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/agent/test_gemini_native_adapter.py

@@ -186,6 +186,43 @@ def test_native_http_error_keeps_status_and_retry_after():
     assert "quota exhausted" in str(err)
 
 
+def test_native_client_accepts_injected_http_client():
+    from agent.gemini_native_adapter import GeminiNativeClient
+
+    injected = SimpleNamespace(close=lambda: None)
+    client = GeminiNativeClient(api_key="AIza-test", http_client=injected)
+    assert client._http is injected
+
+
+@pytest.mark.asyncio
+async def test_async_native_client_streams_without_requiring_async_iterator_from_sync_client():
+    from agent.gemini_native_adapter import AsyncGeminiNativeClient
+
+    chunk = SimpleNamespace(choices=[SimpleNamespace(delta=SimpleNamespace(content="hi"), finish_reason=None)])
+    sync_stream = iter([chunk])
+
+    def _advance(iterator):
+        try:
+            return False, next(iterator)
+        except StopIteration:
+            return True, None
+
+    sync_client = SimpleNamespace(
+        api_key="AIza-test",
+        base_url="https://generativelanguage.googleapis.com/v1beta",
+        chat=SimpleNamespace(completions=SimpleNamespace(create=lambda **kwargs: sync_stream)),
+        _advance_stream_iterator=_advance,
+        close=lambda: None,
+    )
+
+    async_client = AsyncGeminiNativeClient(sync_client)
+    stream = await async_client.chat.completions.create(stream=True)
+    collected = []
+    async for item in stream:
+        collected.append(item)
+    assert collected == [chunk]
+
+
 def test_stream_event_translation_emits_tool_call_delta_with_stable_index():
     from agent.gemini_native_adapter import translate_stream_event
 
@@ -209,4 +246,30 @@ def test_stream_event_translation_emits_tool_call_delta_with_stable_index():
     assert first[0].choices[0].delta.tool_calls[0].index == 0
     assert second[0].choices[0].delta.tool_calls[0].index == 0
     assert first[0].choices[0].delta.tool_calls[0].id == second[0].choices[0].delta.tool_calls[0].id
+    assert first[0].choices[0].delta.tool_calls[0].function.arguments == '{"q": "abc"}'
+    assert second[0].choices[0].delta.tool_calls[0].function.arguments == ""
     assert first[-1].choices[0].finish_reason == "tool_calls"
+
+
+def test_stream_event_translation_keeps_identical_calls_in_distinct_parts():
+    from agent.gemini_native_adapter import translate_stream_event
+
+    event = {
+        "candidates": [
+            {
+                "content": {
+                    "parts": [
+                        {"functionCall": {"name": "search", "args": {"q": "abc"}}},
+                        {"functionCall": {"name": "search", "args": {"q": "abc"}}},
+                    ]
+                },
+                "finishReason": "STOP",
+            }
+        ]
+    }
+
+    chunks = translate_stream_event(event, model="gemini-2.5-flash", tool_call_indices={})
+    tool_chunks = [chunk for chunk in chunks if chunk.choices[0].delta.tool_calls]
+    assert tool_chunks[0].choices[0].delta.tool_calls[0].index == 0
+    assert tool_chunks[1].choices[0].delta.tool_calls[0].index == 1
+    assert tool_chunks[0].choices[0].delta.tool_calls[0].id != tool_chunks[1].choices[0].delta.tool_calls[0].id

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 ──