@providerkit/core 0.26.0 → 0.26.2
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/decisions.js +1 -1
- package/dist/decisions.js.map +1 -1
- package/dist/errors.d.ts +1 -1
- package/dist/errors.d.ts.map +1 -1
- package/dist/errors.js +63 -28
- package/dist/errors.js.map +1 -1
- package/dist/providers/anthropic.js +1 -1
- package/dist/providers/anthropic.js.map +1 -1
- package/dist/providers/factory.d.ts +5 -0
- package/dist/providers/factory.d.ts.map +1 -1
- package/dist/providers/factory.js +1 -0
- package/dist/providers/factory.js.map +1 -1
- package/dist/providers/gemini.js +1 -1
- package/dist/providers/gemini.js.map +1 -1
- package/dist/providers/openai.d.ts +10 -0
- package/dist/providers/openai.d.ts.map +1 -1
- package/dist/providers/openai.js +7 -3
- package/dist/providers/openai.js.map +1 -1
- package/dist/providers/responses.d.ts.map +1 -1
- package/dist/providers/responses.js +3 -5
- package/dist/providers/responses.js.map +1 -1
- package/dist/rate-limit.d.ts.map +1 -1
- package/dist/rate-limit.js +29 -6
- package/dist/rate-limit.js.map +1 -1
- package/package.json +1 -1
- package/src/decisions.ts +1 -1
- package/src/errors.ts +68 -25
- package/src/providers/anthropic.ts +1 -1
- package/src/providers/factory.ts +6 -0
- package/src/providers/gemini.ts +1 -1
- package/src/providers/openai.ts +17 -3
- package/src/providers/responses.ts +6 -8
- package/src/rate-limit.ts +28 -6
package/src/errors.ts
CHANGED
|
@@ -10,7 +10,12 @@
|
|
|
10
10
|
// read before the status — status alone gives the wrong advice ("retry" for an
|
|
11
11
|
// empty balance, "check your key" for a plan that never included the API).
|
|
12
12
|
|
|
13
|
-
import {
|
|
13
|
+
import {
|
|
14
|
+
parseRateLimitReset,
|
|
15
|
+
parseRateLimitResponse,
|
|
16
|
+
parseUsageLimitBody,
|
|
17
|
+
type RateLimitWindow,
|
|
18
|
+
} from "./rate-limit.ts";
|
|
14
19
|
import type { FilePart } from "./types.ts";
|
|
15
20
|
|
|
16
21
|
/** What kind of failure this is, named by what actually fixes it. */
|
|
@@ -164,11 +169,12 @@ export class ProviderError extends Error {
|
|
|
164
169
|
const code = readString(err, "code") ?? readString(err, "type");
|
|
165
170
|
const message = messageOf(err);
|
|
166
171
|
const body = bodyTextOf(err);
|
|
172
|
+
const resetBody = providerBodyOf(err) ?? body;
|
|
167
173
|
return new ProviderError(provider, classify(err, status, body), message, {
|
|
168
174
|
status,
|
|
169
175
|
code,
|
|
170
|
-
retryAfterMs: parseRetryAfterMs(err,
|
|
171
|
-
...parseUsageLimitBody(
|
|
176
|
+
retryAfterMs: parseRetryAfterMs(err, resetBody),
|
|
177
|
+
...parseUsageLimitBody(resetBody),
|
|
172
178
|
body: body.slice(0, 2_000) || undefined,
|
|
173
179
|
cause: err,
|
|
174
180
|
});
|
|
@@ -218,13 +224,17 @@ export function messageOf(err: unknown): string {
|
|
|
218
224
|
*/
|
|
219
225
|
function bodyTextOf(err: unknown): string {
|
|
220
226
|
const message = messageOf(err);
|
|
221
|
-
|
|
222
|
-
|
|
223
|
-
|
|
227
|
+
const body = providerBodyOf(err);
|
|
228
|
+
return body === undefined ? message : `${message} ${body}`;
|
|
229
|
+
}
|
|
230
|
+
|
|
231
|
+
/** Kept as JSON for reset parsing; the searchable text above is not JSON. */
|
|
232
|
+
function providerBodyOf(err: unknown): string | undefined {
|
|
233
|
+
if (typeof err !== "object" || err === null || !("error" in err)) return undefined;
|
|
224
234
|
try {
|
|
225
|
-
return
|
|
235
|
+
return typeof err.error === "string" ? err.error : JSON.stringify(err.error);
|
|
226
236
|
} catch {
|
|
227
|
-
return
|
|
237
|
+
return undefined;
|
|
228
238
|
}
|
|
229
239
|
}
|
|
230
240
|
|
|
@@ -361,7 +371,7 @@ const CONTEXT_PATTERNS: readonly RegExp[] = [
|
|
|
361
371
|
// it, and its wording overlaps both other categories ("plan" appears in all
|
|
362
372
|
// three), so it is checked first.
|
|
363
373
|
const ENTITLEMENT_PATTERNS: readonly RegExp[] = [
|
|
364
|
-
/plan does(?:n't| not) (?:include|support)/i,
|
|
374
|
+
/plan does(?:n't| not) (?:yet )?(?:include|support)/i,
|
|
365
375
|
/not included (?:in|with) your .{0,40}plan/i,
|
|
366
376
|
/upgrade to [\w ]{1,30}(?:or higher|plan)/i,
|
|
367
377
|
/requires? (?:the |a |an )?[\w\s]{1,30}plan/i,
|
|
@@ -381,6 +391,9 @@ const QUOTA_PATTERNS: readonly RegExp[] = [
|
|
|
381
391
|
/insufficient[_\s]quota/i,
|
|
382
392
|
/exceeded your current quota/i,
|
|
383
393
|
/insufficient (?:balance|credits?)/i,
|
|
394
|
+
/insufficient (?:account )?funds/i,
|
|
395
|
+
/\bcredit limit exceeded\b/i,
|
|
396
|
+
/\baccount invoice is overdue\b/i,
|
|
384
397
|
/credit balance is too low/i,
|
|
385
398
|
/(?:no|any|out of) credits/i,
|
|
386
399
|
/(?:credits?|balance) depleted/i,
|
|
@@ -401,6 +414,13 @@ const QUOTA_PATTERNS: readonly RegExp[] = [
|
|
|
401
414
|
// The OpenAI-shape `type` field for billing failures, serialized into the
|
|
402
415
|
// scanned text — a quota signal with no prose wording at all (codex-router).
|
|
403
416
|
/billing_error/i,
|
|
417
|
+
// OpenCode's typed account caps can carry throttle prose or no prose at all.
|
|
418
|
+
// Their names survive both the HTTP and in-band error envelopes.
|
|
419
|
+
/\b(?:GoUsageLimitError|BudgetLimitError|FreeUsageLimitError|CreditLimitExceeded)\b/i,
|
|
420
|
+
/\bbudget exceeded\b/i,
|
|
421
|
+
// Z.ai reports expired coding packages and spent weekly/monthly windows as 429.
|
|
422
|
+
/\bpackage has expired\b/i,
|
|
423
|
+
/\b(?:weekly|monthly) limit exhausted\b/i,
|
|
404
424
|
// Alibaba intl's English spelling of 欠费.
|
|
405
425
|
/\barrears?\b/i,
|
|
406
426
|
/余额不足/,
|
|
@@ -721,19 +741,34 @@ const STATUS_BY_TYPE: ReadonlyMap<string, number> = new Map([
|
|
|
721
741
|
* body does name keeps its own kind, which is what tells the caller to wait,
|
|
722
742
|
* rotate a key, or compact instead of retrying a failure that will repeat.
|
|
723
743
|
*/
|
|
724
|
-
export function streamError(provider: string,
|
|
744
|
+
export function streamError(provider: string, envelope: unknown): ProviderError {
|
|
745
|
+
// Only failure data is searchable or logged. A failed Responses object can
|
|
746
|
+
// echo instructions, tool definitions and output alongside its error.
|
|
747
|
+
const outer =
|
|
748
|
+
typeof envelope === "object" && envelope !== null ? (envelope as Record<string, unknown>) : {};
|
|
749
|
+
const error = outer.error ?? outer.errors ?? envelope;
|
|
725
750
|
const body = JSON.stringify(error ?? {}) ?? "";
|
|
726
751
|
const status = [
|
|
727
752
|
readNumber(error, "code"),
|
|
728
753
|
readNumber(error, "status"),
|
|
729
754
|
readNumber(error, "status_code"),
|
|
755
|
+
readNumber(envelope, "status_code"),
|
|
730
756
|
].find((value) => value !== undefined && value >= 400 && value < 600);
|
|
731
757
|
const code =
|
|
732
758
|
readString(error, "code") ?? readString(error, "status") ?? readString(error, "type");
|
|
733
759
|
const typeStatus =
|
|
734
760
|
STATUS_BY_TYPE.get(readString(error, "type") ?? "") ??
|
|
735
761
|
STATUS_BY_TYPE.get(readString(error, "code") ?? "");
|
|
736
|
-
|
|
762
|
+
// Keep type-derived 4xx precedence. Only a generic 5xx label without a real
|
|
763
|
+
// status may hide an account cap; OpenCode uses server_error for those.
|
|
764
|
+
const classified = classifyHttp(status ?? typeStatus, body);
|
|
765
|
+
const kind =
|
|
766
|
+
status === undefined &&
|
|
767
|
+
typeStatus !== undefined &&
|
|
768
|
+
typeStatus >= 500 &&
|
|
769
|
+
(matches(QUOTA_PATTERNS, body) || matches(ENTITLEMENT_PATTERNS, body))
|
|
770
|
+
? classifyHttp(undefined, body)
|
|
771
|
+
: classified;
|
|
737
772
|
return new ProviderError(
|
|
738
773
|
provider,
|
|
739
774
|
kind === "unknown" ? "overload" : kind,
|
|
@@ -741,27 +776,35 @@ export function streamError(provider: string, error: unknown): ProviderError {
|
|
|
741
776
|
{
|
|
742
777
|
...(status !== undefined ? { status } : {}),
|
|
743
778
|
...(code ? { code } : {}),
|
|
744
|
-
|
|
745
|
-
...readReset(error, body),
|
|
779
|
+
...readReset(envelope, error, body, kind),
|
|
746
780
|
body: body.slice(0, 2_000),
|
|
747
781
|
},
|
|
748
782
|
);
|
|
749
783
|
}
|
|
750
784
|
|
|
751
|
-
/**
|
|
752
|
-
*
|
|
753
|
-
|
|
754
|
-
function readReset(error: unknown, body: string) {
|
|
755
|
-
const raw =
|
|
756
|
-
typeof error === "object" && error !== null
|
|
757
|
-
? (error as { headers?: unknown }).headers
|
|
758
|
-
: undefined;
|
|
759
|
-
if (typeof raw !== "object" || raw === null) return parseUsageLimitBody(body);
|
|
785
|
+
/** Read reset metadata without copying echoed request data into diagnostics.
|
|
786
|
+
* Mirrored account windows describe only rate/quota failures, not a 503. */
|
|
787
|
+
function readReset(envelope: unknown, error: unknown, body: string, kind: ErrorKind) {
|
|
760
788
|
const headers = new Headers();
|
|
761
|
-
for (const [
|
|
762
|
-
if (typeof
|
|
789
|
+
for (const source of [error, envelope]) {
|
|
790
|
+
if (typeof source !== "object" || source === null || !("headers" in source)) continue;
|
|
791
|
+
const raw = source.headers;
|
|
792
|
+
if (typeof raw !== "object" || raw === null) continue;
|
|
793
|
+
for (const [name, value] of Object.entries(raw)) {
|
|
794
|
+
if (typeof value === "string" || typeof value === "number") headers.set(name, String(value));
|
|
795
|
+
}
|
|
763
796
|
}
|
|
764
|
-
|
|
797
|
+
if (kind !== "rate" && kind !== "quota") {
|
|
798
|
+
return {
|
|
799
|
+
retryAfterMs: parseRateLimitReset(headers).retryAfterMs ?? parseRetryAfterMs(error, body),
|
|
800
|
+
};
|
|
801
|
+
}
|
|
802
|
+
const metadata =
|
|
803
|
+
typeof envelope === "object" && envelope !== null && "metadata" in envelope
|
|
804
|
+
? envelope.metadata
|
|
805
|
+
: undefined;
|
|
806
|
+
const reset = parseRateLimitResponse(headers, JSON.stringify({ error, metadata }));
|
|
807
|
+
return { ...reset, retryAfterMs: reset.retryAfterMs ?? parseRetryAfterMs(error, body) };
|
|
765
808
|
}
|
|
766
809
|
|
|
767
810
|
/**
|
|
@@ -578,7 +578,7 @@ export function createAnthropicProvider(config: AnthropicConfig): Provider {
|
|
|
578
578
|
// most of the time, but a prompt found too long mid-stream arrives the
|
|
579
579
|
// same way, and retrying that one only fails it again more slowly.
|
|
580
580
|
case "error":
|
|
581
|
-
throw streamError(id, event
|
|
581
|
+
throw streamError(id, event);
|
|
582
582
|
|
|
583
583
|
case "message_start": {
|
|
584
584
|
const usage = event.message?.usage;
|
package/src/providers/factory.ts
CHANGED
|
@@ -84,6 +84,11 @@ export interface PresetProviderConfig extends ProviderFallbackConfig {
|
|
|
84
84
|
* between rounds. Pinning keeps a conversation's rounds (and their cache)
|
|
85
85
|
* on one host. Empty/absent = default routing. */
|
|
86
86
|
providerOrder?: string[];
|
|
87
|
+
/** OpenRouter-only: `provider.data_collection`. `"deny"` keeps the request
|
|
88
|
+
* on hosts that don't train on the prompt. Not zero retention: that is
|
|
89
|
+
* OpenRouter's separate `zdr`. Absent = the account's own setting. Never
|
|
90
|
+
* sent to another host. */
|
|
91
|
+
dataCollection?: "allow" | "deny";
|
|
87
92
|
/** Default `service_tier` for every call on the OpenAI and Responses shapes;
|
|
88
93
|
* `StreamOptions.serviceTier` wins. Anthropic and Gemini presets refuse it. */
|
|
89
94
|
serviceTier?: ServiceTier;
|
|
@@ -208,6 +213,7 @@ function buildWireAdapter(
|
|
|
208
213
|
...(wire.path ? { path: wire.path } : {}),
|
|
209
214
|
headers,
|
|
210
215
|
...(config.providerOrder ? { providerOrder: config.providerOrder } : {}),
|
|
216
|
+
...(config.dataCollection ? { dataCollection: config.dataCollection } : {}),
|
|
211
217
|
...(preset.sessionHeader ? { sessionHeader: preset.sessionHeader } : {}),
|
|
212
218
|
...(preset.initiatorHeader ? { initiatorHeader: true } : {}),
|
|
213
219
|
...(preset.noSampling ? { noSampling: true } : {}),
|
package/src/providers/gemini.ts
CHANGED
|
@@ -594,7 +594,7 @@ export function createGeminiProvider(config: GeminiConfig): Provider {
|
|
|
594
594
|
// `?alt=sse` commits to 200 the moment the headers go out, so a
|
|
595
595
|
// throttle or an overload landing after that arrives here as a
|
|
596
596
|
// google.rpc.Status in the body rather than as a status line.
|
|
597
|
-
if (chunk.error) throw streamError(id, chunk
|
|
597
|
+
if (chunk.error) throw streamError(id, chunk);
|
|
598
598
|
// A blocked PROMPT has no candidate to finish, so it reads as a cut
|
|
599
599
|
// stream — retried, into the same block. It is a content failure.
|
|
600
600
|
const blocked = chunk.promptFeedback?.blockReason;
|
package/src/providers/openai.ts
CHANGED
|
@@ -103,6 +103,16 @@ export interface OpenAIConfig extends ProviderFallbackConfig {
|
|
|
103
103
|
* when hitting openrouter.ai and providerOrder is not passed. Defaults to true.
|
|
104
104
|
*/
|
|
105
105
|
pinHost?: boolean;
|
|
106
|
+
/**
|
|
107
|
+
* OpenRouter's `provider.data_collection`. `"deny"` routes only to hosts
|
|
108
|
+
* that, by OpenRouter's records, don't train on the prompt; a model with no
|
|
109
|
+
* such host answers 404 instead of falling back to one that does. It is not
|
|
110
|
+
* zero retention: some of those hosts keep prompts for a while, which only
|
|
111
|
+
* OpenRouter's separate `zdr` rules out. Sent on OpenRouter only, with or
|
|
112
|
+
* without a pin; any other host never sees it. Unset leaves the account's
|
|
113
|
+
* own setting in charge.
|
|
114
|
+
*/
|
|
115
|
+
dataCollection?: "allow" | "deny";
|
|
106
116
|
/** Header that carries `StreamOptions.sessionId`. When set and a call has
|
|
107
117
|
* no sessionId, the provider's own per-instance id rides instead, so the
|
|
108
118
|
* gateway never sees a call without one. */
|
|
@@ -654,8 +664,12 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
654
664
|
})()
|
|
655
665
|
: undefined;
|
|
656
666
|
|
|
657
|
-
|
|
658
|
-
|
|
667
|
+
const dataCollection = isOpenRouter(baseUrl) ? config.dataCollection : undefined;
|
|
668
|
+
if (pinOrder?.length || dataCollection) {
|
|
669
|
+
request.provider = {
|
|
670
|
+
...(pinOrder?.length ? { order: pinOrder, allow_fallbacks: true } : {}),
|
|
671
|
+
...(dataCollection ? { data_collection: dataCollection } : {}),
|
|
672
|
+
};
|
|
659
673
|
}
|
|
660
674
|
|
|
661
675
|
const defaultPath = /\/v\d+[^/]*$/i.test(baseUrl.replace(/\/+$/, ""))
|
|
@@ -701,7 +715,7 @@ export function createOpenAIProvider(config: OpenAIConfig): Provider {
|
|
|
701
715
|
// line, and a frame carrying `error` carries no choices: unread, it
|
|
702
716
|
// falls through both branches below and the turn ends as a successful
|
|
703
717
|
// zero-token completion nobody retries.
|
|
704
|
-
if (chunk.error) throw streamError(id, chunk
|
|
718
|
+
if (chunk.error) throw streamError(id, chunk);
|
|
705
719
|
|
|
706
720
|
// A usage-only frame carries no choices — this shape sends it last.
|
|
707
721
|
if (chunk.usage) {
|
|
@@ -689,7 +689,7 @@ export function createResponsesProvider(config: ResponsesConfig): Provider {
|
|
|
689
689
|
// error failed, and one that states `incomplete` was cut. Read as
|
|
690
690
|
// a clean stop, either is a broken turn reported as a finished one.
|
|
691
691
|
if (response?.status === "failed" || response?.error) {
|
|
692
|
-
throw streamError(id, response.error
|
|
692
|
+
throw streamError(id, response.error ? response : { message: "response failed" });
|
|
693
693
|
}
|
|
694
694
|
// Any other status (`cancelled`, `queued`, one added later) says
|
|
695
695
|
// nothing about the answer being whole. Read as a clean stop, it
|
|
@@ -734,7 +734,10 @@ export function createResponsesProvider(config: ResponsesConfig): Provider {
|
|
|
734
734
|
// out: a terminal `response.failed`, or a bare error frame from a
|
|
735
735
|
// gateway in front of it.
|
|
736
736
|
case "response.failed":
|
|
737
|
-
throw streamError(
|
|
737
|
+
throw streamError(
|
|
738
|
+
id,
|
|
739
|
+
event.response?.error ? event.response : { message: "response failed" },
|
|
740
|
+
);
|
|
738
741
|
|
|
739
742
|
case "error":
|
|
740
743
|
case "response.error":
|
|
@@ -742,12 +745,7 @@ export function createResponsesProvider(config: ResponsesConfig): Provider {
|
|
|
742
745
|
// them a spent Codex window reads as an unnamed in-stream failure:
|
|
743
746
|
// floored to overload, then retried and walked across every backup
|
|
744
747
|
// on the same account wall.
|
|
745
|
-
throw streamError(
|
|
746
|
-
id,
|
|
747
|
-
event.error
|
|
748
|
-
? { ...event.error, status_code: event.status_code, headers: event.headers }
|
|
749
|
-
: { message: event.message, code: event.code },
|
|
750
|
-
);
|
|
748
|
+
throw streamError(id, event);
|
|
751
749
|
}
|
|
752
750
|
}
|
|
753
751
|
|
package/src/rate-limit.ts
CHANGED
|
@@ -315,6 +315,28 @@ export function parseUsageLimitBody(bodyText: string, now = Date.now()): RateLim
|
|
|
315
315
|
? (source as Record<string, unknown>)
|
|
316
316
|
: (body as Record<string, unknown>);
|
|
317
317
|
|
|
318
|
+
// OpenCode puts a budget's ISO reset in envelope metadata, or in Google's
|
|
319
|
+
// ErrorInfo details. It names the real horizon, not a short retry hint.
|
|
320
|
+
const envelopes: unknown[] = [
|
|
321
|
+
body,
|
|
322
|
+
error,
|
|
323
|
+
...(Array.isArray(error.details) ? error.details : []),
|
|
324
|
+
];
|
|
325
|
+
let metadataReset = -Infinity;
|
|
326
|
+
for (const envelope of envelopes) {
|
|
327
|
+
if (typeof envelope !== "object" || envelope === null || !("metadata" in envelope)) continue;
|
|
328
|
+
const metadata = envelope.metadata;
|
|
329
|
+
if (
|
|
330
|
+
typeof metadata !== "object" ||
|
|
331
|
+
metadata === null ||
|
|
332
|
+
!("resetsAt" in metadata) ||
|
|
333
|
+
typeof metadata.resetsAt !== "string"
|
|
334
|
+
)
|
|
335
|
+
continue;
|
|
336
|
+
const at = Date.parse(metadata.resetsAt);
|
|
337
|
+
if (Number.isFinite(at)) metadataReset = Math.max(metadataReset, at);
|
|
338
|
+
}
|
|
339
|
+
|
|
318
340
|
const inSeconds = jsonSeconds(error.resets_in_seconds);
|
|
319
341
|
const atSeconds = jsonSeconds(error.resets_at);
|
|
320
342
|
// The plain wait some gateways put in the error body instead of a header,
|
|
@@ -323,12 +345,12 @@ export function parseUsageLimitBody(bodyText: string, now = Date.now()): RateLim
|
|
|
323
345
|
jsonSeconds(error.retry_after) ??
|
|
324
346
|
jsonSeconds(error.retry_after_seconds) ??
|
|
325
347
|
jsonSeconds((body as Record<string, unknown>).retry_after_seconds);
|
|
326
|
-
|
|
327
|
-
if (
|
|
328
|
-
|
|
329
|
-
|
|
330
|
-
if (
|
|
331
|
-
return resetAfter(
|
|
348
|
+
let waitMs: number | undefined;
|
|
349
|
+
if (inSeconds !== undefined && inSeconds >= 0) waitMs = inSeconds * 1000;
|
|
350
|
+
else if (atSeconds !== undefined && atSeconds > 0) waitMs = Math.max(0, atSeconds * 1000 - now);
|
|
351
|
+
else if (afterSeconds !== undefined && afterSeconds >= 0) waitMs = afterSeconds * 1000;
|
|
352
|
+
if (waitMs !== undefined || Number.isFinite(metadataReset)) {
|
|
353
|
+
return resetAfter(Math.max(0, waitMs ?? 0, metadataReset - now), now);
|
|
332
354
|
}
|
|
333
355
|
}
|
|
334
356
|
|