fix(gateway): pass through upstream token usage in streams for all adapters

The prior usage-passthrough fix only covered openai/opencode; the same
empty-choices+usage drop bug remained in the 5 sibling OpenAI-compatible
adapters, and non-OpenAI providers (anthropic/gemini/ollama) never surfaced
streaming usage at all.

- deepseek/github/groq/kimicode/mistral: preserve usage on empty-choices
  chunks and attach it to normal chunks (same pattern as openai.lua)
- anthropic: emit usage from message_start (prompt) and message_delta
  (completion); gateway merges split usage additively
- gemini: read usageMetadata in the stream path
- ollama: fix non-streaming key (usage -> token_usage, matches
  UnifiedResponse json tag) and read prompt_eval_count/eval_count;
  surface counts from the done stream chunk
- gateway: mergeUsage combines usage across chunks (non-zero fields win,
  total recomputed from prompt+completion) so split usage doesn't lose
  the prompt half; single-chunk case (OpenAI) preserved exactly
- usage-only chunks: done=false (no redundant terminal stop), matching
  the Go fallback standardSSEChunk

.gitignore

@@ -17,3 +17,6 @@ cmd/gui/*.log
 .pi-glla/
 .codegraph/
 /build/
+
+# local ops scripts (machine-specific, not for sharing)
+scripts/

Makefile

@@ -9,6 +9,12 @@ export GUI_DIR
 GUI_DIR=cmd/gui
 VERSION ?= $(shell git describe --tags --dirty 2>/dev/null || echo "0.1.0")
 
+# Clean env for GUI packaging: the host may have LD_LIBRARY_PATH polluted by
+# third-party runtimes (e.g. /opt/cangjie) which electron-builder bakes into
+# the produced binary's DT_RPATH/dep resolution -> the packaged app then
+# SIGSEGVs inside ld.so on clean machines.  Strip it for every pack step.
+CLEAN_PACK_ENV = env -u LD_LIBRARY_PATH
+
 all: build
 
 # ---- core binary (server users / embedded core) ----
@@ -26,23 +32,25 @@ gui-dev:
 
 # build the embedded core into the GUI, then run electron-builder
 # targets: deb + AppImage (linux), rpm needs system rpmbuild
+# NOTE: run in a clean env (no LD_LIBRARY_PATH pollution) so the packaged
+# binary does not link against host-only libs and crash on user machines.
 gui-dist:
-	@cd $(GUI_DIR) && npm run dist:linux
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:linux
 
 # deb only (for local sharing), rpm needs rpmbuild
 gui-deb:
-	@cd $(GUI_DIR) && npm run dist:debian
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:debian
 
 # windows nsis installer: host-built core (needs apt install gcc-mingw-w64-x86-64 + wine)
 gui-win:
-	@cd $(GUI_DIR) && npm run dist:win
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:win
 
 # windows nsis installer: full cross-build in docker (self-sufficient, no host mingw)
 gui-win-docker:
-	@cd $(GUI_DIR) && ./scripts/dist-win-docker.sh
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) ./scripts/dist-win-docker.sh
 
 gui-dist-dir:
-	@cd $(GUI_DIR) && npm run dist:dir
+	@cd $(GUI_DIR) && $(CLEAN_PACK_ENV) npm run dist:dir
 
 test:
 	go test ./...

README.md

@@ -89,8 +89,7 @@ GOMODCACHE=... GOPROXY=off go build -tags luajit -o llmsproxy ./cmd/llmsproxy
 
 > **配置不入库**：仓库不携带任何 config 文件（配置文件含密钥）。二进制首次运行时
 > 会在 `-config` 指定的位置生成默认配置：随机 admin key（打印在启动日志中）、
-> 仅绑定 `127.0.0.1:8080`、内置一个免 key 的 opencode zen 源可直接对话。
+> 仅绑定 `127.0.0.1:8080`。首次登录后请在 WebUI「密钥」页更换管理员密钥。
-> 首次登录后请在 WebUI「密钥」页更换管理员密钥。
 
 ```bash
 # 无 key -> 401
@@ -281,19 +280,12 @@ return {
 
 ### 内置适配器
 
-`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`
+`openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama` `kimicode`。
-`opencode`。
 
 `anthropic`/`gemini`/`ollama` 适配器内置多模态转换（`image_url` → 各自上游格式）；若
 源启用 `disable_thinking`，`deepseek` 适配器会把 `extra_body.thinking.type` 置为
 `disabled`。
 
-**opencode** 适配器面向 opencode.ai zen 免费池（`https://opencode.ai/zen/v1`）：
-zen 按 User-Agent 指纹识别官方客户端并分流——非官方 UA 的请求（curl、Go 默认 UA）
-会被丢进匿名池触发 `FreeUsageLimitError`。该适配器固定发送 opencode 客户端 UA，
-配合 `api_key: "public"`（官方无 key 客户端实际发送 `Bearer public`）即可走免费池，
-零成本接入 `deepseek-v4-flash-free` 等免费模型。
-
 **kimicode** 是展示 `build_headers` 的样例：云端校验调用方 app，需要按
 `meta.app_secret` 对时间戳+URL+请求体哈希做 HMAC 签名并附 `X-App-Sign` 等头。
 配好 `sources[].meta.{app_id, app_secret, app_agent}` 即可。

README_EN.md

@@ -77,7 +77,7 @@ GOMODCACHE=... GOPROXY=off go build -tags luajit -o llmsproxy ./cmd/llmsproxy
 > **No config is shipped in the repo** (config files carry real keys). On
 > first run the binary generates a default config at the `-config` path: a
 > random admin key (printed to the startup log), loopback-only `127.0.0.1:8080`,
-> and a keyless opencode zen source ready to chat. Rotate the admin key in the
+> and a keyless source ready to chat. Rotate the admin key in the
 > WebUI after first login.
 
 ```bash
@@ -289,19 +289,12 @@ Shared helpers: `hmac_sha256_hex(key, data)`, `sha256_hex(data)`,
 ### Built-in adapters
 
 `openai` `deepseek` `anthropic` `gemini` `github` `groq` `mistral` `ollama`
-`kimicode` `opencode`.
+`kimicode`.
 
 `anthropic`/`gemini`/`ollama` include multimodal conversion
 (`image_url` → their native format); with `disable_thinking` the `deepseek`
 adapter sets `extra_body.thinking.type` to `disabled`.
 
-**opencode** targets the opencode.ai zen free pool
-(`https://opencode.ai/zen/v1`): zen fingerprints clients by User-Agent and
-routes non-official UAs (curl, Go's default) into an anonymous pool that
-hits `FreeUsageLimitError`. The adapter always sends the opencode client UA;
-combined with `api_key: "public"` (the keyless official client actually sends
-`Bearer public`) it gets the free pool, e.g. `deepseek-v4-flash-free`.
-
 **kimicode** demonstrates `build_headers`: the cloud validates the calling
 app, so you HMAC-sign timestamp+URL+body with `meta.app_secret` and add
 `X-App-Sign`-style headers. Configure `sources[].meta.{app_id, app_secret,

cmd/gui/main.js

@@ -10,6 +10,7 @@ const {
   Tray,
   shell,
   nativeImage,
+  nativeTheme,
 } = require("electron");
 const path = require("path");
 const fs = require("fs");
@@ -46,6 +47,15 @@ const CORE_EXE = path.join(
   "bin",
   process.platform === "win32" ? "llmsproxy.exe" : "llmsproxy",
 );
+// Packaged window icon: must be a REAL file on disk (extraResource), not an
+// asar path — Electron/GTK pulls BrowserWindow icons from the filesystem and
+// crashes (SIGSEGV) when given an asar virtual path on Linux.
+// Dev:  cmd/gui/build/icon.png.
+const WINDOW_ICON = path.join(
+  app.isPackaged ? process.resourcesPath : __dirname,
+  "build",
+  "icon.png",
+);
 // default port for the embedded core (configurable via settings)
 const DEFAULT_PORT = 8787;
 
@@ -53,6 +63,7 @@ let mainWindow = null;
 let tray = null;
 let coreProc = null;
 let coreStartedAt = 0;
+let coreReady = false; // true only after the embedded core actually answers HTTP
 let settings = {}; // { port, autoStart, minimizeToTray }
 let authRules = [];
 
@@ -120,9 +131,9 @@ function coreStarted() {
 
 function writeDefaultEmbeddedConfig() {
   // Generate a self-contained embedded profile with a fresh random admin key.
-  // (Like Clash Verge generates its profile; the core's EnsureDefault default
+  // No upstream sources are preconfigured — the desktop user adds them via
-  // is 127.0.0.1:8080 — we want our own port + key on first run.) If a config
+  // the WebUI's sources page (like Clash Verge's empty-by-default profile).
-  // already exists we do NOT overwrite it (keep user's edits).
+  // If a config already exists we do NOT overwrite it (keep user's edits).
   if (fs.existsSync(CONFIG_FILE)) return;
   const key = "sk-gw-" + crypto.randomBytes(16).toString("hex");
   const yaml = [
@@ -132,16 +143,6 @@ function writeDefaultEmbeddedConfig() {
     "adapter_dir: " + JSON.stringify(path.join(PROFILE_DIR, "adapters")),
     "runtime_file: " + JSON.stringify(path.join(PROFILE_DIR, "runtime.json")),
     "",
-    "sources:",
-    "  - name: zen",
-    "    base_url: https://opencode.ai/zen/v1",
-    "    api_key: public",
-    "    adapter: opencode",
-    "    models:",
-    "      - id: deepseek-v4-flash-free",
-    "        priority: 100",
-    "        kind: chat",
-    "",
     // Admin key as a plain `keys` entry (no seed flag) so the web UI never
     // shows the replace-the-initial-key warning. gateway_keys entries are
     // marked seed:true by the core.
@@ -211,6 +212,7 @@ function startCore() {
   coreProc.on("exit", (code) => {
     const wasManaged = coreStartedAt > 0;
     coreStartedAt = 0;
+    coreReady = false;
     coreProc = null;
     console.log("core exited:", code);
     // auto-restart if still running GUI (crash recovery, max every ~5s)
@@ -223,6 +225,12 @@ function startCore() {
       }, 1500);
     }
   });
+  // Broadcast starting (NOT ready) immediately. The renderer must not
+  // attempt to load the iframe until the core actually answers HTTP —
+  // otherwise the frame load races the listener and fails permanently.
+  // (Fix for blank-screen: running=true was previously broadcast while the
+  // core was still booting, so the iframe ERR_CONNECTION_REFUSED and the
+  // renderer never retried.)
   notifyCoreState();
   // pump until the core answers, then inject the admin key and tell the
   // renderer the web UI is ready (the iframe must load only after auth is set)
@@ -230,6 +238,7 @@ function startCore() {
     .then((ok) => {
       if (ok) {
         setEmbeddedAuth();
+        coreReady = true;
         notifyCoreState();
       }
     })
@@ -255,6 +264,7 @@ let coreLastExit = 0;
 
 function stopCore() {
   app.isQuitting = true;
+  coreReady = false;
   if (coreProc && coreProc.exitCode === null) {
     try {
       coreProc.kill();
@@ -357,17 +367,27 @@ function createTray() {
     );
   }
   tray = new Tray(img);
+  // Rebuild the context menu when the OS theme flips (dark/light) so menu
+  // text colors follow the desktop theme; nativeTheme is the Electron API
+  // that tracks GTK/appindicator dark mode on Linux.
+  const themeRebuild = () => {
+    try {
+      if (global.__rebuildTray) global.__rebuildTray();
+    } catch (e) {}
+  };
+  try {
+    nativeTheme.on("updated", themeRebuild);
+  } catch (e) {}
+  // Status rows used to be enabled:false, which GTK renders in a fixed light
+  // grey that never flips with the system dark/light theme — invisible on
+  // light desktops. They are now enabled (normal theme-aware fg) with a
+  // no-op click so they stay display-only but inherit the menu text color.
+  const statusRow = (text) => ({ label: text, enabled: true, click: () => {} });
   const rebuild = () => {
     const running = coreStarted();
     const menu = Menu.buildFromTemplate([
-      {
+      statusRow("ModelRouter Desktop"),
-        label: "ModelRouter Desktop",
+      statusRow(running ? "● 内核运行中" : "○ 内核已停止"),
-        enabled: false,
-      },
-      {
-        label: running ? "● 内核运行中" : "○ 内核已停止",
-        enabled: false,
-      },
       { type: "separator" },
       {
         label: "显示主窗口",
@@ -438,7 +458,8 @@ function createTray() {
 function notifyCoreState() {
   if (mainWindow && !mainWindow.isDestroyed()) {
     mainWindow.webContents.send("core:state", {
-      running: coreStarted(),
+      running: coreStarted() && coreReady,
+      ready: coreReady,
       startedAt: coreStartedAt,
     });
   }
@@ -496,6 +517,12 @@ function createWindow() {
   const menu = Menu.buildFromTemplate([]);
   Menu.setApplicationMenu(menu);
   mainWindow = new BrowserWindow({
+    // Window icon (taskbar/dock). Packaged: real file under resourcesPath (see
+    // WINDOW_ICON) — an asar path here makes GTK segfault on Linux.
+    icon: WINDOW_ICON,
+    // GNOME taskbar icons match the window's WM_CLASS (lowercase
+    // "modelrouter-gui") against StartupWMClass in the .desktop file.
+    backgroundColor: "#1f2937",
     width: 1280,
     height: 860,
     minWidth: 900,
@@ -560,7 +587,8 @@ ipcMain.handle(
 );
 
 ipcMain.handle("core:state", () => ({
-  running: coreStarted(),
+  running: coreStarted() && coreReady,
+  ready: coreReady,
   startedAt: coreStartedAt,
   baseUrl: embeddedBaseUrl(),
   port: settings.port || DEFAULT_PORT,

cmd/gui/package.json

@@ -22,7 +22,7 @@
   "build": {
     "appId": "com.modelrouter.gui",
     "productName": "ModelRouter",
-    "linux": {
+"linux": {
       "target": [
         "deb",
         "rpm",
@@ -32,10 +32,19 @@
       "icon": "build/icon.png",
       "maintainer": "ModelRouter",
       "synopsis": "ModelRouter Desktop",
+"desktop": {
+        "entry": {
+            "StartupWMClass": "modelrouter-gui"
+        }
+      },
       "extraResources": [
         {
           "from": "bin/llmsproxy",
           "to": "bin/llmsproxy"
+        },
+        {
+          "from": "build/icon.png",
+          "to": "build/icon.png"
         }
       ]
     },
@@ -44,7 +53,13 @@
     },
     "mac": {
       "target": "dmg",
-      "icon": "build/icon.png"
+      "icon": "build/icon.png",
+      "extraResources": [
+        {
+          "from": "build/icon.png",
+          "to": "build/icon.png"
+        }
+      ]
     },
     "win": {
       "target": "nsis",
@@ -57,6 +72,10 @@
         {
           "from": "bin/lua51.dll",
           "to": "bin/lua51.dll"
+        },
+        {
+          "from": "build/icon.png",
+          "to": "build/icon.png"
         }
       ]
     },

cmd/gui/renderer/app.js

@@ -5,6 +5,7 @@
 const state = {
   core: {
     running: false,
+    ready: false,
     baseUrl: "",
     port: 8787,
     profileDir: "",
@@ -18,7 +19,9 @@ const state = {
     minimizeToTray: true,
   },
   theme: localStorage.getItem("mr-theme") || "light",
-  loadedOnce: false,
+  loadedOnce: false, // iframe has been pointed at the web UI at least once
+  frameLoaded: false, // last loadFrame actually finished loading (did-finish-load)
+  loadTries: 0,
 };
 
 const $ = (s) => document.querySelector(s);
@@ -55,7 +58,11 @@ function renderCoreStatus() {
   const core = state.core;
   const dot = $("#conn-dot");
   dot.className = "status-dot" + (core.running ? " ok" : " bad");
-  dot.title = core.running ? "内核运行中" : "内核未运行";
+  dot.title = !core.running
+    ? "内核未运行"
+    : core.ready
+      ? "内核运行中"
+      : "内核启动中…";
 }
 
 function showOffline() {
@@ -70,11 +77,47 @@ function showLoading() {
   $("#console-loading").style.display = "flex";
 }
 
+// probe the core's web root directly (fetch from the renderer). Using
+// fetch() instead of relying on the iframe's own load means the core's
+// readiness is confirmed by an actual HTTP round-trip before we point the
+// iframe at it — this is what prevents the blank-screen race where the
+// frame load happens while the listener is still coming up.
+function coreReachable() {
+  return fetch(
+    state.core.baseUrl || "http://127.0.0.1:" + (state.core.port || 8787),
+    { method: "GET", redirect: "follow", cache: "no-store" },
+  )
+    .then(() => true)
+    .catch(() => false);
+}
+
+// point the iframe at the WebUI, but only after the core is reachable.
+// If the load fails (core still booting / raced), we retry with backoff
+// instead of giving up forever (the old code set loadedOnce blindly and
+// never retried -> permanent blank screen).
+let frameRetryTimer = null;
+function scheduleFrameLoad() {
+  if (frameRetryTimer) return; // already scheduled
+  frameRetryTimer = setTimeout(async () => {
+    frameRetryTimer = null;
+    if (!state.core.running) return; // core stopped meanwhile
+    if (!(await coreReachable())) {
+      // core not reachable yet: wait a bit and try again
+      state.loadTries += 1;
+      showLoading();
+      scheduleFrameLoad();
+      return;
+    }
+    loadFrame(state.core.baseUrl || "http://127.0.0.1:" + state.core.port);
+  }, state.loadTries === 0 ? 800 : Math.min(3000, 400 * state.loadTries));
+}
+
 function loadFrame(url) {
   const f = frame();
   f.style.display = "block";
   $("#offline").style.display = "none";
   $("#console-loading").style.display = "none";
+  if (state.loadedOnce && f.src === url) return; // already pointing there
   f.src = url;
   state.loadedOnce = true;
 }
@@ -82,8 +125,9 @@ function loadFrame(url) {
 async function ensureCore() {
   await refreshCore();
   if (state.core.running) {
-    loadFrame(state.core.baseUrl);
+    scheduleFrameLoad();
   } else {
+    state.loadedOnce = false;
     showOffline();
   }
 }
@@ -141,10 +185,21 @@ async function saveSettings() {
   toast("设置已保存");
   await refreshCore();
   if (state.core.running) {
-    loadFrame(state.core.baseUrl);
+    scheduleFrameLoad();
+  } else {
+    state.loadedOnce = false;
+    showOffline();
   }
 }
 
+// ===== theme =====
+function applyTheme() {
+  document.documentElement.dataset.theme = state.theme;
+  try {
+    localStorage.setItem("mr-theme", state.theme);
+  } catch (e) {}
+}
+
 // ===== init =====
 function init() {
   $("#tb-min").onclick = () => window.modelrouter.win.minimize();
@@ -171,15 +226,30 @@ function init() {
     showLoading();
     await window.modelrouter.core.start();
     await refreshCore();
-    if (state.core.running) loadFrame(state.core.baseUrl);
+    if (state.core.running) scheduleFrameLoad();
     else showOffline();
   };
-  frame().onload = () => refreshCore();
+  // when the iframe finishes loading, mark loaded and refresh core status so
+  // the conn-dot reflects the actual WebUI (not just the core reachability)
+  frame().onload = () => {
+    state.frameLoaded = true;
+    refreshCore();
+  };
+  // load failure (e.g. the frame was pointed at the core just before it
+  // finished booting) -> clear loadedOnce and let scheduleFrameLoad retry
+  frame().onerror = () => {
+    state.loadedOnce = false;
+    state.frameLoaded = false;
+    scheduleFrameLoad();
+  };
   window.modelrouter.core.onState((d) => {
     state.core = Object.assign({}, state.core, d);
     renderCoreStatus();
-    if (d.running && !state.loadedOnce) loadFrame(state.core.baseUrl);
+    if (d.running && !state.loadedOnce) scheduleFrameLoad();
-    if (!d.running) showOffline();
+    if (!d.running) {
+      state.loadedOnce = false;
+      showOffline();
+    }
   });
   // Escape closes overlay
   document.addEventListener("keydown", (e) => {
@@ -189,6 +259,7 @@ function init() {
 }
 
 async function boot() {
+  applyTheme();
   init();
   state.settings = await window.modelrouter.settings.get();
   renderRail();

internal/config/config.go

@@ -103,9 +103,10 @@ func EnsureDefault(path string) (bool, error) {
 }
 
 // writeDefaultConfig writes a minimal, safe-by-default config: loopback-only
-// listen, a fresh random admin key, and a no-key zen source that works out of
+// listen and a fresh random admin key. No upstream sources are preconfigured
-// the box. adapter_dir / runtime_file live next to the config file so the
+// (the repo ships none — add them via WebUI or config.yaml). adapter_dir /
-// binary works regardless of the working directory it is started from.
+// runtime_file live next to the config file so the binary works regardless of
+// the working directory it is started from.
 func writeDefaultConfig(path string) error {
 	key, err := NewGatewayKey()
 	if err != nil {
@@ -122,15 +123,6 @@ func writeDefaultConfig(path string) error {
 		DefaultModel: "AUTO",
 		AdapterDir:   filepath.Join(abs, "adapters"),
 		RuntimeFile:  filepath.Join(abs, "runtime.json"),
-		Sources: []Source{
-			{
-				Name:    "zen",
-				BaseURL: "https://opencode.ai/zen/v1",
-				APIKey:  "public", // zen 免费池：官方无 key 客户端实际发送 Bearer public
-				Adapter: "opencode",
-				Models:  []Model{{ID: "deepseek-v4-flash-free", Priority: 100, Kind: "chat"}},
-			},
-		},
 	}
 	out, err := yaml.Marshal(&cfg)
 	if err != nil {

internal/config/config_test.go

@@ -91,14 +91,8 @@ func TestEnsureDefaultGeneratesOnMissingFile(t *testing.T) {
 	if cfg.Listen != "127.0.0.1:8080" {
 		t.Fatalf("listen = %q, want loopback-only", cfg.Listen)
 	}
-	if len(cfg.Sources) != 1 || cfg.Sources[0].Name != "zen" || cfg.Sources[0].Adapter != "opencode" {
+	if len(cfg.Sources) != 0 {
-		t.Fatalf("default sources = %+v", cfg.Sources)
+		t.Fatalf("default config must ship no sources (user adds them), got %+v", cfg.Sources)
-	}
-	if cfg.Sources[0].APIKey != "public" {
-		t.Fatalf("zen api_key = %q", cfg.Sources[0].APIKey)
-	}
-	if cfg.Sources[0].Models[0].ID != "deepseek-v4-flash-free" {
-		t.Fatalf("default model = %+v", cfg.Sources[0].Models)
 	}
 	// adapter_dir / runtime_file resolve next to the config file
 	if !strings.HasPrefix(cfg.AdapterDir, dir) || !strings.HasPrefix(cfg.RuntimeFile, dir) {

internal/gateway/chat.go

@@ -53,11 +53,14 @@ type RespMessage struct {
 }
 
 type ChatChunk struct {
-	ID      string        `json:"id"`
+	ID      string            `json:"id"`
-	Object  string        `json:"object"`
+	Object  string            `json:"object"`
-	Created int64         `json:"created"`
+	Created int64             `json:"created"`
-	Model   string        `json:"model"`
+	Model   string            `json:"model"`
-	Choices []ChunkChoice `json:"choices"`
+	Choices []ChunkChoice     `json:"choices"`
+	// Usage is sent in the final chunk of a stream (empty choices) so
+	// OpenAI-compatible clients can read token usage.
+	Usage *types.TokenUsage `json:"usage,omitempty"`
 }
 
 type ChunkChoice struct {
@@ -577,6 +580,32 @@ func (g *Gateway) writeRec(rec *Req) {
 	g.stats.Record(*rec)
 }
 
+// mergeUsage combines token usage across stream chunks additively. Some
+// providers split usage across chunks (e.g. Anthropic reports prompt tokens
+// in message_start and the final completion tokens in message_delta); a plain
+// "last non-nil wins" would discard the prompt half. Non-zero fields from cur
+// override prev; total is recomputed from the merged parts so a partial later
+// chunk can't shrink it. For the common single-chunk case (OpenAI's terminal
+// empty-choices+usage chunk) upstream totals are preserved exactly.
+func mergeUsage(prev, cur *types.TokenUsage) *types.TokenUsage {
+	if prev == nil {
+		u := *cur
+		if u.Total == 0 && (u.Prompt > 0 || u.Completion > 0) {
+			u.Total = u.Prompt + u.Completion
+		}
+		return &u
+	}
+	out := *prev
+	if cur.Prompt > 0 {
+		out.Prompt = cur.Prompt
+	}
+	if cur.Completion > 0 {
+		out.Completion = cur.Completion
+	}
+	out.Total = out.Prompt + out.Completion
+	return &out
+}
+
 func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands []*provider.Provider, req *types.ChatRequest, effective string, rec *Req) {
 	rec.LatMs = 0
 	t0 := time.Now()
@@ -626,7 +655,11 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 	}) {
 		return
 	}
+	var lastUsage *types.TokenUsage
 	for ck := range chunks {
+		if ck.Usage != nil {
+			lastUsage = mergeUsage(lastUsage, ck.Usage)
+		}
 		chunk := ChatChunk{
 			ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
 		}
@@ -653,6 +686,27 @@ func (g *Gateway) streamChat(w http.ResponseWriter, ctx context.Context, cands [
 		ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
 		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &stop}},
 	})
+	// Final usage chunk (OpenAI standard: empty choices + usage before [DONE]).
+	// Prefer the upstream's exact usage if the stream carried it; fall back to
+	// the gateway's estimate otherwise.
+	var tut *types.TokenUsage
+	if lastUsage != nil {
+		tut = lastUsage
+	} else if rec.Prompt+rec.Compl > 0 {
+		u := types.TokenUsage{
+			Prompt:     int(rec.Prompt),
+			Completion: int(rec.Compl),
+			Total:      int(rec.Prompt + rec.Compl),
+		}
+		tut = &u
+	}
+	if tut != nil {
+		send(ChatChunk{
+			ID: id, Object: "chat.completion.chunk", Created: created, Model: effective,
+			Choices: []ChunkChoice{},
+			Usage:   tut,
+		})
+	}
 	fmt.Fprintf(w, "data: [DONE]\n\n")
 	if flusher != nil {
 		flusher.Flush()
@@ -771,7 +825,11 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 	}) {
 		return
 	}
+	var lastUsage *types.TokenUsage
 	for ck := range chunks {
+		if ck.Usage != nil {
+			lastUsage = mergeUsage(lastUsage, ck.Usage)
+		}
 		chunk := ChatChunk{
 			ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
 		}
@@ -798,6 +856,27 @@ func (g *Gateway) streamChatAuto(w http.ResponseWriter, ctx context.Context, cha
 		ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
 		Choices: []ChunkChoice{{Index: 0, Delta: RespMessage{}, FinishReason: &stop}},
 	})
+	// Final usage chunk (OpenAI standard: empty choices + usage before [DONE]).
+	// Prefer the upstream's exact usage if the stream carried it; fall back to
+	// the gateway's estimate otherwise.
+	var tut *types.TokenUsage
+	if lastUsage != nil {
+		tut = lastUsage
+	} else if rec.Prompt+rec.Compl > 0 {
+		u := types.TokenUsage{
+			Prompt:     int(rec.Prompt),
+			Completion: int(rec.Compl),
+			Total:      int(rec.Prompt + rec.Compl),
+		}
+		tut = &u
+	}
+	if tut != nil {
+		send(ChatChunk{
+			ID: id, Object: "chat.completion.chunk", Created: created, Model: rec.Model,
+			Choices: []ChunkChoice{},
+			Usage:   tut,
+		})
+	}
 	fmt.Fprintf(w, "data: [DONE]\n\n")
 	if flusher != nil {
 		flusher.Flush()

internal/gateway/server.go

@@ -343,10 +343,10 @@ func (g *Gateway) handleLoginAPI(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	http.SetCookie(w, &http.Cookie{
-		Name:   "gw_key",
+		Name:     "gw_key",
-		Value:  body.Key,
+		Value:    body.Key,
-		Path:   "/",
+		Path:     "/",
-		MaxAge: 86400 * 30,
+		MaxAge:   86400 * 30,
 		HttpOnly: true,
 		SameSite: http.SameSiteLaxMode,
 	})
@@ -456,7 +456,15 @@ func (g *Gateway) handleStatusAPI(w http.ResponseWriter, r *http.Request) {
 		"gateway_keys":  ks,
 	}
 	if reqRole(r.Context()) == "admin" {
-		resp["sources"] = g.core.Registry().Status()
+		sts := g.core.Registry().Status()
+		recent := g.stats.SourceRecent(300)
+		for i := range sts {
+			if v, ok := recent[sts[i].Name]; ok {
+				sts[i].RecentOK = v[0]
+				sts[i].RecentErr = v[1]
+			}
+		}
+		resp["sources"] = sts
 		resp["adapters"] = g.core.ListAdapters()
 	}
 	writeJSON(w, http.StatusOK, resp)

internal/gateway/stats.go

@@ -26,7 +26,7 @@ type Req struct {
 	Compl int64 `json:"completion_tokens"`
 	// LatMs total handling time ms
 	LatMs int64 `json:"latency_ms"`
-	OK    bool   `json:"ok"`
+	OK    bool  `json:"ok"`
 	// Status http status code
 	Status int `json:"status"`
 	// Err short error message
@@ -445,6 +445,33 @@ func (s *Stats) AuditRecords(from, to int64, key string) []Req {
 	return out
 }
 
+// SourceRecent counts real gateway requests per source within the last
+// window (unix seconds). It reads only the in-memory ring, so it is cheap and
+// reflects live traffic — used by /api/status so the source status column is
+// driven by what actually happens, not just a probe.
+func (s *Stats) SourceRecent(windowSec int64) map[string][2]int64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	cut := time.Now().Unix() - windowSec
+	out := map[string][2]int64{}
+	for _, r := range s.recs {
+		if r.Time/1000 < cut {
+			continue
+		}
+		if r.Source == "" {
+			continue
+		}
+		v := out[r.Source]
+		if r.OK {
+			v[0]++
+		} else {
+			v[1]++
+		}
+		out[r.Source] = v
+	}
+	return out
+}
+
 // Snapshot returns the whole dashboard payload; when key != "" the records
 // and aggregate views are restricted to that gateway key.
 func (s *Stats) Snapshot(limit int, key string) map[string]interface{} {

internal/gateway/ui/index.html

@@ -1393,6 +1393,63 @@
         return `${p(d.getMonth() + 1)}-${p(d.getDate())} ${p(d.getHours())}:${p(d.getMinutes())}:${p(d.getSeconds())}`;
       };
       let statsKeyF = ""; // active key filter for records ('' = all)
+
+      // ---- Source status column (refreshable) ----
+      // The column is driven by a combination of signals so a source that is
+      // actually serving traffic can never be shown as "down/backoff" just
+      // because /models probing got rate-limited by the upstream (common with
+      // free pools). Priority: real recent traffic > live probe > scheduling
+      // cooldown.
+      function srcStatusCell(x) {
+        const ok = x.recent_ok || 0;
+        const err = x.recent_err || 0;
+        const cooled = x.backoff_until && x.backoff_until * 1000 > Date.now();
+        const probeOK = !!x.live_available;
+        const lastErr = x.last_error ? esc(x.last_error) : "";
+        let cell;
+        if (ok + err > 0) {
+          // real traffic: show a success rate, probe only annotates
+          const rate = Math.round((ok * 100) / (ok + err));
+          const cls =
+            rate >= 95 ? "tag-green" : rate >= 50 ? "tag-amber" : "tag-red";
+          const tip = `最近流量 ${ok} 成功 / ${err} 失败 (探测: ${probeOK ? "可达" : "探测失败"} ${cooled ? "· 冷却中" : ""})`;
+          cell = `<span class="tag ${cls}" title="${escAttr(tip)}"><i class="net-dot"></i>${rate}%（${ok}/${ok + err}）${cooled ? "·冷却" : ""}</span>`;
+        } else if (probeOK) {
+          // no recent traffic but probe alive
+          const tip = `探测正常 ${cooled ? "· 调度冷却中（可自动恢复）" : ""}`;
+          cell = `<span class="tag tag-green" title="${escAttr(tip)}"><i class="net-dot"></i>${t("online")}${cooled ? "·冷却" : ""}</span>`;
+        } else if (cooled) {
+          cell = `<span class="tag tag-amber" title="${escAttr("调度冷却中（不影响探测）" + (lastErr ? " " + lastErr : ""))}"><i class="net-dot"></i>冷却中</span>`;
+        } else {
+          cell = `<span class="tag tag-red" title="${escAttr(lastErr || "探测失败")}"><i class="net-dot"></i>${t("offline")}</span>`;
+        }
+        return cell;
+      }
+
+      // Re-fetch /api/status and repaint only the source status column. Called
+      // from the 5s stats loop and the manual refresh button so the column is
+      // never frozen at the first render.
+      async function paintSourceStatus() {
+        const tb = $("#src-body");
+        if (!tb) return;
+        try {
+          const s = await api("/api/status");
+          const srcs = s.sources || [];
+          tb.innerHTML = srcs
+            .map(
+              (x) =>
+                `<tr><td>${srcStatusCell(x)}</td>
+     <td><b>${esc(x.name)}</b></td><td>${esc(x.adapter)}</td>
+     <td><div class="src-models">${(x.models || []).map((m) => `<span class="tag tag-blue modelchip" onclick="showModelConfig('${escAttr(x.name)}','${escAttr(m)}')">${esc(m)}</span>`).join("")}</div></td>
+     <td><span class="muted">${esc(x.base_url || "")}</span></td>
+     <td>${x.max_concurrent}</td></tr>`,
+            )
+            .join("");
+        } catch (e) {
+          // status fetch failing (e.g. session) is surfaced by other loops
+        }
+      }
+
       async function renderStatus() {
         const s = await api("/api/status");
         const base = (window._base = s.base_url || location.origin + "/v1");
@@ -1402,24 +1459,16 @@
         const pane = $("#tab-status");
         if (pane && pane.dataset.built) {
           if (statsTimerId) clearInterval(statsTimerId);
+          await paintSourceStatus();
           await paintStats();
           statsTimerId = setInterval(() => {
-            if (!document.hidden) paintStats();
+            if (!document.hidden) {
+              paintStats();
+              paintSourceStatus();
+            }
           }, 5000);
           return;
         }
-        const srcRows = s.sources
-          ? s.sources
-              .map(
-                (x) =>
-                  `<tr><td>${x.live_available ? `<span class="tag tag-green"><i class="net-dot"></i>${t("online")}</span>` : `<span class="tag tag-red" title="${esc(x.last_error || "")}"><i class="net-dot"></i>${t("offline")}</span>`}</td>
-     <td><b>${esc(x.name)}</b></td><td>${esc(x.adapter)}</td>
-     <td><div class="src-models">${x.models.map((m) => `<span class="tag tag-blue modelchip" onclick="showModelConfig('${escAttr(x.name)}','${escAttr(m)}')">${esc(m)}</span>`).join("")}</div></td>
-     <td><span class="muted">${esc(x.base_url || "")}</span></td>
-     <td>${x.max_concurrent}</td></tr>`,
-              )
-              .join("")
-          : "";
         pane.innerHTML = `
     <div class="kpis" id="kpi-row"><div class="kpi-skeletons" aria-hidden="true">${Array(
       5,
@@ -1441,7 +1490,8 @@
     ${
       s.sources
         ? `<div class="card"><h2>${t("srcTitle")} (${s.sources.length})</h2>
-      <div class="tbl-wrap"><table><tr><th>${t("tConn")}</th><th>${t("tName")}</th><th>${t("tAdapter")}</th><th>${t("tModels")}</th><th>${t("tURL")}</th><th>${t("tConc")}</th></tr>${srcRows || `<tr><td colspan="6" class="empty">${t("srcEmpty")}</td></tr>`}</table></div>
+      <div class="tbl-wrap"><table><tr><th>${t("tConn")}</th><th>${t("tName")}</th><th>${t("tAdapter")}</th><th>${t("tModels")}</th><th>${t("tURL")}</th><th>${t("tConc")}</th></tr><tbody id="src-body"></tbody></table></div>
+      <p class="muted" style="padding:6px 2px 0"><button class="ghost small" onclick="paintSourceStatus()">${t("refreshSrc") || "刷新源状态"}</button> 状态列每 5 秒自动刷新</p>
     </div>`
         : ""
     }
@@ -1475,9 +1525,13 @@
           )
           .join("");
         if (statsTimerId) clearInterval(statsTimerId);
+        await paintSourceStatus();
         await paintStats();
         statsTimerId = setInterval(() => {
-          if (!document.hidden) paintStats();
+          if (!document.hidden) {
+            paintStats();
+            paintSourceStatus();
+          }
         }, 5000);
       }
       function renderKeySelect(keys, keyNames) {

internal/lua/adapters/anthropic.lua

@@ -99,9 +99,37 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if chunk.type == "message_start" then return "" end
+    if chunk.type == "message_start" then
+        local uses = nil
+        if chunk.message and type(chunk.message.usage) == "table" then
+            local u = chunk.message.usage
+            local p = u.input_tokens or 0
+            local c = u.output_tokens or 0
+            if p > 0 or c > 0 then
+                uses = { prompt = p, completion = c, total = p + c }
+            end
+        end
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     if chunk.type == "message_delta" then
-        return json.encode({ content = "", done = (chunk.delta and chunk.delta.stop_reason ~= nil) })
+        local uses = nil
+        if type(chunk.usage) == "table" then
+            local u = chunk.usage
+            local p = u.input_tokens or 0
+            local c = u.output_tokens or 0
+            if p > 0 or c > 0 then
+                uses = { prompt = p, completion = c, total = p + c }
+            end
+        end
+        local done = (chunk.delta and chunk.delta.stop_reason ~= nil)
+        if uses ~= nil then
+            -- completion is final here; prompt is merged from message_start
+            return json.encode({ content = "", done = done, usage = uses })
+        end
+        return json.encode({ content = "", done = done })
     end
     if chunk.type == "content_block_start" and chunk.content_block
         and chunk.content_block.type == "tool_use" then

internal/lua/adapters/deepseek.lua

@@ -82,7 +82,27 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -90,6 +110,9 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end

internal/lua/adapters/gemini.lua

@@ -91,7 +91,24 @@ function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
 
-    if not chunk.candidates or #chunk.candidates == 0 then return "" end
+    -- Gemini attaches usageMetadata to the final chunk (alongside or after
+    -- candidates). Keys map to Go's TokenUsage json tags (prompt/...).
+    local uses = nil
+    if type(chunk.usageMetadata) == "table" then
+        local p = chunk.usageMetadata.promptTokenCount or 0
+        local c = chunk.usageMetadata.candidatesTokenCount or 0
+        local t = chunk.usageMetadata.totalTokenCount or 0
+        if p > 0 or c > 0 or t > 0 then
+            uses = { prompt = p, completion = c, total = t }
+        end
+    end
+
+    if not chunk.candidates or #chunk.candidates == 0 then
+        if uses ~= nil then
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local cand = chunk.candidates[1]
     local unified = { content = "", done = (cand.finishReason ~= nil) }
     local reasoning = ""
@@ -117,6 +134,9 @@ function adapter.transform_stream_chunk(raw_chunk)
     end
     if reasoning ~= "" then unified.reasoning_content = reasoning end
     if #tools > 0 then unified.tool_calls = tools end
+    if uses ~= nil then
+        unified.usage = uses
+    end
     return json.encode(unified)
 end
 

internal/lua/adapters/github.lua

@@ -63,7 +63,27 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -71,6 +91,9 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end

internal/lua/adapters/groq.lua

@@ -62,7 +62,27 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -70,6 +90,9 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end

internal/lua/adapters/kimicode.lua

@@ -95,7 +95,27 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -103,6 +123,9 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end

internal/lua/adapters/mistral.lua

@@ -62,7 +62,27 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -70,6 +90,9 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end

internal/lua/adapters/ollama.lua

@@ -53,11 +53,14 @@ function adapter.transform_response(raw_body)
     local ok, resp = pcall(json.decode, raw_body)
     if not ok then return raw_body end
 
+    local p = resp.prompt_eval_count or 0
+    local c = resp.eval_count or 0
     local unified = {
         content = "",
         finish_reason = resp.done_reason or "",
         tool_calls = {},
-        usage = { prompt = 0, completion = 0, total = 0 }
+        -- key must be token_usage to match Go's UnifiedResponse json tag
+        token_usage = { prompt = p, completion = c, total = p + c }
     }
 
     if resp.message then
@@ -70,12 +73,32 @@ end
 function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
-    if not chunk.message then return "" end
+
+    -- Ollama's terminal chunk (done=true) carries token counts but may omit
+    -- message; pass them through so the gateway emits real usage.
+    local uses = nil
+    if chunk.done then
+        local p = chunk.prompt_eval_count or 0
+        local c = chunk.eval_count or 0
+        if p > 0 or c > 0 then
+            uses = { prompt = p, completion = c, total = p + c }
+        end
+    end
+
+    if not chunk.message then
+        if uses ~= nil then
+            return json.encode({ content = "", done = true, usage = uses })
+        end
+        return ""
+    end
 
     local unified = {
         content = chunk.message.content or "",
         done = chunk.done or false
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if chunk.message.reasoning_content then
         unified.reasoning_content = chunk.message.reasoning_content
     end

internal/lua/adapters/openai.lua

@@ -67,7 +67,26 @@ function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
 
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -75,11 +94,13 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end
     if delta.tool_calls then
-        -- pass raw streaming fragments through; OpenAI clients accumulate index+id+name+arguments
         unified.tool_calls = delta.tool_calls
     end
     return json.encode(unified)

internal/lua/adapters/opencode.lua

@@ -112,7 +112,27 @@ function adapter.transform_stream_chunk(raw_chunk)
     local ok, chunk = pcall(json.decode, raw_chunk)
     if not ok then return "" end
 
-    if not chunk.choices or #chunk.choices == 0 then return "" end
+    -- OpenAI-style streams may attach usage to a chunk with empty choices
+    -- (the final usage chunk). Preserve it; the gateway emits it as the
+    -- terminal usage chunk. Note: keys must match Go's TokenUsage json tags
+    -- (prompt/completion/total); the gateway re-emits standard *_tokens.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        uses = {
+            prompt = chunk.usage.prompt_tokens or chunk.usage.prompt or 0,
+            completion = chunk.usage.completion_tokens or chunk.usage.completion or 0,
+            total = chunk.usage.total_tokens or chunk.usage.total or 0,
+        }
+    end
+
+    if not chunk.choices or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- usage-only chunk is not a content/finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
     local delta = chunk.choices[1].delta or {}
     local fr = chunk.choices[1].finish_reason
 
@@ -120,6 +140,9 @@ function adapter.transform_stream_chunk(raw_chunk)
         content = delta.content or "",
         done = (fr ~= nil)
     }
+    if uses ~= nil then
+        unified.usage = uses
+    end
     if delta.reasoning_content then
         unified.reasoning_content = delta.reasoning_content
     end

internal/lua/vm_test.go

@@ -135,6 +135,39 @@ func TestOpenCodeAdapterFingerprint(t *testing.T) {
 	}
 }
 
+func TestOpenCodeNormalizesRoles(t *testing.T) {
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+	// zen 上游只接受 system/user/assistant/tool/latest_reminder；
+	// OpenAI 系客户端发 `developer`（Claude Code 还会用 `function`）必须归一化成 system。
+	body := `{"model":"x","messages":[
+		{"role":"developer","content":"you are helpful"},
+		{"role":"system","content":"sys"},
+		{"role":"user","content":"hi"},
+		{"role":"assistant","content":"ok"},
+		{"role":"tool","tool_call_id":"t1","content":"{\"ok\":true}"}
+	]}`
+	out, err := vm.Transform("opencode", "transform_request", body)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(out, `"developer"`) {
+		t.Fatalf("developer role must be normalized: %s", out)
+	}
+	if !strings.Contains(out, `"system"`) || !strings.Contains(out, `"user"`) {
+		t.Fatalf("allowed roles must survive: %s", out)
+	}
+	if !strings.Contains(out, `"assistant"`) {
+		t.Fatalf("assistant role must survive: %s", out)
+	}
+	if !strings.Contains(out, `"tool"`) {
+		t.Fatalf("tool role must survive: %s", out)
+	}
+}
+
 func TestOpenCodeStripsMultiModalParts(t *testing.T) {
 	vm := NewVM(freshAdapterDir(t))
 	if err := vm.Start(); err != nil {

internal/provider/provider.go

@@ -901,20 +901,52 @@ func standardSSEChunk(data string) string {
 			} `json:"delta"`
 			FinishReason *string `json:"finish_reason"`
 		} `json:"choices"`
+		UpstreamUsage struct {
+			PromptTokens     int `json:"prompt_tokens"`
+			CompletionTokens int `json:"completion_tokens"`
+			TotalTokens      int `json:"total_tokens"`
+			Prompt           int `json:"prompt"`
+			Completion       int `json:"completion"`
+			Total            int `json:"total"`
+		} `json:"usage"`
 	}
 	if err := json.Unmarshal([]byte(data), &raw); err != nil {
 		return ""
 	}
-	if len(raw.Choices) == 0 {
+	if len(raw.Choices) == 0 && raw.UpstreamUsage.Total == 0 && raw.UpstreamUsage.TotalTokens == 0 {
 		return ""
 	}
+	var usage *types.TokenUsage
+	pu := raw.UpstreamUsage
+	if pu.Total > 0 || pu.TotalTokens > 0 {
+		usage = &types.TokenUsage{
+			Prompt:     pickFirst(pu.PromptTokens, pu.Prompt),
+			Completion: pickFirst(pu.CompletionTokens, pu.Completion),
+			Total:      pickFirst(pu.TotalTokens, pu.Total),
+		}
+	}
 	out, _ := json.Marshal(types.UnifiedChunk{
-		Content: raw.Choices[0].Delta.Content,
+		Content: func() string {
-		Done:    raw.Choices[0].FinishReason != nil,
+			if len(raw.Choices) > 0 {
+				return raw.Choices[0].Delta.Content
+			}
+			return ""
+		}(),
+		Done:  len(raw.Choices) > 0 && raw.Choices[0].FinishReason != nil,
+		Usage: usage,
 	})
 	return string(out)
 }
 
+// pickFirst returns a if non-zero, else b (for usage keys that may appear in
+// either standard *_tokens or legacy short form).
+func pickFirst(a, b int) int {
+	if a != 0 {
+		return a
+	}
+	return b
+}
+
 func truncate(s string, n int) string {
 	if len(s) <= n {
 		return s

internal/provider/registry.go

@@ -211,6 +211,12 @@ type SourceStatus struct {
 	FailCount     int      `json:"fail_count,omitempty"`
 	BackoffUntil  int64    `json:"backoff_until,omitempty"`
 	Permanent     bool     `json:"permanent,omitempty"`
+	// RecentOK / RecentErr count real gateway requests served by this source
+	// within the last 300s (drives the status column so a source that is
+	// actually serving traffic can never be shown as down just because a
+	// probe was rate-limited by the upstream).
+	RecentOK  int64 `json:"recent_ok,omitempty"`
+	RecentErr int64 `json:"recent_err,omitempty"`
 }
 
 // ProbeAll runs a live reachability check for every provider (in parallel).

internal/types/types.go

@@ -85,6 +85,28 @@ type TokenUsage struct {
 	Total      int `json:"total"`
 }
 
+// MarshalJSON emits both the legacy short keys (prompt/completion/total, used
+// by the internal unified representation and older clients) and the OpenAI
+// standard keys (prompt_tokens/completion_tokens/total_tokens). Standard
+// clients such as DSH and DevEco Code read the *_tokens fields.
+func (t TokenUsage) MarshalJSON() ([]byte, error) {
+	return json.Marshal(struct {
+		PromptTokens     int `json:"prompt_tokens"`
+		CompletionTokens int `json:"completion_tokens"`
+		TotalTokens      int `json:"total_tokens"`
+		Prompt           int `json:"prompt"`
+		Completion       int `json:"completion"`
+		Total            int `json:"total"`
+	}{
+		PromptTokens:     t.Prompt,
+		CompletionTokens: t.Completion,
+		TotalTokens:      t.Total,
+		Prompt:           t.Prompt,
+		Completion:       t.Completion,
+		Total:            t.Total,
+	})
+}
+
 type ImageData struct {
 	B64JSON string `json:"b64_json,omitempty"`
 	URL     string `json:"url,omitempty"`
@@ -116,6 +138,11 @@ type UnifiedChunk struct {
 	Done             bool            `json:"done"`
 	ToolCalls        json.RawMessage `json:"tool_calls,omitempty"`
 	ReasoningContent string          `json:"reasoning_content,omitempty"`
+	// Usage carries the upstream token usage when the stream chunk provides
+	// it (OpenAI-style streams attach usage to some chunks; the final one
+	// often has empty choices). Gateway uses it to emit exact usage in the
+	// final stream chunk instead of estimates.
+	Usage *TokenUsage `json:"usage,omitempty"`
 }
 
 // Meta passed to Lua build_headers hook