@linxiraos/pi-catalog 1.4.7 → 1.4.9
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/CHANGELOG.md +4 -23
- package/dist/types/compat/auth-ids.d.ts +2 -2
- package/dist/types/compat/catalog-policy.d.ts +6 -4
- package/dist/types/compat/factory-droid.d.ts +7 -1
- package/dist/types/compat/provider-ids.d.ts +1 -1
- package/dist/types/compat/resolve.d.ts +1 -1
- package/dist/types/compat/revision.d.ts +2 -0
- package/dist/types/compat/server-side-fallback.d.ts +6 -4
- package/dist/types/compat/types.d.ts +3 -1
- package/dist/types/discovery/cursor-proto.d.ts +8 -0
- package/dist/types/discovery/openai-compatible.d.ts +6 -0
- package/dist/types/model-manager.d.ts +11 -1
- package/dist/types/model-thinking.d.ts +4 -0
- package/dist/types/models.d.ts +16 -0
- package/dist/types/pricing.d.ts +66 -6
- package/dist/types/provider-models/openai-compat.d.ts +18 -13
- package/dist/types/types.d.ts +14 -19
- package/dist/types/utils.d.ts +8 -0
- package/package.json +4 -4
- package/src/compat/auth-ids.ts +5 -0
- package/src/compat/axes.ts +21 -9
- package/src/compat/catalog-policy.d.ts +9 -0
- package/src/compat/catalog-policy.ts +14 -10
- package/src/compat/collapse.ts +1 -1
- package/src/compat/factory-droid.d.ts +7 -1
- package/src/compat/factory-droid.ts +9 -1
- package/src/compat/provider-ids.d.ts +1 -1
- package/src/compat/provider-ids.ts +2 -0
- package/src/compat/resolve.d.ts +1 -1
- package/src/compat/resolve.ts +4 -5
- package/src/compat/revision.d.ts +2 -0
- package/src/compat/revision.ts +10 -0
- package/src/compat/rules/README.md +3 -1
- package/src/compat/rules/auth/_order.kdl +1 -1
- package/src/compat/rules/auth/mistral.kdl +4 -0
- package/src/compat/rules/auth/modular.kdl +9 -0
- package/src/compat/rules/auth/scaleway.kdl +17 -0
- package/src/compat/rules/classes/anthropic.kdl +5 -0
- package/src/compat/rules/classes/deepseek.kdl +1 -1
- package/src/compat/rules/classes/gemini.kdl +1 -1
- package/src/compat/rules/classes/xai.kdl +1 -1
- package/src/compat/rules/providers/anthropic.kdl +8 -4
- package/src/compat/rules/providers/cursor.kdl +10 -0
- package/src/compat/rules/providers/deepseek.kdl +2 -2
- package/src/compat/rules/providers/factory-droid.kdl +219 -13
- package/src/compat/rules/providers/modular.kdl +35 -0
- package/src/compat/rules/providers/openai-codex.kdl +1 -1
- package/src/compat/rules/providers/sambanova.kdl +2 -0
- package/src/compat/rules/providers/scaleway.kdl +179 -0
- package/src/compat/rules/runtime/behavior.kdl +13 -0
- package/src/compat/rules.json +1 -1
- package/src/compat/server-side-fallback.d.ts +9 -0
- package/src/compat/server-side-fallback.ts +7 -5
- package/src/compat/types.d.ts +3 -1
- package/src/compat/types.ts +3 -1
- package/src/discovery/codex.ts +2 -9
- package/src/discovery/cursor-proto.d.ts +8 -0
- package/src/discovery/cursor-proto.ts +12 -1
- package/src/discovery/factory-droid.ts +35 -11
- package/src/discovery/openai-compatible.d.ts +6 -0
- package/src/discovery/openai-compatible.ts +1 -1
- package/src/model-cache.ts +32 -58
- package/src/model-catalog-format.ts +5 -8
- package/src/model-manager.d.ts +11 -1
- package/src/model-manager.ts +85 -143
- package/src/model-thinking.d.ts +4 -0
- package/src/model-thinking.ts +10 -0
- package/src/model-tokenizer.ts +13 -12
- package/src/models.d.ts +16 -0
- package/src/models.json +1 -1
- package/src/models.ts +32 -1
- package/src/pricing.d.ts +66 -6
- package/src/pricing.ts +59 -94
- package/src/provider-models/catalog-snapshot.json.zst +0 -0
- package/src/provider-models/catalog-snapshot.json.zst.d.ts +3 -0
- package/src/provider-models/descriptors.ts +2 -0
- package/src/provider-models/openai-compat.d.ts +18 -13
- package/src/provider-models/openai-compat.ts +120 -71
- package/src/types.d.ts +14 -19
- package/src/types.ts +11 -18
- package/src/utils.d.ts +8 -0
- package/src/utils.ts +18 -0
- package/src/wire/factory-droid.ts +4 -3
package/src/model-manager.ts
CHANGED
|
@@ -1,19 +1,12 @@
|
|
|
1
|
+
import { type } from "@linxiraos/pi-omptype";
|
|
1
2
|
import { logger, VERSION } from "@linxiraos/pi-utils";
|
|
2
3
|
import { buildModel } from "./build";
|
|
3
4
|
import { collapseBuiltVariants } from "./compat/collapse";
|
|
4
5
|
import { applyCatalogMetrics, CatalogMetricsIndex } from "./identity/metrics";
|
|
5
6
|
import { readModelCache, writeModelCache } from "./model-cache";
|
|
6
7
|
import { type GeneratedProvider, getBundledModels } from "./models";
|
|
7
|
-
import {
|
|
8
|
-
import {
|
|
9
|
-
type Api,
|
|
10
|
-
type Model,
|
|
11
|
-
type ModelCost,
|
|
12
|
-
modelKind,
|
|
13
|
-
type ModelSpec,
|
|
14
|
-
type Provider,
|
|
15
|
-
type TokenCost,
|
|
16
|
-
} from "./types";
|
|
8
|
+
import { TimeBasedCostShape } from "./pricing";
|
|
9
|
+
import { type Api, type Model, modelKind, type ModelSpec, type Provider } from "./types";
|
|
17
10
|
import { isRecord } from "./utils";
|
|
18
11
|
|
|
19
12
|
const DEFAULT_CACHE_TTL_MS = 2 * 60 * 60 * 1000;
|
|
@@ -267,7 +260,6 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
267
260
|
const cache = readModelCache<TApi>(cacheProviderId, ttlMs, now, dbPath);
|
|
268
261
|
const dynamicModelsAuthoritative = options.dynamicModelsAuthoritative ?? false;
|
|
269
262
|
const isAuthoritativeDynamicCache = dynamicModelsAuthoritative && (cache?.payloadAuthoritative ?? false);
|
|
270
|
-
const cacheAdditiveStaticModelIds = isAuthoritativeDynamicCache ? undefined : modelsDevAdditiveStaticModelIds;
|
|
271
263
|
const restoredCache = restoreCachedModelHeaders(
|
|
272
264
|
cache?.models ?? [],
|
|
273
265
|
staticModels,
|
|
@@ -286,7 +278,12 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
286
278
|
staticNames: options.keepStaticNames ? new Map(staticModels.map(model => [model.id, model.name])) : undefined,
|
|
287
279
|
};
|
|
288
280
|
const cacheDropIds = options.dropCachedModelIdsOnStaticMismatch;
|
|
289
|
-
|
|
281
|
+
// Additive mode is cache identity: a snapshot written while shared-catalog
|
|
282
|
+
// rows could replace bundled metadata must not match an additive resolver.
|
|
283
|
+
const staticCatalogFingerprint = fingerprintStaticModels(staticModels, {
|
|
284
|
+
authoritative: dynamicModelsAuthoritative,
|
|
285
|
+
additive: modelsDevAdditiveStaticModelIds !== undefined,
|
|
286
|
+
});
|
|
290
287
|
// Endpoint-migration policy is cache identity: adding an id must invalidate
|
|
291
288
|
// matching-static-catalog caches written by the prior resolver.
|
|
292
289
|
const staticFingerprint =
|
|
@@ -294,6 +291,13 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
294
291
|
? `${staticCatalogFingerprint}:drop:${Bun.hash(cacheDropIds.join("\0")).toString(36)}`
|
|
295
292
|
: staticCatalogFingerprint;
|
|
296
293
|
const cacheFingerprintMatches = cache?.staticFingerprint === staticFingerprint && staticFingerprint.length > 0;
|
|
294
|
+
// A matching cache was written against this static catalog under the same
|
|
295
|
+
// additive policy, so its same-id rows are live discovery merged over the
|
|
296
|
+
// bundled rows, never shared-catalog overrides (#17481). Strip them only on
|
|
297
|
+
// a mismatch: a static-catalog upgrade or a cache written before additive
|
|
298
|
+
// semantics were enabled.
|
|
299
|
+
const cacheAdditiveStaticModelIds =
|
|
300
|
+
isAuthoritativeDynamicCache || cacheFingerprintMatches ? undefined : modelsDevAdditiveStaticModelIds;
|
|
297
301
|
const cacheNeedsModelMigration =
|
|
298
302
|
!cacheFingerprintMatches &&
|
|
299
303
|
cacheDropIds !== undefined &&
|
|
@@ -315,8 +319,8 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
315
319
|
// Cold-start fast path: when a fresh, authoritative cache exists, the network
|
|
316
320
|
// fetch is skipped, AND the static catalog slice is byte-identical to what
|
|
317
321
|
// was merged in last time, the cache row IS the authoritative merge result.
|
|
318
|
-
//
|
|
319
|
-
//
|
|
322
|
+
// A mismatched additive cache never reaches here; its same-id rows are
|
|
323
|
+
// stripped below.
|
|
320
324
|
if (
|
|
321
325
|
!shouldFetchFromNetwork &&
|
|
322
326
|
cache?.fresh &&
|
|
@@ -324,18 +328,9 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
324
328
|
cacheFingerprintMatches &&
|
|
325
329
|
!cacheHasUnresolvedHeaders
|
|
326
330
|
) {
|
|
327
|
-
const
|
|
328
|
-
? restoredCache.models.filter(model => !cacheAdditiveStaticModelIds.has(model.id))
|
|
329
|
-
: restoredCache.models;
|
|
330
|
-
const cachedModels = cacheAdditiveStaticModelIds
|
|
331
|
-
? mergeCatalogMetrics(
|
|
332
|
-
mergeDynamicModels(staticModels, cacheContribution, mergeAuthority),
|
|
333
|
-
restoredCache.models,
|
|
334
|
-
)
|
|
335
|
-
: restoredCache.models;
|
|
336
|
-
const source: ModelResolutionSource = cacheContribution.length > 0 ? "cache" : "bundled";
|
|
331
|
+
const source: ModelResolutionSource = restoredCache.models.length > 0 ? "cache" : "bundled";
|
|
337
332
|
return {
|
|
338
|
-
models: collapseBuiltVariants(
|
|
333
|
+
models: collapseBuiltVariants(restoredCache.models),
|
|
339
334
|
stale: false,
|
|
340
335
|
source,
|
|
341
336
|
...(source === "cache" ? { updatedAt: cache.updatedAt } : {}),
|
|
@@ -382,7 +377,10 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
382
377
|
dynamicModels.length > 0 &&
|
|
383
378
|
(dynamicModelsAuthoritative || !hasModelsDevFetcher || modelsDevFetchSucceeded)
|
|
384
379
|
: modelsDevFetchSucceeded;
|
|
385
|
-
const mergedWithCache =
|
|
380
|
+
const mergedWithCache = mergeCachedModels(staticModels, cacheModels, {
|
|
381
|
+
authority: mergeAuthority,
|
|
382
|
+
matchesStatic: cacheFingerprintMatches,
|
|
383
|
+
});
|
|
386
384
|
const mergedWithModelsDev = mergeDynamicModels(
|
|
387
385
|
mergedWithCache,
|
|
388
386
|
modelsDevModels,
|
|
@@ -455,7 +453,10 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
|
|
|
455
453
|
? preparedLatestCacheModels.filter(model => !cacheAdditiveStaticModelIds.has(model.id))
|
|
456
454
|
: preparedLatestCacheModels;
|
|
457
455
|
const mergedFallbackModels = mergeDynamicModels(
|
|
458
|
-
|
|
456
|
+
mergeCachedModels(staticModels, latestCacheModels, {
|
|
457
|
+
authority: mergeAuthority,
|
|
458
|
+
matchesStatic: cacheFingerprintMatches,
|
|
459
|
+
}),
|
|
459
460
|
modelsDevModels,
|
|
460
461
|
mergeAuthority,
|
|
461
462
|
);
|
|
@@ -631,6 +632,24 @@ function mergeDynamicModels<TApi extends Api>(
|
|
|
631
632
|
return Array.from(merged.values());
|
|
632
633
|
}
|
|
633
634
|
|
|
635
|
+
/**
|
|
636
|
+
* Lays cached rows over the static catalog. A snapshot written against this
|
|
637
|
+
* static catalog (`matchesStatic`) is already that catalog merged with live
|
|
638
|
+
* discovery, so its rows replace same-id static rows verbatim; re-merging would
|
|
639
|
+
* only rebuild them and churn the persisted payload. A snapshot from another
|
|
640
|
+
* static catalog re-merges so upgraded bundled metadata applies.
|
|
641
|
+
*/
|
|
642
|
+
function mergeCachedModels<TApi extends Api>(
|
|
643
|
+
staticModels: readonly Model<TApi>[],
|
|
644
|
+
cacheModels: readonly Model<TApi>[],
|
|
645
|
+
options: { authority: DynamicMergeAuthority; matchesStatic: boolean },
|
|
646
|
+
): Model<TApi>[] {
|
|
647
|
+
if (!options.matchesStatic) return mergeDynamicModels(staticModels, cacheModels, options.authority);
|
|
648
|
+
const merged = new Map<string, Model<TApi>>(staticModels.map(model => [model.id, model]));
|
|
649
|
+
for (const cachedModel of cacheModels) merged.set(cachedModel.id, cachedModel);
|
|
650
|
+
return Array.from(merged.values());
|
|
651
|
+
}
|
|
652
|
+
|
|
634
653
|
function retainModelIds<TApi extends Api>(
|
|
635
654
|
models: readonly Model<TApi>[],
|
|
636
655
|
retainedModels: readonly Model<TApi>[],
|
|
@@ -642,6 +661,17 @@ function retainModelIds<TApi extends Api>(
|
|
|
642
661
|
|
|
643
662
|
const MODEL_CACHE_FINGERPRINT_VERSION = `merge-v5:${VERSION}`;
|
|
644
663
|
|
|
664
|
+
/**
|
|
665
|
+
* Merge policy that is part of a cache's identity in
|
|
666
|
+
* {@link fingerprintStaticModels}: `authoritative` when dynamic discovery owns
|
|
667
|
+
* the catalog, `additive` when the shared catalog may only add ids. The flags
|
|
668
|
+
* compose; snapshots written under different policies never match.
|
|
669
|
+
*/
|
|
670
|
+
export interface StaticFingerprintPolicy {
|
|
671
|
+
authoritative?: boolean;
|
|
672
|
+
additive?: boolean;
|
|
673
|
+
}
|
|
674
|
+
|
|
645
675
|
/**
|
|
646
676
|
* Return the versioned content identity for a static provider slice. Callers
|
|
647
677
|
* may mutate or freeze their arrays, so every call observes current content
|
|
@@ -649,12 +679,11 @@ const MODEL_CACHE_FINGERPRINT_VERSION = `merge-v5:${VERSION}`;
|
|
|
649
679
|
*/
|
|
650
680
|
export function fingerprintStaticModels<TApi extends Api>(
|
|
651
681
|
models: readonly (ModelSpec<TApi> | Model<TApi>)[],
|
|
652
|
-
|
|
682
|
+
policy: StaticFingerprintPolicy = {},
|
|
653
683
|
): string {
|
|
654
|
-
|
|
655
|
-
if (
|
|
656
|
-
|
|
657
|
-
return `${MODEL_CACHE_FINGERPRINT_VERSION}:${Bun.hash(JSON.stringify(models)).toString(36)}`;
|
|
684
|
+
const policyKey = `${policy.authoritative ? ":authoritative" : ""}${policy.additive ? ":additive" : ""}`;
|
|
685
|
+
if (models.length === 0) return `${MODEL_CACHE_FINGERPRINT_VERSION}${policyKey}:empty`;
|
|
686
|
+
return `${MODEL_CACHE_FINGERPRINT_VERSION}${policyKey}:${Bun.hash(JSON.stringify(models)).toString(36)}`;
|
|
658
687
|
}
|
|
659
688
|
|
|
660
689
|
function mergeDynamicModel<TApi extends Api>(
|
|
@@ -782,18 +811,19 @@ function preferDiscoveryLimit(discoveryLimit: number | null, fallbackLimit: numb
|
|
|
782
811
|
* caller treats the source as failed (cache/static fallback) rather than as an
|
|
783
812
|
* intentionally empty catalog that an authoritative provider would prune to.
|
|
784
813
|
*/
|
|
785
|
-
function buildDiscoveredModelSpecs<TApi extends Api>(
|
|
814
|
+
function buildDiscoveredModelSpecs<TApi extends Api>(values: unknown): DiscoveredModelSet<TApi> {
|
|
786
815
|
const models: Model<TApi>[] = [];
|
|
787
816
|
const explicitKindModels = new Set<Model<TApi>>();
|
|
788
|
-
if (!Array.isArray(
|
|
817
|
+
if (!Array.isArray(values)) {
|
|
789
818
|
return { models, explicitKindModels };
|
|
790
819
|
}
|
|
791
820
|
const failures: unknown[] = [];
|
|
792
|
-
for (const
|
|
793
|
-
if (!
|
|
821
|
+
for (const value of values) {
|
|
822
|
+
if (!ModelSpecShape.allows(value)) continue;
|
|
823
|
+
const item = value as ModelSpec<TApi>;
|
|
794
824
|
let model: Model<TApi>;
|
|
795
825
|
try {
|
|
796
|
-
model = buildModel(item
|
|
826
|
+
model = buildModel(item);
|
|
797
827
|
} catch (error) {
|
|
798
828
|
failures.push(error);
|
|
799
829
|
logger.warn("model discovery skipped a model that failed to build", {
|
|
@@ -810,109 +840,21 @@ function buildDiscoveredModelSpecs<TApi extends Api>(value: unknown): Discovered
|
|
|
810
840
|
return { models, explicitKindModels };
|
|
811
841
|
}
|
|
812
842
|
|
|
813
|
-
|
|
814
|
-
|
|
815
|
-
|
|
816
|
-
|
|
817
|
-
|
|
818
|
-
|
|
819
|
-
|
|
820
|
-
|
|
821
|
-
|
|
822
|
-
|
|
823
|
-
|
|
824
|
-
|
|
825
|
-
|
|
826
|
-
|
|
827
|
-
|
|
828
|
-
|
|
829
|
-
|
|
830
|
-
|
|
831
|
-
}
|
|
832
|
-
if (typeof v.name !== "string" || v.name.length === 0) {
|
|
833
|
-
return false;
|
|
834
|
-
}
|
|
835
|
-
if (typeof v.api !== "string" || v.api.length === 0) {
|
|
836
|
-
return false;
|
|
837
|
-
}
|
|
838
|
-
if (typeof v.provider !== "string" || v.provider.length === 0) {
|
|
839
|
-
return false;
|
|
840
|
-
}
|
|
841
|
-
if (typeof v.baseUrl !== "string" || v.baseUrl.length === 0) {
|
|
842
|
-
return false;
|
|
843
|
-
}
|
|
844
|
-
if (typeof v.reasoning !== "boolean") {
|
|
845
|
-
return false;
|
|
846
|
-
}
|
|
847
|
-
if (!isModelInputArray(v.input)) {
|
|
848
|
-
return false;
|
|
849
|
-
}
|
|
850
|
-
if (!isModelCost(v.cost)) {
|
|
851
|
-
return false;
|
|
852
|
-
}
|
|
853
|
-
// Finite positive: NaN > 0 is false, +Infinity < Infinity is false.
|
|
854
|
-
const cw = v.contextWindow;
|
|
855
|
-
if (cw !== null && (typeof cw !== "number" || !(cw > 0 && cw < Infinity))) {
|
|
856
|
-
return false;
|
|
857
|
-
}
|
|
858
|
-
const mt = v.maxTokens;
|
|
859
|
-
if (mt !== null && (typeof mt !== "number" || !(mt > 0 && mt < Infinity))) {
|
|
860
|
-
return false;
|
|
861
|
-
}
|
|
862
|
-
return true;
|
|
863
|
-
}
|
|
864
|
-
|
|
865
|
-
function isModelInputArray(value: unknown): value is ("text" | "image")[] {
|
|
866
|
-
if (!Array.isArray(value) || value.length === 0) {
|
|
867
|
-
return false;
|
|
868
|
-
}
|
|
869
|
-
for (let i = 0; i < value.length; i++) {
|
|
870
|
-
const item = value[i];
|
|
871
|
-
if (item !== "text" && item !== "image") {
|
|
872
|
-
return false;
|
|
873
|
-
}
|
|
874
|
-
}
|
|
875
|
-
return true;
|
|
876
|
-
}
|
|
877
|
-
|
|
878
|
-
function isTokenCost(value: unknown): value is TokenCost {
|
|
879
|
-
if (!isRecord(value)) {
|
|
880
|
-
return false;
|
|
881
|
-
}
|
|
882
|
-
const c = value as {
|
|
883
|
-
input?: unknown;
|
|
884
|
-
output?: unknown;
|
|
885
|
-
cacheRead?: unknown;
|
|
886
|
-
cacheWrite?: unknown;
|
|
887
|
-
};
|
|
888
|
-
// Finite (NaN-safe): -Infinity < x < Infinity rejects NaN and both infinities.
|
|
889
|
-
// Preserves original behavior: 0 and negatives remain valid.
|
|
890
|
-
const ci = c.input;
|
|
891
|
-
if (typeof ci !== "number" || !(ci > -Infinity && ci < Infinity)) {
|
|
892
|
-
return false;
|
|
893
|
-
}
|
|
894
|
-
const co = c.output;
|
|
895
|
-
if (typeof co !== "number" || !(co > -Infinity && co < Infinity)) {
|
|
896
|
-
return false;
|
|
897
|
-
}
|
|
898
|
-
const cr = c.cacheRead;
|
|
899
|
-
if (typeof cr !== "number" || !(cr > -Infinity && cr < Infinity)) {
|
|
900
|
-
return false;
|
|
901
|
-
}
|
|
902
|
-
const cw = c.cacheWrite;
|
|
903
|
-
if (typeof cw !== "number" || !(cw > -Infinity && cw < Infinity)) {
|
|
904
|
-
return false;
|
|
905
|
-
}
|
|
906
|
-
return true;
|
|
907
|
-
}
|
|
908
|
-
|
|
909
|
-
function isModelCost(value: unknown): value is ModelCost {
|
|
910
|
-
if (!isTokenCost(value)) return false;
|
|
911
|
-
const cost = value as TokenCost & { longContext?: unknown; timeBased?: unknown };
|
|
912
|
-
if (cost.timeBased !== undefined && !isTimeBasedCost(cost.timeBased)) return false;
|
|
913
|
-
const longContext = cost.longContext;
|
|
914
|
-
if (longContext === undefined) return true;
|
|
915
|
-
if (!isTokenCost(longContext) || !isRecord(longContext)) return false;
|
|
916
|
-
const threshold = longContext.inputThreshold;
|
|
917
|
-
return typeof threshold === "number" && threshold > 0 && threshold < Infinity;
|
|
918
|
-
}
|
|
843
|
+
/** Finite rates; 0 and negatives stay valid. */
|
|
844
|
+
const FiniteTokenCost = type({ input: "number", output: "number", cacheRead: "number", cacheWrite: "number" });
|
|
845
|
+
/** Discovery output is built in memory, so optional fields may be present as `undefined`. */
|
|
846
|
+
const ModelSpecShape = type({
|
|
847
|
+
id: "string > 0",
|
|
848
|
+
name: "string > 0",
|
|
849
|
+
api: "string > 0",
|
|
850
|
+
provider: "string > 0",
|
|
851
|
+
baseUrl: "string > 0",
|
|
852
|
+
reasoning: "boolean",
|
|
853
|
+
input: "('text' | 'image')[] > 0",
|
|
854
|
+
cost: FiniteTokenCost.and({
|
|
855
|
+
"timeBased?": TimeBasedCostShape.or("undefined"),
|
|
856
|
+
"longContext?": FiniteTokenCost.and({ inputThreshold: "number > 0" }).or("undefined"),
|
|
857
|
+
}),
|
|
858
|
+
contextWindow: "number > 0 | null",
|
|
859
|
+
maxTokens: "number > 0 | null",
|
|
860
|
+
});
|
package/src/model-thinking.d.ts
CHANGED
|
@@ -32,6 +32,10 @@ export declare function withoutUnverifiedThinking<T extends {
|
|
|
32
32
|
/**
|
|
33
33
|
* Clamps a requested thinking level against explicit model metadata.
|
|
34
34
|
*
|
|
35
|
+
* An unserved level first follows `thinking.effortMap` when the map sends it to
|
|
36
|
+
* a served tier that reaches the wire unchanged (a gateway answering `medium`
|
|
37
|
+
* as `high`); otherwise it climbs down to the nearest lower tier.
|
|
38
|
+
*
|
|
35
39
|
* Non-reasoning models always resolve to `undefined`.
|
|
36
40
|
*/
|
|
37
41
|
export declare function clampThinkingLevelForModel<TApi extends Api>(model: ApiModel<TApi> | undefined, requested: Effort | undefined): Effort | undefined;
|
package/src/model-thinking.ts
CHANGED
|
@@ -42,6 +42,10 @@ export function withoutUnverifiedThinking<T extends { reasoning: boolean; thinki
|
|
|
42
42
|
/**
|
|
43
43
|
* Clamps a requested thinking level against explicit model metadata.
|
|
44
44
|
*
|
|
45
|
+
* An unserved level first follows `thinking.effortMap` when the map sends it to
|
|
46
|
+
* a served tier that reaches the wire unchanged (a gateway answering `medium`
|
|
47
|
+
* as `high`); otherwise it climbs down to the nearest lower tier.
|
|
48
|
+
*
|
|
45
49
|
* Non-reasoning models always resolve to `undefined`.
|
|
46
50
|
*/
|
|
47
51
|
export function clampThinkingLevelForModel<TApi extends Api>(
|
|
@@ -60,6 +64,12 @@ export function clampThinkingLevelForModel<TApi extends Api>(
|
|
|
60
64
|
return requested;
|
|
61
65
|
}
|
|
62
66
|
|
|
67
|
+
const effortMap = model.thinking?.effortMap;
|
|
68
|
+
const aliased = effortMap?.[requested] as Effort | undefined;
|
|
69
|
+
if (aliased !== undefined && levels.includes(aliased) && (effortMap![aliased] ?? aliased) === aliased) {
|
|
70
|
+
return aliased;
|
|
71
|
+
}
|
|
72
|
+
|
|
63
73
|
const requestedIndex = THINKING_EFFORTS.indexOf(requested);
|
|
64
74
|
if (requestedIndex === -1) {
|
|
65
75
|
return undefined;
|
package/src/model-tokenizer.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import {
|
|
1
|
+
import { revisionAtLeast } from "./compat/revision";
|
|
2
2
|
import { classifyModel } from "./compat/taxonomy";
|
|
3
3
|
import type { ModelIdentity } from "./compat/types";
|
|
4
4
|
import { bareModelId } from "./identity/id";
|
|
@@ -8,15 +8,6 @@ import type { ModelTokenizer } from "./types";
|
|
|
8
8
|
const MAX_TOKENIZER_CACHE_ENTRIES = 2_048;
|
|
9
9
|
const modelTokenizerCache = new Map<string, ModelTokenizer | null>();
|
|
10
10
|
|
|
11
|
-
function revisionAtLeast(revision: string | undefined, floor: string): boolean {
|
|
12
|
-
if (revision === undefined) return false;
|
|
13
|
-
const parsedRevision = parseRevision(revision);
|
|
14
|
-
const parsedFloor = parseRevision(floor);
|
|
15
|
-
return (
|
|
16
|
-
parsedRevision !== undefined && parsedFloor !== undefined && compareRevision(parsedRevision, parsedFloor) >= 0
|
|
17
|
-
);
|
|
18
|
-
}
|
|
19
|
-
|
|
20
11
|
function claudeTokenizer(identity: ModelIdentity): ModelTokenizer | undefined {
|
|
21
12
|
if (identity.class !== "anthropic") return undefined;
|
|
22
13
|
if (identity.family === "opus") {
|
|
@@ -49,9 +40,19 @@ function glmTokenizer(identity: ModelIdentity): ModelTokenizer | undefined {
|
|
|
49
40
|
return identity.class === "glm" && revisionAtLeast(identity.revision, "5") ? "glm5" : undefined;
|
|
50
41
|
}
|
|
51
42
|
|
|
52
|
-
// GPT-5 and later bill text in o200k_base
|
|
43
|
+
// GPT-5 and later bill text in o200k_base, GPT-3.5 and GPT-4 (Turbo) in cl100k_base. GPT-4.1, GPT-4.5, gpt-4o and the
|
|
44
|
+
// o-series are o200k too, which is the default for a model without a tokenizer.
|
|
53
45
|
function openAITokenizer(identity: ModelIdentity): ModelTokenizer | undefined {
|
|
54
|
-
|
|
46
|
+
if (identity.class !== "openai") return undefined;
|
|
47
|
+
if (revisionAtLeast(identity.revision, "5")) return "o200k";
|
|
48
|
+
if (
|
|
49
|
+
identity.family === "gpt" &&
|
|
50
|
+
revisionAtLeast(identity.revision, "3.5") &&
|
|
51
|
+
!revisionAtLeast(identity.revision, "4.1")
|
|
52
|
+
) {
|
|
53
|
+
return "cl100k";
|
|
54
|
+
}
|
|
55
|
+
return undefined;
|
|
55
56
|
}
|
|
56
57
|
|
|
57
58
|
/**
|
package/src/models.d.ts
CHANGED
|
@@ -11,6 +11,22 @@ export declare function getBundledProviders(): readonly KnownProvider[];
|
|
|
11
11
|
export declare function getBundledModels(provider: GeneratedProvider): Model<Api>[];
|
|
12
12
|
/** Resolve display semantics without mistaking an absent rate card for a free model. */
|
|
13
13
|
export declare function getModelPricingStatus(model: Pick<Model<Api>, "cost" | "pricingStatus">): ModelPricingStatus;
|
|
14
|
+
/** Factory Droid credit rate in force at a time, and the next dated change to it. */
|
|
15
|
+
export interface FactoryDroidCreditRate {
|
|
16
|
+
rate: number;
|
|
17
|
+
/** The earliest later `credit-rates` card that changes `rate`: the rate it sets and when (Unix ms). */
|
|
18
|
+
next?: {
|
|
19
|
+
rate: number;
|
|
20
|
+
at: number;
|
|
21
|
+
};
|
|
22
|
+
}
|
|
23
|
+
/**
|
|
24
|
+
* Factory Droid Standard Credits rate at `timestamp` (Unix ms, default now):
|
|
25
|
+
* the `input` of the `credit-rates` card due then (the latest at or before
|
|
26
|
+
* it, as for cost cards), else the base rate. Undefined for models without a
|
|
27
|
+
* credit rate.
|
|
28
|
+
*/
|
|
29
|
+
export declare function getFactoryDroidCreditRate(model: Pick<Model, "factoryDroidCredits" | "factoryDroidCreditCards">, timestamp?: number): FactoryDroidCreditRate | undefined;
|
|
14
30
|
/**
|
|
15
31
|
* The base token card in force at `timestamp` (Unix ms, default now): the
|
|
16
32
|
* latest due `timeBased` effective rate, else the row's own legs. Price
|