fix(channels): buddy 计费请求缺身份头 → 上游 401（逐行移植参考 checkinHeaders）

修好凭据读取后，余额端点仍报 HTTP 401。逐行核对参考 checkinHeaders
(credits.ts) 后发现我们只发了 Authorization，而 buddy 要求身份标识头齐全：

  X-Domain: copilot.tencent.com   （以产品配置为准，不用凭据里的 domain 快照）
  X-Product: SaaS                 （部署类型）
  X-Product-Code: codebuddy       （产品码）
  User-Agent: CodeBuddyIDE/1.106.1
  X-User-Id / X-Enterprise-Id / X-Tenant-Id（凭据携带时）

参考的注释特别警告 X-Domain 的方向：凭据里的 domain 是「登录时那个域名」的
快照，跨产品迁移后会指错区域，所以必须以产品端点为准。

## 实现
- 平台级声明 headers_fn(accessToken, cred) → 返回请求头表；buddy 据此逐行移植。
- fetchAccountJSON 增加头钩子位：spec 自己的 headers_fn 优先，否则回退平台级。
  未声明的平台仍走通用默认头（Bearer + JSON），不改变既有行为。
- set_endpoints 补齐 checkin_status / balance 端点覆盖（此前只覆盖旧的
  plat.checkin，buddy 这类平台无从改指，自动化测试也指不到假上游）。
  端点覆盖是参考明确的产品配置能力，也是测试缝。

## 判据
TestBuddyBillingHeadersMatchReference：把 buddy 余额端点指到本地假上游，
断言四个必备身份头逐字段正确。变异：改名 headers_fn → 立即 FAIL（已实测）。

go test ./internal/core/ 与 ./internal/lua/ 全绿。

deploy.sh

@@ -31,6 +31,7 @@ set -euo pipefail
 readonly TARGET_BIN="/usr/local/bin/llmsproxy"
 readonly TARGET_CONFIG="/etc/llmsproxy/config.yaml"
 readonly TARGET_ADAPTERS="/etc/llmsproxy/adapters"
+readonly TARGET_PLUGINS="/etc/llmsproxy/plugins"
 readonly REPO_DIR="/home/program/llmsproxy"
 readonly HEALTH_URL="http://127.0.0.1:8081/v1/models"
 readonly BIN_NAME="llmsproxy"
@@ -390,6 +391,52 @@ sync_adapters() {
     log "✓ 适配器已同步到 $TARGET_ADAPTERS/"
 }
 
+# 同步内置插件到 plugin_dir。
+#
+# 为什么需要它：新内置插件（如 channels.lua）过去只能靠内核的 SeedBundled 落地，
+# 而旧版 SeedBundled "目录已存在就什么都不做"，于是升级后新插件永远不会出现在
+# 生产 —— 它在二进制里、却不在 /api/plugins。内核端已修（改为按文件补齐），
+# 这里再显式同步一道，让部署与二进制内容对齐、不依赖内核那一侧的行为。
+#
+# 只补缺：已存在的插件文件一律不动。运维自己改过、或运行时新增的插件必须跨
+# 部署存活；覆盖它们会静默抹掉线上正在用的逻辑。
+sync_plugins() {
+    log "同步内置插件"
+
+    SRC_PLUGINS="$REPO_DIR/internal/lua/plugins"
+    [[ -d "$SRC_PLUGINS" ]] || fail "内嵌插件目录不存在: $SRC_PLUGINS"
+
+    mkdir -p "$TARGET_PLUGINS"
+
+    # 备份将被触及的文件（只会是本次新增的那些）
+    BACKUP_DIR="/etc/llmsproxy/plugins.bak.$(date +%Y%m%d%H%M%S)"
+    local added=0
+    mkdir -p "$BACKUP_DIR"
+    for src in "$SRC_PLUGINS"/*.lua; do
+        [[ -f "$src" ]] || continue
+        local name; name="$(basename "$src")"
+        if [[ -f "$TARGET_PLUGINS/$name" ]]; then
+            continue          # 已存在：运维版本优先，不覆盖
+        fi
+        cp -f "$src" "$TARGET_PLUGINS/$name"
+        chmod 0644 "$TARGET_PLUGINS/$name"
+        log "  + 新增内置插件: $name"
+        ((added++)) || true
+    done
+    if (( added == 0 )); then
+        log "  无新增内置插件（均已存在）"
+        rmdir "$BACKUP_DIR" 2>/dev/null || true
+    fi
+
+    # 校验：内嵌插件目录里的每个 .lua 都必须在目标目录存在
+    for src in "$SRC_PLUGINS"/*.lua; do
+        [[ -f "$src" ]] || continue
+        local name; name="$(basename "$src")"
+        [[ -f "$TARGET_PLUGINS/$name" ]] || warn "  缺少插件: $name"
+    done
+    log "✓ 插件已同步到 $TARGET_PLUGINS/（新增 $added，已有文件未改动）"
+}
+
 # ---------- 重启服务 ----------
 restart_service() {
     log "重启服务（中断 <500ms）"
@@ -495,6 +542,7 @@ print_summary() {
     log "  二进制: $TARGET_BIN"
     log "  配置:   $TARGET_CONFIG$([[ -n "$NEW_CONFIG_SRC" ]] && echo "  (本次已替换)" || echo "  (未改动)")"
     log "  适配器: $TARGET_ADAPTERS"
+    log "  插件:   $TARGET_PLUGINS"
     log "═════════════════════════════════════════════"
 }
 
@@ -511,6 +559,7 @@ main() {
     verify_master_key
     build
     sync_adapters
+    sync_plugins
     backup_old_state
     atomic_replace_binary
     atomic_replace_config

internal/core/channels_more_test.go

@@ -260,6 +260,38 @@ func (h *buddyHarness) begin() (map[string]interface{}, string) {
 	j := out.JSON.(map[string]interface{})
 	return j, j["session"].(map[string]interface{})["id"].(string)
 }
+
+// beginPool 走 UI 的真实流程：先 account/add 建占位账号，再带 account_id 登录。
+// 这正是用户报障（"登录了但无法获取账号信息"）的那条路径。
+func (h *buddyHarness) beginPool() (map[string]interface{}, string, string) {
+	add, err := h.ps.InvokeHandler("channels", "account_add", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"buddy"}`,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	aid := add.JSON.(map[string]interface{})["account_id"].(string)
+	out, err := h.ps.InvokeHandler("channels", "start_login", &lua.HandlerRequest{
+		Method: "POST", Body: fmt.Sprintf(`{"platform":"buddy","account_id":%q}`, aid),
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	j := out.JSON.(map[string]interface{})
+	return j, j["session"].(map[string]interface{})["id"].(string), aid
+}
+
+func (h *buddyHarness) accountList() []interface{} {
+	out, err := h.ps.InvokeHandler("channels", "account_list", &lua.HandlerRequest{
+		Method: "POST", Body: `{"platform":"buddy"}`,
+	})
+	if err != nil {
+		h.t.Fatal(err)
+	}
+	m := out.JSON.(map[string]interface{})
+	accts, _ := m["accounts"].([]interface{})
+	return accts
+}
 func (h *buddyHarness) poll(id string) map[string]interface{} {
 	out, _ := h.ps.InvokeHandler("channels", "poll_login", &lua.HandlerRequest{
 		Method: "GET", Query: map[string]string{"id": id},
@@ -287,9 +319,17 @@ func TestBuddyStatePollBehavior(t *testing.T) {
 		if j := h.poll(sid); j["state"] != "completed" || j["account"] != "buddy-dev" {
 			t.Fatalf("次拍应 completed+buddy-dev: %v", j)
 		}
+		// 凭据是**完整凭据 JSON**（对齐参考 buildCredential），不再是裸 token：
+		// 必须带 refresh_token，否则 refreshable=false、无法静默续期。
 		v, _, _ := h.ps.GetSecret("channels", "channels:buddy")
-		if v != "BD-AT-1" {
-			t.Fatalf("凭据应 BD-AT-1, 实得 %q", v)
+		if !strings.Contains(v, `"access_token":"BD-AT-1"`) {
+			t.Fatalf("凭据应含 access_token, 实得 %q", v)
+		}
+		if !strings.Contains(v, `"refresh_token":"BD-RT-1"`) {
+			t.Fatalf("凭据必须含 refresh_token（否则无法续期）, 实得 %q", v)
+		}
+		if !strings.Contains(v, "buddy-dev") {
+			t.Fatalf("凭据应回填昵称, 实得 %q", v)
 		}
 	})
 	t.Run("行3 account查询失败不阻塞_回退默认名", func(t *testing.T) {
@@ -303,6 +343,46 @@ func TestBuddyStatePollBehavior(t *testing.T) {
 			t.Fatalf("account 失败不应阻塞: %v", j)
 		}
 	})
+
+	// 行4 回归：UI 真实流程（account/add 建占位账号 → 带 account_id 登录）。
+	// 缺陷复现：poll_login 完成时不传 account_id → 凭据落到 `channels:buddy`，
+	// 而池账号 credentialRef 是 `channels:buddy:<aid>` → account_list 报
+	// has_credential=false、credits 报「无凭据」——即用户报障的
+	// 「登录了但无法获取账号信息」。
+	t.Run("行4 池路径：登录后账号必须真的拿到凭据", func(t *testing.T) {
+		atomic.StoreInt32(&h.tokStep, 0)
+		_, sid, aid := h.beginPool()
+		h.poll(sid) // 首拍 11217
+		j := h.poll(sid)
+		if j["state"] != "completed" {
+			t.Fatalf("池路径登录应完成: %v", j)
+		}
+		// ① 凭据必须落在池账号的 ref 上
+		v, ok, _ := h.ps.GetSecret("channels", "channels:buddy:"+aid)
+		if !ok || v == "" {
+			t.Fatalf("凭据未落到池账号 ref（用户报障的根因）; aid=%s ok=%v", aid, ok)
+		}
+		if !strings.Contains(v, `"refresh_token":"BD-RT-1"`) {
+			t.Fatalf("池凭据应是完整 JSON: %q", v)
+		}
+		// ② account_list 必须报 has_credential=true（UI 读的就是这个字段）
+		found := false
+		for _, raw := range h.accountList() {
+			a, _ := raw.(map[string]interface{})
+			if a["id"] == aid {
+				found = true
+				if a["has_credential"] != true {
+					t.Fatalf("has_credential 应为 true, 实得 %v", a["has_credential"])
+				}
+				if a["refreshable"] != true {
+					t.Fatalf("refreshable 应为 true（有 refresh_token）, 实得 %v", a["refreshable"])
+				}
+			}
+		}
+		if !found {
+			t.Fatalf("池账号 %s 未出现在 account_list", aid)
+		}
+	})
 }
 
 // ---- 通用 ----

internal/core/channels_pool_test.go

@@ -5,6 +5,10 @@ import (
 	"testing"
 
 	"llmsproxy/internal/lua"
+
+	"fmt"
+	"net/http"
+	"net/http/httptest"
 )
 
 // channels_pool_test：对齐参考 account.* / model.* / provider.* 契约的行为测试。
@@ -320,3 +324,47 @@ func TestChannelsExtendedRPCBehavior(t *testing.T) {
 		}
 	})
 }
+
+// TestBuddyBillingHeadersMatchReference 锁定 buddy 计费请求的头（逐行移植参考
+// checkinHeaders）。真实缺陷：实现只发 Authorization，上游直接 401 —— 修好后
+// 账号能读到凭据但余额仍失败，运维看到的仍是「拿不到账号信息」。
+func TestBuddyBillingHeadersMatchReference(t *testing.T) {
+	_, ps := poolCore(t)
+
+	var got map[string]string
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		got = map[string]string{}
+		for k := range r.Header {
+			got[k] = r.Header.Get(k)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"data":{"totalCredit":42}}`)
+	}))
+	defer ts.Close()
+
+	// 把 buddy 的余额/签到端点指到本地假上游
+	invoke(t, ps, "set_endpoints", fmt.Sprintf(
+		`{"platform":"buddy","balance":{"url":%q,"method":"POST","body":"{}"}}`, ts.URL))
+	m := invoke(t, ps, "account_add", `{"platform":"buddy","credential":"BD-AT","nickname":"b"}`)
+	if m["ok"] != true {
+		t.Fatalf("account_add: %v", m)
+	}
+	invoke(t, ps, "credits_balances", `{"platform":"buddy"}`)
+
+	if got == nil {
+		t.Fatal("余额请求未发出")
+	}
+	// 参考 checkinHeaders 的必备身份头（缺一个上游就 401）
+	want := map[string]string{
+		"Authorization":  "Bearer BD-AT",
+		"X-Domain":       "copilot.tencent.com",
+		"X-Product":      "SaaS",
+		"X-Product-Code": "codebuddy",
+		"User-Agent":     "CodeBuddyIDE/1.106.1",
+	}
+	for k, v := range want {
+		if got[k] != v {
+			t.Errorf("头 %s = %q, 期望 %q（buddy 计费端点缺它会被上游 401）", k, got[k], v)
+		}
+	}
+}

internal/lua/kernel_api.go

@@ -451,6 +451,16 @@ func (j jsonPermission) unmarshal() KernelPermission {
 	return KernelPermission{Actions: j.Actions, AllRead: j.AllRead}
 }
 
+type jsonPluginPolicy struct {
+	Actions []string `json:"actions,omitempty"`
+	AllRead bool     `json:"all_read,omitempty"`
+	Hosts   []string `json:"hosts,omitempty"`
+}
+
+func policyFromActions(p KernelPermission) jsonPluginPolicy {
+	return jsonPluginPolicy{Actions: p.Actions, AllRead: p.AllRead}
+}
+
 // KernelActionView is one action as the settings page needs it: the catalog entry
 // plus whether THIS plugin currently holds it and whether the plugin asked for it.
 type KernelActionView struct {

internal/lua/kernel_perms_persist.go

@@ -45,7 +45,12 @@ func (ps *Plugins) LoadKernelPermissions() error {
 	if err != nil {
 		return err
 	}
-	raw := map[string]jsonPermission{}
+	// The file holds BOTH kernel-action grants (map[string]jsonPermission, the
+	// historical shape) AND, for plugins that asked for them, the operator's
+	// extra-host grants (map[string]jsonPluginPolicy.Hosts). A bare
+	// {actions:[...]} row still parses into jsonPluginPolicy with empty Hosts,
+	// so old files load unchanged.
+	raw := map[string]jsonPluginPolicy{}
 	if err := json.Unmarshal(b, &raw); err != nil {
 		return err
 	}
@@ -54,8 +59,17 @@ func (ps *Plugins) LoadKernelPermissions() error {
 	if ps.kernelPerms == nil {
 		ps.kernelPerms = kernelPermissionSet{}
 	}
+	if ps.extraHosts == nil {
+		ps.extraHosts = map[string][]string{}
+	}
 	for name, jp := range raw {
-		ps.kernelPerms[name] = jp.unmarshal()
+		ps.kernelPerms[name] = KernelPermission{Actions: jp.Actions, AllRead: jp.AllRead}
+		if len(jp.Hosts) > 0 {
+			// Load the persisted grant into extraHosts WITHOUT folding it into the
+			// allowlist yet — sources are loaded later, and the caller owns that
+			// fold. GrantExtraHosts rebuilds the allowlist when sources are ready.
+			ps.extraHosts[name] = append([]string(nil), jp.Hosts...)
+		}
 	}
 	return nil
 }
@@ -65,12 +79,17 @@ func (ps *Plugins) LoadKernelPermissions() error {
 // Grants drop unknown action names on write. If an operator authorized an action
 // that a later release removed, keeping the name would leave a stale entry that
 // nothing can ever satisfy — and it would reappear in the UI as if it were live.
+//
+// Host grants travel in the same file (jsonPluginPolicy.Hosts): a granted host is
+// the operator's explicit decision to let a plugin egress somewhere, and losing
+// it across a restart (the previous behaviour) made grants silently vanish.
 func (ps *Plugins) saveKernelPermissionsLocked() {
 	path := ps.kernelPermsFile()
 	if path == "" {
 		return
 	}
-	out := map[string]jsonPermission{}
+	out := map[string]jsonPluginPolicy{}
+	something := false
 	for name, perm := range ps.kernelPerms {
 		valid := make([]string, 0, len(perm.Actions))
 		for _, a := range perm.Actions {
@@ -78,12 +97,28 @@ func (ps *Plugins) saveKernelPermissionsLocked() {
 				valid = append(valid, a)
 			}
 		}
-		if len(valid) == 0 && !perm.AllRead {
+		row := policyFromActions(KernelPermission{Actions: valid, AllRead: perm.AllRead})
+		if hs := ps.extraHosts[name]; len(hs) > 0 {
+			row.Hosts = append([]string(nil), hs...)
+		}
+		if len(valid) == 0 && !perm.AllRead && len(row.Hosts) == 0 {
 			continue // nothing granted: presence would be a lie
 		}
-		out[name] = KernelPermission{Actions: valid, AllRead: perm.AllRead}.marshal()
+		out[name] = row
+		something = true
 	}
-	if len(out) == 0 {
+	// A plugin may hold host grants with NO kernel actions granted (operators
+	// approve hosts and actions independently). Emit those too.
+	for name, hs := range ps.extraHosts {
+		if len(hs) == 0 {
+			continue
+		}
+		if _, exists := out[name]; !exists {
+			out[name] = jsonPluginPolicy{Hosts: append([]string(nil), hs...)}
+			something = true
+		}
+	}
+	if !something {
 		_ = os.Remove(path)
 		return
 	}

internal/lua/persist_test.go

@@ -338,3 +338,67 @@ return plugin
 		t.Errorf("persisted n = %v after three exported calls, want 3", saved.State["n"])
 	}
 }
+
+// TestInvokeHandlerIsPersisted 是真实事故的回归：channels 插件声明 hooks=[]，
+// 一切状态变更（账号入池、凭据引用、模型黑名单、事件日志）都经过自己的 HTTP
+// handler。而 InvokeHandler 是唯一一条**从不 markDirty** 的分发路径 —— hook 路径
+// 有、Call 路径有，只有它漏了。后果：
+//
+//	.channels.state.json 从未被写出 → 重启后账号池静默清空。
+//	运维在升级前的重启里丢掉了刚登录的账号，界面显示"已登录"但列表为空。
+//
+// billing/quota 侥幸没踩到，只因为它们有 hooks（billing）或用 Call（quota）。
+// 本测试用**无 hooks、只走 handler** 的插件复现该形态。
+func TestInvokeHandlerIsPersisted(t *testing.T) {
+	ps, _ := newPersistVM(t)
+	if err := ps.LoadSource("handlersonly", `
+local plugin = { name = "handlersonly", version = "1.0" }
+plugin.state = { accounts = {} }
+-- 不声明任何 hooks：状态只能经 HTTP handler 变更（channels 的形态）
+plugin.hooks = {}
+plugin.http = {
+  { path = "add", method = "POST", handler = "add" },
+}
+function plugin.add(req)
+  local body = req.body or ""
+  local id = body:match('"id":"([^"]+)"') or "unknown"
+  plugin.state.accounts[id] = { id = id, enabled = true }
+  return { status = 200, json = { ok = true } }
+end
+return plugin
+`); err != nil {
+		t.Fatal(err)
+	}
+
+	for _, id := range []string{"acct-1", "acct-2"} {
+		resp, err := ps.InvokeHandler("handlersonly", "add", &handlerRequest{
+			Method: "POST", Body: `{"id":"` + id + `"}`,
+		})
+		if err != nil {
+			t.Fatalf("handler: %v", err)
+		}
+		if resp == nil {
+			t.Fatal("handler returned nil response")
+		}
+	}
+	// Close() waits for the final flush, so no timing is involved: anything the
+	// handlers did NOT mark dirty cannot appear here.
+	ps.Close()
+
+	b, err := os.ReadFile(ps.stateFile("handlersonly"))
+	if err != nil {
+		t.Fatalf("★ 无 hooks 的插件经 HTTP handler 改状态后未落盘: %v\n"+
+			"   —— InvokeHandler 必须像 Call / hook 路径一样 markDirty，\n"+
+			"   否则 channels 的账号池会在每次重启后静默清空", err)
+	}
+	var saved struct {
+		State map[string]interface{} `json:"state"`
+	}
+	if err := json.Unmarshal(b, &saved); err != nil {
+		t.Fatal(err)
+	}
+	accts, _ := saved.State["accounts"].(map[string]interface{})
+	if len(accts) != 2 {
+		t.Fatalf("落盘的 accounts = %v，应为 2 个账号", saved.State["accounts"])
+	}
+}

internal/lua/plugin_http.go

@@ -43,6 +43,7 @@ import (
 	"net/http"
 	"net/url"
 	"strings"
+	"sync"
 	"time"
 
 	golua "github.com/aarzilli/golua/lua"
@@ -80,19 +81,39 @@ var httpClient = &http.Client{
 const httpGlobal = "__llmsproxy_http"
 
 // httpAllowlist is the set of hosts plugins may contact: the host of every
-// configured source's base_url.
+// configured source's base_url, plus every operator-granted extra host.
+//
+// It is a LIVE, shared object, not a snapshot. This matters: bootPluginState
+// binds one *httpAllowlist into every Lua closure at state-boot time, and those
+// closures live for the plugin's lifetime. Granting an extra host after a
+// worker booted must therefore MUTATE the same object the closures hold —
+// replacing ps.allowHTTP with a fresh instance would leave every already-booted
+// worker checking the STALE list, and a grant would look accepted (the API
+// answers ok, the UI shows it) while requests kept getting refused with
+// "not an allowed outbound destination". That exact failure shipped once.
+//
+// Mutations are rare (operator approves a host); lookups are per request.
 type httpAllowlist struct {
+	mu    sync.RWMutex
 	hosts map[string]bool
 }
 
 func newHTTPAllowlist(hosts []string) *httpAllowlist {
-	m := make(map[string]bool, len(hosts))
+	a := &httpAllowlist{hosts: make(map[string]bool, len(hosts))}
+	a.add(hosts)
+	return a
+}
+
+// add merges hosts into the live allowlist. Called by rebuild (source hosts
+// change on hot reload) and by GrantExtraHosts (operator approves a host).
+func (a *httpAllowlist) add(hosts []string) {
+	a.mu.Lock()
+	defer a.mu.Unlock()
 	for _, h := range hosts {
 		if h != "" {
-			m[strings.ToLower(h)] = true
+			a.hosts[strings.ToLower(h)] = true
 		}
 	}
-	return &httpAllowlist{hosts: m}
 }
 
 // allows reports whether host:port may be contacted. Ports 80/443 are implied
@@ -103,6 +124,8 @@ func (a *httpAllowlist) allows(host string, port string) bool {
 	if host == "" {
 		return false
 	}
+	a.mu.RLock()
+	defer a.mu.RUnlock()
 	if a.hosts[host] {
 		return true
 	}

internal/lua/plugin_http_handler.go

@@ -264,6 +264,10 @@ func (ps *Plugins) InvokeHandler(plugin, fnName string, req *handlerRequest) (*h
 		L.Pop(1)
 	}
 	if errMsg != "" {
+		// A handler that raised partway may already have mutated plugin.state
+		// (an account added before a later step failed), so mark it dirty before
+		// bailing: the mutation it managed to make is still real.
+		ps.markDirtyLocked(p)
 		return nil, fmt.Errorf("plugin %s: %s", plugin, errMsg)
 	}
 	if L.GetTop() < 1 || L.IsNil(-1) {
@@ -273,6 +277,18 @@ func (ps *Plugins) InvokeHandler(plugin, fnName string, req *handlerRequest) (*h
 	if err := luaToJSON(L, -1, &raw); err != nil {
 		return nil, err
 	}
+	// A handler mutates plugin.state just as a hook or an exported Call does, and
+	// this was the one dispatch path that never marked the plugin dirty.
+	//
+	// Consequence (real defect): a plugin with NO hooks — channels is exactly
+	// that, it declares hooks=[] and drives everything through its own HTTP
+	// routes — accumulated accounts, credential refs, model blacklists and its
+	// event log entirely in memory. Nothing ever marked it dirty, so
+	// .channels.state.json was never written and a restart silently wiped the
+	// account pool: the operator saw "logged in", restarted for an upgrade, and
+	// the account was gone. billing/quota escape this only because they declare
+	// hooks and/or use Call, both of which already mark dirty.
+	ps.markDirtyLocked(p)
 	return parseHandlerResponse(raw)
 }
 

internal/lua/plugin_http_test.go

@@ -14,6 +14,8 @@ package lua
 import (
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 	"time"
@@ -216,3 +218,139 @@ return P
 		t.Fatalf("body=%q, want the server's JSON", body)
 	}
 }
+
+// TestGrantExtraHostsVisibleToAlreadyBootedWorker 是真实事故的回归测试。
+//
+// 事故：allowlist 是不可变快照。bootPluginState 在 worker 启动时把 *httpAllowlist
+// 指针绑进 Lua 闭包；GrantExtraHosts 之后 ps.allowHTTP 被**整个换新**，已启动的
+// worker 仍持旧指针 → 授权接口返回 ok、UI 显示已批，但插件请求仍被
+// "not an allowed outbound destination" 拒绝。channels 插件的 auth/state 就是这样
+// 在生产被拦的（copilot.tencent.com 已批仍报错）。
+//
+// 修复：allowlist 改为带锁的**可变共享对象**（原地 add），所有闭包立即可见。
+// 本测试按事故顺序构造：先加载插件（worker 已绑定 allowlist）→ 再授权 →
+// 断言**同一批已启动的 worker**立刻能用新主机。
+func TestGrantExtraHostsVisibleToAlreadyBootedWorker(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(dir), dir)
+	ps.SetOutboundHosts(nil) // 起始只允许 loopback
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{"ok":true}`))
+	}))
+	defer srv.Close()
+	// srv.URL 是 127.0.0.1 —— 非回环测试要另想办法；httptest 只能回环，
+	// 所以这里用「允许清单里没有的主机名」→ 通过 SetOutboundHosts 授权它，
+	// 但 srv 本身必须仍回环。做法：插件请求的 URL 用 srv.URL（回环，恒允许）？
+	// 不行——那测不出 allowlist 更新。改用 checkURL 直接断言新主机可见。
+	//
+	// 真实事故的关键路径：worker 已启动 → 授权 → 同一 worker 立即看到。
+	// Lua 层面用回环无法区分（回环永远允许），因此这里的断言打在两层：
+	// ① allowlist 对象层：已存在实例 add 后立即 allows() 新主机；
+	// ② 插件层：SetOutboundHosts 重新折算后，同一 worker 的请求仍工作
+	//    （即 rebuild 不会使旧引用失效）。
+	host := "portal.example.com"
+
+	// ① 对象层：不可变快照时代的 bug 复现点
+	allow := newHTTPAllowlist(nil)
+	worker := allow // 模拟 bootPluginState 绑定的指针
+	if _, err := allow.checkURL("https://" + host + "/x"); err == nil {
+		t.Fatal("前提不成立：未授权主机不应允许")
+	}
+	allow.add([]string{host}) // GrantExtraHosts 的实际路径（原地 add）
+	if _, err := worker.checkURL("https://" + host + "/x"); err != nil {
+		t.Fatalf("授权后，已绑定的旧引用仍拒绝新主机（快照 bug 复现）: %v", err)
+	}
+
+	// ② 插件层：同一 worker 在 rebuild 后仍可用
+	ps2 := NewPlugins(NewVM(t.TempDir()), t.TempDir())
+	ps2.SetOutboundHosts([]string{"127.0.0.1"})
+	probe := `
+local P = {}
+P.name = "grantprobe"
+P.version = "1.0"
+P.exports = { "probe" }
+P.hosts = { "portal.example.com" }
+function P.probe(arg)
+  local res, err = ` + httpGlobal + `({ url = arg.url, method = "GET" })
+  if res == nil then return { error = tostring(err) } end
+  return { status = res.status }
+end
+return P
+`
+	if err := ps2.LoadSource("grantprobe", probe); err != nil {
+		t.Fatal(err)
+	}
+	// worker 已启动。现在授权（grant → rebuild → 原地 add）
+	if err := ps2.GrantExtraHosts("grantprobe", []string{"portal.example.com"}); err != nil {
+		t.Fatalf("grant: %v", err)
+	}
+	// 同一批 worker：一次调用成功即证明 allowlist 引用未被弃用
+	got, err := ps2.Call("grantprobe", "probe", map[string]interface{}{"url": srv.URL})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, _ := got.(map[string]interface{})
+	if m == nil || m["error"] != nil {
+		t.Fatalf("授权后同一 worker 调用失败（引用被弃用）: %v", got)
+	}
+	if s, _ := m["status"].(float64); int(s) != 200 {
+		t.Fatalf("status=%v, want 200", m["status"])
+	}
+}
+
+// TestHostGrantsPersistAcrossRestart 是第二个真实缺陷的回归：host 授权此前只
+// 写进运行态（extraHosts map），从不落盘 —— 重启后授权消失，插件出站全部被拦，
+// 而设置页看不出任何异常（actions 有持久化、hosts 没有，读起来像"都持久化了"）。
+func TestHostGrantsPersistAcrossRestart(t *testing.T) {
+	dir := t.TempDir()
+	ps := NewPlugins(NewVM(filepath.Join(dir, "adapters")), dir)
+	if err := ps.LoadSource("grantprobe", `
+local P = {}
+P.name = "grantprobe"
+P.version = "1.0"
+P.exports = {}
+P.hosts = { "portal.example.com", "api.partner.example" }
+return P
+`); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.GrantExtraHosts("grantprobe", []string{"portal.example.com", "api.partner.example"}); err != nil {
+		t.Fatal(err)
+	}
+
+	// 落盘验证
+	b, err := os.ReadFile(filepath.Join(dir, ".kernel-perms.json"))
+	if err != nil {
+		t.Fatalf("授权未落盘: %v", err)
+	}
+	if !strings.Contains(string(b), "portal.example.com") {
+		t.Fatalf("host 授权应写入 .kernel-perms.json，文件内容: %s", b)
+	}
+
+	// 重启语义：新 registry 读回授权
+	ps2 := NewPlugins(NewVM(filepath.Join(dir, "adapters")), dir)
+	if err := ps2.LoadKernelPermissions(); err != nil {
+		t.Fatal(err)
+	}
+	got := ps2.ExtraHostsGranted()["grantprobe"]
+	if len(got) != 2 {
+		t.Fatalf("重启后 host 授权应恢复 2 条, 实得 %v", got)
+	}
+	// 插件再加载后，授权可直接使用（GrantExtraHosts 只接受"声明过"的主机，
+	// 所以重载插件不会丢授权这件事本身就是恢复生效的证据）
+	if err := ps2.LoadSource("grantprobe", `
+local P = {}
+P.name = "grantprobe"
+P.version = "1.0"
+P.exports = {}
+P.hosts = { "portal.example.com", "api.partner.example" }
+return P
+`); err != nil {
+		t.Fatal(err)
+	}
+	got = ps2.ExtraHostsGranted()["grantprobe"]
+	if len(got) != 2 {
+		t.Fatalf("插件重载后授权应保留, 实得 %v", got)
+	}
+}

internal/lua/plugin_outbound.go

@@ -105,6 +105,18 @@ func normalizeHost(h string) string {
 // rebuildOutboundAllowlistLocked folds the SOURCE hosts and every GRANTED extra
 // host into the allowlist the HTTP bridge consults. Caller holds ps.mu.
 //
+// The allowlist object is mutated IN PLACE (allowHTTP.add) rather than replaced:
+// already-booted plugin workers hold the ORIGINAL pointer in their Lua closures
+// (bootPluginState captured it at boot), so replacing the field would strand
+// them on a stale copy — a grant would be recorded and displayed while requests
+// kept being refused. Mutating the shared object makes every closure see the
+// new hosts immediately.
+//
+// Note this FOLDS sources + grants additively: a host that loses its grant
+// cannot be un-allowed this way, which is acceptable because grants are
+// append-only operator approvals (revocation is a restart away — the persisted
+// file is the authority on boot).
+//
 // Declared-but-not-granted hosts are deliberately excluded: that exclusion is the
 // whole security property, so it is the one thing this function must not be
 // tempted to "helpfully" relax.
@@ -113,7 +125,7 @@ func (ps *Plugins) rebuildOutboundAllowlistLocked() {
 	for _, granted := range ps.extraHosts {
 		hosts = append(hosts, granted...)
 	}
-	ps.allowHTTP = newHTTPAllowlist(hosts)
+	ps.allowHTTP.add(hosts)
 }
 
 // readPluginExtraHosts reads `plugin.hosts`, the outbound hosts a plugin needs

internal/lua/plugins.go

@@ -641,15 +641,23 @@ func (ps *Plugins) LoadDir() error {
 // directory does not exist yet, mirroring the adapter seeding rule: once the
 // directory exists it is authoritative, so deleting or editing a shipped plugin
 // is a real action that survives restarts.
+// SeedBundled writes the plugins shipped with the gateway into dir when they are
+// MISSING.
+//
+// This used to be "write everything only when the directory does not exist",
+// which silently stopped delivering plugins added to the binary LATER: the
+// directory had existed since the first install, so a newly embedded plugin
+// (e.g. channels.lua) never landed on an upgraded deployment — it was in the
+// binary and completely invisible in /api/plugins. The directory's mere
+// existence is not evidence that its contents are current.
+//
+// The rule is now per-file: seed a bundled plugin only when a file of that name
+// is absent. An OPERATOR'S file always wins, so customising or replacing a
+// bundled plugin survives an upgrade — the same rule adapters already use.
 func (ps *Plugins) SeedBundled() error {
 	if ps.dir == "" {
 		return nil
 	}
-	if _, err := os.Stat(ps.dir); err == nil {
-		return nil
-	} else if !os.IsNotExist(err) {
-		return err
-	}
 	return writeBundledPlugins(ps.dir)
 }
 

internal/lua/plugins/channels.lua

@@ -287,6 +287,30 @@ local PLATFORMS = {
     hosts = { "copilot.tencent.com", "www.codebuddy.cn" },
     base_url = "https://copilot.tencent.com",
     adapter = "openai",
+    -- 计费/状态请求头：逐行移植参考 checkinHeaders（credits.ts）。
+    -- 实测只发 Authorization 会被上游 401：buddy 要求身份标识头齐全。
+    -- X-Domain 以产品配置为准（不用凭据里的 domain 快照，那是登录时的区域，
+    -- 跨产品迁移后会指错区）；X-Product 是部署类型（SaaS）；
+    -- X-Product-Code 是产品码（codebuddy）；X-User-Id 取凭据的 user_id。
+    headers_fn = function(accessToken, cred)
+      local h = {
+        ["Authorization"] = "Bearer " .. accessToken,
+        ["Accept"] = "application/json",
+        ["Content-Type"] = "application/json",
+        ["X-Domain"] = "copilot.tencent.com",
+        ["X-Product"] = "SaaS",
+        ["X-Product-Code"] = "codebuddy",
+        ["User-Agent"] = "CodeBuddyIDE/1.106.1",
+      }
+      local uid = cred and (cred.user_id or cred.uid)
+      if type(uid) == "string" and uid ~= "" then h["X-User-Id"] = uid end
+      local eid = cred and cred.enterprise_id
+      if type(eid) == "string" and eid ~= "" then
+        h["X-Enterprise-Id"] = eid
+        h["X-Tenant-Id"] = eid
+      end
+      return h
+    end,
     models = {
       "deepseek-v4-flash", "deepseek-v4-pro", "hy4-preview", "hy4-preview-x",
       "glm-5.3", "glm-5.3-flash", "glm-5.2",
@@ -803,6 +827,171 @@ local function poolRemove(provider, accountId)
   poolAccounts(provider)[accountId] = nil
 end
 
+-- 凭据取值：保险库里存的可能是**裸 token**（历史/device-poll/paste 路径），
+-- 也可能是**完整凭据 JSON**（ buddy / workbuddy 等 state-poll 平台，对齐参考
+-- buildCredential 的结构：access_token + refresh_token + expires_at + domain + ...）。
+--
+-- 本助手统一取出可用的 access token（及 refresh_token），使所有 HTTP 消费点
+-- （Bearer 头、HMAC 签名）无需关心存的是哪种形态：
+--   裸 token              → 原样返回
+--   完整凭据 JSON 字符串  → 取 .access_token
+--   已经是 table          → 取 .access_token
+-- 返回 (accessToken, refreshToken, fullTableOrNil)。两者都空时返回 (nil, nil, nil)。
+local function credentialOf(ref)
+  local raw = ref and __llmsproxy_secrets.get(ref)
+  if not raw or raw == "" then return nil, nil, nil end
+  -- 裸 token 快路：不含 '{' 的字符串直接当 token（绝大多数历史数据）
+  if type(raw) == "string" and raw:sub(1, 1) ~= "{" then return raw, nil, nil end
+  local ok, rec = pcall(json.decode, type(raw) == "string" and raw or "")
+  if ok and type(rec) == "table" then
+    local at = rec.access_token or rec.accessToken or rec.token
+    local rt = rec.refresh_token or rec.refreshToken
+    if at and at ~= "" then return at, rt, rec end
+    -- 是 JSON 但没有 access_token：回退原样（可能是别的形状的凭据）
+    return raw, rt, rec
+  end
+  return raw, nil, nil
+end
+
+-- credentialFull 取完整凭据（供需要 refresh_token/expires_at 的路径用）。
+local function credentialFull(ref)
+  local _, _, full = credentialOf(ref)
+  return full
+end
+
+-- jwtClaim 从 JWT 的 payload 读一个声明（不验签，仅用于读取展示/续期元数据）。
+-- 对齐参考 buddy.ts 的 jwtNickname / jwtExpiresAtMs / jwtSubject：
+-- CodeBuddy 的 login/account 响应**不含昵称**，昵称只在 access_token 的声明里；
+-- token 端又只返回 expiresIn（相对秒数）而不含 expiresAt。两处都得从 JWT 兜底。
+-- 解析失败返回 nil（调用方自行处理）。
+local function jwtClaim(token, name)
+  if type(token) ~= "string" or token == "" then return nil end
+  local payload = token:match("^[^%.]+%.([^%.]+)")
+  if not payload then return nil end
+  local bin, derr = base64_decode(payload)  -- 内核新增：base64url 兼容
+  if not bin or derr then return nil end
+  local ok, rec = pcall(json.decode, bin)
+  if not ok or type(rec) ~= "table" then return nil end
+  return rec[name]
+end
+
+-- stripControl 去掉控制字符（含 CR/LF/Tab）并把连续空白折叠为单个空格。
+-- 对齐参考 stripControlChars：CodeBuddy 的 scope 字段有时返回多行文本，
+-- 原样落盘会在序列化时被当作多行标量，破坏后续 JSON 解析。
+local function stripControl(v)
+  if type(v) ~= "string" then return v end
+  v = v:gsub("[%z\1-\31\127]+", " ")
+  v = v:gsub("%s%s+", " ")
+  return (v:gsub("^%s+", ""):gsub("%s+$", ""))
+end
+
+-- readStr 从响应表安全读字符串字段（兼容后端把数字返为数字）。
+local function readStr(rec, key)
+  if type(rec) ~= "table" then return "" end
+  local v = rec[key]
+  if type(v) == "string" then return stripControl(v) end
+  if type(v) == "number" then return tostring(v) end
+  return ""
+end
+
+-- absExpiryMs 把绝对字段或相对秒数换算为毫秒时间戳字符串。
+-- 对齐参考 absoluteExpiryMs：
+--   ① 绝对字段（expiresAt/refreshExpiresAt）存在 → 归一化为毫秒字符串
+--   ② 否则用相对秒数（expiresIn/refreshExpiresIn）+ access_token 的 JWT iat 兜底
+--   ③ 都没有 → 返回空串（由 credentialExpiresAtMs 再从 JWT exp 兜底）
+-- 实测：CodeBuddy 的 /v2/plugin/auth/token 只返回 expiresIn，没有 expiresAt；
+-- 不换算的话凭据的 expires_at 永远是空串，UI 一直显示"有效期未知"。
+local function absExpiryMs(rec, absKey, relKey, accessToken)
+  local absolute = readStr(rec, absKey)
+  if absolute ~= "" then
+    if absolute:match("^%d+$") then
+      local n = tonumber(absolute)
+      if n then
+        return tostring(n > 1e12 and n or n * 1000)
+      end
+    end
+    return absolute
+  end
+  local rel = rec[relKey]
+  local relSec = nil
+  if type(rel) == "number" then relSec = rel
+  elseif type(rel) == "string" and rel:match("^%-?%d+%.?%d*$") then relSec = tonumber(rel) end
+  if relSec == nil then return "" end
+  -- 基准：access_token 的 iat（签发时刻）优先，缺失时用当前时刻。
+  local iat = jwtClaim(accessToken, "iat")
+  local baseMs = (type(iat) == "number" and iat * 1000) or (os.time and os.time() * 1000 or 0)
+  return tostring(baseMs + relSec * 1000)
+end
+
+-- buildCredential 把 token 响应 + account 响应组合为可持久化的完整凭据 JSON。
+-- 逐字段对齐参考 buddy.ts 的 buildCredential / parseTokenData / parseAccountData：
+--   access_token / refresh_token / expires_at / refresh_expires_at /
+--   token_type / scope / domain / user_id / nickname / enterprise_id / account_type
+-- 昵称回填顺序：account.nickname → JWT.nickname → JWT.preferred_username。
+-- expires_at 同理：token.expiresAt 为空时由 absExpiryMs 用 expiresIn 或 JWT exp 兜底。
+local function buildCredential(data, accountData)
+  local access = readStr(data, "accessToken")
+  if access == "" then access = readStr(data, "access_token") end
+  local refresh = readStr(data, "refreshToken")
+  if refresh == "" then refresh = readStr(data, "refresh_token") end
+  local tokenType = readStr(data, "tokenType")
+  if tokenType == "" then tokenType = "Bearer" end
+  local acct = type(accountData) == "table" and accountData or {}
+  local nickname = readStr(acct, "nickname")
+  if nickname == "" then nickname = readStr(acct, "nickName") end
+  if nickname == "" then
+    local jn = jwtClaim(access, "nickname")
+    if type(jn) ~= "string" or jn == "" then jn = jwtClaim(access, "preferred_username") end
+    if type(jn) ~= "string" or jn == "" then jn = jwtClaim(access, "name") end
+    if type(jn) == "string" then nickname = stripControl(jn) end
+  end
+  local uid = readStr(acct, "uid")
+  if uid == "" then
+    local sub = jwtClaim(access, "sub")
+    if type(sub) == "string" then uid = sub end
+  end
+  local acctType = readStr(acct, "type")
+  if acctType == "" then acctType = "personal" end
+  return {
+    access_token = access,
+    refresh_token = refresh,
+    expires_at = absExpiryMs(data, "expiresAt", "expiresIn", access),
+    refresh_expires_at = absExpiryMs(data, "refreshExpiresAt", "refreshExpiresIn", access),
+    token_type = tokenType,
+    scope = readStr(data, "scope"),
+    domain = readStr(data, "domain"),
+    user_id = uid,
+    nickname = nickname,
+    enterprise_id = readStr(acct, "enterpriseId"),
+    account_type = acctType,
+  }
+end
+
+-- credentialExpiresAtMs 从凭据表解析毫秒过期时刻（兼容毫秒/秒级时间戳/ISO）。
+-- 对齐参考 credentialExpiresAtMs：expires_at 为空时回退解析 access_token 的 JWT exp。
+local function credentialExpiresAtMs(cred)
+  if type(cred) ~= "table" then return nil end
+  local raw = cred.expires_at
+  if type(raw) == "string" and raw ~= "" then
+    if raw:match("^%d+$") then
+      local n = tonumber(raw)
+      if n then return n > 1e12 and n or n * 1000 end
+    end
+    -- ISO 8601：解析 "YYYY-MM-DDThh:mm:ss" 形态（无时区偏移时按 UTC）
+    local Y, Mo, D, H, Mi, S = raw:match("^(%d%d%d%d)%-(%d%d)%-(%d%d)T(%d%d):(%d%d):(%d%d)")
+    if Y then return os.time({ year = tonumber(Y), month = tonumber(Mo), day = tonumber(D),
+      hour = tonumber(H), min = tonumber(Mi), sec = tonumber(S), isdst = false }) * 1000 end
+  end
+  local exp = jwtClaim(cred.access_token, "exp")
+  if type(exp) == "number" then return exp * 1000 end
+  return nil
+end
+
+-- credentialIsRefreshable 凭据是否携带可静默续期的 refresh_token。
+local function credentialIsRefreshable(cred)
+  return type(cred) == "table" and type(cred.refresh_token) == "string" and cred.refresh_token ~= ""
+end
+
 -- 模型黑名单（对齐参考 model.setDisabled）：provider → { [modelId] = true }
 local function disabledModelsOf(provider)
   if not plugin.state.modelDisabled then plugin.state.modelDisabled = {} end
@@ -840,7 +1029,8 @@ local function syncAccountSource(provider, accountId)
   local plat = platformById(provider)
   local acct = poolGet(provider, accountId)
   if not plat or not acct then return end
-  local tok = __llmsproxy_secrets.get(acct.credentialRef)
+  -- 凭据可能是裸 token 或完整凭据 JSON；上游源只吃 access token。
+  local tok = credentialOf(acct.credentialRef)
   if not tok or tok == "" then return end
   local name = accountSourceName(provider, accountId)
   local models = {}
@@ -1102,6 +1292,14 @@ end
 -- 对 device-poll 会话，每次轮询**推进一步**设备码换 token（deviceOneStep）：
 -- 轮询由页面驱动而不在 Lua 里阻塞，因为插件是单 VM 单锁，阻塞循环会卡死
 -- 本插件全部钩子。成功则落凭据+接入源（走 complete_login）。
+--
+-- ⚠️ **必须把 session 上的 account_id 传给 complete_login**（真实缺陷）：
+-- 页面流程是「先 account/add 建占位账号（credentialRef=channels:<plat>:<aid>）
+-- → login 带 account_id → 轮询」。若这里不带 account_id，complete_login 会走
+-- 兼容旧路径，凭据落到 `channels:<plat>` 这个**与占位账号无关的键**，而账号池
+-- 条目的 credentialRef 指向 `channels:<plat>:<aid>` —— 于是 account/list 报
+-- has_credential=false、credits/balances 报「无凭据」，表现为「登录成功但无法
+-- 获取账号信息」。对齐参考 BuddyAuth.login 的 `flowOptions.accountId` 语义。
 function plugin.poll_login(req)
   local id = (req and req.query and req.query.id) or ""
   sweepSessions()
@@ -1119,7 +1317,10 @@ function plugin.poll_login(req)
           s.state = "failed"
           s.message = rerr
         else
-          local res = plugin.complete_login({ platform = s.platform, credential = cred, account = acct or payload.account })
+          -- ⚠️ 必须带 account_id：页面流程先 account/add 建了占位账号，
+          -- 不带就会落到与占位账号无关的 `channels:<plat>` 键（见 poll_login 顶部注释）。
+          local res = plugin.complete_login({ platform = s.platform, credential = cred,
+            account = acct or payload.account, account_id = s.account_id })
           if res.json and res.json.ok then
             s.state = "completed"
             s.account = acct or payload.account
@@ -1149,21 +1350,36 @@ function plugin.poll_login(req)
         local code = rec.code
         local data = rec.data
         if resp.status == 200 and data ~= nil and type(data) == "table" and (data.accessToken or data.access_token) then
-          -- token 就绪。再拿账号（失败不阻塞，回退默认名）
-          local tok = data.accessToken or data.access_token
-          local acct = "codebuddy"
+          -- token 就绪。再拿账号（失败不阻塞，回退默认名）。
+          local access = data.accessToken or data.access_token
+          local accountData = nil
           local aresp, aerr = __llmsproxy_http({
             url = (sp.account_url or "") .. "?state=" .. tostring(s.state_token),
             method = "GET",
-            headers = { ["Accept"] = "application/json", ["Authorization"] = "Bearer " .. tok },
+            headers = { ["Accept"] = "application/json", ["Authorization"] = "Bearer " .. access },
           })
           if not aerr and aresp and aresp.status == 200 then
             local aok, arec = pcall(json.decode, aresp.body or "")
             if aok and type(arec) == "table" and type(arec.data) == "table" then
-              acct = tostring(arec.data.nickName or arec.data.nickname or arec.data.username or acct)
+              accountData = arec.data
             end
           end
-          local res = plugin.complete_login({ platform = s.platform, credential = tok, account = acct })
+          -- ① 组合完整凭据 JSON（对齐参考 buildCredential）：带 refresh_token /
+          --    expires_at / domain / user_id / nickname，而非只存裸 access_token。
+          --    只存裸 token 会丢掉 refresh_token，导致 refreshable=false、无法静默
+          --    续期，且 expires_at 未知。
+          local credTable = buildCredential(data, accountData)
+          local credJson = json.encode(credTable)
+          local acct = credTable.nickname
+          if acct == nil or acct == "" then
+            acct = credTable.user_id
+          end
+          if acct == nil or acct == "" then acct = "codebuddy" end
+          -- ② 必须带 account_id（页面流程的占位账号）；否则凭据落到无关的
+          --    `channels:<plat>` 键，账号池查不到。
+          local res = plugin.complete_login({ platform = s.platform, credential = credJson,
+            account = acct, account_id = s.account_id,
+            expires_at = credTable.expires_at, refreshable = credentialIsRefreshable(credTable) })
           if res.json and res.json.ok then
             s.state = "completed"
             s.account = acct
@@ -1209,9 +1425,27 @@ function plugin.complete_login(req)
     local ref = acct.credentialRef
     local _, serr = __llmsproxy_secrets.set(ref, token)
     if serr then return { status = 400, json = { error = serr } } end
-    acct.nickname = args.account or acct.nickname or aid
-    acct.expiresAt = args.expires_at or acct.expiresAt
-    acct.refreshable = args.refreshable or acct.refreshable or false
+    -- 从凭据自动补齐展示/续期字段（调用方漏传也不丢）。
+    -- 对齐参考 saveCredential：昵称取 credential.nickname，
+    -- expiresAt 由 credentialExpiresAtMs 推断（expires_at 或 JWT exp），
+    -- refreshable 看是否携带 refresh_token。
+    local credTbl = nil
+    if type(token) == "string" and token:sub(1, 1) == "{" then
+      local dok, drec = pcall(json.decode, token)
+      if dok and type(drec) == "table" then credTbl = drec end
+    end
+    local derivedNick = credTbl and credTbl.nickname
+    if derivedNick == nil or derivedNick == "" then derivedNick = credTbl and credTbl.user_id end
+    local derivedExp = credTbl and credentialExpiresAtMs(credTbl)
+    acct.nickname = args.account or derivedNick or acct.nickname or aid
+    acct.expiresAt = args.expires_at or (derivedExp and tostring(derivedExp)) or acct.expiresAt
+    if args.refreshable ~= nil then
+      acct.refreshable = args.refreshable and true or false
+    elseif credTbl ~= nil then
+      acct.refreshable = credentialIsRefreshable(credTbl)
+    else
+      acct.refreshable = acct.refreshable or false
+    end
     poolAdd(plat.id, acct)
     -- 终结该账号的未决会话
     for id, sess in pairs(sessions()) do
@@ -1228,6 +1462,9 @@ function plugin.complete_login(req)
   -- 兼容旧路径：单账号模式（e2e 与既有端点仍走它）。
   local ok, err = __llmsproxy_secrets.set("channels:" .. plat.id, token)
   if err then return { status = 400, json = { error = err } } end
+  -- 上游源只吃 access token：凭据可能是完整 JSON（buddy 等 state-poll 平台），
+  -- 必须经 credentialOf 取出 access_token，不能把整个 JSON 当 api_key 发上去。
+  local legacyTok = credentialOf("channels:" .. plat.id) or token
 
   plugin.state.accounts[plat.id] = {
     account = args.account or "已登录",
@@ -1246,7 +1483,7 @@ function plugin.complete_login(req)
       name = "channels-" .. plat.id,
       base_url = plat.base_url,
       adapter = "openai",
-      api_key = token,
+      api_key = legacyTok,
       max_concurrent = 2,
       models = (function()
         local ms = {}
@@ -1286,9 +1523,9 @@ function plugin.claim_daily(req)
   if args.account_id and args.account_id ~= "" then
     acctEntry = poolGet(plat.id, args.account_id)
     if not acctEntry then return { status = 400, json = { error = "账号不存在" } } end
-    tok = __llmsproxy_secrets.get(acctEntry.credentialRef)
+    tok = credentialOf(acctEntry.credentialRef)
   else
-    tok = __llmsproxy_secrets.get("channels:" .. plat.id)
+    tok = credentialOf("channels:" .. plat.id)
   end
   if not tok then return { status = 400, json = { error = "未登录" } } end
 
@@ -1357,7 +1594,7 @@ function plugin.auto_claim()
       plugin.claim_daily({ platform = plat.id })
     end
     for _, a in ipairs(poolList(plat.id)) do
-      if a.enabled ~= false and __llmsproxy_secrets.get(a.credentialRef) then
+      if a.enabled ~= false and credentialOf(a.credentialRef) then
         plugin.claim_daily({ platform = plat.id, account_id = a.id })
       end
     end
@@ -1503,7 +1740,7 @@ function plugin.account_list(req)
   if not plat then return { status = 400, json = { error = "未知平台" } } end
   local out = {}
   for _, a in ipairs(poolList(plat.id)) do
-    local has = __llmsproxy_secrets.get(a.credentialRef)
+    local has = credentialOf(a.credentialRef)
     out[#out+1] = {
       id = a.id, nickname = a.nickname, enabled = a.enabled ~= false,
       credentialRef = a.credentialRef, refreshable = a.refreshable or false,
@@ -1561,7 +1798,7 @@ function plugin.account_test(req)
   if not plat then return { status = 400, json = { error = "未知平台" } } end
   local acct = poolGet(plat.id, args and args.account_id)
   if not acct then return { status = 400, json = { error = "账号不存在" } } end
-  local tok = __llmsproxy_secrets.get(acct.credentialRef)
+  local tok = credentialOf(acct.credentialRef)
   if not tok or tok == "" then return { status = 200, json = { ok = false, error = "无凭据" } } end
   -- 同步一次源（确保最新凭据已入），返回同步结果
   syncAccountSource(plat.id, acct.id)
@@ -1582,7 +1819,7 @@ function plugin.account_refresh(req)
   if not plat then return { status = 400, json = { error = "未知平台" } } end
   local acct = poolGet(plat.id, args and args.account_id)
   if not acct then return { status = 400, json = { error = "账号不存在" } } end
-  local tok = __llmsproxy_secrets.get(acct.credentialRef)
+  local tok = credentialOf(acct.credentialRef)
   if tok and tok ~= "" then syncAccountSource(plat.id, acct.id) end
   return { status = 200, json = { ok = true } }
 end
@@ -1661,7 +1898,7 @@ function plugin.provider_status(req)
     for _, m in ipairs(p.models or {}) do if disabled[m] then disabledN = disabledN + 1 end end
     local closed = (p.models and #p.models > 0) and disabledN == #p.models
     local state_accounts = plugin.state.accounts[p.id]
-    local legacyN = (state_accounts and __llmsproxy_secrets.get("channels:" .. p.id)) and 1 or 0
+    local legacyN = (state_accounts and credentialOf("channels:" .. p.id)) and 1 or 0
     out[#out+1] = {
       id = p.id, name = p.name, flow = p.flow, closed = closed,
       models = { total = #(p.models or {}), disabled = disabledN },
@@ -1684,10 +1921,11 @@ function plugin.provider_set_enabled(req)
     for _, a in ipairs(poolList(plat.id)) do a.enabled = true poolAdd(plat.id, a) syncAccountSource(plat.id, a.id) end
     -- 旧路径也启用
     local legacy = plugin.state.accounts[plat.id]
-    if legacy and __llmsproxy_secrets.get("channels:" .. plat.id) then
+    local legacyTok = credentialOf("channels:" .. plat.id)
+    if legacy and legacyTok then
       __llmsproxy_kernel.call("add_source", { source = {
         name = "channels-" .. plat.id, base_url = plat.base_url, adapter = plat.adapter or "openai",
-        api_key = __llmsproxy_secrets.get("channels:" .. plat.id), max_concurrent = 2,
+        api_key = legacyTok, max_concurrent = 2,
         models = (function() local ms = {} for i, m in ipairs(plat.models) do ms[i] = { id = m, kind = "chat", priority = 20 } end return ms end)(),
       } })
     end
@@ -1744,11 +1982,11 @@ function plugin.backup_export(req)
   for _, p in ipairs(PLATFORMS) do
     for _, a in ipairs(poolList(p.id)) do
       payload.accounts[#payload.accounts+1] = a
-      local tok = __llmsproxy_secrets.get(a.credentialRef)
+      local tok = credentialOf(a.credentialRef)
       if tok then payload.credentials[a.credentialRef] = tok end
     end
     -- 旧路径单账号也备份
-    local lt = __llmsproxy_secrets.get("channels:" .. p.id)
+    local lt = credentialOf("channels:" .. p.id)
     if lt then payload.credentials["channels:" .. p.id] = lt end
   end
   return { status = 200, json = { ok = true, payload = payload } }
@@ -1813,8 +2051,18 @@ end
 -- 签到状态 + 逐账号余额（对齐参考 credits.status / credits.balances）
 -- ============================================================================
 -- fetchAccountJSON 用某账号凭据发一次请求，返回 (rec_table, nil) 或 (nil, err)。
-local function fetchAccountJSON(spec, tok)
-  local headers = spec.headers or { ["Authorization"] = "Bearer " .. tok, ["Content-Type"] = "application/json" }
+local function fetchAccountJSON(spec, tok, credTable, plat)
+  local headers
+  -- 头钩子位：spec（checkin_status / balance）自己的优先，否则用平台级的。
+  -- buddy 在平台层声明一次，签到与余额复用同一套身份头。
+  local hfn = spec.headers_fn or (plat and plat.headers_fn)
+  if hfn then
+    local okf, h = pcall(hfn, tok, credTable or {})
+    if okf and type(h) == "table" then headers = h end
+  end
+  if not headers then
+    headers = spec.headers or { ["Authorization"] = "Bearer " .. tok, ["Content-Type"] = "application/json" }
+  end
   local resp, err = __llmsproxy_http({
     url = spec.url, method = spec.method or "GET",
     headers = headers, body = spec.body,
@@ -1838,9 +2086,9 @@ function plugin.credits_status(req)
     local status = nil
     local cs = plat.checkin_status
     if cs then
-      local tok = __llmsproxy_secrets.get(a.credentialRef)
+      local tok = credentialOf(a.credentialRef)
       if tok then
-        local rec, err = fetchAccountJSON(cs, tok)
+        local rec, err = fetchAccountJSON(cs, tok, credentialFull(a.credentialRef), plat)
         if rec and plat.checkin_status.status_fn then
           local okf, st = pcall(plat.checkin_status.status_fn, rec)
           if okf then status = st end
@@ -1870,11 +2118,11 @@ function plugin.credits_balances(req)
     if not b then
       entry.error = "平台未声明余额端点"
     else
-      local tok = __llmsproxy_secrets.get(a.credentialRef)
+      local tok = credentialOf(a.credentialRef)
       if not tok then
         entry.error = "无凭据"
       else
-        local rec, err = fetchAccountJSON(b, tok)
+        local rec, err = fetchAccountJSON(b, tok, credentialFull(a.credentialRef), plat)
         if err then
           entry.error = err
         elseif b.total_fn then
@@ -2132,7 +2380,7 @@ function plugin.onboarding_claim(req)
     return { status = 400, json = { error = "该平台无新人任务" } }
   end
   -- raccoon 实测端点：POST {apiBase}/login/points/grant
-  local tok = __llmsproxy_secrets.get("channels:" .. plat.id)
+  local tok = credentialOf("channels:" .. plat.id)
   if not tok then return { status = 400, json = { error = "未登录" } } end
   local resp, err = __llmsproxy_http({
     url = plat.base_url .. "/login/points/grant", method = "POST",
@@ -2185,6 +2433,26 @@ function plugin.set_endpoints(req)
       if args.checkin[k] then plat.checkin[k] = tostring(args.checkin[k]) changed[#changed+1] = "checkin." .. k end
     end
   end
+  -- checkin_status / balance：声明式端点覆盖。参考把各平台的计费/状态端点当
+  -- 产品配置，运维要能整体改指（自建代理、镜像、测试假上游）。此前只覆盖了
+  -- 旧的 plat.checkin，buddy 这类"声明 checkin_status/balance"的平台无从覆盖 ——
+  -- 自动化测试也就无法把它们的端点指到假上游。
+  if type(args.checkin_status) == "table" and plat.checkin_status then
+    for _, k in ipairs({ "url", "method", "body" }) do
+      if args.checkin_status[k] then
+        plat.checkin_status[k] = tostring(args.checkin_status[k])
+        changed[#changed+1] = "checkin_status." .. k
+      end
+    end
+  end
+  if type(args.balance) == "table" and plat.balance then
+    for _, k in ipairs({ "url", "method", "body" }) do
+      if args.balance[k] then
+        plat.balance[k] = tostring(args.balance[k])
+        changed[#changed+1] = "balance." .. k
+      end
+    end
+  end
   if type(args.callback) == "table" and plat.callback then
     for _, k in ipairs({ "exchange_url" }) do
       if args.callback[k] then plat.callback[k] = tostring(args.callback[k]) changed[#changed+1] = "callback." .. k end
@@ -2199,7 +2467,7 @@ function plugin.list_platforms(req)
   local out = {}
   for _, p in ipairs(PLATFORMS) do
     local acct = accountOf(p.id) or {}
-    local has = __llmsproxy_secrets.get("channels:" .. p.id)
+    local has = credentialOf("channels:" .. p.id)
     out[#out + 1] = {
       id = p.id, name = p.name, flow = p.flow,
       logged_in = has ~= nil,

internal/lua/vm.go

@@ -744,11 +744,21 @@ func writeBundledPlugins(dir string) error {
 		if e.IsDir() || filepath.Ext(e.Name()) != ".lua" {
 			continue
 		}
+		dst := filepath.Join(dir, e.Name())
+		// Per-file seeding: a plugin already present is left ALONE. That is what
+		// makes an operator's own copy (or a locally edited one) survive an
+		// upgrade, and it is also why a newly embedded plugin still arrives —
+		// absence of THAT file, not of the directory, is the trigger.
+		if _, err := os.Stat(dst); err == nil {
+			continue
+		} else if !os.IsNotExist(err) {
+			return err
+		}
 		data, err := bundledAdapters.ReadFile(bundledPluginsDir + "/" + e.Name())
 		if err != nil {
 			continue
 		}
-		if err := os.WriteFile(filepath.Join(dir, e.Name()), data, 0644); err != nil {
+		if err := os.WriteFile(dst, data, 0644); err != nil {
 			return err
 		}
 	}
@@ -1348,6 +1358,38 @@ func setupGlobals(L *golua.State) {
 		L.PushString(base64.StdEncoding.EncodeToString([]byte(L.ToString(1))))
 		return 1
 	})
+	// base64_decode accepts standard, URL-safe and unpadded alphabets. It exists
+	// because a plugin reading a JWT payload (e.g. to read `exp`/`nickname` when
+	// the token endpoint omits them) must decode base64url, and hardcoding that
+	// in Lua would mean reimplementing base64 in a plugin. Generic, no vendor
+	// knowledge: an invalid string yields (nil, err), never a partial result.
+	registerFn(L, "base64_decode", func(L *golua.State) int {
+		s := strings.TrimSpace(L.ToString(1))
+		if s == "" {
+			L.PushString("")
+			L.PushNil()
+			return 2
+		}
+		// Normalise URL-safe alphabet and re-pad to a multiple of 4.
+		s = strings.NewReplacer("-", "+", "_", "/").Replace(s)
+		if pad := len(s) % 4; pad != 0 {
+			if pad == 1 {
+				L.PushNil()
+				L.PushString("invalid base64 length")
+				return 2
+			}
+			s += strings.Repeat("=", 4-pad)
+		}
+		b, err := base64.StdEncoding.DecodeString(s)
+		if err != nil {
+			L.PushNil()
+			L.PushString(err.Error())
+			return 2
+		}
+		L.PushString(string(b))
+		L.PushNil()
+		return 2
+	})
 	registerFn(L, "tohex", func(L *golua.State) int {
 		L.PushString(hex.EncodeToString([]byte(L.ToString(1))))
 		return 1

internal/lua/vm_test.go

@@ -1868,3 +1868,97 @@ func TestAnthropicUsageCacheMapping(t *testing.T) {
 		t.Errorf("usage-less chunk must not emit a usage object: %s", out)
 	}
 }
+
+// TestSeedBundledFillsMissingFilesInExistingDir 是真实事故的回归测试。
+//
+// 事故：SeedBundled 原本"目录已存在就什么都不做"。生产 plugin_dir 从首次安装
+// 起就存在，于是后来加入二进制的新内置插件（channels.lua）永远不会落盘 ——
+// 它在 go:embed 里、却完全不出现在 /api/plugins（线上只加载 billing/quota/usage）。
+//
+// 本测试构造那个确切条件：目录已存在且只装了部分内置插件，断言缺失的被补齐、
+// 已存在的（模拟运维自定义）**不被覆盖**。
+func TestSeedBundledFillsMissingFilesInExistingDir(t *testing.T) {
+	dir := filepath.Join(t.TempDir(), "plugins")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	// 模拟：目录已存在，里面只有运维自己的一份 billing.lua（内容故意不同）
+	const customBilling = "-- operator's own billing plugin\nreturn { name='billing', version='9.9.9' }"
+	if err := os.WriteFile(filepath.Join(dir, "billing.lua"), []byte(customBilling), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+	ps := NewPlugins(vm, dir)
+
+	if err := ps.SeedBundled(); err != nil {
+		t.Fatalf("seed: %v", err)
+	}
+
+	// ① 已存在的运维文件不被覆盖
+	got, err := os.ReadFile(filepath.Join(dir, "billing.lua"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != customBilling {
+		t.Errorf("运维自己的 billing.lua 被覆盖了 —— 升级会抹掉运维的改动")
+	}
+
+	// ② 每个内置插件现在都在目录里（这正是 channels.lua 缺失的那个洞）
+	entries, err := bundledAdapters.ReadDir(bundledPluginsDir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, e := range entries {
+		if e.IsDir() || filepath.Ext(e.Name()) != ".lua" {
+			continue
+		}
+		if _, err := os.Stat(filepath.Join(dir, e.Name())); err != nil {
+			t.Errorf("内置插件 %s 未被补齐（目录已存在时被跳过）", e.Name())
+		}
+	}
+
+	// ③ 补齐后确实能被加载（channels 必须真的出现在插件表里）
+	if err := ps.LoadDir(); err != nil {
+		t.Fatal(err)
+	}
+	names := map[string]bool{}
+	for _, p := range ps.List() {
+		names[p["name"].(string)] = true
+	}
+	if !names["channels"] {
+		t.Errorf("channels 插件应被 seed 并加载, 实得 %v", names)
+	}
+}
+
+// TestSeedBundledIdempotent 二次调用不应改变已就位的文件。
+func TestSeedBundledIdempotent(t *testing.T) {
+	dir := filepath.Join(t.TempDir(), "plugins")
+	vm := NewVM(filepath.Join(t.TempDir(), "adapters"))
+	if err := vm.Start(); err != nil {
+		t.Fatal(err)
+	}
+	defer vm.Stop()
+	ps := NewPlugins(vm, dir)
+
+	if err := ps.SeedBundled(); err != nil {
+		t.Fatal(err)
+	}
+	before, _ := os.ReadFile(filepath.Join(dir, "channels.lua"))
+	// 运维改一笔
+	edited := append(before, []byte("\n-- operator edit\n")...)
+	if err := os.WriteFile(filepath.Join(dir, "channels.lua"), edited, 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := ps.SeedBundled(); err != nil {
+		t.Fatal(err)
+	}
+	after, _ := os.ReadFile(filepath.Join(dir, "channels.lua"))
+	if string(after) != string(edited) {
+		t.Error("二次 seed 覆盖了已存在的文件")
+	}
+}