fix(catalog): preserve opaque gateway model ids

Co-authored-by: Matheus Gois <matheus.gois@doordash.com>

packages/catalog/src/identity/reference.ts

@@ -8,7 +8,7 @@
  * may strip `search`-style markers and prefers cache-pricing-complete
  * references, both of which would be wrong for canonical coalescing.
  */
-import type { Api, Model } from "../types";
+import type { Api, Model, ThinkingConfig } from "../types";
 import { getBracketStrippedModelIdCandidates, getLongestModelLikeIdSegment, getModelLikeIdSegments } from "./id";
 import { REFERENCE_TRAILING_MARKER_PATTERN } from "./markers";
 
@@ -137,8 +137,27 @@ function getReferenceCandidateIds(modelId: string): string[] {
 	return [...candidates];
 }
 
+/**
+ * Inherit bundled reference thinking only for same-provider matches. Wire routing
+ * (`effortRouting`) is provider-specific; cross-provider inheritance can rewrite
+ * gateway ids (e.g. Portkey `@modal/GLM-5-2-FP8` → devin `glm-5-2`).
+ */
+export function inheritReferenceThinking(
+	modelThinking: ThinkingConfig | undefined,
+	reference: Pick<Model<Api>, "provider" | "thinking"> | undefined,
+	provider: string,
+): ThinkingConfig | undefined {
+	if (modelThinking !== undefined) return modelThinking;
+	if (!reference?.thinking) return undefined;
+	if (reference.provider !== provider) return undefined;
+	return reference.thinking;
+}
+
 /** Resolve a (possibly proxied/affixed) model id to its bundled upstream reference. */
 export function resolveModelReference(modelId: string, index: ModelReferenceIndex): Model<Api> | undefined {
+	// Portkey/gateway wire ids (`@provider/model`) are opaque; fuzzy matching would
+	// map them to unrelated bundled entries (e.g. `@modal/GLM-5-2-FP8` → devin/glm-5-2).
+	if (modelId.startsWith("@")) return undefined;
 	for (const candidate of getReferenceCandidateIds(modelId)) {
 		const key = normalizeReferenceKey(candidate);
 		const reference = index.exact.get(key) ?? index.suffixAlias.get(key);

packages/catalog/test/gateway-reference.test.ts

@@ -0,0 +1,17 @@
+import { describe, expect, test } from "bun:test";
+import { getBundledModelReferenceIndex } from "../src/identity/bundled";
+import { inheritReferenceThinking, resolveModelReference } from "../src/identity/reference";
+
+describe("Portkey gateway model references", () => {
+	test("@modal ids do not fuzzy-match bundled catalog entries", () => {
+		const index = getBundledModelReferenceIndex();
+		expect(resolveModelReference("@modal/GLM-5-2-FP8", index)).toBeUndefined();
+	});
+
+	test("cross-provider references do not inherit wire routing thinking", () => {
+		const index = getBundledModelReferenceIndex();
+		const devinGlm = resolveModelReference("glm-5-2", index);
+		expect(devinGlm?.provider).toBe("devin");
+		expect(inheritReferenceThinking(undefined, devinGlm, "gateway")).toBeUndefined();
+	});
+});

packages/coding-agent/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Fixed
+
+- Fixed Portkey/gateway custom models whose ids start with `@` (e.g. `@modal/GLM-5-2-FP8`) being rewritten to unrelated bundled wire ids (e.g. `glm-5-2`), which caused `400` responses requiring `x-portkey-config` or `x-portkey-provider`.
+
 ## [17.0.6] - 2026-07-20
 
 - Fixed failed plan-mode exits leaving the session on the restored execution model while plan mode remained active and silently changing ambient `xd://` tool presentation; rollback now restores the plan model, thinking level, and exact top-level-versus-mounted tool partition so exit can be retried safely ([#6013](https://github.com/can1357/oh-my-pi/pull/6013)).

packages/coding-agent/src/config/model-discovery.ts

@@ -10,6 +10,7 @@ import type { Api, Model, RemoteCompactionConfig } from "@oh-my-pi/pi-ai/types";
 import { buildModel } from "@oh-my-pi/pi-catalog/build";
 import {
 	getBundledModelReferenceIndex,
+	inheritReferenceThinking,
 	isQwenModelId,
 	resolveModelReference,
 	stripBracketedModelIdAffixes,
@@ -752,7 +753,7 @@ export async function discoverOpenAIModelsList(
 				provider: providerConfig.provider,
 				baseUrl,
 				reasoning: reference?.reasoning ?? false,
-				thinking: reference?.thinking,
+				thinking: inheritReferenceThinking(undefined, reference, providerConfig.provider),
 				input: nativeMetadataForModel?.input ?? reference?.input ?? ["text"],
 				...(providerConfig.discovery.type === "lm-studio" ? { imageInputDecoder: "stb" as const } : {}),
 				// Proxy/gateway pricing is provider-specific and rarely matches
@@ -907,7 +908,7 @@ export async function discoverProxyModels(
 				provider: providerConfig.provider,
 				baseUrl,
 				reasoning: reference?.reasoning ?? false,
-				thinking: reference?.thinking,
+				thinking: inheritReferenceThinking(undefined, reference, providerConfig.provider),
 				input: reference?.input ?? ["text"],
 				// Proxy pricing is provider-specific and usually does not match
 				// upstream bundled catalogs, so keep costs local-unknown even when

packages/coding-agent/src/config/model-registry.ts

@@ -64,7 +64,11 @@ const BUILT_IN_DISCOVERY_NON_AUTHORITATIVE_RETRY_MS = 5 * 60 * 1000;
 import type { ApiKeyResolver, FetchImpl } from "@oh-my-pi/pi-ai";
 import { registerOAuthProvider, unregisterOAuthProviders } from "@oh-my-pi/pi-ai/oauth";
 import type { OAuthCredentials, OAuthLoginCallbacks } from "@oh-my-pi/pi-ai/oauth/types";
-import { getBundledModelReferenceIndex, resolveModelReference } from "@oh-my-pi/pi-catalog/identity";
+import {
+	getBundledModelReferenceIndex,
+	inheritReferenceThinking,
+	resolveModelReference,
+} from "@oh-my-pi/pi-catalog/identity";
 import { isBunTestRuntime, isRecord, logger, wrapFetchForExtraCa } from "@oh-my-pi/pi-utils";
 import { parseModelString, resolveProviderModelReference } from "../config/model-resolver";
 import type { AuthStorage, OAuthCredential } from "../session/auth-storage";
@@ -667,7 +671,7 @@ function finalizeCustomModel(model: CustomModelOverlay, options: CustomModelBuil
 		provider: resolvedModel.provider,
 		baseUrl: resolvedModel.baseUrl,
 		reasoning: resolvedModel.reasoning ?? reference?.reasoning ?? (options.useDefaults ? false : undefined),
-		thinking: resolvedModel.thinking ?? reference?.thinking,
+		thinking: inheritReferenceThinking(resolvedModel.thinking, reference, resolvedModel.provider),
 		input: input as ("text" | "image")[],
 		...(supportsTools !== undefined ? { supportsTools } : {}),
 		cost,

packages/coding-agent/test/issue-portkey-gateway-wire-model.test.ts

@@ -0,0 +1,68 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+import { Effort } from "@oh-my-pi/pi-catalog/effort";
+import { resolveWireModelId } from "@oh-my-pi/pi-catalog/model-thinking";
+import { ModelRegistry } from "@oh-my-pi/pi-coding-agent/config/model-registry";
+import { resetSettingsForTest } from "@oh-my-pi/pi-coding-agent/config/settings";
+import { AuthStorage } from "@oh-my-pi/pi-coding-agent/session/auth-storage";
+import { removeSyncWithRetries, Snowflake } from "@oh-my-pi/pi-utils";
+
+describe("Portkey gateway custom models", () => {
+	let tempDir: string;
+	let modelsPath: string;
+	let authStorage: AuthStorage;
+
+	beforeEach(async () => {
+		resetSettingsForTest();
+		tempDir = path.join(os.tmpdir(), `pi-test-portkey-gateway-${Snowflake.next()}`);
+		fs.mkdirSync(tempDir, { recursive: true });
+		modelsPath = path.join(tempDir, "models.yml");
+		authStorage = await AuthStorage.create(":memory:");
+	});
+
+	afterEach(() => {
+		resetSettingsForTest();
+		authStorage.close();
+		if (tempDir && fs.existsSync(tempDir)) {
+			removeSyncWithRetries(tempDir);
+		}
+	});
+
+	test("gateway @modal/GLM-5-2-FP8 keeps full wire id (not devin glm-5-2)", () => {
+		fs.writeFileSync(
+			modelsPath,
+			`providers:
+  gateway:
+    baseUrl: https://gateway.example.com/v1
+    api: openai-completions
+    apiKey: test
+    authHeader: true
+    headers:
+      x-portkey-api-key: test
+    models:
+      - id: "@modal/GLM-5-2-FP8"
+        name: glm-5p2 (modal)
+        reasoning: true
+        input: [text]
+        contextWindow: 1048576
+        maxTokens: 131072
+        compat:
+          thinkingFormat: openai
+          supportsReasoningEffort: true
+          reasoningEffortMap:
+            minimal: none
+            low: low
+            medium: medium
+            high: high
+            xhigh: max
+`,
+		);
+		const registry = new ModelRegistry(authStorage, modelsPath);
+		const model = registry.find("gateway", "@modal/GLM-5-2-FP8");
+		expect(model).toBeDefined();
+		expect(resolveWireModelId(model!, Effort.High)).toBe("@modal/GLM-5-2-FP8");
+		expect(resolveWireModelId(model!, undefined)).toBe("@modal/GLM-5-2-FP8");
+	});
+});