Route live research through llm-connect and record pilot findings

Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a0cffe-206f-7323-89d3-b20ca4cf4bf9

tests/test_research_bench.py

@@ -154,3 +154,80 @@ def test_cli_refuses_fixture_output_on_real_source(tmp_path):
     assert result.returncode != 0
     assert "only run against synthetic" in result.stderr
     assert not (root / "reports/research-run.json").exists()
+
+
+@pytest.mark.parametrize('finish,content,complete', [
+    ('stop', '# Supported claim\n\nProvisional.', True),
+    ('length', '# Cut short', False),
+    ('stop', '', False),
+])
+def test_llm_connect_service_contract(tmp_path, finish, content, complete):
+    from llm_connect.models import LLMResponse
+    from llm_connect.server import LLMServer
+
+    calls = []
+
+    class ServiceAdapter:
+        def validate_config(self, config):
+            return True
+
+        def execute_prompt(self, prompt, config):
+            calls.append((prompt, config))
+            return LLMResponse(content=content, model='resolved/model',
+                               usage={'total_tokens': 42}, finish_reason=finish,
+                               metadata={'provider': 'test', 'response_id': 'r1',
+                                         'debug': 'DO-NOT-PERSIST'})
+
+    server = LLMServer(ServiceAdapter(), port=0)
+    server.start()
+    root = initialize(tmp_path)
+    try:
+        live = bench.LLMConnectAdapter(f'http://127.0.0.1:{server.port}', 'requested/model')
+        if complete:
+            assert bench.execute(root, live)['status'] == 'completed'
+            assert len(calls) == 4
+        else:
+            with pytest.raises(ValueError, match='empty or incomplete'):
+                bench.execute(root, live)
+            assert len(calls) == 1
+        receipt_text = (root / 'reports/research-run.json').read_text()
+        receipt = json.loads(receipt_text)
+        assert receipt['responses'][0]['model'] == 'resolved/model'
+        assert receipt['responses'][0]['usage']['total_tokens'] == 42
+        assert receipt['responses'][0]['finish_reason'] == finish
+        assert receipt['service_requests_attempted'] == len(calls)
+        assert calls[0][1].model_name == 'requested/model'
+        assert calls[0][1].max_tokens == 4000
+        assert 'DO-NOT-PERSIST' not in receipt_text
+    finally:
+        server.stop()
+
+
+def test_llm_connect_http_failure_is_safe_and_not_retried(tmp_path):
+    from llm_connect.exceptions import LLMAPIError
+    from llm_connect.server import LLMServer
+
+    calls = []
+
+    class FailingService:
+        def validate_config(self, config):
+            return True
+
+        def execute_prompt(self, prompt, config):
+            calls.append(prompt)
+            raise LLMAPIError('SECRET-SENTINEL', status_code=401,
+                              response_body='SECRET-SENTINEL')
+
+    server = LLMServer(FailingService(), port=0)
+    server.start()
+    root = initialize(tmp_path)
+    try:
+        live = bench.LLMConnectAdapter(f'http://127.0.0.1:{server.port}', 'test')
+        with pytest.raises(bench.InfospaceError) as raised:
+            bench.execute(root, live)
+        text = (root / 'reports/research-run.json').read_text()
+        assert len(calls) == 1
+        assert json.loads(text)['http_status'] >= 400
+        assert 'SECRET-SENTINEL' not in text + str(raised.value)
+    finally:
+        server.stop()