@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.
Files changed (83) hide show
  1. package/CHANGELOG.md +4 -23
  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/cursor-proto.d.ts +8 -0
  11. package/dist/types/discovery/openai-compatible.d.ts +6 -0
  12. package/dist/types/model-manager.d.ts +11 -1
  13. package/dist/types/model-thinking.d.ts +4 -0
  14. package/dist/types/models.d.ts +16 -0
  15. package/dist/types/pricing.d.ts +66 -6
  16. package/dist/types/provider-models/openai-compat.d.ts +18 -13
  17. package/dist/types/types.d.ts +14 -19
  18. package/dist/types/utils.d.ts +8 -0
  19. package/package.json +4 -4
  20. package/src/compat/auth-ids.ts +5 -0
  21. package/src/compat/axes.ts +21 -9
  22. package/src/compat/catalog-policy.d.ts +9 -0
  23. package/src/compat/catalog-policy.ts +14 -10
  24. package/src/compat/collapse.ts +1 -1
  25. package/src/compat/factory-droid.d.ts +7 -1
  26. package/src/compat/factory-droid.ts +9 -1
  27. package/src/compat/provider-ids.d.ts +1 -1
  28. package/src/compat/provider-ids.ts +2 -0
  29. package/src/compat/resolve.d.ts +1 -1
  30. package/src/compat/resolve.ts +4 -5
  31. package/src/compat/revision.d.ts +2 -0
  32. package/src/compat/revision.ts +10 -0
  33. package/src/compat/rules/README.md +3 -1
  34. package/src/compat/rules/auth/_order.kdl +1 -1
  35. package/src/compat/rules/auth/mistral.kdl +4 -0
  36. package/src/compat/rules/auth/modular.kdl +9 -0
  37. package/src/compat/rules/auth/scaleway.kdl +17 -0
  38. package/src/compat/rules/classes/anthropic.kdl +5 -0
  39. package/src/compat/rules/classes/deepseek.kdl +1 -1
  40. package/src/compat/rules/classes/gemini.kdl +1 -1
  41. package/src/compat/rules/classes/xai.kdl +1 -1
  42. package/src/compat/rules/providers/anthropic.kdl +8 -4
  43. package/src/compat/rules/providers/cursor.kdl +10 -0
  44. package/src/compat/rules/providers/deepseek.kdl +2 -2
  45. package/src/compat/rules/providers/factory-droid.kdl +219 -13
  46. package/src/compat/rules/providers/modular.kdl +35 -0
  47. package/src/compat/rules/providers/openai-codex.kdl +1 -1
  48. package/src/compat/rules/providers/sambanova.kdl +2 -0
  49. package/src/compat/rules/providers/scaleway.kdl +179 -0
  50. package/src/compat/rules/runtime/behavior.kdl +13 -0
  51. package/src/compat/rules.json +1 -1
  52. package/src/compat/server-side-fallback.d.ts +9 -0
  53. package/src/compat/server-side-fallback.ts +7 -5
  54. package/src/compat/types.d.ts +3 -1
  55. package/src/compat/types.ts +3 -1
  56. package/src/discovery/codex.ts +2 -9
  57. package/src/discovery/cursor-proto.d.ts +8 -0
  58. package/src/discovery/cursor-proto.ts +12 -1
  59. package/src/discovery/factory-droid.ts +35 -11
  60. package/src/discovery/openai-compatible.d.ts +6 -0
  61. package/src/discovery/openai-compatible.ts +1 -1
  62. package/src/model-cache.ts +32 -58
  63. package/src/model-catalog-format.ts +5 -8
  64. package/src/model-manager.d.ts +11 -1
  65. package/src/model-manager.ts +85 -143
  66. package/src/model-thinking.d.ts +4 -0
  67. package/src/model-thinking.ts +10 -0
  68. package/src/model-tokenizer.ts +13 -12
  69. package/src/models.d.ts +16 -0
  70. package/src/models.json +1 -1
  71. package/src/models.ts +32 -1
  72. package/src/pricing.d.ts +66 -6
  73. package/src/pricing.ts +59 -94
  74. package/src/provider-models/catalog-snapshot.json.zst +0 -0
  75. package/src/provider-models/catalog-snapshot.json.zst.d.ts +3 -0
  76. package/src/provider-models/descriptors.ts +2 -0
  77. package/src/provider-models/openai-compat.d.ts +18 -13
  78. package/src/provider-models/openai-compat.ts +120 -71
  79. package/src/types.d.ts +14 -19
  80. package/src/types.ts +11 -18
  81. package/src/utils.d.ts +8 -0
  82. package/src/utils.ts +18 -0
  83. package/src/wire/factory-droid.ts +4 -3
@@ -1,4 +1,4 @@
1
- import { getInstallId, USER_AGENT } from "@linxiraos/pi-utils";
1
+ import { USER_AGENT, getInstallId } from "@linxiraos/pi-utils";
2
2
  import * as logger from "@linxiraos/pi-utils/logger";
3
3
  import { toClinePassPublicModelId } from "../cline-pass-model-id";
4
4
  import {
@@ -13,7 +13,7 @@ import {
13
13
  import { resolveAnthropicApiVersion } from "../compat/anthropic";
14
14
  import { xaiResponsesReasoningEffortMap } from "../compat/openai";
15
15
  import { hasModelScopedEffortLadder, resolveModelPolicy } from "../compat/resolve";
16
- import { compareRevision, parseRevision } from "../compat/revision";
16
+ import { compareRevision, parseRevision, revisionAtLeast } from "../compat/revision";
17
17
  import { providerEntries, seedModels } from "../compat/providers";
18
18
  import { billingVariantPlain, classifyModel, discoveryVocabulary } from "../compat/taxonomy";
19
19
  import {
@@ -21,6 +21,7 @@ import {
21
21
  fetchOpenAICompatibleModels,
22
22
  type OpenAICompatibleModelMapperContext,
23
23
  type OpenAICompatibleModelRecord,
24
+ withOpenAICompatibleDiscoveryTimeout,
24
25
  } from "../discovery/openai-compatible";
25
26
  import { Effort, THINKING_EFFORTS } from "../effort";
26
27
  import { FIREWORKS_FAST_SUFFIX, toFireworksPublicModelId } from "../fireworks-model-id";
@@ -43,15 +44,7 @@ import {
43
44
  type TokenCost,
44
45
  runnerApiKind,
45
46
  } from "../types";
46
- import {
47
- discoveryFetch,
48
- isAnthropicOAuthToken,
49
- isRecord,
50
- toBoolean,
51
- toNumber,
52
- toPositiveNumber,
53
- toPositiveNumberOrNull,
54
- } from "../utils";
47
+ import { discoveryFetch, isAnthropicOAuthToken, isRecord, toBoolean, toNumber, toPositiveNumber } from "../utils";
55
48
  import { ALIBABA_TOKEN_PLAN_BASE_URL, parseAlibabaTokenPlanCredential } from "../wire/alibaba-token-plan";
56
49
  import { normalizeCharmHyperBaseUrl } from "../wire/charm-hyper";
57
50
  import { CLINEPASS_API_BASE_URL, clinePassClientHeaders } from "../wire/cline-pass";
@@ -73,20 +66,12 @@ import {
73
66
  } from "../wire/singularityapi";
74
67
  import { isXiaomiTokenPlanKey } from "../wire/xiaomi-token-plan";
75
68
  import { createBundledReferenceMap, createReferenceResolver, toModelSpec } from "./bundled-references";
69
+ import catalogSnapshotPath from "./catalog-snapshot.json.zst" with { type: "file" };
76
70
  import { getDefaultModelDiscoveryBaseUrl, resolveModelCacheProviderId } from "./cache-provider-id";
77
71
  import { getClinePassModelMetadata } from "./cline-pass";
78
72
  import type { ModelManagerConfig } from "./descriptor-types";
79
73
  import { filterModelsDevCatalogRows } from "./models-dev-policies";
80
74
 
81
- function revisionAtLeast(revision: string | undefined, floor: string): boolean {
82
- if (revision === undefined) return false;
83
- const parsedRevision = parseRevision(revision);
84
- const parsedFloor = parseRevision(floor);
85
- return (
86
- parsedRevision !== undefined && parsedFloor !== undefined && compareRevision(parsedRevision, parsedFloor) >= 0
87
- );
88
- }
89
-
90
75
  function isGlmReasoningIdentity(provider: string, modelId: string, floor: string): boolean {
91
76
  const identity = classifyModel(provider, modelId, { lenient: true });
92
77
  if (identity.class !== "glm" || !revisionAtLeast(identity.revision, floor)) return false;
@@ -99,24 +84,6 @@ const MODELS_DEV_URL = "https://catalog.stencil.so/models.json.zstd";
99
84
  /** Little-endian magic number opening every zstd frame (RFC 8878). */
100
85
  const ZSTD_MAGIC = 0xfd2fb528;
101
86
 
102
- /**
103
- * Uses a cancellable timer rather than the native abort-timeout helper so
104
- * successful fast discovery requests do not leave armed timeout signals for
105
- * concurrent GC to trip over later.
106
- */
107
- async function withCatalogDiscoveryTimeout<T>(timeoutMs: number, run: (signal: AbortSignal) => Promise<T>): Promise<T> {
108
- const controller = new AbortController();
109
- const timer = setTimeout(
110
- () => controller.abort(new DOMException("The operation timed out.", "TimeoutError")),
111
- timeoutMs,
112
- );
113
- try {
114
- return await run(controller.signal);
115
- } finally {
116
- clearTimeout(timer);
117
- }
118
- }
119
-
120
87
  const ANTHROPIC_BASE_URL = "https://api.anthropic.com/v1";
121
88
  const ANTHROPIC_OAUTH_BETA =
122
89
  "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,redact-thinking-2026-02-12,context-management-2025-06-27,prompt-caching-scope-2026-01-05,mid-conversation-system-2026-04-07,advanced-tool-use-2025-11-20,effort-2025-11-24,extended-cache-ttl-2025-04-11";
@@ -215,7 +182,37 @@ const CATALOG_USER_AGENT = USER_AGENT;
215
182
  * Transient failures reuse the last in-memory payload for callers that only
216
183
  * need best-effort metadata.
217
184
  */
185
+ async function parseCatalogBytes(bytes: Uint8Array): Promise<unknown> {
186
+ const isZstd = bytes.length >= 4 && new DataView(bytes.buffer, bytes.byteOffset).getUint32(0, true) === ZSTD_MAGIC;
187
+ const text = new TextDecoder().decode(isZstd ? await Bun.zstdDecompress(bytes) : bytes);
188
+ return JSON.parse(text);
189
+ }
190
+
191
+ /**
192
+ * The committed snapshot, decompressed once per process.
193
+ *
194
+ * `gen:models` writes the exact payload shape this function returns, so the
195
+ * offline source and the live source are interchangeable downstream. A
196
+ * snapshot that fails to parse is reported, not cached, so the next caller
197
+ * retries instead of inheriting a poisoned cache.
198
+ */
199
+ let catalogSnapshot: unknown;
200
+ let catalogSnapshotInflight: Promise<unknown> | undefined;
201
+
202
+ function readCatalogSnapshot(): Promise<unknown> {
203
+ if (catalogSnapshot !== undefined) return Promise.resolve(catalogSnapshot);
204
+ catalogSnapshotInflight ??= (async () => {
205
+ const bytes = new Uint8Array(await Bun.file(catalogSnapshotPath).arrayBuffer());
206
+ catalogSnapshot = await parseCatalogBytes(bytes);
207
+ return catalogSnapshot;
208
+ })().finally(() => {
209
+ catalogSnapshotInflight = undefined;
210
+ });
211
+ return catalogSnapshotInflight;
212
+ }
213
+
218
214
  export function fetchWellKnownModels(fetchImpl?: FetchImpl, signal?: AbortSignal): Promise<unknown> {
215
+ if (!fetchImpl) return waitForCatalogRequest(readCatalogSnapshot(), signal);
219
216
  const session = getCatalogSession(fetchImpl);
220
217
  return fetchRevalidatedWellKnownModels(fetchImpl, signal).catch(error => {
221
218
  if (session.hasPayload) return session.payload;
@@ -226,8 +223,9 @@ export function fetchWellKnownModels(fetchImpl?: FetchImpl, signal?: AbortSignal
226
223
  function fetchRevalidatedWellKnownModels(fetchImpl?: FetchImpl, signal?: AbortSignal): Promise<unknown> {
227
224
  const session = getCatalogSession(fetchImpl);
228
225
  if (!session.inflight) {
229
- session.inflight = withCatalogDiscoveryTimeout(DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS, transportSignal =>
230
- fetchCatalogPayload(fetchImpl ?? discoveryFetch(), session, transportSignal),
226
+ session.inflight = withOpenAICompatibleDiscoveryTimeout(
227
+ DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS,
228
+ transportSignal => fetchCatalogPayload(fetchImpl ?? discoveryFetch(), session, transportSignal),
231
229
  ).finally(() => {
232
230
  session.inflight = null;
233
231
  });
@@ -239,7 +237,9 @@ function fetchRevalidatedWellKnownModelsWithTimeout(
239
237
  fetchImpl?: FetchImpl,
240
238
  timeoutMs = DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS,
241
239
  ): Promise<unknown> {
242
- return withCatalogDiscoveryTimeout(timeoutMs, signal => fetchRevalidatedWellKnownModels(fetchImpl, signal));
240
+ // Same source rule as {@link fetchWellKnownModels}: no transport, no network.
241
+ if (!fetchImpl) return readCatalogSnapshot();
242
+ return withOpenAICompatibleDiscoveryTimeout(timeoutMs, signal => fetchRevalidatedWellKnownModels(fetchImpl, signal));
243
243
  }
244
244
 
245
245
  async function fetchCatalogPayload(
@@ -545,6 +545,8 @@ function mapWithBundledReference<TApi extends Api>(
545
545
  return {
546
546
  ...defaults,
547
547
  name,
548
+ contextWindow: toPositiveNumber(entry.context_length, defaults.contextWindow),
549
+ maxTokens: toPositiveNumber(entry.max_completion_tokens, defaults.maxTokens),
548
550
  };
549
551
  }
550
552
  // Generic `/models` rows describe the chat roster. Do not make a bundled
@@ -672,7 +674,7 @@ async function fetchOllamaNativeModels(
672
674
  * Ollama's cloud catalog reports for stock models.
673
675
  */
674
676
  const OLLAMA_FALLBACK_CONTEXT_WINDOW = 128_000;
675
- /** Cap max output tokens at a value that matches ZETA's other openai-responses defaults. */
677
+ /** Cap max output tokens at a value that matches OMP's other openai-responses defaults. */
676
678
  const OLLAMA_DEFAULT_MAX_TOKENS = 8192;
677
679
 
678
680
  interface OllamaResolvedMetadata {
@@ -1594,13 +1596,12 @@ export function novitaModelManagerOptions(
1594
1596
  export const DEEPINFRA_BASE_URL = "https://api.deepinfra.com/v1/openai";
1595
1597
  /**
1596
1598
  * `filter=with_meta` attaches per-model `metadata` (limits, pricing, tags);
1597
- * `sort_by=omp` asks DeepInfra to return models in omp-priority order
1598
- * (upstream-negotiated API key; never rebrand this outbound value)
1599
+ * `sort_by=zeta` asks DeepInfra to return models in zeta-priority order
1599
1600
  * (earlier = better). The mapper does not stamp `priority` yet — see
1600
1601
  * `mapDeepinfraModel` — but the params are sent so discovery picks the
1601
1602
  * ordering up as soon as the server honors it.
1602
1603
  */
1603
- const DEEPINFRA_MODELS_QUERY = "?filter=with_meta&sort_by=omp";
1604
+ const DEEPINFRA_MODELS_QUERY = "?filter=with_meta&sort_by=zeta";
1604
1605
  const DEEPINFRA_EFFORTS = [Effort.Low, Effort.Medium, Effort.High] as const;
1605
1606
 
1606
1607
  /** DeepInfra OpenAI-compatible discovery configuration. */
@@ -1710,7 +1711,7 @@ function mapDeepinfraModel(
1710
1711
  * Bespoke fetch instead of `fetchOpenAICompatibleModels`: the shared helper
1711
1712
  * cannot carry the `filter`/`sort_by` query params and re-sorts results by id,
1712
1713
  * which would destroy DeepInfra's priority ordering once the server honors
1713
- * `sort_by=omp`. Response order is preserved (dedupe keeps the first, i.e.
1714
+ * `sort_by=zeta`. Response order is preserved (dedupe keeps the first, i.e.
1714
1715
  * highest-priority, occurrence).
1715
1716
  */
1716
1717
  async function fetchDeepinfraModels(options: {
@@ -1726,8 +1727,9 @@ async function fetchDeepinfraModels(options: {
1726
1727
  const fetchImpl = discoveryFetch(options.fetch);
1727
1728
  let payload: unknown;
1728
1729
  try {
1729
- const response = await withCatalogDiscoveryTimeout(DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS, signal =>
1730
- fetchImpl(`${options.baseUrl}/models${DEEPINFRA_MODELS_QUERY}`, { method: "GET", headers, signal }),
1730
+ const response = await withOpenAICompatibleDiscoveryTimeout(
1731
+ DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS,
1732
+ signal => fetchImpl(`${options.baseUrl}/models${DEEPINFRA_MODELS_QUERY}`, { method: "GET", headers, signal }),
1731
1733
  );
1732
1734
  if (!response.ok) {
1733
1735
  return null;
@@ -2285,7 +2287,7 @@ async function loadSiliconFlowModelsDevReferences(
2285
2287
  try {
2286
2288
  // Bounded: this enrichment is optional, so a stalled models.dev must not
2287
2289
  // hold back the authoritative endpoint request that runs after it.
2288
- const payload = await withCatalogDiscoveryTimeout(SILICONFLOW_MODELS_DEV_REFERENCE_TIMEOUT_MS, signal =>
2290
+ const payload = await withOpenAICompatibleDiscoveryTimeout(SILICONFLOW_MODELS_DEV_REFERENCE_TIMEOUT_MS, signal =>
2289
2291
  fetchWellKnownModels(fetchImpl, signal),
2290
2292
  );
2291
2293
  return createModelsDevReferenceMap<"openai-completions">(
@@ -2829,7 +2831,7 @@ const CLINE_PASS_LIVE_CATALOG_URL = "https://openrouter.ai/api/v1/models";
2829
2831
  /** Enrichment tier, never authoritative: any failure degrades to the bundled path. */
2830
2832
  async function fetchClinePassLiveCatalog(fetchImpl: FetchImpl): Promise<ClinePassLiveCatalog | undefined> {
2831
2833
  try {
2832
- const response = await withCatalogDiscoveryTimeout(5_000, signal =>
2834
+ const response = await withOpenAICompatibleDiscoveryTimeout(5_000, signal =>
2833
2835
  fetchImpl(CLINE_PASS_LIVE_CATALOG_URL, { signal }),
2834
2836
  );
2835
2837
  if (!response.ok) return undefined;
@@ -3008,7 +3010,7 @@ async function fetchClinePassModels(
3008
3010
  // The roster is authoritative for ids; the live reference catalog is a pure
3009
3011
  // enrichment tier fetched alongside and tolerated away (offline → bundle).
3010
3012
  const [response, liveCatalog] = await Promise.all([
3011
- withCatalogDiscoveryTimeout(5_000, signal =>
3013
+ withOpenAICompatibleDiscoveryTimeout(5_000, signal =>
3012
3014
  fetchImpl(CLINEPASS_MODELS_URL, { signal, headers: clinePassClientHeaders() }),
3013
3015
  ),
3014
3016
  fetchClinePassLiveCatalog(fetchImpl),
@@ -4423,7 +4425,7 @@ export async function fetchLmStudioNativeModelMetadata(
4423
4425
  if (options?.signal !== undefined) {
4424
4426
  return fetchMetadata(options.signal);
4425
4427
  }
4426
- return withCatalogDiscoveryTimeout(LM_STUDIO_NATIVE_METADATA_TIMEOUT_MS, fetchMetadata);
4428
+ return withOpenAICompatibleDiscoveryTimeout(LM_STUDIO_NATIVE_METADATA_TIMEOUT_MS, fetchMetadata);
4427
4429
  }
4428
4430
 
4429
4431
  export interface LmStudioModelManagerConfig {
@@ -4516,7 +4518,7 @@ function toSyntheticStringList(value: unknown): readonly string[] {
4516
4518
 
4517
4519
  /**
4518
4520
  * Translate Synthetic's per-model `reasoning_effort` vocabulary into an effort
4519
- * ladder. Every advertised value that names an ZETA tier maps verbatim; `none`
4521
+ * ladder. Every advertised value that names an OMP tier maps verbatim; `none`
4520
4522
  * is the thinking-off state rather than a tier of its own, so it backs the
4521
4523
  * `minimal` selector through the wire map (same shape as the Fireworks
4522
4524
  * `minimal → none` map) and gives these routes a real no-thinking tier.
@@ -4715,9 +4717,9 @@ export interface BasetenModelManagerConfig {
4715
4717
  fetch?: FetchImpl;
4716
4718
  }
4717
4719
 
4718
- // A previous version of ZETA shipped these models without reasoning levels.
4720
+ // A previous version of OMP shipped these models without reasoning levels.
4719
4721
  // We've since fixed that (V4-generation whitelist). This const lets us bust
4720
- // the cache so that users on that version of ZETA pick up the reasoning levels
4722
+ // the cache so that users on that version of OMP pick up the reasoning levels
4721
4723
  // immediately.
4722
4724
  const BASETEN_CACHE_MIGRATION_MODEL_IDS = [
4723
4725
  "zai-org/GLM-5.3",
@@ -4747,10 +4749,10 @@ export function basetenModelManagerOptions(
4747
4749
  const features = Array.isArray(raw.supported_features) ? raw.supported_features : [];
4748
4750
  const modalities = Array.isArray(raw.input_modalities) ? raw.input_modalities : [];
4749
4751
 
4750
- // Baseten's discovery flags are not enough to enable ZETA reasoning for every
4752
+ // Baseten's discovery flags are not enough to enable OMP reasoning for every
4751
4753
  // model. Only models with a verified Baseten reasoning policy are enabled
4752
4754
  // here; an unknown model may use a different reasoning wire shape or effort
4753
- // vocabulary, which ZETA must not guess.
4755
+ // vocabulary, which OMP must not guess.
4754
4756
  const identity = classifyModel("baseten", defaults.id, { lenient: true });
4755
4757
  const isSupportedBasetenReasoningModel =
4756
4758
  (identity.class === "kimi" && identity.family === "k3") ||
@@ -4771,8 +4773,6 @@ export function basetenModelManagerOptions(
4771
4773
  cacheWrite: 0,
4772
4774
  };
4773
4775
 
4774
- const contextWindow = toPositiveNumber(raw.context_length, reference?.contextWindow ?? defaults.contextWindow);
4775
- const maxTokens = toPositiveNumber(raw.max_completion_tokens, reference?.maxTokens ?? defaults.maxTokens);
4776
4776
  const baseModel = mapWithBundledReference(entry, defaults, reference);
4777
4777
 
4778
4778
  return {
@@ -4780,8 +4780,6 @@ export function basetenModelManagerOptions(
4780
4780
  reasoning,
4781
4781
  input: vision ? ["text", "image"] : ["text"],
4782
4782
  cost,
4783
- contextWindow,
4784
- maxTokens,
4785
4783
  ...(supportsTools === false ? { supportsTools } : {}),
4786
4784
  };
4787
4785
  },
@@ -5446,7 +5444,7 @@ interface StepfunModelRecord extends OpenAICompatibleModelRecord {
5446
5444
 
5447
5445
  /**
5448
5446
  * Translate StepFun's per-model `reasoning_effort_support_list` into a ladder.
5449
- * Every advertised value that names an ZETA tier maps verbatim, in ZETA's tier
5447
+ * Every advertised value that names an OMP tier maps verbatim, in OMP's tier
5450
5448
  * order; a row advertising nothing (or only tiers this client does not know)
5451
5449
  * resolves to no thinking, so the wire path never sends a `reasoning_effort`
5452
5450
  * the endpoint rejects. Same shape as `mapOpenRouterThinking` for OpenRouter's
@@ -6394,7 +6392,9 @@ async function fetchLiteLLMRichModelsInternal<TApi extends Api>(
6394
6392
  if (options.signal !== undefined) {
6395
6393
  return fetchModels(options.signal);
6396
6394
  }
6397
- return options.timeoutMs !== undefined ? withCatalogDiscoveryTimeout(options.timeoutMs, fetchModels) : fetchModels();
6395
+ return options.timeoutMs !== undefined
6396
+ ? withOpenAICompatibleDiscoveryTimeout(options.timeoutMs, fetchModels)
6397
+ : fetchModels();
6398
6398
  }
6399
6399
 
6400
6400
  export async function fetchLiteLLMRichModels<TApi extends Api>(
@@ -6545,13 +6545,7 @@ export function nanoGptModelManagerOptions(
6545
6545
  const maxOutputTokens = entry.max_output_tokens ?? entry.max_completion_tokens;
6546
6546
  const mapped = mapWithBundledReference(
6547
6547
  { ...entry, max_completion_tokens: maxOutputTokens },
6548
- reference
6549
- ? defaults
6550
- : {
6551
- ...defaults,
6552
- contextWindow: toPositiveNumberOrNull(entry.context_length),
6553
- maxTokens: toPositiveNumberOrNull(maxOutputTokens),
6554
- },
6548
+ defaults,
6555
6549
  reference,
6556
6550
  );
6557
6551
  return {
@@ -6784,7 +6778,7 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
6784
6778
  fetchDynamicModels: async () => {
6785
6779
  const fetchImpl = discoveryFetch(config?.fetch);
6786
6780
  const requestBaseUrl = isPersonalGitHubCopilotBaseUrl(baseUrl)
6787
- ? ((await withCatalogDiscoveryTimeout(DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS, signal =>
6781
+ ? ((await withOpenAICompatibleDiscoveryTimeout(DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS, signal =>
6788
6782
  discoverGitHubCopilotApiEndpoint(apiKey, fetchImpl, signal),
6789
6783
  )) ?? baseUrl)
6790
6784
  : baseUrl;
@@ -6926,6 +6920,22 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
6926
6920
  }
6927
6921
  : {}),
6928
6922
  };
6923
+ // Cross-provider fallback references (e.g. a Cursor
6924
+ // collapsed family for an enterprise-only sibling id)
6925
+ // carry provider-specific wire routing that must not
6926
+ // transfer: the off-tier `requestModelId` pin would send
6927
+ // every Copilot request under the `-none` sibling id
6928
+ // regardless of thinking level.
6929
+ if (reference && reference.provider !== "github-copilot") {
6930
+ delete base.requestModelId;
6931
+ if (base.thinking) {
6932
+ // `base` is a shallow copy of the shared global
6933
+ // reference: clone before deleting or the bundled
6934
+ // entry loses its routing process-wide.
6935
+ base.thinking = { ...base.thinking };
6936
+ delete base.thinking.effortRouting;
6937
+ }
6938
+ }
6929
6939
  const defaultCost = copilotTierCost(tokenPrices.defaultTier);
6930
6940
  if (defaultCost) {
6931
6941
  // Cache writes are not reported per tier; retain the bundled provider rate.
@@ -7779,6 +7789,8 @@ const MODELS_DEV_PROVIDER_DESCRIPTORS_SPECIALIZED: readonly ModelsDevProviderDes
7779
7789
  return { api: "openai-completions" as const, baseUrl: ZENMUX_OPENAI_BASE_URL };
7780
7790
  },
7781
7791
  }),
7792
+ // --- Modular ---
7793
+ openAiCompletionsDescriptor("modular", "modular", "https://api.modular.com/v1"),
7782
7794
  ];
7783
7795
  /** All provider descriptors for models.dev data mapping in generate-models.ts. */
7784
7796
  export const MODELS_DEV_PROVIDER_DESCRIPTORS: readonly ModelsDevProviderDescriptor[] = [
@@ -8223,3 +8235,40 @@ export function singularityApiTechModelManagerOptions(
8223
8235
  ): ModelManagerOptions<Api> {
8224
8236
  return singularityApiModelManagerOptions("singularityapi-tech", SINGULARITYAPI_TECH_API_BASE_URL, config);
8225
8237
  }
8238
+
8239
+ // ---------------------------------------------------------------------------
8240
+ // Scaleway
8241
+ // ---------------------------------------------------------------------------
8242
+
8243
+ /**
8244
+ * Scaleway discovery configuration: the API key plus optional base-URL and
8245
+ * fetch overrides. Consumed by {@link scalewayModelManagerOptions}, and
8246
+ * exported for extensions and tests that construct the manager directly.
8247
+ */
8248
+ export interface ScalewayModelManagerConfig {
8249
+ apiKey?: string;
8250
+ baseUrl?: string;
8251
+ fetch?: FetchImpl;
8252
+ }
8253
+
8254
+ /**
8255
+ * Scaleway model manager: plain OpenAI-compatible chat completions at
8256
+ * `api.scaleway.ai/v1` (fr-par serverless Generative APIs). `/v1/models`
8257
+ * returns bare ids, so the authored seed in `providers/scaleway.kdl` is the
8258
+ * reference each live row merges names, modalities, prices, and limits from;
8259
+ * a successful fetch is the complete roster and prunes retired rows.
8260
+ */
8261
+ export function scalewayModelManagerOptions(
8262
+ config?: ScalewayModelManagerConfig,
8263
+ ): ModelManagerOptions<"openai-completions"> {
8264
+ return createOpenAICompatibleModelManagerOptions({
8265
+ api: "openai-completions",
8266
+ providerId: "scaleway",
8267
+ defaultBaseUrl: "https://api.scaleway.ai/v1",
8268
+ config,
8269
+ dynamicModelsAuthoritative: true,
8270
+ requireApiKey: true,
8271
+ filterModel: (_entry, model) => !isExcludedModel("scaleway", model.id),
8272
+ mapModel: mapWithBundledReference,
8273
+ });
8274
+ }
package/src/types.d.ts CHANGED
@@ -1,5 +1,6 @@
1
1
  import type { ModelIdentity } from "./compat/types";
2
2
  import type { Effort } from "./effort";
3
+ import type { DatedCostCard } from "./pricing";
3
4
  export type { FetchImpl } from "@linxiraos/pi-utils";
4
5
  export type { KnownProvider } from "./provider-models/descriptors";
5
6
  export type KnownApi = "openai-completions" | "openai-responses" | "openrouter" | "openai-codex-responses" | "azure-openai-responses" | "anthropic-messages" | "bedrock-converse-stream" | "google-generative-ai" | "google-gemini-cli" | "google-vertex" | "ollama-chat" | "cursor-agent" | "factory-droid-agent" | "gitlab-duo-agent" | "devin-agent" | "apple-foundation-models";
@@ -477,8 +478,6 @@ export interface OpenAICompat {
477
478
  clampOutputToModelMax?: boolean;
478
479
  /** Strip image inputs before encoding (text-only serving of a multimodal id). */
479
480
  stripImageInput?: boolean;
480
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
481
- thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
482
481
  }
483
482
  /**
484
483
  * Compatibility settings for anthropic-messages API.
@@ -676,8 +675,6 @@ export interface AnthropicCompat {
676
675
  * contract override it. Default: 10485760.
677
676
  */
678
677
  maxImagePayloadBytes?: number;
679
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
680
- thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
681
678
  /**
682
679
  * Drop the enabled `thinking` config and replayed thinking blocks when the
683
680
  * conversation is not thinking-led (a proxy contract for non-interleaved
@@ -887,8 +884,6 @@ export interface ResolvedOpenAISharedCompat {
887
884
  toolSchemaFlavor?: OpenAICompat["toolSchemaFlavor"];
888
885
  /** Strip image inputs before encoding (text-only serving of a multimodal id). */
889
886
  stripImageInput: boolean;
890
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
891
- thinkingLoopGuard?: OpenAICompat["thinkingLoopGuard"];
892
887
  /** Flatten/reject leftover root `anyOf`/`oneOf` unions in strict tool schemas (xAI's function-calling validator 400s on them). */
893
888
  rejectRootObjectUnion: boolean;
894
889
  /** Retry without strict tools when the host rejects a strict grammar as too large (OpenRouter-Anthropic compiled-grammar overflow). */
@@ -900,7 +895,7 @@ export interface ResolvedOpenAISharedCompat {
900
895
  * `buildModel`; request handlers read fields and never detect, resolve, or
901
896
  * allocate.
902
897
  */
903
- export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "trustExplicitThinkingOnly" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "statefulResponses" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "mistralReasoningContentParts" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "syntheticReasoningContentFallback" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "stripThinkingHistory" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "sessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "supportsSteering" | "stripImageInput" | "thinkingLoopGuard" | "whenThinking">> & {
898
+ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "trustExplicitThinkingOnly" | "reasoningEffortMap" | "supportsReasoningParams" | "supportsReasoningSummary" | "statefulResponses" | "supportsSamplingParams" | "supportsPenaltyAndStopParams" | "thinkingFormat" | "kimiApiFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "disableReasoningWithTools" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "mistralReasoningContentParts" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "syntheticReasoningContentFallback" | "replayReasoningContent" | "qwenPreserveThinking" | "qwenTemplateReasoningEffort" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "stripThinkingHistory" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "sessionHeader" | "supportsPromptCacheBreakpoints" | "promptCacheBreakpointTtl" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "clampOutputToModelMax" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "streamRevision" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "supportsConfigurationUpdate" | "supportsSteering" | "stripImageInput" | "whenThinking">> & {
904
899
  vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
905
900
  extraBody?: OpenAICompat["extraBody"];
906
901
  cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -992,7 +987,7 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
992
987
  */
993
988
  export type ResolvedOpenRouterCompat = ResolvedOpenAICompat & ResolvedOpenAIResponsesCompat;
994
989
  /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
995
- export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "thinkingLoopGuard" | "bedrockMessagesApi" | "effortBeta" | "extendedCacheTtlBeta" | "disabledThinking" | "betweenToolsEffort" | "stripThinkingHistory" | "fastMode" | "maxImageDimension" | "maxImagePayloadBytes">> & {
990
+ export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdleTimeoutMs" | "bedrockMessagesApi" | "effortBeta" | "extendedCacheTtlBeta" | "disabledThinking" | "betweenToolsEffort" | "stripThinkingHistory" | "fastMode" | "maxImageDimension" | "maxImagePayloadBytes">> & {
996
991
  /** Effort-beta override; undefined keeps the transport's legacy heuristic. */
997
992
  effortBeta?: AnthropicCompat["effortBeta"];
998
993
  /** Extended-cache-ttl beta opt-out; undefined sends it where a 1h breakpoint needs it. */
@@ -1016,8 +1011,6 @@ export type ResolvedAnthropicCompat = Required<Omit<AnthropicCompat, "streamIdle
1016
1011
  * request builder.
1017
1012
  */
1018
1013
  maxImagePayloadBytes?: number;
1019
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
1020
- thinkingLoopGuard?: AnthropicCompat["thinkingLoopGuard"];
1021
1014
  /**
1022
1015
  * Bedrock's Anthropic Messages API (see {@link AnthropicCompat.bedrockMessagesApi}).
1023
1016
  * `true` when detected or configured; otherwise unset, so rows baked before
@@ -1116,16 +1109,13 @@ export interface GoogleCompat {
1116
1109
  * {@link AnthropicCompat.maxImagePayloadBytes}).
1117
1110
  */
1118
1111
  maxImagePayloadBytes?: number;
1119
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
1120
- thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
1121
1112
  /** See {@link BedrockCompat.supportsSamplingParams}. */
1122
1113
  supportsSamplingParams?: boolean;
1123
1114
  }
1124
1115
  /** Fully-resolved google-API compat view, materialized once by `buildModel`. */
1125
- export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "thinkingLoopGuard" | "antigravityUsageLabel" | "supportsSamplingParams" | "maxImageDimension" | "maxImagePayloadBytes">> & {
1116
+ export type ResolvedGoogleCompat = Required<Omit<GoogleCompat, "streamFirstEventTimeoutMs" | "streamIdleTimeoutMs" | "antigravityUsageLabel" | "supportsSamplingParams" | "maxImageDimension" | "maxImagePayloadBytes">> & {
1126
1117
  streamFirstEventTimeoutMs?: number;
1127
1118
  streamIdleTimeoutMs?: number;
1128
- thinkingLoopGuard?: GoogleCompat["thinkingLoopGuard"];
1129
1119
  antigravityUsageLabel?: string;
1130
1120
  supportsSamplingParams?: boolean;
1131
1121
  maxImageDimension?: number;
@@ -1209,12 +1199,12 @@ export interface ModelCost extends TokenCost {
1209
1199
  /**
1210
1200
  * Exact local content tokenizer family for a model.
1211
1201
  *
1212
- * Absent means no first-party local tokenizer is known and consumers retain
1213
- * their estimate/default-tokenizer policy. The values name tokenizer
1202
+ * Absent means no first-party local tokenizer is known and consumers count
1203
+ * in o200k_base. The values name tokenizer
1214
1204
  * generations rather than providers: DeepSeek V3 through V4 share
1215
1205
  * `"deepseek-v3"`; Kimi K2 through K3 share `"kimi-k2"`.
1216
1206
  */
1217
- export type ModelTokenizer = "claude-v3" | "claude-v47" | "claude-v5" | "claude-v5-sonnet" | "qwen3" | "deepseek-v3" | "kimi-k2" | "glm5" | "o200k";
1207
+ export type ModelTokenizer = "claude-v3" | "claude-v47" | "claude-v5" | "claude-v5-sonnet" | "qwen3" | "deepseek-v3" | "kimi-k2" | "glm5" | "cl100k" | "o200k";
1218
1208
  /** One account's discovered entitlements on a model; see {@link Model.accountAccess}. */
1219
1209
  export interface ModelAccountAccess {
1220
1210
  /**
@@ -1404,8 +1394,12 @@ export interface Model<TApi extends Api = Api> {
1404
1394
  factoryDroidApiProviders?: string[];
1405
1395
  /** Factory Droid: `configured_order` when live routing chose the rotation; absent means registry order. */
1406
1396
  factoryDroidRoutingSource?: "configured_order";
1397
+ /** Factory Droid: server-side refusal fallbacks the account can run; absent when none are named. */
1398
+ factoryDroidRefusalFallbacks?: string[];
1407
1399
  /** Factory Droid Standard Credits base rate (relative per-token weight, not dollars). */
1408
1400
  factoryDroidCredits?: number;
1401
+ /** Factory Droid dated `credit-rates` cards, ascending; `getFactoryDroidCreditRate` picks the one due now. */
1402
+ factoryDroidCreditCards?: readonly DatedCostCard[];
1409
1403
  /** Canonical organization resolved during Factory discovery. */
1410
1404
  factoryDroidOrgId?: string;
1411
1405
  cost: ModelCost;
@@ -1478,8 +1472,9 @@ export interface Model<TApi extends Api = Api> {
1478
1472
  useResponsesLite?: boolean;
1479
1473
  /**
1480
1474
  * Service-tier ids the provider advertises for this model (Codex discovery
1481
- * `service_tiers[].id`, e.g. `priority`, `ultrafast`). Absent when the
1482
- * provider publishes no per-model tier list.
1475
+ * `service_tiers[].id`, or an OpenAI-compatible `/v1/models` row's
1476
+ * `service_tiers`, e.g. `priority`, `ultrafast`). Absent when the provider
1477
+ * publishes no per-model tier list.
1483
1478
  */
1484
1479
  serviceTiers?: readonly string[];
1485
1480
  /** Preferred model to switch to when context promotion is triggered (model id or provider/id). */
package/src/types.ts CHANGED
@@ -1,5 +1,6 @@
1
1
  import type { ModelIdentity } from "./compat/types";
2
2
  import type { Effort } from "./effort";
3
+ import type { DatedCostCard } from "./pricing";
3
4
 
4
5
  // Re-exported from @linxiraos/pi-utils so the whole workspace shares one
5
6
  // `fetch`-compatible signature (tls-fetch's wrappers produce/accept it).
@@ -598,8 +599,6 @@ export interface OpenAICompat {
598
599
  clampOutputToModelMax?: boolean;
599
600
  /** Strip image inputs before encoding (text-only serving of a multimodal id). */
600
601
  stripImageInput?: boolean;
601
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
602
- thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
603
602
  }
604
603
 
605
604
  /**
@@ -798,8 +797,6 @@ export interface AnthropicCompat {
798
797
  * contract override it. Default: 10485760.
799
798
  */
800
799
  maxImagePayloadBytes?: number;
801
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
802
- thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
803
800
  /**
804
801
  * Drop the enabled `thinking` config and replayed thinking blocks when the
805
802
  * conversation is not thinking-led (a proxy contract for non-interleaved
@@ -1015,8 +1012,6 @@ export interface ResolvedOpenAISharedCompat {
1015
1012
  toolSchemaFlavor?: OpenAICompat["toolSchemaFlavor"];
1016
1013
  /** Strip image inputs before encoding (text-only serving of a multimodal id). */
1017
1014
  stripImageInput: boolean;
1018
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
1019
- thinkingLoopGuard?: OpenAICompat["thinkingLoopGuard"];
1020
1015
  /** Flatten/reject leftover root `anyOf`/`oneOf` unions in strict tool schemas (xAI's function-calling validator 400s on them). */
1021
1016
  rejectRootObjectUnion: boolean;
1022
1017
  /** Retry without strict tools when the host rejects a strict grammar as too large (OpenRouter-Anthropic compiled-grammar overflow). */
@@ -1099,7 +1094,6 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
1099
1094
  | "supportsConfigurationUpdate"
1100
1095
  | "supportsSteering"
1101
1096
  | "stripImageInput"
1102
- | "thinkingLoopGuard"
1103
1097
  | "whenThinking"
1104
1098
  >
1105
1099
  > & {
@@ -1201,7 +1195,6 @@ export type ResolvedAnthropicCompat = Required<
1201
1195
  Omit<
1202
1196
  AnthropicCompat,
1203
1197
  | "streamIdleTimeoutMs"
1204
- | "thinkingLoopGuard"
1205
1198
  | "bedrockMessagesApi"
1206
1199
  | "effortBeta"
1207
1200
  | "extendedCacheTtlBeta"
@@ -1236,8 +1229,6 @@ export type ResolvedAnthropicCompat = Required<
1236
1229
  * request builder.
1237
1230
  */
1238
1231
  maxImagePayloadBytes?: number;
1239
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
1240
- thinkingLoopGuard?: AnthropicCompat["thinkingLoopGuard"];
1241
1232
  /**
1242
1233
  * Bedrock's Anthropic Messages API (see {@link AnthropicCompat.bedrockMessagesApi}).
1243
1234
  * `true` when detected or configured; otherwise unset, so rows baked before
@@ -1339,8 +1330,6 @@ export interface GoogleCompat {
1339
1330
  * {@link AnthropicCompat.maxImagePayloadBytes}).
1340
1331
  */
1341
1332
  maxImagePayloadBytes?: number;
1342
- /** Thinking-loop watchdog guard family applied to streamed reasoning. */
1343
- thinkingLoopGuard?: "gemini" | "deepseek" | "xai";
1344
1333
  /** See {@link BedrockCompat.supportsSamplingParams}. */
1345
1334
  supportsSamplingParams?: boolean;
1346
1335
  }
@@ -1351,7 +1340,6 @@ export type ResolvedGoogleCompat = Required<
1351
1340
  GoogleCompat,
1352
1341
  | "streamFirstEventTimeoutMs"
1353
1342
  | "streamIdleTimeoutMs"
1354
- | "thinkingLoopGuard"
1355
1343
  | "antigravityUsageLabel"
1356
1344
  | "supportsSamplingParams"
1357
1345
  | "maxImageDimension"
@@ -1360,7 +1348,6 @@ export type ResolvedGoogleCompat = Required<
1360
1348
  > & {
1361
1349
  streamFirstEventTimeoutMs?: number;
1362
1350
  streamIdleTimeoutMs?: number;
1363
- thinkingLoopGuard?: GoogleCompat["thinkingLoopGuard"];
1364
1351
  antigravityUsageLabel?: string;
1365
1352
  supportsSamplingParams?: boolean;
1366
1353
  maxImageDimension?: number;
@@ -1491,8 +1478,8 @@ export interface ModelCost extends TokenCost {
1491
1478
  /**
1492
1479
  * Exact local content tokenizer family for a model.
1493
1480
  *
1494
- * Absent means no first-party local tokenizer is known and consumers retain
1495
- * their estimate/default-tokenizer policy. The values name tokenizer
1481
+ * Absent means no first-party local tokenizer is known and consumers count
1482
+ * in o200k_base. The values name tokenizer
1496
1483
  * generations rather than providers: DeepSeek V3 through V4 share
1497
1484
  * `"deepseek-v3"`; Kimi K2 through K3 share `"kimi-k2"`.
1498
1485
  */
@@ -1505,6 +1492,7 @@ export type ModelTokenizer =
1505
1492
  | "deepseek-v3"
1506
1493
  | "kimi-k2"
1507
1494
  | "glm5"
1495
+ | "cl100k"
1508
1496
  | "o200k";
1509
1497
 
1510
1498
  /** One account's discovered entitlements on a model; see {@link Model.accountAccess}. */
@@ -1700,8 +1688,12 @@ export interface Model<TApi extends Api = Api> {
1700
1688
  factoryDroidApiProviders?: string[];
1701
1689
  /** Factory Droid: `configured_order` when live routing chose the rotation; absent means registry order. */
1702
1690
  factoryDroidRoutingSource?: "configured_order";
1691
+ /** Factory Droid: server-side refusal fallbacks the account can run; absent when none are named. */
1692
+ factoryDroidRefusalFallbacks?: string[];
1703
1693
  /** Factory Droid Standard Credits base rate (relative per-token weight, not dollars). */
1704
1694
  factoryDroidCredits?: number;
1695
+ /** Factory Droid dated `credit-rates` cards, ascending; `getFactoryDroidCreditRate` picks the one due now. */
1696
+ factoryDroidCreditCards?: readonly DatedCostCard[];
1705
1697
  /** Canonical organization resolved during Factory discovery. */
1706
1698
  factoryDroidOrgId?: string;
1707
1699
  cost: ModelCost;
@@ -1774,8 +1766,9 @@ export interface Model<TApi extends Api = Api> {
1774
1766
  useResponsesLite?: boolean;
1775
1767
  /**
1776
1768
  * Service-tier ids the provider advertises for this model (Codex discovery
1777
- * `service_tiers[].id`, e.g. `priority`, `ultrafast`). Absent when the
1778
- * provider publishes no per-model tier list.
1769
+ * `service_tiers[].id`, or an OpenAI-compatible `/v1/models` row's
1770
+ * `service_tiers`, e.g. `priority`, `ultrafast`). Absent when the provider
1771
+ * publishes no per-model tier list.
1779
1772
  */
1780
1773
  serviceTiers?: readonly string[];
1781
1774
  /** Preferred model to switch to when context promotion is triggered (model id or provider/id). */