fix(gemini): tighten native routing and streaming replay

- only use the native adapter for the canonical Gemini native endpoint
- keep custom and /openai base URLs on the OpenAI-compatible path
- preserve Hermes keepalive transport injection for native Gemini clients
- stabilize streaming tool-call replay across repeated SSE events
- add follow-up tests for base_url precedence, async streaming, and duplicate tool-call chunks

tests/hermes_cli/test_config_validation.py

@@ -13,7 +13,7 @@ class TestCustomProvidersValidation:
         issues = validate_config_structure({
             "custom_providers": {
                 "name": "Generativelanguage.googleapis.com",
-                "base_url": "https://generativelanguage.googleapis.com/v1beta/openai",
+                "base_url": "https://generativelanguage.googleapis.com/v1beta",
                 "api_key": "xxx",
                 "model": "models/gemini-2.5-flash",
                 "rate_limit_delay": 2.0,

tests/hermes_cli/test_gemini_provider.py

@@ -210,8 +210,10 @@ class TestGeminiAgentInit:
     def test_gemini_agent_uses_native_client(self, monkeypatch):
         monkeypatch.setenv("GOOGLE_API_KEY", "AIzaSy_REAL_KEY")
         with patch("agent.gemini_native_adapter.GeminiNativeClient") as mock_client, \
-             patch("run_agent.OpenAI") as mock_openai:
+             patch("run_agent.OpenAI") as mock_openai, \
+             patch("run_agent.ContextCompressor") as mock_compressor:
             mock_client.return_value = MagicMock()
+            mock_compressor.return_value = MagicMock(context_length=1048576, threshold_tokens=524288)
             from run_agent import AIAgent
             AIAgent(
                 model="gemini-2.5-flash",
@@ -222,6 +224,38 @@ class TestGeminiAgentInit:
         assert mock_client.called
         mock_openai.assert_not_called()
 
+    def test_gemini_custom_base_url_keeps_openai_client(self, monkeypatch):
+        monkeypatch.setenv("GOOGLE_API_KEY", "AIzaSy_REAL_KEY")
+        with patch("agent.gemini_native_adapter.GeminiNativeClient") as mock_client, \
+             patch("run_agent.OpenAI") as mock_openai, \
+             patch("run_agent.ContextCompressor") as mock_compressor:
+            mock_openai.return_value = MagicMock()
+            mock_compressor.return_value = MagicMock(context_length=128000, threshold_tokens=64000)
+            from run_agent import AIAgent
+            AIAgent(
+                model="gemini-2.5-flash",
+                provider="gemini",
+                api_key="AIzaSy_REAL_KEY",
+                base_url="https://proxy.example.com/v1",
+            )
+        mock_openai.assert_called_once()
+
+    def test_gemini_openai_compat_base_url_keeps_openai_client(self, monkeypatch):
+        monkeypatch.setenv("GOOGLE_API_KEY", "AIzaSy_REAL_KEY")
+        with patch("agent.gemini_native_adapter.GeminiNativeClient") as mock_client, \
+             patch("run_agent.OpenAI") as mock_openai, \
+             patch("run_agent.ContextCompressor") as mock_compressor:
+            mock_openai.return_value = MagicMock()
+            mock_compressor.return_value = MagicMock(context_length=1048576, threshold_tokens=524288)
+            from run_agent import AIAgent
+            AIAgent(
+                model="gemini-2.5-flash",
+                provider="gemini",
+                api_key="AIzaSy_REAL_KEY",
+                base_url="https://generativelanguage.googleapis.com/v1beta/openai",
+            )
+        mock_openai.assert_called_once()
+
     def test_gemini_resolve_provider_client_uses_native_client(self, monkeypatch):
         """resolve_provider_client('gemini') should build GeminiNativeClient."""
         monkeypatch.setenv("GEMINI_API_KEY", "AIzaSy_TEST_KEY")
@@ -233,6 +267,16 @@ class TestGeminiAgentInit:
         assert mock_client.called
         mock_openai.assert_not_called()
 
+    def test_gemini_resolve_provider_client_keeps_openai_for_non_native_base_url(self, monkeypatch):
+        monkeypatch.setenv("GOOGLE_API_KEY", "AIzaSy_TEST_KEY")
+        monkeypatch.setenv("GEMINI_BASE_URL", "https://proxy.example.com/v1")
+        with patch("agent.gemini_native_adapter.GeminiNativeClient") as mock_client, \
+             patch("agent.auxiliary_client.OpenAI") as mock_openai:
+            mock_openai.return_value = MagicMock()
+            from agent.auxiliary_client import resolve_provider_client
+            resolve_provider_client("gemini")
+        mock_openai.assert_called_once()
+
 
 # ── models.dev Integration ──