@linxiraos/pi-catalog 1.4.6 → 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 (104) hide show
  1. package/CHANGELOG.md +6 -2
  2. package/README.md +3 -3
  3. package/dist/types/build.d.ts +12 -4
  4. package/dist/types/compat/auth-ids.d.ts +2 -2
  5. package/dist/types/compat/catalog-policy.d.ts +6 -4
  6. package/dist/types/compat/factory-droid.d.ts +7 -1
  7. package/dist/types/compat/provider-ids.d.ts +1 -1
  8. package/dist/types/compat/resolve.d.ts +1 -1
  9. package/dist/types/compat/revision.d.ts +2 -0
  10. package/dist/types/compat/server-side-fallback.d.ts +6 -4
  11. package/dist/types/compat/types.d.ts +6 -2
  12. package/dist/types/discovery/codex.d.ts +4 -5
  13. package/dist/types/discovery/cursor-proto.d.ts +8 -0
  14. package/dist/types/discovery/openai-compatible.d.ts +6 -0
  15. package/dist/types/model-cache.d.ts +8 -0
  16. package/dist/types/model-catalog-format.d.ts +37 -0
  17. package/dist/types/model-manager.d.ts +11 -1
  18. package/dist/types/model-thinking.d.ts +4 -0
  19. package/dist/types/models.d.ts +22 -3
  20. package/dist/types/pricing.d.ts +66 -6
  21. package/dist/types/provider-models/openai-compat.d.ts +18 -13
  22. package/dist/types/types.d.ts +14 -19
  23. package/dist/types/utils.d.ts +8 -0
  24. package/dist/types/wire/codex.d.ts +1 -1
  25. package/package.json +6 -10
  26. package/src/build.d.ts +12 -4
  27. package/src/build.ts +50 -3
  28. package/src/compat/auth-ids.ts +5 -0
  29. package/src/compat/axes.ts +27 -9
  30. package/src/compat/catalog-policy.d.ts +9 -0
  31. package/src/compat/catalog-policy.ts +14 -10
  32. package/src/compat/collapse.ts +1 -1
  33. package/src/compat/factory-droid.d.ts +7 -1
  34. package/src/compat/factory-droid.ts +9 -1
  35. package/src/compat/provider-ids.d.ts +1 -1
  36. package/src/compat/provider-ids.ts +2 -0
  37. package/src/compat/resolve.d.ts +1 -1
  38. package/src/compat/resolve.ts +4 -5
  39. package/src/compat/revision.d.ts +2 -0
  40. package/src/compat/revision.ts +10 -0
  41. package/src/compat/rules/README.md +14 -12
  42. package/src/compat/rules/auth/_order.kdl +1 -1
  43. package/src/compat/rules/auth/mistral.kdl +4 -0
  44. package/src/compat/rules/auth/modular.kdl +9 -0
  45. package/src/compat/rules/auth/scaleway.kdl +17 -0
  46. package/src/compat/rules/auth/typesafe.kdl +2 -1
  47. package/src/compat/rules/classes/anthropic.kdl +5 -0
  48. package/src/compat/rules/classes/deepseek.kdl +3 -2
  49. package/src/compat/rules/classes/gemini.kdl +1 -1
  50. package/src/compat/rules/classes/openai.kdl +2 -2
  51. package/src/compat/rules/classes/xai.kdl +1 -1
  52. package/src/compat/rules/providers/anthropic.kdl +8 -4
  53. package/src/compat/rules/providers/cursor.kdl +10 -0
  54. package/src/compat/rules/providers/deepseek.kdl +2 -2
  55. package/src/compat/rules/providers/factory-droid.kdl +243 -13
  56. package/src/compat/rules/providers/google-vertex.kdl +3 -0
  57. package/src/compat/rules/providers/kimi-code.kdl +6 -0
  58. package/src/compat/rules/providers/mistral.kdl +8 -0
  59. package/src/compat/rules/providers/modular.kdl +35 -0
  60. package/src/compat/rules/providers/openai-codex.kdl +1 -1
  61. package/src/compat/rules/providers/sambanova.kdl +2 -0
  62. package/src/compat/rules/providers/scaleway.kdl +179 -0
  63. package/src/compat/rules/runtime/behavior.kdl +22 -6
  64. package/src/compat/rules.json +1 -1
  65. package/src/compat/server-side-fallback.d.ts +9 -0
  66. package/src/compat/server-side-fallback.ts +7 -5
  67. package/src/compat/types.d.ts +6 -2
  68. package/src/compat/types.ts +6 -2
  69. package/src/discovery/codex.d.ts +4 -5
  70. package/src/discovery/codex.ts +7 -15
  71. package/src/discovery/cursor-proto.d.ts +8 -0
  72. package/src/discovery/cursor-proto.ts +12 -1
  73. package/src/discovery/factory-droid.ts +35 -11
  74. package/src/discovery/openai-compatible.d.ts +6 -0
  75. package/src/discovery/openai-compatible.ts +1 -1
  76. package/src/model-cache.d.ts +8 -0
  77. package/src/model-cache.ts +83 -61
  78. package/src/model-catalog-format.d.ts +37 -0
  79. package/src/model-catalog-format.ts +296 -0
  80. package/src/model-manager.d.ts +11 -1
  81. package/src/model-manager.ts +85 -143
  82. package/src/model-thinking.d.ts +4 -0
  83. package/src/model-thinking.ts +10 -0
  84. package/src/model-tokenizer.ts +13 -12
  85. package/src/models.d.ts +22 -3
  86. package/src/models.json +1 -1
  87. package/src/models.ts +48 -11
  88. package/src/pricing.d.ts +66 -6
  89. package/src/pricing.ts +59 -94
  90. package/src/provider-models/cache-provider-id.ts +23 -6
  91. package/src/provider-models/catalog-snapshot.json.zst +0 -0
  92. package/src/provider-models/catalog-snapshot.json.zst.d.ts +3 -0
  93. package/src/provider-models/descriptors.ts +2 -0
  94. package/src/provider-models/openai-compat.d.ts +18 -13
  95. package/src/provider-models/openai-compat.ts +203 -94
  96. package/src/provider-models/special.ts +5 -6
  97. package/src/types.d.ts +14 -19
  98. package/src/types.ts +11 -18
  99. package/src/utils.d.ts +8 -0
  100. package/src/utils.ts +18 -0
  101. package/src/wire/codex.d.ts +1 -1
  102. package/src/wire/codex.ts +1 -1
  103. package/src/wire/factory-droid.ts +4 -3
  104. package/src/models.json.d.ts +0 -9
@@ -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),
@@ -4222,9 +4224,11 @@ export function kimiCodeModelManagerOptions(
4222
4224
  config?: KimiCodeModelManagerConfig,
4223
4225
  ): ModelManagerOptions<"openai-completions"> {
4224
4226
  const apiKey = config?.apiKey;
4225
- const baseUrl = config?.baseUrl ?? "https://api.kimi.com/coding/v1";
4227
+ const baseUrl = config?.baseUrl ?? getDefaultModelDiscoveryBaseUrl("kimi-code")!;
4226
4228
  return {
4227
4229
  providerId: "kimi-code",
4230
+ cacheProviderId: resolveModelCacheProviderId("kimi-code", { apiKey, baseUrl }),
4231
+ dynamicModelsAuthoritative: true,
4228
4232
  ...(apiKey && {
4229
4233
  fetchDynamicModels: () =>
4230
4234
  fetchOpenAICompatibleModels({
@@ -4421,7 +4425,7 @@ export async function fetchLmStudioNativeModelMetadata(
4421
4425
  if (options?.signal !== undefined) {
4422
4426
  return fetchMetadata(options.signal);
4423
4427
  }
4424
- return withCatalogDiscoveryTimeout(LM_STUDIO_NATIVE_METADATA_TIMEOUT_MS, fetchMetadata);
4428
+ return withOpenAICompatibleDiscoveryTimeout(LM_STUDIO_NATIVE_METADATA_TIMEOUT_MS, fetchMetadata);
4425
4429
  }
4426
4430
 
4427
4431
  export interface LmStudioModelManagerConfig {
@@ -4514,7 +4518,7 @@ function toSyntheticStringList(value: unknown): readonly string[] {
4514
4518
 
4515
4519
  /**
4516
4520
  * Translate Synthetic's per-model `reasoning_effort` vocabulary into an effort
4517
- * 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`
4518
4522
  * is the thinking-off state rather than a tier of its own, so it backs the
4519
4523
  * `minimal` selector through the wire map (same shape as the Fireworks
4520
4524
  * `minimal → none` map) and gives these routes a real no-thinking tier.
@@ -4713,9 +4717,9 @@ export interface BasetenModelManagerConfig {
4713
4717
  fetch?: FetchImpl;
4714
4718
  }
4715
4719
 
4716
- // A previous version of ZETA shipped these models without reasoning levels.
4720
+ // A previous version of OMP shipped these models without reasoning levels.
4717
4721
  // We've since fixed that (V4-generation whitelist). This const lets us bust
4718
- // 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
4719
4723
  // immediately.
4720
4724
  const BASETEN_CACHE_MIGRATION_MODEL_IDS = [
4721
4725
  "zai-org/GLM-5.3",
@@ -4745,10 +4749,10 @@ export function basetenModelManagerOptions(
4745
4749
  const features = Array.isArray(raw.supported_features) ? raw.supported_features : [];
4746
4750
  const modalities = Array.isArray(raw.input_modalities) ? raw.input_modalities : [];
4747
4751
 
4748
- // 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
4749
4753
  // model. Only models with a verified Baseten reasoning policy are enabled
4750
4754
  // here; an unknown model may use a different reasoning wire shape or effort
4751
- // vocabulary, which ZETA must not guess.
4755
+ // vocabulary, which OMP must not guess.
4752
4756
  const identity = classifyModel("baseten", defaults.id, { lenient: true });
4753
4757
  const isSupportedBasetenReasoningModel =
4754
4758
  (identity.class === "kimi" && identity.family === "k3") ||
@@ -4769,8 +4773,6 @@ export function basetenModelManagerOptions(
4769
4773
  cacheWrite: 0,
4770
4774
  };
4771
4775
 
4772
- const contextWindow = toPositiveNumber(raw.context_length, reference?.contextWindow ?? defaults.contextWindow);
4773
- const maxTokens = toPositiveNumber(raw.max_completion_tokens, reference?.maxTokens ?? defaults.maxTokens);
4774
4776
  const baseModel = mapWithBundledReference(entry, defaults, reference);
4775
4777
 
4776
4778
  return {
@@ -4778,8 +4780,6 @@ export function basetenModelManagerOptions(
4778
4780
  reasoning,
4779
4781
  input: vision ? ["text", "image"] : ["text"],
4780
4782
  cost,
4781
- contextWindow,
4782
- maxTokens,
4783
4783
  ...(supportsTools === false ? { supportsTools } : {}),
4784
4784
  };
4785
4785
  },
@@ -5444,7 +5444,7 @@ interface StepfunModelRecord extends OpenAICompatibleModelRecord {
5444
5444
 
5445
5445
  /**
5446
5446
  * Translate StepFun's per-model `reasoning_effort_support_list` into a ladder.
5447
- * 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
5448
5448
  * order; a row advertising nothing (or only tiers this client does not know)
5449
5449
  * resolves to no thinking, so the wire path never sends a `reasoning_effort`
5450
5450
  * the endpoint rejects. Same shape as `mapOpenRouterThinking` for OpenRouter's
@@ -5679,7 +5679,7 @@ type LiteLLMRichEndpointModel<TApi extends Api> = {
5679
5679
  };
5680
5680
  type LiteLLMRichEndpointFailure = {
5681
5681
  endpoint: string;
5682
- reason: "http-status" | "invalid-json" | "network-error";
5682
+ reason: "http-status" | "invalid-json" | "invalid-pagination" | "network-error";
5683
5683
  status?: number;
5684
5684
  error?: unknown;
5685
5685
  };
@@ -5692,6 +5692,8 @@ type LiteLLMRichEndpointResult<TApi extends Api> =
5692
5692
  | { failure: LiteLLMRichEndpointFailure };
5693
5693
 
5694
5694
  const LITELLM_RICH_ENDPOINTS = ["/model_group/info", "/v2/model/info", "/model/info", "/v1/model/info"] as const;
5695
+ // Bound paginated management responses before trusting their advertised page count.
5696
+ const LITELLM_MAX_RICH_PAGES = 100;
5695
5697
  export const OPENAI_COMPAT_DISCOVERY_DEFAULT_CONTEXT_WINDOW = 128_000;
5696
5698
  export const OPENAI_COMPAT_DISCOVERY_DEFAULT_MAX_TOKENS = 32_768;
5697
5699
  const UNKNOWN_PROXY_COST = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } as const;
@@ -6171,29 +6173,83 @@ async function fetchLiteLLMRichEndpoint<TApi extends Api>(
6171
6173
  if (options.apiKey) {
6172
6174
  requestHeaders.Authorization = `Bearer ${options.apiKey}`;
6173
6175
  }
6174
- let response: Response;
6175
- try {
6176
- response = await fetchImpl(`${managementBaseUrl}${endpoint}`, {
6177
- method: "GET",
6178
- headers: requestHeaders,
6179
- signal,
6180
- });
6181
- } catch (error) {
6182
- return { failure: { endpoint, reason: "network-error", error } };
6183
- }
6184
- if (!response.ok) {
6185
- return response.status === 404 ? null : { failure: { endpoint, reason: "http-status", status: response.status } };
6186
- }
6187
- let payload: unknown;
6188
- try {
6189
- payload = await response.json();
6190
- } catch (error) {
6191
- return { failure: { endpoint, reason: "invalid-json", status: response.status, error } };
6176
+ const fetchPage = async (
6177
+ page?: number,
6178
+ ): Promise<{ payload: unknown; status: number } | { failure: LiteLLMRichEndpointFailure } | null> => {
6179
+ let response: Response;
6180
+ try {
6181
+ response = await fetchImpl(`${managementBaseUrl}${endpoint}${page === undefined ? "" : `?page=${page}`}`, {
6182
+ method: "GET",
6183
+ headers: requestHeaders,
6184
+ signal,
6185
+ });
6186
+ } catch (error) {
6187
+ return { failure: { endpoint, reason: "network-error", error } };
6188
+ }
6189
+ if (!response.ok) {
6190
+ return response.status === 404 && page === undefined
6191
+ ? null
6192
+ : { failure: { endpoint, reason: "http-status", status: response.status } };
6193
+ }
6194
+ try {
6195
+ return { payload: await response.json(), status: response.status };
6196
+ } catch (error) {
6197
+ return { failure: { endpoint, reason: "invalid-json", status: response.status, error } };
6198
+ }
6199
+ };
6200
+ const firstPage = await fetchPage();
6201
+ if (!firstPage || "failure" in firstPage) {
6202
+ return firstPage;
6192
6203
  }
6193
- const entries = extractLiteLLMRichEntries(payload);
6204
+ const entries = extractLiteLLMRichEntries(firstPage.payload);
6194
6205
  if (!entries || entries.length === 0) {
6195
6206
  return null;
6196
6207
  }
6208
+ // /v2/model/info returns {current_page, total_pages, data}; older
6209
+ // endpoints and unpaginated /v2 proxies must keep their existing behavior.
6210
+ if (
6211
+ endpoint === "/v2/model/info" &&
6212
+ isRecord(firstPage.payload) &&
6213
+ ("current_page" in firstPage.payload || "total_pages" in firstPage.payload)
6214
+ ) {
6215
+ const { current_page: currentPage, total_pages: totalPages, total_count: totalCount } = firstPage.payload;
6216
+ const invalidPage = (status: number): LiteLLMRichEndpointResult<TApi> => ({
6217
+ failure: { endpoint, reason: "invalid-pagination", status },
6218
+ });
6219
+ if (
6220
+ currentPage !== 1 ||
6221
+ typeof totalPages !== "number" ||
6222
+ !Number.isSafeInteger(totalPages) ||
6223
+ totalPages < 1 ||
6224
+ totalPages > LITELLM_MAX_RICH_PAGES ||
6225
+ (totalCount !== undefined &&
6226
+ (typeof totalCount !== "number" || !Number.isSafeInteger(totalCount) || totalCount < entries.length))
6227
+ ) {
6228
+ return invalidPage(firstPage.status);
6229
+ }
6230
+ const remainingPages = await Promise.all(
6231
+ Array.from({ length: totalPages - 1 }, (_, index) => fetchPage(index + 2)),
6232
+ );
6233
+ for (const [index, nextPage] of remainingPages.entries()) {
6234
+ if (!nextPage || "failure" in nextPage) {
6235
+ return nextPage ?? invalidPage(firstPage.status);
6236
+ }
6237
+ const nextEntries = extractLiteLLMRichEntries(nextPage.payload);
6238
+ if (
6239
+ !isRecord(nextPage.payload) ||
6240
+ nextPage.payload.current_page !== index + 2 ||
6241
+ nextPage.payload.total_pages !== totalPages ||
6242
+ (totalCount !== undefined && nextPage.payload.total_count !== totalCount) ||
6243
+ !nextEntries?.length
6244
+ ) {
6245
+ return invalidPage(nextPage.status);
6246
+ }
6247
+ entries.push(...nextEntries);
6248
+ }
6249
+ if (totalCount !== undefined && entries.length !== totalCount) {
6250
+ return invalidPage(firstPage.status);
6251
+ }
6252
+ }
6197
6253
  const deduped = new Map<string, LiteLLMRichEndpointModel<TApi>>();
6198
6254
  const excludedModelIds = new Set<string>();
6199
6255
  for (const entry of entries) {
@@ -6336,7 +6392,9 @@ async function fetchLiteLLMRichModelsInternal<TApi extends Api>(
6336
6392
  if (options.signal !== undefined) {
6337
6393
  return fetchModels(options.signal);
6338
6394
  }
6339
- return options.timeoutMs !== undefined ? withCatalogDiscoveryTimeout(options.timeoutMs, fetchModels) : fetchModels();
6395
+ return options.timeoutMs !== undefined
6396
+ ? withOpenAICompatibleDiscoveryTimeout(options.timeoutMs, fetchModels)
6397
+ : fetchModels();
6340
6398
  }
6341
6399
 
6342
6400
  export async function fetchLiteLLMRichModels<TApi extends Api>(
@@ -6350,8 +6408,10 @@ export function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig):
6350
6408
  const baseUrl = config?.baseUrl ?? getDefaultModelDiscoveryBaseUrl("litellm")!;
6351
6409
  return {
6352
6410
  providerId: "litellm",
6353
- // rich-v13 routes `mode: decisions` groups to the Decisions runner as judge
6354
- // rows. rich-v12 invalidates namespaced proxy ids that missed bare catalog
6411
+ // rich-v14 fetches all /v2/model/info pages so cached proxy rosters
6412
+ // include models beyond page one. rich-v13 routes `mode: decisions`
6413
+ // groups to the Decisions runner as judge rows.
6414
+ // rich-v12 invalidates namespaced proxy ids that missed bare catalog
6355
6415
  // references. rich-v11 excluded ClinePass gateway metadata (issue #10932).
6356
6416
  // rich-v10 filtered known non-conversational LiteLLM modes, keyed the
6357
6417
  // deployment's `supports_vision` declaration into cached compat, and
@@ -6485,13 +6545,7 @@ export function nanoGptModelManagerOptions(
6485
6545
  const maxOutputTokens = entry.max_output_tokens ?? entry.max_completion_tokens;
6486
6546
  const mapped = mapWithBundledReference(
6487
6547
  { ...entry, max_completion_tokens: maxOutputTokens },
6488
- reference
6489
- ? defaults
6490
- : {
6491
- ...defaults,
6492
- contextWindow: toPositiveNumberOrNull(entry.context_length),
6493
- maxTokens: toPositiveNumberOrNull(maxOutputTokens),
6494
- },
6548
+ defaults,
6495
6549
  reference,
6496
6550
  );
6497
6551
  return {
@@ -6724,7 +6778,7 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
6724
6778
  fetchDynamicModels: async () => {
6725
6779
  const fetchImpl = discoveryFetch(config?.fetch);
6726
6780
  const requestBaseUrl = isPersonalGitHubCopilotBaseUrl(baseUrl)
6727
- ? ((await withCatalogDiscoveryTimeout(DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS, signal =>
6781
+ ? ((await withOpenAICompatibleDiscoveryTimeout(DEFAULT_OPENAI_COMPATIBLE_DISCOVERY_TIMEOUT_MS, signal =>
6728
6782
  discoverGitHubCopilotApiEndpoint(apiKey, fetchImpl, signal),
6729
6783
  )) ?? baseUrl)
6730
6784
  : baseUrl;
@@ -6866,6 +6920,22 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
6866
6920
  }
6867
6921
  : {}),
6868
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
+ }
6869
6939
  const defaultCost = copilotTierCost(tokenPrices.defaultTier);
6870
6940
  if (defaultCost) {
6871
6941
  // Cache writes are not reported per tier; retain the bundled provider rate.
@@ -7719,6 +7789,8 @@ const MODELS_DEV_PROVIDER_DESCRIPTORS_SPECIALIZED: readonly ModelsDevProviderDes
7719
7789
  return { api: "openai-completions" as const, baseUrl: ZENMUX_OPENAI_BASE_URL };
7720
7790
  },
7721
7791
  }),
7792
+ // --- Modular ---
7793
+ openAiCompletionsDescriptor("modular", "modular", "https://api.modular.com/v1"),
7722
7794
  ];
7723
7795
  /** All provider descriptors for models.dev data mapping in generate-models.ts. */
7724
7796
  export const MODELS_DEV_PROVIDER_DESCRIPTORS: readonly ModelsDevProviderDescriptor[] = [
@@ -8163,3 +8235,40 @@ export function singularityApiTechModelManagerOptions(
8163
8235
  ): ModelManagerOptions<Api> {
8164
8236
  return singularityApiModelManagerOptions("singularityapi-tech", SINGULARITYAPI_TECH_API_BASE_URL, config);
8165
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
+ }
@@ -94,11 +94,10 @@ export function openaiCodexModelManagerOptions(
94
94
  * {@link unionAccountCatalogs}.
95
95
  *
96
96
  * Returns `null` when any account's fetch failed transiently, so a partial list
97
- * cannot replace the previous or bundled authoritative catalog. An account
98
- * whose credential the backend rejected outright (401/403 — revoked or
99
- * deauthorized) contributes nothing and is skipped instead: it would otherwise
100
- * veto every sibling account's models until the user removes it. If no account
101
- * produced a catalog, discovery aborts the same way.
97
+ * cannot replace the previous or bundled authoritative catalog. A definitive
98
+ * account denial (401/403 auth or 402 payment) contributes nothing and is skipped
99
+ * instead: it would otherwise veto every sibling account's models until the
100
+ * account recovers. If no account produced a catalog, discovery aborts.
102
101
  */
103
102
  function unionCodexModels(
104
103
  results: readonly { accountId: string | undefined; result: CodexModelDiscoveryResult | null }[],
@@ -107,7 +106,7 @@ function unionCodexModels(
107
106
  for (const { accountId, result } of results) {
108
107
  if (!result) return null;
109
108
  if (result.rejectedStatus !== undefined) {
110
- logger.warn("Codex model discovery skipped an account whose credential was rejected", {
109
+ logger.warn("Codex model discovery skipped an account denied by the backend", {
111
110
  accountId,
112
111
  status: result.rejectedStatus,
113
112
  });