@providerkit/core 0.30.0 → 0.32.0
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.
- package/README.md +101 -5
- package/THIRD_PARTY_NOTICES +144 -0
- package/dist/auth/chatgpt-models.js.map +1 -1
- package/dist/auth/oauth.d.ts.map +1 -1
- package/dist/auth/oauth.js.map +1 -1
- package/dist/capability.js.map +1 -1
- package/dist/context.d.ts +4 -3
- package/dist/context.d.ts.map +1 -1
- package/dist/context.js +19 -5
- package/dist/context.js.map +1 -1
- package/dist/decisions.js.map +1 -1
- package/dist/errors.d.ts.map +1 -1
- package/dist/errors.js.map +1 -1
- package/dist/fallback.d.ts +3 -0
- package/dist/fallback.d.ts.map +1 -1
- package/dist/fallback.js +9 -5
- package/dist/fallback.js.map +1 -1
- package/dist/gate.js.map +1 -1
- package/dist/hosted/anthropic.d.ts +30 -0
- package/dist/hosted/anthropic.d.ts.map +1 -0
- package/dist/hosted/anthropic.js +465 -0
- package/dist/hosted/anthropic.js.map +1 -0
- package/dist/hosted/gemini.d.ts +22 -0
- package/dist/hosted/gemini.d.ts.map +1 -0
- package/dist/hosted/gemini.js +440 -0
- package/dist/hosted/gemini.js.map +1 -0
- package/dist/hosted/openai.d.ts +33 -0
- package/dist/hosted/openai.d.ts.map +1 -0
- package/dist/hosted/openai.js +107 -0
- package/dist/hosted/openai.js.map +1 -0
- package/dist/hosted/responses.d.ts +28 -0
- package/dist/hosted/responses.d.ts.map +1 -0
- package/dist/hosted/responses.js +358 -0
- package/dist/hosted/responses.js.map +1 -0
- package/dist/hosted/shared.d.ts +107 -0
- package/dist/hosted/shared.d.ts.map +1 -0
- package/dist/hosted/shared.js +264 -0
- package/dist/hosted/shared.js.map +1 -0
- package/dist/hosted/types.d.ts +191 -0
- package/dist/hosted/types.d.ts.map +1 -0
- package/dist/hosted/types.js +10 -0
- package/dist/hosted/types.js.map +1 -0
- package/dist/hosted.d.ts +33 -0
- package/dist/hosted.d.ts.map +1 -0
- package/dist/hosted.js +61 -0
- package/dist/hosted.js.map +1 -0
- package/dist/index.d.ts +8 -7
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +8 -7
- package/dist/index.js.map +1 -1
- package/dist/key-pool.d.ts.map +1 -1
- package/dist/key-pool.js.map +1 -1
- package/dist/otel.d.ts +26 -0
- package/dist/otel.d.ts.map +1 -0
- package/dist/otel.js +263 -0
- package/dist/otel.js.map +1 -0
- package/dist/presets.d.ts +2 -2
- package/dist/presets.d.ts.map +1 -1
- package/dist/providers/anthropic.d.ts +75 -1
- package/dist/providers/anthropic.d.ts.map +1 -1
- package/dist/providers/anthropic.js +207 -135
- package/dist/providers/anthropic.js.map +1 -1
- package/dist/providers/factory.d.ts +10 -0
- package/dist/providers/factory.d.ts.map +1 -1
- package/dist/providers/factory.js +31 -23
- package/dist/providers/factory.js.map +1 -1
- package/dist/providers/gemini.d.ts +123 -1
- package/dist/providers/gemini.d.ts.map +1 -1
- package/dist/providers/gemini.js +114 -54
- package/dist/providers/gemini.js.map +1 -1
- package/dist/providers/hosted-codecs.d.ts +10 -0
- package/dist/providers/hosted-codecs.d.ts.map +1 -0
- package/dist/providers/hosted-codecs.js +8 -0
- package/dist/providers/hosted-codecs.js.map +1 -0
- package/dist/providers/openai.d.ts +69 -0
- package/dist/providers/openai.d.ts.map +1 -1
- package/dist/providers/openai.js +38 -14
- package/dist/providers/openai.js.map +1 -1
- package/dist/providers/responses.d.ts +123 -0
- package/dist/providers/responses.d.ts.map +1 -1
- package/dist/providers/responses.js +57 -22
- package/dist/providers/responses.js.map +1 -1
- package/dist/providers/speech-factory.js.map +1 -1
- package/dist/registry.d.ts +3 -0
- package/dist/registry.d.ts.map +1 -1
- package/dist/registry.js +8 -4
- package/dist/registry.js.map +1 -1
- package/dist/retry.js.map +1 -1
- package/dist/speech-core.d.ts.map +1 -1
- package/dist/speech-core.js.map +1 -1
- package/dist/tools.d.ts.map +1 -1
- package/dist/tools.js.map +1 -1
- package/dist/transcription-core.d.ts.map +1 -1
- package/dist/transcription-core.js.map +1 -1
- package/dist/transport.js.map +1 -1
- package/dist/turn.d.ts +10 -1
- package/dist/turn.d.ts.map +1 -1
- package/dist/turn.js +53 -1
- package/dist/turn.js.map +1 -1
- package/dist/types.d.ts +51 -0
- package/dist/types.d.ts.map +1 -1
- package/dist/types.js.map +1 -1
- package/dist/usage.d.ts +10 -1
- package/dist/usage.d.ts.map +1 -1
- package/dist/usage.js +40 -2
- package/dist/usage.js.map +1 -1
- package/dist/watchdog.d.ts +6 -0
- package/dist/watchdog.d.ts.map +1 -1
- package/dist/watchdog.js +10 -2
- package/dist/watchdog.js.map +1 -1
- package/package.json +19 -5
- package/src/context.ts +21 -5
- package/src/fallback.ts +21 -6
- package/src/hosted/anthropic.ts +512 -0
- package/src/hosted/gemini.ts +510 -0
- package/src/hosted/openai.ts +137 -0
- package/src/hosted/responses.ts +420 -0
- package/src/hosted/shared.ts +341 -0
- package/src/hosted/types.ts +214 -0
- package/src/hosted.ts +101 -0
- package/src/index.ts +64 -7
- package/src/otel.ts +306 -0
- package/src/providers/anthropic.ts +252 -143
- package/src/providers/factory.ts +127 -82
- package/src/providers/gemini.ts +184 -60
- package/src/providers/hosted-codecs.ts +29 -0
- package/src/providers/openai.ts +52 -15
- package/src/providers/responses.ts +93 -24
- package/src/registry.ts +14 -4
- package/src/turn.ts +55 -1
- package/src/types.ts +62 -1
- package/src/usage.ts +32 -2
- package/src/watchdog.ts +12 -2
- package/dist/jev.d.ts +0 -123
- package/dist/jev.d.ts.map +0 -1
- package/dist/jev.js +0 -278
- package/dist/jev.js.map +0 -1
package/src/providers/gemini.ts
CHANGED
|
@@ -5,6 +5,11 @@
|
|
|
5
5
|
// its own chain of thought on the next turn. Written straight against the REST
|
|
6
6
|
// wire rather than @google/genai, because this package ships zero dependencies
|
|
7
7
|
// and the SDK is a Node-shaped one.
|
|
8
|
+
//
|
|
9
|
+
// Built-in tool shapes, parts and grounding fields follow Vercel's AI SDK
|
|
10
|
+
// (Apache-2.0, Copyright 2023 Vercel, Inc.: google-prepare-tools.ts,
|
|
11
|
+
// google-language-model.ts, convert-to-google-messages.ts); the byte offsets
|
|
12
|
+
// and segment checks follow PydanticAI (MIT). See THIRD_PARTY_NOTICES.
|
|
8
13
|
import { repairHistory } from "../context.ts";
|
|
9
14
|
import {
|
|
10
15
|
ProviderError,
|
|
@@ -13,10 +18,12 @@ import {
|
|
|
13
18
|
validateMaxAttempts,
|
|
14
19
|
type ErrorKind,
|
|
15
20
|
} from "../errors.ts";
|
|
21
|
+
import type { HostedToolType } from "../hosted/types.ts";
|
|
22
|
+
import { refuseHostedTools, type HostedCodecs } from "./hosted-codecs.ts";
|
|
16
23
|
import { schemaPrompt, toAntigravityToolSchema, toGeminiToolSchema } from "../schema.ts";
|
|
17
24
|
import { parseToolArgs, settleToolCalls } from "../tool-args.ts";
|
|
18
25
|
import { streamSse, apiUrl, type FetchLike } from "../transport.ts";
|
|
19
|
-
import {
|
|
26
|
+
import { configuredFallbacks, type ProviderFallbackConfig } from "../fallback.ts";
|
|
20
27
|
import type {
|
|
21
28
|
ChatMessage,
|
|
22
29
|
Effort,
|
|
@@ -25,6 +32,8 @@ import type {
|
|
|
25
32
|
Provider,
|
|
26
33
|
ProviderChunk,
|
|
27
34
|
StreamOptions,
|
|
35
|
+
TokenUsage,
|
|
36
|
+
ToolCall,
|
|
28
37
|
ToolCallDelta,
|
|
29
38
|
ToolChoice,
|
|
30
39
|
ToolDefinition,
|
|
@@ -64,6 +73,15 @@ export interface GeminiConfig extends ProviderFallbackConfig {
|
|
|
64
73
|
maxTokens?: number;
|
|
65
74
|
fetchImpl?: FetchLike;
|
|
66
75
|
headers?: Record<string, string>;
|
|
76
|
+
/**
|
|
77
|
+
* For providers imported from `@providerkit/core/hosted`.
|
|
78
|
+
* The hosted tools this endpoint runs. Defaults to web search, web fetch,
|
|
79
|
+
* code execution and file search on the Gemini API
|
|
80
|
+
* (generativelanguage.googleapis.com), and to none anywhere else. Vertex
|
|
81
|
+
* and the gateways differ, and one that ignores a tool answers unsearched
|
|
82
|
+
* without saying so. Antigravity runs none.
|
|
83
|
+
*/
|
|
84
|
+
servesHostedTools?: readonly HostedToolType[];
|
|
67
85
|
}
|
|
68
86
|
|
|
69
87
|
const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com";
|
|
@@ -193,6 +211,16 @@ function parseArgs(json: string): Record<string, unknown> {
|
|
|
193
211
|
return parseToolArgs(json);
|
|
194
212
|
}
|
|
195
213
|
|
|
214
|
+
function functionCallPart(call: ToolCall): unknown {
|
|
215
|
+
return {
|
|
216
|
+
functionCall: { id: call.id, name: call.name, args: parseArgs(call.arguments) },
|
|
217
|
+
// Replayed verbatim: this is the model's own reasoning token, and without
|
|
218
|
+
// it the next turn starts from a chain of thought that no longer includes
|
|
219
|
+
// the call it just made.
|
|
220
|
+
...(call.thoughtSignature ? { thoughtSignature: call.thoughtSignature } : {}),
|
|
221
|
+
};
|
|
222
|
+
}
|
|
223
|
+
|
|
196
224
|
/** Same rule for a tool RESULT, which `functionResponse.response` also requires
|
|
197
225
|
* as an object. Plain text (the common case) rides under `output`. */
|
|
198
226
|
function toResponseObject(content: string): Record<string, unknown> {
|
|
@@ -219,8 +247,25 @@ function toResponseObject(content: string): Record<string, unknown> {
|
|
|
219
247
|
* that the response count does not match the call count. The responses go
|
|
220
248
|
* first, in call order, and the run's images follow them all, the way the
|
|
221
249
|
* OpenAI dialect places them.
|
|
250
|
+
*
|
|
251
|
+
* A model turn that ran built-in tools goes back to `provider`, the one that
|
|
252
|
+
* made it, as the parts it came in: the code and its result, the server
|
|
253
|
+
* tool's call and response with their thought signatures, which Gemini needs
|
|
254
|
+
* returned unchanged when built-ins and functions share a turn.
|
|
222
255
|
*/
|
|
223
|
-
export function toGeminiContents(
|
|
256
|
+
export function toGeminiContents(
|
|
257
|
+
messages: readonly ChatMessage[],
|
|
258
|
+
provider?: string,
|
|
259
|
+
): { system?: string; contents: GeminiContent[] } {
|
|
260
|
+
return geminiContents(messages, provider);
|
|
261
|
+
}
|
|
262
|
+
|
|
263
|
+
/** @internal Hosted replay stays outside the plain converter. */
|
|
264
|
+
export function geminiContents(
|
|
265
|
+
messages: readonly ChatMessage[],
|
|
266
|
+
provider?: string,
|
|
267
|
+
replay?: HostedCodecs["gemini"]["replay"],
|
|
268
|
+
): {
|
|
224
269
|
system?: string;
|
|
225
270
|
contents: GeminiContent[];
|
|
226
271
|
} {
|
|
@@ -253,17 +298,15 @@ export function toGeminiContents(messages: readonly ChatMessage[]): {
|
|
|
253
298
|
break;
|
|
254
299
|
|
|
255
300
|
case "assistant": {
|
|
301
|
+
const replayed =
|
|
302
|
+
provider === undefined ? undefined : replay?.(message, provider, functionCallPart);
|
|
303
|
+
if (replayed) {
|
|
304
|
+
contents.push({ role: "model", parts: replayed });
|
|
305
|
+
break;
|
|
306
|
+
}
|
|
256
307
|
const parts: unknown[] = [];
|
|
257
308
|
if (message.content) parts.push({ text: message.content });
|
|
258
|
-
for (const call of message.toolCalls ?? [])
|
|
259
|
-
parts.push({
|
|
260
|
-
functionCall: { id: call.id, name: call.name, args: parseArgs(call.arguments) },
|
|
261
|
-
// Replayed verbatim: this is the model's own reasoning token, and
|
|
262
|
-
// without it the next turn starts from a chain of thought that
|
|
263
|
-
// no longer includes the call it just made.
|
|
264
|
-
...(call.thoughtSignature ? { thoughtSignature: call.thoughtSignature } : {}),
|
|
265
|
-
});
|
|
266
|
-
}
|
|
309
|
+
for (const call of message.toolCalls ?? []) parts.push(functionCallPart(call));
|
|
267
310
|
// An assistant turn with neither text nor calls has no representation
|
|
268
311
|
// here, and an empty `parts` is a 400.
|
|
269
312
|
if (parts.length > 0) contents.push({ role: "model", parts });
|
|
@@ -369,12 +412,39 @@ function mapFinishReason(provider: string, reason: string): FinishReason {
|
|
|
369
412
|
}
|
|
370
413
|
}
|
|
371
414
|
|
|
372
|
-
interface GeminiPart {
|
|
415
|
+
export interface GeminiPart {
|
|
373
416
|
text?: string;
|
|
374
417
|
/** Marks the parts that are the model's reasoning. Absent on the answer. */
|
|
375
418
|
thought?: boolean;
|
|
376
419
|
thoughtSignature?: string;
|
|
377
420
|
functionCall?: { id?: string; name?: string; args?: Record<string, unknown> };
|
|
421
|
+
/** Code the code execution tool ran, and what came of it, paired by order. */
|
|
422
|
+
executableCode?: { language?: string; code?: string };
|
|
423
|
+
codeExecutionResult?: { outcome?: string; output?: string };
|
|
424
|
+
/** A built-in tool's call and response, on Gemini 3 with function tools. */
|
|
425
|
+
toolCall?: { id?: string; toolType?: string; args?: Record<string, unknown> };
|
|
426
|
+
toolResponse?: { id?: string; toolType?: string; response?: unknown };
|
|
427
|
+
}
|
|
428
|
+
|
|
429
|
+
export interface GroundingSegment {
|
|
430
|
+
/** Offsets count UTF-8 bytes in the named Part, or the answer when absent. */
|
|
431
|
+
partIndex?: number;
|
|
432
|
+
startIndex?: number;
|
|
433
|
+
endIndex?: number;
|
|
434
|
+
text?: string;
|
|
435
|
+
}
|
|
436
|
+
|
|
437
|
+
export interface GroundingChunk {
|
|
438
|
+
web?: { uri?: string; title?: string };
|
|
439
|
+
retrievedContext?: { uri?: string; title?: string; text?: string };
|
|
440
|
+
}
|
|
441
|
+
|
|
442
|
+
/** What a grounded answer read, and which of its sentences each source backs. */
|
|
443
|
+
export interface GroundingMetadata {
|
|
444
|
+
webSearchQueries?: string[];
|
|
445
|
+
groundingChunks?: GroundingChunk[];
|
|
446
|
+
groundingSupports?: { segment?: GroundingSegment; groundingChunkIndices?: number[] }[];
|
|
447
|
+
searchEntryPoint?: { renderedContent?: string };
|
|
378
448
|
}
|
|
379
449
|
|
|
380
450
|
/** Google's `google.rpc.Status`: an HTTP `code`, the canonical `status` name,
|
|
@@ -386,16 +456,20 @@ interface GeminiStatus {
|
|
|
386
456
|
details?: unknown;
|
|
387
457
|
}
|
|
388
458
|
|
|
389
|
-
interface GeminiResponse {
|
|
459
|
+
export interface GeminiResponse {
|
|
390
460
|
candidates?: {
|
|
391
461
|
content?: { parts?: GeminiPart[] };
|
|
392
462
|
finishReason?: string;
|
|
463
|
+
groundingMetadata?: GroundingMetadata;
|
|
464
|
+
urlContextMetadata?: { urlMetadata?: { retrievedUrl?: string; urlRetrievalStatus?: string }[] };
|
|
393
465
|
}[];
|
|
394
466
|
usageMetadata?: {
|
|
395
467
|
promptTokenCount?: number;
|
|
396
468
|
candidatesTokenCount?: number;
|
|
397
469
|
thoughtsTokenCount?: number;
|
|
398
470
|
cachedContentTokenCount?: number;
|
|
471
|
+
/** The prompt the built-in tools' results added, billed as input. */
|
|
472
|
+
toolUsePromptTokenCount?: number;
|
|
399
473
|
};
|
|
400
474
|
/** Present only on the in-band failure below. Never on a real candidate. */
|
|
401
475
|
error?: GeminiStatus;
|
|
@@ -404,6 +478,11 @@ interface GeminiResponse {
|
|
|
404
478
|
}
|
|
405
479
|
|
|
406
480
|
export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
481
|
+
return geminiProvider(config);
|
|
482
|
+
}
|
|
483
|
+
|
|
484
|
+
/** @internal Shared by the plain and hosted entry points. */
|
|
485
|
+
export function geminiProvider(config: GeminiConfig, codecs?: HostedCodecs): Provider {
|
|
407
486
|
const baseUrl = config.baseUrl ?? DEFAULT_BASE_URL;
|
|
408
487
|
const id = config.id ?? "gemini";
|
|
409
488
|
const isAntigravity = id === "antigravity";
|
|
@@ -421,11 +500,16 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
421
500
|
): AsyncIterable<ProviderChunk> {
|
|
422
501
|
const { maxAttempts } = opts;
|
|
423
502
|
validateMaxAttempts(maxAttempts);
|
|
503
|
+
if (!codecs) refuseHostedTools(id, opts, messages);
|
|
424
504
|
const effort = opts.effort ?? config.effort;
|
|
425
505
|
const maxTokens = opts.maxTokens ?? config.maxTokens;
|
|
426
506
|
// Repaired first: a result with no call, or a call with no result, is a
|
|
427
507
|
// 400 that loses the turn. See repairHistory.
|
|
428
|
-
const { system, contents } =
|
|
508
|
+
const { system, contents } = geminiContents(
|
|
509
|
+
repairHistory(messages).messages,
|
|
510
|
+
id,
|
|
511
|
+
codecs?.gemini.replay,
|
|
512
|
+
);
|
|
429
513
|
|
|
430
514
|
// A model id copied out of Gemini's docs is often already `models/…`, and
|
|
431
515
|
// the doubled segment 404s as "model not found" — a confusing way to
|
|
@@ -436,6 +520,15 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
436
520
|
// Not `!isGemini3`: `gemini-flash-latest` and models newer than 3 are not
|
|
437
521
|
// Gemini 3 either, and they must get a level.
|
|
438
522
|
const isGemini2 = /gemini-2/i.test(model);
|
|
523
|
+
const hosted = codecs?.gemini.plan(
|
|
524
|
+
config,
|
|
525
|
+
baseUrl,
|
|
526
|
+
id,
|
|
527
|
+
isAntigravity,
|
|
528
|
+
isGemini2,
|
|
529
|
+
tools,
|
|
530
|
+
opts,
|
|
531
|
+
);
|
|
439
532
|
|
|
440
533
|
const generationConfig: Record<string, unknown> = {};
|
|
441
534
|
if (maxTokens !== undefined) generationConfig.maxOutputTokens = maxTokens;
|
|
@@ -460,9 +553,10 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
460
553
|
// outright with `400 "Function calling with a response mime type:
|
|
461
554
|
// 'application/json' is unsupported"`, so the prompt is the only way a
|
|
462
555
|
// schema travels beside tools there. Gemini 3 serves both, and there the
|
|
463
|
-
// setting is the caller's measured preference.
|
|
556
|
+
// setting is the caller's measured preference. A built-in tool is a tool
|
|
557
|
+
// here too: the prompt is the one route that works beside it everywhere.
|
|
464
558
|
const promptCarried =
|
|
465
|
-
tools.length > 0 &&
|
|
559
|
+
(tools.length > 0 || (hosted?.tools.length ?? 0) > 0) &&
|
|
466
560
|
((opts.jsonWithTools ?? config.jsonWithTools) === "prompt" || !isGemini3);
|
|
467
561
|
if (opts.json && !promptCarried) {
|
|
468
562
|
generationConfig.responseMimeType = "application/json";
|
|
@@ -490,34 +584,51 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
490
584
|
...(opts.json && promptCarried ? [{ text: schemaPrompt(opts.json.schema) }] : []),
|
|
491
585
|
];
|
|
492
586
|
if (systemParts.length > 0) request.systemInstruction = { parts: systemParts };
|
|
493
|
-
|
|
494
|
-
|
|
495
|
-
|
|
496
|
-
|
|
497
|
-
|
|
498
|
-
|
|
499
|
-
|
|
500
|
-
|
|
501
|
-
|
|
502
|
-
|
|
503
|
-
|
|
587
|
+
// Gemini has no switch for one call per turn. Sending the request anyway
|
|
588
|
+
// would hand back the parallel calls the caller ruled out.
|
|
589
|
+
if (tools.length > 0 && opts.parallelToolCalls === false) {
|
|
590
|
+
throw new ProviderError(
|
|
591
|
+
id,
|
|
592
|
+
"invalid",
|
|
593
|
+
`${id}: Gemini can't limit a turn to one tool call. Leave parallelToolCalls unset ` +
|
|
594
|
+
"for Gemini models, or run the calls it returns one at a time.",
|
|
595
|
+
);
|
|
596
|
+
}
|
|
597
|
+
const builtIns = hosted?.tools ?? [];
|
|
598
|
+
if (tools.length > 0 || builtIns.length > 0) {
|
|
504
599
|
request.tools = [
|
|
505
|
-
|
|
506
|
-
|
|
507
|
-
|
|
508
|
-
|
|
509
|
-
|
|
510
|
-
|
|
511
|
-
|
|
512
|
-
|
|
513
|
-
|
|
514
|
-
|
|
515
|
-
|
|
516
|
-
|
|
600
|
+
...(tools.length > 0
|
|
601
|
+
? [
|
|
602
|
+
{
|
|
603
|
+
functionDeclarations: tools.map((tool) => ({
|
|
604
|
+
name: tool.name,
|
|
605
|
+
description: tool.description,
|
|
606
|
+
// Antigravity (Cloud Code Assist) still expects the OpenAPI-subset
|
|
607
|
+
// `parameters` field; the Gemini API and OpenCode Zen take standard
|
|
608
|
+
// JSON Schema in `parametersJsonSchema` (cc-proxy 0.2.8).
|
|
609
|
+
...(isAntigravity
|
|
610
|
+
? { parameters: toAntigravityToolSchema(tool.inputSchema) }
|
|
611
|
+
: { parametersJsonSchema: toGeminiToolSchema(tool.inputSchema) }),
|
|
612
|
+
})),
|
|
613
|
+
},
|
|
614
|
+
]
|
|
615
|
+
: []),
|
|
616
|
+
...builtIns,
|
|
517
617
|
];
|
|
618
|
+
}
|
|
619
|
+
if (tools.length > 0) {
|
|
518
620
|
// Pointless without declarations, and Gemini 400s on a tool config that
|
|
519
|
-
// names a function it was never given.
|
|
520
|
-
|
|
621
|
+
// names a function it was never given. Built-ins beside functions need
|
|
622
|
+
// the mode that lets both run and the server's own calls streamed back,
|
|
623
|
+
// which then go back with the turn (google-prepare-tools.ts:187-231).
|
|
624
|
+
// A forced choice never gets here with built-ins: it is refused.
|
|
625
|
+
const toolConfig =
|
|
626
|
+
builtIns.length > 0
|
|
627
|
+
? {
|
|
628
|
+
functionCallingConfig: { mode: "VALIDATED" },
|
|
629
|
+
includeServerSideToolInvocations: true,
|
|
630
|
+
}
|
|
631
|
+
: toToolConfig(opts.toolChoice);
|
|
521
632
|
if (toolConfig) request.toolConfig = toolConfig;
|
|
522
633
|
}
|
|
523
634
|
|
|
@@ -571,6 +682,7 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
571
682
|
let callIndex = 0;
|
|
572
683
|
let finished = false;
|
|
573
684
|
let emitted = false;
|
|
685
|
+
const hostedStream = codecs?.gemini.stream(id, hosted?.hosted ?? [], isGemini2);
|
|
574
686
|
|
|
575
687
|
for (let baseIdx = 0; baseIdx < bases.length; baseIdx++) {
|
|
576
688
|
const currentBase = bases[baseIdx]!;
|
|
@@ -615,33 +727,43 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
615
727
|
if (chunk.usageMetadata) {
|
|
616
728
|
const usage = chunk.usageMetadata;
|
|
617
729
|
emitted = true;
|
|
618
|
-
|
|
619
|
-
|
|
620
|
-
|
|
621
|
-
|
|
622
|
-
|
|
623
|
-
|
|
624
|
-
|
|
625
|
-
|
|
626
|
-
|
|
627
|
-
|
|
628
|
-
|
|
629
|
-
|
|
630
|
-
|
|
730
|
+
const tokenUsage: TokenUsage = {
|
|
731
|
+
// What the built-in tools' results added to the prompt bills as
|
|
732
|
+
// input too, and Gemini reports it apart
|
|
733
|
+
// (Vercel convert-google-usage.ts:31-36).
|
|
734
|
+
inputTokens: (usage.promptTokenCount ?? 0) + (usage.toolUsePromptTokenCount ?? 0),
|
|
735
|
+
// Thoughts bill as OUTPUT, and Gemini reports them OUTSIDE
|
|
736
|
+
// candidatesTokenCount, so leaving them out undercounts a thinking
|
|
737
|
+
// turn by most of what it cost.
|
|
738
|
+
outputTokens: (usage.candidatesTokenCount ?? 0) + (usage.thoughtsTokenCount ?? 0),
|
|
739
|
+
cachedInputTokens: usage.cachedContentTokenCount ?? 0,
|
|
740
|
+
...(usage.thoughtsTokenCount !== undefined
|
|
741
|
+
? { reasoningTokens: usage.thoughtsTokenCount }
|
|
742
|
+
: {}),
|
|
631
743
|
};
|
|
744
|
+
yield { type: "usage", usage: hostedStream?.usage(tokenUsage) ?? tokenUsage };
|
|
632
745
|
}
|
|
633
746
|
|
|
634
747
|
const candidate = chunk.candidates?.[0];
|
|
635
748
|
if (!candidate) continue;
|
|
749
|
+
hostedStream?.candidate(candidate);
|
|
636
750
|
|
|
637
|
-
//
|
|
638
|
-
//
|
|
639
|
-
//
|
|
751
|
+
// Coalesce text per frame, but emit hosted work at once: validating
|
|
752
|
+
// a later function call can throw, and must never hide a paid run
|
|
753
|
+
// from retry and fallback (invariant 19).
|
|
640
754
|
let reasoning = "";
|
|
641
755
|
let content = "";
|
|
642
756
|
const toolCalls: ToolCallDelta[] = [];
|
|
643
|
-
for (const part of candidate.content?.parts ?? []) {
|
|
644
|
-
|
|
757
|
+
for (const [partIndex, part] of (candidate.content?.parts ?? []).entries()) {
|
|
758
|
+
const ran = hostedStream?.part(part, partIndex, content.length);
|
|
759
|
+
if (ran) {
|
|
760
|
+
if (ran.delta) {
|
|
761
|
+
emitted = true;
|
|
762
|
+
yield { type: "delta", hosted: [ran.delta] };
|
|
763
|
+
}
|
|
764
|
+
continue;
|
|
765
|
+
}
|
|
766
|
+
if (part.text !== undefined) {
|
|
645
767
|
if (part.thought) reasoning += part.text;
|
|
646
768
|
else content += part.text;
|
|
647
769
|
continue;
|
|
@@ -670,6 +792,7 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
670
792
|
}
|
|
671
793
|
if (content) {
|
|
672
794
|
emitted = true;
|
|
795
|
+
hostedStream?.text(content);
|
|
673
796
|
yield { type: "delta", content };
|
|
674
797
|
}
|
|
675
798
|
if (toolCalls.length > 0) {
|
|
@@ -683,6 +806,7 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
683
806
|
const finishReason = mapFinishReason(id, candidate.finishReason);
|
|
684
807
|
finished = true;
|
|
685
808
|
emitted = true;
|
|
809
|
+
if (hostedStream) yield* hostedStream.close();
|
|
686
810
|
yield {
|
|
687
811
|
type: "finish",
|
|
688
812
|
// Gemini routinely reports STOP on a turn that called tools, and a
|
|
@@ -718,5 +842,5 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
718
842
|
},
|
|
719
843
|
};
|
|
720
844
|
|
|
721
|
-
return
|
|
845
|
+
return configuredFallbacks(provider, config, codecs);
|
|
722
846
|
}
|
|
@@ -0,0 +1,29 @@
|
|
|
1
|
+
import { ProviderError } from "../errors.ts";
|
|
2
|
+
import type { ChatMessage, StreamOptions } from "../types.ts";
|
|
3
|
+
|
|
4
|
+
/** Internal wiring only. Type imports never pull the codecs into plain chat. */
|
|
5
|
+
export interface HostedCodecs {
|
|
6
|
+
anthropic: typeof import("../hosted/anthropic.ts").anthropicHosted;
|
|
7
|
+
responses: typeof import("../hosted/responses.ts").responsesHosted;
|
|
8
|
+
gemini: typeof import("../hosted/gemini.ts").geminiHosted;
|
|
9
|
+
openai: typeof import("../hosted/openai.ts").openaiHosted;
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
export function refuseHostedTools(
|
|
13
|
+
provider: string,
|
|
14
|
+
opts: StreamOptions,
|
|
15
|
+
messages: readonly ChatMessage[],
|
|
16
|
+
): void {
|
|
17
|
+
if (
|
|
18
|
+
opts.hostedTools?.length ||
|
|
19
|
+
messages.some(
|
|
20
|
+
(message) => message.role === "assistant" && message.hostedItems?.provider === provider,
|
|
21
|
+
)
|
|
22
|
+
) {
|
|
23
|
+
throw new ProviderError(
|
|
24
|
+
provider,
|
|
25
|
+
"invalid",
|
|
26
|
+
`${provider}: hosted tools need the opt-in provider. Import its constructor from @providerkit/core/hosted.`,
|
|
27
|
+
);
|
|
28
|
+
}
|
|
29
|
+
}
|
package/src/providers/openai.ts
CHANGED
|
@@ -11,6 +11,8 @@ import {
|
|
|
11
11
|
streamError,
|
|
12
12
|
validateMaxAttempts,
|
|
13
13
|
} from "../errors.ts";
|
|
14
|
+
import type { HostedToolType } from "../hosted/types.ts";
|
|
15
|
+
import { refuseHostedTools, type HostedCodecs } from "./hosted-codecs.ts";
|
|
14
16
|
import { streamSse, apiUrl, parseFrame, type FetchLike } from "../transport.ts";
|
|
15
17
|
import { settleToolCalls, type ToolCallIdentity } from "../tool-args.ts";
|
|
16
18
|
import { attributionHeaders } from "../attribution.ts";
|
|
@@ -28,7 +30,7 @@ import type {
|
|
|
28
30
|
} from "../types.ts";
|
|
29
31
|
import { toDataUri } from "../types.ts";
|
|
30
32
|
import { isStrictSchema, schemaPrompt } from "../schema.ts";
|
|
31
|
-
import {
|
|
33
|
+
import { configuredFallbacks, type ProviderFallbackConfig } from "../fallback.ts";
|
|
32
34
|
|
|
33
35
|
export interface OpenAIConfig extends ProviderFallbackConfig {
|
|
34
36
|
apiKey: string;
|
|
@@ -139,6 +141,16 @@ export interface OpenAIConfig extends ProviderFallbackConfig {
|
|
|
139
141
|
noSampling?: boolean;
|
|
140
142
|
/** Default `service_tier` for every call; `StreamOptions.serviceTier` wins. */
|
|
141
143
|
serviceTier?: ServiceTier;
|
|
144
|
+
/**
|
|
145
|
+
* For providers imported from `@providerkit/core/hosted`.
|
|
146
|
+
* The hosted tools this endpoint runs. This wire has one, web search.
|
|
147
|
+
* Defaults to it on openrouter.ai, which runs it as its own server tool,
|
|
148
|
+
* and on api.openai.com for the `-search-preview` models, and to none
|
|
149
|
+
* anywhere else: a gateway that ignores the tool answers unsearched
|
|
150
|
+
* without saying so. A host other than OpenRouter that runs it is sent
|
|
151
|
+
* OpenAI's `web_search_options`.
|
|
152
|
+
*/
|
|
153
|
+
servesHostedTools?: readonly HostedToolType[];
|
|
142
154
|
}
|
|
143
155
|
|
|
144
156
|
const DEFAULT_BASE_URL = "https://api.openai.com";
|
|
@@ -476,7 +488,7 @@ export function toOpenAIMessages(messages: readonly ChatMessage[], provider = "o
|
|
|
476
488
|
return out;
|
|
477
489
|
}
|
|
478
490
|
|
|
479
|
-
interface OpenAIChunk {
|
|
491
|
+
export interface OpenAIChunk {
|
|
480
492
|
choices?: {
|
|
481
493
|
delta?: {
|
|
482
494
|
content?: string | null;
|
|
@@ -484,6 +496,8 @@ interface OpenAIChunk {
|
|
|
484
496
|
reasoning?: string | null;
|
|
485
497
|
/** OpenRouter's normalized reasoning payload, on the final delta. */
|
|
486
498
|
reasoning_details?: unknown[];
|
|
499
|
+
/** The sources a searched answer cites. */
|
|
500
|
+
annotations?: import("../hosted/openai.ts").ChatAnnotation[];
|
|
487
501
|
tool_calls?: {
|
|
488
502
|
index?: number;
|
|
489
503
|
id?: string;
|
|
@@ -507,6 +521,8 @@ interface OpenAIChunk {
|
|
|
507
521
|
cost?: unknown;
|
|
508
522
|
is_byok?: boolean;
|
|
509
523
|
cost_details?: { upstream_inference_cost?: unknown } | null;
|
|
524
|
+
/** The server tool calls OpenRouter ran for the turn. */
|
|
525
|
+
server_tool_use_details?: { web_search_requests?: number } | null;
|
|
510
526
|
} | null;
|
|
511
527
|
/** Present only on the in-band failure below. Never beside a choice.
|
|
512
528
|
* `code` is the numeric HTTP status on the gateways, a slug on OpenAI. */
|
|
@@ -550,6 +566,11 @@ export function openRouterCostUsd(usage: {
|
|
|
550
566
|
}
|
|
551
567
|
|
|
552
568
|
export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
569
|
+
return openaiProvider(config);
|
|
570
|
+
}
|
|
571
|
+
|
|
572
|
+
/** @internal Shared by the plain and hosted entry points. */
|
|
573
|
+
export function openaiProvider(config: OpenAIConfig, codecs?: HostedCodecs): Provider {
|
|
553
574
|
const baseUrl = config.baseUrl ?? DEFAULT_BASE_URL;
|
|
554
575
|
const id = config.id ?? "openai";
|
|
555
576
|
// Keyed on the host, not the id: the host is who sends the bill, and
|
|
@@ -570,6 +591,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
570
591
|
opts: StreamOptions = {},
|
|
571
592
|
): AsyncIterable<ProviderChunk> {
|
|
572
593
|
validateMaxAttempts(opts.maxAttempts);
|
|
594
|
+
if (!codecs) refuseHostedTools(id, opts, messages);
|
|
573
595
|
const effort = opts.effort ?? config.effort;
|
|
574
596
|
const model = opts.model ?? config.model;
|
|
575
597
|
|
|
@@ -594,18 +616,25 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
594
616
|
const serviceTier = opts.serviceTier ?? config.serviceTier;
|
|
595
617
|
if (serviceTier) request.service_tier = serviceTier;
|
|
596
618
|
Object.assign(request, effortParams(config.effortDialect ?? dialectFor(id), effort, model));
|
|
597
|
-
|
|
598
|
-
|
|
599
|
-
|
|
600
|
-
|
|
601
|
-
|
|
602
|
-
|
|
603
|
-
|
|
604
|
-
|
|
605
|
-
|
|
619
|
+
const openRouter = isOpenRouter(baseUrl);
|
|
620
|
+
const hosted = codecs?.openai.plan(config, baseUrl, id, model, openRouter, opts);
|
|
621
|
+
if (tools.length > 0 || hosted?.tools.length) {
|
|
622
|
+
request.tools = [
|
|
623
|
+
...tools.map((tool) => ({
|
|
624
|
+
type: "function",
|
|
625
|
+
function: {
|
|
626
|
+
name: tool.name,
|
|
627
|
+
description: tool.description,
|
|
628
|
+
parameters: tool.inputSchema,
|
|
629
|
+
},
|
|
630
|
+
})),
|
|
631
|
+
...(hosted?.tools ?? []),
|
|
632
|
+
];
|
|
606
633
|
// Only beside tools: OpenAI refuses the field on a request without any.
|
|
607
|
-
if (opts.parallelToolCalls === false)
|
|
634
|
+
if (tools.length > 0 && opts.parallelToolCalls === false)
|
|
635
|
+
request.parallel_tool_calls = false;
|
|
608
636
|
}
|
|
637
|
+
if (hosted?.searchOptions) request.web_search_options = hosted.searchOptions;
|
|
609
638
|
if (opts.toolChoice && opts.toolChoice !== "auto") {
|
|
610
639
|
request.tool_choice =
|
|
611
640
|
typeof opts.toolChoice === "string"
|
|
@@ -620,7 +649,8 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
620
649
|
// here the only job is to leave the response format off those calls, so
|
|
621
650
|
// the schema reaches the model the one way nothing can suppress.
|
|
622
651
|
const promptCarried =
|
|
623
|
-
tools.length > 0
|
|
652
|
+
(tools.length > 0 || hosted?.active) &&
|
|
653
|
+
(opts.jsonWithTools ?? config.jsonWithTools) === "prompt";
|
|
624
654
|
const enforced =
|
|
625
655
|
!promptCarried &&
|
|
626
656
|
(config.jsonMode ?? (id === "openai" ? "schema" : "object")) === "schema";
|
|
@@ -687,6 +717,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
687
717
|
// Each call as its deltas named it, by index. The finish checks them
|
|
688
718
|
// before it goes out, and they are what says the turn called a tool.
|
|
689
719
|
const calls = new Map<number, ToolCallIdentity>();
|
|
720
|
+
const hostedStream = codecs?.openai.stream();
|
|
690
721
|
for await (const data of streamSse({
|
|
691
722
|
url: apiUrl(baseUrl, config.path ?? defaultPath),
|
|
692
723
|
headers: {
|
|
@@ -738,6 +769,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
738
769
|
chunk.usage.cached_tokens ??
|
|
739
770
|
0;
|
|
740
771
|
const reportedCostUsd = reportsCost ? openRouterCostUsd(chunk.usage) : undefined;
|
|
772
|
+
const hostedCalls = codecs?.openai.counts(chunk.usage);
|
|
741
773
|
yield {
|
|
742
774
|
type: "usage",
|
|
743
775
|
usage: {
|
|
@@ -750,6 +782,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
750
782
|
? { reasoningTokens: chunk.usage.completion_tokens_details.reasoning_tokens }
|
|
751
783
|
: {}),
|
|
752
784
|
...(reportedCostUsd !== undefined ? { reportedCostUsd } : {}),
|
|
785
|
+
...(hostedCalls ? { hostedCalls } : {}),
|
|
753
786
|
},
|
|
754
787
|
};
|
|
755
788
|
}
|
|
@@ -765,6 +798,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
765
798
|
out.content = delta.content;
|
|
766
799
|
has = true;
|
|
767
800
|
}
|
|
801
|
+
hostedStream?.delta(delta);
|
|
768
802
|
// `reasoning_content` is DeepSeek's field; `reasoning` is
|
|
769
803
|
// OpenRouter's normalized one. Whichever arrives is the same thing.
|
|
770
804
|
const reasoning = delta.reasoning_content ?? delta.reasoning;
|
|
@@ -799,6 +833,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
799
833
|
|
|
800
834
|
if (choice.finish_reason) {
|
|
801
835
|
const stated = mapFinishReason(id, choice.finish_reason);
|
|
836
|
+
if (hostedStream) yield* hostedStream.close();
|
|
802
837
|
const minted = settleToolCalls(id, calls.values());
|
|
803
838
|
if (minted.length > 0) yield { type: "delta", toolCalls: minted };
|
|
804
839
|
ended = true;
|
|
@@ -812,11 +847,13 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
812
847
|
// Either proof will do: some gateways send `[DONE]` with no finish, and
|
|
813
848
|
// some a finish with no `[DONE]`. Neither means the socket closed early.
|
|
814
849
|
if (!ended) throw streamCut(id);
|
|
815
|
-
// A `[DONE]` with no finish stated still owes its calls the same check
|
|
850
|
+
// A `[DONE]` with no finish stated still owes its calls the same check,
|
|
851
|
+
// and its citations their place.
|
|
852
|
+
if (hostedStream) yield* hostedStream.close();
|
|
816
853
|
const minted = settleToolCalls(id, calls.values());
|
|
817
854
|
if (minted.length > 0) yield { type: "delta", toolCalls: minted };
|
|
818
855
|
},
|
|
819
856
|
};
|
|
820
857
|
|
|
821
|
-
return
|
|
858
|
+
return configuredFallbacks(provider, config, codecs);
|
|
822
859
|
}
|