fix: keep model wire format provider-scoped

open-sse/config/providerModels.ts

@@ -170,8 +170,29 @@ export function findModelName(aliasOrId: string, modelId: string): string {
 }
 
 export function getModelTargetFormat(aliasOrId: string, modelId: string): string | null {
-  const models = PROVIDER_MODELS[aliasOrId];
-  const found = models?.find((m) => m.id === modelId) || getGlobalModel(modelId);
+  const alias = PROVIDER_ID_TO_ALIAS[aliasOrId] || aliasOrId;
+  const models = PROVIDER_MODELS[alias] || PROVIDER_MODELS[aliasOrId];
+  // Wire-format metadata is provider-scoped. A model's capabilities can be
+  // shared globally, but an OpenAI registry entry's Responses target must not
+  // leak into another provider's Chat Completions connection (e.g.
+  // Nous/OpenRouter `openai/gpt-5.6-luna`).
+  // Preserve the provider-prefix normalization used before the global model
+  // fallback was added: `openai/openai/gpt-5.6-luna` is an OpenAI-local
+  // spelling, not a reason to consult another provider's metadata.
+  const prefixes = [aliasOrId, alias].filter(
+    (value, index, values) => values.indexOf(value) === index
+  );
+  const bareModelId =
+    typeof modelId === "string"
+      ? prefixes.reduce(
+          (current, prefix) =>
+            current === modelId && modelId.startsWith(`${prefix}/`)
+              ? modelId.slice(prefix.length + 1)
+              : current,
+          modelId
+        )
+      : modelId;
+  const found = models?.find((m) => m.id === bareModelId);
   if (found?.targetFormat) return found.targetFormat;
   // #5842: OpenAI "*-pro" reasoning models (o1-pro, gpt-5.x-pro) are only served by
   // the native /v1/responses endpoint — /v1/chat/completions 404s ("only supported

tests/unit/openai-gpt56-responses-routing.test.ts

@@ -15,8 +15,13 @@
 import test from "node:test";
 import assert from "node:assert/strict";
 
-import { getModelTargetFormat } from "../../open-sse/config/providerModels.ts";
+import {
+  getModelTargetFormat,
+  getProviderModel,
+  PROVIDER_ID_TO_ALIAS,
+} from "../../open-sse/config/providerModels.ts";
 import { DefaultExecutor } from "../../open-sse/executors/default.ts";
+import { resolveChatCoreTargetFormat } from "../../open-sse/handlers/chatCore/targetFormat.ts";
 
 test("getModelTargetFormat routes the public OpenAI GPT-5.6 family through Responses", () => {
   for (const modelId of ["gpt-5.6", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna"]) {
@@ -43,3 +48,40 @@ test("DefaultExecutor keeps /v1/chat/completions for gpt-5.4", () => {
   const url = executor.buildUrl("gpt-5.4", true, 0, null);
   assert.equal(url, "https://api.openai.com/v1/chat/completions");
 });
+
+test("cross-provider GPT-5.6 keeps the provider's Chat Completions wire format", () => {
+  for (const provider of ["nous", "openrouter"]) {
+    assert.equal(
+      getModelTargetFormat(provider, "openai/gpt-5.6-luna"),
+      null,
+      `${provider} must not inherit OpenAI's Responses target format`
+    );
+    assert.equal(
+      getModelTargetFormat(provider, "gpt-5.6-luna"),
+      null,
+      `${provider} must not inherit OpenAI's Responses target format for bare ids`
+    );
+  }
+});
+
+test("chatCore target-format resolution keeps provider defaults for cross-provider GPT-5.6", () => {
+  for (const provider of ["nous", "openrouter"]) {
+    const result = resolveChatCoreTargetFormat({
+      provider,
+      resolvedModel: "openai/gpt-5.6-luna",
+      apiFormat: undefined,
+      sourceFormat: "openai",
+      customModelTargetFormat: undefined,
+      providerSpecificData: undefined,
+    });
+    assert.equal(result.alias, PROVIDER_ID_TO_ALIAS[provider] || provider);
+    assert.equal(result.targetFormat, "openai");
+  }
+});
+
+test("global capability fallback remains available without leaking wire format", () => {
+  const metadata = getProviderModel("nous", "gpt-5.6-luna");
+  assert.equal(metadata?.supportsVision, true);
+  assert.equal(metadata?.supportsReasoning, true);
+  assert.equal(getModelTargetFormat("nous", "gpt-5.6-luna"), null);
+});