@oh-my-pi/pi-catalog 18.14.2 → 18.16.0

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.
Files changed (52) hide show
  1. package/CHANGELOG.md +30 -0
  2. package/dist/types/compat/auth-ids.d.ts +2 -2
  3. package/dist/types/compat/catalog-policy.d.ts +6 -4
  4. package/dist/types/compat/factory-droid.d.ts +7 -1
  5. package/dist/types/compat/provider-ids.d.ts +1 -1
  6. package/dist/types/compat/resolve.d.ts +1 -1
  7. package/dist/types/compat/revision.d.ts +2 -0
  8. package/dist/types/compat/server-side-fallback.d.ts +6 -4
  9. package/dist/types/compat/types.d.ts +3 -1
  10. package/dist/types/discovery/openai-compatible.d.ts +6 -0
  11. package/dist/types/models.d.ts +16 -0
  12. package/dist/types/pricing.d.ts +66 -6
  13. package/dist/types/provider-models/openai-compat.d.ts +18 -0
  14. package/dist/types/types.d.ts +11 -5
  15. package/dist/types/utils.d.ts +8 -0
  16. package/package.json +4 -4
  17. package/src/compat/auth-ids.ts +5 -0
  18. package/src/compat/axes.ts +5 -2
  19. package/src/compat/catalog-policy.ts +14 -10
  20. package/src/compat/collapse.ts +1 -1
  21. package/src/compat/factory-droid.ts +9 -1
  22. package/src/compat/provider-ids.ts +2 -0
  23. package/src/compat/resolve.ts +4 -2
  24. package/src/compat/revision.ts +10 -0
  25. package/src/compat/rules/README.md +2 -0
  26. package/src/compat/rules/auth/_order.kdl +1 -1
  27. package/src/compat/rules/auth/mistral.kdl +4 -0
  28. package/src/compat/rules/auth/modular.kdl +9 -0
  29. package/src/compat/rules/auth/scaleway.kdl +17 -0
  30. package/src/compat/rules/providers/factory-droid.kdl +219 -13
  31. package/src/compat/rules/providers/modular.kdl +35 -0
  32. package/src/compat/rules/providers/sambanova.kdl +2 -0
  33. package/src/compat/rules/providers/scaleway.kdl +179 -0
  34. package/src/compat/rules/runtime/behavior.kdl +13 -0
  35. package/src/compat/rules.json +1 -1
  36. package/src/compat/server-side-fallback.ts +7 -5
  37. package/src/compat/types.ts +3 -1
  38. package/src/discovery/codex.ts +2 -9
  39. package/src/discovery/factory-droid.ts +35 -11
  40. package/src/discovery/openai-compatible.ts +1 -1
  41. package/src/model-cache.ts +32 -58
  42. package/src/model-catalog-format.ts +5 -8
  43. package/src/model-manager.ts +27 -121
  44. package/src/model-tokenizer.ts +13 -12
  45. package/src/models.json +1 -1
  46. package/src/models.ts +32 -1
  47. package/src/pricing.ts +59 -94
  48. package/src/provider-models/descriptors.ts +2 -0
  49. package/src/provider-models/openai-compat.ts +60 -59
  50. package/src/types.ts +11 -4
  51. package/src/utils.ts +18 -0
  52. package/src/wire/factory-droid.ts +4 -3
@@ -1,13 +1,15 @@
1
- import type { Model } from "../types";
1
+ import type { CatalogPolicyInputs } from "../types";
2
2
  import { resolveCatalogPolicy } from "./catalog-policy";
3
+ import type { ResolveRoute } from "./resolve";
3
4
 
4
5
  /**
5
6
  * Ordered model ids for Anthropic's server-side refusal `fallbacks` chain,
6
- * authored on the `server-side-fallback-models` catalog axis. Empty when no
7
- * rule assigns a chain (the model or host does not support the feature).
7
+ * authored on the `server-side-fallback-models` catalog axis, on `route`'s
8
+ * upstream when it names one. Empty when no rule assigns a chain (the model,
9
+ * host or upstream does not support the feature).
8
10
  */
9
- export function serverSideFallbackModels(model: Model): readonly string[] {
10
- const value = resolveCatalogPolicy(model).serverSideFallbackModels;
11
+ export function serverSideFallbackModels(model: CatalogPolicyInputs, route?: ResolveRoute): readonly string[] {
12
+ const value = resolveCatalogPolicy(model, route).serverSideFallbackModels;
11
13
  if (!Array.isArray(value)) return [];
12
14
  return value.filter((entry): entry is string => typeof entry === "string");
13
15
  }
@@ -753,6 +753,8 @@ export interface ResolvedAxes {
753
753
  export interface RequestPolicy {
754
754
  completionsReasoningMode?: "none" | "effort" | "opt-in" | "forced-on";
755
755
  completionsReasoningHistory?: "omit" | "preserved" | "interleaved";
756
+ /** Send `allow_missing_thinking` with the effort so tool-call turns may arrive without reasoning. */
757
+ completionsAllowMissingThinking?: boolean;
756
758
  anthropicThinking?: "adaptive" | "adaptive-summarized" | "budget-interleaved" | "budget-effort";
757
759
  /** Advertise `fine-grained-tool-streaming-2025-05-14` on requests that carry tools. */
758
760
  anthropicToolStreamingBeta?: boolean;
@@ -760,7 +762,7 @@ export interface RequestPolicy {
760
762
  openaiPlatformHeader?: string;
761
763
  responsesCacheRetention?: boolean;
762
764
  responsesVerbosity?: "low";
763
- responsesServiceTier?: "priority";
765
+ responsesServiceTier?: "priority" | "ultrafast";
764
766
  responsesParallelToolCalls?: boolean;
765
767
  responsesSafetyIdentifier?: boolean;
766
768
  /** Default `tool_choice` to `auto` when the request carries tools and the caller picks none. */
@@ -3,7 +3,7 @@ import { compareRevision, parseRevision } from "../compat/revision";
3
3
  import { classifyModel } from "../compat/taxonomy";
4
4
  import { getBundledModels } from "../models";
5
5
  import type { FetchImpl, ModelSpec } from "../types";
6
- import { discoveryFetch } from "../utils";
6
+ import { discoveryFetch, parseServiceTierIds } from "../utils";
7
7
  import { CODEX_BASE_URL, CODEX_CLIENT_VERSION, OPENAI_HEADER_VALUES, OPENAI_HEADERS } from "../wire/codex";
8
8
 
9
9
  const DEFAULT_MODEL_LIST_PATHS = ["/codex/models", "/models"] as const;
@@ -334,14 +334,7 @@ function parseCodexModelEntry(entry: unknown): ParsedCodexModelEntry | null {
334
334
 
335
335
  // codex-rs `ModelServiceTier { id, name, description }`; only the id reaches the wire.
336
336
  // An explicit empty array is kept: it means the model offers no optional tier.
337
- let serviceTiers: string[] | undefined;
338
- if (Array.isArray(payload.service_tiers)) {
339
- serviceTiers = [];
340
- for (const tier of payload.service_tiers) {
341
- const id = tier !== null && typeof tier === "object" && "id" in tier ? toNonEmptyString(tier.id) : null;
342
- if (id && !serviceTiers.includes(id)) serviceTiers.push(id);
343
- }
344
- }
337
+ const serviceTiers = parseServiceTierIds(payload.service_tiers);
345
338
 
346
339
  return {
347
340
  slug,
@@ -6,6 +6,7 @@ import {
6
6
  resolveFactoryDroidRotation,
7
7
  } from "../compat/factory-droid";
8
8
  import { hasModelScopedEffortLadder, resolveModelPolicy } from "../compat/resolve";
9
+ import { serverSideFallbackModels } from "../compat/server-side-fallback";
9
10
  import type { RequestPolicy } from "../compat/types";
10
11
  import { getBundledModel, isGeneratedProvider } from "../models";
11
12
  import type { FetchImpl, ModelSpec, ThinkingConfig, ThinkingControlMode } from "../types";
@@ -166,16 +167,35 @@ export interface FactoryDroidModelDiscoveryOptions extends AccountScope {
166
167
  */
167
168
  export function factoryDroidSeedModels(scope: AccountScope): ModelSpec<"factory-droid-agent">[] {
168
169
  const inferenceRegion = resolveFactoryDroidInferenceRegion(scope);
169
- return factoryDroidRegistry()
170
- .filter(model => isModelAvailable(model, {}, null, inferenceRegion, null))
171
- .map(model =>
172
- buildFactoryDroidModel(model, {
173
- region: scope.region,
174
- inferenceRegion,
175
- orgId: scope.orgId,
176
- apiProviders: resolveFactoryDroidRotation(model.policy, inferenceRegion),
177
- }),
178
- );
170
+ return withAvailableRefusalFallbacks(
171
+ factoryDroidRegistry()
172
+ .filter(model => isModelAvailable(model, {}, null, inferenceRegion, null))
173
+ .map(model =>
174
+ buildFactoryDroidModel(model, {
175
+ region: scope.region,
176
+ inferenceRegion,
177
+ orgId: scope.orgId,
178
+ apiProviders: resolveFactoryDroidRotation(model.policy, inferenceRegion),
179
+ }),
180
+ ),
181
+ );
182
+ }
183
+
184
+ /**
185
+ * Native offers a server-side refusal fallback only when the account can run
186
+ * that model, so each model keeps the fallbacks its upstreams name that are
187
+ * themselves in the account's roster.
188
+ */
189
+ function withAvailableRefusalFallbacks(models: ModelSpec<"factory-droid-agent">[]): ModelSpec<"factory-droid-agent">[] {
190
+ const available = new Set(models.map(model => model.id));
191
+ for (const model of models) {
192
+ const named = new Set<string>();
193
+ for (const upstream of model.factoryDroidApiProviders ?? []) {
194
+ for (const id of serverSideFallbackModels(model, { upstream })) named.add(id);
195
+ }
196
+ if (named.size > 0) model.factoryDroidRefusalFallbacks = [...named].filter(id => available.has(id));
197
+ }
198
+ return models;
179
199
  }
180
200
 
181
201
  /**
@@ -219,7 +239,7 @@ export async function fetchFactoryDroidModels(
219
239
  } catch {
220
240
  return null;
221
241
  }
222
- return factoryDroidRegistry().flatMap(model => {
242
+ const models = factoryDroidRegistry().flatMap(model => {
223
243
  if (!isModelAvailable(model, flags, policy, servingRegion, images)) return [];
224
244
  const { id } = model.spec;
225
245
  const family = model.policy.routingFamily;
@@ -236,6 +256,7 @@ export async function fetchFactoryDroidModels(
236
256
  if (routed !== undefined) spec.factoryDroidRoutingSource = "configured_order";
237
257
  return [spec];
238
258
  });
259
+ return withAvailableRefusalFallbacks(models);
239
260
  }
240
261
 
241
262
  /**
@@ -282,6 +303,9 @@ export function buildFactoryDroidModel(
282
303
  ...factoryDroidRegionalLimits(policy, resolveFactoryDroidInferenceRegion(options)),
283
304
  ...(options.orgId ? { factoryDroidOrgId: options.orgId } : {}),
284
305
  ...(policy.creditRates ? { factoryDroidCredits: policy.creditRates.input } : {}),
306
+ ...(policy.creditRates?.effectiveRates
307
+ ? { factoryDroidCreditCards: policy.creditRates.effectiveRates.map(card => ({ ...card })) }
308
+ : {}),
285
309
  ...(options.apiProviders?.length ? { factoryDroidApiProviders: [...options.apiProviders] } : {}),
286
310
  };
287
311
  const resolved = resolveModelPolicy(model);
@@ -21,7 +21,7 @@ export const DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS = 10_000;
21
21
  * successful fast discovery requests do not leave armed timeout signals for
22
22
  * concurrent GC to trip over later.
23
23
  */
24
- async function withOpenAICompatibleDiscoveryTimeout<T>(
24
+ export async function withOpenAICompatibleDiscoveryTimeout<T>(
25
25
  timeoutMs: number,
26
26
  run: (signal: AbortSignal) => Promise<T>,
27
27
  ): Promise<T> {
@@ -4,15 +4,17 @@
4
4
  */
5
5
  import type { Database } from "bun:sqlite";
6
6
  import * as path from "node:path";
7
+ import { type } from "@oh-my-pi/omptype";
7
8
  import { getModelDbPath, isSqliteCorruptionError, logger, openSqliteDatabaseSync, VERSION } from "@oh-my-pi/pi-utils";
8
9
  import RULES from "./compat/rules.json" with { type: "json" };
9
10
  import type { Api, Model } from "./types";
10
11
 
11
12
  // Rows persist fully materialized Models so compatible warm-cache reads do not
12
13
  // rerun the compat cascade. Request headers are intentionally omitted:
13
- // arbitrary provider-defined header names can carry credentials. v14
14
- // invalidates Antigravity rows predating per-SKU `antigravityWireProfiles`
15
- // (model_enum label, output cap, SKU thinking). v13 replaces
14
+ // arbitrary provider-defined header names can carry credentials. v15
15
+ // invalidates unbundled OpenAI-compatible rows that ignored provider-reported
16
+ // token limits. v14 invalidates Antigravity rows predating per-SKU
17
+ // `antigravityWireProfiles` (model_enum label, output cap, SKU thinking). v13 replaces
16
18
  // sparse ModelSpec rows with materialized rows and records the exact app/rules/
17
19
  // builder policy that produced them. v12 invalidated Kimi Code rows carrying the blanket
18
20
  // maxTokens: 32000 that predate per-family output caps (k3/k3-256k -> 131072,
@@ -27,7 +29,7 @@ import type { Api, Model } from "./types";
27
29
  // retired unknown-limit sentinels (222222/8888); v5 invalidated rows predating
28
30
  // effort-tier variant collapsing (raw `-low`/`-high`/`-thinking` member ids);
29
31
  // v4 dropped the pre-efforts ThinkingConfig shape.
30
- const CACHE_SCHEMA_VERSION = 14;
32
+ const CACHE_SCHEMA_VERSION = 15;
31
33
  /** Oldest schema whose rows are guaranteed never to carry request headers. */
32
34
  const FIRST_HEADERLESS_SCHEMA_VERSION = 11;
33
35
  const HEADER_RESTORE_VERSION = 1;
@@ -444,61 +446,33 @@ function migrateCacheSchema(db: Database): void {
444
446
  }
445
447
  }
446
448
 
449
+ /**
450
+ * A cacheable materialized model. Also checked on the write side against
451
+ * in-memory models, so optional keys may be present as `undefined`.
452
+ */
453
+ const MaterializedModel = type({
454
+ id: "string > 0",
455
+ name: "string > 0",
456
+ api: "string > 0",
457
+ provider: "string > 0",
458
+ // Empty is legitimate: Azure-style providers resolve the endpoint from
459
+ // configuration at request time, and their bundled rows carry "".
460
+ baseUrl: "string",
461
+ reasoning: "boolean",
462
+ input: "('text' | 'image')[] > 0",
463
+ "headers?": "undefined",
464
+ supportsComputerUseConfig: "boolean | null",
465
+ contextWindow: "number > 0 | null",
466
+ maxTokens: "number > 0 | null",
467
+ identity: { class: "string" },
468
+ "compat?": type("object")
469
+ .narrow(compat => !Array.isArray(compat))
470
+ .or("undefined"),
471
+ cost: { input: "number", output: "number", cacheRead: "number", cacheWrite: "number" },
472
+ });
473
+
447
474
  function isMaterializedModel(value: unknown): value is PersistedModel<Api> {
448
- if (value === null || typeof value !== "object" || Array.isArray(value)) return false;
449
- const model = value as Partial<PersistedModel<Api>>;
450
- if (
451
- typeof model.id !== "string" ||
452
- model.id.length === 0 ||
453
- typeof model.name !== "string" ||
454
- model.name.length === 0 ||
455
- typeof model.api !== "string" ||
456
- model.api.length === 0 ||
457
- typeof model.provider !== "string" ||
458
- model.provider.length === 0 ||
459
- // Empty is legitimate: Azure-style providers resolve the endpoint from
460
- // configuration at request time, and their bundled rows carry "".
461
- typeof model.baseUrl !== "string" ||
462
- typeof model.reasoning !== "boolean" ||
463
- !Array.isArray(model.input) ||
464
- model.input.length === 0 ||
465
- model.input.some(input => input !== "text" && input !== "image") ||
466
- model.headers !== undefined ||
467
- !Object.hasOwn(model, "supportsComputerUseConfig") ||
468
- (model.supportsComputerUseConfig !== null && typeof model.supportsComputerUseConfig !== "boolean")
469
- ) {
470
- return false;
471
- }
472
- const contextWindow = model.contextWindow;
473
- const maxTokens = model.maxTokens;
474
- if (
475
- (contextWindow !== null &&
476
- (typeof contextWindow !== "number" || !Number.isFinite(contextWindow) || contextWindow <= 0)) ||
477
- (maxTokens !== null && (typeof maxTokens !== "number" || !Number.isFinite(maxTokens) || maxTokens <= 0))
478
- ) {
479
- return false;
480
- }
481
- if (
482
- model.identity === null ||
483
- typeof model.identity !== "object" ||
484
- typeof model.identity.class !== "string" ||
485
- (model.compat !== undefined &&
486
- (model.compat === null || typeof model.compat !== "object" || Array.isArray(model.compat)))
487
- ) {
488
- return false;
489
- }
490
- if (model.cost === null || typeof model.cost !== "object") return false;
491
- const cost = model.cost as Partial<Model<Api>["cost"]>;
492
- return (
493
- typeof cost.input === "number" &&
494
- Number.isFinite(cost.input) &&
495
- typeof cost.output === "number" &&
496
- Number.isFinite(cost.output) &&
497
- typeof cost.cacheRead === "number" &&
498
- Number.isFinite(cost.cacheRead) &&
499
- typeof cost.cacheWrite === "number" &&
500
- Number.isFinite(cost.cacheWrite)
501
- );
475
+ return MaterializedModel.allows(value);
502
476
  }
503
477
 
504
478
  function parseMaterializedModels<TApi extends Api>(serialized: string): Model<TApi>[] | null {
@@ -1,4 +1,5 @@
1
1
  import type { Api, Model } from "./types";
2
+ import { isRecord } from "./utils";
2
3
 
3
4
  // References are distinguished from JSON scalar values by their one-element array.
4
5
  // -1 represents an explicitly present undefined value.
@@ -32,16 +33,12 @@ export interface ModelCatalog {
32
33
  readonly compatValues: Map<string, Map<number, Map<string, Record<string, unknown>>>>;
33
34
  }
34
35
 
35
- function isObject(value: unknown): value is Record<string, unknown> {
36
- return value !== null && typeof value === "object" && !Array.isArray(value);
37
- }
38
-
39
36
  // JSON serialization with sorted object keys gives both deterministic ordering
40
37
  // and value-based interning, independent of the input objects' insertion order.
41
38
  function canonical(value: unknown): string {
42
39
  if (value === undefined) return "u";
43
40
  if (Array.isArray(value)) return `[${value.map(canonical).join(",")}]`;
44
- if (isObject(value)) {
41
+ if (isRecord(value)) {
45
42
  return `{${Object.keys(value)
46
43
  .sort()
47
44
  .map(key => `${JSON.stringify(key)}:${canonical(value[key])}`)
@@ -80,7 +77,7 @@ export function encodeModelCatalog(models: Record<string, Record<string, Model<A
80
77
  let node: Node;
81
78
  if (Array.isArray(value)) {
82
79
  node = [1, value.map(encode)];
83
- } else if (isObject(value)) {
80
+ } else if (isRecord(value)) {
84
81
  const properties: EncodedValue[] = [];
85
82
  const boolKeys: number[] = [];
86
83
  const masks: number[] = [];
@@ -119,7 +116,7 @@ export function encodeModelCatalog(models: Record<string, Record<string, Model<A
119
116
  for (const provider of Object.keys(models).sort()) {
120
117
  for (const id of Object.keys(models[provider]).sort()) {
121
118
  const model = models[provider][id];
122
- if (typeof model.api !== "string" || !isObject(model.compat)) continue;
119
+ if (typeof model.api !== "string" || !isRecord(model.compat)) continue;
123
120
  let counts = compatCounts.get(model.api);
124
121
  if (!counts) compatCounts.set(model.api, (counts = new Map()));
125
122
  const signature = canonical(model.compat);
@@ -161,7 +158,7 @@ export function encodeModelCatalog(models: Record<string, Record<string, Model<A
161
158
  const rest: Record<string, unknown> = Object.create(null);
162
159
  const source = model as unknown as Record<string, unknown>;
163
160
  const base = typeof model.api === "string" ? bases.get(model.api) : undefined;
164
- const useBase = base !== undefined && isObject(model.compat);
161
+ const useBase = base !== undefined && isRecord(model.compat);
165
162
  for (const key of Object.keys(source).sort()) {
166
163
  if (Object.hasOwn(constants, key) || (key === "id" && source.id === id) || (key === "compat" && useBase))
167
164
  continue;
@@ -1,19 +1,12 @@
1
+ import { type } from "@oh-my-pi/omptype";
1
2
  import { logger, VERSION } from "@oh-my-pi/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 { isTimeBasedCost } from "./pricing";
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;
@@ -818,18 +811,19 @@ function preferDiscoveryLimit(discoveryLimit: number | null, fallbackLimit: numb
818
811
  * caller treats the source as failed (cache/static fallback) rather than as an
819
812
  * intentionally empty catalog that an authoritative provider would prune to.
820
813
  */
821
- function buildDiscoveredModelSpecs<TApi extends Api>(value: unknown): DiscoveredModelSet<TApi> {
814
+ function buildDiscoveredModelSpecs<TApi extends Api>(values: unknown): DiscoveredModelSet<TApi> {
822
815
  const models: Model<TApi>[] = [];
823
816
  const explicitKindModels = new Set<Model<TApi>>();
824
- if (!Array.isArray(value)) {
817
+ if (!Array.isArray(values)) {
825
818
  return { models, explicitKindModels };
826
819
  }
827
820
  const failures: unknown[] = [];
828
- for (const item of value) {
829
- if (!isModelLike(item)) continue;
821
+ for (const value of values) {
822
+ if (!ModelSpecShape.allows(value)) continue;
823
+ const item = value as ModelSpec<TApi>;
830
824
  let model: Model<TApi>;
831
825
  try {
832
- model = buildModel(item as ModelSpec<TApi>);
826
+ model = buildModel(item);
833
827
  } catch (error) {
834
828
  failures.push(error);
835
829
  logger.warn("model discovery skipped a model that failed to build", {
@@ -846,109 +840,21 @@ function buildDiscoveredModelSpecs<TApi extends Api>(value: unknown): Discovered
846
840
  return { models, explicitKindModels };
847
841
  }
848
842
 
849
- function isModelLike(value: unknown): value is ModelSpec<Api> {
850
- if (!isRecord(value)) {
851
- return false;
852
- }
853
- const v = value as {
854
- id?: unknown;
855
- name?: unknown;
856
- api?: unknown;
857
- provider?: unknown;
858
- baseUrl?: unknown;
859
- reasoning?: unknown;
860
- input?: unknown;
861
- cost?: unknown;
862
- contextWindow?: unknown;
863
- maxTokens?: unknown;
864
- };
865
- if (typeof v.id !== "string" || v.id.length === 0) {
866
- return false;
867
- }
868
- if (typeof v.name !== "string" || v.name.length === 0) {
869
- return false;
870
- }
871
- if (typeof v.api !== "string" || v.api.length === 0) {
872
- return false;
873
- }
874
- if (typeof v.provider !== "string" || v.provider.length === 0) {
875
- return false;
876
- }
877
- if (typeof v.baseUrl !== "string" || v.baseUrl.length === 0) {
878
- return false;
879
- }
880
- if (typeof v.reasoning !== "boolean") {
881
- return false;
882
- }
883
- if (!isModelInputArray(v.input)) {
884
- return false;
885
- }
886
- if (!isModelCost(v.cost)) {
887
- return false;
888
- }
889
- // Finite positive: NaN > 0 is false, +Infinity < Infinity is false.
890
- const cw = v.contextWindow;
891
- if (cw !== null && (typeof cw !== "number" || !(cw > 0 && cw < Infinity))) {
892
- return false;
893
- }
894
- const mt = v.maxTokens;
895
- if (mt !== null && (typeof mt !== "number" || !(mt > 0 && mt < Infinity))) {
896
- return false;
897
- }
898
- return true;
899
- }
900
-
901
- function isModelInputArray(value: unknown): value is ("text" | "image")[] {
902
- if (!Array.isArray(value) || value.length === 0) {
903
- return false;
904
- }
905
- for (let i = 0; i < value.length; i++) {
906
- const item = value[i];
907
- if (item !== "text" && item !== "image") {
908
- return false;
909
- }
910
- }
911
- return true;
912
- }
913
-
914
- function isTokenCost(value: unknown): value is TokenCost {
915
- if (!isRecord(value)) {
916
- return false;
917
- }
918
- const c = value as {
919
- input?: unknown;
920
- output?: unknown;
921
- cacheRead?: unknown;
922
- cacheWrite?: unknown;
923
- };
924
- // Finite (NaN-safe): -Infinity < x < Infinity rejects NaN and both infinities.
925
- // Preserves original behavior: 0 and negatives remain valid.
926
- const ci = c.input;
927
- if (typeof ci !== "number" || !(ci > -Infinity && ci < Infinity)) {
928
- return false;
929
- }
930
- const co = c.output;
931
- if (typeof co !== "number" || !(co > -Infinity && co < Infinity)) {
932
- return false;
933
- }
934
- const cr = c.cacheRead;
935
- if (typeof cr !== "number" || !(cr > -Infinity && cr < Infinity)) {
936
- return false;
937
- }
938
- const cw = c.cacheWrite;
939
- if (typeof cw !== "number" || !(cw > -Infinity && cw < Infinity)) {
940
- return false;
941
- }
942
- return true;
943
- }
944
-
945
- function isModelCost(value: unknown): value is ModelCost {
946
- if (!isTokenCost(value)) return false;
947
- const cost = value as TokenCost & { longContext?: unknown; timeBased?: unknown };
948
- if (cost.timeBased !== undefined && !isTimeBasedCost(cost.timeBased)) return false;
949
- const longContext = cost.longContext;
950
- if (longContext === undefined) return true;
951
- if (!isTokenCost(longContext) || !isRecord(longContext)) return false;
952
- const threshold = longContext.inputThreshold;
953
- return typeof threshold === "number" && threshold > 0 && threshold < Infinity;
954
- }
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
+ });
@@ -1,4 +1,4 @@
1
- import { compareRevision, parseRevision } from "./compat/revision";
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; older generations (cl100k GPT-4, unparsed gpt-4o) keep the estimate.
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
- return identity.class === "openai" && revisionAtLeast(identity.revision, "5") ? "o200k" : undefined;
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
  /**