Let OAI endpoints override default aliases if set

src/utils/model/model.auth.test.ts

@@ -6,6 +6,7 @@ import {
   getDefaultMainLoopModelSetting,
   getDefaultOpusModel,
   isOpus1mMergeEnabled,
+  parseUserSpecifiedModel,
 } from './model.js'
 import { GLM_5_2_MODEL } from './ncodeModels.js'
 
@@ -73,6 +74,16 @@ function restoreEnv(): void {
   } else {
     process.env.NOUMENA_DEFAULT_HAIKU_MODEL = originalDefaultHaiku
   }
+  if (originalOpenAIApiKey === undefined) {
+    delete process.env.OPENAI_API_KEY
+  } else {
+    process.env.OPENAI_API_KEY = originalOpenAIApiKey
+  }
+  if (originalOpenAIModel === undefined) {
+    delete process.env.OPENAI_MODEL
+  } else {
+    process.env.OPENAI_MODEL = originalOpenAIModel
+  }
 }
 
 function makeSession(
@@ -270,4 +281,40 @@ describe('model auth session gating', () => {
     })
   })
 
+  it('forwards BYOK OpenAI-compat model ids verbatim, even when they collide with a managed alias', () => {
+    process.env.CLAUDE_CODE_ENTRYPOINT = 'cli'
+    process.env.USER_TYPE = 'test'
+    delete process.env.NCODE_BUILD_MODE
+    delete process.env.CLAUDE_CODE_USE_BEDROCK
+    delete process.env.CLAUDE_CODE_USE_VERTEX
+    delete process.env.CLAUDE_CODE_USE_FOUNDRY
+    process.env.OPENAI_API_KEY = 'openai-key'
+
+    const session = makeSession({
+      principalKind: 'api_key_user',
+      principalSource: 'direct_api_key_env',
+      sessionState: 'usable',
+      headersKind: 'none',
+      providerAuthKind: 'byok_static_env',
+      providerPlan: {
+        mode: 'byok_static_env',
+        source: 'direct_api_key_env',
+        staticKeyEnvVarName: 'OPENAI_API_KEY',
+      },
+      hasUsableApiKey: true,
+      apiKey: 'openai-key',
+      rawApiKeySource: 'OPENAI_API_KEY',
+    })
+
+    withMockCurrentSession(session, () => {
+      // `glm-5.2` / `k2.7` are reserved NCode managed aliases, but here they name
+      // a model on the user's own OpenAI-compatible server. They must reach the
+      // wire verbatim rather than being rewritten to an internal deployment path
+      // (which the user's server would 404).
+      expect(parseUserSpecifiedModel('glm-5.2')).toBe('glm-5.2')
+      expect(parseUserSpecifiedModel('k2.7')).toBe('k2.7')
+      expect(parseUserSpecifiedModel('my-own-model')).toBe('my-own-model')
+    })
+  })
+
 })

src/utils/model/model.ts

@@ -566,9 +566,16 @@ export function parseUserSpecifiedModel(
     ? normalizedModel.replace(/\[1m]$/i, '').trim()
     : normalizedModel
 
-  const exactNcodeModel = resolveNCodeManagedModel(normalizedModel)
-  if (exactNcodeModel) {
-    return exactNcodeModel.model
+  // BYOK OpenAI-compat: the model id names a model on the user's own server and
+  // must pass through verbatim. Never resolve it against the built-in NCODE
+  // managed-model aliases — an id that collides with a reserved alias (e.g.
+  // `glm-5.2`, `k2.7`) would otherwise be rewritten to an internal deployment
+  // path the user's server does not recognize, causing a 404.
+  if (!isOpenAICompatByokActive()) {
+    const exactNcodeModel = resolveNCodeManagedModel(normalizedModel)
+    if (exactNcodeModel) {
+      return exactNcodeModel.model
+    }
   }
 
   if (isModelAlias(modelString)) {
@@ -587,9 +594,13 @@ export function parseUserSpecifiedModel(
     }
   }
 
-  const ncodeModel = resolveNCodeManagedModel(modelString)
-  if (ncodeModel) {
-    return ncodeModel.model
+  // See the BYOK note above: managed-alias resolution is skipped entirely for
+  // BYOK OpenAI-compat sessions so the user's model id is forwarded verbatim.
+  if (!isOpenAICompatByokActive()) {
+    const ncodeModel = resolveNCodeManagedModel(modelString)
+    if (ncodeModel) {
+      return ncodeModel.model
+    }
   }
 
   // Opus 4/4.1 are no longer available on the first-party API (same as