feat: enable full AI automation and safe startup updates

tests/unit/test_llm_gateway.py

@@ -239,3 +239,57 @@ async def test_mid_stream_failure_records_stream_reset_payload(monkeypatch):
     assert payload["reason"] == "mid_stream_failure"
     assert "stream_exec_id" in payload
     assert "lower API priority" in payload["message"]
+
+
+@pytest.mark.asyncio
+async def test_configured_retry_budget_is_shared_by_identical_fallback(monkeypatch):
+    from unittest.mock import AsyncMock
+    from artemis.llm.reliability import retry_policy_for
+    from artemis.llm.router import ModelEndpoint
+
+    monkeypatch.setattr(llm_service, "retry_policy_for", retry_policy_for)
+    monkeypatch.setattr(llm_service.settings, "OPENAI_MAX_RETRIES", 5)
+    monkeypatch.setattr(llm_service.settings, "OPENAI_RETRY_INTERVAL_SECONDS", 10)
+    monkeypatch.setattr(llm_service, "_wait_for_resume", AsyncMock(return_value=False))
+    sleep = AsyncMock()
+    monkeypatch.setattr(llm_service.asyncio, "sleep", sleep)
+
+    class Fails:
+        def __init__(self):
+            self.calls = 0
+
+        async def ainvoke(self, *args, **kwargs):
+            self.calls += 1
+            if self.calls % 2:
+                raise ConnectionError("connection reset")
+            raise TimeoutError("timeout")
+
+    main, fallback = Fails(), Fails()
+    endpoint = ModelEndpoint(model_name="test", api_base="https://example.test/v1", max_retries=5)
+    primary = RobustChatModelWrapper(main, endpoint=endpoint)
+    secondary = RobustChatModelWrapper(fallback, endpoint=endpoint.model_copy())
+    with pytest.raises(LLMExhaustedError):
+        await with_fallback(lambda: primary.ainvoke([]), lambda: secondary.ainvoke([]))
+    assert main.calls == 6
+    assert fallback.calls == 0
+    assert [call.args[0] for call in sleep.call_args_list] == [10] * 5
+    assert llm_service._FAILED_ENDPOINTS.get() is None
+
+
+@pytest.mark.asyncio
+async def test_request_deadline_cancels_attempt_and_respects_zero_retries(monkeypatch):
+    import asyncio
+    from unittest.mock import AsyncMock
+
+    monkeypatch.setattr(llm_service, "_wait_for_resume", AsyncMock(return_value=False))
+    cancelled = asyncio.Event()
+
+    async def pending():
+        try:
+            await asyncio.Event().wait()
+        finally:
+            cancelled.set()
+
+    with pytest.raises(LLMExhaustedError):
+        await llm_service._run_with_recovery(pending, request_timeout=0.005, max_retries=0)
+    assert cancelled.is_set()