@modelprofile.com/flexharness-providers 13.1.1 → 13.2.0

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.
@@ -1,4 +1,5 @@
1
1
  import * as plugins from './plugins.js';
2
+ import { readOpenAiEffectiveCachePolicy } from './smartai.cache.openai.js';
2
3
  import { classifyOpenAiLimit } from './smartai.limits.openai.js';
3
4
  export { classifyOpenAiLimit, readOpenAiError } from './smartai.limits.openai.js';
4
5
  export { classifyOpenAiChatGptLimit } from './smartai.chatgpt.limits.js';
@@ -22,20 +23,33 @@ export const createOpenAiModelProvider = () => ({
22
23
  const connection = options.connection;
23
24
  // An explicit address, and for a connection an explicit key: the SDK would otherwise take
24
25
  // OPENAI_BASE_URL and OPENAI_API_KEY from the environment. A connection without a key authenticates its own way.
25
- const openai = plugins.createOpenAI(connection
26
+ const settings = connection
26
27
  ? { ...connection.settings, apiKey: connection.settings.apiKey ?? '', baseURL: connection.settings.baseURL || OPENAI_API_BASE_URL }
27
- : { apiKey: options.apiKey, baseURL: options.baseUrl || OPENAI_API_BASE_URL });
28
- const model = api === 'chat' ? openai.chat(options.model) : openai(options.model);
29
- const limits = plugins.createModelLimitMiddleware({
30
- provider: 'openai',
31
- classify: (failure) => connection?.classifyLimit?.(failure) ?? classifyOpenAiLimit(failure),
32
- });
33
- // OpenAI reads only images and PDFs as files, on either API; a text file goes as text
34
- const textFiles = plugins.createTextFilePartMiddleware();
35
- return plugins.wrapLanguageModel({
36
- model,
37
- middleware: connection?.middleware ? [limits, textFiles, connection.middleware] : [limits, textFiles],
38
- });
28
+ : { apiKey: options.apiKey, baseURL: options.baseUrl || OPENAI_API_BASE_URL };
29
+ const createModel = (observer) => {
30
+ const fetchWithObservation = async (input, init) => {
31
+ init?.signal?.throwIfAborted();
32
+ observer({
33
+ dispatchedAt: Date.now(),
34
+ policy: readOpenAiEffectiveCachePolicy(input instanceof Request ? input.url : String(input), init?.body, settings.fetch === undefined),
35
+ });
36
+ return (settings.fetch ?? globalThis.fetch)(input, init);
37
+ };
38
+ const openai = plugins.createOpenAI(observer ? { ...settings, fetch: fetchWithObservation } : settings);
39
+ const model = api === 'chat' ? openai.chat(options.model) : openai(options.model);
40
+ const limits = plugins.createModelLimitMiddleware({
41
+ provider: 'openai',
42
+ classify: (failure) => connection?.classifyLimit?.(failure) ?? classifyOpenAiLimit(failure),
43
+ });
44
+ // OpenAI reads only images and PDFs as files, on either API; a text file goes as text
45
+ const textFiles = plugins.createTextFilePartMiddleware();
46
+ return plugins.wrapLanguageModel({
47
+ model,
48
+ middleware: connection?.middleware ? [limits, textFiles, connection.middleware] : [limits, textFiles],
49
+ });
50
+ };
51
+ // Rebuild the owning pipeline with a request-local observer, as the Anthropic provider does.
52
+ return plugins.registerSmartAiCachePolicyObserver(createModel(), createModel);
39
53
  },
40
54
  });
41
- //# sourceMappingURL=data:application/json;base64,eyJ2ZXJzaW9uIjozLCJmaWxlIjoiaW5kZXguanMiLCJzb3VyY2VSb290IjoiIiwic291cmNlcyI6WyIuLi8uLi90c19wcm92aWRlcnMvb3BlbmFpL2luZGV4LnRzIl0sIm5hbWVzIjpbXSwibWFwcGluZ3MiOiJBQUFBLE9BQU8sS0FBSyxPQUFPLE1BQU0sY0FBYyxDQUFDO0FBQ3hDLE9BQU8sRUFBRSxtQkFBbUIsRUFBRSxNQUFNLDRCQUE0QixDQUFDO0FBRWpFLE9BQU8sRUFBRSxtQkFBbUIsRUFBRSxlQUFlLEVBQTRCLE1BQU0sNEJBQTRCLENBQUM7QUFDNUcsT0FBTyxFQUFFLDBCQUEwQixFQUFFLE1BQU0sNkJBQTZCLENBQUM7QUFDekUsT0FBTyxFQUFFLHlDQUF5QyxFQUFFLHNDQUFzQyxFQUFFLE1BQU0saUNBQWlDLENBQUM7QUFDcEksT0FBTyxFQUFFLHlCQUF5QixFQUFvQyxNQUFNLCtCQUErQixDQUFDO0FBQzVHLE9BQU8sRUFDTCxnQ0FBZ0MsRUFDaEMsd0NBQXdDLEVBQ3hDLDJDQUEyQyxFQUMzQyw2QkFBNkIsRUFDN0IsaUNBQWlDLEVBQ2pDLHNCQUFzQixHQVN2QixNQUFNLDZCQUE2QixDQUFDO0FBQ3JDLE9BQU8sRUFDTCx1QkFBdUIsRUFDdkIscUNBQXFDLEVBQ3JDLDJCQUEyQixHQU81QixNQUFNLDhCQUE4QixDQUFDO0FBZ0N0QyxNQUFNLG1CQUFtQixHQUFHLDJCQUEyQixDQUFDO0FBRXhELE1BQU0sQ0FBQyxNQUFNLHlCQUF5QixHQUFHLEdBQXlFLEVBQUUsQ0FBQyxDQUFDO0lBQ3BILEVBQUUsRUFBRSxRQUFRO0lBQ1osUUFBUSxFQUFFLENBQUMsT0FBTyxFQUFFLEVBQUU7UUFDcEIsTUFBTSxHQUFHLEdBQUcsT0FBTyxDQUFDLEdBQUcsSUFBSSxXQUFXLENBQUM7UUFDdkMsSUFBSSxHQUFHLEtBQUssV0FBVyxJQUFJLEdBQUcsS0FBSyxNQUFNLEVBQUUsQ0FBQztZQUMxQyxNQUFNLElBQUksS0FBSyxDQUFDLHVCQUF1QixNQUFNLENBQUMsR0FBRyxDQUFDLCtCQUErQixDQUFDLENBQUM7UUFDckYsQ0FBQztRQUNELE1BQU0sVUFBVSxHQUFHLE9BQU8sQ0FBQyxVQUFVLEVBQUUsSUFBSSxDQUFDO1FBQzVDLElBQUksVUFBVSxJQUFJLENBQUMsVUFBVSxDQUFDLFFBQVEsQ0FBQyxHQUFHLENBQUMsRUFBRSxDQUFDO1lBQzVDLE1BQU0sTUFBTSxHQUFHLFVBQVUsQ0FBQyxHQUFHLENBQUMsQ0FBQyxLQUFLLEVBQUUsRUFBRSxDQUFDLElBQUksS0FBSyxHQUFHLENBQUMsQ0FBQyxJQUFJLENBQUMsSUFBSSxDQUFDLENBQUM7WUFDbEUsTUFBTSxJQUFJLEtBQUssQ0FBQyxtQ0FBbUMsTUFBTSxjQUFjLEdBQUcsMkJBQTJCLENBQUMsQ0FBQztRQUN6RyxDQUFDO1FBQ0QsTUFBTSxVQUFVLEdBQUcsT0FBTyxDQUFDLFVBQVUsQ0FBQztRQUN0QywwRkFBMEY7UUFDMUYsaUhBQWlIO1FBQ2pILE1BQU0sTUFBTSxHQUFHLE9BQU8sQ0FBQyxZQUFZLENBQUMsVUFBVTtZQUM1QyxDQUFDLENBQUMsRUFBRSxHQUFHLFVBQVUsQ0FBQyxRQUFRLEVBQUUsTUFBTSxFQUFFLFVBQVUsQ0FBQyxRQUFRLENBQUMsTUFBTSxJQUFJLEVBQUUsRUFBRSxPQUFPLEVBQUUsVUFBVSxDQUFDLFFBQVEsQ0FBQyxPQUFPLElBQUksbUJBQW1CLEVBQUU7WUFDbkksQ0FBQyxDQUFDLEVBQUUsTUFBTSxFQUFFLE9BQU8sQ0FBQyxNQUFNLEVBQUUsT0FBTyxFQUFFLE9BQU8sQ0FBQyxPQUFPLElBQUksbUJBQW1CLEVBQUUsQ0FBQyxDQUFDO1FBQ2pGLE1BQU0sS0FBSyxHQUFHLEdBQUcsS0FBSyxNQUFNLENBQUMsQ0FBQyxDQUFDLE1BQU0sQ0FBQyxJQUFJLENBQUMsT0FBTyxDQUFDLEtBQUssQ0FBQyxDQUFDLENBQUMsQ0FBQyxNQUFNLENBQUMsT0FBTyxDQUFDLEtBQUssQ0FBQyxDQUFDO1FBQ2xGLE1BQU0sTUFBTSxHQUFHLE9BQU8sQ0FBQywwQkFBMEIsQ0FBQztZQUNoRCxRQUFRLEVBQUUsUUFBUTtZQUNsQixRQUFRLEVBQUUsQ0FBQyxPQUFPLEVBQUUsRUFBRSxDQUFDLFVBQVUsRUFBRSxhQUFhLEVBQUUsQ0FBQyxPQUFPLENBQUMsSUFBSSxtQkFBbUIsQ0FBQyxPQUFPLENBQUM7U0FDNUYsQ0FBQyxDQUFDO1FBQ0gsc0ZBQXNGO1FBQ3RGLE1BQU0sU0FBUyxHQUFHLE9BQU8sQ0FBQyw0QkFBNEIsRUFBRSxDQUFDO1FBQ3pELE9BQU8sT0FBTyxDQUFDLGlCQUFpQixDQUFDO1lBQy9CLEtBQUs7WUFDTCxVQUFVLEVBQUUsVUFBVSxFQUFFLFVBQVUsQ0FBQyxDQUFDLENBQUMsQ0FBQyxNQUFNLEVBQUUsU0FBUyxFQUFFLFVBQVUsQ0FBQyxVQUFVLENBQUMsQ0FBQyxDQUFDLENBQUMsQ0FBQyxNQUFNLEVBQUUsU0FBUyxDQUFDO1NBQ3RHLENBQUMsQ0FBQztJQUNMLENBQUM7Q0FDRixDQUFDLENBQUMifQ==
55
+ //# sourceMappingURL=data:application/json;base64,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
@@ -4,6 +4,6 @@ export { createOpenAI, openai, wrapLanguageModel };
4
4
  export type { Tool } from 'ai';
5
5
  export type { OpenAIChatLanguageModelOptions as OpenAIChatProviderOptions, OpenAIProviderSettings, OpenAIResponsesProviderOptions, } from '@ai-sdk/openai';
6
6
  export type { LanguageModelV4CallOptions, LanguageModelV4Middleware } from '@ai-sdk/provider';
7
- import { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord } from '@modelprofile.com/flexharness-models';
8
- export { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord };
9
- export type { IModelCallFailure, IModelProvider, ISmartAiOptions, JSONObject, LanguageModelV4, TModelLimitClassifier, } from '@modelprofile.com/flexharness-models';
7
+ import { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord, registerSmartAiCachePolicyObserver } from '@modelprofile.com/flexharness-models';
8
+ export { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord, registerSmartAiCachePolicyObserver, };
9
+ export type { IModelCallFailure, IModelProvider, ISmartAiOpenAiEffectiveCachePolicy, ISmartAiOptions, JSONObject, LanguageModelV4, TModelLimitClassifier, TSmartAiCachePolicyObserver, } from '@modelprofile.com/flexharness-models';
@@ -1,6 +1,6 @@
1
1
  import { createOpenAI, openai } from '@ai-sdk/openai';
2
2
  import { wrapLanguageModel } from 'ai';
3
3
  export { createOpenAI, openai, wrapLanguageModel };
4
- import { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord } from '@modelprofile.com/flexharness-models';
5
- export { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord };
6
- //# sourceMappingURL=data:application/json;base64,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
4
+ import { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord, registerSmartAiCachePolicyObserver, } from '@modelprofile.com/flexharness-models';
5
+ export { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord, registerSmartAiCachePolicyObserver, };
6
+ //# sourceMappingURL=data:application/json;base64,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
@@ -0,0 +1,7 @@
1
+ import type * as plugins from './plugins.js';
2
+ /**
3
+ * Read the serialized request, after the adapter's own transforms. OpenAI's API counts only
4
+ * through a transport known not to change the body; the ChatGPT backend counts through any, since
5
+ * its budget is already the shortest one OpenAI documents.
6
+ */
7
+ export declare function readOpenAiEffectiveCachePolicy(url: string, body: unknown, transparentTransport: boolean): plugins.ISmartAiOpenAiEffectiveCachePolicy | undefined;
@@ -0,0 +1,38 @@
1
+ import { OPENAI_CHATGPT_CODEX_BASE_URL } from './smartai.chatgpt.openai.js';
2
+ const isRecord = (value) => typeof value === 'object' && value !== null && !Array.isArray(value);
3
+ const apiPaths = new Set(['/v1/responses', '/v1/chat/completions']);
4
+ /**
5
+ * Read the serialized request, after the adapter's own transforms. OpenAI's API counts only
6
+ * through a transport known not to change the body; the ChatGPT backend counts through any, since
7
+ * its budget is already the shortest one OpenAI documents.
8
+ */
9
+ export function readOpenAiEffectiveCachePolicy(url, body, transparentTransport) {
10
+ const endpointUrl = new URL(url);
11
+ if (endpointUrl.username !== '' || endpointUrl.password !== '' || typeof body !== 'string')
12
+ return undefined;
13
+ const endpoint = endpointUrl.href.startsWith(`${OPENAI_CHATGPT_CODEX_BASE_URL}/`)
14
+ ? 'chatgpt'
15
+ : transparentTransport && endpointUrl.origin === 'https://api.openai.com' && apiPaths.has(endpointUrl.pathname)
16
+ ? 'api'
17
+ : undefined;
18
+ if (!endpoint)
19
+ return undefined;
20
+ let request;
21
+ try {
22
+ request = JSON.parse(body);
23
+ }
24
+ catch {
25
+ return undefined;
26
+ }
27
+ if (!isRecord(request) || typeof request.model !== 'string')
28
+ return undefined;
29
+ const options = request.prompt_cache_options;
30
+ return {
31
+ provider: 'openai',
32
+ endpoint,
33
+ model: request.model,
34
+ ...(request.prompt_cache_retention === undefined ? {} : { retention: request.prompt_cache_retention }),
35
+ ...(isRecord(options) && options.ttl !== undefined ? { ttl: options.ttl } : {}),
36
+ };
37
+ }
38
+ //# sourceMappingURL=data:application/json;base64,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
package/package.json CHANGED
@@ -2,7 +2,7 @@
2
2
  "author": "Task Venture Capital GmbH",
3
3
  "license": "MIT",
4
4
  "name": "@modelprofile.com/flexharness-providers",
5
- "version": "13.1.1",
5
+ "version": "13.2.0",
6
6
  "type": "module",
7
7
  "description": "Model providers, authentication, account management, media and research integrations in one package.",
8
8
  "main": "./dist_ts_providers/index.js",
@@ -83,7 +83,7 @@
83
83
  "@ai-sdk/provider": "^4.0.17",
84
84
  "@ai-sdk/xai": "^5.0.5",
85
85
  "@anthropic-ai/sdk": "0.128.0",
86
- "@modelprofile.com/flexharness-models": "13.1.1",
86
+ "@modelprofile.com/flexharness-models": "13.2.0",
87
87
  "ai": "^7.0.111",
88
88
  "openai": "^7.22.0",
89
89
  "proper-lockfile": "^4.1.2"
@@ -1,4 +1,5 @@
1
1
  import * as plugins from './plugins.js';
2
+ import { readOpenAiEffectiveCachePolicy } from './smartai.cache.openai.js';
2
3
  import { classifyOpenAiLimit } from './smartai.limits.openai.js';
3
4
 
4
5
  export { classifyOpenAiLimit, readOpenAiError, type IOpenAiErrorDetails } from './smartai.limits.openai.js';
@@ -80,19 +81,36 @@ export const createOpenAiModelProvider = (): plugins.IModelProvider<IOpenAiModel
80
81
  const connection = options.connection;
81
82
  // An explicit address, and for a connection an explicit key: the SDK would otherwise take
82
83
  // OPENAI_BASE_URL and OPENAI_API_KEY from the environment. A connection without a key authenticates its own way.
83
- const openai = plugins.createOpenAI(connection
84
+ const settings: plugins.OpenAIProviderSettings = connection
84
85
  ? { ...connection.settings, apiKey: connection.settings.apiKey ?? '', baseURL: connection.settings.baseURL || OPENAI_API_BASE_URL }
85
- : { apiKey: options.apiKey, baseURL: options.baseUrl || OPENAI_API_BASE_URL });
86
- const model = api === 'chat' ? openai.chat(options.model) : openai(options.model);
87
- const limits = plugins.createModelLimitMiddleware({
88
- provider: 'openai',
89
- classify: (failure) => connection?.classifyLimit?.(failure) ?? classifyOpenAiLimit(failure),
90
- });
91
- // OpenAI reads only images and PDFs as files, on either API; a text file goes as text
92
- const textFiles = plugins.createTextFilePartMiddleware();
93
- return plugins.wrapLanguageModel({
94
- model,
95
- middleware: connection?.middleware ? [limits, textFiles, connection.middleware] : [limits, textFiles],
96
- });
86
+ : { apiKey: options.apiKey, baseURL: options.baseUrl || OPENAI_API_BASE_URL };
87
+ const createModel = (observer?: plugins.TSmartAiCachePolicyObserver): plugins.LanguageModelV4 => {
88
+ const fetchWithObservation: NonNullable<plugins.OpenAIProviderSettings['fetch']> = async (input, init) => {
89
+ init?.signal?.throwIfAborted();
90
+ observer!({
91
+ dispatchedAt: Date.now(),
92
+ policy: readOpenAiEffectiveCachePolicy(
93
+ input instanceof Request ? input.url : String(input),
94
+ init?.body,
95
+ settings.fetch === undefined,
96
+ ),
97
+ });
98
+ return (settings.fetch ?? globalThis.fetch)(input, init);
99
+ };
100
+ const openai = plugins.createOpenAI(observer ? { ...settings, fetch: fetchWithObservation } : settings);
101
+ const model = api === 'chat' ? openai.chat(options.model) : openai(options.model);
102
+ const limits = plugins.createModelLimitMiddleware({
103
+ provider: 'openai',
104
+ classify: (failure) => connection?.classifyLimit?.(failure) ?? classifyOpenAiLimit(failure),
105
+ });
106
+ // OpenAI reads only images and PDFs as files, on either API; a text file goes as text
107
+ const textFiles = plugins.createTextFilePartMiddleware();
108
+ return plugins.wrapLanguageModel({
109
+ model,
110
+ middleware: connection?.middleware ? [limits, textFiles, connection.middleware] : [limits, textFiles],
111
+ });
112
+ };
113
+ // Rebuild the owning pipeline with a request-local observer, as the Anthropic provider does.
114
+ return plugins.registerSmartAiCachePolicyObserver(createModel(), createModel);
97
115
  },
98
116
  });
@@ -8,13 +8,27 @@ export type {
8
8
  OpenAIResponsesProviderOptions,
9
9
  } from '@ai-sdk/openai';
10
10
  export type { LanguageModelV4CallOptions, LanguageModelV4Middleware } from '@ai-sdk/provider';
11
- import { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord } from '@modelprofile.com/flexharness-models';
12
- export { createModelLimitMiddleware, createTextFilePartMiddleware, readEpochSeconds, readJsonRecord };
11
+ import {
12
+ createModelLimitMiddleware,
13
+ createTextFilePartMiddleware,
14
+ readEpochSeconds,
15
+ readJsonRecord,
16
+ registerSmartAiCachePolicyObserver,
17
+ } from '@modelprofile.com/flexharness-models';
18
+ export {
19
+ createModelLimitMiddleware,
20
+ createTextFilePartMiddleware,
21
+ readEpochSeconds,
22
+ readJsonRecord,
23
+ registerSmartAiCachePolicyObserver,
24
+ };
13
25
  export type {
14
26
  IModelCallFailure,
15
27
  IModelProvider,
28
+ ISmartAiOpenAiEffectiveCachePolicy,
16
29
  ISmartAiOptions,
17
30
  JSONObject,
18
31
  LanguageModelV4,
19
32
  TModelLimitClassifier,
33
+ TSmartAiCachePolicyObserver,
20
34
  } from '@modelprofile.com/flexharness-models';
@@ -120,6 +120,21 @@ The `/auth` subpath keeps its connections, which derive identity from the token
120
120
  re-exports the ChatGPT limit classifier, the instructions middleware, the Codex base URL, the
121
121
  default originator and `OpenAiChatGptAuthError` from here.
122
122
 
123
+ ## Prompt cache timing
124
+
125
+ Each model call reports the cache settings of the request it sent to an observer
126
+ of `@modelprofile.com/flexharness-models` (`bindSmartAiCachePolicyObserver`): the
127
+ endpoint (`api` for `https://api.openai.com/v1`, `chatgpt` for the ChatGPT Codex
128
+ backend), the `model`, and `prompt_cache_retention` and
129
+ `prompt_cache_options.ttl` when sent. `observeSmartAiCacheDispatch` turns that
130
+ into the documented idle budget: 30 minutes for GPT-5.6 and later on the API, 30
131
+ minutes for `24h` retention on earlier models, and 5 minutes for in-memory
132
+ retention, for no retention and for every ChatGPT request, since that backend
133
+ takes no retention and documents none. API requests count only without a custom
134
+ `fetch`, which may change the body; other endpoints report nothing. The
135
+ response's `input_tokens_details.cached_tokens` arrives as the usage's
136
+ `cacheReadTokens`.
137
+
123
138
  ## Text files
124
139
 
125
140
  OpenAI reads only images and PDFs as files, so the provider sends every other text file of a user
@@ -0,0 +1,38 @@
1
+ import type * as plugins from './plugins.js';
2
+ import { OPENAI_CHATGPT_CODEX_BASE_URL } from './smartai.chatgpt.openai.js';
3
+
4
+ const isRecord = (value: unknown): value is Record<string, unknown> =>
5
+ typeof value === 'object' && value !== null && !Array.isArray(value);
6
+
7
+ const apiPaths = new Set(['/v1/responses', '/v1/chat/completions']);
8
+
9
+ /**
10
+ * Read the serialized request, after the adapter's own transforms. OpenAI's API counts only
11
+ * through a transport known not to change the body; the ChatGPT backend counts through any, since
12
+ * its budget is already the shortest one OpenAI documents.
13
+ */
14
+ export function readOpenAiEffectiveCachePolicy(
15
+ url: string,
16
+ body: unknown,
17
+ transparentTransport: boolean,
18
+ ): plugins.ISmartAiOpenAiEffectiveCachePolicy | undefined {
19
+ const endpointUrl = new URL(url);
20
+ if (endpointUrl.username !== '' || endpointUrl.password !== '' || typeof body !== 'string') return undefined;
21
+ const endpoint = endpointUrl.href.startsWith(`${OPENAI_CHATGPT_CODEX_BASE_URL}/`)
22
+ ? 'chatgpt'
23
+ : transparentTransport && endpointUrl.origin === 'https://api.openai.com' && apiPaths.has(endpointUrl.pathname)
24
+ ? 'api'
25
+ : undefined;
26
+ if (!endpoint) return undefined;
27
+ let request: unknown;
28
+ try { request = JSON.parse(body); } catch { return undefined; }
29
+ if (!isRecord(request) || typeof request.model !== 'string') return undefined;
30
+ const options = request.prompt_cache_options;
31
+ return {
32
+ provider: 'openai',
33
+ endpoint,
34
+ model: request.model,
35
+ ...(request.prompt_cache_retention === undefined ? {} : { retention: request.prompt_cache_retention }),
36
+ ...(isRecord(options) && options.ttl !== undefined ? { ttl: options.ttl } : {}),
37
+ };
38
+ }