diff --git a/changelog.d/features/8908-model-token-limit-overrides.md b/changelog.d/features/8908-model-token-limit-overrides.md
new file mode 100644
index 0000000000..8d32a9e1f0
--- /dev/null
+++ b/changelog.d/features/8908-model-token-limit-overrides.md
@@ -0,0 +1 @@
+- **feat(models):** add exact per-model `context_length`, `max_input_tokens`, and `max_output_tokens` overrides across model discovery and runtime enforcement, with automatic migration from the retired output-only `max_token` key ([#8908](https://github.com/diegosouzapw/OmniRoute/pull/8908)) — thanks @xz-dev
diff --git a/open-sse/handlers/chatCore.ts b/open-sse/handlers/chatCore.ts
index 913716487e..b74ec33890 100644
--- a/open-sse/handlers/chatCore.ts
+++ b/open-sse/handlers/chatCore.ts
@@ -139,6 +139,7 @@ import {
   supportsMaxTokens,
   getResolvedModelCapabilities,
   getExplicitModelOutputCap,
+  resolveInputTokenCapForGate,
 } from "@/lib/modelCapabilities.ts";
 import { toPositiveInteger } from "../services/reasoningTokenBuffer.ts";
 import { normalizeThinkingForModel } from "@/shared/constants/modelSpecs.ts";
@@ -1857,11 +1858,6 @@ export async function handleChatCore({
     }
   }
 
-  // Key the lookup by { provider, model } — the bare-string form resolves to
-  // `provider: null`, which skips both the registry cap and the operator's
-  // `max_token` capability override (#6524), the documented escape hatch for a
-  // wrong synced `limit_output`. Clamping against a stale spec while the operator
-  // raised the ceiling would silently truncate output.
   const modelOutputCap = toPositiveInteger(
     getExplicitModelOutputCap({ provider, model: effectiveModel })
   );
@@ -1870,13 +1866,15 @@ export async function handleChatCore({
     finalEstimatedInputTokens,
     finalContextLimit,
     targetFormat === FORMATS.CLAUDE && sourceFormat !== FORMATS.CLAUDE ? DEFAULT_MAX_TOKENS : 0,
-    modelOutputCap
+    modelOutputCap,
+    toPositiveInteger(resolveInputTokenCapForGate({ provider, model: effectiveModel }, { isCombo }))
   );
   if (!outputBudget.ok) {
+    const exceededInputCap = outputBudget.maxInputTokens !== undefined;
     const message =
-      `Input exceeds the context window for ${provider}/${effectiveModel}: ` +
-      `estimated ${outputBudget.estimatedInputTokens} input tokens, limit ${outputBudget.contextLimit}. ` +
-      "Reduce the prompt or route to a model with a larger context window.";
+      `Input exceeds ${exceededInputCap ? "maximum input tokens" : "context window"} for ${provider}/${effectiveModel}: ` +
+      `estimated ${outputBudget.estimatedInputTokens} input tokens, ${exceededInputCap ? `max input ${outputBudget.maxInputTokens}` : `limit ${outputBudget.contextLimit}`}. ` +
+      `Reduce the prompt or route to a model with a larger ${exceededInputCap ? "input limit" : "context window"}.`;
     log?.warn?.("CONTEXT", message);
     trackPendingRequest(model, provider, connectionId, false);
     return createErrorResult(
diff --git a/open-sse/handlers/chatCore/outputTokenBudget.ts b/open-sse/handlers/chatCore/outputTokenBudget.ts
index 62752e42f4..3adb97fd81 100644
--- a/open-sse/handlers/chatCore/outputTokenBudget.ts
+++ b/open-sse/handlers/chatCore/outputTokenBudget.ts
@@ -15,6 +15,7 @@ export type OutputTokenBudgetResult =
       ok: false;
       estimatedInputTokens: number;
       contextLimit: number;
+      maxInputTokens?: number | null;
     };
 
 type OutputTokenAdjustment = { field: string; value?: number; remove?: boolean };
@@ -74,19 +75,43 @@ function adjustOutputTokenFields(
  * cap limits how much is requested, not whether the request fits. Absent /
  * null / non-positive cap values leave behavior byte-identical to before this
  * parameter existed (fail-open).
+ *
+ * `maxInputTokenCap` (the model's own input ceiling, `maxInputTokens`) is an
+ * additional, independent input-only bound enforced on the accept/reject
+ * decision. The total-window check (`contextLimit - input >= 1`) stays in place
+ * and remains responsible for reserving output room; the input cap never
+ * double-counts a requested output. Absent / null / non-positive input caps
+ * leave behavior byte-identical (fail-open).
  */
 export function enforceOutputTokenBudget(
   body: Record<string, unknown> | null | undefined,
   estimatedInputTokens: number,
   contextLimit: number,
   defaultOutputTokens = 0,
-  maxOutputTokenCap?: number | null
+  maxOutputTokenCap?: number | null,
+  maxInputTokenCap?: number | null
 ): OutputTokenBudgetResult {
   const normalizedInputTokens = Math.max(0, Math.ceil(estimatedInputTokens));
   const normalizedContextLimit = Math.max(1, Math.floor(contextLimit));
   const normalizedDefaultOutputTokens = Math.max(0, Math.floor(defaultOutputTokens));
   const availableOutputTokens = normalizedContextLimit - normalizedInputTokens;
 
+  // Independent input-only ceiling: reject when the prompt alone exceeds the
+  // model's declared max input, regardless of remaining output room.
+  const normalizedInputCap = maxInputTokenCap == null ? null : Math.floor(maxInputTokenCap);
+  if (
+    normalizedInputCap !== null &&
+    normalizedInputCap > 0 &&
+    normalizedInputTokens > normalizedInputCap
+  ) {
+    return {
+      ok: false,
+      estimatedInputTokens: normalizedInputTokens,
+      contextLimit: normalizedContextLimit,
+      maxInputTokens: normalizedInputCap,
+    };
+  }
+
   if (availableOutputTokens < 1) {
     return {
       ok: false,
diff --git a/open-sse/services/combo/contextOverrideGate.ts b/open-sse/services/combo/contextOverrideGate.ts
index 605c27c13f..4f03978a45 100644
--- a/open-sse/services/combo/contextOverrideGate.ts
+++ b/open-sse/services/combo/contextOverrideGate.ts
@@ -16,14 +16,13 @@
  * pool to one provider and producing a hard 503 with no fallback once that
  * provider's quota is exhausted. An operator-set or auto-discovered override
  * reflects the real capacity, so it supersedes both catalog limits. Uses the
- * raw override (`getModelContextOverride` returns `null` when none is set) —
+ * resolved exact override (`getResolvedModelContextOverride` returns `null` when none is set) —
  * NOT `getModelContextLimitForModelString`, which falls back to
  * `contextWindow` and would therefore bypass the `maxInputTokens` cap for
  * every model, not just overridden ones.
  */
 
-import { getModelContextOverride } from "../../../src/lib/db/modelContextOverrides";
-import { parseModel } from "../model.ts";
+import { getResolvedModelContextOverride } from "../../../src/lib/modelCapabilities";
 
 /**
  * Resolve the context-fit verdict from a persisted per-model override, if one
@@ -36,8 +35,7 @@ function resolveContextOverrideVerdict(
   requiredContextTokens: number
 ): boolean | undefined {
   if (!modelStr) return undefined;
-  const parsed = parseModel(modelStr);
-  const override = getModelContextOverride(parsed.provider, parsed.model);
+  const override = getResolvedModelContextOverride(modelStr);
   if (override == null) return undefined;
   return override >= requiredContextTokens;
 }
diff --git a/src/app/(dashboard)/dashboard/settings/components/ModelCapabilityOverridesTab.tsx b/src/app/(dashboard)/dashboard/settings/components/ModelCapabilityOverridesTab.tsx
index 04d182b6b5..a4d3c2cb14 100644
--- a/src/app/(dashboard)/dashboard/settings/components/ModelCapabilityOverridesTab.tsx
+++ b/src/app/(dashboard)/dashboard/settings/components/ModelCapabilityOverridesTab.tsx
@@ -5,7 +5,7 @@ import { useTranslations } from "next-intl";
 import { Card, Button } from "@/shared/components";
 import { matchesSearch } from "@/shared/utils/turkishText";
 
-type ModelOverrideKey = "max_token";
+type ModelOverrideKey = "context_length" | "max_input_tokens" | "max_output_tokens";
 type StatusTone = "success" | "error" | "info";
 
 type ModelOverrideTarget = {
@@ -337,7 +337,7 @@ function ModelOverrideForm({
   onSave: (target: string, key: ModelOverrideKey, value: number) => void;
 }) {
   const t = useTranslations("settings");
-  const [key, setKey] = useState<ModelOverrideKey>("max_token");
+  const [key, setKey] = useState<ModelOverrideKey>("context_length");
   const [value, setValue] = useState("");
   const numericValue = Number(value);
   const saveDisabled = !activeTarget || !Number.isInteger(numericValue) || numericValue <= 0;
@@ -349,7 +349,9 @@ function ModelOverrideForm({
         onChange={(event) => setKey(event.target.value as ModelOverrideKey)}
         className="sm:w-40 px-2 py-2 text-xs bg-bg-base border border-border rounded-md focus:outline-none focus:border-primary"
       >
-        <option value="max_token">max_token</option>
+        <option value="context_length">context_length</option>
+        <option value="max_input_tokens">max_input_tokens</option>
+        <option value="max_output_tokens">max_output_tokens</option>
       </select>
       <input
         type="number"
diff --git a/src/app/api/model-capability-overrides/route.ts b/src/app/api/model-capability-overrides/route.ts
index 8b7b456579..0cb73d7a4f 100644
--- a/src/app/api/model-capability-overrides/route.ts
+++ b/src/app/api/model-capability-overrides/route.ts
@@ -6,10 +6,33 @@ import {
   listModelCapabilityOverrides,
   removeModelCapabilityOverride,
   setModelCapabilityOverride,
+  type ModelCapabilityOverride,
   type ModelCapabilityOverrideKey,
 } from "@/lib/db/modelCapabilityOverrides";
+import {
+  listModelContextOverrides,
+  removeModelContextOverride,
+  setModelContextOverride,
+} from "@/lib/db/modelContextOverrides";
 
-const overrideKeySchema = z.enum(["max_token"]);
+const overrideKeySchema = z.enum(["context_length", "max_input_tokens", "max_output_tokens"]);
+type PublicOverrideKey = z.infer<typeof overrideKeySchema>;
+type PublicOverride = Omit<ModelCapabilityOverride, "key"> & { key: PublicOverrideKey };
+
+function listPublicOverrides(): PublicOverride[] {
+  const capabilityOverrides = listModelCapabilityOverrides() as PublicOverride[];
+  const contextOverrides = listModelContextOverrides().map((override): PublicOverride => ({
+    provider: override.provider,
+    modelId: override.modelId,
+    target: `${override.provider}/${override.modelId}`,
+    key: "context_length",
+    value: override.realContext,
+    refreshedAt: override.refreshedAt,
+  }));
+  return [...capabilityOverrides, ...contextOverrides].sort((left, right) =>
+    right.refreshedAt.localeCompare(left.refreshedAt)
+  );
+}
 
 const upsertOverrideSchema = z.object({
   target: z.string().min(3),
@@ -33,7 +56,7 @@ export async function GET(request: Request) {
   const authError = await requireManagementAuth(request);
   if (authError) return authError;
 
-  return NextResponse.json({ overrides: listModelCapabilityOverrides() });
+  return NextResponse.json({ overrides: listPublicOverrides() });
 }
 
 export async function PATCH(request: Request) {
@@ -57,12 +80,20 @@ export async function PATCH(request: Request) {
     return NextResponse.json({ error: "Invalid model capability override" }, { status: 400 });
   }
 
-  const written = setModelCapabilityOverride(target, parsed.data.key, parsed.data.value);
+  const targetParts = target.split(/\/(.*)/s);
+  const written =
+    parsed.data.key === "context_length"
+      ? setModelContextOverride(targetParts[0], targetParts[1], parsed.data.value, "manual")
+      : setModelCapabilityOverride(
+          target,
+          parsed.data.key as ModelCapabilityOverrideKey,
+          parsed.data.value
+        );
   if (!written) {
     return NextResponse.json({ error: "Invalid model capability override" }, { status: 400 });
   }
 
-  return NextResponse.json({ overrides: listModelCapabilityOverrides() });
+  return NextResponse.json({ overrides: listPublicOverrides() });
 }
 
 export async function DELETE(request: Request) {
@@ -78,6 +109,11 @@ export async function DELETE(request: Request) {
     return NextResponse.json({ error: "target and key are required" }, { status: 400 });
   }
 
-  removeModelCapabilityOverride(target, parsedKey.data as ModelCapabilityOverrideKey);
-  return NextResponse.json({ overrides: listModelCapabilityOverrides() });
+  if (parsedKey.data === "context_length") {
+    const targetParts = target.split(/\/(.*)/s);
+    removeModelContextOverride(targetParts[0], targetParts[1]);
+  } else {
+    removeModelCapabilityOverride(target, parsedKey.data as ModelCapabilityOverrideKey);
+  }
+  return NextResponse.json({ overrides: listPublicOverrides() });
 }
diff --git a/src/app/api/v1/models/catalogCache.ts b/src/app/api/v1/models/catalogCache.ts
index 316120293c..1cff2fa64f 100644
--- a/src/app/api/v1/models/catalogCache.ts
+++ b/src/app/api/v1/models/catalogCache.ts
@@ -66,7 +66,17 @@ export const CATALOG_STALE_WHILE_REVALIDATE_MS = 30_000;
 export const CATALOG_CACHE_TTL_MS_DEFAULT = 60_000;
 
 const catalogCache = new Map<string, CachedCatalog>();
-const catalogInFlight = new Map<string, Promise<CachedCatalog>>();
+
+/**
+ * An in-flight build is bound to the catalog-state generation it started from
+ * (`getModelCatalogCacheVersion()` at launch). After a write invalidates the
+ * catalog, the generation moves on: a stale in-flight build must neither be
+ * joined by new requests nor repopulate the now-current cache when it finishes.
+ * It still resolves to its own original caller (that request legitimately waits
+ * on it), just without being persisted.
+ */
+type InFlightBuild = { generation: number; promise: Promise<CachedCatalog> };
+const catalogInFlight = new Map<string, InFlightBuild>();
 
 let _catalogBuilderRuns = 0;
 
@@ -90,10 +100,12 @@ function dropCatalogCacheIfStateChanged(): void {
   if (currentVersion === lastSeenCatalogCacheVersion) return;
   lastSeenCatalogCacheVersion = currentVersion;
   catalogCache.clear();
-  // Deliberately NOT clearing catalogInFlight: an in-flight build already reads live
-  // DB/settings state as of when it started, so letting it finish and populate the
-  // (now-current) cache entry is correct — clearing it would just force a redundant
-  // second builder run for requests that arrive mid-flight.
+  // Deliberately NOT clearing catalogInFlight: an in-flight build bound to the
+  // previous generation is left to finish for its original caller, but the
+  // generation check in the join path (below) keeps new requests from joining
+  // it, and the generation check in storePayload keeps it from repopulating
+  // the now-current cache. Clearing it here would just detach the entry while
+  // the build still ran — wasted work with no correctness gain.
 }
 
 // Header sources mix Title-Case keys (diagnostic/cors headers built by app code) with
@@ -116,14 +128,27 @@ export function mergeCatalogHeaders(
   return merged;
 }
 
-function storePayload(cacheKey: string, payload: CatalogPayload): CachedCatalog {
+/**
+ * Persist a freshly built payload — but only when the build still belongs to the
+ * current catalog-state generation. A build that started before a write
+ * invalidation (its `buildGeneration` is older than `getModelCatalogCacheVersion()`)
+ * returns its entry to its original caller but must NOT repopulate the cache: the
+ * payload reflects pre-write state and caching it would serve stale data.
+ */
+function storePayload(
+  cacheKey: string,
+  payload: CatalogPayload,
+  buildGeneration: number
+): CachedCatalog {
   const entry: CachedCatalog = {
     body: payload.body,
     headers: payload.headers,
     status: payload.status,
     expiresAt: Date.now() + payload.cacheTTL,
   };
-  catalogCache.set(cacheKey, entry);
+  if (buildGeneration === getModelCatalogCacheVersion()) {
+    catalogCache.set(cacheKey, entry);
+  }
   return entry;
 }
 
@@ -151,10 +176,11 @@ function scheduleBackgroundRefresh(
 ): void {
   if (catalogInFlight.has(cacheKey)) return; // a refresh for this key is already running
 
+  const generation = getModelCatalogCacheVersion();
   const refreshPromise: Promise<CachedCatalog> = new Promise((resolve, reject) => {
     setTimeout(() => {
       runBuilder(buildPayload, request)
-        .then((payload) => resolve(storePayload(cacheKey, payload)))
+        .then((payload) => resolve(storePayload(cacheKey, payload, generation)))
         .catch((err) => {
           console.error(
             `[catalog] Background stale-while-revalidate refresh failed for key "${cacheKey}":`,
@@ -170,11 +196,12 @@ function scheduleBackgroundRefresh(
   // observes the failure.
   refreshPromise.catch(() => {});
 
-  catalogInFlight.set(cacheKey, refreshPromise);
+  catalogInFlight.set(cacheKey, { generation, promise: refreshPromise });
   refreshPromise
     .catch(() => {})
     .finally(() => {
-      if (catalogInFlight.get(cacheKey) === refreshPromise) catalogInFlight.delete(cacheKey);
+      if (catalogInFlight.get(cacheKey)?.promise === refreshPromise)
+        catalogInFlight.delete(cacheKey);
     });
 }
 
@@ -229,16 +256,24 @@ export async function resolveCachedCatalogResponse(
     });
   }
 
+  const currentGeneration = getModelCatalogCacheVersion();
   let inflight = catalogInFlight.get(cacheKey);
-  if (!inflight) {
-    inflight = runBuilder(buildPayload, request).then((payload) => storePayload(cacheKey, payload));
+  // Only join an in-flight build from the CURRENT generation. A build bound to an
+  // older (pre-write) generation reflects stale state, so a new request starts a
+  // fresh build instead of joining it.
+  if (!inflight || inflight.generation !== currentGeneration) {
+    const generation = currentGeneration;
+    const promise = runBuilder(buildPayload, request).then((payload) =>
+      storePayload(cacheKey, payload, generation)
+    );
+    inflight = { generation, promise };
     catalogInFlight.set(cacheKey, inflight);
-    inflight.finally(() => {
-      if (catalogInFlight.get(cacheKey) === inflight) catalogInFlight.delete(cacheKey);
+    promise.finally(() => {
+      if (catalogInFlight.get(cacheKey)?.promise === promise) catalogInFlight.delete(cacheKey);
     });
   }
 
-  const payload = await inflight;
+  const payload = await inflight.promise;
   return new Response(payload.body, {
     status: payload.status,
     headers: mergeCatalogHeaders(corsHeaders, payload.headers, diagnosticHeaders),
@@ -284,7 +319,7 @@ export function __setCatalogCacheEntryForTest(request: Request, entry: CachedCat
 
 /** Awaits any background refresh in flight, instead of guessing at a real-time sleep. */
 export async function __flushCatalogBackgroundRefreshForTest(): Promise<void> {
-  await Promise.all([...catalogInFlight.values()].map((p) => p.catch(() => {})));
+  await Promise.all([...catalogInFlight.values()].map((entry) => entry.promise.catch(() => {})));
 }
 
 /**
@@ -301,5 +336,10 @@ export async function __flushCatalogBackgroundRefreshForTest(): Promise<void> {
 export function __forceCatalogInFlightRejectionForTest(request: Request, error: unknown): void {
   const rejected: Promise<CachedCatalog> = Promise.reject(error);
   rejected.catch(() => {}); // mark as handled — avoids an unhandledRejection warning
-  catalogInFlight.set(buildCatalogCacheKey(request), rejected);
+  // Bind to the current generation so the cold path still joins it (a stale
+  // generation would be skipped as pre-write state and never awaited).
+  catalogInFlight.set(buildCatalogCacheKey(request), {
+    generation: getModelCatalogCacheVersion(),
+    promise: rejected,
+  });
 }
diff --git a/src/lib/contextWindowResolver.ts b/src/lib/contextWindowResolver.ts
index bcd6cc132f..8e7124ea85 100644
--- a/src/lib/contextWindowResolver.ts
+++ b/src/lib/contextWindowResolver.ts
@@ -91,8 +91,12 @@ export async function runContextWindowReconcile(): Promise<ReconcileResult> {
   const byProvider = await getAllSyncedAvailableModels();
   const discovered = toDiscoveredWindows(byProvider);
   return reconcileContextWindows(discovered, {
+    // Compare against the override-free catalog view. A persisted override must
+    // never feed back into the comparison that (re)writes it, or the reconciler
+    // oscillates (write → equal → remove → differ → write ...).
     getCatalogWindow: (provider, modelId) =>
-      getResolvedModelCapabilities({ provider, model: modelId }).contextWindow,
+      getResolvedModelCapabilities({ provider, model: modelId }, { persistedOverrides: false })
+        .contextWindow,
     getExistingSource: (provider, modelId) =>
       getModelContextOverrideRecord(provider, modelId)?.source ?? null,
     writeAuto: (provider, modelId, window) => {
diff --git a/src/lib/db/migrations/135_migrate_model_capability_max_token.sql b/src/lib/db/migrations/135_migrate_model_capability_max_token.sql
new file mode 100644
index 0000000000..be8fbecfda
--- /dev/null
+++ b/src/lib/db/migrations/135_migrate_model_capability_max_token.sql
@@ -0,0 +1,24 @@
+-- 135_migrate_model_capability_max_token.sql
+-- `max_token` historically meant the maximum output token count. Promote legacy
+-- rows to the explicit `max_output_tokens` key without replacing an operator's
+-- existing modern value, then remove the retired key.
+
+INSERT INTO model_capability_overrides (
+  provider,
+  model_id,
+  override_key,
+  override_value,
+  refreshed_at
+)
+SELECT
+  provider,
+  model_id,
+  'max_output_tokens',
+  override_value,
+  refreshed_at
+FROM model_capability_overrides
+WHERE override_key = 'max_token'
+ON CONFLICT (provider, model_id, override_key) DO NOTHING;
+
+DELETE FROM model_capability_overrides
+WHERE override_key = 'max_token';
diff --git a/src/lib/db/modelCapabilityOverrides.ts b/src/lib/db/modelCapabilityOverrides.ts
index 36ad9364e8..387f76698a 100644
--- a/src/lib/db/modelCapabilityOverrides.ts
+++ b/src/lib/db/modelCapabilityOverrides.ts
@@ -1,6 +1,7 @@
 import { getDbInstance } from "./core";
+import { invalidateDbCache } from "./readCache";
 
-export type ModelCapabilityOverrideKey = "max_token";
+export type ModelCapabilityOverrideKey = "max_input_tokens" | "max_output_tokens";
 
 export interface ModelCapabilityOverride {
   provider: string;
@@ -20,7 +21,7 @@ interface OverrideRow {
 }
 
 function isSupportedKey(value: unknown): value is ModelCapabilityOverrideKey {
-  return value === "max_token";
+  return value === "max_input_tokens" || value === "max_output_tokens";
 }
 
 function isPositiveInteger(value: unknown): value is number {
@@ -99,6 +100,7 @@ export function setModelCapabilityOverride(
         "VALUES (?, ?, ?, ?, datetime('now'))"
     )
     .run(parsedTarget.provider, parsedTarget.modelId, key, JSON.stringify(value));
+  invalidateDbCache("model-capabilities");
   return true;
 }
 
@@ -115,6 +117,7 @@ export function removeModelCapabilityOverride(
         "WHERE provider = ? AND model_id = ? AND override_key = ?"
     )
     .run(parsedTarget.provider, parsedTarget.modelId, key);
+  if (info.changes > 0) invalidateDbCache("model-capabilities");
   return info.changes > 0;
 }
 
diff --git a/src/lib/db/modelContextOverrides.ts b/src/lib/db/modelContextOverrides.ts
index 40b7326a1e..7c531cc7d3 100644
--- a/src/lib/db/modelContextOverrides.ts
+++ b/src/lib/db/modelContextOverrides.ts
@@ -1,4 +1,5 @@
 import { getDbInstance } from "./core";
+import { invalidateDbCache } from "./readCache";
 
 /**
  * Feature 5004 — self-correcting context-window overrides.
@@ -36,7 +37,10 @@ function isPositiveInteger(value: unknown): value is number {
   return typeof value === "number" && Number.isInteger(value) && value > 0;
 }
 
-function normalizeKey(provider: unknown, modelId: unknown): { provider: string; modelId: string } | null {
+function normalizeKey(
+  provider: unknown,
+  modelId: unknown
+): { provider: string; modelId: string } | null {
   const p = typeof provider === "string" ? provider.trim() : "";
   const m = typeof modelId === "string" ? modelId.trim() : "";
   if (!p || !m) return null;
@@ -104,6 +108,7 @@ export function setModelContextOverride(
         "VALUES (?, ?, ?, ?, datetime('now'))"
     )
     .run(key.provider, key.modelId, realContext, normalizedSource);
+  invalidateDbCache("model-capabilities");
   return true;
 }
 
@@ -114,6 +119,7 @@ export function removeModelContextOverride(provider: string, modelId: string): b
   const info = getDbInstance()
     .prepare("DELETE FROM model_context_overrides WHERE provider = ? AND model_id = ?")
     .run(key.provider, key.modelId);
+  if (info.changes > 0) invalidateDbCache("model-capabilities");
   return info.changes > 0;
 }
 
diff --git a/src/lib/db/readCache.ts b/src/lib/db/readCache.ts
index b8f520fe19..79fff8c2d2 100644
--- a/src/lib/db/readCache.ts
+++ b/src/lib/db/readCache.ts
@@ -102,9 +102,7 @@ export async function getCachedPricing(): Promise<Record<string, unknown>> {
 export async function getCachedProviderConnections(
   filter?: Record<string, unknown>
 ): Promise<unknown[]> {
-  const cacheKey = filter && Object.keys(filter).length > 0
-    ? JSON.stringify(filter)
-    : "all";
+  const cacheKey = filter && Object.keys(filter).length > 0 ? JSON.stringify(filter) : "all";
 
   const cached = connectionsCache.get(cacheKey);
   if (cached) return cached;
@@ -136,7 +134,10 @@ export async function getCachedRawProviderConnections(
   return rows;
 }
 
-const connectionByIdCache = new TTLCache<Record<string, unknown> | null>(CONNECTIONS_TTL_MS, 10_000);
+const connectionByIdCache = new TTLCache<Record<string, unknown> | null>(
+  CONNECTIONS_TTL_MS,
+  10_000
+);
 const nodesCache = new TTLCache<(Record<string, unknown> | null)[]>(CONNECTIONS_TTL_MS);
 
 /**
@@ -263,7 +264,7 @@ export function getModelCatalogCacheVersion(): number {
  * cannot be selectively invalidated).
  */
 export function invalidateDbCache(
-  scope?: "settings" | "pricing" | "connections" | "combos" | "nodes",
+  scope?: "settings" | "pricing" | "connections" | "combos" | "nodes" | "model-capabilities",
   id?: string
 ): void {
   if (!scope || scope === "settings") settingsCache.invalidate();
diff --git a/src/lib/modelCapabilities.ts b/src/lib/modelCapabilities.ts
index 3deb72832b..c8152b18fb 100644
--- a/src/lib/modelCapabilities.ts
+++ b/src/lib/modelCapabilities.ts
@@ -100,6 +100,16 @@ type CapabilityInput =
 
 type SyncedCapabilities = ReturnType<typeof getSyncedCapability>;
 
+/**
+ * Controls whether persisted operator/discovery overrides participate in resolution.
+ * Omit it (the public default) to resolve effective runtime capabilities. Catalog
+ * reconciliation alone uses `persistedOverrides: false` to compare discovery with
+ * static/synced catalog data without feeding an existing override back into itself.
+ */
+export interface ResolveModelCapabilitiesOptions {
+  persistedOverrides?: boolean;
+}
+
 export interface ResolvedModelCapabilities {
   provider: string | null;
   model: string | null;
@@ -494,7 +504,7 @@ function resolveVisionCapability(
 }
 
 /**
- * Issue #6524: an operator-set `max_token` capability override (see
+ * Issue #6524: an operator-set `max_output_tokens` capability override (see
  * `src/lib/db/modelCapabilityOverrides.ts`) is the manual escape hatch for a
  * wrong/stale synced `limit_output` value (e.g. a provider's models.dev catalog
  * row reporting `limit_output` equal to `limit_context`). It already won over the
@@ -502,22 +512,66 @@ function resolveVisionCapability(
  * makes `getExplicitModelOutputCap()` (used by the reasoning-token-buffer clamp)
  * consult the same override so both read paths agree.
  */
-function getMaxTokenCapabilityOverride(resolved: {
+/**
+ * Exact-match capability override lookup with intentional raw-alias fallback.
+ *
+ * An override may be stored under either the canonical model id or the exact
+ * provider-scoped raw alias the operator used (e.g. `github/claude-opus-4.5`
+ * resolving to canonical `claude-opus-4-5-20251101`). We consult the canonical
+ * id first, then the raw alias — both are exact provider/model matches. There is
+ * deliberately NO suffix/effort/family inheritance: an override for
+ * `codex/gpt-5.6` never applies to `codex/gpt-5.6-high`.
+ */
+function getCapabilityOverride(
+  resolved: { provider: string | null; model: string | null; rawModel: string | null },
+  key: "max_input_tokens" | "max_output_tokens"
+): number | null {
+  const canonical = getModelCapabilityOverride(resolved.provider, resolved.model, key);
+  if (canonical !== null) return canonical;
+  return resolved.rawModel && resolved.rawModel !== resolved.model
+    ? getModelCapabilityOverride(resolved.provider, resolved.rawModel, key)
+    : null;
+}
+
+function getContextOverride(resolved: {
   provider: string | null;
   model: string | null;
   rawModel: string | null;
 }): number | null {
-  return (
-    getModelCapabilityOverride(resolved.provider, resolved.model, "max_token") ??
-    (resolved.rawModel && resolved.rawModel !== resolved.model
-      ? getModelCapabilityOverride(resolved.provider, resolved.rawModel, "max_token")
-      : null)
-  );
+  const canonical = getModelContextOverride(resolved.provider, resolved.model);
+  if (canonical !== null) return canonical;
+  return resolved.rawModel && resolved.rawModel !== resolved.model
+    ? getModelContextOverride(resolved.provider, resolved.rawModel)
+    : null;
+}
+
+/**
+ * Resolve a persisted context override by canonical id, then by the exact raw
+ * alias supplied by the caller. Neither lookup inherits to related models.
+ */
+export function getResolvedModelContextOverride(input: CapabilityInput): number | null {
+  return getContextOverride(resolveCapabilityInput(input));
+}
+
+function getInputTokenCapabilityOverride(resolved: {
+  provider: string | null;
+  model: string | null;
+  rawModel: string | null;
+}): number | null {
+  return getCapabilityOverride(resolved, "max_input_tokens");
+}
+
+function getOutputTokenCapabilityOverride(resolved: {
+  provider: string | null;
+  model: string | null;
+  rawModel: string | null;
+}): number | null {
+  return getCapabilityOverride(resolved, "max_output_tokens");
 }
 
 export function getExplicitModelOutputCap(input: CapabilityInput): number | null {
   const resolved = resolveCapabilityInput(input);
-  const maxTokenOverride = getMaxTokenCapabilityOverride(resolved);
+  const maxTokenOverride = getOutputTokenCapabilityOverride(resolved);
   if (maxTokenOverride !== null) return maxTokenOverride;
 
   const synced = getSyncedCapabilityForResolved(
@@ -534,7 +588,13 @@ export function getExplicitModelOutputCap(input: CapabilityInput): number | null
   return spec?.maxOutputTokens ?? null;
 }
 
-export function getResolvedModelCapabilities(input: CapabilityInput): ResolvedModelCapabilities {
+export function getResolvedModelCapabilities(
+  input: CapabilityInput,
+  options?: ResolveModelCapabilitiesOptions
+): ResolvedModelCapabilities {
+  // Reconciliation / auto-discovery needs the override-free catalog view so a
+  // persisted override never feeds back into the comparison that (re)writes it.
+  const usePersistedOverrides = options?.persistedOverrides !== false;
   const resolved = resolveCapabilityInput(input);
   const spec = getStaticSpec(resolved.model, resolved.rawModel);
   const registryModel = getRegistryModel(resolved.provider, resolved.model);
@@ -585,14 +645,23 @@ export function getResolvedModelCapabilities(input: CapabilityInput): ResolvedMo
     resolved.model,
     resolved.rawModel
   );
+  // A persisted context-window override (operator-set or auto-discovered)
+  // reflects the real *total* window and wins over every static/synced source.
+  // `maxInputTokens` still follows its own precedence chain; only when that
+  // chain has no narrower source does it naturally fall back to this window.
+  const persistedContextWindow = usePersistedOverrides ? getContextOverride(resolved) : null;
   const contextWindow =
+    persistedContextWindow ??
     authoritativeContextWindow ??
     synced?.limit_context ??
     (typeof registryModel?.contextLength === "number" ? registryModel.contextLength : null) ??
     spec?.contextWindow ??
     null;
 
-  const maxTokenOverride = getMaxTokenCapabilityOverride(resolved);
+  const maxInputOverride = usePersistedOverrides ? getInputTokenCapabilityOverride(resolved) : null;
+  const maxTokenOverride = usePersistedOverrides
+    ? getOutputTokenCapabilityOverride(resolved)
+    : null;
 
   // Vision consults leaf static metadata for path-shaped ids; other capability
   // fields keep using the non-leaf `spec` from getStaticSpec() above.
@@ -628,11 +697,22 @@ export function getResolvedModelCapabilities(input: CapabilityInput): ResolvedMo
     structuredOutput: synced?.structured_output ?? null,
     temperature: synced?.temperature ?? null,
     contextWindow,
-    maxInputTokens:
-      (typeof registryModel?.maxInputTokens === "number" ? registryModel.maxInputTokens : null) ??
-      authoritativeContextWindow ??
-      synced?.limit_input ??
-      contextWindow,
+    maxInputTokens: (() => {
+      // Input cap is input-only. An explicit `max_input_tokens` override wins;
+      // otherwise fall back to the existing per-source input limits, then to the
+      // total window. The effective cap can never exceed the total window
+      // (input + output), so clamp it — but never double-count a requested
+      // output against this input cap.
+      const candidate =
+        maxInputOverride ??
+        (typeof registryModel?.maxInputTokens === "number" ? registryModel.maxInputTokens : null) ??
+        authoritativeContextWindow ??
+        synced?.limit_input ??
+        contextWindow;
+      return candidate !== null && contextWindow !== null
+        ? Math.min(candidate, contextWindow)
+        : candidate;
+    })(),
     maxOutputTokens:
       maxTokenOverride ??
       synced?.limit_output ??
@@ -657,6 +737,51 @@ export function getResolvedModelCapabilities(input: CapabilityInput): ResolvedMo
   };
 }
 
+/**
+ * Input cap enforced at the request-time hard gate, with explicit combo semantics.
+ *
+ * Feature 5004 lets a raw, exact `model_context_overrides` entry supersede a
+ * deliberately smaller catalog/client input hint for COMBO routing: the combo
+ * compatibility filter (`open-sse/services/combo/contextOverrideGate.ts`) already
+ * rescues such targets, so the final hard gate in handleChatCore must not turn
+ * around and reject the rescued target on the very hint the filter bypassed.
+ *
+ * Semantics (deliberately narrow — NO suffix/effort/family inheritance):
+ * - An explicit `max_input_tokens` capability override is ALWAYS enforced, for
+ *   direct and combo requests alike. It is the operator's input-only ceiling and
+ *   must never be bypassed by a context-window override.
+ * - Otherwise, for a COMBO request with an exact persisted context override, that
+ *   context override is the input cap (it reflects the real window; the smaller
+ *   catalog hint does not apply). Direct requests ignore this branch.
+ * - Otherwise the canonical `maxInputTokens` chain applies (registry input hint →
+ *   authoritative window → synced limit_input → total window), clamped to the
+ *   total window.
+ *
+ * `isCombo` selects the combo-rescue branch; pass `false`/omit for direct calls.
+ * The returned cap is still only an *input* bound — the total-window/output
+ * reserve check is enforced separately by `enforceOutputTokenBudget`.
+ */
+export function resolveInputTokenCapForGate(
+  input: CapabilityInput,
+  { isCombo = false }: { isCombo?: boolean } = {}
+): number | null {
+  const resolved = resolveCapabilityInput(input);
+
+  // 1. An explicit `max_input_tokens` override always wins and is never bypassed.
+  const explicitInputOverride = getInputTokenCapabilityOverride(resolved);
+  if (explicitInputOverride !== null) return explicitInputOverride;
+
+  // 2. Combo rescue: an exact persisted context override supersedes the smaller
+  //    catalog/client input hint (mirrors contextOverrideGate.evaluateContextLimit).
+  if (isCombo) {
+    const contextOverride = getContextOverride(resolved);
+    if (contextOverride !== null) return contextOverride;
+  }
+
+  // 3. Canonical chain (already clamped to the total window by the resolver).
+  return getResolvedModelCapabilities(input).maxInputTokens;
+}
+
 export function supportsToolCalling(input: CapabilityInput): boolean {
   if (typeof input === "string" && !String(input || "").trim()) return false;
   return getResolvedModelCapabilities(input).toolCalling;
@@ -732,9 +857,5 @@ export function getModelContextLimit(
     typeof providerOrInput === "string" && modelId !== undefined
       ? getResolvedModelCapabilities({ provider: providerOrInput, model: modelId })
       : getResolvedModelCapabilities(providerOrInput);
-  // Feature 5004: a persisted override (operator-set or auto-discovered) wins over the
-  // static catalog / models.dev sync. `getResolvedModelCapabilities` stays override-free
-  // so the reconciler can compare the catalog value against provider-declared windows.
-  const override = getModelContextOverride(resolved.provider, resolved.model);
-  return override ?? resolved.contextWindow;
+  return resolved.contextWindow;
 }
diff --git a/src/lib/modelMetadataRegistry.ts b/src/lib/modelMetadataRegistry.ts
index e0d49968b7..72fd26c67f 100644
--- a/src/lib/modelMetadataRegistry.ts
+++ b/src/lib/modelMetadataRegistry.ts
@@ -2,7 +2,11 @@ import { randomUUID } from "node:crypto";
 import { parseModel } from "@omniroute/open-sse/services/model.ts";
 import { getModelInfo } from "@/sse/services/model";
 import { getModelAliases } from "@/lib/db/models";
-import { getResolvedModelCapabilities, isNonChatCatalogSurface } from "@/lib/modelCapabilities";
+import {
+  getResolvedModelCapabilities,
+  getResolvedModelContextOverride,
+  isNonChatCatalogSurface,
+} from "@/lib/modelCapabilities";
 import {
   getAuthoritativeContextWindow,
   getAuthoritativeProviderContextWindow,
@@ -418,6 +422,7 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
     getAuthoritativeContextWindow(metadata.model) ??
     getAuthoritativeContextWindow(model);
   const specialtySurface = isNonChatCatalogSurface(entry.type);
+  const persistedContextWindow = getResolvedModelContextOverride({ provider, model });
   const capabilityFields = {
     ...(typeof metadata.capabilities.vision === "boolean"
       ? { vision: metadata.capabilities.vision }
@@ -486,16 +491,21 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
 
   if (
     !specialtySurface &&
-    (typeof nextEntry.context_length !== "number" || authoritativeContextWindow !== null) &&
+    (typeof nextEntry.context_length !== "number" ||
+      authoritativeContextWindow !== null ||
+      persistedContextWindow !== null) &&
     typeof metadata.limits.contextWindow === "number"
   ) {
     nextEntry.context_length = metadata.limits.contextWindow;
+  } else if (specialtySurface && persistedContextWindow !== null) {
+    // Exact persisted overrides are authoritative for every surface of the model.
+    nextEntry.context_length = persistedContextWindow;
   } else if (
     specialtySurface &&
     authoritativeContextWindow !== null &&
     typeof authoritativeContextWindow === "number"
   ) {
-    // Only authoritative static windows may decorate specialty rows.
+    // Only authoritative static windows may otherwise decorate specialty rows.
     nextEntry.context_length = authoritativeContextWindow;
   } else if (specialtySurface && typeof nextEntry.context_length === "number") {
     // Keep an explicit source-provided context if the emitter already set one.
diff --git a/tests/unit/chatcore-combo-context-override-rescue.test.ts b/tests/unit/chatcore-combo-context-override-rescue.test.ts
new file mode 100644
index 0000000000..67b716ed3c
--- /dev/null
+++ b/tests/unit/chatcore-combo-context-override-rescue.test.ts
@@ -0,0 +1,137 @@
+// Black-box regression for the Feature-5004 combo context-override rescue at the
+// handleChatCore final input gate.
+//
+// The combo compatibility filter (open-sse/services/combo/contextOverrideGate.ts)
+// deliberately lets a raw, exact `model_context_overrides` entry supersede a smaller
+// catalog/client input hint so a capable fallback target is not wrongly dropped for a
+// large prompt. But handleChatCore then enforced the canonical `maxInputTokens` again
+// at the final hard boundary — so a target the filter just rescued was rejected before
+// the upstream fetch (#5004 combo rescue defeated at the gate).
+//
+// Fix: the final gate resolves the input cap through `resolveInputTokenCapForGate`,
+// which mirrors the filter's semantics — an exact context override supersedes the
+// smaller catalog input hint for combo requests, while an explicit `max_input_tokens`
+// capability override is ALWAYS enforced (combo and direct alike). No inheritance is
+// generalized.
+//
+// GitHub's Claude alias canonicalizes before capability lookup, so a prompt sized
+// between the catalog input hint and an exact raw-alias context override exercises
+// the rescue the filter performs. These tests drive the public handleChatCore with a
+// temp DB and an upstream fetch stub and assert on the
+// dispatch signal (did the upstream fetch fire, and was the rejection the 400
+// input-cap error) — the precise contract of the gate.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-combo-ctx-rescue-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const contextOverrides = await import("../../src/lib/db/modelContextOverrides.ts");
+const capabilityOverrides = await import("../../src/lib/db/modelCapabilityOverrides.ts");
+const { handleChatCore } = await import("../../open-sse/handlers/chatCore.ts");
+
+const PROVIDER = "github";
+const MODEL = "claude-opus-4.5";
+const CONTEXT_OVERRIDE = 1_000_000;
+// ~3.7M chars ≈ 925k estimated tokens — above the catalog input hint, below the override.
+const BIG_PROMPT = "x".repeat(3_700_000);
+
+const originalFetch = globalThis.fetch;
+let fetchCalls = 0;
+const silentLog = { debug() {}, info() {}, warn() {}, error() {} };
+
+function buildRequest(isCombo: boolean, content: string) {
+  const body = {
+    model: MODEL,
+    messages: [{ role: "user", content }],
+    max_tokens: 10,
+    stream: false,
+  };
+  return {
+    body,
+    modelInfo: { provider: PROVIDER, model: MODEL, extendedContext: false },
+    credentials: {
+      apiKey: "sk-test",
+      providerSpecificData: { baseUrl: "https://combo-ctx-rescue.example.test" },
+    },
+    clientRawRequest: {
+      endpoint: "/v1/chat/completions",
+      body,
+      headers: new Headers({ accept: "application/json" }),
+    },
+    userAgent: "unit-test",
+    isCombo,
+    log: silentLog,
+  };
+}
+
+test.before(() => {
+  core.resetDbInstance();
+  globalThis.fetch = async () => {
+    fetchCalls += 1;
+    return new Response(
+      JSON.stringify({
+        id: "chatcmpl-combo-ctx",
+        choices: [
+          { index: 0, message: { role: "assistant", content: "ok" }, finish_reason: "stop" },
+        ],
+        usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 },
+      }),
+      { status: 200, headers: { "content-type": "application/json" } }
+    );
+  };
+});
+
+test.after(() => {
+  globalThis.fetch = originalFetch;
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+test.beforeEach(() => {
+  fetchCalls = 0;
+  // Isolate each scenario: clear any overrides a prior test wrote so ordering does
+  // not couple the assertions.
+  contextOverrides.removeModelContextOverride(PROVIDER, MODEL);
+  capabilityOverrides.removeModelCapabilityOverride(`${PROVIDER}/${MODEL}`, "max_input_tokens");
+});
+
+test("raw-alias combo request rescued by an exact context override dispatches", async () => {
+  assert.equal(contextOverrides.setModelContextOverride(PROVIDER, MODEL, CONTEXT_OVERRIDE), true);
+  // The prompt exceeds the catalog input hint but fits the context override. A combo
+  // request must NOT be rejected at the gate — the upstream fetch must fire.
+  const result = await handleChatCore(buildRequest(true, BIG_PROMPT));
+  assert.ok(fetchCalls > 0, "rescued combo target must reach the upstream fetch");
+  assert.notEqual(
+    result.status,
+    400,
+    "rescued combo target must not be rejected with the input-cap 400"
+  );
+});
+
+test("an exact raw-alias max_input_tokens override is enforced even in a combo", async () => {
+  // An explicit max_input_tokens capability override must NEVER be bypassed by a
+  // context override — combo included.
+  assert.equal(contextOverrides.setModelContextOverride(PROVIDER, MODEL, CONTEXT_OVERRIDE), true);
+  assert.equal(
+    capabilityOverrides.setModelCapabilityOverride(
+      `${PROVIDER}/${MODEL}`,
+      "max_input_tokens",
+      1000
+    ),
+    true
+  );
+  const result = await handleChatCore(buildRequest(true, BIG_PROMPT));
+  assert.equal(result.status, 400);
+  assert.equal(
+    fetchCalls,
+    0,
+    "explicit max_input override must reject with zero fetch even in combo"
+  );
+  assert.match(JSON.stringify(result), /maximum input tokens/i);
+});
diff --git a/tests/unit/chatcore-model-output-cap-wiring.test.ts b/tests/unit/chatcore-model-output-cap-wiring.test.ts
index f50caa3958..0331aeda67 100644
--- a/tests/unit/chatcore-model-output-cap-wiring.test.ts
+++ b/tests/unit/chatcore-model-output-cap-wiring.test.ts
@@ -1,46 +1,33 @@
-// Wiring guard for the model-output-cap clamp: output-token-budget-model-cap.test.ts
-// exercises enforceOutputTokenBudget() directly and therefore cannot catch a
-// callsite regression — drop the cap argument in handleChatCore and every one of
-// those unit tests still passes. This test drives handleChatCore() end to end
-// (stubbed fetch, temp DB) and asserts the body actually dispatched upstream.
-//
-// The cap is supplied through an operator `max_token` capability override
-// (src/lib/db/modelCapabilityOverrides.ts, issue #6524) rather than a catalog
-// model, which pins two things at once and keeps the test independent of
-// provider-catalog drift:
-//   1. the clamp runs on the single-model (non-combo) path;
-//   2. the cap lookup is keyed by { provider, model } — the bare-string form
-//      resolves to provider: null, and the override table is keyed by provider,
-//      so a string-keyed lookup silently misses it and no clamp happens.
+// Direct output-budget tests cannot prove handleChatCore passes capability caps
+// through its non-combo runtime path. This drives that public handler with a
+// temporary DB and upstream fetch stub.
 import test from "node:test";
 import assert from "node:assert/strict";
 import fs from "node:fs";
 import os from "node:os";
 import path from "node:path";
 
-const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-model-output-cap-"));
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-model-cap-wiring-"));
 process.env.DATA_DIR = TEST_DATA_DIR;
 
 const core = await import("../../src/lib/db/core.ts");
 const overridesDb = await import("../../src/lib/db/modelCapabilityOverrides.ts");
 const { handleChatCore } = await import("../../open-sse/handlers/chatCore.ts");
 
-// Distinctive enough that it can never collide with a provider registered in
-// open-sse/config/providerRegistry.ts, so nothing but the override supplies a cap.
 const PROVIDER = "capwire-testprov";
 const MODEL = "capwire-testmodel";
 const OUTPUT_CAP = 1000;
+const INPUT_CAP = 10;
 const REQUESTED_MAX_TOKENS = 50_000;
-
 const originalFetch = globalThis.fetch;
 let dispatchedBody: Record<string, unknown> | null = null;
-
+let fetchCalls = 0;
 const silentLog = { debug() {}, info() {}, warn() {}, error() {} };
 
-function buildRequest(maxTokens: number) {
+function buildRequest(maxTokens: number, content = "hello") {
   const body = {
     model: MODEL,
-    messages: [{ role: "user", content: "hello" }],
+    messages: [{ role: "user", content }],
     max_tokens: maxTokens,
     stream: false,
   };
@@ -65,12 +52,15 @@ function buildRequest(maxTokens: number) {
 test.before(() => {
   core.resetDbInstance();
   assert.equal(
-    overridesDb.setModelCapabilityOverride(`${PROVIDER}/${MODEL}`, "max_token", OUTPUT_CAP),
-    true,
-    "the operator override must be persisted for this test to mean anything"
+    overridesDb.setModelCapabilityOverride(`${PROVIDER}/${MODEL}`, "max_output_tokens", OUTPUT_CAP),
+    true
+  );
+  assert.equal(
+    overridesDb.setModelCapabilityOverride(`${PROVIDER}/${MODEL}`, "max_input_tokens", INPUT_CAP),
+    true
   );
-
   globalThis.fetch = async (_input: RequestInfo | URL, init?: RequestInit) => {
+    fetchCalls += 1;
     dispatchedBody = init?.body ? JSON.parse(String(init.body)) : null;
     return new Response(
       JSON.stringify({
@@ -93,25 +83,34 @@ test.after(() => {
 
 test("handleChatCore clamps an over-cap max_tokens to the model's output cap before dispatch", async () => {
   dispatchedBody = null;
+  fetchCalls = 0;
   await handleChatCore(buildRequest(REQUESTED_MAX_TOKENS));
-
-  assert.ok(dispatchedBody, "expected the request to reach the upstream fetch");
-  assert.equal(
-    dispatchedBody?.max_tokens,
-    OUTPUT_CAP,
-    `expected max_tokens clamped to the ${OUTPUT_CAP}-token operator cap, got ${dispatchedBody?.max_tokens}`
-  );
+  assert.equal(fetchCalls, 1);
+  assert.equal(dispatchedBody?.max_tokens, OUTPUT_CAP);
 });
 
 test("handleChatCore leaves a max_tokens below the cap untouched", async () => {
   dispatchedBody = null;
+  fetchCalls = 0;
   const underCap = OUTPUT_CAP - 1;
   await handleChatCore(buildRequest(underCap));
+  assert.equal(fetchCalls, 1);
+  assert.equal(dispatchedBody?.max_tokens, underCap);
+});
 
-  assert.ok(dispatchedBody, "expected the request to reach the upstream fetch");
-  assert.equal(
-    dispatchedBody?.max_tokens,
-    underCap,
-    "a request below the cap must never be raised to it"
-  );
+test("handleChatCore rejects input over the model input cap before upstream dispatch", async () => {
+  dispatchedBody = null;
+  fetchCalls = 0;
+  const result = await handleChatCore(buildRequest(1, "x".repeat(200)));
+  assert.equal(result.status, 400);
+  assert.equal(fetchCalls, 0, "input-cap rejection must not call the upstream");
+  assert.match(JSON.stringify(result), /maximum input tokens/i);
+});
+
+test("handleChatCore dispatches input within the model input cap", async () => {
+  dispatchedBody = null;
+  fetchCalls = 0;
+  await handleChatCore(buildRequest(1, "x"));
+  assert.equal(fetchCalls, 1);
+  assert.ok(dispatchedBody, "input below the cap must reach the upstream");
 });
diff --git a/tests/unit/context-window-reconcile-persisted-overrides.test.ts b/tests/unit/context-window-reconcile-persisted-overrides.test.ts
new file mode 100644
index 0000000000..13285d398f
--- /dev/null
+++ b/tests/unit/context-window-reconcile-persisted-overrides.test.ts
@@ -0,0 +1,55 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(
+  path.join(os.tmpdir(), "omniroute-context-reconcile-runtime-")
+);
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const models = await import("../../src/lib/db/models.ts");
+const overrides = await import("../../src/lib/db/modelContextOverrides.ts");
+const { runContextWindowReconcile } = await import("../../src/lib/contextWindowResolver.ts");
+
+test.beforeEach(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
+});
+
+test.after(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+test("runContextWindowReconcile retains an auto override across repeated synced discovery", async () => {
+  // gpt-4o's static catalog window is 128K; discovery reports a real 372K.
+  // This uses the live DB discovery and resolver seams, not injected pure deps.
+  await models.replaceSyncedAvailableModelsForConnection("openai", "reconcile-test", [
+    { id: "gpt-4o", inputTokenLimit: 372000 },
+  ]);
+
+  const first = await runContextWindowReconcile();
+  assert.deepEqual(first, { scanned: 1, written: 1, removed: 0, skippedManual: 0 });
+  assert.deepEqual(
+    (() => {
+      const record = overrides.getModelContextOverrideRecord("openai", "gpt-4o");
+      return record && { realContext: record.realContext, source: record.source };
+    })(),
+    { realContext: 372000, source: "auto:discovery" }
+  );
+
+  const second = await runContextWindowReconcile();
+  assert.deepEqual(second, { scanned: 1, written: 1, removed: 0, skippedManual: 0 });
+  assert.deepEqual(
+    (() => {
+      const record = overrides.getModelContextOverrideRecord("openai", "gpt-4o");
+      return record && { realContext: record.realContext, source: record.source };
+    })(),
+    { realContext: 372000, source: "auto:discovery" },
+    "the persisted auto override remains rather than being removed after its first write"
+  );
+});
diff --git a/tests/unit/model-capability-overrides.test.ts b/tests/unit/model-capability-overrides.test.ts
index 2857327e8f..788b48c9cf 100644
--- a/tests/unit/model-capability-overrides.test.ts
+++ b/tests/unit/model-capability-overrides.test.ts
@@ -10,6 +10,8 @@ process.env.DATA_DIR = moduleDataDir;
 const coreDb = await import("../../src/lib/db/core.ts");
 const caps = await import("../../src/lib/modelCapabilities.ts");
 const overrides = await import("../../src/lib/db/modelCapabilityOverrides.ts");
+const contextOverrides = await import("../../src/lib/db/modelContextOverrides.ts");
+const route = await import("../../src/app/api/model-capability-overrides/route.ts");
 
 beforeEach(() => {
   coreDb.resetDbInstance();
@@ -23,8 +25,18 @@ after(() => {
   fs.rmSync(moduleDataDir, { recursive: true, force: true });
 });
 
+function patchOverride(key: string, value: unknown) {
+  return route.PATCH(
+    new Request("http://localhost/api/model-capability-overrides", {
+      method: "PATCH",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({ target: "codex/gpt-5.6", key, value }),
+    })
+  );
+}
+
 describe("model capability overrides", () => {
-  it("stores, lists, removes, and applies a provider/model max_token override", () => {
+  it("stores, lists, removes, and applies an exact max_output_tokens override", () => {
     const withoutOverride = caps.getResolvedModelCapabilities({
       provider: "openai",
       model: "gpt-4o",
@@ -32,7 +44,7 @@ describe("model capability overrides", () => {
     const distinct = (withoutOverride ?? 0) + 12345;
 
     assert.equal(
-      overrides.setModelCapabilityOverride("openai/gpt-4o", "max_token", distinct),
+      overrides.setModelCapabilityOverride("openai/gpt-4o", "max_output_tokens", distinct),
       true
     );
     assert.deepEqual(
@@ -41,32 +53,56 @@ describe("model capability overrides", () => {
         key: entry.key,
         value: entry.value,
       })),
-      [{ target: "openai/gpt-4o", key: "max_token", value: distinct }]
+      [{ target: "openai/gpt-4o", key: "max_output_tokens", value: distinct }]
     );
-
     assert.equal(
       caps.getResolvedModelCapabilities({ provider: "openai", model: "gpt-4o" }).maxOutputTokens,
       distinct
     );
     assert.notEqual(
       caps.getResolvedModelCapabilities({ provider: "anthropic", model: "gpt-4o" }).maxOutputTokens,
-      distinct,
-      "override must be scoped by provider/model, not bare model id"
+      distinct
+    );
+    assert.equal(
+      overrides.removeModelCapabilityOverride("openai/gpt-4o", "max_output_tokens"),
+      true
     );
-
-    assert.equal(overrides.removeModelCapabilityOverride("openai/gpt-4o", "max_token"), true);
     assert.equal(
       caps.getResolvedModelCapabilities({ provider: "openai", model: "gpt-4o" }).maxOutputTokens,
       withoutOverride
     );
   });
 
+  it("uses exact input/output overrides, clamps input to context, and isolates effort variants", () => {
+    const target = "codex/gpt-5.6";
+    const variant = "codex/gpt-5.6-high";
+    assert.equal(contextOverrides.setModelContextOverride("codex", "gpt-5.6", 372000), true);
+    assert.equal(overrides.setModelCapabilityOverride(target, "max_input_tokens", 999999), true);
+    assert.equal(overrides.setModelCapabilityOverride(target, "max_output_tokens", 123456), true);
+
+    const base = caps.getResolvedModelCapabilities(target);
+    assert.deepEqual(
+      {
+        contextWindow: base.contextWindow,
+        maxInputTokens: base.maxInputTokens,
+        maxOutputTokens: base.maxOutputTokens,
+      },
+      { contextWindow: 372000, maxInputTokens: 372000, maxOutputTokens: 123456 }
+    );
+    assert.equal(overrides.removeModelCapabilityOverride(target, "max_output_tokens"), true);
+    assert.notEqual(caps.getResolvedModelCapabilities(target).maxOutputTokens, 123456);
+
+    const effort = caps.getResolvedModelCapabilities(variant);
+    assert.notEqual(effort.contextWindow, 372000);
+    assert.notEqual(effort.maxInputTokens, 372000);
+    assert.notEqual(effort.maxOutputTokens, 123456);
+  });
+
   it("applies overrides stored under provider-scoped model aliases", () => {
     assert.equal(
-      overrides.setModelCapabilityOverride("github/claude-opus-4.5", "max_token", 77777),
+      overrides.setModelCapabilityOverride("github/claude-opus-4.5", "max_output_tokens", 77777),
       true
     );
-
     assert.equal(
       caps.getResolvedModelCapabilities({ provider: "github", model: "claude-opus-4.5" })
         .maxOutputTokens,
@@ -74,10 +110,143 @@ describe("model capability overrides", () => {
     );
   });
 
+  it("accepts exactly the three public token-limit keys through the API", async () => {
+    assert.equal(
+      contextOverrides.setModelContextOverride("codex", "gpt-5.6", 272000, "auto:discovery"),
+      true
+    );
+    const discoveredResponse = await route.GET(
+      new Request("http://localhost/api/model-capability-overrides")
+    );
+    const discoveredPayload = (await discoveredResponse.json()) as {
+      overrides: Array<{ target: string; key: string; value: number }>;
+    };
+    assert.ok(
+      discoveredPayload.overrides.some(
+        (override) =>
+          override.target === "codex/gpt-5.6" &&
+          override.key === "context_length" &&
+          override.value === 272000
+      ),
+      "the unified surface exposes an auto-discovered context window before manual replacement"
+    );
+
+    assert.equal((await patchOverride("context_length", 372000)).status, 200);
+    assert.equal((await patchOverride("max_input_tokens", 353400)).status, 200);
+    assert.equal((await patchOverride("max_output_tokens", 128000)).status, 200);
+    assert.equal((await patchOverride("max_token", 77777)).status, 400, "legacy key");
+    assert.equal((await patchOverride("unknown", 1)).status, 400, "unsupported key");
+    assert.equal((await patchOverride("max_input_tokens", 0)).status, 400, "non-positive integer");
+    assert.equal(
+      (await patchOverride("max_input_tokens", Number.POSITIVE_INFINITY)).status,
+      400,
+      "JSON serializes Infinity as null; route rejects the resulting non-number"
+    );
+
+    const response = await route.GET(
+      new Request("http://localhost/api/model-capability-overrides")
+    );
+    const payload = (await response.json()) as {
+      overrides: Array<{ target: string; key: string; value: number }>;
+    };
+    assert.deepEqual(
+      payload.overrides
+        .map(({ target, key, value }) => ({ target, key, value }))
+        .sort((left, right) => left.key.localeCompare(right.key)),
+      [
+        { target: "codex/gpt-5.6", key: "context_length", value: 372000 },
+        { target: "codex/gpt-5.6", key: "max_input_tokens", value: 353400 },
+        { target: "codex/gpt-5.6", key: "max_output_tokens", value: 128000 },
+      ]
+    );
+
+    assert.equal(
+      contextOverrides.getModelContextOverrideRecord("codex", "gpt-5.6")?.source,
+      "manual"
+    );
+
+    const resolved = caps.getResolvedModelCapabilities("codex/gpt-5.6");
+    assert.deepEqual(
+      {
+        contextWindow: resolved.contextWindow,
+        maxInputTokens: resolved.maxInputTokens,
+        maxOutputTokens: resolved.maxOutputTokens,
+      },
+      { contextWindow: 372000, maxInputTokens: 353400, maxOutputTokens: 128000 }
+    );
+
+    const removed = await route.DELETE(
+      new Request(
+        "http://localhost/api/model-capability-overrides?target=codex/gpt-5.6&key=context_length",
+        { method: "DELETE" }
+      )
+    );
+    assert.equal(removed.status, 200);
+    assert.equal(contextOverrides.getModelContextOverride("codex", "gpt-5.6"), null);
+
+    const rejectedDelete = await route.DELETE(
+      new Request(
+        "http://localhost/api/model-capability-overrides?target=codex/gpt-5.6&key=max_token",
+        { method: "DELETE" }
+      )
+    );
+    assert.equal(rejectedDelete.status, 400);
+  });
+
   it("rejects invalid targets and non-positive values", () => {
-    assert.equal(overrides.setModelCapabilityOverride("gpt-4o", "max_token", 1000), false);
-    assert.equal(overrides.setModelCapabilityOverride("openai/gpt-4o", "max_token", 0), false);
-    assert.equal(overrides.setModelCapabilityOverride("openai/gpt-4o", "max_token", 1.5), false);
+    assert.equal(overrides.setModelCapabilityOverride("gpt-4o", "max_output_tokens", 1000), false);
+    assert.equal(
+      overrides.setModelCapabilityOverride("openai/gpt-4o", "max_output_tokens", 0),
+      false
+    );
+    assert.equal(
+      overrides.setModelCapabilityOverride("openai/gpt-4o", "max_output_tokens", 1.5),
+      false
+    );
     assert.deepEqual(overrides.listModelCapabilityOverrides(), []);
   });
+
+  it("migrates legacy max_token rows without replacing a modern output override", () => {
+    const db = coreDb.getDbInstance();
+    const insert = db.prepare(
+      "INSERT INTO model_capability_overrides " +
+        "(provider, model_id, override_key, override_value, refreshed_at) VALUES (?, ?, ?, ?, ?)"
+    );
+    insert.run("legacy", "legacy-only", "max_token", "64000", "2026-01-01 00:00:00");
+    insert.run("legacy", "collision", "max_token", "64000", "2026-01-01 00:00:00");
+    insert.run("legacy", "collision", "max_output_tokens", "128000", "2026-02-01 00:00:00");
+
+    const migration = fs.readFileSync(
+      path.resolve("src/lib/db/migrations/135_migrate_model_capability_max_token.sql"),
+      "utf8"
+    );
+    db.exec(migration);
+    db.exec(migration);
+
+    assert.deepEqual(
+      db
+        .prepare(
+          "SELECT provider, model_id, override_key, override_value, refreshed_at " +
+            "FROM model_capability_overrides WHERE provider = 'legacy' " +
+            "ORDER BY model_id, override_key"
+        )
+        .all(),
+      [
+        {
+          provider: "legacy",
+          model_id: "collision",
+          override_key: "max_output_tokens",
+          override_value: "128000",
+          refreshed_at: "2026-02-01 00:00:00",
+        },
+        {
+          provider: "legacy",
+          model_id: "legacy-only",
+          override_key: "max_output_tokens",
+          override_value: "64000",
+          refreshed_at: "2026-01-01 00:00:00",
+        },
+      ]
+    );
+  });
 });
diff --git a/tests/unit/model-context-override-readpath.test.ts b/tests/unit/model-context-override-readpath.test.ts
index e9aef36538..9cd6a6ea3e 100644
--- a/tests/unit/model-context-override-readpath.test.ts
+++ b/tests/unit/model-context-override-readpath.test.ts
@@ -28,8 +28,10 @@ after(() => {
 describe("getModelContextLimit override precedence (5004)", () => {
   it("an override wins over the catalog, and removing it falls back to the catalog", () => {
     // Read the override-free catalog value dynamically (non-brittle for any model).
-    const catalog = caps.getResolvedModelCapabilities({ provider: "openai", model: "gpt-4o" })
-      .contextWindow;
+    const catalog = caps.getResolvedModelCapabilities({
+      provider: "openai",
+      model: "gpt-4o",
+    }).contextWindow;
     const distinct = (catalog ?? 0) + 12345;
 
     mco.setModelContextOverride("openai", "gpt-4o", distinct);
@@ -49,11 +51,61 @@ describe("getModelContextLimit override precedence (5004)", () => {
     assert.equal(caps.getModelContextLimit("custom-local", "my-7b-128k"), 131072);
   });
 
-  it("leaves getResolvedModelCapabilities override-free (so the reconciler sees the catalog)", () => {
+  it("resolves exact raw aliases after canonical rows, without effort inheritance", () => {
+    const provider = "github";
+    const rawAlias = "claude-opus-4.5";
+    const canonical = "claude-opus-4-5-20251101";
+
+    assert.equal(mco.setModelContextOverride(provider, rawAlias, 333333), true);
+    assert.equal(
+      caps.getResolvedModelCapabilities({ provider, model: rawAlias }).contextWindow,
+      333333,
+      "an exact raw alias override must be effective"
+    );
+    assert.notEqual(
+      caps.getResolvedModelCapabilities(
+        { provider, model: rawAlias },
+        { persistedOverrides: false }
+      ).contextWindow,
+      333333,
+      "override-free resolution must not read the raw alias row"
+    );
+
+    assert.equal(mco.setModelContextOverride(provider, canonical, 444444), true);
+    assert.equal(
+      caps.getResolvedModelCapabilities({ provider, model: rawAlias }).contextWindow,
+      444444,
+      "the canonical row must win over the exact raw alias row"
+    );
+    assert.notEqual(
+      caps.getResolvedModelCapabilities({ provider, model: `${rawAlias}-high` }).contextWindow,
+      444444,
+      "an exact alias override must not inherit to an effort variant"
+    );
+  });
+
+  it("default getResolvedModelCapabilities reflects the override; persistedOverrides:false returns the catalog", () => {
     mco.setModelContextOverride("openai", "gpt-4o", 999999, "auto:discovery");
-    const catalog = caps.getResolvedModelCapabilities({ provider: "openai", model: "gpt-4o" })
-      .contextWindow;
-    assert.notEqual(catalog, 999999, "getResolvedModelCapabilities must not reflect the override");
-    assert.equal(caps.getModelContextLimit("openai", "gpt-4o"), 999999, "but getModelContextLimit does");
+    // Default resolution is the effective runtime view: the persisted context
+    // override wins over the catalog.
+    const effective = caps.getResolvedModelCapabilities({
+      provider: "openai",
+      model: "gpt-4o",
+    }).contextWindow;
+    assert.equal(effective, 999999, "default resolution must reflect the persisted override");
+
+    // The override-free catalog view (used by the reconciler) excludes the override.
+    const catalog = caps.getResolvedModelCapabilities(
+      { provider: "openai", model: "gpt-4o" },
+      { persistedOverrides: false }
+    ).contextWindow;
+    assert.notEqual(catalog, 999999, "persistedOverrides:false must return the catalog value");
+
+    // getModelContextLimit always follows the effective override.
+    assert.equal(
+      caps.getModelContextLimit("openai", "gpt-4o"),
+      999999,
+      "getModelContextLimit reflects the override"
+    );
   });
 });
diff --git a/tests/unit/model-token-limit-catalog.test.ts b/tests/unit/model-token-limit-catalog.test.ts
new file mode 100644
index 0000000000..b92649a01d
--- /dev/null
+++ b/tests/unit/model-token-limit-catalog.test.ts
@@ -0,0 +1,164 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-token-limit-catalog-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const contextOverrides = await import("../../src/lib/db/modelContextOverrides.ts");
+const capabilityOverrides = await import("../../src/lib/db/modelCapabilityOverrides.ts");
+const models = await import("../../src/lib/db/models.ts");
+const providers = await import("../../src/lib/db/providers.ts");
+const catalog = await import("../../src/app/api/v1/models/catalog.ts");
+
+const TARGET = "openai/gpt-5.6";
+const LIMITS = { context: 372000, input: 353400, output: 128000 };
+
+test.beforeEach(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
+  catalog.__resetCatalogBuilderRunsForTest();
+});
+
+test.after(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+async function getModel(target = TARGET) {
+  const response = await catalog.getUnifiedModelsResponse(
+    new Request("http://localhost/api/v1/models")
+  );
+  const body = (await response.json()) as { data: Array<Record<string, unknown>> };
+  return body.data.find((model) => model.id === target);
+}
+
+test("v1 model catalog projects effective context, input, and output overrides and invalidates on delete", async () => {
+  assert.equal(contextOverrides.setModelContextOverride("openai", "gpt-5.6", LIMITS.context), true);
+  assert.equal(
+    capabilityOverrides.setModelCapabilityOverride(TARGET, "max_input_tokens", LIMITS.input),
+    true
+  );
+  assert.equal(
+    capabilityOverrides.setModelCapabilityOverride(TARGET, "max_output_tokens", LIMITS.output),
+    true
+  );
+  await providers.createProviderConnection({
+    provider: "openai",
+    authType: "api_key",
+    name: "token-limit-catalog",
+    apiKey: "sk-test",
+  });
+
+  const initial = await getModel();
+  assert.ok(initial);
+  assert.deepEqual(
+    {
+      context_length: initial.context_length,
+      max_input_tokens: initial.max_input_tokens,
+      max_output_tokens: initial.max_output_tokens,
+    },
+    {
+      context_length: LIMITS.context,
+      max_input_tokens: LIMITS.input,
+      max_output_tokens: LIMITS.output,
+    }
+  );
+
+  const sentinelOutput = 111111;
+  assert.equal(
+    capabilityOverrides.setModelCapabilityOverride(TARGET, "max_output_tokens", sentinelOutput),
+    true
+  );
+  assert.equal((await getModel())?.max_output_tokens, sentinelOutput);
+  assert.equal(
+    capabilityOverrides.removeModelCapabilityOverride(TARGET, "max_output_tokens"),
+    true
+  );
+  assert.notEqual((await getModel())?.max_output_tokens, sentinelOutput);
+});
+
+test("v1 model catalog projects a synced Codex context to both public aliases", async () => {
+  const target = "codex/gpt-5.6-sol";
+  const connection = await providers.createProviderConnection({
+    provider: "codex",
+    authType: "oauth",
+    name: "codex-token-limit-catalog",
+    accessToken: "test-access-token",
+  });
+  assert.equal(typeof connection.id, "string");
+  await models.replaceSyncedAvailableModelsForConnection("codex", connection.id as string, [
+    {
+      id: "gpt-5.6-sol",
+      name: "GPT 5.6 Sol",
+      source: "provider",
+      supportedEndpoints: ["responses"],
+      inputTokenLimit: 272000,
+      outputTokenLimit: 128000,
+    },
+  ]);
+  assert.equal(
+    contextOverrides.setModelContextOverride("codex", "gpt-5.6-sol", LIMITS.context),
+    true
+  );
+  assert.equal(
+    capabilityOverrides.setModelCapabilityOverride(target, "max_input_tokens", LIMITS.input),
+    true
+  );
+  assert.equal(
+    capabilityOverrides.setModelCapabilityOverride(target, "max_output_tokens", LIMITS.output),
+    true
+  );
+
+  for (const publicId of ["cx/gpt-5.6-sol", "codex/gpt-5.6-sol"]) {
+    const projected = await getModel(publicId);
+    assert.ok(projected, `expected ${publicId} in the public catalog`);
+    assert.deepEqual(
+      {
+        context_length: projected.context_length,
+        max_input_tokens: projected.max_input_tokens,
+        max_output_tokens: projected.max_output_tokens,
+      },
+      {
+        context_length: LIMITS.context,
+        max_input_tokens: LIMITS.input,
+        max_output_tokens: LIMITS.output,
+      },
+      `${publicId} must retain the persisted Codex token-limit overrides`
+    );
+  }
+
+  const canonical = await getModel("codex/gpt-5.6-sol");
+  assert.equal(canonical?.type, "image");
+  assert.deepEqual(canonical?.output_modalities, ["image"]);
+  assert.ok(Array.isArray(canonical?.supported_sizes));
+
+  assert.equal(contextOverrides.removeModelContextOverride("codex", "gpt-5.6-sol"), true);
+  assert.equal(
+    (await getModel("codex/gpt-5.6-sol"))?.context_length,
+    undefined,
+    "the specialty row must not inherit the synced chat context after override removal"
+  );
+  assert.equal((await getModel("cx/gpt-5.6-sol"))?.context_length, 272000);
+});
+
+test("v1 model catalog projects an exact raw-alias context override", async () => {
+  const target = "github/claude-opus-4.5";
+  assert.equal(contextOverrides.setModelContextOverride("github", "claude-opus-4.5", 333333), true);
+  await providers.createProviderConnection({
+    provider: "github",
+    authType: "api_key",
+    name: "raw-alias-token-limit-catalog",
+    apiKey: "ghp-test",
+  });
+
+  assert.equal(
+    (await getModel(target))?.context_length,
+    333333,
+    "the catalog entry keeps its raw alias and must project that exact override"
+  );
+});
diff --git a/tests/unit/output-token-budget-model-cap.test.ts b/tests/unit/output-token-budget-model-cap.test.ts
index dd7dc42b17..4015f90189 100644
--- a/tests/unit/output-token-budget-model-cap.test.ts
+++ b/tests/unit/output-token-budget-model-cap.test.ts
@@ -118,3 +118,73 @@ test("a sub-token cap is treated as absent, never as a cap of zero", () => {
   assert.equal(result.body.max_tokens, 8_000, "sub-token cap must leave the request untouched");
   assert.deepEqual(result.adjustedFields, []);
 });
+
+test("rejects when estimated input exceeds the model input cap, even with output room", () => {
+  // Input cap is input-only and independent of the (larger) total window.
+  const result = enforceOutputTokenBudget(
+    { max_tokens: 1_000 },
+    354_000,
+    372_000,
+    0,
+    null,
+    353_400
+  );
+
+  assert.equal(result.ok, false);
+  if (result.ok) return;
+  assert.equal(result.estimatedInputTokens, 354_000);
+  assert.equal(result.maxInputTokens, 353_400);
+});
+
+test("accepts input at the model input cap without double-counting requested output", () => {
+  // estimatedInput == input cap; a large requested output must not be re-counted
+  // against the input-only cap (only the total window bounds input + output).
+  const result = enforceOutputTokenBudget(
+    { max_tokens: 128_000 },
+    353_400,
+    372_000,
+    0,
+    null,
+    353_400
+  );
+
+  assert.equal(result.ok, true);
+  if (!result.ok) return;
+  // Output clamped to remaining window room (372000 - 353400 = 18600).
+  assert.equal(result.body.max_tokens, 18_600);
+});
+
+test("input cap is fail-open when absent, null, or non-positive", () => {
+  const base = enforceOutputTokenBudget({ max_tokens: 1_000 }, 100_000, 200_000, 0, null);
+  const undefinedCap = enforceOutputTokenBudget(
+    { max_tokens: 1_000 },
+    100_000,
+    200_000,
+    0,
+    null,
+    undefined
+  );
+  const nullCap = enforceOutputTokenBudget({ max_tokens: 1_000 }, 100_000, 200_000, 0, null, null);
+  const zeroCap = enforceOutputTokenBudget({ max_tokens: 1_000 }, 100_000, 200_000, 0, null, 0);
+
+  assert.deepEqual(undefinedCap, base);
+  assert.deepEqual(nullCap, base);
+  assert.deepEqual(zeroCap, base);
+  assert.equal(base.ok, true);
+});
+
+test("the total window still rejects when input fits the input cap but no output room remains", () => {
+  // Input (150k) fits a 353k input cap but leaves no room in the 128k window.
+  const result = enforceOutputTokenBudget(
+    { max_tokens: 1_000 },
+    150_000,
+    128_000,
+    0,
+    null,
+    353_400
+  );
+
+  assert.equal(result.ok, false);
+  if (result.ok) return;
+  assert.equal(result.maxInputTokens, undefined, "window rejection, not input-cap rejection");
+});
diff --git a/tests/unit/repro-6524.test.ts b/tests/unit/repro-6524.test.ts
index 1d4ac3a626..d99dcec855 100644
--- a/tests/unit/repro-6524.test.ts
+++ b/tests/unit/repro-6524.test.ts
@@ -11,14 +11,14 @@
  * `getExplicitModelOutputCap` clamp source): the clamp math itself is correct, but
  * `getExplicitModelOutputCap()` only ever read the unvalidated synced
  * `limit_output` (or registry/static fallbacks) — it ignored the operator-settable
- * `max_token` capability override (`src/lib/db/modelCapabilityOverrides.ts`,
+ * `max_output_tokens` capability override (`src/lib/db/modelCapabilityOverrides.ts`,
  * `/api/model-capability-overrides`) that `getResolvedModelCapabilities()` already
  * consulted. That inconsistency meant an operator manually correcting a bad synced
  * output cap (the existing, already-shipped remediation path for wrong catalog
  * data) had no effect on the reasoning buffer, which kept inflating past the real
  * cap regardless.
  *
- * Fix: `getExplicitModelOutputCap()` now checks the same `max_token` override
+ * Fix: `getExplicitModelOutputCap()` now checks the same `max_output_tokens` override
  * before falling back to synced/registry/static data, via a helper shared with
  * `getResolvedModelCapabilities()`.
  */
@@ -32,15 +32,12 @@ const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-repro-652
 process.env.DATA_DIR = TEST_DATA_DIR;
 
 const core = await import("../../src/lib/db/core.ts");
-const { saveModelsDevCapabilities, clearModelsDevCapabilities } = await import(
-  "../../src/lib/modelsDevSync.ts"
-);
-const { setModelCapabilityOverride, removeModelCapabilityOverride } = await import(
-  "../../src/lib/db/modelCapabilityOverrides.ts"
-);
-const { resolveReasoningBufferedMaxTokens } = await import(
-  "../../open-sse/services/reasoningTokenBuffer.ts"
-);
+const { saveModelsDevCapabilities, clearModelsDevCapabilities } =
+  await import("../../src/lib/modelsDevSync.ts");
+const { setModelCapabilityOverride, removeModelCapabilityOverride } =
+  await import("../../src/lib/db/modelCapabilityOverrides.ts");
+const { resolveReasoningBufferedMaxTokens } =
+  await import("../../open-sse/services/reasoningTokenBuffer.ts");
 
 const PROVIDER = "ollama-cloud";
 const MODEL = "deepseek-v4-flash";
@@ -87,10 +84,10 @@ test("#6524: with only the (wrong) synced catalog data, the buffer no longer inf
   assert.equal(result, 64000);
 });
 
-test("#6524: an operator-set max_token override now clamps the reasoning buffer to the real cap", () => {
+test("#6524: an operator-set max_output_tokens override clamps the reasoning buffer", () => {
   assert.ok(
-    setModelCapabilityOverride(TARGET, "max_token", REAL_UPSTREAM_OUTPUT_CAP),
-    "expected the max_token override to be written"
+    setModelCapabilityOverride(TARGET, "max_output_tokens", REAL_UPSTREAM_OUTPUT_CAP),
+    "expected the max_output_tokens override to be written"
   );
   try {
     const result = resolveReasoningBufferedMaxTokens(TARGET, 64000);
@@ -100,6 +97,6 @@ test("#6524: an operator-set max_token override now clamps the reasoning buffer
         `(reproduces reported 64000 -> 96000 inflation, upstream then 400s)`
     );
   } finally {
-    removeModelCapabilityOverride(TARGET, "max_token");
+    removeModelCapabilityOverride(TARGET, "max_output_tokens");
   }
 });
diff --git a/tests/unit/v1-models-catalog-generation-race.test.ts b/tests/unit/v1-models-catalog-generation-race.test.ts
new file mode 100644
index 0000000000..1c1e519164
--- /dev/null
+++ b/tests/unit/v1-models-catalog-generation-race.test.ts
@@ -0,0 +1,153 @@
+// Regression guard — catalog in-flight build survived invalidation and repopulated
+// the cache with pre-write state (generation race).
+//
+// catalogCache coalesces concurrent GET /v1/models onto one in-flight builder and
+// drops completed entries when invalidateDbCache() bumps modelCatalogCacheVersion.
+// But the in-flight map was keyed only by request shape: a build that STARTED before
+// a write could still be JOINED by a request that arrived after it, and when the
+// stale build finished it called storePayload unconditionally — repopulating the
+// now-current cache with a body built from pre-write DB state.
+//
+// Fix: every in-flight build is bound to the catalog-state generation it started
+// from. New requests only join a current-generation build, and storePayload only
+// persists a payload whose build generation still equals the live version. A stale
+// build still resolves to its own original caller (that request is legitimately
+// waiting on it) but never overwrites the fresh cache.
+//
+// These tests drive resolveCachedCatalogResponse directly with a deferred builder so
+// the interleaving is fully deterministic — no sleeps, no real-time races.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-catalog-genrace-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+process.env.API_KEY_SECRET = process.env.API_KEY_SECRET || "catalog-genrace-test-secret";
+
+const core = await import("../../src/lib/db/core.ts");
+const apiKeysDb = await import("../../src/lib/db/apiKeys.ts");
+const readCache = await import("../../src/lib/db/readCache.ts");
+const catalogCache = await import("../../src/app/api/v1/models/catalogCache.ts");
+
+const HEADER_SOURCES = { corsHeaders: {}, diagnosticHeaders: {} };
+
+function makeRequest() {
+  return new Request("http://localhost/v1/models");
+}
+
+function makePayload(body: string): catalogCache.CatalogPayload {
+  return { body, headers: { "content-type": "application/json" }, status: 200, cacheTTL: 60_000 };
+}
+
+/** A builder whose single in-flight promise resolves only when release() is called. */
+function deferredBuilder(body: string) {
+  let release!: (payload: catalogCache.CatalogPayload) => void;
+  let calls = 0;
+  const gate = new Promise<catalogCache.CatalogPayload>((resolve) => {
+    release = resolve;
+  });
+  const builder = () => {
+    calls += 1;
+    return gate;
+  };
+  return {
+    builder,
+    release: () => release(makePayload(body)),
+    calls: () => calls,
+  };
+}
+
+test.beforeEach(() => {
+  core.resetDbInstance();
+  apiKeysDb.resetApiKeyState();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
+  catalogCache.__resetCatalogBuilderRunsForTest();
+});
+
+test.after(() => {
+  core.resetDbInstance();
+  apiKeysDb.resetApiKeyState();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+test("a build that started before invalidation is not joined and does not repopulate the cache", async () => {
+  // 1. Start the OLD build and hold it in flight.
+  const oldBuild = deferredBuilder("OLD");
+  const oldResponsePromise = catalogCache.resolveCachedCatalogResponse(
+    makeRequest(),
+    HEADER_SOURCES,
+    oldBuild.builder
+  );
+  // Let the cold path register its in-flight entry (it awaits inflight.promise).
+  await Promise.resolve();
+  assert.equal(oldBuild.calls(), 1, "first request starts the old build");
+
+  // 2. Invalidate the catalog state (simulates a settings/connections/combos write).
+  readCache.invalidateDbCache("model-capabilities");
+
+  // 3. A new request after the write must NOT join the stale old build — it starts a
+  //    FRESH build at the new generation.
+  const newBuild = deferredBuilder("NEW");
+  const newResponsePromise = catalogCache.resolveCachedCatalogResponse(
+    makeRequest(),
+    HEADER_SOURCES,
+    newBuild.builder
+  );
+  await Promise.resolve();
+  assert.equal(
+    newBuild.calls(),
+    1,
+    "post-invalidation request must run a fresh builder, not join the stale old build"
+  );
+
+  // 4. Release the NEW build first — it is current-generation and must populate the cache.
+  newBuild.release();
+  const newResponse = await newResponsePromise;
+  assert.equal(await newResponse.text(), "NEW");
+
+  // 5. Now release the OLD (stale) build. It still resolves to its original caller,
+  //    but must NOT overwrite the fresh cache entry.
+  oldBuild.release();
+  const oldResponse = await oldResponsePromise;
+  assert.equal(
+    await oldResponse.text(),
+    "OLD",
+    "stale build still returns its own payload to its original caller"
+  );
+
+  // 6. The next request must be served from cache with the FRESH body — proving the
+  //    stale old build never repopulated the cache. A cache hit means no builder runs.
+  const freshBuild = deferredBuilder("SHOULD-NOT-RUN");
+  const cachedResponse = await catalogCache.resolveCachedCatalogResponse(
+    makeRequest(),
+    HEADER_SOURCES,
+    freshBuild.builder
+  );
+  assert.equal(await cachedResponse.text(), "NEW", "cache must hold the fresh payload");
+  assert.equal(freshBuild.calls(), 0, "a cache hit must not run the builder again");
+});
+
+test("concurrent requests in the SAME generation still coalesce onto one build", async () => {
+  const build = deferredBuilder("SHARED");
+  const p1 = catalogCache.resolveCachedCatalogResponse(
+    makeRequest(),
+    HEADER_SOURCES,
+    build.builder
+  );
+  const p2 = catalogCache.resolveCachedCatalogResponse(
+    makeRequest(),
+    HEADER_SOURCES,
+    build.builder
+  );
+  await Promise.resolve();
+  assert.equal(build.calls(), 1, "same-generation concurrent requests share one build");
+
+  build.release();
+  const [r1, r2] = await Promise.all([p1, p2]);
+  assert.equal(await r1.text(), "SHARED");
+  assert.equal(await r2.text(), "SHARED");
+});