[codex] Add native audio input support

agent/models_dev.py

@@ -480,7 +480,8 @@ def get_model_capabilities(provider: str, model: str) -> Optional[ModelCapabilit
     else:
         input_mods = None
     if isinstance(input_mods, list):
-        supports_vision = "image" in input_mods
+        input_mod_values = {str(mod).strip().lower() for mod in input_mods}
+        supports_vision = "image" in input_mod_values
     else:
         supports_vision = bool(entry.get("attachment", False))
     supports_reasoning = bool(entry.get("reasoning", False))

hermes_cli/runtime_provider.py

@@ -1431,6 +1431,7 @@ def resolve_runtime_provider(
     explicit_api_key: Optional[str] = None,
     explicit_base_url: Optional[str] = None,
     target_model: Optional[str] = None,
+    allow_auto_codex_fallback: bool = True,
 ) -> Dict[str, Any]:
     """Resolve runtime provider credentials for agent execution.
 
@@ -1441,6 +1442,11 @@ def resolve_runtime_provider(
     api_mode is derived from the model they are switching TO, not the stale
     persisted default. Other callers can leave it None to preserve existing
     behavior (api_mode derived from config).
+
+    allow_auto_codex_fallback: When False, an auto-detected Codex provider with
+    invalid credentials raises instead of falling through to another provider.
+    The normal chat path keeps the historical fallback behavior; utility
+    endpoints that must stay inside the Codex auth boundary can opt out.
     """
     requested_provider = resolve_requested_provider(requested)
 
@@ -1662,7 +1668,7 @@ def resolve_runtime_provider(
                 "requested_provider": requested_provider,
             }
         except AuthError:
-            if requested_provider != "auto":
+            if requested_provider != "auto" or not allow_auto_codex_fallback:
                 raise
             # Auto-detected Codex but credentials are stale/revoked —
             # fall through to env-var providers (e.g. OpenRouter).

run_agent.py

@@ -1690,13 +1690,15 @@ def _flush_messages_to_session_db(self, messages: List[Dict], conversation_histo
                 if _is_multimodal_tool_result(content):
                     content = _multimodal_text_summary(content)
                 elif isinstance(content, list):
-                    # List of OpenAI-style content parts: strip images, keep text.
+                    # List of OpenAI-style content parts: strip media, keep text.
                     _txt = []
                     for p in content:
                         if isinstance(p, dict) and p.get("type") == "text":
                             _txt.append(str(p.get("text", "")))
                         elif isinstance(p, dict) and p.get("type") in {"image", "image_url", "input_image"}:
                             _txt.append("[screenshot]")
+                        elif isinstance(p, dict) and p.get("type") in {"audio", "input_audio"}:
+                            _txt.append("[audio]")
                     content = "\n".join(_txt) if _txt else None
                 tool_calls_data = None
                 if hasattr(msg, "tool_calls") and isinstance(msg.tool_calls, list) and msg.tool_calls:

tests/agent/test_models_dev.py

@@ -401,3 +401,8 @@ def test_model_not_found_returns_none(self):
         with patch("agent.models_dev.fetch_models_dev", return_value=CAPS_REGISTRY):
             caps = get_model_capabilities("anthropic", "nonexistent-model")
         assert caps is None
+
+    def test_provider_not_found_returns_none(self):
+        with patch("agent.models_dev.fetch_models_dev", return_value=CAPS_REGISTRY):
+            caps = get_model_capabilities("nonexistent-provider", "gemma-4-31b-it")
+        assert caps is None

tests/gateway/test_api_server.py

@@ -534,8 +534,10 @@ def _create_app(adapter: APIServerAdapter) -> web.Application:
     app.router.add_get("/v1/capabilities", adapter._handle_capabilities)
     app.router.add_get("/v1/skills", adapter._handle_skills)
     app.router.add_get("/v1/toolsets", adapter._handle_toolsets)
+    app.router.add_post("/v1/audio/transcriptions", adapter._handle_audio_transcriptions)
     app.router.add_post("/v1/chat/completions", adapter._handle_chat_completions)
     app.router.add_post("/v1/responses", adapter._handle_responses)
+    app.router.add_post("/v1/runs", adapter._handle_runs)
     app.router.add_get("/v1/responses/{response_id}", adapter._handle_get_response)
     app.router.add_delete("/v1/responses/{response_id}", adapter._handle_delete_response)
     return app
@@ -798,11 +800,35 @@ async def test_capabilities_advertises_plugin_safe_contract(self, adapter):
             assert data["features"]["chat_completions"] is True
             assert data["features"]["run_status"] is True
             assert data["features"]["run_events_sse"] is True
+            assert data["features"]["audio_api"] is True
+            assert data["audio"]["transcription"] is True
+            assert data["audio"]["native_model"] is False
+            assert data["audio"]["max_bytes"] == 26214400
+            assert "ogg" in data["audio"]["formats"]
             assert data["features"]["session_continuity_header"] == "X-Hermes-Session-Id"
             assert data["endpoints"]["run_status"]["path"] == "/v1/runs/{run_id}"
+            assert data["endpoints"]["audio_transcriptions"] == {
+                "method": "POST",
+                "path": "/v1/audio/transcriptions",
+            }
             assert data["endpoints"]["skills"] == {"method": "GET", "path": "/v1/skills"}
             assert data["endpoints"]["toolsets"] == {"method": "GET", "path": "/v1/toolsets"}
 
+    @pytest.mark.asyncio
+    async def test_capabilities_never_reports_native_audio_from_runtime(self, adapter):
+        app = _create_app(adapter)
+        with patch("gateway.run._load_gateway_config", return_value={}), \
+             patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"provider": "openai"}), \
+             patch("gateway.run._resolve_gateway_model", return_value="gpt-audio"):
+            async with TestClient(TestServer(app)) as cli:
+                resp = await cli.get("/v1/capabilities")
+                assert resp.status == 200
+                data = await resp.json()
+
+        assert data["features"]["audio_api"] is True
+        assert data["audio"]["transcription"] is True
+        assert data["audio"]["native_model"] is False
+
     @pytest.mark.asyncio
     async def test_capabilities_requires_auth_when_key_configured(self, auth_adapter):
         app = _create_app(auth_adapter)
@@ -819,6 +845,69 @@ async def test_capabilities_requires_auth_when_key_configured(self, auth_adapter
             assert data["auth"]["required"] is True
 
 
+# ---------------------------------------------------------------------------
+# /v1/runs audio preflight
+# ---------------------------------------------------------------------------
+
+
+class TestRunsAudioPreflight:
+    @pytest.mark.asyncio
+    async def test_runs_rejects_unsupported_audio_before_run_state(self, adapter):
+        app = _create_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post(
+                "/v1/runs",
+                json={
+                    "input": [
+                        {
+                            "role": "user",
+                            "content": [
+                                {"type": "input_audio", "input_audio": {"data": "ZmFrZQ==", "format": "ogg"}},
+                            ],
+                        }
+                    ],
+                },
+            )
+            data = await resp.json()
+
+        assert resp.status == 400
+        assert data["error"]["code"] == "unsupported_audio_input"
+        assert adapter._run_streams == {}
+        assert adapter._run_statuses == {}
+
+    @pytest.mark.asyncio
+    async def test_runs_rejects_previous_response_assistant_audio_before_run_state(self, adapter):
+        adapter._response_store.put(
+            "resp_prev",
+            {
+                "conversation_history": [
+                    {
+                        "role": "assistant",
+                        "content": [
+                            {"type": "input_audio", "input_audio": {"data": "ZmFrZQ==", "format": "ogg"}},
+                        ],
+                    },
+                ],
+                "session_id": "session_prev",
+                "instructions": None,
+            },
+        )
+
+        app = _create_app(adapter)
+        async with TestClient(TestServer(app)) as cli:
+            resp = await cli.post(
+                "/v1/runs",
+                json={"previous_response_id": "resp_prev", "input": "next"},
+            )
+            data = await resp.json()
+
+        assert resp.status == 400
+        assert data["error"]["code"] == "unsupported_audio_input"
+        assert data["error"]["param"] == "previous_response_id"
+        assert adapter._run_streams == {}
+        assert adapter._run_statuses == {}
+
+
 # ---------------------------------------------------------------------------
 # /v1/skills and /v1/toolsets endpoints
 # ---------------------------------------------------------------------------