fix(models): tighten audio capability + correct openai oauth content

docs/plans/state.json

@@ -89,6 +89,16 @@
       ],
       "test_status": "pnpm test:run src/auth/anthropic.test.ts + pnpm typecheck passing"
     },
+    "openai-oauth-and-audio-capability-tweaks": {
+      "status": "completed",
+      "date": "2026-02-14",
+      "summary": "Fixed OpenAI OAuth (Codex) message shaping by using output_text for assistant turns, and made native-audio detection conservative via a model allowlist so Flynn does not attempt to send raw audio to models that cannot consume it.",
+      "files_modified": [
+        "src/models/openai.ts",
+        "src/models/capabilities.ts"
+      ],
+      "test_status": "pnpm typecheck passing (no new tests added in this change)"
+    },
     "p0-p1-implementation-plan": {
       "file": "2026-02-06-p0-p1-implementation-plan.md",
       "status": "completed",

src/models/capabilities.ts

@@ -7,22 +7,26 @@
 
 export type ModelProvider = 'anthropic' | 'openai' | 'gemini' | 'bedrock' | 'github' | 'ollama' | 'llamacpp' | 'openrouter' | 'zhipuai' | 'xai';
 
-/** Providers that support native audio input in their API. */
+/**
-const AUDIO_CAPABLE_PROVIDERS = new Set<string>([
+ * Models known to support native audio input via their API.
-  'gemini',
+ * We use an allowlist (not a provider-level blanket) to avoid silently
-  'openai',
+ * dropping audio for models that don't actually handle audio content parts.
-  'github',  // GitHub Models uses OpenAI-compatible API
+ */
+const AUDIO_CAPABLE_MODELS = new Set<string>([
+  // Gemini — all current models support audio
+  'gemini-2.0-flash',
+  'gemini-2.0-pro',
+  'gemini-1.5-flash',
+  'gemini-1.5-pro',
+  // OpenAI — only multimodal audio models
+  'gpt-4o',
+  'gpt-4o-mini',
+  'gpt-4o-audio-preview',
 ]);
 
-/**
+/** Providers where all models support audio (e.g. Gemini). */
- * Models known NOT to support audio despite their provider supporting it.
+const AUDIO_CAPABLE_PROVIDERS = new Set<string>([
- * For example, older OpenAI models or specialized models.
+  'gemini',
- */
-const AUDIO_INCAPABLE_MODELS = new Set<string>([
-  // Older OpenAI models that predate audio input support
-  'gpt-3.5-turbo',
-  'gpt-4',
-  'gpt-4-turbo',
 ]);
 
 /**
@@ -34,15 +38,15 @@ const AUDIO_INCAPABLE_MODELS = new Set<string>([
 export function supportsAudioInput(provider: string, model: string, override?: boolean): boolean {
   if (override !== undefined) {return override;}
 
-  // Provider must be in the capable set
+  // Provider-level blanket (all models support audio)
-  if (!AUDIO_CAPABLE_PROVIDERS.has(provider)) {
+  if (AUDIO_CAPABLE_PROVIDERS.has(provider)) {
-    return false;
+    return true;
   }
 
-  // Check model-specific exclusions
+  // Model-level allowlist
-  if (AUDIO_INCAPABLE_MODELS.has(model)) {
+  if (AUDIO_CAPABLE_MODELS.has(model)) {
-    return false;
+    return true;
   }
 
-  return true;
+  return false;
 }

src/models/openai.ts

@@ -96,9 +96,10 @@ export class OpenAIClient implements ModelClient {
       .map((m) => {
         const text = getMessageTextWithTools(m);
         if (!text) {return null;}
+        const contentType = m.role === 'assistant' ? 'output_text' : 'input_text';
         return {
           role: m.role,
-          content: [{ type: 'input_text', text }],
+          content: [{ type: contentType, text }],
         };
       })
       .filter((x): x is NonNullable<typeof x> => Boolean(x));