@sixb/azure-ai-foundry 0.2.5 → 0.2.7

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.
package/dist/index.js CHANGED
@@ -5,7 +5,7 @@ import {
5
5
  MODEL_REASONING_LEVELS,
6
6
  ModelCatalogUnavailableError as ModelCatalogUnavailableError4,
7
7
  ModelProviderError as ModelProviderError7,
8
- UnsupportedModelFeatureError as UnsupportedModelFeatureError9
8
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError10
9
9
  } from "@sixb/core/models";
10
10
  import { chatEvents } from "@sixb/model-protocols/chat";
11
11
  import { isStrictToolLimitError, messagesEvents } from "@sixb/model-protocols/messages";
@@ -39,19 +39,20 @@ function foundryUsage(raw, reliableReasoning) {
39
39
  const usage = responsesUsage(raw);
40
40
  const { reasoningOutputTokens, ...totals } = usage;
41
41
  const cached = usage.cacheReadInputTokens;
42
+ const written = counter(object(raw?.input_tokens_details)?.cache_write_tokens);
42
43
  const input = usage.inputTokens;
43
44
  const output = usage.outputTokens;
44
45
  const knownReasoning = reasoningOutputTokens !== undefined && output !== undefined && reasoningOutputTokens <= output && (reliableReasoning === true || reliableReasoning === undefined && reasoningOutputTokens > 0);
45
46
  return {
46
47
  ...totals,
47
- ...object(raw?.input_tokens_details)?.cache_write_tokens === 0 ? { cacheWriteInputTokens: 0 } : {},
48
- ...input !== undefined && cached !== undefined && cached <= input && onlyTokenMeters(raw) ? { uncachedInputTokens: input - cached } : {},
48
+ ...written === undefined ? {} : { cacheWriteInputTokens: written },
49
+ ...input !== undefined && cached !== undefined && cached + (written ?? 0) <= input && onlyTokenMeters(raw) ? { uncachedInputTokens: input - cached - (written ?? 0) } : {},
49
50
  ...knownReasoning ? { reasoningOutputTokens } : {},
50
51
  ...knownReasoning && reasoningOutputTokens !== undefined && output !== undefined && reasoningOutputTokens <= output && onlyTokenMeters(raw) ? { textOutputTokens: output - reasoningOutputTokens } : {}
51
52
  };
52
53
  }
53
54
  function foundryEstimator(card, request, modelName, modelVersion, aliases) {
54
- return tokenEstimator(card, request, ["temperature", "top_p", "metadata"], onlyTokenMeters, modelName, modelVersion, aliases);
55
+ return tokenEstimator(card, request, ["temperature", "top_p", "metadata", "prompt_cache_key", "prompt_cache_retention"], onlyTokenMeters, modelName, modelVersion, aliases);
55
56
  }
56
57
  function tokenEstimator(card, request, requestKeys, tokenMeters, modelName, modelVersion, aliases) {
57
58
  const rateCard = card && defineModelRateCard(card);
@@ -78,7 +79,38 @@ function onlyTokenMeters(raw) {
78
79
  "output_tokens_details"
79
80
  ].includes(key)))
80
81
  return false;
81
- return Object.entries(object(raw.input_tokens_details) ?? {}).every(([key, value]) => key === "cached_tokens" || key === "cache_write_tokens" && value === 0) && Object.keys(object(raw.output_tokens_details) ?? {}).every((key) => key === "reasoning_tokens");
82
+ return Object.entries(object(raw.input_tokens_details) ?? {}).every(([key, value]) => key === "cached_tokens" || key === "cache_write_tokens" || key === "cache_write_tokens_details" && defaultLifetimeWrites(value)) && Object.keys(object(raw.output_tokens_details) ?? {}).every((key) => key === "reasoning_tokens");
83
+ }
84
+ function defaultLifetimeWrites(details) {
85
+ const lifetimes = object(details);
86
+ return lifetimes !== undefined && Object.entries(lifetimes).every(([key, value]) => key === "ttl_30m_tokens" && counter(value) !== undefined || key === "ttl_12h_tokens" && value === 0);
87
+ }
88
+
89
+ // src/caching.ts
90
+ import { billsPromptCacheWrites } from "@sixb/model-protocols/responses";
91
+ var CACHING_FIELDS = ["prompt_cache_key", "prompt_cache_retention", "prompt_cache_options"];
92
+ var refusals = new Set;
93
+ function promptCaching(input) {
94
+ const { caching, deployment } = input;
95
+ const openai = deployment?.modelPublisher.toLowerCase() === "openai";
96
+ const billsWrites = openai && billsPromptCacheWrites(deployment.modelName);
97
+ const reusable = caching === "auto" || caching !== "off" && (input.localTools > 0 || input.messages.some((message) => message.role === "assistant"));
98
+ const intent = typeof caching === "object" ? caching : undefined;
99
+ const settings = {
100
+ ...intent?.key ? { prompt_cache_key: intent.key } : {},
101
+ ...!reusable && billsWrites ? { prompt_cache_options: { mode: "explicit" } } : {},
102
+ ...reusable && openai && !billsWrites && intent?.lifetime === "long" ? { prompt_cache_retention: "24h" } : {}
103
+ };
104
+ return Object.fromEntries(Object.entries(settings).filter(([field]) => !Object.hasOwn(input.request ?? {}, field) && !refusals.has(`${input.scope}\x00${field}`)));
105
+ }
106
+ function refusedCaching(error, body, request, scope) {
107
+ if (error.status !== 400)
108
+ return false;
109
+ const field = CACHING_FIELDS.find((name) => Object.hasOwn(body, name) && !Object.hasOwn(request ?? {}, name) && error.message.includes(name));
110
+ if (!field)
111
+ return false;
112
+ refusals.add(`${scope}\x00${field}`);
113
+ return true;
82
114
  }
83
115
 
84
116
  // src/catalog.ts
@@ -88,7 +120,11 @@ import {
88
120
  MODEL_REASONING_EFFORTS,
89
121
  ModelCatalogUnavailableError
90
122
  } from "@sixb/core/models";
91
- import { downloadModelsDev, modelsDevRateCard } from "@sixb/model-protocols/models-dev";
123
+ import {
124
+ downloadModelsDev,
125
+ modelsDevRateCard,
126
+ scaleTokenPrice
127
+ } from "@sixb/model-protocols/models-dev";
92
128
 
93
129
  // src/publisher.ts
94
130
  var publishers = {
@@ -278,7 +314,11 @@ function rates(raw, protocol) {
278
314
  if (protocol !== "messages" || card?.cacheWriteInput === undefined)
279
315
  return card;
280
316
  const { cacheWriteInput, ...rest } = card;
281
- return defineModelRateCard2({ ...rest, cacheWriteInput5m: cacheWriteInput });
317
+ return defineModelRateCard2({
318
+ ...rest,
319
+ cacheWriteInput5m: cacheWriteInput,
320
+ ...rest.input === undefined ? {} : { cacheWriteInput1h: scaleTokenPrice(rest.input, [2n, 1n]) }
321
+ });
282
322
  }
283
323
 
284
324
  // src/chat-accounting.ts
@@ -286,6 +326,7 @@ function foundryChatUsage(raw, reliableReasoning) {
286
326
  const input = counter(raw?.prompt_tokens);
287
327
  const output = counter(raw?.completion_tokens);
288
328
  const read = counter(object(raw?.prompt_tokens_details)?.cached_tokens);
329
+ const written = counter(object(raw?.prompt_tokens_details)?.cache_write_tokens);
289
330
  const reported = reasoningTokens(raw);
290
331
  const reasoning = reported !== undefined && reported !== false && (reliableReasoning === true || reliableReasoning === undefined && reported > 0) ? reported : undefined;
291
332
  const known = tokenMeters(raw);
@@ -293,15 +334,15 @@ function foundryChatUsage(raw, reliableReasoning) {
293
334
  ...input === undefined ? {} : { inputTokens: input },
294
335
  ...output === undefined ? {} : { outputTokens: output },
295
336
  ...read === undefined ? {} : { cacheReadInputTokens: read },
296
- ...object(raw?.prompt_tokens_details)?.cache_write_tokens === 0 ? { cacheWriteInputTokens: 0 } : {},
297
- ...input !== undefined && read !== undefined && read <= input && known ? { uncachedInputTokens: input - read } : {},
337
+ ...written === undefined ? {} : { cacheWriteInputTokens: written },
338
+ ...input !== undefined && read !== undefined && read + (written ?? 0) <= input && known ? { uncachedInputTokens: input - read - (written ?? 0) } : {},
298
339
  ...reasoning === undefined ? {} : { reasoningOutputTokens: reasoning },
299
340
  ...output !== undefined && reasoning !== undefined && reasoning <= output && known ? { textOutputTokens: output - reasoning } : {},
300
341
  ...raw ? { raw } : {}
301
342
  };
302
343
  }
303
344
  function foundryChatEstimator(card, request, modelName, modelVersion, aliases) {
304
- const estimator = tokenEstimator(card, request, ["temperature", "top_p", "metadata"], tokenMeters, modelName, modelVersion, aliases);
345
+ const estimator = tokenEstimator(card, request, ["temperature", "top_p", "metadata", "prompt_cache_key", "prompt_cache_retention"], tokenMeters, modelName, modelVersion, aliases);
305
346
  return {
306
347
  ...estimator.rateCard ? { rateCard: estimator.rateCard } : {},
307
348
  estimateReservation: estimator.estimateReservation,
@@ -327,9 +368,12 @@ function tokenMeters(raw) {
327
368
  if (!details)
328
369
  return false;
329
370
  for (const [meter, count] of Object.entries(details)) {
330
- if (key === "prompt_tokens_details" && meter === "cached_tokens" || key !== "prompt_tokens_details" && meter === "reasoning_tokens") {
371
+ if (key === "prompt_tokens_details" && (meter === "cached_tokens" || meter === "cache_write_tokens") || key !== "prompt_tokens_details" && meter === "reasoning_tokens") {
331
372
  if (counter(count) === undefined)
332
373
  return false;
374
+ } else if (key === "prompt_tokens_details" && meter === "cache_write_tokens_details") {
375
+ if (!defaultLifetimeWrites(count))
376
+ return false;
333
377
  } else if (key === "prompt_tokens_details" && meter === "image_tokens") {
334
378
  const image = counter(count);
335
379
  const input = counter(raw.prompt_tokens);
@@ -337,7 +381,7 @@ function tokenMeters(raw) {
337
381
  return false;
338
382
  } else if (![
339
383
  "audio_tokens",
340
- ...key === "prompt_tokens_details" ? ["cache_write_tokens"] : ["accepted_prediction_tokens", "rejected_prediction_tokens"]
384
+ ...key === "prompt_tokens_details" ? [] : ["accepted_prediction_tokens", "rejected_prediction_tokens"]
341
385
  ].includes(meter) || count !== 0)
342
386
  return false;
343
387
  }
@@ -526,8 +570,9 @@ function validateOptions(options) {
526
570
  throw new TypeError(`${PREFIX} Invalid reasoningSummary.`);
527
571
  }
528
572
  function responsesRequest(request, definition, options, scope, deployment) {
529
- const max = prepareRequest(request, definition, options, scope);
530
- const input = responsesInput(request.messages, definition.providerId).map((item) => {
573
+ const { max, messages } = prepareRequest(request, definition, options, scope);
574
+ const replayed = responsesInput(messages, definition.providerId, definition.modelId);
575
+ const input = replayed.map((item) => {
531
576
  if (!isJsonObject3(item))
532
577
  return item;
533
578
  const content = Array.isArray(item.content) ? item.content.map((part) => {
@@ -575,6 +620,14 @@ function responsesRequest(request, definition, options, scope, deployment) {
575
620
  stream: true,
576
621
  store: false,
577
622
  ...options.encryptedReasoning === false ? {} : { include: ["reasoning.encrypted_content"] },
623
+ ...promptCaching({
624
+ caching: request.caching,
625
+ messages: request.messages,
626
+ localTools: request.tools.length,
627
+ deployment,
628
+ request: options.request,
629
+ scope
630
+ }),
578
631
  ...max === undefined ? {} : { max_output_tokens: max },
579
632
  ...tools.length ? {
580
633
  tools,
@@ -602,28 +655,44 @@ function responsesRequest(request, definition, options, scope, deployment) {
602
655
  }
603
656
  function prepareRequest(request, definition, options, scope, remotePdf = false) {
604
657
  positiveInteger(request.maxOutputTokens, "maxOutputTokens");
605
- validateMessages(request.messages, definition, options.maxInputFileBytes ?? 20 * 1024 * 1024, scope, remotePdf);
658
+ validateMessages(request.messages, definition, options.maxInputFileBytes ?? 20 * 1024 * 1024, remotePdf);
606
659
  const limits = [
607
660
  definition.maxOutputTokens,
608
661
  options.maxOutputTokens,
609
662
  request.maxOutputTokens
610
663
  ].filter((n) => n !== undefined);
611
- return limits.length ? Math.min(...limits) : undefined;
664
+ return {
665
+ max: limits.length ? Math.min(...limits) : undefined,
666
+ messages: ownReplayState(request.messages, definition.providerId, scope)
667
+ };
612
668
  }
613
- function validateMessages(messages, definition, budget, scope, remotePdf = false) {
669
+ function ownReplayState(messages, providerId, scope) {
670
+ const foreign = (data) => data !== undefined && object(data)?.scope !== scope;
671
+ const own = (part) => {
672
+ if (!part.providerData || !foreign(part.providerData[providerId]))
673
+ return part;
674
+ const { [providerId]: _foreign, ...others } = part.providerData;
675
+ return { ...part, providerData: others };
676
+ };
677
+ return messages.map((message) => {
678
+ if (message.role === "system")
679
+ return own(message);
680
+ if (message.role === "user")
681
+ return { ...message, content: message.content.map(own) };
682
+ if (message.role === "tool")
683
+ return { ...message, content: message.content.map(own) };
684
+ return {
685
+ ...message,
686
+ content: message.content.flatMap((part) => part.type !== "provider-state" ? [own(part)] : part.providerId === providerId && foreign(part.data) ? [] : [part])
687
+ };
688
+ });
689
+ }
690
+ function validateMessages(messages, definition, budget, remotePdf = false) {
614
691
  let bytes = 0;
615
692
  for (const message of messages) {
616
- if (message.role === "system") {
617
- checkScope(message.providerData?.[definition.providerId], scope);
693
+ if (message.role === "system")
618
694
  continue;
619
- }
620
695
  for (const part of message.content) {
621
- if (part.type === "provider-state") {
622
- if (part.providerId === definition.providerId)
623
- checkScope(part.data, scope);
624
- continue;
625
- }
626
- checkScope(part.providerData?.[definition.providerId], scope);
627
696
  if (part.type !== "file")
628
697
  continue;
629
698
  const allowed = definition.capabilities.inputMediaTypes;
@@ -650,11 +719,6 @@ function validateMessages(messages, definition, budget, scope, remotePdf = false
650
719
  }
651
720
  }
652
721
  }
653
- function checkScope(data, scope) {
654
- if (data !== undefined && object(data)?.scope !== scope) {
655
- throw new UnsupportedModelFeatureError(`${PREFIX} Replay state belongs to a different endpoint, deployment, or protocol.`);
656
- }
657
- }
658
722
  function foundryStrictTool(tool) {
659
723
  if (!tool.strict)
660
724
  return;
@@ -702,8 +766,8 @@ function validateChatOptions(options) {
702
766
  if (options.includeUsage !== undefined && typeof options.includeUsage !== "boolean")
703
767
  throw new TypeError(`${PREFIX} includeUsage must be boolean.`);
704
768
  }
705
- function foundryChatRequest(request, definition, options, scope) {
706
- const max = prepareRequest(request, definition, options, scope);
769
+ function foundryChatRequest(request, definition, options, scope, deployment) {
770
+ const { max, messages } = prepareRequest(request, definition, options, scope);
707
771
  const profile = options.profile ?? "openai";
708
772
  if (profile !== "deepseek" && options.reasoningReplay === "tool-continuation")
709
773
  throw new UnsupportedModelFeatureError2(`${PREFIX} Reasoning-content replay requires the DeepSeek Chat profile.`);
@@ -735,7 +799,7 @@ function foundryChatRequest(request, definition, options, scope) {
735
799
  return {
736
800
  ...options.request,
737
801
  model: definition.modelId,
738
- messages: chatInput(request.messages, {
802
+ messages: chatInput(messages, {
739
803
  providerId: definition.providerId,
740
804
  errorPrefix: PREFIX,
741
805
  systemRole: options.systemRole,
@@ -743,6 +807,14 @@ function foundryChatRequest(request, definition, options, scope) {
743
807
  }),
744
808
  stream: true,
745
809
  ...options.includeUsage === false ? {} : { stream_options: { include_usage: true } },
810
+ ...promptCaching({
811
+ caching: request.caching,
812
+ messages: request.messages,
813
+ localTools: request.tools.length,
814
+ deployment,
815
+ request: options.request,
816
+ scope
817
+ }),
746
818
  ...max === undefined ? {} : { [options.maxTokensParameter ?? "max_completion_tokens"]: max },
747
819
  ...explicitReasoning && typeof reasoning === "string" ? { reasoning_effort: reasoning } : {},
748
820
  ...tools.length ? {
@@ -776,7 +848,8 @@ import { systemOneRequest, systemOneResult } from "@sixb/model-protocols/systemo
776
848
  // src/discovery.ts
777
849
  import {
778
850
  ModelCatalogUnavailableError as ModelCatalogUnavailableError2,
779
- ModelProviderError as ModelProviderError3
851
+ ModelProviderError as ModelProviderError3,
852
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError3
780
853
  } from "@sixb/core/models";
781
854
 
782
855
  // src/diagnostics.ts
@@ -826,28 +899,38 @@ var ONE_SHOT = new Set([
826
899
  ]);
827
900
  var AUDIO_API_VERSION = "2024-10-21";
828
901
  var IMAGE_API_VERSION = "2025-04-01-preview";
902
+ var ENTRA_SCOPE = "https://ai.azure.com/.default";
903
+ var RENEW_BEFORE_EXPIRY_MS = 2 * 60000;
904
+ var EXPIRING_MS = 1000;
905
+ var RENEW_RETRY_MS = 3000;
906
+ var TOKEN_TIMEOUT_MS = 30000;
829
907
 
830
908
  class FoundryTransport {
831
909
  diagnostics = new WeakMap;
832
- baseUrl;
833
910
  projectUrl;
911
+ resourceUrl;
834
912
  options;
835
- constructor(options, target = "project") {
913
+ token;
914
+ renewal;
915
+ constructor(options) {
836
916
  if (!URL.canParse(options.endpoint))
837
- throw new TypeError(`${PREFIX} endpoint must be a valid ${target} URL.`);
917
+ throw new TypeError(`${PREFIX} endpoint must be a valid project URL.`);
838
918
  const url = new URL(options.endpoint);
839
919
  if (!["https:", "http:"].includes(url.protocol) || url.username || url.password || url.search || url.hash) {
840
- throw new TypeError(`${PREFIX} endpoint must be an HTTP(S) ${target} URL without credentials, query, or fragment.`);
920
+ throw new TypeError(`${PREFIX} endpoint must be an HTTP(S) project URL without credentials, query, or fragment.`);
841
921
  }
842
922
  const path = url.pathname.replace(/\/+$/, "");
843
- if (target === "project" && !/^\/api\/projects\/[^/]+$/.test(path))
923
+ if (!/^\/api\/projects\/[^/]+$/.test(path))
844
924
  throw new TypeError(`${PREFIX} endpoint must be a Foundry project URL: https://<resource>.services.ai.azure.com/api/projects/<project>.`);
845
- if (target === "resource" && path !== "" && path !== "/openai/v1")
846
- throw new TypeError(`${PREFIX} embeddings.endpoint must be a resource origin or /openai/v1 URL.`);
847
925
  this.projectUrl = `${url.origin}${path}`;
848
- this.baseUrl = target === "project" ? `${this.projectUrl}/openai/v1` : `${url.origin}/openai/v1`;
849
- if (typeof options.apiKey !== "function" && (typeof options.apiKey !== "string" || !options.apiKey.trim()))
850
- throw new TypeError(`${PREFIX} apiKey is required.`);
926
+ this.resourceUrl = url.origin;
927
+ if (options.credential !== undefined) {
928
+ if (options.apiKey !== undefined)
929
+ throw new TypeError(`${PREFIX} Configure either apiKey or credential, not both.`);
930
+ if (typeof options.credential?.getToken !== "function")
931
+ throw new TypeError(`${PREFIX} credential must have a getToken method, such as DefaultAzureCredential from @azure/identity.`);
932
+ } else if (typeof options.apiKey !== "function" && (typeof options.apiKey !== "string" || !options.apiKey.trim()))
933
+ throw new TypeError(`${PREFIX} apiKey or credential is required.`);
851
934
  if (options.maxRetries !== undefined && (!Number.isSafeInteger(options.maxRetries) || options.maxRetries < 0)) {
852
935
  throw new TypeError(`${PREFIX} maxRetries must be a nonnegative safe integer.`);
853
936
  }
@@ -915,35 +998,78 @@ class FoundryTransport {
915
998
  diagnostics.add(value);
916
999
  for (const reserved of ["authorization", "api-key", "x-api-key"])
917
1000
  if (headers.has(reserved))
918
- throw new TypeError(`${PREFIX} Use apiKey instead of '${reserved}' headers.`);
1001
+ throw new TypeError(`${PREFIX} Use apiKey or credential instead of '${reserved}' headers.`);
1002
+ if (this.options.credential) {
1003
+ const token = await this.bearerToken(signal);
1004
+ diagnostics.add(token);
1005
+ headers.set("authorization", `Bearer ${token}`);
1006
+ return headers;
1007
+ }
919
1008
  const key = await resolve(this.options.apiKey, signal);
920
1009
  if (typeof key !== "string" || !key.trim())
921
- throw new TypeError(`${PREFIX} apiKey must resolve to a nonempty string.`);
1010
+ throw new TypeError(`${PREFIX} apiKey must resolve to a nonempty string: pass apiKey or credential, or set AZURE_AI_FOUNDRY_API_KEY.`);
922
1011
  diagnostics.add(key);
923
1012
  headers.set(keyHeader, key);
924
1013
  return headers;
925
1014
  }
1015
+ bearerToken(signal) {
1016
+ const token = this.token;
1017
+ const now = Date.now();
1018
+ if (!token || now >= token.expiresAt - EXPIRING_MS)
1019
+ return abortable(() => this.renewToken(), signal);
1020
+ if (now >= token.renewAt)
1021
+ this.renewToken().catch(() => {
1022
+ if (this.token === token)
1023
+ this.token = { ...token, renewAt: Date.now() + RENEW_RETRY_MS };
1024
+ });
1025
+ return Promise.resolve(token.value);
1026
+ }
1027
+ renewToken() {
1028
+ this.renewal ??= (async () => {
1029
+ const controller = new AbortController;
1030
+ const timer = setTimeout(() => controller.abort(new Error(`${PREFIX} credential.getToken timed out.`)), TOKEN_TIMEOUT_MS);
1031
+ try {
1032
+ const signal = controller.signal;
1033
+ const token = await abortable(() => this.options.credential.getToken(ENTRA_SCOPE, { abortSignal: signal }), signal);
1034
+ if (typeof token?.token !== "string" || !token.token.trim())
1035
+ throw new TypeError(`${PREFIX} credential returned no token for ${ENTRA_SCOPE}.`);
1036
+ const expiresAt = token.expiresOnTimestamp ?? 0;
1037
+ this.token = {
1038
+ value: token.token,
1039
+ expiresAt,
1040
+ renewAt: Math.min(token.refreshAfterTimestamp ?? expiresAt, expiresAt - RENEW_BEFORE_EXPIRY_MS)
1041
+ };
1042
+ return token.token;
1043
+ } finally {
1044
+ clearTimeout(timer);
1045
+ this.renewal = undefined;
1046
+ }
1047
+ })();
1048
+ return this.renewal;
1049
+ }
926
1050
  fetch(url, init) {
927
1051
  return (this.options.fetch ?? fetch)(url, init);
928
1052
  }
929
1053
  url(protocol, deployment) {
930
- if (protocol === "embeddings")
931
- return `${this.baseUrl}/embeddings`;
1054
+ const resource = this.resourceUrl;
1055
+ const path = encodeURIComponent(deployment ?? "");
932
1056
  if (protocol === "transcription")
933
- return `${new URL(this.baseUrl).origin}/openai/deployments/${encodeURIComponent(deployment ?? "")}/audio/transcriptions?api-version=${AUDIO_API_VERSION}`;
934
- if (protocol === "decision")
935
- return `${new URL(this.baseUrl).origin}/providers/microsoft/v1/systemone`;
1057
+ return `${resource}/openai/deployments/${path}/audio/transcriptions?api-version=${AUDIO_API_VERSION}`;
936
1058
  if (protocol === "image" || protocol === "image-edit")
937
- return `${new URL(this.baseUrl).origin}/openai/deployments/${encodeURIComponent(deployment ?? "")}/images/${protocol === "image" ? "generations" : "edits"}?api-version=${IMAGE_API_VERSION}`;
1059
+ return `${resource}/openai/deployments/${path}/images/${protocol === "image" ? "generations" : "edits"}?api-version=${IMAGE_API_VERSION}`;
1060
+ if (protocol === "decision")
1061
+ return `${resource}/providers/microsoft/v1/systemone`;
938
1062
  if (protocol === "reranking")
939
- return `${new URL(this.baseUrl).origin}/providers/cohere/v2/rerank`;
1063
+ return `${resource}/providers/cohere/v2/rerank`;
1064
+ if (protocol === "messages")
1065
+ return `${resource}/anthropic/v1/messages`;
1066
+ if (protocol === "embeddings")
1067
+ return `${resource}/openai/v1/embeddings`;
940
1068
  if (protocol === "speech")
941
- return `${new URL(this.baseUrl).origin}/openai/v1/audio/speech`;
942
- if (protocol === "responses")
943
- return `${this.baseUrl}/responses`;
1069
+ return `${resource}/openai/v1/audio/speech`;
944
1070
  if (protocol === "chat")
945
- return `${this.baseUrl}/chat/completions`;
946
- return `${new URL(this.baseUrl).origin}/anthropic/v1/messages`;
1071
+ return `${resource}/openai/v1/chat/completions`;
1072
+ return `${resource}/openai/v1/responses`;
947
1073
  }
948
1074
  }
949
1075
  async function readJson(response, signal, maxBytes) {
@@ -1108,6 +1234,8 @@ class FoundryDeployments {
1108
1234
  const deployment = snapshot?.byName.get(name);
1109
1235
  if (!snapshot || !deployment)
1110
1236
  throw this.invalid(`Deployment '${name}' was not found in ${offline ? "cached project deployments; resolve online first" : "this project"}.`);
1237
+ if (deployment.connectionName)
1238
+ throw new UnsupportedModelFeatureError3(`${PREFIX} Deployment '${name}' comes from project connection '${deployment.connectionName}', on another resource. Configure a provider with that resource's project URL.`);
1111
1239
  return {
1112
1240
  deployment,
1113
1241
  discoveredAt: snapshot.fetchedAt
@@ -1331,7 +1459,7 @@ function deploymentRecord(raw, invalid) {
1331
1459
  }
1332
1460
 
1333
1461
  // src/kinds.ts
1334
- import { UnsupportedModelFeatureError as UnsupportedModelFeatureError3 } from "@sixb/core/models";
1462
+ import { UnsupportedModelFeatureError as UnsupportedModelFeatureError4 } from "@sixb/core/models";
1335
1463
  var KIND_BINDINGS = {
1336
1464
  language: "foundry() or foundry.chat()",
1337
1465
  embedding: "foundry.embedding()",
@@ -1359,16 +1487,11 @@ function deploymentKind(deployment) {
1359
1487
  function assertDeploymentKind(deployment, deploymentName, kind) {
1360
1488
  const actual = deploymentKind(deployment);
1361
1489
  if (actual !== kind)
1362
- throw new UnsupportedModelFeatureError3(`${PREFIX} Deployment '${deploymentName}' runs ${deployment.modelName}, not ${article(kind)} ${kind} model. Use ${KIND_BINDINGS[actual]}.`);
1363
- if (deployment.connectionName)
1364
- throw new UnsupportedModelFeatureError3(`${PREFIX} ${capitalize(article(kind))} ${kind} deployment from a project connection cannot be routed; use the owning resource's project URL and API key.`);
1490
+ throw new UnsupportedModelFeatureError4(`${PREFIX} Deployment '${deploymentName}' runs ${deployment.modelName}, not ${article(kind)} ${kind} model. Use ${KIND_BINDINGS[actual]}.`);
1365
1491
  }
1366
1492
  function article(kind) {
1367
1493
  return /^[aeiou]/.test(kind) ? "an" : "a";
1368
1494
  }
1369
- function capitalize(word) {
1370
- return word.charAt(0).toUpperCase() + word.slice(1);
1371
- }
1372
1495
  function isEmbeddingDeployment(deployment) {
1373
1496
  return deployment.capabilities.embeddings === "true" || deployment.capabilities.embedding === "true" || /^(text-embedding-|cohere-embed)/i.test(deployment.modelName);
1374
1497
  }
@@ -1490,7 +1613,7 @@ import {
1490
1613
  defineModelRateCard as defineModelRateCard4,
1491
1614
  EmbeddingModelResponseError,
1492
1615
  ModelCatalogUnavailableError as ModelCatalogUnavailableError3,
1493
- UnsupportedModelFeatureError as UnsupportedModelFeatureError4
1616
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError5
1494
1617
  } from "@sixb/core/models";
1495
1618
  function createFoundryEmbedding(providerId, modelId, options, transport, deployments, catalog, resolution) {
1496
1619
  const dimensions = options.dimensions;
@@ -1589,7 +1712,7 @@ function validateDeployment(resolved, dimensions) {
1589
1712
  const name = deployment.modelName.toLowerCase();
1590
1713
  const maxDimensions = name === "text-embedding-3-small" ? 1536 : name === "text-embedding-3-large" ? 3072 : name === "text-embedding-ada-002" ? 1536 : undefined;
1591
1714
  if (deployment.modelPublisher.toLowerCase() !== "openai" || !maxDimensions || deployment.capabilities.embeddings === "false" || deployment.capabilities.embedding === "false")
1592
- throw new UnsupportedModelFeatureError4(`${PREFIX} This embedding deployment is not supported. Use an OpenAI text-embedding-3-small, text-embedding-3-large, or text-embedding-ada-002 deployment.`);
1715
+ throw new UnsupportedModelFeatureError5(`${PREFIX} This embedding deployment is not supported. Use an OpenAI text-embedding-3-small, text-embedding-3-large, or text-embedding-ada-002 deployment.`);
1593
1716
  const configurable = name !== "text-embedding-ada-002";
1594
1717
  if (dimensions > maxDimensions || !configurable && dimensions !== maxDimensions)
1595
1718
  throw new TypeError(`${PREFIX} ${name} requires ${configurable ? `1\u2013${maxDimensions}` : maxDimensions} dimensions.`);
@@ -1673,7 +1796,7 @@ import {
1673
1796
  estimateModelReservation as estimateModelReservation2,
1674
1797
  ImageModelResponseError,
1675
1798
  rateModelCall as rateModelCall2,
1676
- UnsupportedModelFeatureError as UnsupportedModelFeatureError5
1799
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError6
1677
1800
  } from "@sixb/core/models";
1678
1801
  var MAX_RESPONSE_BYTES2 = 512 * 1024 * 1024;
1679
1802
  var MAX_TIMEOUT_MS2 = 2 ** 31 - 1;
@@ -1712,7 +1835,7 @@ function family(deployment) {
1712
1835
  };
1713
1836
  if (name === "flux-1.1-pro")
1714
1837
  return { gpt: false, definition: { maxImagesPerCall: 1 } };
1715
- throw new UnsupportedModelFeatureError5(`${PREFIX} ${deployment.modelName} is not supported: use a GPT Image, FLUX.1 Kontext [pro] or FLUX1.1 [pro] deployment.`);
1838
+ throw new UnsupportedModelFeatureError6(`${PREFIX} ${deployment.modelName} is not supported: use a GPT Image, FLUX.1 Kontext [pro] or FLUX1.1 [pro] deployment.`);
1716
1839
  }
1717
1840
  var gpt = (input, imageOutput, textOutput) => ({
1718
1841
  card: textOutput === undefined ? { currency: "USD", input, output: imageOutput } : { currency: "USD", input, output: textOutput, imageOutput },
@@ -2013,7 +2136,7 @@ function tokenMeters2(raw) {
2013
2136
  const tools = object(value);
2014
2137
  if (!tools || Object.entries(tools).some(([k, v]) => !["web_search_requests", "web_fetch_requests"].includes(k) || v !== 0))
2015
2138
  return false;
2016
- } else if (key === "service_tier" && value === "standard") {} else
2139
+ } else if (key === "service_tier" && value === "standard" || key === "speed" && value === "standard" || key === "inference_geo" && value === "global") {} else
2017
2140
  return false;
2018
2141
  }
2019
2142
  return true;
@@ -2022,10 +2145,11 @@ function tokenMeters2(raw) {
2022
2145
  // src/messages-request.ts
2023
2146
  import {
2024
2147
  modelReasoningSupportIssue as modelReasoningSupportIssue3,
2025
- UnsupportedModelFeatureError as UnsupportedModelFeatureError6
2148
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError7
2026
2149
  } from "@sixb/core/models";
2027
2150
  import {
2028
2151
  messagesInput,
2152
+ messagesPromptCaching,
2029
2153
  messagesStrictTools,
2030
2154
  StrictToolRefusals,
2031
2155
  strictToolsKey
@@ -2140,11 +2264,11 @@ function validateMessagesOptions(options) {
2140
2264
  }
2141
2265
  }
2142
2266
  function foundryMessagesRequest(request, definition, options, scope, unconstrained = false) {
2143
- const max = prepareRequest(request, definition, options, scope, true);
2267
+ const { max, messages } = prepareRequest(request, definition, options, scope, true);
2144
2268
  let conversation = false;
2145
- for (const message of request.messages) {
2269
+ for (const message of messages) {
2146
2270
  if (message.role === "system" && conversation)
2147
- throw new UnsupportedModelFeatureError6(`${PREFIX} Messages system instructions must precede the conversation.`);
2271
+ throw new UnsupportedModelFeatureError7(`${PREFIX} Messages system instructions must precede the conversation.`);
2148
2272
  if (message.role !== "system")
2149
2273
  conversation = true;
2150
2274
  }
@@ -2152,14 +2276,14 @@ function foundryMessagesRequest(request, definition, options, scope, unconstrain
2152
2276
  throw new TypeError(`${PREFIX} Native Messages requires maxOutputTokens in the definition, binding, or request.`);
2153
2277
  const caps = definition.capabilities;
2154
2278
  if (request.tools.length && caps.localTools !== true)
2155
- throw new UnsupportedModelFeatureError6(`${PREFIX} Configure localTools capability before using Messages tools.`);
2279
+ throw new UnsupportedModelFeatureError7(`${PREFIX} Configure localTools capability before using Messages tools.`);
2156
2280
  const schema = request.responseFormat && messagesOutputSchema(request.responseFormat.schema);
2157
2281
  if (request.responseFormat && (caps.nativeStructuredOutput !== true || !schema))
2158
- throw new UnsupportedModelFeatureError6(`${PREFIX} Messages structured output requires a declared capability and a supported closed-object Claude schema.`);
2282
+ throw new UnsupportedModelFeatureError7(`${PREFIX} Messages structured output requires a declared capability and a supported closed-object Claude schema.`);
2159
2283
  const thinking = reasoningRequest(request, definition, options, max);
2160
2284
  if (thinking.thinking && object(thinking.thinking)?.type !== "disabled" && ["temperature", "top_p", "top_k"].some((key) => options.request?.[key] !== undefined))
2161
- throw new UnsupportedModelFeatureError6(`${PREFIX} Explicit sampling controls cannot be combined with Messages thinking.`);
2162
- const mapped = messagesInput(request.messages, definition.providerId, PREFIX);
2285
+ throw new UnsupportedModelFeatureError7(`${PREFIX} Explicit sampling controls cannot be combined with Messages thinking.`);
2286
+ const mapped = messagesInput(messages, definition.providerId, PREFIX);
2163
2287
  const output = {
2164
2288
  ...thinking.effort === undefined ? {} : { effort: thinking.effort },
2165
2289
  ...schema ? { format: { type: "json_schema", schema } } : {}
@@ -2174,7 +2298,7 @@ function foundryMessagesRequest(request, definition, options, scope, unconstrain
2174
2298
  const refused = unconstrained || key !== undefined && strictToolRefusals.has(key);
2175
2299
  const strict = chosen.map((choice) => choice && !refused);
2176
2300
  const strictTools = refused ? undefined : key;
2177
- const body = {
2301
+ const body = messagesPromptCaching({
2178
2302
  ...options.request,
2179
2303
  model: definition.modelId,
2180
2304
  messages: mapped.messages,
@@ -2190,9 +2314,12 @@ function foundryMessagesRequest(request, definition, options, scope, unconstrain
2190
2314
  input_schema: tool.inputSchema,
2191
2315
  ...strict[index] ? { strict: true } : {}
2192
2316
  })),
2193
- tool_choice: { type: "auto", disable_parallel_tool_use: caps.parallelToolCalls !== true }
2317
+ tool_choice: {
2318
+ type: "auto",
2319
+ disable_parallel_tool_use: caps.parallelToolCalls !== true
2320
+ }
2194
2321
  } : {}
2195
- };
2322
+ }, request.caching, request.tools.length);
2196
2323
  return { body, ...strictTools === undefined ? {} : { strictTools } };
2197
2324
  }
2198
2325
  function reasoningRequest(request, definition, options, max) {
@@ -2202,7 +2329,7 @@ function reasoningRequest(request, definition, options, max) {
2202
2329
  const caps = definition.capabilities.reasoning;
2203
2330
  const issue = modelReasoningSupportIssue3(caps, reasoning);
2204
2331
  if (caps === undefined || issue)
2205
- throw new UnsupportedModelFeatureError6(`${PREFIX} ${issue ?? "Configure reasoning capabilities for this deployment"}.`);
2332
+ throw new UnsupportedModelFeatureError7(`${PREFIX} ${issue ?? "Configure reasoning capabilities for this deployment"}.`);
2206
2333
  if (reasoning === "none") {
2207
2334
  return {
2208
2335
  thinking: { type: "disabled" }
@@ -2210,13 +2337,13 @@ function reasoningRequest(request, definition, options, max) {
2210
2337
  }
2211
2338
  if (typeof reasoning === "string") {
2212
2339
  if (reasoning === "minimal" || options.thinkingMode === "manual")
2213
- throw new UnsupportedModelFeatureError6(`${PREFIX} Named efforts require an adaptive-thinking Claude offering.`);
2340
+ throw new UnsupportedModelFeatureError7(`${PREFIX} Named efforts require an adaptive-thinking Claude offering.`);
2214
2341
  return { thinking: { type: "adaptive" }, effort: reasoning };
2215
2342
  }
2216
2343
  if (options.thinkingMode === "adaptive")
2217
- throw new UnsupportedModelFeatureError6(`${PREFIX} This Claude offering only supports adaptive thinking, not manual budgets.`);
2344
+ throw new UnsupportedModelFeatureError7(`${PREFIX} This Claude offering only supports adaptive thinking, not manual budgets.`);
2218
2345
  if (reasoning.budgetTokens < 1024 || reasoning.budgetTokens >= max)
2219
- throw new UnsupportedModelFeatureError6(`${PREFIX} Messages reasoning budget must be at least 1024 and below maxOutputTokens (${max}).`);
2346
+ throw new UnsupportedModelFeatureError7(`${PREFIX} Messages reasoning budget must be at least 1024 and below maxOutputTokens (${max}).`);
2220
2347
  return { thinking: { type: "enabled", budget_tokens: reasoning.budgetTokens } };
2221
2348
  }
2222
2349
 
@@ -2376,7 +2503,7 @@ import {
2376
2503
  defineLanguageModel as defineLanguageModel2,
2377
2504
  defineLanguageModelLabel,
2378
2505
  defineModelRateCard as defineModelRateCard7,
2379
- UnsupportedModelFeatureError as UnsupportedModelFeatureError7
2506
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError8
2380
2507
  } from "@sixb/core/models";
2381
2508
  function resolveModel(input) {
2382
2509
  const { deployment, options } = input;
@@ -2391,12 +2518,10 @@ function resolveModel(input) {
2391
2518
  const entry = input.catalogModel;
2392
2519
  const deploymentPublisher = deployment && azurePublisher(deployment.modelPublisher);
2393
2520
  const publisher = deploymentPublisher?.id === "fireworks-ai" ? entry?.definition.publisher ?? deploymentPublisher : deploymentPublisher ?? entry?.definition.publisher;
2394
- const protocol = input.protocol ?? entry?.protocol ?? (deployment?.capabilities.responses === "true" ? "responses" : deployment?.capabilities.chatCompletion === "true" || deployment?.capabilities.chat_completion === "true" ? "chat" : deployment?.capabilities.messages === "true" ? "messages" : "responses");
2395
- if (protocol === "messages" && deployment?.connectionName)
2396
- throw new UnsupportedModelFeatureError7(`${PREFIX} Native Messages cannot route a deployment from a project connection; use the owning resource's project URL and API key.`);
2521
+ const protocol = input.protocol ?? (deployment ? publisherProtocol(deployment.modelPublisher) : "responses");
2397
2522
  const flag = deployment?.capabilities[protocol === "chat" ? "chatCompletion" : protocol] ?? (protocol === "chat" ? deployment?.capabilities.chat_completion : undefined);
2398
2523
  if (flag === "false")
2399
- throw new UnsupportedModelFeatureError7(`${PREFIX} Deployment '${input.modelId}' reports that ${protocol === "chat" ? "Chat" : protocol === "responses" ? "Responses" : "Messages"} is unsupported.`);
2524
+ throw new UnsupportedModelFeatureError8(`${PREFIX} Deployment '${input.modelId}' reports that ${protocol === "chat" ? "Chat" : protocol === "responses" ? "Responses" : "Messages"} is unsupported.`);
2400
2525
  const definition = defineLanguageModel2({
2401
2526
  ...entry?.definition,
2402
2527
  publisher,
@@ -2447,6 +2572,10 @@ function resolveModel(input) {
2447
2572
  ])
2448
2573
  };
2449
2574
  }
2575
+ function publisherProtocol(publisher) {
2576
+ const name = publisher.toLowerCase();
2577
+ return name === "openai" ? "responses" : name === "anthropic" ? "messages" : "chat";
2578
+ }
2450
2579
  function effectiveDefinition(definition, protocol, options) {
2451
2580
  const caps = definition.capabilities;
2452
2581
  const media = caps.inputMediaTypes;
@@ -2488,7 +2617,7 @@ import {
2488
2617
  defineModelRateCard as defineModelRateCard8,
2489
2618
  rateModelCall as rateModelCall4,
2490
2619
  SpeechModelResponseError,
2491
- UnsupportedModelFeatureError as UnsupportedModelFeatureError8
2620
+ UnsupportedModelFeatureError as UnsupportedModelFeatureError9
2492
2621
  } from "@sixb/core/models";
2493
2622
  var FORMATS2 = {
2494
2623
  "audio/mpeg": "mp3",
@@ -2558,7 +2687,7 @@ function speechBinding(providerId, modelId, settings, transport, deployments, re
2558
2687
  const resolved = await deployments.resolve(modelId, input?.offline === true);
2559
2688
  assertDeploymentKind(resolved.deployment, modelId, "speech");
2560
2689
  if (!/^gpt-4o-mini-tts$/i.test(resolved.deployment.modelName))
2561
- throw new UnsupportedModelFeatureError8(`${PREFIX} ${resolved.deployment.modelName} is not supported: deploy gpt-4o-mini-tts, which reports its usage.`);
2690
+ throw new UnsupportedModelFeatureError9(`${PREFIX} ${resolved.deployment.modelName} is not supported: deploy gpt-4o-mini-tts, which reports its usage.`);
2562
2691
  const rateCard = settings.costEstimator ? undefined : settings.rateCard ?? publishedRateCard2(resolved.deployment);
2563
2692
  return speechBinding(providerId, modelId, settings, transport, deployments, {
2564
2693
  ...resolved,
@@ -2940,19 +3069,23 @@ function transcriptionEstimator(rateCard) {
2940
3069
  }
2941
3070
 
2942
3071
  // src/provider.ts
2943
- function createAzureAIFoundry(options) {
2944
- const transport = new FoundryTransport(options);
3072
+ function createAzureAIFoundry(options = {}) {
3073
+ if ("embeddings" in options)
3074
+ throw new TypeError(`${PREFIX} Remove the embeddings option: embedding deployments are called on the project's resource with its credentials.`);
3075
+ const endpoint = options.endpoint ?? process.env.AZURE_AI_FOUNDRY_ENDPOINT;
3076
+ if (!endpoint)
3077
+ throw new TypeError(`${PREFIX} endpoint is required: pass the project endpoint or set AZURE_AI_FOUNDRY_ENDPOINT.`);
3078
+ const transport = new FoundryTransport({
3079
+ ...options,
3080
+ endpoint,
3081
+ ...options.credential || options.apiKey !== undefined ? {} : { apiKey: () => process.env.AZURE_AI_FOUNDRY_API_KEY }
3082
+ });
2945
3083
  const providerId = options.providerId ?? "azure-ai-foundry";
2946
3084
  if (!providerId.trim())
2947
3085
  throw new TypeError(`${PREFIX} providerId must not be empty.`);
2948
3086
  const deployments = new FoundryDeployments(providerId, transport, options.discovery);
2949
3087
  const modelCatalog = new RemoteModelsDevCatalog(options.catalog);
2950
- const embeddingTransport = options.embeddings ? new FoundryTransport(options.embeddings, "resource") : undefined;
2951
- const embedding = (name, input) => {
2952
- if (!embeddingTransport)
2953
- throw new TypeError(`${PREFIX} Configure embeddings.endpoint and embeddings.apiKey before declaring an embedding model.`);
2954
- return createFoundryEmbedding(providerId, name, input, embeddingTransport, deployments, modelCatalog);
2955
- };
3088
+ const embedding = (name, input) => createFoundryEmbedding(providerId, name, input, transport, deployments, modelCatalog);
2956
3089
  const model = (protocol, deploymentName, modelOptions) => {
2957
3090
  if (!deploymentName.trim() || deploymentName !== deploymentName.trim())
2958
3091
  throw new TypeError(`${PREFIX} deploymentName must be nonempty without surrounding whitespace.`);
@@ -2974,7 +3107,7 @@ function createAzureAIFoundry(options) {
2974
3107
  const results = [];
2975
3108
  for (const deployment of records) {
2976
3109
  const name = deployment.name;
2977
- if (deploymentKind(deployment) !== "language" || await modelCatalog.getEmbedding(deployment.modelName, true))
3110
+ if (deployment.connectionName || deploymentKind(deployment) !== "language" || await modelCatalog.getEmbedding(deployment.modelName, true))
2978
3111
  continue;
2979
3112
  try {
2980
3113
  const profile = resolveModel({
@@ -2986,11 +3119,12 @@ function createAzureAIFoundry(options) {
2986
3119
  catalogModel: lookup?.(deployment.modelName)
2987
3120
  });
2988
3121
  const flags = deployment.capabilities;
2989
- const supported = profile.metadata.catalog || (profile.protocol === "responses" ? flags?.responses === "true" : profile.protocol === "chat" ? flags?.chatCompletion === "true" || flags?.chat_completion === "true" : flags?.messages === "true");
3122
+ const offers = (offered) => offered === "chat" ? flags.chatCompletion === "true" || flags.chat_completion === "true" : flags[offered] === "true";
3123
+ const supported = profile.metadata.catalog || (protocol ? offers(protocol) : ["responses", "chat", "messages"].some(offers));
2990
3124
  if (supported)
2991
3125
  results.push(profile.definition);
2992
3126
  } catch (error) {
2993
- if (!(error instanceof UnsupportedModelFeatureError9))
3127
+ if (!(error instanceof UnsupportedModelFeatureError10))
2994
3128
  throw error;
2995
3129
  }
2996
3130
  }
@@ -3090,11 +3224,7 @@ class FoundryModel {
3090
3224
  ...structuredClone(serializable),
3091
3225
  ...costEstimator ? { costEstimator } : {}
3092
3226
  };
3093
- this.scope = JSON.stringify([
3094
- protocol === "messages" ? transport.url(protocol) : transport.baseUrl,
3095
- this.protocol,
3096
- modelId
3097
- ]);
3227
+ this.scope = JSON.stringify([transport.url(protocol), protocol, modelId]);
3098
3228
  this.reliableReasoning = options.reasoningUsage === "unknown" ? false : options.reasoningUsage === "reported" || this.metadata.publisher?.toLowerCase() === "openai" ? true : undefined;
3099
3229
  this.estimator = options.costEstimator ?? (protocol === "messages" ? foundryMessagesEstimator : protocol === "chat" ? foundryChatEstimator : foundryEstimator)(this.profile.rateCard, this.options.request, this.metadata.modelName, this.metadata.modelVersion, this.profile.aliases);
3100
3230
  Object.freeze(this);
@@ -3105,7 +3235,7 @@ class FoundryModel {
3105
3235
  const resolution = await this.deployments.resolve(this.modelId, options?.offline === true);
3106
3236
  const kind = deploymentKind(resolution.deployment);
3107
3237
  if (kind !== "language")
3108
- throw new UnsupportedModelFeatureError9(`${PREFIX} Use ${KIND_BINDINGS[kind]} for ${article(kind)} ${kind} deployment.`);
3238
+ throw new UnsupportedModelFeatureError10(`${PREFIX} Use ${KIND_BINDINGS[kind]} for ${article(kind)} ${kind} deployment.`);
3109
3239
  let catalogModel;
3110
3240
  try {
3111
3241
  catalogModel = await this.modelCatalog.get(resolution.deployment.modelName, options?.offline);
@@ -3117,7 +3247,7 @@ class FoundryModel {
3117
3247
  throw error;
3118
3248
  }
3119
3249
  if (await this.modelCatalog.getEmbedding(resolution.deployment.modelName, true))
3120
- throw new UnsupportedModelFeatureError9(`${PREFIX} Use foundry.embedding() for an embedding deployment.`);
3250
+ throw new UnsupportedModelFeatureError10(`${PREFIX} Use foundry.embedding() for an embedding deployment.`);
3121
3251
  return new FoundryModel(this.requestedProtocol, this.transport, this.deployments, this.providerId, this.modelId, this.options, this.modelCatalog, {
3122
3252
  ...resolution,
3123
3253
  catalogModel
@@ -3136,7 +3266,7 @@ class FoundryModel {
3136
3266
  const model = await abortable(() => this.#pendingExecution, request.signal);
3137
3267
  return model.stream(request);
3138
3268
  }
3139
- const response = this.protocol === "messages" ? await this.postMessages(request) : await this.transport.post(JSON.stringify(this.protocol === "chat" ? foundryChatRequest(request, this.definition, this.options, this.scope) : responsesRequest(request, this.definition, this.options, this.scope, this.resolution?.deployment)), request.signal, this.providerId, this.modelId, this.protocol);
3269
+ const response = this.protocol === "messages" ? await this.postMessages(request) : await this.postOpenAICompatible(request);
3140
3270
  const id = this.transport.redactText(response, requestId(response));
3141
3271
  if (!response.body)
3142
3272
  throw new ModelProviderError7(`${PREFIX} Provider returned an empty streaming response.`, this.providerId, this.modelId, { requestId: id, status: response.status });
@@ -3174,6 +3304,18 @@ class FoundryModel {
3174
3304
  }), response)
3175
3305
  };
3176
3306
  }
3307
+ async postOpenAICompatible(request) {
3308
+ const deployment = this.resolution?.deployment;
3309
+ for (let attempt = 0;; attempt++) {
3310
+ const body = this.protocol === "chat" ? foundryChatRequest(request, this.definition, this.options, this.scope, deployment) : responsesRequest(request, this.definition, this.options, this.scope, deployment);
3311
+ try {
3312
+ return await this.transport.post(JSON.stringify(body), request.signal, this.providerId, this.modelId, this.protocol);
3313
+ } catch (error) {
3314
+ if (attempt >= 3 || !(error instanceof ModelProviderError7) || !refusedCaching(error, body, this.options.request, this.scope))
3315
+ throw error;
3316
+ }
3317
+ }
3318
+ }
3177
3319
  async postMessages(request) {
3178
3320
  const prepared = foundryMessagesRequest(request, this.definition, this.options, this.scope);
3179
3321
  const post = (body) => this.transport.post(JSON.stringify(body), request.signal, this.providerId, this.modelId, "messages");
@@ -3216,4 +3358,4 @@ export {
3216
3358
  createAzureAIFoundry
3217
3359
  };
3218
3360
 
3219
- //# debugId=9D607CB81AD6E4F564756E2164756E21
3361
+ //# debugId=31DEE47D5A3829A864756E2164756E21