feat(ai): exposed catalog metadata in auth gateway model list

- Added context length, max tokens, input modalities, and tool support flags to GET /v1/models response in auth gateway server.
- Updated auth gateway model list tests to verify catalog metadata fields in output rows.

packages/ai/CHANGELOG.md

@@ -4,6 +4,7 @@
 
 ### Added
 
+- Added model metadata fields (`context_length`, `max_output_tokens`, `input_modalities`, etc.) to auth gateway model listing responses
 - Added `Tool.coerceArguments` (default `true`): setting it `false` opts a tool out of every LLM-quirk argument repair pass (JSON-string parsing, object→string stringification, unrecognized-key dropping, singleton array wrapping) so validation runs verbatim. Tools whose arguments are the deliverable payload — like the subagent `yield` tool — use it to keep lossy repairs from silently corrupting data their own validate-and-retry loop is designed to correct.
 
 ### Fixed

packages/ai/src/auth-gateway/server.ts

@@ -730,19 +730,45 @@ async function handleCredentialsCheck(storage: AuthStorage, signal: AbortSignal)
 	return json(200, { generatedAt: Date.now(), credentials });
 }
 
+/**
+ * Row shape for `GET /v1/models`. Beyond the OpenAI-standard `id`/`object`/
+ * `owned_by`, rows advertise the catalog metadata OpenAI-compatible clients
+ * (omp's own proxy discovery, Zed's openai_compatible provider, ...) read to
+ * size and capability-gate discovered models: `context_length`,
+ * `max_output_tokens`, `input_modalities`, and `supports_tools` (only emitted
+ * when the catalog explicitly reports `false`; absent means usable).
+ */
+interface ModelListRow {
+	id: string;
+	object: "model";
+	owned_by: string;
+	api: Api;
+	display_name: string;
+	context_length?: number;
+	max_output_tokens?: number;
+	input_modalities: ("text" | "image")[];
+	supports_tools?: boolean;
+}
+
 function handleModelsList(opts: AuthGatewayBootOptions): Response {
 	const seen = new Set<string>();
-	const data: Array<{ id: string; object: "model"; owned_by: string; api: Api }> = [];
+	const data: ModelListRow[] = [];
 	for (const model of opts.listModels?.() ?? []) {
 		const id = `${model.provider}/${model.id}`;
 		if (seen.has(id)) continue;
 		seen.add(id);
-		data.push({
+		const row: ModelListRow = {
 			id,
 			object: "model",
 			owned_by: model.provider,
 			api: model.api,
-		});
+			display_name: model.name,
+			input_modalities: model.input,
+		};
+		if (model.contextWindow != null) row.context_length = model.contextWindow;
+		if (model.maxTokens != null) row.max_output_tokens = model.maxTokens;
+		if (model.supportsTools === false) row.supports_tools = false;
+		data.push(row);
 	}
 	return json(200, { object: "list", data });
 }

packages/ai/test/auth-gateway-model-list.test.ts

@@ -26,8 +26,26 @@ test("model listing exposes one provider-qualified route per upstream model", as
 		expect(await response.json()).toEqual({
 			object: "list",
 			data: [
-				{ id: "anthropic/shared-model", object: "model", owned_by: "anthropic", api: "mock" },
-				{ id: "devin/shared-model", object: "model", owned_by: "devin", api: "mock" },
+				{
+					id: "anthropic/shared-model",
+					object: "model",
+					owned_by: "anthropic",
+					api: "mock",
+					display_name: "shared-model",
+					context_length: 200_000,
+					max_output_tokens: 32_768,
+					input_modalities: ["text"],
+				},
+				{
+					id: "devin/shared-model",
+					object: "model",
+					owned_by: "devin",
+					api: "mock",
+					display_name: "shared-model",
+					context_length: 200_000,
+					max_output_tokens: 32_768,
+					input_modalities: ["text"],
+				},
 			],
 		});
 	} finally {