Reduce entity lookup context and configure reasoning effort

tests/test_llm_context_efficiency.py

@@ -0,0 +1,52 @@
+import asyncio
+import copy
+import json
+
+import pytest
+
+from test_entity_resolution import SETTINGS
+from test_llm_gateway import http_endpoint, request
+from viano_bug_analysis.analysis.entity_resolution import request_for, policy
+from viano_bug_analysis.llm import GatewayConfig, LLMError, LLMGateway, ProviderConfig
+
+
+@pytest.mark.parametrize("effort", [None, "low", "medium", "high"])
+def test_reasoning_and_output_limits_reach_transport(http_endpoint, effort):
+    url, seen, _ = http_endpoint
+    provider = ProviderConfig(id="test", model="fixture", base_url=url,
+                              reasoning_effort=effort, max_output_tokens=16384)
+    asyncio.run(LLMGateway(GatewayConfig((provider,))).complete_round(request()))
+    body = seen[0]["body"]
+    assert body[provider.token_parameter] == 16384
+    assert body.get("reasoning_effort") == effort
+    assert ("reasoning_effort" in body) == (effort is not None)
+    snapshot = GatewayConfig((provider,)).snapshot()["providers"][0]
+    assert ("reasoning_effort" in snapshot) == (effort is not None)
+
+
+@pytest.mark.parametrize("effort", [True, 1024, [], "unlimited"])
+def test_invalid_reasoning_configuration(effort):
+    with pytest.raises(LLMError, match="LLM_CONFIG"):
+        ProviderConfig(id="test", model="fixture", base_url="http://localhost", reasoning_effort=effort)
+
+
+@pytest.mark.parametrize("stage", ["plan", "assessment"])
+def test_identity_projection_keeps_readable_evidence_and_legacy_requests(stage):
+    context = {"sources": {
+        "O0001": {"kind":"observation", "data":{"text":"The switch is missing"}},
+        "F0001": {"kind":"fact", "data":{"name":"description", "value":"Switch operator"}},
+        "N0001": {"kind":"entity", "data":{"id":"42"}},
+        "X0001": {"kind":"unread_reference", "data":{"path":"raw/" + "x" * 10000}},
+        "F0002": {"kind":"fact_reference", "data":{"path":"raw/fact"}}}}
+    before = copy.deepcopy(context)
+    args = dict(plan={"searches":[]}, captures=[], candidates=[]) if stage == "assessment" else {}
+    old = request_for(stage, context, policy(SETTINGS), protocol_version=2, **args)
+    new = request_for(stage, context, policy(SETTINGS), protocol_version=3, **args)
+    old_data = json.loads(old["messages"][1]["content"])
+    data = json.loads(new["messages"][1]["content"])
+    assert old_data["context"] == context == before
+    assert set(data["context"]["sources"]) == {"O0001", "F0001", "N0001"}
+    for key, value in data["context"]["sources"].items():
+        assert value == context["sources"][key]
+    assert len(new["messages"][1]["content"]) < len(old["messages"][1]["content"]) / 2
+    assert data["output_schema"] == new["output_schema"]