@opengeni/config 1.4.0 → 1.4.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@opengeni/config",
3
- "version": "1.4.0",
3
+ "version": "1.4.4",
4
4
  "description": "OpenGeni runtime configuration: settings resolution, deployment knobs, and config validation shared across the server packages.",
5
5
  "license": "Apache-2.0",
6
6
  "repository": {
@@ -33,9 +33,9 @@
33
33
  "prepublishOnly": "bash ../../scripts/prepublish-guard"
34
34
  },
35
35
  "dependencies": {
36
- "@opengeni/codex": "1.4.0",
37
- "@opengeni/contracts": "1.4.0",
38
- "@opengeni/xai-subscription": "1.4.0",
36
+ "@opengeni/codex": "1.4.4",
37
+ "@opengeni/contracts": "1.4.4",
38
+ "@opengeni/xai-subscription": "1.4.4",
39
39
  "zod": "^4.2.1"
40
40
  },
41
41
  "engines": {
package/src/index.ts CHANGED
@@ -1,8 +1,4 @@
1
1
  export { managedUserEmailAllowed } from "./managed-user-admission";
2
- import {
3
- parseRealtimeVoicePricingJson,
4
- parseRealtimeVoicePricingTableJson,
5
- } from "./realtime-voice-pricing";
6
2
  export * from "./realtime-voice-pricing";
7
3
  export * from "./web-search";
8
4
  import { EnvCreditPromotionPolicy } from "./credit-promotions";
@@ -71,6 +67,24 @@ import { createHash } from "node:crypto";
71
67
  import { isIP } from "node:net";
72
68
  import { z } from "zod";
73
69
 
70
+ /**
71
+ * Default maximum provider silence on a generic OpenAI-compatible model stream
72
+ * while the worker is waiting for the next chunk. Matches the Codex transport
73
+ * and SuperGrok idle defaults (5 minutes): long enough for a reasoning model's
74
+ * silent thinking gaps, short enough that a wedged stream recovers the same
75
+ * turn instead of hanging until a worker roll.
76
+ */
77
+ export const DEFAULT_MODEL_STREAM_IDLE_TIMEOUT_MS = 300_000;
78
+ /**
79
+ * Default maximum wait across keepalive-only traffic (SSE comments, `ping` /
80
+ * `keepalive` events, repeated `response.in_progress`) for the next
81
+ * model-progress event on a generic model stream. A provider can keep a
82
+ * connection alive while its generation is wedged; this bounds that case.
83
+ * Ten minutes leaves room for a reasoning model's longest silent thinking
84
+ * window (reasoning summaries stream as progress).
85
+ */
86
+ export const DEFAULT_MODEL_STREAM_PROGRESS_TIMEOUT_MS = 600_000;
87
+
74
88
  const envName = /^[A-Za-z_][A-Za-z0-9_]*$/;
75
89
  const registryId = /^[A-Za-z0-9_-]+$/;
76
90
  export const DEFAULT_OPENROUTER_MODEL_ID =
@@ -813,6 +827,29 @@ const SettingsSchema = z.object({
813
827
  .positive()
814
828
  .max(24 * 60 * 60_000)
815
829
  .default(XAI_RESPONSE_STREAM_IDLE_TIMEOUT_MS),
830
+ // Maximum silence between response bytes on a generic OpenAI-compatible
831
+ // model stream (built-in OpenAI/Azure and registry chat/responses providers).
832
+ // Not a request/run duration cap: every received byte resets the timer, and
833
+ // it measures only time spent waiting on the provider for the next chunk. A
834
+ // stalled stream fails as a retryable provider failure so finite same-turn
835
+ // recovery replaces it. A registry provider may override it with
836
+ // `streamIdleTimeoutMs`. Codex, SuperGrok, and Anthropic Messages own their
837
+ // transport-specific idle timers.
838
+ modelStreamIdleTimeoutMs: z.coerce
839
+ .number()
840
+ .int()
841
+ .positive()
842
+ .max(24 * 60 * 60_000)
843
+ .default(DEFAULT_MODEL_STREAM_IDLE_TIMEOUT_MS),
844
+ // Maximum wait across keepalive-only stream traffic for the next
845
+ // model-progress event (registry override: `streamProgressTimeoutMs`).
846
+ // Never shorter than modelStreamIdleTimeoutMs.
847
+ modelStreamProgressTimeoutMs: z.coerce
848
+ .number()
849
+ .int()
850
+ .positive()
851
+ .max(24 * 60 * 60_000)
852
+ .default(DEFAULT_MODEL_STREAM_PROGRESS_TIMEOUT_MS),
816
853
  // Expose the connected apps attached to a Codex subscription through the
817
854
  // synthetic codex_apps MCP server. Independent from subscription routing so
818
855
  // operators can use Codex models without exposing ChatGPT connectors.
@@ -2686,6 +2723,20 @@ const RegistryProviderSchema = z
2686
2723
  label: z.string().min(1).optional(),
2687
2724
  api: ModelProviderApi.default("chat"),
2688
2725
  anthropic: AnthropicProviderOptions.optional(),
2726
+ // Generic OpenAI-compatible stream silence bound; overrides the deployment
2727
+ // `modelStreamIdleTimeoutMs` for this provider only.
2728
+ streamIdleTimeoutMs: z
2729
+ .number()
2730
+ .int()
2731
+ .positive()
2732
+ .max(24 * 60 * 60_000)
2733
+ .optional(),
2734
+ streamProgressTimeoutMs: z
2735
+ .number()
2736
+ .int()
2737
+ .positive()
2738
+ .max(24 * 60 * 60_000)
2739
+ .optional(),
2689
2740
  wireProfile: ModelProviderWireProfile.default("openai"),
2690
2741
  baseUrl: z.string().url(),
2691
2742
  apiKey: z.string().optional(), // inline key (pragmatic) ...
@@ -3178,6 +3229,10 @@ export interface ResolvedModelProvider {
3178
3229
  defaultHeaders?: Record<string, string> | undefined;
3179
3230
  publicDefaultQueryNames?: string[] | undefined;
3180
3231
  publicDefaultHeaderNames?: string[] | undefined;
3232
+ /** Generic OpenAI-compatible stream silence bound override (ms). */
3233
+ streamIdleTimeoutMs?: number | undefined;
3234
+ /** Generic OpenAI-compatible keepalive-only progress bound override (ms). */
3235
+ streamProgressTimeoutMs?: number | undefined;
3181
3236
  credentialSource: CredentialSourceV1;
3182
3237
  billing: BillingAttributionV1;
3183
3238
  }
@@ -4053,6 +4108,8 @@ export function getSettings(source: NodeJS.ProcessEnv = process.env): Settings {
4053
4108
  supergrokResponseStreamIdleTimeoutMs: optional(
4054
4109
  "OPENGENI_SUPERGROK_RESPONSE_STREAM_IDLE_TIMEOUT_MS",
4055
4110
  ),
4111
+ modelStreamIdleTimeoutMs: optional("OPENGENI_MODEL_STREAM_IDLE_TIMEOUT_MS"),
4112
+ modelStreamProgressTimeoutMs: optional("OPENGENI_MODEL_STREAM_PROGRESS_TIMEOUT_MS"),
4056
4113
  codexConnectedAppsEnabled: optional("OPENGENI_CODEX_CONNECTED_APPS_ENABLED"),
4057
4114
  codexToolSearchEnabled: optional("OPENGENI_CODEX_TOOL_SEARCH_ENABLED"),
4058
4115
  lazyToolSearchEnabled: optional("OPENGENI_LAZY_TOOL_SEARCH_ENABLED"),
@@ -6010,6 +6067,12 @@ export function configuredProviders(
6010
6067
  defaultHeaders: provider.defaultHeaders,
6011
6068
  publicDefaultQueryNames: provider.publicDefaultQueryNames,
6012
6069
  publicDefaultHeaderNames: provider.publicDefaultHeaderNames,
6070
+ ...(provider.streamIdleTimeoutMs !== undefined
6071
+ ? { streamIdleTimeoutMs: provider.streamIdleTimeoutMs }
6072
+ : {}),
6073
+ ...(provider.streamProgressTimeoutMs !== undefined
6074
+ ? { streamProgressTimeoutMs: provider.streamProgressTimeoutMs }
6075
+ : {}),
6013
6076
  credentialSource: registryCredentialSource(provider),
6014
6077
  billing: registryBilling(provider),
6015
6078
  }),
@@ -8103,6 +8166,28 @@ export function parseIntegrationsOauthClientsJson(
8103
8166
  return out;
8104
8167
  }
8105
8168
 
8169
+ /**
8170
+ * The operator-registered client for one authorization server, matched by
8171
+ * exact issuer/URL key first and then trailing-slash-insensitively. The OAuth
8172
+ * start flow and the connector catalog's connectability projection share this
8173
+ * so a catalog row is offered exactly when a start would find a client.
8174
+ */
8175
+ export function findIntegrationsOauthClient(
8176
+ configured: Readonly<Record<string, IntegrationOAuthClientConfig>>,
8177
+ candidates: readonly string[],
8178
+ ): IntegrationOAuthClientConfig | null {
8179
+ const normalize = (value: string) => value.replace(/\/+$/, "");
8180
+ for (const key of new Set(candidates.flatMap((candidate) => [candidate, normalize(candidate)]))) {
8181
+ const entry = configured[key];
8182
+ if (entry) return entry;
8183
+ }
8184
+ const normalizedCandidates = new Set(candidates.map(normalize));
8185
+ for (const [key, entry] of Object.entries(configured)) {
8186
+ if (normalizedCandidates.has(normalize(key))) return entry;
8187
+ }
8188
+ return null;
8189
+ }
8190
+
8106
8191
  export const SocialOAuthClientConfigSchema = z.object({
8107
8192
  clientId: z.string().min(1),
8108
8193
  clientSecret: z.string().min(1).optional(),
@@ -8900,11 +8985,6 @@ function validateSettings(settings: Settings, source: NodeJS.ProcessEnv = proces
8900
8985
  // Malformed voice-input pricing deliberately does not fail boot: every
8901
8986
  // process shares this config, and a voice-only typo must not crash-loop
8902
8987
  // workers. The provider is withheld and the API logs voiceInputPricingIssues.
8903
- parseRealtimeVoicePricingJson(settings.azureLivePricingJson, "OPENGENI_AZURE_LIVE_PRICING_JSON");
8904
- parseRealtimeVoicePricingTableJson(
8905
- settings.aiGatewayRealtimePricingJson,
8906
- "OPENGENI_AI_GATEWAY_REALTIME_PRICING_JSON",
8907
- );
8908
8988
  sandboxEnvironmentVariableNames(settings);
8909
8989
  sandboxLifecycleHookIds(settings);
8910
8990
  // Fail fast on a malformed warm-rate table (P2.1).