fix(compat): merge partial model spec overrides

src/compat.ts

@@ -2153,11 +2153,11 @@ export function getEffectiveAntigravityModels(): CompatModelEntry[] {
   const seen = new Set(MODEL_CATALOG.map((model) => model.id.toLowerCase()));
   const result: CompatModelEntry[] = MODEL_CATALOG.map((model) => {
     const spec = getModelSpec(model.id);
-    const contextSpec =
-      getModelSpecOverride(model.id) ?? dynamicCatalog.getModelSpec(model.id);
+    const contextWindow = getModelSpecOverride(model.id)?.contextWindow ??
+      dynamicCatalog.getModelSpec(model.id)?.contextWindow;
     return {
       ...model,
-      ctx: contextSpec?.contextWindow ?? model.ctx,
+      ctx: contextWindow ?? model.ctx,
       maxOutputTokens: spec.maxOutputTokens,
       thinkingBudget: spec.thinkingBudget,
       minThinkingBudget: spec.minThinkingBudget,

src/compat/model-specs.ts

@@ -9,6 +9,8 @@ export interface ModelSpec {
 	contextWindow?: number;
 }
 
+export type ModelSpecOverride = Partial<ModelSpec>;
+
 export const DEFAULT_MODEL_SPECS: Record<string, ModelSpec> = {
 	"gemini-pro-agent":          { maxOutputTokens: 65535, thinkingBudget: 10001, isThinking: true, contextWindow: 1_000_000 },
 	"gemini-3-flash-agent":      { maxOutputTokens: 65536, thinkingBudget: 10000, isThinking: true, contextWindow: 1_000_000 },
@@ -41,13 +43,13 @@ export const DEFAULT_MODEL_SPECS: Record<string, ModelSpec> = {
 	"gpt-oss-120b":              { maxOutputTokens: 32768, thinkingBudget: 8192,  isThinking: true, contextWindow: 131_072 },
 };
 
-let modelSpecsOverride: Record<string, ModelSpec> | null = null;
+let modelSpecsOverride: Record<string, ModelSpecOverride> | null = null;
 
 /**
- * Replace the bundled model spec table with operator-provided overrides.
+ * Apply operator-provided partial overrides over effective model specs.
  * Pass `null` to restore defaults. Called once at startup from index.ts.
  */
-export function setModelSpecsOverride(specs: Record<string, ModelSpec> | null): void {
+export function setModelSpecsOverride(specs: Record<string, ModelSpecOverride> | null): void {
 	modelSpecsOverride = specs && Object.keys(specs).length > 0
 		? Object.fromEntries(
 			Object.entries(specs).map(([key, spec]) => [key.toLowerCase(), spec]),
@@ -56,7 +58,10 @@ export function setModelSpecsOverride(specs: Record<string, ModelSpec> | null):
 }
 
 export function getActiveModelSpecs(): Record<string, ModelSpec> {
-	return modelSpecsOverride ?? DEFAULT_MODEL_SPECS;
+	if (!modelSpecsOverride) return DEFAULT_MODEL_SPECS;
+	return Object.fromEntries(
+		Object.keys(modelSpecsOverride).map((model) => [model, getModelSpec(model)]),
+	);
 }
 
 const GEMINI_MAX_OUTPUT_TOKENS = 65536;
@@ -84,7 +89,7 @@ export function getStaticModelSpec(model: string): ModelSpec | undefined {
 	return best?.spec;
 }
 
-export function getModelSpecOverride(model: string): ModelSpec | undefined {
+export function getModelSpecOverride(model: string): ModelSpecOverride | undefined {
 	if (!modelSpecsOverride) return undefined;
 	const lower = model.toLowerCase();
 	if (modelSpecsOverride[lower]) return modelSpecsOverride[lower];
@@ -94,19 +99,52 @@ export function getModelSpecOverride(model: string): ModelSpec | undefined {
 	return undefined;
 }
 
+function mergeModelSpec(
+	defaults: ModelSpec,
+	override: ModelSpecOverride,
+): ModelSpec {
+	const minThinkingBudget = typeof override.minThinkingBudget === "number" &&
+		Number.isFinite(override.minThinkingBudget) &&
+		override.minThinkingBudget >= 0
+		? override.minThinkingBudget
+		: defaults.minThinkingBudget;
+	const contextWindow = typeof override.contextWindow === "number" &&
+		Number.isFinite(override.contextWindow) &&
+		override.contextWindow > 0
+		? override.contextWindow
+		: defaults.contextWindow;
+	return {
+		maxOutputTokens: typeof override.maxOutputTokens === "number" &&
+			Number.isFinite(override.maxOutputTokens) &&
+			override.maxOutputTokens > 0
+			? override.maxOutputTokens
+			: defaults.maxOutputTokens,
+		thinkingBudget: typeof override.thinkingBudget === "number" &&
+			Number.isFinite(override.thinkingBudget)
+			? override.thinkingBudget
+			: defaults.thinkingBudget,
+		...(minThinkingBudget !== undefined ? { minThinkingBudget } : {}),
+		isThinking: typeof override.isThinking === "boolean"
+			? override.isThinking
+			: defaults.isThinking,
+		...(contextWindow !== undefined ? { contextWindow } : {}),
+	};
+}
+
 export function getModelSpec(model: string): ModelSpec {
 	const lower = model.toLowerCase();
-	const override = getModelSpecOverride(lower);
-	if (override) return override;
 	const dynamicSpec = dynamicCatalog.getModelSpec(lower);
-	if (dynamicSpec) return dynamicSpec;
 	const staticSpec = getStaticModelSpec(lower);
-	if (staticSpec) return staticSpec;
-	const family = getModelFamily(model);
-	if (family === "claude") return { maxOutputTokens: CLAUDE_MAX_OUTPUT_TOKENS, thinkingBudget: CLAUDE_DEFAULT_THINKING_BUDGET, isThinking: true, contextWindow: 1_000_000 };
-	if (family === "gemini") return { maxOutputTokens: GEMINI_MAX_OUTPUT_TOKENS, thinkingBudget: FALLBACK_THINKING_BUDGET, isThinking: true, contextWindow: 1_000_000 };
-	return { maxOutputTokens: 65536, thinkingBudget: FALLBACK_THINKING_BUDGET, isThinking: false, contextWindow: 128_000 };
-}
+	const family = getModelFamily(model);
+	const defaults = dynamicSpec ?? staticSpec ??
+		(family === "claude"
+			? { maxOutputTokens: CLAUDE_MAX_OUTPUT_TOKENS, thinkingBudget: CLAUDE_DEFAULT_THINKING_BUDGET, isThinking: true, contextWindow: 1_000_000 }
+			: family === "gemini"
+				? { maxOutputTokens: GEMINI_MAX_OUTPUT_TOKENS, thinkingBudget: FALLBACK_THINKING_BUDGET, isThinking: true, contextWindow: 1_000_000 }
+				: { maxOutputTokens: 65536, thinkingBudget: FALLBACK_THINKING_BUDGET, isThinking: false, contextWindow: 128_000 });
+	const override = getModelSpecOverride(lower);
+	return override ? mergeModelSpec(defaults, override) : defaults;
+}
 
 export function isThinkingModel(model: string): boolean {
 	return getModelSpec(model).isThinking;

src/config-defaults.ts

@@ -55,7 +55,8 @@ export function applyConfigDefaults(config: Config): Config {
 		idempotencyWindowMs: config.idempotencyWindowMs ?? 2000,
 		streamRecoveryMaxRetries: safeStreamRecoveryMaxRetries(config.streamRecoveryMaxRetries),
 		compressionMode: config.compressionMode ?? "off",
-accounts: config.accounts ? config.accounts.map((account) => ({
+		modelSpecs: config.modelSpecs,
+		accounts: config.accounts ? config.accounts.map((account) => ({
 			...normalizeAccountConfig(account),
 			tier: account.tier || "unknown",
 		})) : [],

src/providers/google-antigravity/catalog.ts

@@ -10,6 +10,7 @@
 // separately in `src/compat/model-specs.ts` via `maxOutputTokens`.
 
 import { dynamicCatalog } from "./dynamic-catalog.js";
+import { getModelSpecOverride } from "../../compat/model-specs.js";
 
 export interface AntigravityModelSpec {
   id: string;
@@ -73,16 +74,21 @@ const FALLBACK_CONTEXT_WINDOW = 128_000;
  * Resolve the upstream-published context window for an Antigravity model id.
  *
  * Lookup order:
- *   1. Exact id match in static table (lowercased).
- *   2. Dynamic catalog entry from live Antigravity endpoint.
- *   3. Substring match across the table (longest key wins via the order here).
- *   4. Family defaults: claude -> 1M, gemini -> 1M, gpt-oss -> 131_072.
- *   5. Defensive fallback: 128_000.
+ *   1. Operator exact/substring override.
+ *   2. Exact id match in static table (lowercased).
+ *   3. Dynamic catalog entry from live Antigravity endpoint.
+ *   4. Substring match across the table (longest key wins via the order here).
+ *   5. Family defaults: claude -> 1M, gemini -> 1M, gpt-oss -> 131_072.
+ *   6. Defensive fallback: 128_000.
  */
 export function getAntigravityContextWindow(model: string): number {
   if (!model) return FALLBACK_CONTEXT_WINDOW;
   const lower = model.toLowerCase().trim();
   if (!lower) return FALLBACK_CONTEXT_WINDOW;
+  const overrideCtx = getModelSpecOverride(lower)?.contextWindow;
+  if (typeof overrideCtx === "number" && Number.isFinite(overrideCtx) && overrideCtx > 0) {
+    return overrideCtx;
+  }
   const exact = ANTIGRAVITY_CONTEXT_WINDOWS[lower];
   if (typeof exact === "number") return exact;
   const dynamicCtx = dynamicCatalog.getContextWindow(lower);

src/types.ts

@@ -142,8 +142,8 @@ export interface Config {
   streamRecoveryMaxRetries?: number;
   // Prompt compression mode ("off" | "lite" | "rtk" | "rtk+lite").
   compressionMode?: "off" | "lite" | "rtk" | "rtk+lite";
-  // Override per-model specs used by the compat layer. Keys are model id substrings
-  // matched case-insensitively. When set, replaces the bundled defaults entirely.
+  // Partially override per-model specs used by the compat layer. Keys are model id
+  // substrings matched case-insensitively; omitted fields inherit effective defaults.
   modelSpecs?: Record<string, ModelSpecConfig>;
   // Override model-id aliases used to translate the operator-facing name
   // (e.g. "gemini-3.1-pro-high") to the upstream Antigravity name
@@ -224,10 +224,11 @@ export interface GoogleQuotaResponse {
 // Per-model thinking/output spec used by the compat layer.
 // Operators can override defaults via the `modelSpecs` field in accounts.json.
 export interface ModelSpecConfig {
-  maxOutputTokens: number;
-  thinkingBudget: number; // -1 = adaptive (model decides), >=0 = fixed
+  maxOutputTokens?: number;
+  thinkingBudget?: number; // -1 = adaptive (model decides), >=0 = fixed
   minThinkingBudget?: number;
-  isThinking: boolean;
+  isThinking?: boolean;
+  contextWindow?: number;
 }
 
 // Per-model quota info for an account

src/validators.ts

@@ -179,6 +179,9 @@ export function validateConfig(value: unknown): ValidationResult<Config> {
 				if (spec.isThinking !== undefined && typeof spec.isThinking !== "boolean") {
 					errors.push(`config.modelSpecs.${key}.isThinking must be a boolean`);
 				}
+				if (spec.contextWindow !== undefined && !isPositiveNumber(spec.contextWindow)) {
+					errors.push(`config.modelSpecs.${key}.contextWindow must be a positive number`);
+				}
 			}
 		}
 	}

test/context-windows.test.ts

@@ -7,7 +7,10 @@ import {
   OPENCODE_ZEN_FREE_MODELS,
   getOpenCodeZenContextWindow,
 } from "../src/providers/opencode-zen/catalog.js";
-import { getModelSpec } from "../src/compat/model-specs.js";
+import {
+	getModelSpec,
+	setModelSpecsOverride,
+} from "../src/compat/model-specs.js";
 
 /**
  * Verify each catalog model has a positive, finite context window.
@@ -69,6 +72,30 @@ describe("official context windows", () => {
 			assert.equal(getModelSpec("gemini-3.1-pro-low").contextWindow, 1_000_000);
 			assert.equal(getModelSpec("gpt-oss-120b-medium").contextWindow, 131_072);
 		});
+
+		it("honors an exact operator context window override", () => {
+			setModelSpecsOverride({
+				"gemini-3.8-flash-high": { contextWindow: 222_222 },
+			});
+			try {
+				assert.equal(getModelSpec("gemini-3.8-flash-high").contextWindow, 222_222);
+				assert.equal(getAntigravityContextWindow("gemini-3.8-flash-high"), 222_222);
+			} finally {
+				setModelSpecsOverride(null);
+			}
+		});
+
+		it("honors a substring operator context window override", () => {
+			setModelSpecsOverride({
+				"gemini-3.8": { contextWindow: 333_333 },
+			});
+			try {
+				assert.equal(getModelSpec("gemini-3.8-flash-high").contextWindow, 333_333);
+				assert.equal(getAntigravityContextWindow("gemini-3.8-flash-high"), 333_333);
+			} finally {
+				setModelSpecsOverride(null);
+			}
+		});
 	});
 
 	describe("openai codex", () => {

test/db-store.test.ts

@@ -19,6 +19,8 @@ import {
 } from "../src/db-store.js";
 import type { Config, PersistedState, TokenUsageTiered } from "../src/types.js";
 import type { PersistedResponsesStore } from "../src/db-store.js";
+import { applyConfigDefaults } from "../src/config-defaults.js";
+import { validateConfig } from "../src/validators.js";
 
 describe("db-store helpers", () => {
   before(async () => {
@@ -96,6 +98,35 @@ describe("db-store helpers", () => {
     assert.equal(cached.proxyPort, 51200);
   });
 
+  it("preserves partial model specs through normalization and persistence", async () => {
+    const modelSpecs = {
+      "gemini-3.8": {
+        maxOutputTokens: 1_000,
+        contextWindow: 222_222,
+      },
+    };
+    const candidate = {
+      accounts: [],
+      modelSpecs,
+      unknownTopLevelField: "discard me",
+    };
+    const validation = validateConfig(candidate);
+
+    assert.equal(validation.ok, true);
+    assert.ok(validation.value);
+
+    const normalized = applyConfigDefaults(validation.value);
+    assert.deepEqual(normalized.modelSpecs, modelSpecs);
+    assert.equal("unknownTopLevelField" in normalized, false);
+
+    await setCachedConfig(validation.value);
+    const cached = getCachedConfig();
+    assert.ok(cached);
+    assert.deepEqual(cached.modelSpecs, modelSpecs);
+    assert.equal("unknownTopLevelField" in cached, false);
+    assert.deepEqual(candidate.modelSpecs, modelSpecs);
+  });
+
   it("can cache and retrieve admin token", async () => {
     await setCachedAdminToken("test-token-value");
     assert.equal(getCachedAdminToken(), "test-token-value");

test/dynamic-catalog.test.ts

@@ -356,6 +356,70 @@ describe("DynamicModelRegistry", () => {
     }
   });
 
+  it("merges partial substring overrides over bundled specs without invalid payload numbers", () => {
+    setModelSpecsOverride({
+      "gemini-3.8": {
+        maxOutputTokens: 1000,
+        isThinking: true,
+      },
+    });
+
+    try {
+      assert.deepEqual(getModelSpec("gemini-3.8-flash-high"), {
+        maxOutputTokens: 1000,
+        thinkingBudget: -1,
+        isThinking: true,
+        contextWindow: 1_000_000,
+      });
+
+      const body = openAIToAntigravityBody({
+        model: "gemini-3.8-flash-high",
+        messages: [{ role: "user", content: "ping" }],
+      }) as { request: { generationConfig?: Record<string, unknown> } };
+      const serialized = JSON.parse(JSON.stringify(body)) as typeof body;
+      assert.deepEqual(serialized.request.generationConfig?.thinkingConfig, {
+        includeThoughts: true,
+      });
+    } finally {
+      setModelSpecsOverride(null);
+    }
+  });
+
+  it("merges partial exact overrides over runtime-discovered specs", () => {
+    dynamicCatalog.updateFromEndpointResponse({
+      models: {
+        "gemini-5.0-ultra": {
+          maxTokens: 5_000_000,
+          maxOutputTokens: 100_000,
+          supportsThinking: true,
+          thinkingBudget: 50_000,
+          minThinkingBudget: 2_000,
+          quotaInfo: { remainingFraction: 1 },
+        },
+      },
+    });
+    setModelSpecsOverride({
+      "gemini-5.0-ultra": { maxOutputTokens: 60_000 },
+    });
+
+    try {
+      assert.deepEqual(getModelSpec("gemini-5.0-ultra"), {
+        maxOutputTokens: 60_000,
+        thinkingBudget: 50_000,
+        minThinkingBudget: 2_000,
+        isThinking: true,
+        contextWindow: 5_000_000,
+      });
+
+      setModelSpecsOverride({
+        "gemini-5.0-ultra": { thinkingBudget: 0 },
+      });
+      assert.equal(getModelSpec("gemini-5.0-ultra").thinkingBudget, 0);
+    } finally {
+      setModelSpecsOverride(null);
+    }
+  });
+
   it("exposes dynamic context windows through getAntigravityContextWindow", () => {
     dynamicCatalog.updateFromEndpointResponse({
       models: {