@bitkyc08/opencodex 2.24.0 → 2.24.2

package/src/server/management/shared.ts

@@ -192,11 +192,11 @@ export interface GrokCandidateModel {
  * from the same two sources as the sync so the two can never disagree.
  */
 export async function fetchGrokCandidateModels(config: OcxConfig): Promise<GrokCandidateModel[]> {
-  const { filterCatalogVisibleModels, nativeOpenAiContextWindow, visibleNativeSlugs } = await import("../../codex/catalog");
+  const { filterCatalogVisibleModels, nativeContextLimits, nativeOpenAiContextWindow, visibleNativeSlugs } = await import("../../codex/catalog");
   const routed = filterCatalogVisibleModels(await fetchAllModels(config), config);
   return [
     ...visibleNativeSlugs(config).map(id => {
-      const contextWindow = nativeOpenAiContextWindow(id, providerContextCap(config, OPENAI_CODEX_PROVIDER_ID));
+      const contextWindow = nativeOpenAiContextWindow(id, nativeContextLimits(config));
       return { id, native: true, ...(contextWindow !== undefined ? { contextWindow } : {}) };
     }),
     ...routed.map(m => ({
@@ -221,7 +221,7 @@ export function stripRegistryOnlyStaticHeaders(name: string, provider: OcxProvid
 
 /** Shared Desktop profile DTO builder for the management API and CLI. */
 export async function buildClaudeDesktopState(config: OcxConfig, stored?: OcxClaudeDesktopProfile) {
-  const { filterCatalogVisibleModels, nativeOpenAiContextWindow, desktopVisibleNativeSlugs } = await import("../../codex/catalog");
+  const { filterCatalogVisibleModels, nativeContextLimits, nativeOpenAiContextWindow, desktopVisibleNativeSlugs } = await import("../../codex/catalog");
   const { DESKTOP_SUPPORTS_1M_THRESHOLD } = await import("../../claude/desktop-3p");
   const { reconcileDesktopProfile, renderDesktopProfile } = await import("../../claude/desktop-profile");
   const routed = filterCatalogVisibleModels(await fetchAllModels(config), config);
@@ -229,7 +229,7 @@ export async function buildClaudeDesktopState(config: OcxConfig, stored?: OcxCla
     // Native rows carry their real context window from the same accessor the Grok sync
     // uses — otherwise Sol's 372k and gpt-5.5's 272k render as blank on Desktop.
     ...desktopVisibleNativeSlugs(config).map(id => {
-      const contextWindow = nativeOpenAiContextWindow(id, providerContextCap(config, OPENAI_CODEX_PROVIDER_ID));
+      const contextWindow = nativeOpenAiContextWindow(id, nativeContextLimits(config));
       return { route: `native/${id}`, label: `${id} (native)`,
         ...(contextWindow !== undefined ? { contextWindow } : {}) };
     }),

package/src/server/management-api.ts

@@ -1,7 +1,7 @@
 import { randomUUID } from "node:crypto";
 import { readFileSync } from "node:fs";
 import type { CatalogModel } from "../codex/catalog";
-import { catalogModelSlug, invalidateCodexModelsCache, nativeModelRows, uniqueCatalogModelsForPublicList } from "../codex/catalog";
+import { catalogModelSlug, invalidateCodexModelsCache, nativeContextLimits, nativeModelRows, uniqueCatalogModelsForPublicList } from "../codex/catalog";
 import {
   DEFAULT_SUBAGENT_MODELS,
   codexAutoStartEnabled,
@@ -205,7 +205,7 @@ export async function handleManagementAPI(
           import("../claude/context-windows"),
           import("../codex/catalog"),
         ]);
-        injectClaudeAgentDefs(config, buildClaudeContextWindows([...visibleNativeSlugs(config)], models, providerContextCap(config, OPENAI_CODEX_PROVIDER_ID)));
+        injectClaudeAgentDefs(config, buildClaudeContextWindows([...visibleNativeSlugs(config)], models, nativeContextLimits(config)));
       } catch {
         // Keep routes available through a provider-discovery blip. A later
         // launch-time sync restores any context markers missing from this pass.

package/src/server/responses/core.ts

@@ -2,6 +2,7 @@ import type { Server } from "bun";
 import { bridgeToResponsesSSE, buildResponseJSON, formatErrorResponse, type ResponsesTerminalStatus } from "../../bridge";
 import { formatPassthroughUpstreamError } from "./passthrough-error";
 import { checkInputAdmission } from "./input-admission";
+import { nativeContextLimits } from "../../codex/catalog";
 import { describeUpstreamConnectFailure } from "./upstream-error";
 import {
   getConfigPath,
@@ -1994,7 +1995,7 @@ async function handleResponsesInner(
   // refusing the turn that shrinks the context would deadlock the client against the very
   // limit this gate reports — it would be told to compact and then denied the compaction.
   if (parsed._compactionRequest !== true) {
-    const inputAdmission = checkInputAdmission(parsed, route.provider, route.providerName, parsed.modelId, providerContextCap(config, OPENAI_CODEX_PROVIDER_ID));
+    const inputAdmission = checkInputAdmission(parsed, route.provider, route.providerName, parsed.modelId, nativeContextLimits(config));
     if (!inputAdmission.admitted) {
       // #1524: this is a LOCAL preflight refusal, not an upstream verdict. A policy or combo
       // fallback must be able to skip this candidate and try one whose context window fits,

package/src/server/responses/input-admission.ts

@@ -10,7 +10,7 @@
  * catches the pathological case and stays out of the way otherwise. Every uncertainty
  * resolves toward admitting.
  */
-import { nativeOpenAiContextWindow, nativeOpenAiMaxInputTokens } from "../../codex/catalog/metadata";
+import { nativeOpenAiContextWindow, nativeOpenAiMaxInputTokens, type NativeContextLimitsInput } from "../../codex/catalog/metadata";
 import { estimateTokens } from "../../lib/token-estimate";
 import { isCanonicalOpenAiForwardProvider, OPENAI_CODEX_PROVIDER_ID } from "../../providers/openai-tiers";
 import type { OcxContentPart, OcxParsedRequest, OcxProviderConfig } from "../../types";
@@ -131,7 +131,7 @@ export function resolveInputCeiling(
   modelId: string,
   // Operator cap for the canonical native provider. Passed in rather than read from a
   // config here so this stays pure: no filesystem, no catalog, no registry scan.
-  nativeContextCap?: number,
+  nativeContextCap?: NativeContextLimitsInput,
 ): number | null {
   const configured = positive(provider.modelContextWindows?.[modelId]) ?? positive(provider.contextWindow);
 
@@ -143,15 +143,18 @@ export function resolveInputCeiling(
   const canonicalNativeBare = providerName === OPENAI_CODEX_PROVIDER_ID
     && isCanonicalOpenAiForwardProvider(provider)
     && !modelId.includes("/");
-  const native = configured === null && canonicalNativeBare
-    ? positive(nativeOpenAiContextWindow(modelId, nativeContextCap))
+  const nativeLimits = canonicalNativeBare && configured !== null
+    ? {
+        ...(typeof nativeContextCap === "number" ? { cap: nativeContextCap } : (nativeContextCap ?? {})),
+        modelWindows: { [modelId]: configured },
+      }
+    : nativeContextCap;
+  const native = canonicalNativeBare
+    ? positive(nativeOpenAiContextWindow(modelId, nativeLimits))
     : null;
-  // The model's own measured input ceiling applies even when the operator configured a
-  // window: a 1,050,000 override must not raise the gate above the 922,000 the upstream
-  // actually accepts. Computed independently of `configured` for exactly that reason.
-  const nativeMaxInput = canonicalNativeBare ? positive(nativeOpenAiMaxInputTokens(modelId, nativeContextCap)) : null;
+  const nativeMaxInput = canonicalNativeBare ? positive(nativeOpenAiMaxInputTokens(modelId, nativeLimits)) : null;
 
-  const window = configured ?? native;
+  const window = canonicalNativeBare ? native : configured;
   // modelMaxInputTokens is an input-only cap, so it can only tighten the window.
   const configuredMaxInput = positive(provider.modelMaxInputTokens?.[modelId]);
   const limits = [window, configuredMaxInput, nativeMaxInput].filter((v): v is number => v !== null);
@@ -168,7 +171,7 @@ export function checkInputAdmission(
   provider: OcxProviderConfig,
   providerName: string,
   modelId: string,
-  nativeContextCap?: number,
+  nativeContextCap?: NativeContextLimitsInput,
 ): InputAdmissionResult {
   const ceiling = resolveInputCeiling(provider, providerName, modelId, nativeContextCap);
   if (ceiling === null) return { admitted: true, estimatedTokens: 0, ceiling: null };

package/src/server/system-env.ts

@@ -237,12 +237,12 @@ function rollbackInjectedKeys(port: number, injectedKeys: string[]): void {
 async function computeEffectiveModelEnv(config: OcxConfig, auto?: AutoContextMode): Promise<{ modelEnv: Record<string, string>; windows: Record<string, number> }> {
   const { boundedContextWindows, buildClaudeContextWindows, effectiveModelEnv } = await import("../claude/context-windows");
   const windows = await boundedContextWindows(async () => {
-    const { gatherRoutedModels, visibleNativeSlugs } = await import("../codex/catalog");
+    const { gatherRoutedModels, nativeContextLimits, visibleNativeSlugs } = await import("../codex/catalog");
     try {
-      return buildClaudeContextWindows([...visibleNativeSlugs(config)], await gatherRoutedModels(config), providerContextCap(config, OPENAI_CODEX_PROVIDER_ID));
+      return buildClaudeContextWindows([...visibleNativeSlugs(config)], await gatherRoutedModels(config), nativeContextLimits(config));
     } catch (error) {
       if (error && typeof error === "object" && (error as { code?: unknown }).code === "catalog_busy") {
-        return buildClaudeContextWindows([...visibleNativeSlugs(config)], [], providerContextCap(config, OPENAI_CODEX_PROVIDER_ID));
+        return buildClaudeContextWindows([...visibleNativeSlugs(config)], [], nativeContextLimits(config));
       }
       throw error;
     }