examples: Tier 2 claude-agent-service fallback

tests/test_examples_llm_extract.py

@@ -9,10 +9,14 @@ import httpx
 import pytest
 import respx
 
-from fire_planner.examples.llm_extract import extract_with_qwen
+from fire_planner.examples.llm_extract import (
+    extract_with_fallback,
+    extract_with_qwen,
+)
 from fire_planner.examples.models import RawPost
 
 LLAMA_URL = "http://llama-cpp.llama-cpp.svc.cluster.local:8000/v1/chat/completions"
+CLAUDE_URL = "http://claude-agent-service.claude-agent.svc.cluster.local:8080/v1/chat/completions"
 
 
 def _post() -> RawPost:
@@ -79,3 +83,91 @@ async def test_extract_with_qwen_returns_none_on_http_error() -> None:
         out = await extract_with_qwen(_post(), llama_url=LLAMA_URL, client=client)
 
     assert out is None
+
+
+@respx.mock
+@pytest.mark.asyncio
+async def test_fallback_escalates_when_qwen_returns_none() -> None:
+    respx.post(LLAMA_URL).respond(500)  # qwen down
+    claude_payload = {
+        "country": "Philippines",
+        "city": "Manila",
+        "confidence": 0.95,
+    }
+    respx.post(CLAUDE_URL).respond(
+        200,
+        json={"choices": [{"message": {"content": json.dumps(claude_payload)}}]},
+    )
+
+    async with httpx.AsyncClient() as client:
+        out = await extract_with_fallback(
+            _post(),
+            llama_url=LLAMA_URL,
+            claude_url=CLAUDE_URL,
+            claude_bearer="t",
+            client=client,
+        )
+
+    assert out is not None
+    assert out.llm_model == "claude-haiku-4-5"
+    assert out.country == "Philippines"
+
+
+@respx.mock
+@pytest.mark.asyncio
+async def test_fallback_escalates_on_low_confidence() -> None:
+    qwen_payload = {"country": None, "confidence": 0.2}
+    respx.post(LLAMA_URL).respond(
+        200,
+        json={"choices": [{"message": {"content": json.dumps(qwen_payload)}}]},
+    )
+    claude_payload = {"country": "Thailand", "city": "Bangkok", "confidence": 0.9}
+    respx.post(CLAUDE_URL).respond(
+        200,
+        json={"choices": [{"message": {"content": json.dumps(claude_payload)}}]},
+    )
+
+    async with httpx.AsyncClient() as client:
+        out = await extract_with_fallback(
+            _post(),
+            llama_url=LLAMA_URL,
+            claude_url=CLAUDE_URL,
+            claude_bearer="t",
+            client=client,
+            confidence_threshold=Decimal("0.5"),
+        )
+
+    assert out is not None
+    assert out.country == "Thailand"
+    assert out.llm_model == "claude-haiku-4-5"
+
+
+@respx.mock
+@pytest.mark.asyncio
+async def test_fallback_keeps_high_confidence_qwen_result() -> None:
+    payload = {
+        "country": "Philippines",
+        "confidence": 0.9,
+    }
+    respx.post(LLAMA_URL).respond(
+        200,
+        json={"choices": [{"message": {"content": json.dumps(payload)}}]},
+    )
+    claude_route = respx.post(CLAUDE_URL).respond(
+        200,
+        json={"choices": [{"message": {"content": "{}"}}]},
+    )
+
+    async with httpx.AsyncClient() as client:
+        out = await extract_with_fallback(
+            _post(),
+            llama_url=LLAMA_URL,
+            claude_url=CLAUDE_URL,
+            claude_bearer="t",
+            client=client,
+            confidence_threshold=Decimal("0.5"),
+        )
+
+    assert out is not None
+    assert out.llm_model == "qwen3-8b"
+    assert claude_route.called is False  # high-confidence qwen → claude not hit