Merge pull request 'The call record: every model call, for the operator's study (M6.6a)' (#179) from call-record into main

nisse-dev/tests/test_bench.py

@@ -467,3 +467,26 @@ def test_runs_from_before_the_driver_read_as_its_defaults() -> None:
     old["tuning"]["executor"] = old["tuning"].pop("driver")
     [row] = summarise([old], current)
     assert row.tuning == ""
+
+
+def test_a_kept_run_leaves_its_state_and_call_record(tmp_path: Path) -> None:
+    # As the agent records itself (L7), so the status page can open a
+    # benchmark run (X6); each run gets its own record, and the client is
+    # left as it was.
+    from nisse_core.models.calls import CallReader
+
+    [task] = load_tasks(["write-function"])
+    client = Scripted(list(SOLVE_SLUGIFY))
+    record = run_task(task, FACTS, client, Tuning(), host=HOST, keep=tmp_path / "runs")
+
+    [kept] = list((tmp_path / "runs").iterdir())
+    assert kept.name.startswith("write-function-")
+    assert (kept / "state.db").exists()
+    [run] = CallReader(kept / "state.calls.db").runs()
+    assert (run["work_item"], run["role"], run["outcome"]) == (
+        "bench",
+        "driver",
+        "wait",
+    )
+    assert record.success
+    assert not hasattr(client, "call_recorder")