chore(models): refresh model catalog

crates/agent-gateway/test/webui/web-settings.test.mjs

@@ -6,6 +6,7 @@ const loader = createWebModuleLoader();
 const webSettings = loader.loadModule("src/lib/webSettings.ts");
 const settings = loader.loadModule("@/lib/settings/index.ts");
 const settingsSync = loader.loadModule("@liveagent/ui/lib/settings/sync.ts");
+const catalog = loader.loadModule("@liveagent/ui/lib/models/modelCatalog.ts");
 const chatHelpers = loader.loadModule("@/lib/chat/chatPageHelpers.ts");
 const adminApi = loader.loadModule("@/lib/adminApi.ts");
 const RIGHT_DOCK_TAB_IDS = settings.RIGHT_DOCK_SINGLETON_TAB_IDS;
@@ -1879,21 +1880,32 @@ test("xai model limits use the generated catalog without changing thinking detec
 });
 
 test("xai thinking levels come from the catalog per model, thinking always on", () => {
-  // 档位按型号差异化（目录真值）：grok-4.5 只有 low/medium/high；
-  // grok-4.20-multi-agent-0309 声明到 xhigh。xai 思考一律恒开
-  //（wire 无法表达 off），目录的 off 声明对 xai 供应商不生效。
-  assert.deepEqual(settings.getKnownModelThinkingLevels("xai", "grok-4.5"), [
-    "low",
-    "medium",
-    "high",
-  ]);
-  assert.equal(settings.isThinkingAlwaysOnForModel("xai", "grok-4.5"), true);
-  const multiAgent = settings.getKnownModelThinkingLevels("xai", "grok-4.20-multi-agent-0309");
-  assert.ok(multiAgent.includes("xhigh"));
-  assert.equal(settings.isThinkingAlwaysOnForModel("xai", "grok-4.3"), true);
-  // 钳制路径：xhigh 超出 grok-4.5 档位表时压回默认 high。
+  // 档位按型号取目录真值：逐条对照快照里每个 xai 条目，不写死档位表（目录每日
+  // 自动刷新）。单点对照分不出"按型号读目录"和"整个供应商一张固定表"，逐条
+  // 对照可以（快照里有档位为空的条目）。xai 思考一律恒开（wire 无法表达 off），
+  // 目录里 off: true 的条目对 xai 供应商同样不生效。
+  const entries = catalog.MODEL_CATALOG.xai.filter((model) => model.thinking);
+  assert.ok(entries.length >= 2, "xai thinking entries missing from the catalog snapshot");
+  for (const entry of entries) {
+    assert.deepEqual(
+      settings.getKnownModelThinkingLevels("xai", entry.id),
+      [...entry.thinking.levels],
+      `${entry.id} levels`,
+    );
+    assert.equal(
+      settings.isThinkingAlwaysOnForModel("xai", entry.id),
+      true,
+      `${entry.id} always on`,
+    );
+  }
+  // 钳制路径：表外档位（取档位阶梯里最高的表外档）压回默认 high。
+  const grok45Levels = settings.getKnownModelThinkingLevels("xai", "grok-4.5");
+  const outside = ["minimal", "low", "medium", "high", "xhigh", "max"]
+    .filter((level) => !grok45Levels.includes(level))
+    .at(-1);
+  assert.ok(outside, "grok-4.5 covers the whole level ladder; clamp needs another model");
   const clamped = settings.normalizeChatRuntimeControlsForProvider(
-    { reasoning: "xhigh", reasoningByProvider: { xai: "xhigh" } },
+    { reasoning: outside, reasoningByProvider: { xai: outside } },
     { providerId: "xai", modelId: "grok-4.5" },
   );
   assert.equal(clamped.reasoning, "high");

crates/agent-gui/test/models/model-catalog.test.mjs

@@ -172,7 +172,11 @@ test("normalizeModelIdCandidates strips a month-day date suffix only when no ful
 test("a dated relay id resolves to the catalog entry, an exact dated id keeps its own entry", () => {
   const relayed = catalog.findCatalogModelAcrossProviders("deepseek-ai/deepseek-v4-pro-0813");
   assert.equal(relayed?.id, "deepseek-v4-pro");
-  assert.equal(relayed?.contextWindow, 1000000);
+  // 限额对照目录快照本身，不写死数值：目录每日自动刷新会改动上游限额。
+  assert.deepEqual(
+    relayed,
+    catalog.MODEL_CATALOG.deepseek.find((model) => model.id === "deepseek-v4-pro"),
+  );
   assert.equal(catalog.findCatalogModelAcrossProviders("deepseek-r1-0528")?.id, "deepseek-r1-0528");
   assert.equal(catalog.findCatalogModelAcrossProviders("deepseek-r1-0601")?.id, "deepseek-r1");
 });

crates/agent-gui/test/settings/normalization.test.mjs

@@ -6,8 +6,21 @@ const loader = createTsModuleLoader();
 const settings = loader.loadModule("src/lib/settings/index.ts");
 const normalize = loader.loadModule("@liveagent/ui/lib/settings/normalize.ts");
 const sync = loader.loadModule("@liveagent/ui/lib/settings/sync.ts");
+const modelCatalog = loader.loadModule("@liveagent/ui/lib/models/modelCatalog.ts");
 const RIGHT_DOCK_TAB_IDS = settings.RIGHT_DOCK_SINGLETON_TAB_IDS;
 
+// 目录每日自动刷新会改动上游限额：断言"取到目录真值"的用例对照快照条目，
+// 不写死数值。
+function catalogLimits(section, modelId) {
+  const entry = modelCatalog.MODEL_CATALOG[section].find((model) => model.id === modelId);
+  assert.ok(entry, `${section}/${modelId} missing from the catalog snapshot`);
+  return { contextWindow: entry.contextWindow, maxOutputToken: entry.maxOutputToken };
+}
+
+function pickLimits(limits) {
+  return { contextWindow: limits.contextWindow, maxOutputToken: limits.maxOutputToken };
+}
+
 async function withNavigator(value, task) {
   const previous = Object.getOwnPropertyDescriptor(globalThis, "navigator");
   Object.defineProperty(globalThis, "navigator", {
@@ -315,8 +328,11 @@ test("DeepSeek provider normalization keeps native routing and native search", (
   assert.equal(provider.promptCachingEnabled, false);
   assert.equal(provider.promptCacheHintMode, undefined);
   assert.equal(provider.nativeWebSearchEnabled, true);
-  assert.equal(provider.models[0].contextWindow, 1_000_000);
-  assert.equal(provider.models[0].maxOutputToken, 393_216);
+  assert.deepEqual(pickLimits(provider.models[0]), catalogLimits("deepseek", "deepseek-v4-flash"));
+  assert.notDeepEqual(
+    pickLimits(provider.models[0]),
+    modelCatalog.PROVIDER_FALLBACK_LIMITS.deepseek,
+  );
 });
 
 test("legacy Codex-group DeepSeek configs stay untouched — migration is user-driven", () => {
@@ -3044,8 +3060,14 @@ test("cross-provider models resolve real catalog limits instead of provider fall
     "claude_code",
     "deepseek-v4-flash",
   );
-  assert.equal(deepseekUnderClaude.contextWindow, 1_000_000);
-  assert.equal(deepseekUnderClaude.maxOutputToken, 393_216);
+  assert.deepEqual(
+    pickLimits(deepseekUnderClaude),
+    catalogLimits("deepseek", "deepseek-v4-flash"),
+  );
+  assert.notDeepEqual(
+    pickLimits(deepseekUnderClaude),
+    modelCatalog.PROVIDER_FALLBACK_LIMITS.claude_code,
+  );
   const glmUnderCodex = settings.getProviderModelDefaults("codex", "glm-4.7");
   assert.equal(glmUnderCodex.contextWindow, 204_800);
   assert.equal(glmUnderCodex.maxOutputToken, 131_072);