npm - @bitkyc08/opencodex - Versions diffs - 2.23.0 → 2.24.0 - Mend

@bitkyc08/opencodex 2.23.0 → 2.24.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (18) hide show

package/gui/dist/assets/{index-rFWrIE11.js → index-CFqJKF2L.js} +12 -12
package/gui/dist/index.html +1 -1
package/package.json +1 -1
package/src/claude/context-windows.ts +15 -3
package/src/claude/model-info.ts +1 -1
package/src/codex/catalog/metadata.ts +100 -24
package/src/codex/catalog/parsing.ts +45 -11
package/src/codex/catalog/provider-fetch.ts +4 -2
package/src/codex/catalog/sync.ts +9 -7
package/src/codex/catalog.ts +1 -1
package/src/generated/compatibility-version.json +15 -15
package/src/lib/pinned-http.ts +7 -2
package/src/lib/windows-elevation.ts +3 -3
package/src/server/auth-cors.ts +40 -0
package/src/server/chat-native.ts +32 -6
package/src/server/management/config-routes.ts +79 -2
package/src/service.ts +8 -0
package/src/types.ts +1 -1

package/gui/dist/index.html CHANGED Viewed

@@ -16,7 +16,7 @@
         } catch (e) {}
       })();
     </script>
-    <script type="module" crossorigin src="/assets/index-rFWrIE11.js"></script>
+    <script type="module" crossorigin src="/assets/index-CFqJKF2L.js"></script>
     <link rel="stylesheet" crossorigin href="/assets/index-DQsMZzI5.css">
   </head>
   <body>

package/package.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "@bitkyc08/opencodex",
-  "version": "2.23.0",
+  "version": "2.24.0",
   "description": "Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code",
   "type": "module",
   "main": "./bin/package-main.mjs",

package/src/claude/context-windows.ts CHANGED Viewed

@@ -14,8 +14,20 @@ import { nativeOpenAiContextWindow, type CatalogModel } from "../codex/catalog";
 const ONE_MILLION = 1_000_000;
-/** Auto-context defaults (devlog 260712 020, user-approved). */
-export const AUTO_COMPACT_WINDOW_DEFAULT = 350_000;
+/**
+ * Auto-context defaults (devlog 260712 020, user-approved).
+ *
+ * The compact window is the token count at which Claude Code starts compacting, and it is
+ * also the floor `shouldMarkOneMillion` uses — a model may only carry the marker if it can
+ * host this window. 350,000 was chosen when the widest native row advertised 372,000.
+ *
+ * It now matches the auto-compaction limit the Codex catalog ships for the same models
+ * (`nativeAutoCompactLimit`: 829,800 against the 922,000 native window). Leaving the two
+ * apart meant one model compacting at 350k under Claude Code and at 829,800 under Codex.
+ * The value stays clear of the measured 922,000 ceiling by ~92k, so compaction still has
+ * room to run before the upstream refuses.
+ */
+export const AUTO_COMPACT_WINDOW_DEFAULT = 829_800;
 export const AUTO_CONTEXT_FLOOR = 200_000;
 /** Binary-verified accepted range for CLAUDE_CODE_AUTO_COMPACT_WINDOW (2.1.207: pSo=1e5, yDs=1e6). */
 export const AUTO_COMPACT_WINDOW_MIN = 100_000;
@@ -59,7 +71,7 @@ function inAutoCompactRange(value: number): boolean {
  * predicate so marker and threshold never separate (audit 021 #2); an invalid
  * value disables auto marking entirely (the CLI would ignore it, leaving marked
  * sub-1M models without their safety net). Out-of-range CONFIG values fall back
- * to the 350k default (the management API rejects them; this guards hand-edits).
+ * to AUTO_COMPACT_WINDOW_DEFAULT (the management API rejects them; this guards hand-edits).
  */
 export function resolveAutoContext(claudeCode: AutoContextConfigSlice | undefined, envOverride?: string): AutoContextMode {
   if (claudeCode?.autoContext === false) return AUTO_CONTEXT_OFF;

package/src/claude/model-info.ts CHANGED Viewed

@@ -129,7 +129,7 @@ export function buildAnthropicModelInfos(
     if (seen.has(id)) return;
     seen.add(id);
     // The marker fixes Claude Code's accounting at 1e6, but a model may accept less input
-    // than that: GPT-5.6 advertises a 1,050,000 window while refusing anything past 922,000
+    // than that — a routed GPT-5.6 row runs a 1,050,000 window while refusing past 922,000
     // (measured — see devlog/_plan/260817_native_gpt56_1m_context/001_measurement_evidence.md).
     // Advertising the flat 1e6 there would invite mid-session context_length_exceeded, so the
     // variant reports whichever of the two is smaller.

package/src/codex/catalog/metadata.ts CHANGED Viewed

@@ -100,25 +100,37 @@ export function isUnsupportedOpenAiNativeSlug(slug: string): boolean {
 }
 /**
- * Advertised total context for the Codex-login native GPT-5.6 family.
+ * Advertised context for the Codex-login native GPT-5.6 family.
  *
- * Measured on 2026-08-17 against a real Codex-login account, not taken from the live
- * catalog: `GET /backend-api/codex/models` reports `context_window: 272000` /
- * `max_context_window: 872000` for these slugs, yet a `POST /backend-api/codex/responses`
- * probe admitted 921,508 input tokens and refused 922,013 with
- * `error.code: context_length_exceeded` on sol, terra and luna alike. That boundary is
- * exactly the 922,000 max-input the API-key side of this repo already declares
- * (`src/providers/registry.ts`), and 1,050,000 = 922,000 input + 128,000 output.
- * Evidence: devlog/_plan/260817_native_gpt56_1m_context/001_measurement_evidence.md.
+ * This is an OPERATING CAP, not the hard ceiling — the same shape upstream uses. The live
+ * catalog reports `context_window: 272000` against a `max_context_window: 872000` for these
+ * slugs, and gpt-5.4 runs 272,000 against 1,000,000: the advertised window is always well
+ * inside what the model can take.
+ *
+ * The hard ceiling here was measured on 2026-08-17 against a real Codex-login account:
+ * `POST /backend-api/codex/responses` admitted 921,508 input tokens and refused 922,013 with
+ * `error.code: context_length_exceeded` on sol, terra and luna alike.
+ *
+ * Codex spends `context_window * effective_context_window_percent`, which defaults to 95%
+ * (codex-rs `openai_models.rs` / `turn_context.rs`). So this value yields a 875,900-token
+ * budget and leaves ~46k of headroom under the measured ceiling. An earlier release shipped
+ * 1,050,000 here, which spent 997,500 — past what the upstream accepts.
+ *
+ * Do NOT back-solve this from the ceiling (970,000 would land the budget at 921,500, inside
+ * the 1,840-token gap between the last success and the first refusal). The 95% is a safety
+ * margin to keep, not a discount to cancel out.
+ *
+ * Evidence: devlog/_plan/260817_native_gpt56_1m_context/001_measurement_evidence.md
+ * and 014_final_922k_with_margin.md.
  */
-export const NATIVE_GPT56_CONTEXT_WINDOW = 1_050_000;
+export const NATIVE_GPT56_CONTEXT_WINDOW = 922_000;
 /**
- * Largest input the native GPT-5.6 family actually accepts (measured; see above).
+ * Hard ceiling: the largest input the native GPT-5.6 family actually accepts (measured).
  *
- * This is deliberately NOT `contextWindow * 0.9`: that would be 945,000, which the
- * upstream refuses. Every derived limit — auto-compaction, admission ceilings, the
- * Anthropic `max_input_tokens` surface — has to clamp to this instead.
+ * Equal to the advertised window above rather than below it, because that window is already
+ * capped under this ceiling. The clamp stays because routed and API-key rows carry the same
+ * family at a 1,050,000 window, where 90% (945,000) WOULD overshoot this limit.
  */
 export const NATIVE_GPT56_MAX_INPUT_TOKENS = 922_000;
@@ -158,12 +170,74 @@ const PINNED_NATIVE_CAPABILITY_ENTRIES: Map<string, RawEntry> = new Map(
   }),
 );
-export function nativeOpenAiContextWindow(slug: string, contextCap?: number): number | undefined {
+/**
+ * The user-owned levers that narrow a native window, carried together.
+ *
+ * Both only ever lower: the authoritative window is measured against what the upstream
+ * accepts, so a user value above it would re-create the over-advertising this unit fixed.
+ *
+ * This travels as an ARGUMENT rather than module state on purpose. `grok/sync.ts` runs in
+ * the `ocx ensure` parent process, outside the server, so an injected global would never
+ * reach it — that failure is recorded in
+ * devlog/_plan/260817_native_gpt56_1m_context/006_root_cause_replan.md. Every call site
+ * already holds a config or a cap, so passing one more field costs nothing.
+ *
+ * A bare number is still accepted for the many call sites that only know the cap.
+ */
+export interface NativeContextLimits {
+  /** `providerContextCaps.openai` */
+  readonly cap?: number;
+  /** `providers.openai.contextWindow` — a floor-wide user override. */
+  readonly providerWindow?: number;
+  /** `providers.openai.modelContextWindows` — per-model, wins over `providerWindow`. */
+  readonly modelWindows?: Readonly<Record<string, number>>;
+}
+export type NativeContextLimitsInput = NativeContextLimits | number | undefined;
+function positiveInt(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : undefined;
+}
+function asLimits(input: NativeContextLimitsInput): NativeContextLimits {
+  if (input === undefined) return {};
+  return typeof input === "number" ? { cap: input } : input;
+}
+/** Read both levers out of a config once, for call sites that hold one. */
+export function nativeContextLimits(
+  config: Pick<OcxConfig, "providers" | "providerContextCaps">,
+): NativeContextLimits {
+  const provider = config.providers?.[OPENAI_CODEX_PROVIDER_ID];
+  const modelWindows: Record<string, number> = {};
+  for (const [slug, value] of Object.entries(provider?.modelContextWindows ?? {})) {
+    const window = positiveInt(value);
+    if (window !== undefined) modelWindows[slug] = window;
+  }
+  return {
+    ...(positiveInt(providerContextCap(config, OPENAI_CODEX_PROVIDER_ID)) !== undefined
+      ? { cap: providerContextCap(config, OPENAI_CODEX_PROVIDER_ID) }
+      : {}),
+    ...(positiveInt(provider?.contextWindow) !== undefined ? { providerWindow: provider!.contextWindow } : {}),
+    ...(Object.keys(modelWindows).length > 0 ? { modelWindows } : {}),
+  };
+}
+/** Apply the user levers to an authoritative value. Lowering only, in a fixed order. */
+function narrowToLimits(raw: number | undefined, slug: string, input: NativeContextLimitsInput): number | undefined {
+  if (raw === undefined) return undefined;
+  const limits = asLimits(input);
+  const overlay = positiveInt(limits.modelWindows?.[slug]) ?? positiveInt(limits.providerWindow);
+  const narrowed = overlay === undefined ? raw : Math.min(raw, overlay);
+  return applyProviderContextCap(narrowed, limits.cap) ?? narrowed;
+}
+export function nativeOpenAiContextWindow(slug: string, limits?: NativeContextLimitsInput): number | undefined {
   const raw = NATIVE_OPENAI_CONTEXT_OVERRIDES[slug]?.contextWindow
     ?? (typeof PINNED_NATIVE_CAPABILITY_ENTRIES.get(slug)?.context_window === "number"
       ? PINNED_NATIVE_CAPABILITY_ENTRIES.get(slug)!.context_window as number
       : undefined);
-  return applyProviderContextCap(raw, contextCap) ?? raw;
+  return narrowToLimits(raw, slug, limits);
 }
 /**
@@ -173,12 +247,12 @@ export function nativeOpenAiContextWindow(slug: string, contextCap?: number): nu
  * A provider context cap lowers this too: a capped 272k window must not keep advertising a
  * 922k input ceiling, or the cap would be cosmetic on every input-side surface.
  */
-export function nativeOpenAiMaxInputTokens(slug: string, contextCap?: number): number | undefined {
+export function nativeOpenAiMaxInputTokens(slug: string, limits?: NativeContextLimitsInput): number | undefined {
   const raw = NATIVE_OPENAI_CONTEXT_OVERRIDES[slug]?.maxInputTokens;
   if (raw === undefined) return undefined;
-  const window = nativeOpenAiContextWindow(slug, contextCap);
-  const capped = applyProviderContextCap(raw, contextCap) ?? raw;
-  return window === undefined ? capped : Math.min(capped, window);
+  const window = nativeOpenAiContextWindow(slug, limits);
+  const narrowed = narrowToLimits(raw, slug, limits) ?? raw;
+  return window === undefined ? narrowed : Math.min(narrowed, window);
 }
 export function nativeInputModalities(slug: string): string[] {
@@ -291,13 +365,15 @@ export function desktopVisibleNativeSlugs(
   ]);
 }
-export function nativeModelRows(config: Pick<OcxConfig, "disabledModels" | "combos" | "providerContextCaps">): Array<{ slug: string; disabled: boolean; contextWindow?: number; maxInputTokens?: number }> {
+export function nativeModelRows(config: Pick<OcxConfig, "disabledModels" | "combos" | "providerContextCaps" | "providers">): Array<{ slug: string; disabled: boolean; contextWindow?: number; maxInputTokens?: number }> {
   const disabled = disabledNativeSlugs(config);
   const shadowed = configuredNativeAliasSlugs(config);
-  const openaiContextCap = providerContextCap(config, OPENAI_CODEX_PROVIDER_ID);
+  // Both user levers, not just the cap: a per-model window set from the dashboard has to show
+  // up on the row the dashboard itself renders.
+  const limits = nativeContextLimits(config);
   return NATIVE_OPENAI_MODELS.filter(slug => !shadowed.has(slug)).map(slug => {
-    const contextWindow = nativeOpenAiContextWindow(slug, openaiContextCap);
-    const maxInputTokens = nativeOpenAiMaxInputTokens(slug, openaiContextCap);
+    const contextWindow = nativeOpenAiContextWindow(slug, limits);
+    const maxInputTokens = nativeOpenAiMaxInputTokens(slug, limits);
     return {
       slug,
       disabled: disabled.has(slug),

package/src/codex/catalog/parsing.ts CHANGED Viewed

@@ -31,7 +31,7 @@ import { redactSecretString } from "../../lib/redact";
 import upstreamModelsSnapshot from "../data/upstream-models.json";
-import { NATIVE_OPENAI_CONTEXT_OVERRIDES, SUPPORTED_NATIVE_OPENAI_SLUGS, UPSTREAM_NATIVE_ENTRIES, isNativeOpenAiCapabilityAliasModel, nativeMultiAgentVersion } from "./metadata";
+import { NATIVE_OPENAI_CONTEXT_OVERRIDES, SUPPORTED_NATIVE_OPENAI_SLUGS, UPSTREAM_NATIVE_ENTRIES, isNativeOpenAiCapabilityAliasModel, nativeMultiAgentVersion, nativeOpenAiContextWindow, nativeOpenAiMaxInputTokens, type NativeContextLimitsInput } from "./metadata";
 import { trustedAccountBoundNativeCatalogSlug } from "./account-models";
 import { CODEX_NATIVE_ALIAS_CATALOG_KIND } from "./kinds";
@@ -269,10 +269,11 @@ export function isNativeOpenAiEntry(entry: RawEntry): boolean {
 /**
  * Auto-compaction threshold for a native row.
  *
- * The usual rule is 90% of the window, but a model whose measured input ceiling sits below
- * that (GPT-5.6: 922,000 against a 1,050,000 window, where 90% would be 945,000) has to
+ * The usual rule is 90% of the window, but a row whose input ceiling sits below that has to
  * clamp to the ceiling instead — otherwise the client keeps filling until upstream answers
- * `context_length_exceeded` and compaction never gets a chance to run.
+ * `context_length_exceeded` and compaction never gets a chance to run. Native GPT-5.6 no
+ * longer trips this (922,000 window, 829,800 at 90%), but the routed and API-key rows carry
+ * the same family at a 1,050,000 window where 90% would be 945,000 — past the ceiling.
  */
 function nativeAutoCompactLimit(contextWindow: number, maxInputTokens: number | undefined, contextCap?: number): number {
   const ninety = Math.floor(contextWindow * 0.9);
@@ -281,32 +282,65 @@ function nativeAutoCompactLimit(contextWindow: number, maxInputTokens: number |
   return Math.min(ninety, cappedMaxInput, contextWindow);
 }
-export function applyNativeOpenAiContextOverride(entry: RawEntry, contextCap?: number): void {
+/**
+ * Narrow any already-resolved native window by the user levers.
+ *
+ * Used for the fields the accessors do not own (`max_context_window`, and preserved rows
+ * that carry no static override) so every field on a row lands at the same width.
+ */
+function narrowNativeMaxContextWindow(
+  slug: string,
+  value: number | undefined,
+  limits?: NativeContextLimitsInput,
+): number | undefined {
+  if (typeof value !== "number" || value <= 0) return value;
+  const resolved = nativeOpenAiContextWindow(slug, limits);
+  const authoritative = nativeOpenAiContextWindow(slug);
+  // The accessor pair tells us how far the levers moved this slug; apply the same delta to a
+  // field the accessor does not model, without ever raising it.
+  if (resolved === undefined || authoritative === undefined) return value;
+  return Math.min(value, Math.max(resolved, 1));
+}
+export function applyNativeOpenAiContextOverride(entry: RawEntry, limits?: NativeContextLimitsInput): void {
   const nativeSlug = trustedAccountBoundNativeCatalogSlug(entry)
     ?? (isNativeOpenAiEntry(entry) ? entry.slug as string : undefined);
   if (!nativeSlug) return;
   const override = NATIVE_OPENAI_CONTEXT_OVERRIDES[nativeSlug];
   if (override) {
+    // Read the effective values through the accessors rather than re-deriving them from the
+    // static table: this function used to apply only the provider cap, so a per-model window
+    // the dashboard had already accepted was silently written back at full width here.
     if (typeof override.contextWindow === "number") {
-      const contextWindow = applyProviderContextCap(override.contextWindow, contextCap) ?? override.contextWindow;
+      const contextWindow = nativeOpenAiContextWindow(nativeSlug, limits) ?? override.contextWindow;
       entry.context_window = contextWindow;
-      entry.auto_compact_token_limit = nativeAutoCompactLimit(contextWindow, override.maxInputTokens, contextCap);
+      entry.auto_compact_token_limit = nativeAutoCompactLimit(
+        contextWindow,
+        nativeOpenAiMaxInputTokens(nativeSlug, limits) ?? override.maxInputTokens,
+        undefined,
+      );
     }
     if (typeof override.maxContextWindow === "number") {
-      entry.max_context_window = applyProviderContextCap(override.maxContextWindow, contextCap) ?? override.maxContextWindow;
+      const maxContextWindow = narrowNativeMaxContextWindow(nativeSlug, override.maxContextWindow, limits);
+      entry.max_context_window = maxContextWindow;
     }
   }
   // providerContextCaps.openai is a ceiling for native OpenAI rows regardless of where the
   // advertised window came from (#1430): preserved rows without a hardcoded override (e.g.
   // gpt-5.4-mini) must stay under the cap too, and auto-compaction follows the capped window.
+  // The per-model window narrows the same rows for the same reason.
   const currentContext = typeof entry.context_window === "number" ? entry.context_window : undefined;
-  const cappedContext = applyProviderContextCap(currentContext, contextCap);
+  const cappedContext = narrowNativeMaxContextWindow(nativeSlug, currentContext, limits);
   if (cappedContext !== currentContext && typeof cappedContext === "number") {
     entry.context_window = cappedContext;
-    entry.auto_compact_token_limit = nativeAutoCompactLimit(cappedContext, override?.maxInputTokens, contextCap);
+    entry.auto_compact_token_limit = nativeAutoCompactLimit(
+      cappedContext,
+      nativeOpenAiMaxInputTokens(nativeSlug, limits) ?? override?.maxInputTokens,
+      undefined,
+    );
   }
   const currentMax = typeof entry.max_context_window === "number" ? entry.max_context_window : undefined;
-  const cappedMax = applyProviderContextCap(currentMax, contextCap);
+  const cappedMax = narrowNativeMaxContextWindow(nativeSlug, currentMax, limits);
   if (cappedMax !== currentMax) {
     entry.max_context_window = cappedMax;
   }

package/src/codex/catalog/provider-fetch.ts CHANGED Viewed

@@ -1725,8 +1725,10 @@ async function gatherRoutedModelsUncached(
         id: slug,
         owned_by: "openai",
         contextWindow,
-        // Input limit, not the total window: GPT-5.6 advertises 1,050,000 but refuses past
-        // 922,000 (measured). Falls back to the window for slugs with no separate ceiling.
+        // Input limit, not the total window. These coincide for native GPT-5.6 today (the
+        // advertised 922,000 window is already capped at its measured ceiling), but the two
+        // stay separate fields because routed/API rows of the same family run a wider window.
+        // Falls back to the window for slugs with no separate ceiling.
         maxInputTokens: Math.min(nativeOpenAiMaxInputTokens(slug, openaiContextCap) ?? contextWindow, contextWindow),
         inputModalities: nativeInputModalities(slug),
         reasoningEfforts: nativeReasoningEfforts(slug),

package/src/codex/catalog/sync.ts CHANGED Viewed

@@ -35,7 +35,7 @@ import { codexAccountNamespaceEntries, isMainCodexAccountTarget } from "../accou
 import { CODEX_CUSTOM_MODEL_CATALOG_KIND, CODEX_PROVIDER_MODEL_CATALOG_KIND, activeCodexModelsCachePath, applyCatalogMetadata, applyMultiAgentMode, applyNativeOpenAiContextOverride, applyRoutedCodexToolMode, catalogBackupPathFor, catalogHasRoutedEntries, catalogModelSlug, ensureStrictCatalogFields, findNativeTemplate, isDefaultCatalogPath, isRoutedModelCompatibilityExcluded, legacyCatalogBackupPath, normalizeRoutedCatalogEntry, normalizeServiceTiers, readCatalog, readCatalogBackup, readCodexCatalogPath, readNativeBaseline } from "./parsing";
 import type { CatalogModel, MultiAgentMode, RawCatalog, RawEntry } from "./parsing";
-import { accountBoundNativeOpenAiSlugs, accountBoundNativeOpenAiSlugsBySelector, applyNativeVisibility, CODEX_NATIVE_ALIAS_CATALOG_KIND, desktopAllowlistSuppressedNativeSlugs, disabledNativeSlugs, isNativeAliasCatalogEntry, isUnsupportedOpenAiNativeSlug, NATIVE_OPENAI_MODELS, observedAccountBoundNativeEntries, shouldIncludeAccountBoundNativeOpenAi, shouldIncludeNativeOpenAi, shouldUpgradeToUpstreamEntry, SUPPORTED_NATIVE_OPENAI_SLUGS, upstreamNativeEntry } from "./metadata";
+import { accountBoundNativeOpenAiSlugs, accountBoundNativeOpenAiSlugsBySelector, applyNativeVisibility, CODEX_NATIVE_ALIAS_CATALOG_KIND, desktopAllowlistSuppressedNativeSlugs, disabledNativeSlugs, isNativeAliasCatalogEntry, isUnsupportedOpenAiNativeSlug, NATIVE_OPENAI_MODELS, nativeContextLimits, observedAccountBoundNativeEntries, shouldIncludeAccountBoundNativeOpenAi, shouldIncludeNativeOpenAi, shouldUpgradeToUpstreamEntry, SUPPORTED_NATIVE_OPENAI_SLUGS, upstreamNativeEntry, type NativeContextLimitsInput } from "./metadata";
 import {
   bundledCatalogCacheState,
   loadBundledCodexCatalog,
@@ -226,7 +226,7 @@ export function effectiveSubagentRoster(
   return { candidates, advertised, excluded };
 }
-export function finishUpstreamNativeEntry(clone: RawEntry, priority: number, contextCap?: number): RawEntry {
+export function finishUpstreamNativeEntry(clone: RawEntry, priority: number, contextCap?: NativeContextLimitsInput): RawEntry {
   if (priority !== 9) clone.priority = priority;
   applyNativeOpenAiContextOverride(clone, contextCap);
   // GPT-5.6 natives keep their exact upstream ladders (e.g. luna has max but no ultra).
@@ -279,7 +279,7 @@ export function deriveEntry(
   priority: number,
   model?: CatalogModel,
   exactComboSlugs: ReadonlySet<string> = new Set(),
-  contextCap?: number,
+  contextCap?: NativeContextLimitsInput,
 ): RawEntry {
   const preserveExact = isExactComboCatalogModel(model, exactComboSlugs);
   const codexForwardNativeCapabilityAlias = model?.codexForwardNativeCapabilityAlias === true
@@ -405,7 +405,7 @@ export interface ObservedCatalogEntryBuildInput {
   readonly disabledNativeAccountSlugs: ReadonlySet<string>;
   readonly multiAgentV2Enabled: boolean;
   readonly keepNativeChatGptOnV1?: boolean;
-  readonly openaiContextCap?: number;
+  readonly openaiContextCap?: NativeContextLimitsInput;
   /** Additional native ids to clone under account selectors, without creating bare rows. */
   readonly accountNativeSlugs?: readonly string[];
   /** Per-selector account ids; unknown observations must not be copied to unrelated accounts. */
@@ -424,7 +424,7 @@ export function buildCatalogEntries(
   accountSelectors: readonly string[] = [],
   suppressedBareNativeSlugs: ReadonlySet<string> = new Set(),
   disabledNativeAccountSlugs: ReadonlySet<string> = new Set(),
-  contextCap?: number,
+  contextCap?: NativeContextLimitsInput,
   accountNativeSlugs?: readonly string[],
   accountNativeSlugsBySelector?: ReadonlyMap<string, readonly string[]>,
   keepNativeChatGptOnV1 = false,
@@ -760,7 +760,7 @@ export interface ObservedCatalogMergeInput {
   readonly accountBoundEntries: readonly RawEntry[];
   readonly suppressedBareNativeSlugs?: ReadonlySet<string>;
   readonly policy: ObservedCatalogMergePolicy;
-  readonly openaiContextCap?: number;
+  readonly openaiContextCap?: NativeContextLimitsInput;
 }
 /**
@@ -1424,7 +1424,9 @@ function writeRetainedCatalogSync({
   const hasPhysicalComboProvider = Object.hasOwn(config.providers, COMBO_NAMESPACE);
   const includeNativeOpenAi = shouldIncludeNativeOpenAi(config);
   const includeAccountBoundNativeOpenAi = shouldIncludeAccountBoundNativeOpenAi(config);
-  const openaiContextCap = providerContextCap(config, OPENAI_CODEX_PROVIDER_ID);
+  // Both user levers. Passing only the cap here is what let a per-model window the dashboard
+  // had accepted get written back at full width in the on-disk catalog.
+  const openaiContextCap = nativeContextLimits(config);
   const accountSelectors = includeAccountBoundNativeOpenAi
     ? visibleCodexAccountSelectors(config)
     : [];

package/src/codex/catalog.ts CHANGED Viewed

@@ -2,7 +2,7 @@
 // Public surface preserved exactly; importers keep using "src/codex/catalog".
 export { isMediaGenerationModelId, shouldExposeRoutedModel, readCodexCatalogPath, readCatalog, normalizeRoutedCatalogEntry, catalogModelSlug, filterSupportedNativeSlugs, catalogModelSupportsReasoningSummaries } from "./catalog/parsing";
 export type { CatalogModel, MultiAgentMode } from "./catalog/parsing";
-export { accountBoundNativeOpenAiSlugs, accountBoundNativeOpenAiSlugsBySelector, CODEX_NATIVE_ALIAS_CATALOG_KIND, NATIVE_DAYBREAK_BLUE_MODEL, NATIVE_GPT56_MAX_INPUT_TOKENS, NATIVE_OPENAI_CAPABILITY_ALIAS_MODELS, NATIVE_OPENAI_MODELS, configuredNativeAliasSlugs, desktopAllowlistSuppressedNativeSlugs, isNativeAliasCatalogEntry, isNativeOpenAiCapabilityAliasModel, nativeOpenAiCapabilitySourceSlug, nativeOpenAiContextWindow, nativeOpenAiMaxInputTokens, disabledNativeSlugs, visibleNativeSlugs, desktopVisibleNativeSlugs, nativeModelRows, applyNativeVisibility, observedAccountBoundNativeEntries, observedAccountBoundNativeOpenAiSlugs, upstreamNativeEntry, nativeOpenAiSlugs, listCatalogNativeSlugs, nativeInputModalities, nativeReasoningEfforts, nativeDefaultReasoningEffort, shouldIncludeAccountBoundNativeOpenAi, shouldIncludeNativeOpenAi } from "./catalog/metadata";
+export { accountBoundNativeOpenAiSlugs, accountBoundNativeOpenAiSlugsBySelector, CODEX_NATIVE_ALIAS_CATALOG_KIND, NATIVE_DAYBREAK_BLUE_MODEL, NATIVE_GPT56_MAX_INPUT_TOKENS, NATIVE_OPENAI_CAPABILITY_ALIAS_MODELS, NATIVE_OPENAI_MODELS, configuredNativeAliasSlugs, desktopAllowlistSuppressedNativeSlugs, isNativeAliasCatalogEntry, isNativeOpenAiCapabilityAliasModel, nativeContextLimits, nativeOpenAiCapabilitySourceSlug, nativeOpenAiContextWindow, nativeOpenAiMaxInputTokens, disabledNativeSlugs, visibleNativeSlugs, desktopVisibleNativeSlugs, nativeModelRows, applyNativeVisibility, observedAccountBoundNativeEntries, observedAccountBoundNativeOpenAiSlugs, upstreamNativeEntry, nativeOpenAiSlugs, listCatalogNativeSlugs, nativeInputModalities, nativeReasoningEfforts, nativeDefaultReasoningEffort, shouldIncludeAccountBoundNativeOpenAi, shouldIncludeNativeOpenAi, type NativeContextLimits, type NativeContextLimitsInput } from "./catalog/metadata";
 export { isSpawnableCodexCandidate, codexExecInvocation, loadBundledCodexCatalog, materializeBundledCodexCatalog, loadCatalogTemplate } from "./catalog/bundled";
 export { nativeEffortClamp, shouldApplyNativeEffortClamp, catalogModelEfforts, codexSupportedReasoningEfforts, clampedDefaultEffort, clampEntryToCodexSupportedEfforts, clampCatalogModelsToCodexSupport } from "./catalog/effort";
 export { applyProviderConfigHints, isDatedVariantId, filterCatalogVisibleModels, gatherRoutedModels, clearGatherRoutedModelsInflight, augmentRoutedModelsWithRegistryOpenAiApiRows, augmentRoutedModelsWithMetadata, resolveComboCatalogMember, configuredComboTargetModelsByProvider } from "./catalog/provider-fetch";

package/src/generated/compatibility-version.json CHANGED Viewed

@@ -10,7 +10,7 @@
     },
     {
       "path": "package.json",
-      "sha256": "57bd46dc1a69c71fc5d775e860bfa1dee2852d2f8151ff3a22b3c269fe47cf70"
+      "sha256": "82a2b976e7b6e80afbf2b72f446758665ef0c78c83adf116f1cad43f2cfbc34b"
     },
     {
       "path": "scripts/model-metadata.source.json",
@@ -342,7 +342,7 @@
     },
     {
       "path": "src/claude/context-windows.ts",
-      "sha256": "586219c2f9c4e69085e7484d5c081d48b1139a05cb1569ce820ff3b6c41b4843"
+      "sha256": "41c66bc19175eace1d367f349a3311a691c3445d5b9f3f16fab7cd351939cb9c"
     },
     {
       "path": "src/claude/desktop-3p-guard.ts",
@@ -378,7 +378,7 @@
     },
     {
       "path": "src/claude/model-info.ts",
-      "sha256": "dae5d67dc82592f2d994bb3681dfa561957379e6b6124b532d797714dc9eedc9"
+      "sha256": "f83181da3d11d47ebdbffcbbfbdfbc860ed2c296f30cdc36d1c7a3bec7061134"
     },
     {
       "path": "src/claude/outbound.ts",
@@ -662,7 +662,7 @@
     },
     {
       "path": "src/codex/catalog.ts",
-      "sha256": "8597b4cac76055772d9ef207b5661d0e77713f406e5b4a849725c4f26ef6a9a1"
+      "sha256": "0c76625e56687be734c75c6be80586e298d8cbb3c471707417708ee522afb5b8"
     },
     {
       "path": "src/codex/catalog/account-models.ts",
@@ -690,7 +690,7 @@
     },
     {
       "path": "src/codex/catalog/metadata.ts",
-      "sha256": "0c519602bb24c836ad960bf26835bc23a3e17e62b5d4c3d9e056a29f666b1e12"
+      "sha256": "66d9f4c603a77d34ab44b11b9b4397cc53c6a1e88a4bfb47cbc2a6b25ef20fbc"
     },
     {
       "path": "src/codex/catalog/native-models.ts",
@@ -698,15 +698,15 @@
     },
     {
       "path": "src/codex/catalog/parsing.ts",
-      "sha256": "55375cc7a7162b609a27c4d16bf87ed204737e977fc357dbea285483e8f8d3a5"
+      "sha256": "18f9a24b0770c52d203f8acc446e43e7fba9e42408cc29b612470acd1f0fee8b"
     },
     {
       "path": "src/codex/catalog/provider-fetch.ts",
-      "sha256": "fc052262bc324234b53f65af433c24de5eb71faa954b2038f33fa4fc5854aa9b"
+      "sha256": "cfd71e3d50149f8f04c4d272bb1dd42f1f26bbf276cabb13bb9f21fa5ebf7dc6"
     },
     {
       "path": "src/codex/catalog/sync.ts",
-      "sha256": "221eb6f979016e8d1183529eb83b374223f0b106af441992c87c740ff67cfc36"
+      "sha256": "1d944106ef7674207ed73ae1c7c62befafdd6a4144f4436991af50fe2ada8344"
     },
     {
       "path": "src/codex/codex-write-lock.ts",
@@ -1754,7 +1754,7 @@
     },
     {
       "path": "src/lib/pinned-http.ts",
-      "sha256": "1719469445c480bca840a041ff742a0745339f07adc1da798bd0540acf46ad81"
+      "sha256": "ac1cd748e53b16109aaae86194fddac88424b508f00bb2e7c7090869f58c9b22"
     },
     {
       "path": "src/lib/privacy.ts",
@@ -1854,7 +1854,7 @@
     },
     {
       "path": "src/lib/windows-elevation.ts",
-      "sha256": "effa576dab1f241f1a2d3f6a9f35feb592859d2039ea69f8dab513f0172c3b68"
+      "sha256": "0dc9864f87be8eed6c8dff80bec938f73fa2836c9f606f0e81c4c020494a8850"
     },
     {
       "path": "src/lib/windows-secret-acl.ts",
@@ -2282,7 +2282,7 @@
     },
     {
       "path": "src/server/auth-cors.ts",
-      "sha256": "f96a03d8dc0c4b2021aaeefae404bfec369bf4c3979774836d6ee58de2835c79"
+      "sha256": "1a228cb899ba4b2d10598cae4eaf41de4c0dc27dc0b18b94943f65cf651f386f"
     },
     {
       "path": "src/server/background-lifecycle.ts",
@@ -2298,7 +2298,7 @@
     },
     {
       "path": "src/server/chat-native.ts",
-      "sha256": "00b1e267d7e415af9d279dbf858941ddc16d232bb4c7b52862b78fc9feee4ce1"
+      "sha256": "18ba244fd5ebd4a1b4a78a4b0aead236a4968a347d91e3130f73e2fc1c7f01f6"
     },
     {
       "path": "src/server/claude-messages.ts",
@@ -2378,7 +2378,7 @@
     },
     {
       "path": "src/server/management/config-routes.ts",
-      "sha256": "7225bd064b16d2662cf88f4fc51ad4631fd5c423b62d3c28edbc60414c33571f"
+      "sha256": "90a3722a21edff368f0605300f2e4e2addcc3353ca5d2372d6df99e9ebe03fbd"
     },
     {
       "path": "src/server/management/context.ts",
@@ -2650,7 +2650,7 @@
     },
     {
       "path": "src/service.ts",
-      "sha256": "de767b040554e966d3858bfbdf8b0a4b047ed816bd24b91d7380b7349fe66ba1"
+      "sha256": "2091110dc15b31cf2481875cf894e4b16f1dfc10b5f2c81f55d37ae929049fa1"
     },
     {
       "path": "src/stall-timeout.ts",
@@ -2726,7 +2726,7 @@
     },
     {
       "path": "src/types.ts",
-      "sha256": "36393bc83887926ecb77c8695c1db03e9124e04a982eb2a7d96978438574b9cf"
+      "sha256": "a2b31962f7d7f54f60b1cb9ed1066eeb5b0fb267cd0d3e4053eea6501b53046e"
     },
     {
       "path": "src/update/badge.ts",

package/src/lib/pinned-http.ts CHANGED Viewed

@@ -51,6 +51,7 @@ function pinnedHttpRequest(
     && options?.inactivityTimeoutMs === undefined;
   const firstByteTimeoutMs = options?.firstByteTimeoutMs ?? legacyIdleTimeoutMs;
   const inactivityTimeoutMs = options?.inactivityTimeoutMs ?? legacyIdleTimeoutMs;
+  const legacyFirstByteDisabled = usesLegacyIdleTimeout && legacyIdleTimeoutMs === 0;
   const maxBytes = options?.maxBytes;
   const headers = new Headers(options?.headers);
   headers.set("host", parsed.host);
@@ -88,6 +89,7 @@ function pinnedHttpRequest(
     };
     const startFirstByteTimer = () => {
       clearFirstByteTimer();
+      if (settled || legacyFirstByteDisabled) return;
       firstByteTimer = setTimeout(
         () => fail(new PinnedHttpError("first_byte_timeout", `${context} first byte timed out`)),
         firstByteTimeoutMs,
@@ -194,8 +196,8 @@ function pinnedHttpRequest(
     const requestFn = parsed.protocol === "https:" ? https.request : http.request;
     req = requestFn(requestOptions, onResponse);
+    if (usesLegacyIdleTimeout) startFirstByteTimer();
     const onAbort = () => fail(signal?.reason instanceof Error ? signal.reason : new Error("aborted"));
-    signal?.addEventListener("abort", onAbort, { once: true });
     req.on("socket", (socket) => {
       const connectedEvent = parsed.protocol === "https:" ? "secureConnect" : "connect";
       if (!socket.connecting) {
@@ -233,6 +235,9 @@ function pinnedHttpRequest(
       clearFirstByteTimer();
       signal?.removeEventListener("abort", onAbort);
     });
+    signal?.addEventListener("abort", onAbort, { once: true });
+    if (signal?.aborted && !settled) onAbort();
+    if (settled) return;
     req.end(body);
   });
 }
@@ -262,4 +267,4 @@ export function pinnedHttpPost(
   options?: PinnedHttpRequestOptions,
 ): Promise<Response> {
   return pinnedHttpRequest(url, pinned, "POST", body, signal, options);
-}
+}

package/src/lib/windows-elevation.ts CHANGED Viewed

@@ -648,12 +648,12 @@ export function runWindowsElevatedScheduledTaskRegistration(
       "-EncodedCommand",
       encodedCommand,
     ]))}`,
-    " -Verb RunAs -WindowStyle Hidden -PassThru -Wait",
+    " -Verb RunAs -WindowStyle Hidden -PassThru -Wait;",
     `if ($null -eq $p) { exit ${OCX_ELEVATED_UAC_CANCELLED} }`,
-    "$null = $p.Handle",
+    "$null = $p.Handle;",
     `if ($null -eq $p.ExitCode) { exit ${OCX_ELEVATED_PROTOCOL_FAILED} }`,
     "exit $p.ExitCode",
-  ].join("; ");
+  ].join("");
   return startPowerShellCommand(script).completion.then(result => result.exitCode);
 }