBenchOne({ baseUrl, model, accessToken, prompt, maxTokens, timeoutMs, fetchImpl }) {
11
- const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
12
- let ttfbMs = null;
13
- let totalMs = null;
14
- let content = "";
15
- try {
16
- const sid = `sess_bench_${Date.now()}`;
17
- const res = await tr.request({
18
- url: `${baseUrl}/api/v1/chat/completions`,
19
- method: "POST",
20
- headers: { ...clineHeaders(sid, accessToken), Accept: "text/event-stream" },
21
- body: { model, messages: [{ role: "user", content: prompt }], stream: true, max_tokens: maxTokens, session_id: sid, reasoning_effort: "high" },
22
- stream: true,
23
- });
24
- if (!res.ok) {
25
- let txt = "";
26
- try { txt = await res.text(); } catch {}
27
- const label = res.status === 401 ? "鉴权失败" : res.status === 429 ? "限流" : res.status >= 500 ? `上游错误 ${res.status}` : `HTTP ${res.status}`;
28
- return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: res.totalMs, tps: null, charsPerSec: null, tokens: null };
29
- }
30
- for await (const ev of res.stream()) {
31
- try { content += sseContent(JSON.parse(ev)); } catch {}
32
- }
33
- ttfbMs = res.ttfbMs;
34
- totalMs = res.totalMs;
35
- const chars = content.length;
36
- const { tps, charsPerSec } = computeMetrics({ ttfbMs, totalMs, promptTokens: null, completionTokens: null, chars });
37
- return { id: model, ok: true, status: 200, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: null, chars };
38
- } catch (e) {
39
- const msg = e?.message || String(e);
40
- return { id: model, ok: false, label: /timeout|abort/i.test(msg) ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
41
- }
42
- }
1
+ import { clineHeaders } from "../providers/cline/headers.js";
2
+ import { computeMetrics } from "../metrics.js";
3
+ import { createTransport } from "../transport/index.js";
4
+
5
+ function sseContent(obj) {
6
+ const c = obj?.choices?.[0]?.delta?.content || obj?.choices?.[0]?.message?.content || "";
7
+ return typeof c === "string" ? c : "";
8
+ }
9
+
10
+ export async function clineBenchOne({ baseUrl, model, accessToken, prompt, maxTokens, timeoutMs, fetchImpl }) {
11
+ const tr = createTransport({ fetchImpl, keepAlive: false, retry: {}, timeoutMs });
12
+ let ttfbMs = null;
13
+ let totalMs = null;
14
+ let content = "";
15
+ try {
16
+ const sid = `sess_bench_${Date.now()}`;
17
+ const res = await tr.request({
18
+ url: `${baseUrl}/api/v1/chat/completions`,
19
+ method: "POST",
20
+ headers: { ...clineHeaders(sid, accessToken), Accept: "text/event-stream" },
21
+ body: { model, messages: [{ role: "user", content: prompt }], stream: true, max_tokens: maxTokens, session_id: sid, reasoning_effort: "high" },
22
+ stream: true,
23
+ });
24
+ if (!res.ok) {
25
+ let txt = "";
26
+ try { txt = await res.text(); } catch {}
27
+ const label = res.status === 401 ? "鉴权失败" : res.status === 429 ? "限流" : res.status >= 500 ? `上游错误 ${res.status}` : `HTTP ${res.status}`;
28
+ return { id: model, ok: false, status: res.status, label, error: txt.slice(0, 300), ttfbMs, totalMs: res.totalMs, tps: null, charsPerSec: null, tokens: null };
29
+ }
30
+ for await (const ev of res.stream()) {
31
+ try { content += sseContent(JSON.parse(ev)); } catch {}
32
+ }
33
+ ttfbMs = res.ttfbMs;
34
+ totalMs = res.totalMs;
35
+ const chars = content.length;
36
+ const { tps, charsPerSec } = computeMetrics({ ttfbMs, totalMs, promptTokens: null, completionTokens: null, chars });
37
+ return { id: model, ok: true, status: 200, label: "成功", ttfbMs, totalMs, tps, charsPerSec, tokens: null, chars };
38
+ } catch (e) {
39
+ const msg = e?.message || String(e);
40
+ return { id: model, ok: false, label: /timeout|abort/i.test(msg) ? "超时" : "网络错误", error: msg.slice(0, 300), ttfbMs, totalMs, tps: null, charsPerSec: null, tokens: null };
41
+ }
42
+ }