Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio CI: guard the long-chat gate Glass ceilings against runner noise (#17012)

Daniel Han74b0cbe840a2
diff --git a/tests/studio/test_area_matrix.py b/tests/studio/test_area_matrix.py
index 88443e8c069..9275966a804 100644
--- a/tests/studio/test_area_matrix.py
+++ b/tests/studio/test_area_matrix.py
@@ -129,6 +129,43 @@ def test_appearance_check(base, mode, fails, monkeypatch):
assert [v["fail"] for v in rows] == [fails]
+# Highest readings on the GitHub runners the gate runs on (identical code spans about 1.35x
+# between runs; 25 runs, 2026-11-07).
+RUNNER_READINGS = [
+ ("blurred", "thread_scroll", 8.3),
+ (am.GLASS, "reasoning_toggle", 7.8),
+ ("blurred", "settings_scroll", 14.0),
+ ("frosted", "settings_scroll", 13.5),
+]
+
+
[email protected]("mode, area, ratio", RUNNER_READINGS)
+def test_shipped_ceilings_clear_runner_noise_but_not_a_real_regression(
+ mode, area, ratio, monkeypatch
+):
+ monkeypatch.setattr(am, "MAX_RATIO", 4.0)
+
+ def verdict(r):
+ pages = {
+ (am.SHORT, "default"): _page({area: {"elements": 10, "cpu_ms": 100}}),
+ (am.SHORT, mode): _page({area: {"elements": 10, "cpu_ms": 100 * r}}),
+ }
+ result = {**_result(pages), "areas": [area]}
+ [row] = [
+ v
+ for v in am.judge(result, known = am.KNOWN_DEFAULT)
+ if v["kind"] == "appearance" and v["metric"] == "cpu_ms" and v["mode"] == mode
+ ]
+ return row["fail"]
+
+ assert not verdict(
+ ratio
+ ), f"{mode}:{area} at {ratio}x, a reading main already gets, fails the shipped ceiling"
+ assert verdict(
+ ratio * 1.5
+ ), f"{mode}:{area} at 1.5x main's worst reading still passes; the ceiling catches nothing"
+
+
def test_an_area_that_stops_running_fails():
pages = {
(am.SHORT, "default"): _page({"a": {"elements": 10, "cpu_ms": 10}}),