fix(grok-cli): stop discovery from pinning Grok Build's 256k context window

changelog.d/fixes/15412-grok-cli-discovery-context-window.md

@@ -0,0 +1 @@
+- **fix(grok-cli):** Grok Build discovery prefers the grok-cli registry context window over the 256k that Grok Build `/v1/models` reports for every model, and never stamps a default; the reconciler drops stale `auto:discovery` pins when discovery reports no window, so `grok-4.5`/`4.6`/`4.7` advertise their real 500k window ([#15412](https://github.com/diegosouzapw/OmniRoute/pull/15412))

open-sse/config/grokBuild.ts

@@ -9,7 +9,6 @@ export const GROK_BUILD_DEVICE_CODE_URL = `${GROK_BUILD_OAUTH_ISSUER}/oauth2/dev
 export const GROK_BUILD_TOKEN_URL = `${GROK_BUILD_OAUTH_ISSUER}/oauth2/token`;
 
 export const GROK_BUILD_DEFAULT_CLIENT_VERSION = "1.0.41";
-export const GROK_BUILD_DEFAULT_CONTEXT_WINDOW = 256_000;
 export const GROK_BUILD_DEFAULT_REASONING_EFFORT = "high";
 export const GROK_BUILD_SUPPORTED_REASONING_EFFORTS = Object.freeze([
   "low",

src/app/api/providers/[id]/models/discovery/providerModelsConfig.ts

@@ -1,10 +1,10 @@
 import { getAntigravityModelsDiscoveryUrls } from "@omniroute/open-sse/config/antigravityUpstream.ts";
 import {
-  GROK_BUILD_DEFAULT_CONTEXT_WINDOW,
   getGrokBuildModelsHeaders,
   GROK_BUILD_MODELS_URL,
   GROK_BUILD_SUPPORTED_REASONING_EFFORTS,
 } from "@omniroute/open-sse/config/grokBuild.ts";
+import { grok_cliProvider } from "@omniroute/open-sse/config/providers/registry/grok-cli/index.ts";
 import { getAntigravityContentHeaders } from "@omniroute/open-sse/services/antigravityHeaders.ts";
 import { parseGeminiModelsList } from "@/lib/providerModels/geminiModelsParser";
 import { buildClaudeModelsHeaders } from "@/lib/providerModels/claudeModelsHeaders";
@@ -246,6 +246,37 @@ function grokBuildPositiveNumber(...values: unknown[]): number | undefined {
   );
 }
 
+const GROK_CLI_REGISTRY_CONTEXT_BY_ID = new Map<string, number>(
+  grok_cliProvider.models.flatMap((model) => {
+    const window = model.contextLength;
+    return typeof window === "number" && Number.isInteger(window) && window > 0
+      ? ([[model.id, window]] as Array<[string, number]>)
+      : [];
+  })
+);
+
+function resolveGrokBuildInputTokenLimit(
+  model: GrokBuildModelRecord,
+  metadata: GrokBuildModelRecord,
+  id: string
+): number | undefined {
+  // Registry first. Grok Build's /v1/models advertises contextWindow 256000 for every
+  // model, yet the backend serves grok-4.6 up to 500k (prod: 88 successful requests
+  // with 256k-485k input; upstream rejects at "> 500000 tokens"). Trusting the
+  // advertised number pinned 256k auto:discovery overrides over the verified window.
+  // Models the registry does not know keep the upstream number — under-advertising
+  // only makes clients compact early — and never get an invented default.
+  return (
+    GROK_CLI_REGISTRY_CONTEXT_BY_ID.get(id) ??
+    grokBuildPositiveNumber(
+      model.contextWindow,
+      model.context_window,
+      metadata.contextWindow,
+      metadata.totalContextTokens
+    )
+  );
+}
+
 function getGrokBuildModelItems(data: unknown): unknown[] {
   const envelope = asGrokBuildRecord(data);
   if (Array.isArray(data)) return data;
@@ -349,13 +380,7 @@ function normalizeGrokBuildModel(value: unknown): GrokBuildModelRecord | null {
   // here would route their request shape to the wrong upstream endpoint.
   if (backend !== "responses") return null;
 
-  const inputTokenLimit =
-    grokBuildPositiveNumber(
-      model.contextWindow,
-      model.context_window,
-      metadata.contextWindow,
-      metadata.totalContextTokens
-    ) || GROK_BUILD_DEFAULT_CONTEXT_WINDOW;
+  const inputTokenLimit = resolveGrokBuildInputTokenLimit(model, metadata, id);
   const outputTokenLimit = grokBuildPositiveNumber(
     model.maxCompletionTokens,
     model.max_completion_tokens
@@ -369,7 +394,7 @@ function normalizeGrokBuildModel(value: unknown): GrokBuildModelRecord | null {
     name: grokBuildString(model.name, id) || id,
     owned_by: "grok-cli",
     ...(description ? { description } : {}),
-    inputTokenLimit,
+    ...(typeof inputTokenLimit === "number" ? { inputTokenLimit } : {}),
     ...(outputTokenLimit ? { outputTokenLimit } : {}),
     ...(supportsThinking ? { supportsThinking: true } : {}),
     ...(supportedThinkingEfforts.length > 0 ? { supportedThinkingEfforts } : {}),

src/lib/contextWindowResolver.ts

@@ -52,7 +52,15 @@ export function reconcileContextWindows(
   for (const { provider, modelId, window } of discovered) {
     result.scanned++;
     if (!provider || !modelId) continue;
-    if (typeof window !== "number" || !Number.isInteger(window) || window <= 0) continue;
+    if (typeof window !== "number" || !Number.isInteger(window) || window <= 0) {
+      // Discovery omitted a real window. Drop a stale auto:discovery pin
+      // (e.g. grok-cli's old 256k fallback) so registry/models.dev can win.
+      if (deps.getExistingSource(provider, modelId) === "auto:discovery") {
+        deps.removeOverride(provider, modelId);
+        result.removed++;
+      }
+      continue;
+    }
 
     const existingSource = deps.getExistingSource(provider, modelId);
     if (existingSource === "manual") {

tests/unit/context-window-reconcile.test.ts

@@ -8,10 +8,7 @@ import {
 
 // Feature 5004 — pure reconcile logic (auto:discovery overrides), deps injected.
 
-function makeDeps(
-  catalog: Record<string, number | null>,
-  existing: Record<string, string> = {}
-) {
+function makeDeps(catalog: Record<string, number | null>, existing: Record<string, string> = {}) {
   const writes: Array<[string, string, number]> = [];
   const removes: Array<[string, string]> = [];
   const deps: ReconcileDeps = {
@@ -25,7 +22,9 @@ function makeDeps(
 
 describe("reconcileContextWindows (5004)", () => {
   it("writes an auto override when the discovered window diverges from the catalog", () => {
-    const discovered: DiscoveredWindow[] = [{ provider: "openai", modelId: "gpt-x", window: 400000 }];
+    const discovered: DiscoveredWindow[] = [
+      { provider: "openai", modelId: "gpt-x", window: 400000 },
+    ];
     const { deps, writes } = makeDeps({ "openai/gpt-x": 128000 });
     const r = reconcileContextWindows(discovered, deps);
     assert.deepEqual(writes, [["openai", "gpt-x", 400000]]);
@@ -34,7 +33,9 @@ describe("reconcileContextWindows (5004)", () => {
   });
 
   it("does nothing when the discovered window already matches the catalog", () => {
-    const discovered: DiscoveredWindow[] = [{ provider: "openai", modelId: "gpt-x", window: 128000 }];
+    const discovered: DiscoveredWindow[] = [
+      { provider: "openai", modelId: "gpt-x", window: 128000 },
+    ];
     const { deps, writes, removes } = makeDeps({ "openai/gpt-x": 128000 });
     const r = reconcileContextWindows(discovered, deps);
     assert.deepEqual(writes, []);
@@ -43,16 +44,26 @@ describe("reconcileContextWindows (5004)", () => {
   });
 
   it("self-heals: removes a stale auto override once the catalog catches up", () => {
-    const discovered: DiscoveredWindow[] = [{ provider: "openai", modelId: "gpt-x", window: 128000 }];
-    const { deps, removes } = makeDeps({ "openai/gpt-x": 128000 }, { "openai/gpt-x": "auto:discovery" });
+    const discovered: DiscoveredWindow[] = [
+      { provider: "openai", modelId: "gpt-x", window: 128000 },
+    ];
+    const { deps, removes } = makeDeps(
+      { "openai/gpt-x": 128000 },
+      { "openai/gpt-x": "auto:discovery" }
+    );
     const r = reconcileContextWindows(discovered, deps);
     assert.deepEqual(removes, [["openai", "gpt-x"]]);
     assert.equal(r.removed, 1);
   });
 
   it("never overwrites or removes a manual override", () => {
-    const discovered: DiscoveredWindow[] = [{ provider: "openai", modelId: "gpt-x", window: 999999 }];
-    const { deps, writes, removes } = makeDeps({ "openai/gpt-x": 128000 }, { "openai/gpt-x": "manual" });
+    const discovered: DiscoveredWindow[] = [
+      { provider: "openai", modelId: "gpt-x", window: 999999 },
+    ];
+    const { deps, writes, removes } = makeDeps(
+      { "openai/gpt-x": 128000 },
+      { "openai/gpt-x": "manual" }
+    );
     const r = reconcileContextWindows(discovered, deps);
     assert.deepEqual(writes, []);
     assert.deepEqual(removes, []);
@@ -77,10 +88,39 @@ describe("reconcileContextWindows (5004)", () => {
   });
 
   it("writes an override when the catalog does not know the model (catalog null)", () => {
-    const discovered: DiscoveredWindow[] = [{ provider: "local", modelId: "my-7b", window: 131072 }];
+    const discovered: DiscoveredWindow[] = [
+      { provider: "local", modelId: "my-7b", window: 131072 },
+    ];
     const { deps, writes } = makeDeps({}); // catalog null for everything
     const r = reconcileContextWindows(discovered, deps);
     assert.deepEqual(writes, [["local", "my-7b", 131072]]);
     assert.equal(r.written, 1);
   });
+
+  it("removes a stale auto:discovery pin when discovery omits the window", () => {
+    const discovered: DiscoveredWindow[] = [
+      { provider: "grok-cli", modelId: "grok-4.6", window: null },
+    ];
+    const { deps, writes } = makeDeps(
+      { "grok-cli/grok-4.6": 256000 },
+      { "grok-cli/grok-4.6": "auto:discovery" }
+    );
+    const r = reconcileContextWindows(discovered, deps);
+    assert.deepEqual(writes, []);
+    assert.equal(r.removed, 1);
+    assert.equal(r.written, 0);
+  });
+
+  it("keeps a manual override when discovery omits the window", () => {
+    const discovered: DiscoveredWindow[] = [
+      { provider: "grok-cli", modelId: "grok-4.6", window: null },
+    ];
+    const { deps, writes } = makeDeps(
+      { "grok-cli/grok-4.6": 256000 },
+      { "grok-cli/grok-4.6": "manual" }
+    );
+    const r = reconcileContextWindows(discovered, deps);
+    assert.deepEqual(writes, []);
+    assert.equal(r.removed, 0);
+  });
 });

tests/unit/grok-cli-responses-compat.test.ts

@@ -3,7 +3,6 @@ import assert from "node:assert/strict";
 
 import { grok_cliProvider } from "../../open-sse/config/providers/registry/grok-cli/index.ts";
 import {
-  GROK_BUILD_DEFAULT_CONTEXT_WINDOW,
   getGrokBuildClientVersion,
   getGrokBuildUserAgent,
   GROK_BUILD_MODELS_URL,
@@ -206,9 +205,51 @@ test("grok-cli live model discovery uses the authenticated session contract", ()
       id: "session-only-alias",
       name: "Session Only",
       owned_by: "grok-cli",
-      inputTokenLimit: GROK_BUILD_DEFAULT_CONTEXT_WINDOW,
       apiFormat: "responses",
       supportedEndpoints: ["responses"],
     },
   ]);
 });
+
+test("grok-cli discovery prefers the registry window over Grok Build's advertised one", () => {
+  // Grok Build /v1/models advertises contextWindow 256000 for every model, but the
+  // backend serves grok-4.6 up to 500k: prod logged 88 successful grok-4.6 requests
+  // with 256k-485k input tokens and upstream rejects at "> 500000 tokens". Trusting
+  // the advertised number pinned 256k auto:discovery overrides over the verified
+  // registry window. Unknown models keep the upstream number (under-advertising is
+  // safe), and models with neither stay unadvertised.
+  const config = PROVIDER_MODELS_CONFIG["grok-cli"];
+  const models = config.parseResponse({
+    data: [
+      {
+        id: "grok-4.6",
+        model: "grok-4.6",
+        name: "Grok 4.6",
+        contextWindow: 256000,
+        apiBackend: "responses",
+      },
+      {
+        id: "grok-4.5",
+        model: "grok-4.5",
+        name: "Grok 4.5",
+        apiBackend: "responses",
+      },
+      {
+        id: "grok-next-build",
+        name: "Grok Next Build",
+        contextWindow: 256000,
+        apiBackend: "responses",
+      },
+      {
+        id: "session-only-alias",
+        name: "Session Only",
+        apiBackend: "responses",
+      },
+    ],
+  });
+  const byId = Object.fromEntries(models.map((model) => [model.id, model]));
+  assert.equal(byId["grok-4.6"].inputTokenLimit, 500000);
+  assert.equal(byId["grok-4.5"].inputTokenLimit, 500000);
+  assert.equal(byId["grok-next-build"].inputTokenLimit, 256000);
+  assert.equal("inputTokenLimit" in byId["session-only-alias"], false);
+});