dsh-vision-router 3.0.3 → 3.0.5

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 (35) hide show
  1. package/README.md +183 -162
  2. package/README.zh.md +179 -158
  3. package/cordis.patch.yml +14 -10
  4. package/docs/architecture/dsh-0.2.1-alpha.2-compat.md +41 -0
  5. package/docs/development/dsh-alpha3-compatibility.md +36 -0
  6. package/docs/development/dsh-alpha4-adaptation.md +12 -0
  7. package/docs/development/dsh-alpha4-validation-evidence.md +29 -0
  8. package/docs/development/dsh-alpha5-adaptation.md +31 -0
  9. package/docs/development/dsh-rc1-adaptation.md +29 -0
  10. package/docs/development/dsh-rc2-adaptation.md +24 -0
  11. package/docs/development/v305-real-scenario-acceptance.md +34 -0
  12. package/docs/development/web-live-reload.md +157 -0
  13. package/docs/releases/v3.0.4.md +18 -0
  14. package/docs/releases/v3.0.5.md +31 -0
  15. package/docs/remote-settings.md +4 -0
  16. package/docs/v2-capability-routing.md +25 -9
  17. package/lib/adapter-update-coalescer.js +41 -8
  18. package/lib/client-presentation-boundary-main.js +41 -7
  19. package/lib/client.js +27 -23
  20. package/lib/core-primitives.js +6 -2
  21. package/lib/http-compat.js +33 -2
  22. package/lib/maintenance-routes.js +7 -1
  23. package/lib/session-message-source-compat.js +13 -4
  24. package/lib/sharp-runtime.js +0 -2
  25. package/lib/vision-backend-runtime-policy.js +3 -3
  26. package/lib/vision-background-benchmark.js +33 -5
  27. package/lib/vision-capability-benchmark-service.js +61 -23
  28. package/lib/vision-evidence-guidance.js +7 -1
  29. package/lib/vision-exact-check-client.js +24 -6
  30. package/lib/vision-image-input-verdict.js +5 -3
  31. package/lib/vision-toggle-root-hardening.js +20 -3
  32. package/lib/web-capability-boundary.js +164 -31
  33. package/package.json +6 -4
  34. package/presets/web-connection-ready/index.mjs +48 -0
  35. package/presets/web-connection-ready/package.json +6 -0
@@ -1170,6 +1170,13 @@ export const CLIENT_PRESENTATION_PRELUDE = String.raw`(function(){
1170
1170
  var complete = usable && livePair.mode !== 'unavailable';
1171
1171
  var settledMode = React.useRef(null);
1172
1172
  var toggleTransition = React.useRef(null);
1173
+ // The Host may render a ready directory while the RPC that selected
1174
+ // it is still in flight. Keep only an in-flight transaction fence here,
1175
+ // never a second source of truth for Vision ON/OFF.
1176
+ var selectionTransaction = React.useRef(null);
1177
+ var refreshSelection = React.useState(0)[1];
1178
+ var selectionInFlight = !!(selectionTransaction.current &&
1179
+ selectionTransaction.current.directory === directory);
1173
1180
  if (settledMode.current && settledMode.current.directory !== directory) settledMode.current = null;
1174
1181
  if (toggleTransition.current && toggleTransition.current.directory !== directory) toggleTransition.current = null;
1175
1182
  if (complete) settledMode.current = { directory: directory, mode: livePair.mode };
@@ -1194,14 +1201,25 @@ export const CLIENT_PRESENTATION_PRELUDE = String.raw`(function(){
1194
1201
  var toastState = React.useState(null);
1195
1202
  var toast = toastState[0];
1196
1203
  var setToast = toastState[1];
1197
- var busy = state.status === 'selecting';
1204
+ var busy = state.status === 'selecting' || selectionInFlight;
1198
1205
  var loading = settling;
1206
+ // Host's model-catalog refresh retains the last good current/groups,
1207
+ // but briefly reports 'loading'. When that *complete* pair is still
1208
+ // present, its selection may be sent to Host for fresh validation;
1209
+ // don't flip the native HTML disabled bit in the middle of a user's
1210
+ // pointer gesture (Chromium drops the click even after re-enabling).
1211
+ // A cold load, reconnect, unavailable pair, or selection in flight
1212
+ // remains blocked. We never infer a new mode or optimistically commit.
1213
+ var refreshWithPair = state.status === 'loading' && complete &&
1214
+ !!livePair.target && state.pending == null && state.error == null &&
1215
+ state.routable !== false;
1216
+ var blockingLoad = loading && !refreshWithPair;
1199
1217
  var removed = props.session && props.session.removed === true;
1200
- var disabled = props.available !== true || removed || busy || loading || livePair.mode === 'unavailable';
1218
+ var disabled = props.available !== true || removed || busy || blockingLoad || livePair.mode === 'unavailable';
1201
1219
  var t = typeof props.t === 'function' ? props.t : fallbackTranslate(active);
1202
1220
  var title = busy || ownSettling
1203
1221
  ? t('switching')
1204
- : loading
1222
+ : blockingLoad
1205
1223
  ? t('loading')
1206
1224
  : livePair.mode === 'unavailable'
1207
1225
  ? t('unavailable')
@@ -1229,22 +1247,30 @@ export const CLIENT_PRESENTATION_PRELUDE = String.raw`(function(){
1229
1247
  'data-vision-router-mode-toggle': 'true',
1230
1248
  'aria-pressed': active,
1231
1249
  'aria-label': title,
1232
- 'aria-busy': busy || loading,
1250
+ 'aria-busy': busy || blockingLoad,
1233
1251
  title: title,
1234
1252
  disabled: disabled,
1235
1253
  className: 'vr-vision-toggle',
1236
1254
  'data-active': active ? 'true' : 'false',
1237
- 'data-dimmed': disabled && !ownSettling && (livePair.mode === 'unavailable' || loading) ? 'true' : 'false',
1255
+ 'data-dimmed': disabled && !ownSettling && (livePair.mode === 'unavailable' || blockingLoad) ? 'true' : 'false',
1238
1256
  onClick: function() {
1239
- if (disabled || !pair.target || typeof props.select !== 'function') return;
1257
+ if (disabled || !pair.target || typeof props.select !== 'function' ||
1258
+ (selectionTransaction.current && selectionTransaction.current.directory === directory)) return;
1240
1259
  var target = pair.target;
1260
+ var ticket = { directory: directory, target: target };
1261
+ selectionTransaction.current = ticket;
1262
+ refreshSelection(function(revision){ return revision + 1; });
1241
1263
  toggleTransition.current = {
1242
1264
  directory: directory,
1243
1265
  target: target,
1244
1266
  seenSelecting: false
1245
1267
  };
1246
1268
  setToast(null);
1247
- void props.select(target).then(function(accepted){
1269
+ var pending;
1270
+ try { pending = props.select(target); }
1271
+ catch (error) { pending = Promise.reject(error); }
1272
+ void Promise.resolve(pending).then(function(accepted){
1273
+ if (selectionTransaction.current !== ticket) return;
1248
1274
  if (!accepted) {
1249
1275
  toggleTransition.current = null;
1250
1276
  announceRejectedSelection();
@@ -1255,6 +1281,14 @@ export const CLIENT_PRESENTATION_PRELUDE = String.raw`(function(){
1255
1281
  if (!latest || (latest.status !== 'selecting' && latest.status !== 'idle' && latest.status !== 'loading')) {
1256
1282
  toggleTransition.current = null;
1257
1283
  }
1284
+ }, function(error){
1285
+ if (selectionTransaction.current !== ticket) return;
1286
+ toggleTransition.current = null;
1287
+ announceRejectedSelection();
1288
+ }).finally(function(){
1289
+ if (selectionTransaction.current !== ticket) return;
1290
+ selectionTransaction.current = null;
1291
+ refreshSelection(function(revision){ return revision + 1; });
1258
1292
  });
1259
1293
  }
1260
1294
  },
package/lib/client.js CHANGED
@@ -249,7 +249,7 @@ openLogFolder,
249
249
  builtinFallbackLabel: '内置免费兜底(自动)',
250
250
  builtinFallbackEnabled: '已启用',
251
251
  builtinFallbackDisabled: '已关闭',
252
- builtinFallbackBody: 'OVHcloud 匿名视觉链共 {count} 个模型,首选 {primary}。匿名限额为每 IP、每模型 2 次/分钟;5 个模型独立限额,理论合计约 10 次/分钟,实际以 OVH 限流为准。它固定在上面用户模型之后尝试,免注册、免 Key。',
252
+ builtinFallbackBody: 'OVHcloud 匿名视觉链共 {count} 个模型,首选 {primary}。匿名限额为每 IP、每模型 {perModelRpm} 次/分钟;{count} 个模型独立限额,理论合计约 {totalRpm} 次/分钟,实际以 OVH 限流为准。它固定在上面用户模型之后尝试,免注册、免 Key。',
253
253
  chainLabel: '识图模型',
254
254
  chainHint: '选择 Vision Router 用来读取图片的模型。将按从上到下的顺序尝试,失败时自动切换到下一个。',
255
255
  addFallback: '+ 添加备用识图模型',
@@ -268,18 +268,18 @@ openLogFolder,
268
268
  catalogFallback: '),模型字段已退回手动输入。',
269
269
  catalogPartialFailure: '部分已配置供应商的模型目录加载失败:{detail}。这些供应商可能仍会显示在「设置 → 模型」中,但暂时不会出现在这里。',
270
270
  visionCapsLoading: '正在读取模型图片能力声明(仅用于提示,不影响可选模型)…',
271
- visionCapsError: '视觉能力元数据暂时不可用;模型仍可选择,实际可用性将由供应商已注册的 DSH adapter 在调用时验证。',
272
- visionCapsFiltered: '模型能力声明只用于提示,不再作为准入条件;未声明图片能力或标成仅文本的生成式模型也可选择,调用失败会自动回退。',
273
- visionCapabilityUndeclaredWarning: '此模型暂未附图片能力声明,不影响使用:调用时会实际尝试,失败后自动切换下一个模型。',
274
- visionCapabilityTextOnlyWarning: '此模型被标记为仅文本,仍可直接使用:调用时自动验证图片能力,不支持会自动切换下一个模型。',
275
- visionCapabilityUnknownWarning: '此模型的图片能力会在调用时自动验证,无需额外配置;失败会自动切换下一个模型。',
271
+ visionCapsError: '能力数据不可用,仍可选择模型;实际图片投递取决于Host声明及传输支持。',
272
+ visionCapsFiltered: '能力声明不限制后端选择;Host标为仅文本可能导致图片无法送达。',
273
+ visionCapabilityUndeclaredWarning: '图片能力未知:调用时尝试原生adapter,失败后尝试下一后端。',
274
+ visionCapabilityTextOnlyWarning: 'Host 将此模型标为仅文本,adapter 调用前会剥离图片;仅兼容的 HTTP Chat Completions 可桥接。若支持图片,在 DSH「设置 → 模型 → 输入类型」勾选图片;旧 Host 设 llm-pi-ai 的 input: [text, image] 或 llm-deepseek 的 inputModalities: [text, image] 后重载配置,否则跳过此后端。',
275
+ visionCapabilityUnknownWarning: '图片能力待验证;调用失败会尝试下一后端。',
276
276
  visionCapsEmptyTitle: '没有可选的用户视觉模型',
277
277
  visionCapsEmptyBody: '检测到 {count} 个用户模型,但它们都没有被 DSH 明确标记为支持图片,因此已被安全隐藏。',
278
278
  visionCapsHiddenPrefix: '被隐藏的模型:',
279
279
  visionCapsReasonMissingImage: '未声明 image',
280
280
  visionCapsReasonUnverified: '无法验证图片能力',
281
281
  visionCapsHiddenMore: '另有 {count} 个模型未显示',
282
- visionCapsMissingImageHint: '如果这里有你刚在「设置 → 模型 → 添加自定义提供方」中添加的视觉模型:可以在高级设置「额外视觉模型」的下拉里选中它直接启用;或按 DSH 的方式在 $DSH_HOME/settings.yaml 为该模型补上 input: [text, image](或为整个提供方补 defaultInput: [text, image])。DSH 当前 Web 表单不会写入这个字段。',
282
+ visionCapsMissingImageHint: '若模型确实支持图片,请到 DSH「设置 → 模型 → 输入类型」启用图片并保存;旧 Host 可在 profile 的 cordis.patch.yml 中声明正确的模型输入类型。仅添加额外视觉模型标记不能改变 Host 图片投递。',
283
283
  visionCapsRetry: '重新检测模型',
284
284
  chainInvalidCurrent: '当前保存的后端已不在可调用模型目录中,或属于非生成式/递归路由,运行时会跳过:',
285
285
  retryCatalog: '重试加载目录',
@@ -301,7 +301,7 @@ openLogFolder,
301
301
  extraVisionModelsHint:
302
302
  '这个设置不再用于“解锁”下拉或允许调用;所有可调用生成式模型本来就能被选择。' +
303
303
  '只有当你希望把某个未声明图片能力的模型明确标记为视觉模型时才需要填写。' +
304
- '实际调用仍优先走 DSH adapter;只有明确的 http(s) OpenAI Chat Completions 渠道才可能使用直连兼容兜底。',
304
+ '此标记不会更改 Host 输入声明。仅经验证的 HTTP Chat Completions 渠道可直接桥接;其他纯文本声明会被安全拒绝。',
305
305
  groupRoutes: '路由名称',
306
306
  groupProxy: '网络',
307
307
  networkAuthorityHint: '默认不接管网络出口:Vision Router 沿用 DSH/Host 当前网络路径。下方代理只用于需要给视觉请求单独指定出口的高级覆盖。',
@@ -535,7 +535,7 @@ openLogFolder,
535
535
  builtinFallbackLabel: 'Built-in free fallback (automatic)',
536
536
  builtinFallbackEnabled: 'Enabled',
537
537
  builtinFallbackDisabled: 'Disabled',
538
- builtinFallbackBody: 'The anonymous OVHcloud vision chain contains {count} models, starting with {primary}. Anonymous limits are 2 requests/minute per IP per model; five independent model buckets are about 10 RPM in theory, subject to OVH rate limiting. It always runs after your user models and needs no signup or API key.',
538
+ builtinFallbackBody: 'The anonymous OVHcloud vision chain contains {count} models, starting with {primary}. Anonymous limits are {perModelRpm} requests/minute per IP per model; {count} independent model buckets are about {totalRpm} RPM in theory, subject to OVH rate limiting. It always runs after your user models and needs no signup or API key.',
539
539
  chainLabel: 'Vision model',
540
540
  chainHint: 'Choose the models Vision Router uses to read images. They are tried from top to bottom, with automatic failover.',
541
541
  addFallback: '+ Add fallback vision model',
@@ -554,18 +554,18 @@ openLogFolder,
554
554
  catalogFallback: '); model fields fell back to free-text input.',
555
555
  catalogPartialFailure: 'Some configured providers failed to load their model catalog: {detail}. They may still appear in Settings → Models, but are temporarily unavailable here.',
556
556
  visionCapsLoading: 'Reading model image-capability metadata (advisory only; it does not hide selectable models)…',
557
- visionCapsError: 'Vision capability metadata is temporarily unavailable; models remain selectable and the registered DSH adapter will verify actual support when called.',
558
- visionCapsFiltered: 'Capability metadata is advisory, not an admission gate. Generative models with undeclared or text-only metadata remain selectable and failures automatically fall through.',
559
- visionCapabilityUndeclaredWarning: 'This model has no image-capability declaration, but it stays usable: Vision Router tries it on the call and falls through automatically on failure.',
560
- visionCapabilityTextOnlyWarning: 'This model is marked text-only, but you can still use it: image support is verified on the call and failures fall through automatically.',
561
- visionCapabilityUnknownWarning: 'Image capability is verified automatically on the call — no extra setup needed; failures fall through to the next model.',
557
+ visionCapsError: 'Capabilities unavailable; models remain selectable, but image delivery depends on Host input declaration.',
558
+ visionCapsFiltered: 'Model capability labels do not restrict selection; a Host text-only label may block image delivery.',
559
+ visionCapabilityUndeclaredWarning: 'Image capability unknown; Vision Router tries the adapter and falls through on failure.',
560
+ visionCapabilityTextOnlyWarning: 'Host text-only strips images before adapter; only verified HTTP Chat Completions can bridge. If supported, enable Image in DSH Settings → Models → Input types. Older Hosts: llm-pi-ai input: [text, image]; llm-deepseek inputModalities: [text, image]. Otherwise skip this backend.',
561
+ visionCapabilityUnknownWarning: 'Image capability is checked at call time; failures fall through.',
562
562
  visionCapsEmptyTitle: 'No selectable user vision models',
563
563
  visionCapsEmptyBody: 'DSH reported {count} user models, but none are explicitly marked as accepting images, so Vision Router hid them safely.',
564
564
  visionCapsHiddenPrefix: 'Hidden models:',
565
565
  visionCapsReasonMissingImage: 'image input not declared',
566
566
  visionCapsReasonUnverified: 'image capability could not be verified',
567
567
  visionCapsHiddenMore: '{count} more models not shown',
568
- visionCapsMissingImageHint: 'If one of these is a vision model you just added through Settings → Models → Add custom provider: select it in the “Extra vision models” dropdown under Advanced to enable it directly; or follow the DSH way and add input: [text, image] to that model in $DSH_HOME/settings.yaml (or defaultInput: [text, image] to the provider). The current DSH Web form does not write this field.',
568
+ visionCapsMissingImageHint: 'For a model that truly accepts images, enable Image in DSH Settings → Models → Input types. Older Hosts: set the correct input field in your profile cordis.patch.yml. Extra vision labels cannot change Host pixel delivery.',
569
569
  visionCapsRetry: 'Re-detect models',
570
570
  chainInvalidCurrent: 'This saved backend is no longer callable, or is a non-generative/recursive route, so runtime will skip it: ',
571
571
  retryCatalog: 'Retry catalog',
@@ -587,7 +587,7 @@ openLogFolder,
587
587
  extraVisionModelsHint:
588
588
  'This setting no longer unlocks the picker or admission: every callable generative model is selectable already. ' +
589
589
  'Use it only when you want to explicitly label an undeclared model as visual. ' +
590
- 'Runtime still tries the DSH adapter first; only a confirmed http(s) OpenAI Chat Completions channel may use the direct compatibility bridge.',
590
+ 'This label does not change Host input declarations. Verified HTTP Chat Completions can bridge; other explicit text-only routes fail safely.',
591
591
  groupRoutes: 'Route names',
592
592
  groupProxy: 'Network',
593
593
  networkAuthorityHint: 'Vision Router does not own egress by default: it follows the current DSH/Host network path. The proxy below is only an advanced vision-request override.',
@@ -845,11 +845,12 @@ openLogFolder,
845
845
  function visionCapabilityWarningKey(capability, status) {
846
846
  if (status === 'loading' || status === 'idle') return undefined
847
847
  if (status === 'error' || !capability) return 'visionCapabilityUnknownWarning'
848
- if (capability.attemptable === false || capability.image === true) return undefined
849
- const modalities = Array.isArray(capability.inputModalities) ? capability.inputModalities : []
850
- return modalities.length > 0 && !modalities.includes('image')
851
- ? 'visionCapabilityTextOnlyWarning'
852
- : 'visionCapabilityUndeclaredWarning'
848
+ if (capability.attemptable === false) return undefined
849
+ const declared = Array.isArray(capability.hostInputModalities)
850
+ ? capability.hostInputModalities : capability.inputModalities
851
+ const modalities = Array.isArray(declared) ? declared : []
852
+ if (modalities.length > 0 && !modalities.includes('image')) return 'visionCapabilityTextOnlyWarning'
853
+ return capability.image === true ? undefined : 'visionCapabilityUndeclaredWarning'
853
854
  }
854
855
 
855
856
  // ── field specs ──────────────────────────────────────────────────────────
@@ -3931,13 +3932,16 @@ openLogFolder,
3931
3932
  const list = Array.isArray(visionCaps.builtinFallback) ? visionCaps.builtinFallback : []
3932
3933
  const enabled = format('freeFallback') !== false
3933
3934
  const primary = list[0] && list[0].model ? list[0].model : 'Qwen3.5-397B-A17B'
3934
- const count = list.length > 0 ? list.length : 5
3935
+ const count = list.length > 0 ? list.length : 4
3936
+ const perModelRpm = Number.isFinite(visionCaps.anonymousRpmPerModel)
3937
+ ? Math.max(0, visionCaps.anonymousRpmPerModel)
3938
+ : 2
3935
3939
  return h('div', { className: 'vr-field' },
3936
3940
  h('div', { className: 'vr-field-head' },
3937
3941
  h('span', { className: 'vr-label' }, t('builtinFallbackLabel')),
3938
3942
  h('span', { className: 'vr-badge' }, enabled ? t('builtinFallbackEnabled') : t('builtinFallbackDisabled')),
3939
3943
  ),
3940
- h('p', { className: 'vr-hint' }, t('builtinFallbackBody', { count, primary })),
3944
+ h('p', { className: 'vr-hint' }, t('builtinFallbackBody', { count, primary, perModelRpm, totalRpm: count * perModelRpm })),
3941
3945
  )
3942
3946
  }
3943
3947
  const textProviderEditor = () => {
@@ -1778,12 +1778,12 @@ export async function downscaleImage(bytes, maxPixels, options = {}) {
1778
1778
  */
1779
1779
  export const DEFAULT_HTTP_PROVIDERS = [
1780
1780
  // OVHcloud anonymous quota is per IP AND per model. Keep the free chain
1781
- // ordered largest -> smallest so quality wins first. A 429 on one model can
1781
+ // ordered by preferred quality so quality wins first. Retired models are
1782
+ // removed from this implicit chain; explicit user rows are never rewritten. A 429 on one model can
1782
1783
  // immediately fall through to the next model's independent anonymous bucket.
1783
1784
  { name: 'ovh', baseURL: 'https://oai.endpoints.kepler.ai.cloud.ovh.net/v1', model: 'Qwen3.5-397B-A17B', apiKeyEnv: '', maxTokens: 4096 },
1784
1785
  { name: 'ovh', baseURL: 'https://oai.endpoints.kepler.ai.cloud.ovh.net/v1', model: 'Qwen2.5-VL-72B-Instruct', apiKeyEnv: '', maxTokens: 4096 },
1785
1786
  { name: 'ovh', baseURL: 'https://oai.endpoints.kepler.ai.cloud.ovh.net/v1', model: 'Qwen3.6-27B', apiKeyEnv: '', maxTokens: 4096 },
1786
- { name: 'ovh', baseURL: 'https://oai.endpoints.kepler.ai.cloud.ovh.net/v1', model: 'Mistral-Small-3.2-24B-Instruct-2506', apiKeyEnv: '', maxTokens: 4096 },
1787
1787
  { name: 'ovh', baseURL: 'https://oai.endpoints.kepler.ai.cloud.ovh.net/v1', model: 'Qwen3.5-9B', apiKeyEnv: '', maxTokens: 4096 },
1788
1788
  ]
1789
1789
 
@@ -2768,6 +2768,8 @@ export function decideVisionBackendCapability(info, provider, model, extraVision
2768
2768
  image: true,
2769
2769
  attemptable: true,
2770
2770
  inputModalities: [...new Set([...inputModalities, 'image'])],
2771
+ ...(inputModalities.length > 0 && !inputModalities.includes('image')
2772
+ ? { hostInputModalities: [...inputModalities] } : {}),
2771
2773
  inferred: 'override',
2772
2774
  reason: undefined,
2773
2775
  }
@@ -2789,6 +2791,8 @@ export function decideVisionBackendCapability(info, provider, model, extraVision
2789
2791
  image: true,
2790
2792
  attemptable: true,
2791
2793
  inputModalities: [...new Set([...inputModalities, 'image'])],
2794
+ ...(inputModalities.length > 0 && !inputModalities.includes('image')
2795
+ ? { hostInputModalities: [...inputModalities] } : {}),
2792
2796
  inferred: 'name',
2793
2797
  reason: undefined,
2794
2798
  }
@@ -1,6 +1,9 @@
1
1
 
2
2
  const GLM_4V_FLASH = /^glm-4v-flash(?:$|[-_.])/i
3
3
  export const DEFAULT_HTTP_ERROR_DETAIL_MAX_BYTES = 64 * 1024
4
+ // OVHcloud documents a 10 MB total request-body cap for VLM endpoints.
5
+ // Count actual serialized UTF-8 JSON, not image pixels or pre-base64 bytes.
6
+ export const OVH_VLM_MAX_JSON_BYTES = 10_000_000
4
7
 
5
8
  /**
6
9
  * Known OpenAI-compatible transport quirks. Rules are intentionally sparse:
@@ -110,6 +113,9 @@ export function parseMaxTokensLimit(detail) {
110
113
  return undefined
111
114
  }
112
115
  const patterns = [
116
+ // Two-sided English ranges must use the upper bound, not the first number.
117
+ // The generic max_tokens fallback below would otherwise parse the lower bound.
118
+ /(?:max(?:_completion)?[_ -]?tokens?)[^\d\n]{0,48}(?:between|from|(?:in|within)(?:\s+the)?(?:\s+(?:valid|allowed))?(?:\s+range)?)[^\d\n]{0,12}\d+\s*(?:,|and|to|[-–—])\s*(\d+)/i,
113
119
  /(?:限制数值范围|range)\s*[\[(]\s*\d+\s*,\s*(\d+)\s*[\])]/i,
114
120
  /(?:less than or equal to|at most|maximum(?: allowed)?|max(?:imum)?(?: value)?|<=)\D{0,24}(\d+)/i,
115
121
  /(?:上限|最大(?:值|输出)?)[^\d]{0,16}(\d+)/i,
@@ -306,7 +312,32 @@ export async function fetchWithOpenAICompatibility(fetchImpl, input, init, conte
306
312
  prompt: context.prompt,
307
313
  })
308
314
  let requestBody = prepared.body
309
- let response = await requestFetch(input, { ...init, body: JSON.stringify(requestBody) })
315
+ // The official OVH VLM API caps the *whole HTTP JSON body* at 10 MB.
316
+ // An image under downscaleMaxPixels can still exceed this after base64,
317
+ // and several smaller images can exceed it together. Refuse the doomed
318
+ // request before network egress; never silently degrade OCR quality or
319
+ // apply this vendor-specific limit to a user's other HTTP providers.
320
+ const ovhUrl = asUrl(url)
321
+ const isOfficialOvhVlm = ovhUrl?.protocol === 'https:' &&
322
+ ovhUrl.hostname === 'oai.endpoints.kepler.ai.cloud.ovh.net' &&
323
+ ovhUrl.port === ''
324
+ const serializeWireBody = (body) => {
325
+ const wire = JSON.stringify(body)
326
+ if (isOfficialOvhVlm && Array.isArray(body.messages) &&
327
+ body.messages.some((message) => Array.isArray(message?.content) &&
328
+ message.content.some((block) => block?.type === 'image_url'))) {
329
+ const bytes = Buffer.byteLength(wire, 'utf8')
330
+ if (bytes > OVH_VLM_MAX_JSON_BYTES) {
331
+ const error = new Error(
332
+ `OVHcloud VLM request is ${bytes} bytes, above its 10 MB JSON payload limit; reduce the image file size or image count before retrying`,
333
+ )
334
+ error.code = 'INVALID_REQUEST'
335
+ throw error
336
+ }
337
+ }
338
+ return wire
339
+ }
340
+ let response = await requestFetch(input, { ...init, body: serializeWireBody(requestBody) })
310
341
 
311
342
  for (let correction = 0; correction < 2; correction++) {
312
343
  if (response.ok || (response.status !== 400 && response.status !== 422)) return response
@@ -323,7 +354,7 @@ export async function fetchWithOpenAICompatibility(fetchImpl, input, init, conte
323
354
  }
324
355
 
325
356
  requestBody = retryBody
326
- response = await requestFetch(input, { ...init, body: JSON.stringify(requestBody) })
357
+ response = await requestFetch(input, { ...init, body: serializeWireBody(requestBody) })
327
358
  }
328
359
 
329
360
  return response
@@ -1,5 +1,6 @@
1
1
  import { randomBytes } from 'node:crypto'
2
2
  import { createCachedUpdateChecker } from './update-check.js'
3
+ import { isLocalUiRequest } from './web-capability-boundary.js'
3
4
  import { detectDshSelfUpdatePlan, runDshPluginUpdate } from './self-update.js'
4
5
 
5
6
  function json(res, status, body, headers = {}) {
@@ -66,7 +67,12 @@ export function installVisionMaintenanceRoutes(ctx, options = {}) {
66
67
  res.end()
67
68
  return
68
69
  }
69
- const force = /(?:[?&])force=1(?:&|$)/.test(String(req.url ?? ''))
70
+ // Forced refresh causes outbound Host network traffic. Remote Web may
71
+ // read already-cached update status, but must never bypass the cache
72
+ // with ?force=1. Reuse the existing local-UI provenance boundary;
73
+ // do not trust the query string or caller-supplied proxy headers.
74
+ const requestedForce = /(?:[?&])force=1(?:&|$)/.test(String(req.url ?? ''))
75
+ const force = requestedForce && isLocalUiRequest(req)
70
76
  const result = await updateChecker.check(force)
71
77
  json(res, 200, updateResultForClient(result), { 'cache-control': 'no-store' })
72
78
  },
@@ -23,12 +23,21 @@ export function normalizeVisionRouterMessageSource(message, session) {
23
23
  if (!message || typeof message !== 'object' || Array.isArray(message)) return message
24
24
  if (!isVisionRouterSource(message.source)) return message
25
25
  const source = sourceKindForSession(session)
26
- if (message.source.kind === source.kind
27
- && (source.kind !== 'plugin' || message.source.plugin === source.plugin)
28
- && Object.keys(message.source).length === Object.keys(source).length) {
26
+ // A DSH ContextFormed notice needs both form and summary. Keep those two
27
+ // protocol fields across Session V3/V4 attribution normalization, but do
28
+ // not forward arbitrary source properties from a malformed DVR message.
29
+ const notice = message.source.form === 'notice'
30
+ && typeof message.source.summary === 'string'
31
+ && message.source.summary.trim().length > 0
32
+ const normalized = {
33
+ ...source,
34
+ ...(notice ? { form: 'notice', summary: message.source.summary } : {}),
35
+ }
36
+ if (Object.keys(message.source).length === Object.keys(normalized).length
37
+ && Object.entries(normalized).every(([key, value]) => message.source[key] === value)) {
29
38
  return message
30
39
  }
31
- return { ...message, source: { ...source } }
40
+ return { ...message, source: normalized }
32
41
  }
33
42
 
34
43
  function normalizeMessageList(list, session) {
@@ -2,8 +2,6 @@ import z from '@deepseek-ai/schemastery'
2
2
  import { mkdir, writeFile, readFile, unlink } from 'node:fs/promises'
3
3
  import path from 'node:path'
4
4
  import { tmpdir } from 'node:os'
5
- import { DeepSeekAdapter, resolveAdapterOptions } from '@deepseek-ai/dsh-llm-deepseek'
6
- import { getOrCreateAnonymousUserId } from '@deepseek-ai/dsh-anonymous-user-id'
7
5
  import { existsSync } from 'node:fs'
8
6
  import { execFile } from 'node:child_process'
9
7
  import { Worker } from 'node:worker_threads'
@@ -174,7 +174,7 @@ async function toOpenAIMessages(ctx, messages) {
174
174
  return out
175
175
  }
176
176
 
177
- async function resolvedHostInfo(llm, call) {
177
+ export async function resolvedHostInfo(llm, call) {
178
178
  if (typeof llm?.prepareCall === 'function') {
179
179
  try {
180
180
  const prepared = await llm.prepareCall(
@@ -195,7 +195,7 @@ async function resolvedHostInfo(llm, call) {
195
195
  return undefined
196
196
  }
197
197
 
198
- function localOnlyVisionCallAllowed(core, config, provider, model) {
198
+ export function localOnlyVisionCallAllowed(core, config, provider, model) {
199
199
  if (config?.localOnlyVision !== true) return true
200
200
  if (provider !== 'vision-http') return false
201
201
  const allowed = []
@@ -346,7 +346,7 @@ function llmWithRuntimePolicy(ctx, llm, options) {
346
346
  plan.reason,
347
347
  )
348
348
  yield* failureStream(
349
- `vision backend ${provider}/${model} is declared text-only by the Host and no safe pixel transport is available`,
349
+ `vision backend ${provider}/${model} is declared text-only by the Host, which strips image pixels before the registered adapter runs; no safe HTTP image bridge is available. If this endpoint really supports images, enable Image under DSH Settings > Models > provider > Customized settings > Model options > Input types (when available); on older Hosts set the provider-specific model field (llm-pi-ai: input: [text, image]; llm-deepseek: inputModalities: [text, image]), then reload model configuration. Otherwise choose a declared image-capable backend. Direct bridging is supported only for verified HTTP OpenAI Chat Completions endpoints.`,
350
350
  )
351
351
  return
352
352
  }
@@ -10,6 +10,7 @@ import { resolveVisionRoutingAuthority } from './vision-routing-authority.js'
10
10
  import { redactDiagnosticText } from './diagnostic-redaction.js'
11
11
  import { installVisionExactCheckClient } from './vision-exact-check-client.js'
12
12
  import { createImageInputVerdictStore } from './vision-image-input-verdict.js'
13
+ import { hostImageDeliveryFromInfo, localOnlyVisionCallAllowed, resolvedHostInfo } from './vision-backend-runtime-policy.js'
13
14
  import { createBackgroundBenchmarkStopStore } from './vision-background-stop-store.js'
14
15
  import { backgroundFailurePolicy } from './vision-background-failure-policy.js'
15
16
  import { isLocalUiRequest } from './web-capability-boundary.js'
@@ -288,9 +289,18 @@ function candidateBlockedByFailure(backoff, candidate, at) {
288
289
  return false
289
290
  }
290
291
 
291
- async function imageVerdictFor(imageVerdictStore, candidate) {
292
+ async function currentHostImageDelivery(ctx, provider, model) {
293
+ if (provider === 'vision-http') return 'native-image'
294
+ return hostImageDeliveryFromInfo(await resolvedHostInfo(ctx?.llm, { provider, model }))
295
+ }
296
+
297
+ async function imageVerdictFor(imageVerdictStore, candidate, ctx) {
292
298
  if (!candidate?.endpointFingerprint || typeof imageVerdictStore?.get !== 'function') return undefined
293
- try { return await imageVerdictStore.get(candidate.endpointFingerprint) } catch { return undefined }
299
+ try {
300
+ const verdict = await imageVerdictStore.get(candidate.endpointFingerprint)
301
+ if (!verdict || verdict.hostMode !== 'native-image') return undefined
302
+ return await currentHostImageDelivery(ctx, candidate.provider, candidate.model) === 'native-image' ? verdict : undefined
303
+ } catch { return undefined }
294
304
  }
295
305
 
296
306
  async function chooseNextWork({ ctx, config, core, store, now, backoff, excluded, imageVerdictStore, backgroundStopStore }) {
@@ -302,7 +312,7 @@ async function chooseNextWork({ ctx, config, core, store, now, backoff, excluded
302
312
  await restorePersistentStops(backoff, collected, backgroundStopStore, ctx, now)
303
313
  const candidates = []
304
314
  for (const candidate of collected) {
305
- const verdict = await imageVerdictFor(imageVerdictStore, candidate)
315
+ const verdict = await imageVerdictFor(imageVerdictStore, candidate, ctx)
306
316
  if (verdict?.state === 'unsupported') {
307
317
  excluded.set(candidate.key, { key: candidate.key, reason: 'measured-text-only' })
308
318
  continue
@@ -470,6 +480,13 @@ export async function runExactVisionCheck({ ctx, config, core, store, provider,
470
480
  const wantedModel = typeof model === 'string' ? model.trim() : ''
471
481
  if (!wantedProvider || !wantedModel) throw Object.assign(new Error('provider and model are required'), { code: 'VISION_CHECK_BACKEND_REQUIRED' })
472
482
  const current = activeSettings(ctx, config)
483
+ if (!localOnlyVisionCallAllowed(core, current, wantedProvider, wantedModel)) {
484
+ const error = new Error('local-only vision policy prevents testing a nonlocal backend')
485
+ error.code = 'VISION_LOCAL_ONLY_POLICY'
486
+ error.benchmarkClass = 'host-configuration'
487
+ throw error
488
+ }
489
+ const declaredBefore = await currentHostImageDelivery(ctx, wantedProvider, wantedModel)
473
490
  const candidates = await collectVisionRoutingCandidates(ctx, current, core, store)
474
491
  const candidate = candidates.find((entry) => entry?.provider === wantedProvider && entry?.model === wantedModel)
475
492
  ?? exactAdapterCandidate(ctx, core, wantedProvider, wantedModel)
@@ -513,6 +530,12 @@ export async function runExactVisionCheck({ ctx, config, core, store, provider,
513
530
  remaining(),
514
531
  'exact image check timed out',
515
532
  )
533
+ if (declaredBefore !== await currentHostImageDelivery(ctx, wantedProvider, wantedModel)) {
534
+ const error = new Error('Host image-input declaration changed during the test; retest the model')
535
+ error.code = 'VISION_CHECK_STALE_CONFIG'
536
+ error.benchmarkClass = 'host-configuration'
537
+ throw error
538
+ }
516
539
  return {
517
540
  ok: true,
518
541
  key: candidate.key,
@@ -602,13 +625,15 @@ export function createBackgroundCapabilityProfiler({
602
625
  excluded.set(candidate.key, { key: candidate.key, reason: 'measured-text-only' })
603
626
  clearDeferredForCandidate(candidate)
604
627
  await clearPersistentStopsForCandidate(candidate)
605
- if (candidate.endpointFingerprint && typeof imageVerdictStore?.markUnsupported === 'function') {
628
+ if (candidate.endpointFingerprint && typeof imageVerdictStore?.markUnsupported === 'function'
629
+ && await currentHostImageDelivery(ctx, candidate.provider, candidate.model) === 'native-image') {
606
630
  try {
607
631
  await imageVerdictStore.markUnsupported({
608
632
  fingerprint: candidate.endpointFingerprint,
609
633
  key: candidate.key,
610
634
  provider: candidate.provider,
611
635
  model: candidate.model,
636
+ hostMode: 'native-image',
612
637
  measuredAt: Number(now()),
613
638
  })
614
639
  } catch (error) {
@@ -1140,7 +1165,9 @@ export function installBackgroundCapabilityProfiling(ctx, config, core, store, o
1140
1165
  const timedOut = error?.name === 'TimeoutError' || error?.name === 'AbortError' || /timed out|timeout|deadline/i.test(String(error?.message ?? ''))
1141
1166
  const failureClass = error?.benchmarkClass ?? classifyCapabilityBenchmarkFailure(error)
1142
1167
  const unsupportedImage = !timedOut && failureClass === 'unsupported-image'
1143
- if (unsupportedImage) await profiler.recordImageUnsupported(body.provider, body.model)
1168
+ if (unsupportedImage && await currentHostImageDelivery(ctx, body.provider, body.model) === 'native-image') {
1169
+ await profiler.recordImageUnsupported(body.provider, body.model)
1170
+ }
1144
1171
  const code = timedOut
1145
1172
  ? 'VISION_CHECK_TIMEOUT'
1146
1173
  : unsupportedImage
@@ -1151,6 +1178,7 @@ export function installBackgroundCapabilityProfiling(ctx, config, core, store, o
1151
1178
  'VISION_CHECK_INVALID_JSON',
1152
1179
  'VISION_CHECK_BACKEND_REQUIRED',
1153
1180
  'VISION_CHECK_BACKEND_STALE',
1181
+ 'VISION_CHECK_STALE_CONFIG',
1154
1182
  ]).has(code)
1155
1183
  sendJson(res, unsupportedImage ? 422 : clientError ? 400 : 502, {
1156
1184
  ok: false,