Fix upstream 400: flatten multi-part content arrays and sanitize invalid tool_calls

The Zhanlu upstream gateway returns HTTP 400 (请求消息格式错误) for two
message-structure issues that AI coding tools like MiMoCode produce:

1. Multi-part content arrays — OpenAI SDKs send content as
   [{type:"text",text:"…"}] arrays; the gateway only accepts string
   content. normalizeMessages now flattens text-only arrays into a
   concatenated string (non-text parts like images are preserved).

2. Invalid tool_calls — when a tool call fails, MiMoCode emits
   {name:"invalid", arguments:{"tool":"task","error":"…"}}
   placeholders. The gateway rejects function names not in the tools list
   and non-object arguments (e.g. "-1"). normalizeMessages now
   sanitizes these in place: the real tool name is extracted from the
   arguments' "tool" field (falling back to an arbitrary declared
   tool), and non-JSON-object arguments are replaced with "{}". No
   messages or tool results are removed, preserving the full
   conversation context including error feedback.

Verified with the exact error.md request: 0/10 400 errors after fix
(vs 10/10 before). Model returns valid streaming responses with task
tool calls.

.gitea/workflows/build.yml

@@ -58,8 +58,13 @@ jobs:
             CGO_ENABLED=0 GOOS=linux GOARCH="${arch}" \
               go build -trimpath -ldflags "${LDFLAGS}" \
               -o "zhanlu-proxy-linux-${arch}" ./cmd/zhanlu-proxy
+
+            echo "building windows/${arch}"
+            CGO_ENABLED=0 GOOS=windows GOARCH="${arch}" \
+              go build -trimpath -ldflags "${LDFLAGS}" \
+              -o "zhanlu-proxy-windows-${arch}.exe" ./cmd/zhanlu-proxy
           done
-          ls -lh zhanlu-proxy-linux-*
+          ls -lh zhanlu-proxy-linux-* zhanlu-proxy-windows-*.exe
 
       - name: Publish release assets
         if: startsWith(github.ref, 'refs/tags/v')
@@ -80,10 +85,11 @@ jobs:
           echo "release id: ${RELEASE_ID}"
 
           for arch in amd64 arm64; do
-            f="zhanlu-proxy-linux-${arch}"
-            curl -fsSL -X POST -H "${AUTH}" \
-              -F "attachment=@${f};filename=${f}" \
-              "${API}/releases/${RELEASE_ID}/assets?name=${f}"
+            for f in "zhanlu-proxy-linux-${arch}" "zhanlu-proxy-windows-${arch}.exe"; do
+              curl -fsSL -X POST -H "${AUTH}" \
+                -F "attachment=@${f};filename=${f}" \
+                "${API}/releases/${RELEASE_ID}/assets?name=${f}"
+            done
           done
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.gitignore

@@ -1,4 +1,9 @@
-credentials.json
+zhanlu.db
+zhanlu.db-shm
+zhanlu.db-wal
+*.db
+*.db-shm
+*.db-wal
 extension/
 *.exe
 *.log
@@ -8,3 +13,5 @@ extension/
 tmp/
 temp/
 source/
+output/
+error.md

README.md

@@ -5,11 +5,13 @@
 当前实现包含：
 
 - 登录页支持插件默认的移动云手机号验证码登录：验证码校验后按插件流程调用 `/api/acepilot/zhanlu/v1/login` 获取用户资料，再通过 SM2 签名调用 `/user/api/v2/external/key/get-or-create` 换取模型 API Key，凭据自动保存到本地 JSON。
-- OpenAI 兼容接口：`/v1/models`（优先从 `/gateway/v1/model/info` 拉取模型列表）、`/v1/chat/completions`（携带 `Authorization: Bearer <apiKey>` 请求 `{modelBaseUrl}/chat/completions`）。
+- OpenAI 兼容接口：`/v1/models`（优先从 `/gateway/v1/model/info` 拉取模型列表）、`/v1/chat/completions`（携带 `Authorization: Bearer <apiKey>` 请求 `{modelBaseUrl}/chat/completions`）、`/v1/responses`（Responses API，自动转换请求/响应格式，兼容新版 SDK 与 AI 编码工具）。
 - 湛卢登录签名逻辑：RSA `authorization`、SHA-256 query hash、HMAC-SHA1 `Signature`，用于 v1.4.2 的 v1/login 认证。
 - 模型 API Key 换取签名逻辑：SM3 摘要 + SM2 签名（`X-Auth-Signature`/`X-Auth-Timestamp`/`X-Auth-Nonce`）。
 - 上游 SSE 直接透传为 OpenAI SSE；非流式请求在本地聚合为 OpenAI Chat Completion JSON。
 - OpenAI 函数/工具调用：支持 `tools`、`tool_choice`、流式 `delta.tool_calls`、非流式 `message.tool_calls` 以及 `role: tool` 结果续传。
+- Token 消耗统计：流式与非流式请求均解析上游 `usage`，按模型/按日/最近明细写入本地 SQLite（`zhanlu.db`），凭据也一并持久化在同一个库中。管理页提供 `GET /admin/stats` 可视化与 `GET /api/stats` JSON 接口。
+- 可用模型：管理后台新增「可用模型」tab，实时拉取上游 `/gateway/v1/model/info` 返回的模型列表，并提供单模型可用性及首字延时（TTFT）探测（`GET /api/models`、`POST /api/models/test`）。
 
 ## 运行
 
@@ -37,17 +39,17 @@ http://127.0.0.1:8080/
 /opt/zhanlu-proxy/zhanlu-proxy
 ```
 
-凭据文件放在：
+数据库放在：
 
 ```text
-/opt/zhanlu-proxy/credentials.json
+/opt/zhanlu-proxy/zhanlu.db
 ```
 
 创建环境变量文件 `/etc/zhanlu-proxy/zhanlu-proxy.env`：
 
 ```env
 ZHANLU_LISTEN_ADDR=:8080
-ZHANLU_CREDENTIALS_FILE=/opt/zhanlu-proxy/credentials.json
+ZHANLU_DB_FILE=/opt/zhanlu-proxy/zhanlu.db
 ZHANLU_MOBILE_LOGIN_BASE_URL=https://ecloud.10086.cn
 ZHANLU_MOBILE_MODEL_BASE_URL=https://ecloud.10086.cn/api/query/aigateway
 ZHANLU_UPSTREAM_TIMEOUT=300s
@@ -106,20 +108,22 @@ journalctl -u zhanlu-proxy -f
 - 使用本次 `secret` AES 解密响应中的 `ak`、`sk`、`license`，得到 `AccessKey`、`SecretKey`、`Token`。
 - 按插件 v1.4.2 流程调用 `/api/acepilot/zhanlu/v1/login`（RSA+HmacSHA1 签名 URL + `plugin_type=zhanlu_ide` 请求头）获取用户资料（email/组织/团队）。
 - 用 SM2 私钥签名调用 `{mobileModelBaseUrl}/user/api/v2/external/key/get-or-create` 换取模型 `apiKey`。
-- 凭据（含 `apiKey`、`modelBaseUrl`、email 等）会写入 JSON 文件，后续 OpenAI 兼容接口自动使用。
+- 凭据（含 `apiKey`、`modelBaseUrl`、email 等）会写入本地 SQLite 数据库（`zhanlu.db`），后续 OpenAI 兼容接口自动使用。
 
-默认保存到当前执行目录：
+默认数据库保存在当前执行目录：
 
 ```text
-credentials.json
+zhanlu.db
 ```
 
 可以通过环境变量覆盖：
 
 ```powershell
-$env:ZHANLU_CREDENTIALS_FILE="E:\path\to\credentials.json"
+$env:ZHANLU_DB_FILE="E:\path\to\zhanlu.db"
 ```
 
+凭据仅持久化在数据库中，不再使用 JSON 文件。
+
 手机号验证码登录使用 `ZHANLU_MOBILE_LOGIN_BASE_URL`，默认公网地址来自插件配置（兼容旧环境变量 `ZHANLU_SERVER_BASE_URL`）：
 
 ```powershell
@@ -165,6 +169,30 @@ curl -N http://127.0.0.1:8080/v1/chat/completions `
 
 请求中的 `tools`、`tool_choice` 会传给湛卢模型。流式响应返回增量 `delta.tool_calls`；非流式响应会把分片聚合为完整的 `message.tool_calls`，并保留 `finish_reason: "tool_calls"`。执行工具后，将 assistant 的 `tool_calls` 和 `role: "tool"` 结果放回 `messages` 再发起请求即可得到最终回答。
 
+### Responses API
+
+代理还兼容 OpenAI Responses API（`POST /v1/responses`），支持使用该 API 的客户端（如新版 OpenAI SDK、Cursor 等 AI 编码工具）直接对接。代理会将 Responses API 请求格式转换为上游 chat/completions 格式，再将响应转回 Responses 格式。
+
+请求中的 `input`（字符串或消息数组）、`instructions`（系统提示词）、`max_output_tokens`、`temperature`、`top_p`、`tools`、`tool_choice`、`response_format` 等参数均会翻译并透传给上游。`tools` 格式自动从 Responses 扁平结构转换为 Chat Completions 嵌套 `{function:{…}}` 结构。多轮对话中的 `function_call` 和 `function_call_output` 输入项也会自动转换为对应的 assistant `tool_calls` 和 `role: tool` 消息。
+
+非流式：
+
+```powershell
+curl http://127.0.0.1:8080/v1/responses `
+  -H "Content-Type: application/json" `
+  -d '{"model":"zhanlu/auto","input":"hello","stream":false}'
+```
+
+流式：
+
+```powershell
+curl -N http://127.0.0.1:8080/v1/responses `
+  -H "Content-Type: application/json" `
+  -d '{"model":"zhanlu/auto","input":"hello","stream":true}'
+```
+
+流式响应返回标准 Responses API SSE 事件序列（`response.created` → `response.output_item.added` → `response.output_text.delta` → `response.output_text.done` → `response.output_item.done` → `response.completed`），思考模型额外包含 `response.reasoning_summary_text.delta` 事件，工具调用包含 `response.function_call_arguments.delta` 事件。Token 统计同样记录在 SQLite 中。
+
 如果设置了本地 OpenAI 兼容 API Key，需要带 `Authorization`：
 
 ```powershell
@@ -184,7 +212,8 @@ curl http://127.0.0.1:8080/v1/models `
 | `ZHANLU_MOBILE_LOGIN_BASE_URL` | `https://ecloud.10086.cn` | 移动云公网登录 Base URL（兼容旧变量 `ZHANLU_SERVER_BASE_URL`） |
 | `ZHANLU_MOBILE_MODEL_BASE_URL` | `https://ecloud.10086.cn/api/query/aigateway` | 移动云公网模型网关 Base URL |
 | `ZHANLU_UPSTREAM_PATH` | `/chat/completions` | 模型网关聊天接口路径 |
-| `ZHANLU_CREDENTIALS_FILE` | `credentials.json` | 凭据 JSON 路径，默认当前执行目录 |
+| `ZHANLU_DB_FILE` | `zhanlu.db` | 本地 SQLite 数据库路径，凭据与 token 统计均存于此，默认当前执行目录 |
+| `ZHANLU_STATS_DISABLED` | `false` | 设为 `true` 关闭 token 用量记录（仅停止写入统计，凭据存储不受影响） |
 | `ZHANLU_ACCESS_KEY` | 空 | 直接从环境变量提供 AccessKey |
 | `ZHANLU_SECRET_KEY` | 空 | 直接从环境变量提供 SecretKey |
 | `ZHANLU_TOKEN` | 空 | 直接从环境变量提供 Token |
@@ -207,14 +236,49 @@ curl http://127.0.0.1:8080/v1/models `
 服务启动时按以下优先级加载凭据：
 
 1. 环境变量 `ZHANLU_ACCESS_KEY`、`ZHANLU_SECRET_KEY`、`ZHANLU_TOKEN` 或 `ZHANLU_API_KEY`。
-2. `ZHANLU_CREDENTIALS_FILE` 指向的 JSON 文件。
+2. `ZHANLU_DB_FILE` 数据库中持久化的凭据行。
 
-登录页面保存后，运行中的服务会立即使用新凭据。若环境中只有 AK/SK/Token 而没有 `apiKey`，首次调用聊天接口时会自动按插件流程换取 API Key 并回写凭据文件。
+登录页面保存后，运行中的服务会立即使用新凭据并写入数据库。若环境中只有 AK/SK/Token 而没有 `apiKey`，首次调用聊天接口时会自动按插件流程换取 API Key 并回写数据库。
+
+## Token 消耗统计
+
+代理在 `/v1/chat/completions` 完成后解析上游 `usage`（流式路径在透传 SSE 的同时旁路解析末块 `usage`，非流式路径在聚合时解析），将以下维度写入 `ZHANLU_DB_FILE`：
+
+- 时间、模型、流式/非流式
+- `prompt_tokens` / `completion_tokens` / `total_tokens` / `reasoning_tokens`（思考模型）
+- `cached_tokens`（来自 `prompt_tokens_details.cached_tokens`，提示缓存命中的 token 数）与缓存命中率
+- 请求状态（`success` / `upstream_error`）与耗时
+
+缓存命中率为 `cached_tokens / prompt_tokens`，仅在模型/网关支持 prompt 缓存且上游返回 `cached_tokens` 时非零。
+
+管理页（需先登录管理页面）提供：
+
+- `GET /admin/stats`：可视化页面，展示总览、按模型、按日柱状、最近请求明细，带重置按钮。
+- `GET /api/stats`：JSON 接口，支持 `since`/`until`（RFC3339）、`model`、`limit` 查询参数。
+- `POST /api/stats/reset`：清空统计（凭据不受影响）。
+
+设置 `ZHANLU_STATS_DISABLED=true` 可停止写入统计。
+
+## 可用模型
+
+管理后台「可用模型」tab（位于「Token 统计」之后）实时展示当前上游接口返回的模型列表，并提供单模型可用性与延时探测：
+
+- 进入 tab 时自动拉取一次，也可随时点击「刷新」重新获取。
+- 每个模型行可单独「测试」，或点击「全部测试」依次探测所有模型。
+- 探测向上游 `/chat/completions` 发送一条极简流式请求（`hi`），测量首字延时（TTFT，首个 SSE 数据块到达时间）与总耗时，并在收到首个内容块后立即关闭连接，避免消耗额外 token。
+- 探测请求不计入 Token 统计。
+
+对应 JSON 接口（需先登录管理页面）：
+
+- `GET /api/models`：实时返回上游模型列表 `{"ok":true,"models":["GLM-4.7",...]}`。
+- `POST /api/models/test`：请求体 `{"model":"<model_id>"}`，返回 `{"ok":true,"model":"...","available":true,"ttft_ms":123,"total_ms":456}` 或 `{"ok":true,"model":"...","available":false,"error":"...","total_ms":789}`。
+
+探测超时上限为 30 秒（`modelTestTimeout`）；凭据未配置或 API Key 缺失时 `GET /api/models` 与 `POST /api/models/test` 会按需自动换取 API Key，与聊天接口一致。
 
 ## 安全说明
 
-- `credentials.json` 包含明文 `AccessKey`、`SecretKey`、`Token` 和 `apiKey`，请不要提交到仓库。
-- 默认保存在当前执行目录的 `credentials.json`。
+- `zhanlu.db` 数据库包含明文 `AccessKey`、`SecretKey`、`Token` 和 `apiKey`，请不要提交到仓库。
+- 默认保存在当前执行目录的 `zhanlu.db`（建议通过 `ZHANLU_DB_FILE` 指向受保护路径）。
 - 建议设置 `ZHANLU_LOGIN_PASSWORD`，避免公网暴露的 `/admin/login` 被直接访问。
 - 错误响应默认不会返回签名 URL，避免泄露 `AccessKey`、`authorization`、`Signature`。
 - `ZHANLU_DEBUG=true` 时会返回更详细错误，但仍会对敏感 query 参数脱敏。

cmd/zhanlu-proxy/main.go

@@ -6,6 +6,7 @@ import (
 
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/config"
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/server"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/store"
 )
 
 func main() {
@@ -14,7 +15,20 @@ func main() {
 		log.Fatal(err)
 	}
 
-	h := server.New(cfg)
+	st, err := store.Open(cfg.DBPath)
+	if err != nil {
+		log.Fatalf("open store: %v", err)
+	}
+	defer st.Close()
+
+	// Credential precedence at startup: environment variables > db row.
+	if cfg.Credentials.Validate() != nil && !cfg.Credentials.HasAPIKey() {
+		if dbCreds, derr := st.LoadCredentials(); derr == nil && (dbCreds.HasAPIKey() || dbCreds.Validate() == nil) {
+			cfg.Credentials = dbCreds
+		}
+	}
+
+	h := server.New(cfg, st)
 	log.Printf("zhanlu proxy listening on %s", cfg.ListenAddr)
 	log.Printf("login page: http://127.0.0.1%s/login", cfg.ListenAddr)
 	if err := http.ListenAndServe(cfg.ListenAddr, h); err != nil {

go.mod

@@ -4,4 +4,16 @@ go 1.25.0
 
 require github.com/emmansun/gmsm v0.44.1
 
-require golang.org/x/crypto v0.54.0 // indirect
+require (
+	github.com/dustin/go-humanize v1.0.1 // indirect
+	github.com/google/uuid v1.6.0 // indirect
+	github.com/mattn/go-isatty v0.0.24 // indirect
+	github.com/ncruces/go-strftime v1.0.0 // indirect
+	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
+	golang.org/x/crypto v0.54.0 // indirect
+	golang.org/x/sys v0.47.0 // indirect
+	modernc.org/libc v1.74.4 // indirect
+	modernc.org/mathutil v1.7.1 // indirect
+	modernc.org/memory v1.11.0 // indirect
+	modernc.org/sqlite v1.56.0 // indirect
+)

go.sum

@@ -1,6 +1,24 @@
+github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
+github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
 github.com/emmansun/gmsm v0.44.1 h1:zDTkdtLWFG0vCbhPV+k9pte14tix/eK71At9Iai9fP4=
 github.com/emmansun/gmsm v0.44.1/go.mod h1:p6RIUta0/KboFHrOxr1x8q+pd8RZtdaTO7XNp0RmMQM=
+github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
+github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/mattn/go-isatty v0.0.24 h1:tGZZoVgT/KiqK1c8ocVLeDS8BSWMRd47J3Lbz7vsReI=
+github.com/mattn/go-isatty v0.0.24/go.mod h1:nMCL3Zebbrt45jsMDgnfIwz6ydEQApk5oEI3HqDio6A=
+github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=
+github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=
+github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec h1:W09IVJc94icq4NjY3clb7Lk8O1qJ8BdBEF8z0ibU0rE=
+github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo=
 golang.org/x/crypto v0.54.0 h1:YLIA59K4fiNzHzjnZt2tUJQjQtUWfWbeHBqKtk3eScw=
 golang.org/x/crypto v0.54.0/go.mod h1:KWL8ny2AZdGR2cWmzeHrp2azQPGogOv+HeQaVEXC2dk=
 golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
 golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+modernc.org/libc v1.74.4 h1:fX1Omw4o2/1C2iRkkIsrQTasJQldLhRmuPreXLoWs9k=
+modernc.org/libc v1.74.4/go.mod h1:eeQAS9W3sZeKYMFubydxJpII9ybHWshk+7or7bLG9co=
+modernc.org/mathutil v1.7.1 h1:GCZVGXdaN8gTqB1Mf/usp1Y/hSqgI2vAGGP4jZMCxOU=
+modernc.org/mathutil v1.7.1/go.mod h1:4p5IwJITfppl0G4sUEDtCr4DthTaT47/N3aT6MhfgJg=
+modernc.org/memory v1.11.0 h1:o4QC8aMQzmcwCK3t3Ux/ZHmwFPzE6hf2Y5LbkRs+hbI=
+modernc.org/memory v1.11.0/go.mod h1:/JP4VbVC+K5sU2wZi9bHoq2MAkCnrt2r98UGeSK7Mjw=
+modernc.org/sqlite v1.56.0 h1:/D8e2RfFqoy/Zc6PuC76U28zFwmI/sYx1Kjm4yEn9e0=
+modernc.org/sqlite v1.56.0/go.mod h1:yCJ2cmAaIkHQ25oXWrF8H4O1lIfPYPR26yCEDj2P3pQ=

internal/auth/credentials.go

@@ -1,10 +1,7 @@
 package auth
 
 import (
-	"encoding/json"
 	"errors"
-	"os"
-	"path/filepath"
 	"strings"
 	"time"
 )
@@ -49,30 +46,3 @@ func (c Credentials) Validate() error {
 func (c Credentials) HasAPIKey() bool {
 	return strings.TrimSpace(c.APIKey) != "" && strings.TrimSpace(c.ModelBaseURL) != ""
 }
-
-func LoadCredentials(path string) (Credentials, error) {
-	b, err := os.ReadFile(path)
-	if err != nil {
-		return Credentials{}, err
-	}
-	var c Credentials
-	if err := json.Unmarshal(b, &c); err != nil {
-		return Credentials{}, err
-	}
-	return c, c.Validate()
-}
-
-func SaveCredentials(path string, c Credentials) error {
-	if err := c.Validate(); err != nil {
-		return err
-	}
-	c.SavedAt = time.Now()
-	b, err := json.MarshalIndent(c, "", "  ")
-	if err != nil {
-		return err
-	}
-	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
-		return err
-	}
-	return os.WriteFile(path, b, 0o600)
-}

internal/auth/sso.go

@@ -5,9 +5,12 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"net/http"
 	"strings"
 	"time"
+
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/util"
 )
 
 // ExchangeCode exchanges an SSO auth code for a user profile via the Zhanlu
@@ -39,9 +42,8 @@ func ExchangeCode(client *http.Client, endpoint string, code string, decryptKey
 	}
 	defer resp.Body.Close()
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		b := make([]byte, 1024)
-		n, _ := resp.Body.Read(b)
-		return Profile{}, fmt.Errorf("exchange returned %d: %s", resp.StatusCode, strings.TrimSpace(string(b[:n])))
+		b, _ := io.ReadAll(io.LimitReader(resp.Body, 1024))
+		return Profile{}, fmt.Errorf("exchange returned %d: %s", resp.StatusCode, strings.TrimSpace(string(b)))
 	}
 
 	var exchange ExchangeResponse
@@ -49,11 +51,17 @@ func ExchangeCode(client *http.Client, endpoint string, code string, decryptKey
 		return Profile{}, err
 	}
 	if exchange.ErrorCode != "" && exchange.ErrorCode != "Success" {
-		msg := firstNonEmpty(exchange.ErrorMsg, exchange.Message, exchange.ErrorCode)
+		msg := util.FirstNonEmpty(exchange.ErrorMsg, exchange.Message, exchange.ErrorCode)
+		if msg == "" {
+			msg = "unknown error"
+		}
 		return Profile{}, fmt.Errorf("exchange failed: %s", msg)
 	}
 	if exchange.State != "" && exchange.State != "OK" {
-		msg := firstNonEmpty(exchange.ErrorMsg, exchange.Message, exchange.State)
+		msg := util.FirstNonEmpty(exchange.ErrorMsg, exchange.Message, exchange.State)
+		if msg == "" {
+			msg = "unknown error"
+		}
 		return Profile{}, fmt.Errorf("exchange failed: %s", msg)
 	}
 
@@ -95,12 +103,3 @@ type ExchangeResponse struct {
 	Result    map[string]any `json:"result"`
 	Data      map[string]any `json:"data"`
 }
-
-func firstNonEmpty(values ...string) string {
-	for _, v := range values {
-		if strings.TrimSpace(v) != "" {
-			return v
-		}
-	}
-	return "unknown error"
-}

internal/config/config.go

@@ -2,11 +2,11 @@ package config
 
 import (
 	"os"
-	"path/filepath"
 	"strings"
 	"time"
 
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/auth"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/util"
 )
 
 type Config struct {
@@ -14,7 +14,8 @@ type Config struct {
 	MobileLoginBaseURL string
 	MobileModelBaseURL string
 	UpstreamPath       string
-	CredentialsPath    string
+	DBPath             string
+	StatsDisabled      bool
 	SSOExchangeURL     string
 	SSOBaseURL         string
 	TokenDecryptKey    string
@@ -25,7 +26,7 @@ type Config struct {
 	LoginPassword      string
 	PluginVersion      string
 	UpstreamTimeout    time.Duration
-	StreamIdleTimout   time.Duration
+	StreamIdleTimeout  time.Duration
 	Debug              bool
 	Credentials        auth.Credentials
 }
@@ -33,10 +34,11 @@ type Config struct {
 func Load() (Config, error) {
 	cfg := Config{
 		ListenAddr:         getenv("ZHANLU_LISTEN_ADDR", ":8080"),
-		MobileLoginBaseURL: firstNonEmpty(os.Getenv("ZHANLU_MOBILE_LOGIN_BASE_URL"), getenv("ZHANLU_SERVER_BASE_URL", "https://ecloud.10086.cn")),
+		MobileLoginBaseURL: util.FirstNonEmpty(os.Getenv("ZHANLU_MOBILE_LOGIN_BASE_URL"), getenv("ZHANLU_SERVER_BASE_URL", "https://ecloud.10086.cn")),
 		MobileModelBaseURL: getenv("ZHANLU_MOBILE_MODEL_BASE_URL", "https://ecloud.10086.cn/api/query/aigateway"),
 		UpstreamPath:       getenv("ZHANLU_UPSTREAM_PATH", "/chat/completions"),
-		CredentialsPath:    getenv("ZHANLU_CREDENTIALS_FILE", defaultCredentialsPath()),
+		DBPath:             getenv("ZHANLU_DB_FILE", "zhanlu.db"),
+		StatsDisabled:      strings.EqualFold(os.Getenv("ZHANLU_STATS_DISABLED"), "true"),
 		SSOBaseURL:         getenv("ZHANLU_SSO_BASE_URL", "http://4c.hq.cmcc"),
 		SSOExchangeURL:     getenv("ZHANLU_SSO_EXCHANGE_URL", "http://rdcloud.4c.hq.cmcc/cmdevops-aiplus-agent-gateway/api/acepilot/zhanlu/authToken"),
 		TokenDecryptKey:    getenv("ZHANLU_TOKEN_DECRYPT_KEY", "3jw7woww2rvhla6k"),
@@ -47,7 +49,7 @@ func Load() (Config, error) {
 		LoginPassword:      os.Getenv("ZHANLU_LOGIN_PASSWORD"),
 		PluginVersion:      getenv("ZHANLU_PLUGIN_VERSION", "1.4.2"),
 		UpstreamTimeout:    durationEnv("ZHANLU_UPSTREAM_TIMEOUT", 300*time.Second),
-		StreamIdleTimout:   durationEnv("ZHANLU_STREAM_IDLE_TIMEOUT", 300*time.Second),
+		StreamIdleTimeout:  durationEnv("ZHANLU_STREAM_IDLE_TIMEOUT", 300*time.Second),
 		Debug:              strings.EqualFold(os.Getenv("ZHANLU_DEBUG"), "true"),
 	}
 	cfg.Credentials = auth.Credentials{
@@ -56,12 +58,8 @@ func Load() (Config, error) {
 		Token:     os.Getenv("ZHANLU_TOKEN"),
 		APIKey:    os.Getenv("ZHANLU_API_KEY"),
 	}
-	if cfg.Credentials.Validate() == nil || cfg.Credentials.HasAPIKey() {
-		return cfg, nil
-	}
-	if creds, err := auth.LoadCredentials(cfg.CredentialsPath); err == nil {
-		cfg.Credentials = creds
-	}
+	// Credentials from the environment were incomplete; the store serves the
+	// persisted row at runtime.
 	return cfg, nil
 }
 
@@ -84,19 +82,6 @@ func durationEnv(key string, fallback time.Duration) time.Duration {
 	return d
 }
 
-func firstNonEmpty(values ...string) string {
-	for _, v := range values {
-		if strings.TrimSpace(v) != "" {
-			return v
-		}
-	}
-	return ""
-}
-
-func defaultCredentialsPath() string {
-	return filepath.Join(".", "credentials.json")
-}
-
 const defaultPublicKeyPEM = `-----BEGIN PUBLIC KEY-----
 MIIBIjANBgkqhkiG9w0BAQEFAAOCAQ8AMIIBCgKCAQEAhxudxTewPgljUHEZHkusP7m3I+zA4/RGvuUMt6TtII/m4zwUOm/Y31zHBTmkCCt8k5vj9y+AmO0TsGmHooNQuMebakdmEWdcA5h7YAHHFbF2w5LcxIXjib08vgVpA+m3R5xPbLK+vfHe2aAX36b5nHReDNncY5vAl3U4CgIEBGPqyG67vJytRWqP+sfEdw5+m192Rf4SCGyiBzRmjiVlH3zeEBjdbOrkAnzKOVz6AHBl2q7LPLJKIzxjoAyhEp5qnDjHUFo5VZUgFwUOt83A/jbGMyzmjRoxBuvKcs9tBuorZyUwIsZN6E+rtQk2YqMPj4RkDsZ7LRmj6on8sN2rHQIDAQAB
 -----END PUBLIC KEY-----`

internal/openai/responses.go

@@ -0,0 +1,396 @@
+package openai
+
+import (
+	"encoding/json"
+	"errors"
+	"strings"
+)
+
+// --- Responses API request parsing & conversion ---
+
+// ParseResponsesRequest converts a raw OpenAI Responses API (POST /v1/responses)
+// request body into a ChatCompletionRequest suitable for the upstream gateway.
+// It translates input→messages, instructions→system message, max_output_tokens→
+// max_tokens, and reshapes tools from the Responses flat format to the Chat
+// Completions nested {function:{…}} format.
+func ParseResponsesRequest(body []byte) (ChatCompletionRequest, error) {
+	var raw map[string]json.RawMessage
+	if err := json.Unmarshal(body, &raw); err != nil {
+		return ChatCompletionRequest{}, err
+	}
+
+	var req ChatCompletionRequest
+	if v, ok := raw["model"]; ok {
+		_ = json.Unmarshal(v, &req.Model)
+	}
+	if v, ok := raw["stream"]; ok {
+		_ = json.Unmarshal(v, &req.Stream)
+	}
+
+	// Build messages from instructions + input.
+	messages := make([]map[string]any, 0, 4)
+	if v, ok := raw["instructions"]; ok {
+		var s string
+		if json.Unmarshal(v, &s) == nil && s != "" {
+			messages = append(messages, map[string]any{"role": "system", "content": s})
+		}
+	}
+	if v, ok := raw["input"]; ok {
+		msgs, err := responsesInputToMessages(v)
+		if err != nil {
+			return req, err
+		}
+		messages = append(messages, msgs...)
+	}
+	req.Messages = messages
+
+	// Extra: renamed + pass-through fields sent to upstream as-is.
+	extra := map[string]json.RawMessage{}
+	if v, ok := raw["max_output_tokens"]; ok {
+		extra["max_tokens"] = v
+	}
+	for _, key := range []string{
+		"temperature", "top_p", "top_k", "frequency_penalty",
+		"presence_penalty", "stop", "seed", "user",
+	} {
+		if v, ok := raw[key]; ok {
+			extra[key] = v
+		}
+	}
+	if v, ok := raw["tools"]; ok {
+		if translated, err := translateResponsesTools(v); err == nil {
+			extra["tools"] = translated
+		} else {
+			extra["tools"] = v // fall back to pass-through
+		}
+	}
+	if v, ok := raw["tool_choice"]; ok {
+		extra["tool_choice"] = v
+	}
+	// Structured Outputs: Responses API uses text.format instead of
+	// response_format. Translate text.format → response_format for the
+	// upstream Chat Completions endpoint.
+	if v, ok := raw["text"]; ok {
+		if rf, err := translateTextFormat(v); err == nil {
+			extra["response_format"] = rf
+		}
+	}
+	req.Extra = extra
+	return req, nil
+}
+
+// translateTextFormat converts the Responses API "text" field (containing a
+// "format" sub-object) into a Chat Completions response_format value.
+//   - {text:{format:{type:"json_object"}}}        → {type:"json_object"}
+//   - {text:{format:{type:"json_schema",name,schema,strict}}} →
+//     {type:"json_schema",json_schema:{name,schema,strict}}
+func translateTextFormat(textRaw json.RawMessage) (json.RawMessage, error) {
+	var text struct {
+		Format map[string]any `json:"format"`
+	}
+	if err := json.Unmarshal(textRaw, &text); err != nil {
+		return nil, err
+	}
+	if len(text.Format) == 0 {
+		return nil, errors.New("empty text.format")
+	}
+	t, _ := text.Format["type"].(string)
+	switch t {
+	case "json_object":
+		return json.Marshal(map[string]any{"type": "json_object"})
+	case "json_schema":
+		js := map[string]any{}
+		for _, k := range []string{"name", "schema", "strict"} {
+			if v, ok := text.Format[k]; ok {
+				js[k] = v
+			}
+		}
+		return json.Marshal(map[string]any{"type": "json_schema", "json_schema": js})
+	default:
+		return json.Marshal(text.Format) // pass through unknown types
+	}
+}
+
+// ResponsesMeta holds fields from the Responses API request that should be
+// echoed back in the response object (the OpenAI SDK requires them).
+type ResponsesMeta struct {
+	ParallelToolCalls any
+	ToolChoice        any
+	Tools             any
+	Temperature       any
+	TopP              any
+	MaxOutputTokens   any
+}
+
+// ParseResponsesMeta extracts echo-back fields from the raw request body.
+func ParseResponsesMeta(body []byte) ResponsesMeta {
+	var m map[string]any
+	_ = json.Unmarshal(body, &m)
+	meta := ResponsesMeta{
+		ParallelToolCalls: false,
+		ToolChoice:        "auto",
+		Tools:             []any{},
+	}
+	if v, ok := m["parallel_tool_calls"]; ok {
+		meta.ParallelToolCalls = v
+	}
+	if v, ok := m["tool_choice"]; ok {
+		meta.ToolChoice = v
+	}
+	// Tools are echoed back in the Responses API flat format (not the
+	// translated Chat Completions format).
+	if v, ok := m["tools"]; ok {
+		meta.Tools = v
+	} else {
+		meta.Tools = []any{}
+	}
+	if v, ok := m["temperature"]; ok {
+		meta.Temperature = v
+	}
+	if v, ok := m["top_p"]; ok {
+		meta.TopP = v
+	}
+	if v, ok := m["max_output_tokens"]; ok {
+		meta.MaxOutputTokens = v
+	}
+	return meta
+}
+
+// responsesInputToMessages converts the Responses API "input" field (which may
+// be a plain string or an array of input items) into Chat Completions messages.
+func responsesInputToMessages(input json.RawMessage) ([]map[string]any, error) {
+	// Case 1: input is a plain string.
+	var s string
+	if json.Unmarshal(input, &s) == nil {
+		return []map[string]any{{"role": "user", "content": s}}, nil
+	}
+
+	// Case 2: input is an array of items.
+	var items []map[string]json.RawMessage
+	if err := json.Unmarshal(input, &items); err != nil {
+		return nil, err
+	}
+
+	messages := make([]map[string]any, 0, len(items))
+	for _, item := range items {
+		// Determine the type — most items are message-like with a role.
+		var role string
+		if v, ok := item["role"]; ok {
+			_ = json.Unmarshal(v, &role)
+		}
+		var itemType string
+		if v, ok := item["type"]; ok {
+			_ = json.Unmarshal(v, &itemType)
+		}
+
+		switch {
+		case itemType == "function_call":
+			// An assistant tool call from a previous turn.
+			msg := map[string]any{"role": "assistant"}
+			tc := map[string]any{"type": "function"}
+			inner := map[string]any{}
+			if v, ok := item["name"]; ok {
+				var name string
+				_ = json.Unmarshal(v, &name)
+				inner["name"] = name
+			}
+			if v, ok := item["arguments"]; ok {
+				inner["arguments"] = rawJSONToString(v)
+			}
+			if v, ok := item["call_id"]; ok {
+				var id string
+				_ = json.Unmarshal(v, &id)
+				tc["id"] = id
+			}
+			tc["function"] = inner
+			msg["tool_calls"] = []any{tc}
+			messages = append(messages, msg)
+		case itemType == "function_call_output":
+			// A tool result from a previous turn.
+			msg := map[string]any{"role": "tool"}
+			if v, ok := item["call_id"]; ok {
+				var id string
+				_ = json.Unmarshal(v, &id)
+				msg["tool_call_id"] = id
+			}
+			if v, ok := item["output"]; ok {
+				msg["content"] = outputToString(v)
+			}
+			messages = append(messages, msg)
+		default:
+			// Standard message item with role + content.
+			if role == "" {
+				role = "user"
+			}
+			// "developer" maps to "system" for broad upstream compatibility.
+			if role == "developer" {
+				role = "system"
+			}
+			msg := map[string]any{"role": role}
+			if v, ok := item["content"]; ok {
+				msg["content"] = convertContentParts(v)
+			} else {
+				msg["content"] = ""
+			}
+			messages = append(messages, msg)
+		}
+	}
+	return messages, nil
+}
+
+// convertContentParts converts a Responses API content field (string or array
+// of content parts) into the Chat Completions content format.
+func convertContentParts(raw json.RawMessage) any {
+	// Content is a plain string.
+	var s string
+	if json.Unmarshal(raw, &s) == nil {
+		return s
+	}
+	// Content is an array of parts.
+	var parts []map[string]any
+	if json.Unmarshal(raw, &parts) != nil {
+		return string(raw) // fallback
+	}
+	result := make([]map[string]any, 0, len(parts))
+	for _, p := range parts {
+		pt, _ := p["type"].(string)
+		switch pt {
+		case "input_text", "output_text", "text":
+			result = append(result, map[string]any{"type": "text", "text": p["text"]})
+		case "input_image":
+			img := map[string]any{"type": "image_url"}
+			if url, ok := p["image_url"]; ok {
+				img["image_url"] = map[string]any{"url": url}
+			} else if d, ok := p["image"]; ok {
+				img["image_url"] = map[string]any{"url": d}
+			}
+			result = append(result, img)
+		default:
+			// Pass through unknown part types as-is.
+			result = append(result, p)
+		}
+	}
+	if len(result) == 0 {
+		return ""
+	}
+	return result
+}
+
+// rawJSONToString converts a json.RawMessage to a string. If the value is
+// already a JSON string, it is used directly. Any other JSON value (object,
+// array, number, bool) is re-encoded as a JSON string so it can populate
+// fields that require a string, such as tool_calls[].function.arguments.
+// This prevents silent data loss when a field arrives as a non-string type.
+func rawJSONToString(v json.RawMessage) string {
+	var s string
+	if json.Unmarshal(v, &s) == nil {
+		return s
+	}
+	var anyValue any
+	if json.Unmarshal(v, &anyValue) == nil {
+		if b, err := json.Marshal(anyValue); err == nil {
+			return string(b)
+		}
+	}
+	return string(v) // last resort: raw bytes
+}
+
+// outputToString converts a function_call_output "output" value to a string
+// for the Chat Completions tool message content. The output may be:
+//   - a plain string (used directly)
+//   - an array of content parts (text extracted and concatenated)
+//   - any other JSON value (re-encoded as a JSON string)
+func outputToString(v json.RawMessage) string {
+	var s string
+	if json.Unmarshal(v, &s) == nil {
+		return s
+	}
+	// Try array of content parts — extract text from each part.
+	var parts []map[string]any
+	if json.Unmarshal(v, &parts) == nil {
+		var sb strings.Builder
+		for _, p := range parts {
+			if t, _ := p["text"].(string); t != "" {
+				sb.WriteString(t)
+			}
+		}
+		if sb.Len() > 0 {
+			return sb.String()
+		}
+	}
+	// Fallback: re-encode as a JSON string.
+	var anyValue any
+	if json.Unmarshal(v, &anyValue) == nil {
+		if b, err := json.Marshal(anyValue); err == nil {
+			return string(b)
+		}
+	}
+	return string(v)
+}
+
+// translateResponsesTools converts tools from the Responses API flat format
+// to the Chat Completions nested {type:"function",function:{…}} format.
+func translateResponsesTools(raw json.RawMessage) (json.RawMessage, error) {
+	var tools []map[string]any
+	if err := json.Unmarshal(raw, &tools); err != nil {
+		return nil, err
+	}
+	out := make([]map[string]any, 0, len(tools))
+	for _, t := range tools {
+		tt, _ := t["type"].(string)
+		if tt != "function" && tt != "" {
+			// Non-function tool types (web_search, file_search, etc.) —
+			// pass through as-is; upstream may or may not support them.
+			out = append(out, t)
+			continue
+		}
+		fn := map[string]any{}
+		for _, key := range []string{"name", "description", "parameters", "strict"} {
+			if v, ok := t[key]; ok {
+				fn[key] = v
+			}
+		}
+		out = append(out, map[string]any{"type": "function", "function": fn})
+	}
+	return json.Marshal(out)
+}
+
+// --- Responses API response building ---
+
+// UsageToResponses converts a Chat Completions usage value (as decoded by
+// json.Unmarshal into any) to the Responses API usage field names
+// (input_tokens / output_tokens instead of prompt_tokens / completion_tokens).
+func UsageToResponses(v any) any {
+	if v == nil {
+		return nil
+	}
+	b, err := json.Marshal(v)
+	if err != nil {
+		return v
+	}
+	var m map[string]any
+	if json.Unmarshal(b, &m) != nil {
+		return v
+	}
+	out := map[string]any{}
+	if pt, ok := m["prompt_tokens"]; ok {
+		out["input_tokens"] = pt
+	}
+	if ct, ok := m["completion_tokens"]; ok {
+		out["output_tokens"] = ct
+	}
+	if tt, ok := m["total_tokens"]; ok {
+		out["total_tokens"] = tt
+	}
+	if d, ok := m["prompt_tokens_details"]; ok {
+		out["input_tokens_details"] = d
+	} else if d, ok := m["input_tokens_details"]; ok {
+		out["input_tokens_details"] = d
+	}
+	if d, ok := m["completion_tokens_details"]; ok {
+		out["output_tokens_details"] = d
+	} else if d, ok := m["output_tokens_details"]; ok {
+		out["output_tokens_details"] = d
+	}
+	return out
+}

internal/openai/responses_test.go

@@ -0,0 +1,389 @@
+package openai
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+func TestParseResponsesRequest_StringInput(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":"hello world","stream":false}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if req.Model != "GLM-4.7" {
+		t.Fatalf("model = %q", req.Model)
+	}
+	if len(req.Messages) != 1 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	if req.Messages[0]["role"] != "user" {
+		t.Fatalf("role = %v", req.Messages[0]["role"])
+	}
+	if req.Messages[0]["content"] != "hello world" {
+		t.Fatalf("content = %v", req.Messages[0]["content"])
+	}
+}
+
+func TestParseResponsesRequest_Instructions(t *testing.T) {
+	body := `{"model":"GLM-4.7","instructions":"be helpful","input":"hi"}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 2 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	if req.Messages[0]["role"] != "system" {
+		t.Fatalf("first role = %v", req.Messages[0]["role"])
+	}
+	if req.Messages[0]["content"] != "be helpful" {
+		t.Fatalf("first content = %v", req.Messages[0]["content"])
+	}
+	if req.Messages[1]["role"] != "user" {
+		t.Fatalf("second role = %v", req.Messages[1]["role"])
+	}
+}
+
+func TestParseResponsesRequest_ArrayInput(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"role":"user","content":"hello"},
+		{"role":"assistant","content":"hi there"},
+		{"role":"user","content":"how are you?"}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 3 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	if req.Messages[0]["role"] != "user" || req.Messages[0]["content"] != "hello" {
+		t.Fatalf("msg[0] = %v", req.Messages[0])
+	}
+	if req.Messages[1]["role"] != "assistant" || req.Messages[1]["content"] != "hi there" {
+		t.Fatalf("msg[1] = %v", req.Messages[1])
+	}
+	if req.Messages[2]["role"] != "user" || req.Messages[2]["content"] != "how are you?" {
+		t.Fatalf("msg[2] = %v", req.Messages[2])
+	}
+}
+
+func TestParseResponsesRequest_ContentParts(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"role":"user","content":[
+			{"type":"input_text","text":"describe this"},
+			{"type":"input_image","image_url":"data:image/png;base64,abc"}
+		]}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 1 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	content, ok := req.Messages[0]["content"].([]map[string]any)
+	if !ok {
+		t.Fatalf("content type = %T", req.Messages[0]["content"])
+	}
+	if len(content) != 2 {
+		t.Fatalf("content parts = %d", len(content))
+	}
+	if content[0]["type"] != "text" || content[0]["text"] != "describe this" {
+		t.Fatalf("content[0] = %v", content[0])
+	}
+	if content[1]["type"] != "image_url" {
+		t.Fatalf("content[1] type = %v", content[1])
+	}
+}
+
+func TestParseResponsesRequest_MaxOutputTokens(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":"hi","max_output_tokens":500}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if v, ok := req.Extra["max_tokens"]; !ok {
+		t.Fatal("max_tokens not in Extra")
+	} else {
+		var n int
+		_ = json.Unmarshal(v, &n)
+		if n != 500 {
+			t.Fatalf("max_tokens = %d", n)
+		}
+	}
+}
+
+func TestParseResponsesRequest_Tools(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":"hi","tools":[
+		{"type":"function","name":"get_weather","description":"get weather","parameters":{"type":"object"}}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	raw, ok := req.Extra["tools"]
+	if !ok {
+		t.Fatal("tools not in Extra")
+	}
+	var tools []map[string]any
+	if err := json.Unmarshal(raw, &tools); err != nil {
+		t.Fatal(err)
+	}
+	if len(tools) != 1 {
+		t.Fatalf("tools = %d", len(tools))
+	}
+	if tools[0]["type"] != "function" {
+		t.Fatalf("tool type = %v", tools[0]["type"])
+	}
+	fn, ok := tools[0]["function"].(map[string]any)
+	if !ok {
+		t.Fatalf("function type = %T", tools[0]["function"])
+	}
+	if fn["name"] != "get_weather" {
+		t.Fatalf("function name = %v", fn["name"])
+	}
+}
+
+func TestParseResponsesRequest_FunctionCallInput(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"role":"user","content":"what's the weather?"},
+		{"type":"function_call","call_id":"call_123","name":"get_weather","arguments":"{\"city\":\"NYC\"}"},
+		{"type":"function_call_output","call_id":"call_123","output":"sunny 72F"}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 3 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	// First message is user text
+	if req.Messages[0]["role"] != "user" {
+		t.Fatalf("msg[0] role = %v", req.Messages[0]["role"])
+	}
+	// Second message is assistant with tool_calls
+	if req.Messages[1]["role"] != "assistant" {
+		t.Fatalf("msg[1] role = %v", req.Messages[1]["role"])
+	}
+	tcs, ok := req.Messages[1]["tool_calls"].([]any)
+	if !ok || len(tcs) != 1 {
+		t.Fatalf("msg[1] tool_calls = %v", req.Messages[1]["tool_calls"])
+	}
+	tc, _ := tcs[0].(map[string]any)
+	if tc["id"] != "call_123" {
+		t.Fatalf("tool call id = %v", tc["id"])
+	}
+	fn, _ := tc["function"].(map[string]any)
+	if fn["name"] != "get_weather" {
+		t.Fatalf("function name = %v", fn["name"])
+	}
+	// Third message is tool result
+	if req.Messages[2]["role"] != "tool" {
+		t.Fatalf("msg[2] role = %v", req.Messages[2]["role"])
+	}
+	if req.Messages[2]["tool_call_id"] != "call_123" {
+		t.Fatalf("msg[2] tool_call_id = %v", req.Messages[2]["tool_call_id"])
+	}
+	if req.Messages[2]["content"] != "sunny 72F" {
+		t.Fatalf("msg[2] content = %v", req.Messages[2]["content"])
+	}
+}
+
+// TestParseResponsesRequest_FunctionCallObjectArguments verifies that when
+// function_call.arguments arrives as a JSON object (not a string), it is
+// re-encoded as a JSON string instead of being silently dropped.
+func TestParseResponsesRequest_FunctionCallObjectArguments(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"type":"function_call","call_id":"call_456","name":"task","arguments":{"operation":"create","summary":"test"}}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 1 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	tc, _ := req.Messages[0]["tool_calls"].([]any)
+	if len(tc) != 1 {
+		t.Fatalf("tool_calls = %v", req.Messages[0]["tool_calls"])
+	}
+	fn, _ := tc[0].(map[string]any)["function"].(map[string]any)
+	args, _ := fn["arguments"].(string)
+	if args == "" {
+		t.Fatalf("arguments was dropped (empty)")
+	}
+	// The re-encoded string must be valid JSON containing the original fields.
+	var parsed map[string]any
+	if err := json.Unmarshal([]byte(args), &parsed); err != nil {
+		t.Fatalf("arguments not valid JSON: %v", err)
+	}
+	if parsed["operation"] != "create" {
+		t.Fatalf("operation = %v", parsed["operation"])
+	}
+}
+
+// TestParseResponsesRequest_FunctionCallOutputArray verifies that when
+// function_call_output.output arrives as an array of content parts, the
+// text is extracted instead of being silently dropped.
+func TestParseResponsesRequest_FunctionCallOutputArray(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"type":"function_call","call_id":"call_789","name":"get_weather","arguments":"{\"city\":\"NYC\"}"},
+		{"type":"function_call_output","call_id":"call_789","output":[{"type":"output_text","text":"sunny 72F"}]}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 2 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	toolMsg := req.Messages[1]
+	if toolMsg["role"] != "tool" {
+		t.Fatalf("msg[1] role = %v", toolMsg["role"])
+	}
+	content, _ := toolMsg["content"].(string)
+	if content != "sunny 72F" {
+		t.Fatalf("content = %q, want %q", content, "sunny 72F")
+	}
+}
+
+// TestParseResponsesRequest_FunctionCallOutputObject verifies that when
+// function_call_output.output is a bare JSON object, it is re-encoded as a
+// JSON string instead of being silently dropped.
+func TestParseResponsesRequest_FunctionCallOutputObject(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"type":"function_call","call_id":"call_obj","name":"run","arguments":"{}"},
+		{"type":"function_call_output","call_id":"call_obj","output":{"result":"success","code":200}}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	toolMsg := req.Messages[1]
+	content, _ := toolMsg["content"].(string)
+	if content == "" {
+		t.Fatal("content was dropped (empty)")
+	}
+	var parsed map[string]any
+	if err := json.Unmarshal([]byte(content), &parsed); err != nil {
+		t.Fatalf("content not valid JSON: %v", err)
+	}
+	if parsed["result"] != "success" {
+		t.Fatalf("result = %v", parsed["result"])
+	}
+}
+
+func TestUsageToResponses(t *testing.T) {
+	usage := map[string]any{
+		"prompt_tokens":             10,
+		"completion_tokens":         20,
+		"total_tokens":              30,
+		"prompt_tokens_details":     map[string]any{"cached_tokens": 4},
+		"completion_tokens_details": map[string]any{"reasoning_tokens": 5},
+	}
+	result := UsageToResponses(usage)
+	m, ok := result.(map[string]any)
+	if !ok {
+		t.Fatalf("result type = %T", result)
+	}
+	if m["input_tokens"] != float64(10) {
+		t.Fatalf("input_tokens = %v", m["input_tokens"])
+	}
+	if m["output_tokens"] != float64(20) {
+		t.Fatalf("output_tokens = %v", m["output_tokens"])
+	}
+	if m["total_tokens"] != float64(30) {
+		t.Fatalf("total_tokens = %v", m["total_tokens"])
+	}
+	if d, ok := m["input_tokens_details"].(map[string]any); !ok || d["cached_tokens"] != float64(4) {
+		t.Fatalf("input_tokens_details = %v", m["input_tokens_details"])
+	}
+	if d, ok := m["output_tokens_details"].(map[string]any); !ok || d["reasoning_tokens"] != float64(5) {
+		t.Fatalf("output_tokens_details = %v", m["output_tokens_details"])
+	}
+}
+
+func TestParseResponsesRequest_TextFormat(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":"Jane, 54","text":{"format":{"type":"json_schema","name":"person","strict":true,"schema":{"type":"object"}}}}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	rf, ok := req.Extra["response_format"]
+	if !ok {
+		t.Fatal("response_format not in Extra")
+	}
+	var m map[string]any
+	if err := json.Unmarshal(rf, &m); err != nil {
+		t.Fatal(err)
+	}
+	if m["type"] != "json_schema" {
+		t.Fatalf("type = %v", m["type"])
+	}
+	js, ok := m["json_schema"].(map[string]any)
+	if !ok {
+		t.Fatalf("json_schema = %v", m["json_schema"])
+	}
+	if js["name"] != "person" {
+		t.Fatalf("name = %v", js["name"])
+	}
+	if js["strict"] != true {
+		t.Fatalf("strict = %v", js["strict"])
+	}
+}
+
+func TestParseResponsesRequest_TextFormatJsonObject(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":"hi","text":{"format":{"type":"json_object"}}}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	rf, ok := req.Extra["response_format"]
+	if !ok {
+		t.Fatal("response_format not in Extra")
+	}
+	var m map[string]any
+	if err := json.Unmarshal(rf, &m); err != nil {
+		t.Fatal(err)
+	}
+	if m["type"] != "json_object" {
+		t.Fatalf("type = %v", m["type"])
+	}
+}
+
+func TestParseResponsesMeta_Defaults(t *testing.T) {
+	meta := ParseResponsesMeta([]byte(`{"model":"GLM-4.7","input":"hi"}`))
+	if meta.ParallelToolCalls != false {
+		t.Fatalf("parallel_tool_calls = %v", meta.ParallelToolCalls)
+	}
+	if meta.ToolChoice != "auto" {
+		t.Fatalf("tool_choice = %v", meta.ToolChoice)
+	}
+	tools, ok := meta.Tools.([]any)
+	if !ok || len(tools) != 0 {
+		t.Fatalf("tools = %v", meta.Tools)
+	}
+}
+
+func TestParseResponsesMeta_EchoBack(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":"hi","parallel_tool_calls":true,"tool_choice":"required","tools":[{"type":"function","name":"get_weather"}],"temperature":0.7,"max_output_tokens":500}`
+	meta := ParseResponsesMeta([]byte(body))
+	if meta.ParallelToolCalls != true {
+		t.Fatalf("parallel_tool_calls = %v", meta.ParallelToolCalls)
+	}
+	if meta.ToolChoice != "required" {
+		t.Fatalf("tool_choice = %v", meta.ToolChoice)
+	}
+	tools, ok := meta.Tools.([]any)
+	if !ok || len(tools) != 1 {
+		t.Fatalf("tools = %v", meta.Tools)
+	}
+	if meta.Temperature != 0.7 {
+		t.Fatalf("temperature = %v", meta.Temperature)
+	}
+	if meta.MaxOutputTokens != float64(500) {
+		t.Fatalf("max_output_tokens = %v", meta.MaxOutputTokens)
+	}
+}

internal/openai/types.go

@@ -1,6 +1,9 @@
 package openai
 
-import "encoding/json"
+import (
+	"encoding/json"
+	"strings"
+)
 
 type ChatCompletionRequest struct {
 	Model    string                     `json:"model"`
@@ -30,7 +33,7 @@ func (r *ChatCompletionRequest) UnmarshalJSON(data []byte) error {
 func (r ChatCompletionRequest) MarshalForUpstream() ([]byte, error) {
 	m := map[string]any{
 		"model":          r.Model,
-		"messages":       r.Messages,
+		"messages":       normalizeMessages(r.Messages, extractToolNames(r.Extra)),
 		"stream":         r.Stream,
 		"stream_options": map[string]any{"include_usage": true},
 	}
@@ -40,9 +43,195 @@ func (r ChatCompletionRequest) MarshalForUpstream() ([]byte, error) {
 			m[k] = anyValue
 		}
 	}
+	// Mirror the Zhanlu plugin: GLM models require {tool_stream:true} to
+	// stream tool calls back to the client (matches /glm/i.test(model)).
+	if strings.Contains(strings.ToLower(r.Model), "glm") {
+		m["tool_stream"] = true
+	}
 	return json.Marshal(m)
 }
 
+// extractToolNames returns the set of function names declared in the tools
+// array (Extra["tools"]), used to validate tool_calls in the message history.
+func extractToolNames(extra map[string]json.RawMessage) map[string]bool {
+	names := map[string]bool{}
+	raw, ok := extra["tools"]
+	if !ok {
+		return names
+	}
+	var tools []any
+	if json.Unmarshal(raw, &tools) != nil {
+		return names
+	}
+	for _, t := range tools {
+		tm, ok := t.(map[string]any)
+		if !ok {
+			continue
+		}
+		// Chat Completions format: {type:"function",function:{name:"…"}}
+		if fn, ok := tm["function"].(map[string]any); ok {
+			if name, _ := fn["name"].(string); name != "" {
+				names[name] = true
+			}
+		}
+		// Responses API flat format: {type:"function",name:"…"}
+		if name, _ := tm["name"].(string); name != "" {
+			names[name] = true
+		}
+	}
+	return names
+}
+
+// normalizeMessages prepares the message history for the Zhanlu upstream
+// gateway. It performs two normalizations:
+//
+//  1. Flattens OpenAI multi-part content — an array of {type:"text",text:"…"}
+//     parts — into a plain string. The gateway only accepts string content
+//     and returns HTTP 400 for arrays.
+//
+//  2. Sanitizes tool_calls that the gateway would reject: calls whose
+//     function name is not in the declared tools set (e.g. "invalid"
+//     placeholders from AI coding tools when a tool call fails), or whose
+//     arguments are not a valid JSON object. Instead of removing these
+//     calls (which would lose error feedback the model needs), the function
+//     name is replaced with a valid one and the arguments are replaced with
+//     "{}". The corresponding tool result messages are preserved so the
+//     model still sees the full conversation including errors.
+//
+// Both []any (from JSON decoding of /v1/chat/completions) and
+// []map[string]any (from convertContentParts in the Responses API path)
+// are handled for content arrays.
+func normalizeMessages(messages []map[string]any, validToolNames map[string]bool) []map[string]any {
+	fallbackName := pickFallbackToolName(validToolNames)
+
+	for _, msg := range messages {
+		// --- flatten multi-part text content ---
+		parts := contentAsAnySlice(msg["content"])
+		if len(parts) > 0 {
+			var sb strings.Builder
+			allText := true
+			for _, p := range parts {
+				part, ok := p.(map[string]any)
+				if !ok {
+					allText = false
+					break
+				}
+				pt, _ := part["type"].(string)
+				if pt != "text" && pt != "input_text" && pt != "output_text" {
+					allText = false
+					break
+				}
+				text, _ := part["text"].(string)
+				sb.WriteString(text)
+			}
+			if allText {
+				msg["content"] = sb.String()
+			}
+		}
+
+		// --- sanitize tool_calls in place ---
+		tcs := contentAsAnySlice(msg["tool_calls"])
+		for _, tc := range tcs {
+			if tcMap, ok := tc.(map[string]any); ok {
+				sanitizeToolCall(tcMap, validToolNames, fallbackName)
+			}
+		}
+	}
+	return messages
+}
+
+// sanitizeToolCall fixes a tool_call in place so the upstream gateway
+// accepts it. Two fields are corrected:
+//
+//   - function.name: if the name is not among the declared tools (when the
+//     set is non-empty), it is replaced. The replacement is extracted from
+//     the arguments' "tool" field (MiMoCode stores the real tool name there
+//     in error placeholders); failing that, an arbitrary declared tool name
+//     is used.
+//
+//   - function.arguments: if the value is not a valid JSON object string
+//     (e.g. "-1", "", "true"), it is replaced with "{}".
+//
+// The tool_call id, type, and the tool result messages are left untouched,
+// preserving the full conversation context for the model.
+func sanitizeToolCall(tc map[string]any, validToolNames map[string]bool, fallbackName string) {
+	fn, ok := tc["function"].(map[string]any)
+	if !ok {
+		return
+	}
+	// Fix function name if not in the valid set.
+	name, _ := fn["name"].(string)
+	if len(validToolNames) > 0 && !validToolNames[name] {
+		args, _ := fn["arguments"].(string)
+		if extracted := extractToolNameFromArgs(args); validToolNames[extracted] {
+			fn["name"] = extracted
+		} else if fallbackName != "" {
+			fn["name"] = fallbackName
+		}
+	}
+	// Fix arguments if not a valid JSON object.
+	args, _ := fn["arguments"].(string)
+	if !isValidJSONObject(args) {
+		fn["arguments"] = "{}"
+	}
+}
+
+// extractToolNameFromArgs attempts to read a "tool" field from the JSON
+// object in args. MiMoCode's error placeholders store the real tool name
+// here (e.g. {"tool":"task","error":"…"}).
+func extractToolNameFromArgs(args string) string {
+	var m map[string]any
+	if json.Unmarshal([]byte(args), &m) != nil {
+		return ""
+	}
+	tool, _ := m["tool"].(string)
+	return tool
+}
+
+// pickFallbackToolName returns an arbitrary name from the set for use as a
+// last-resort replacement when no real tool name can be extracted.
+func pickFallbackToolName(names map[string]bool) string {
+	for name := range names {
+		return name
+	}
+	return ""
+}
+
+// isValidJSONObject reports whether s is a non-empty JSON string that decodes
+// to a JSON object (map[string]any). The OpenAI tool_call arguments field
+// must be a JSON object string like {"key":"value"}; the Zhanlu gateway
+// rejects values like "-1", "true", or "" with HTTP 400.
+func isValidJSONObject(s string) bool {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return false
+	}
+	var v any
+	if err := json.Unmarshal([]byte(s), &v); err != nil {
+		return false
+	}
+	_, ok := v.(map[string]any)
+	return ok
+}
+
+// contentAsAnySlice returns the content value as a []any, handling both
+// []any (JSON decoding) and []map[string]any (convertContentParts). It
+// returns nil for non-slice values (strings, nil, etc.).
+func contentAsAnySlice(v any) []any {
+	switch s := v.(type) {
+	case []any:
+		return s
+	case []map[string]any:
+		result := make([]any, len(s))
+		for i, m := range s {
+			result[i] = m
+		}
+		return result
+	default:
+		return nil
+	}
+}
+
 type ErrorResponse struct {
 	Error ErrorBody `json:"error"`
 }

internal/openai/types_test.go

@@ -0,0 +1,486 @@
+package openai
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func TestMarshalForUpstream_StringContent(t *testing.T) {
+	req := ChatCompletionRequest{
+		Model:    "GLM-4.7",
+		Messages:  []map[string]any{{"role": "user", "content": "hello"}},
+		Stream:   true,
+	}
+	body, err := req.MarshalForUpstream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if err := json.Unmarshal(body, &m); err != nil {
+		t.Fatal(err)
+	}
+	messages, _ := m["messages"].([]any)
+	msg, _ := messages[0].(map[string]any)
+	if content, _ := msg["content"].(string); content != "hello" {
+		t.Fatalf("content = %v, want %q", msg["content"], "hello")
+	}
+}
+
+// TestMarshalForUpstream_FlattensArrayContent verifies that multi-part
+// content arrays (the format used by OpenAI SDKs and AI coding tools) are
+// flattened into a plain string so the Zhanlu upstream gateway accepts the
+// request instead of returning HTTP 400.
+func TestMarshalForUpstream_FlattensArrayContent(t *testing.T) {
+	req := ChatCompletionRequest{
+		Model: "zhanlu/deepseek-v4-pro",
+		Messages: []map[string]any{
+			{"role": "user", "content": []any{
+				map[string]any{"type": "text", "text": "first part "},
+				map[string]any{"type": "text", "text": "second part"},
+			}},
+		},
+		Stream: true,
+	}
+	body, err := req.MarshalForUpstream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if err := json.Unmarshal(body, &m); err != nil {
+		t.Fatal(err)
+	}
+	messages, _ := m["messages"].([]any)
+	msg, _ := messages[0].(map[string]any)
+	content, ok := msg["content"].(string)
+	if !ok {
+		t.Fatalf("content type = %T, want string", msg["content"])
+	}
+	if content != "first part second part" {
+		t.Fatalf("content = %q, want %q", content, "first part second part")
+	}
+}
+
+// TestMarshalForUpstream_PreservesNonTextArrayContent verifies that content
+// arrays containing non-text parts (e.g. images) are left intact rather than
+// being flattened, so the upstream can handle multimodal content.
+func TestMarshalForUpstream_PreservesNonTextArrayContent(t *testing.T) {
+	req := ChatCompletionRequest{
+		Model: "GLM-4.7",
+		Messages: []map[string]any{
+			{"role": "user", "content": []any{
+				map[string]any{"type": "text", "text": "describe this"},
+				map[string]any{"type": "image_url", "image_url": map[string]any{"url": "data:image/png;base64,abc"}},
+			}},
+		},
+		Stream: true,
+	}
+	body, err := req.MarshalForUpstream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if err := json.Unmarshal(body, &m); err != nil {
+		t.Fatal(err)
+	}
+	messages, _ := m["messages"].([]any)
+	msg, _ := messages[0].(map[string]any)
+	parts, ok := msg["content"].([]any)
+	if !ok {
+		t.Fatalf("content type = %T, want []any (preserved array)", msg["content"])
+	}
+	if len(parts) != 2 {
+		t.Fatalf("parts = %d, want 2", len(parts))
+	}
+}
+
+// TestMarshalForUpstream_ExtraFields verifies that extra fields (max_tokens,
+// tools, etc.) are passed through to the upstream body.
+func TestMarshalForUpstream_ExtraFields(t *testing.T) {
+	req := ChatCompletionRequest{
+		Model:   "GLM-4.7",
+		Stream:  true,
+		Extra: map[string]json.RawMessage{
+			"max_tokens":  json.RawMessage(`32000`),
+			"tool_choice": json.RawMessage(`"auto"`),
+		},
+	}
+	body, err := req.MarshalForUpstream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if err := json.Unmarshal(body, &m); err != nil {
+		t.Fatal(err)
+	}
+	if v, _ := m["max_tokens"].(float64); v != 32000 {
+		t.Fatalf("max_tokens = %v, want 32000", m["max_tokens"])
+	}
+	if v, _ := m["tool_choice"].(string); v != "auto" {
+		t.Fatalf("tool_choice = %v, want \"auto\"", m["tool_choice"])
+	}
+}
+
+// TestMarshalForUpstream_GLMToolStream verifies that tool_stream is added
+// for GLM models but not for non-GLM models.
+func TestMarshalForUpstream_GLMToolStream(t *testing.T) {
+	cases := []struct {
+		model    string
+		wantTool bool
+	}{
+		{"GLM-4.7", true},
+		{"zhanlu/glm-4-pro", true},
+		{"zhanlu/deepseek-v4-pro", false},
+		{"deepseek-chat", false},
+	}
+	for _, tc := range cases {
+		req := ChatCompletionRequest{Model: tc.model, Stream: true}
+		body, err := req.MarshalForUpstream()
+		if err != nil {
+			t.Fatal(err)
+		}
+		var m map[string]any
+		_ = json.Unmarshal(body, &m)
+		_, has := m["tool_stream"]
+		if has != tc.wantTool {
+			t.Fatalf("model %q: tool_stream present = %v, want %v", tc.model, has, tc.wantTool)
+		}
+	}
+}
+
+// TestNormalizeMessages_MixedContent verifies a mix of string and array
+// content across multiple messages in a single request.
+func TestNormalizeMessages_MixedContent(t *testing.T) {
+	messages := []map[string]any{
+		{"role": "system", "content": "you are helpful"},
+		{"role": "user", "content": []any{
+			map[string]any{"type": "text", "text": "part 1 "},
+			map[string]any{"type": "text", "text": "part 2"},
+		}},
+		{"role": "assistant", "content": "ok"},
+		{"role": "user", "content": []any{
+			map[string]any{"type": "text", "text": "only part"},
+		}},
+	}
+	result := normalizeMessages(messages, nil)
+	// system message unchanged
+	if c, _ := result[0]["content"].(string); c != "you are helpful" {
+		t.Fatalf("msg[0] content = %v", result[0]["content"])
+	}
+	// array content flattened
+	if c, _ := result[1]["content"].(string); c != "part 1 part 2" {
+		t.Fatalf("msg[1] content = %v, want %q", result[1]["content"], "part 1 part 2")
+	}
+	// string content unchanged
+	if c, _ := result[2]["content"].(string); c != "ok" {
+		t.Fatalf("msg[2] content = %v", result[2]["content"])
+	}
+	// single-part array flattened to string
+	if c, _ := result[3]["content"].(string); c != "only part" {
+		t.Fatalf("msg[3] content = %v, want %q", result[3]["content"], "only part")
+	}
+}
+
+// TestNormalizeMessages_EmptyArray verifies that an empty content array
+// becomes an empty string.
+func TestNormalizeMessages_EmptyArray(t *testing.T) {
+	messages := []map[string]any{
+		{"role": "user", "content": []any{}},
+	}
+	result := normalizeMessages(messages, nil)
+	if c, _ := result[0]["content"].(string); c != "" {
+		t.Fatalf("content = %v, want empty string", result[0]["content"])
+	}
+}
+
+// TestNormalizeMessages_MapSliceContent verifies that content stored as
+// []map[string]any (produced by convertContentParts in the Responses API
+// path) is also flattened, not just []any (from JSON decoding).
+func TestNormalizeMessages_MapSliceContent(t *testing.T) {
+	messages := []map[string]any{
+		{"role": "user", "content": []map[string]any{
+			{"type": "text", "text": "map part 1 "},
+			{"type": "text", "text": "map part 2"},
+		}},
+	}
+	result := normalizeMessages(messages, nil)
+	if c, _ := result[0]["content"].(string); c != "map part 1 map part 2" {
+		t.Fatalf("content = %v, want %q", result[0]["content"], "map part 1 map part 2")
+	}
+}
+
+// TestMarshalForUpstream_ErrorRequestReplay is a regression test mirroring
+// the real failing request from error.md: a multi-turn conversation with
+// system prompt, user messages with multi-part content arrays, assistant
+// messages with tool_calls and reasoning_content, and tool result messages.
+// It verifies that no message has array content after marshaling.
+func TestMarshalForUpstream_ErrorRequestReplay(t *testing.T) {
+	req := ChatCompletionRequest{
+		Model: "zhanlu/deepseek-v4-pro",
+		Messages: []map[string]any{
+			{"role": "system", "content": strings.Repeat("system prompt ", 100)},
+			{"role": "user", "content": []any{
+				map[string]any{"type": "text", "text": "user question"},
+				map[string]any{"type": "text", "text": "<system-reminder>skill search</system-reminder>"},
+			}},
+			{"role": "assistant", "content": "I will search.", "reasoning_content": "thinking...", "tool_calls": []any{
+				map[string]any{"id": "call_1", "type": "function", "function": map[string]any{"name": "bash", "arguments": `{"command":"ls"}`}},
+			}},
+			{"role": "tool", "tool_call_id": "call_1", "content": "file1\nfile2"},
+			{"role": "user", "content": []any{
+				map[string]any{"type": "text", "text": "continue"},
+			}},
+		},
+		Stream: true,
+		Extra: map[string]json.RawMessage{
+			"max_tokens":  json.RawMessage(`32000`),
+			"tool_choice": json.RawMessage(`"auto"`),
+		},
+	}
+	body, err := req.MarshalForUpstream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	var m map[string]any
+	if err := json.Unmarshal(body, &m); err != nil {
+		t.Fatal(err)
+	}
+	messages, _ := m["messages"].([]any)
+	for i, raw := range messages {
+		msg, _ := raw.(map[string]any)
+		switch c := msg["content"].(type) {
+		case string:
+			// OK — flattened
+		case []any:
+			t.Fatalf("message [%d] still has array content after marshal", i)
+		default:
+			t.Fatalf("message [%d] content type = %T", i, c)
+		}
+	}
+}
+
+// TestNormalizeMessages_SanitizesInvalidToolCallNames verifies that
+// tool_calls whose function name is not in the declared tools set are
+// fixed in place — the name is replaced with a valid one — rather than
+// removed. This preserves the conversation context including error
+// feedback. MiMoCode emits {name:"invalid"} placeholders when a tool
+// call fails; the Zhanlu gateway rejects unknown function names with
+// HTTP 400.
+func TestNormalizeMessages_SanitizesInvalidToolCallNames(t *testing.T) {
+	validNames := map[string]bool{"bash": true, "read": true}
+	messages := []map[string]any{
+		{"role": "user", "content": "hello"},
+		{"role": "assistant", "content": "I'll run bash.", "tool_calls": []any{
+			map[string]any{"id": "call_ok", "type": "function", "function": map[string]any{"name": "bash", "arguments": `{"command":"ls"}`}},
+			// "invalid" name with "tool" field in args — should be extracted
+			map[string]any{"id": "call_bad", "type": "function", "function": map[string]any{"name": "invalid", "arguments": `{"tool":"read","error":"failed"}`}},
+		}},
+		{"role": "tool", "tool_call_id": "call_ok", "content": "file1"},
+		{"role": "tool", "tool_call_id": "call_bad", "content": "error result"},
+		{"role": "user", "content": "thanks"},
+	}
+	result := normalizeMessages(messages, validNames)
+	// All messages preserved (no removal)
+	if len(result) != 5 {
+		t.Fatalf("messages = %d, want 5 (sanitize preserves all)", len(result))
+	}
+	// Both tool_calls kept; the invalid one's name should be fixed to "read"
+	// (extracted from args.tool)
+	asst := result[1]
+	tcs, _ := asst["tool_calls"].([]any)
+	if len(tcs) != 2 {
+		t.Fatalf("tool_calls = %d items, want 2 (preserved)", len(tcs))
+	}
+	tc1, _ := tcs[0].(map[string]any)
+	fn1, _ := tc1["function"].(map[string]any)
+	if name, _ := fn1["name"].(string); name != "bash" {
+		t.Fatalf("first tool_call name = %q, want %q", name, "bash")
+	}
+	tc2, _ := tcs[1].(map[string]any)
+	fn2, _ := tc2["function"].(map[string]any)
+	if name, _ := fn2["name"].(string); name != "read" {
+		t.Fatalf("sanitized tool_call name = %q, want %q", name, "read")
+	}
+	// Tool results preserved
+	toolMsg := result[3]
+	if id, _ := toolMsg["tool_call_id"].(string); id != "call_bad" {
+		t.Fatalf("tool_call_id = %q, want %q", id, "call_bad")
+	}
+}
+
+// TestNormalizeMessages_SanitizesInvalidArguments verifies that tool_calls
+// whose arguments are not a valid JSON object string (e.g. "-1", "",
+// "true") are fixed to "{}" in place, not removed. The Zhanlu gateway
+// requires arguments to be a JSON object.
+func TestNormalizeMessages_SanitizesInvalidArguments(t *testing.T) {
+	validNames := map[string]bool{"task": true}
+	messages := []map[string]any{
+		{"role": "user", "content": "create tasks"},
+		{"role": "assistant", "content": "creating.", "tool_calls": []any{
+			map[string]any{"id": "call_good", "type": "function", "function": map[string]any{"name": "task", "arguments": `{"operation":"create","summary":"test"}`}},
+			map[string]any{"id": "call_bad1", "type": "function", "function": map[string]any{"name": "task", "arguments": "-1"}},
+			map[string]any{"id": "call_bad2", "type": "function", "function": map[string]any{"name": "task", "arguments": ""}},
+		}},
+		{"role": "tool", "tool_call_id": "call_good", "content": "created"},
+		{"role": "tool", "tool_call_id": "call_bad1", "content": "error"},
+		{"role": "tool", "tool_call_id": "call_bad2", "content": "error"},
+	}
+	result := normalizeMessages(messages, validNames)
+	// All 5 messages preserved
+	if len(result) != 5 {
+		t.Fatalf("messages = %d, want 5 (sanitize preserves all)", len(result))
+	}
+	asst := result[1]
+	tcs, _ := asst["tool_calls"].([]any)
+	if len(tcs) != 3 {
+		t.Fatalf("tool_calls = %d items, want 3 (preserved)", len(tcs))
+	}
+	// Good args unchanged
+	tc0, _ := tcs[0].(map[string]any)
+	fn0, _ := tc0["function"].(map[string]any)
+	if args, _ := fn0["arguments"].(string); args != `{"operation":"create","summary":"test"}` {
+		t.Fatalf("good args changed: %q", args)
+	}
+	// Bad args fixed to "{}"
+	tc1, _ := tcs[1].(map[string]any)
+	fn1, _ := tc1["function"].(map[string]any)
+	if args, _ := fn1["arguments"].(string); args != "{}" {
+		t.Fatalf("bad1 args = %q, want {}", args)
+	}
+	tc2, _ := tcs[2].(map[string]any)
+	fn2, _ := tc2["function"].(map[string]any)
+	if args, _ := fn2["arguments"].(string); args != "{}" {
+		t.Fatalf("bad2 args = %q, want {}", args)
+	}
+}
+
+// TestNormalizeMessages_KeepsValidToolCalls verifies that valid tool_calls
+// (correct name and valid JSON object arguments) are preserved unchanged.
+func TestNormalizeMessages_KeepsValidToolCalls(t *testing.T) {
+	validNames := map[string]bool{"bash": true, "read": true}
+	messages := []map[string]any{
+		{"role": "user", "content": "list files"},
+		{"role": "assistant", "content": "sure.", "tool_calls": []any{
+			map[string]any{"id": "call_1", "type": "function", "function": map[string]any{"name": "bash", "arguments": `{"command":"ls"}`}},
+			map[string]any{"id": "call_2", "type": "function", "function": map[string]any{"name": "read", "arguments": `{"file":"a.txt"}`}},
+		}},
+		{"role": "tool", "tool_call_id": "call_1", "content": "file1"},
+		{"role": "tool", "tool_call_id": "call_2", "content": "content"},
+	}
+	result := normalizeMessages(messages, validNames)
+	if len(result) != 4 {
+		t.Fatalf("messages = %d, want 4", len(result))
+	}
+	asst := result[1]
+	tcs, _ := asst["tool_calls"].([]any)
+	if len(tcs) != 2 {
+		t.Fatalf("tool_calls = %d items, want 2", len(tcs))
+	}
+	// Verify the first tool_call is unchanged
+	tc, _ := tcs[0].(map[string]any)
+	fn, _ := tc["function"].(map[string]any)
+	if name, _ := fn["name"].(string); name != "bash" {
+		t.Fatalf("name = %q, want %q", name, "bash")
+	}
+	if args, _ := fn["arguments"].(string); args != `{"command":"ls"}` {
+		t.Fatalf("args = %q, want %q", args, `{"command":"ls"}`)
+	}
+}
+
+// TestNormalizeMessages_NoToolNamesSkipsNameCheck verifies that when no
+// tools are declared (empty map), tool_calls are not name-sanitized
+// (only argument validity is checked).
+func TestNormalizeMessages_NoToolNamesSkipsNameCheck(t *testing.T) {
+	messages := []map[string]any{
+		{"role": "user", "content": "hi"},
+		{"role": "assistant", "content": "ok.", "tool_calls": []any{
+			map[string]any{"id": "call_1", "type": "function", "function": map[string]any{"name": "custom_fn", "arguments": `{"x":1}`}},
+		}},
+		{"role": "tool", "tool_call_id": "call_1", "content": "result"},
+	}
+	result := normalizeMessages(messages, nil)
+	if len(result) != 3 {
+		t.Fatalf("messages = %d, want 3", len(result))
+	}
+	asst := result[1]
+	tcs, _ := asst["tool_calls"].([]any)
+	tc, _ := tcs[0].(map[string]any)
+	fn, _ := tc["function"].(map[string]any)
+	// Name should be unchanged (no tools to validate against)
+	if name, _ := fn["name"].(string); name != "custom_fn" {
+		t.Fatalf("name = %q, want %q (should not be changed)", name, "custom_fn")
+	}
+}
+
+// TestNormalizeMessages_ErrorMdScenario mirrors the exact error.md
+// request: an assistant message with 5 tool_calls where 4 have
+// name:"invalid" (with args containing {"tool":"task","error":"…"}) and 1
+// has name:"task" with args:"-1". Verifies that after sanitization, all
+// 5 calls have valid names and valid JSON object arguments.
+func TestNormalizeMessages_ErrorMdScenario(t *testing.T) {
+	validNames := map[string]bool{"task": true, "bash": true}
+	messages := []map[string]any{
+		{"role": "user", "content": "create tasks"},
+		{"role": "assistant", "content": "confirming.", "tool_calls": []any{
+			map[string]any{"id": "call_1", "type": "function", "function": map[string]any{"name": "invalid", "arguments": `{"tool":"task","error":"JSON parsing failed"}`}},
+			map[string]any{"id": "call_2", "type": "function", "function": map[string]any{"name": "invalid", "arguments": `{"tool":"task","error":"JSON parsing failed"}`}},
+			map[string]any{"id": "call_3", "type": "function", "function": map[string]any{"name": "task", "arguments": "-1"}},
+			map[string]any{"id": "call_4", "type": "function", "function": map[string]any{"name": "invalid", "arguments": `{"tool":"task","error":"JSON parsing failed"}`}},
+			map[string]any{"id": "call_5", "type": "function", "function": map[string]any{"name": "invalid", "arguments": `{"tool":"task","error":"JSON parsing failed"}`}},
+		}},
+		{"role": "tool", "tool_call_id": "call_1", "content": "error 1"},
+		{"role": "tool", "tool_call_id": "call_2", "content": "error 2"},
+		{"role": "tool", "tool_call_id": "call_3", "content": "error 3"},
+		{"role": "tool", "tool_call_id": "call_4", "content": "error 4"},
+		{"role": "tool", "tool_call_id": "call_5", "content": "error 5"},
+	}
+	result := normalizeMessages(messages, validNames)
+	// All 7 messages preserved (no removal)
+	if len(result) != 7 {
+		t.Fatalf("messages = %d, want 7", len(result))
+	}
+	asst := result[1]
+	tcs, _ := asst["tool_calls"].([]any)
+	if len(tcs) != 5 {
+		t.Fatalf("tool_calls = %d items, want 5", len(tcs))
+	}
+	for i, raw := range tcs {
+		tc, _ := raw.(map[string]any)
+		fn, _ := tc["function"].(map[string]any)
+		name, _ := fn["name"].(string)
+		args, _ := fn["arguments"].(string)
+		// All names should be "task" (extracted from args or already valid)
+		if name != "task" {
+			t.Errorf("tool_call[%d] name = %q, want %q", i, name, "task")
+		}
+		// All args should be valid JSON objects
+		if !isValidJSONObject(args) {
+			t.Errorf("tool_call[%d] args = %q, not a valid JSON object", i, args)
+		}
+	}
+}
+
+// TestIsValidJSONObject verifies the JSON object validation used to filter
+// tool_call arguments.
+func TestIsValidJSONObject(t *testing.T) {
+	cases := []struct {
+		input string
+		want  bool
+	}{
+		{`{"key":"value"}`, true},
+		{`{}`, true},
+		{`{"nested":{"a":1}}`, true},
+		{`-1`, false},
+		{`true`, false},
+		{`"string"`, false},
+		{`[1,2,3]`, false},
+		{``, false},
+		{`   `, false},
+		{`{invalid json}`, false},
+	}
+	for _, tc := range cases {
+		if got := isValidJSONObject(tc.input); got != tc.want {
+			t.Errorf("isValidJSONObject(%q) = %v, want %v", tc.input, got, tc.want)
+		}
+	}
+}

internal/server/recorder.go

@@ -0,0 +1,174 @@
+package server
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"net/http"
+	"sync"
+	"time"
+)
+
+const (
+	maxRecentEntries = 10
+	maxBodyCapture    = 1 << 20 // 1 MB per body — full request/response capture for debugging
+)
+
+// RecentEntry captures a single API request and its response for debugging.
+type RecentEntry struct {
+	Timestamp      time.Time         `json:"timestamp"`
+	Method         string            `json:"method"`
+	Path           string            `json:"path"`
+	Status         int               `json:"status"`
+	DurationMs     int64             `json:"duration_ms"`
+	RequestHeaders map[string]string `json:"request_headers"`
+	RequestBody    string            `json:"request_body"`
+	ResponseBody   string            `json:"response_body"`
+}
+
+// RecentRecorder is an in-memory ring buffer that stores the most recent API
+// requests. It is safe for concurrent use.
+type RecentRecorder struct {
+	mu      sync.Mutex
+	entries []RecentEntry
+}
+
+func NewRecentRecorder() *RecentRecorder {
+	return &RecentRecorder{}
+}
+
+// Record appends an entry, evicting the oldest when the buffer is full.
+func (r *RecentRecorder) Record(e RecentEntry) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.entries = append(r.entries, e)
+	if len(r.entries) > maxRecentEntries {
+		r.entries = r.entries[len(r.entries)-maxRecentEntries:]
+	}
+}
+
+// Entries returns a copy of the buffer in newest-first order.
+func (r *RecentRecorder) Entries() []RecentEntry {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	n := len(r.entries)
+	out := make([]RecentEntry, n)
+	for i, e := range r.entries {
+		out[n-1-i] = e // reverse: newest first
+	}
+	return out
+}
+
+// recordingResponseWriter wraps http.ResponseWriter to capture the status code
+// and a truncated copy of the response body. It implements http.Flusher so
+// streaming handlers can flush through the wrapper.
+type recordingResponseWriter struct {
+	http.ResponseWriter
+	statusCode  int
+	body        bytes.Buffer
+	totalBytes  int
+}
+
+func newRecordingResponseWriter(w http.ResponseWriter) *recordingResponseWriter {
+	return &recordingResponseWriter{ResponseWriter: w, statusCode: http.StatusOK}
+}
+
+func (w *recordingResponseWriter) WriteHeader(code int) {
+	w.statusCode = code
+	w.ResponseWriter.WriteHeader(code)
+}
+
+func (w *recordingResponseWriter) Write(b []byte) (int, error) {
+	w.totalBytes += len(b)
+	if w.body.Len() < maxBodyCapture {
+		remaining := maxBodyCapture - w.body.Len()
+		if len(b) <= remaining {
+			w.body.Write(b)
+		} else {
+			w.body.Write(b[:remaining])
+		}
+	}
+	return w.ResponseWriter.Write(b)
+}
+
+func (w *recordingResponseWriter) Flush() {
+	if f, ok := w.ResponseWriter.(http.Flusher); ok {
+		f.Flush()
+	}
+}
+
+// captureRequestHeaders extracts selected request headers, masking the
+// Authorization value to avoid leaking API keys.
+func captureRequestHeaders(r *http.Request) map[string]string {
+	headers := map[string]string{}
+	for _, key := range []string{"Content-Type", "User-Agent", "Accept", "Authorization"} {
+		if v := r.Header.Get(key); v != "" {
+			if key == "Authorization" {
+				headers[key] = mask(v)
+			} else {
+				headers[key] = v
+			}
+		}
+	}
+	return headers
+}
+
+// truncateBody returns the string form of b, truncated to maxBodyCapture
+// bytes with a marker if the original was longer.
+func truncateBody(b []byte) string {
+	if len(b) > maxBodyCapture {
+		return string(b[:maxBodyCapture]) + fmt.Sprintf("\n...(truncated, total %d bytes)", len(b))
+	}
+	return string(b)
+}
+
+// formatResponseBody returns the captured response body, with a truncation
+// marker if the full response exceeded the capture limit.
+func formatResponseBody(buf *bytes.Buffer, totalBytes int) string {
+	s := buf.String()
+	if totalBytes > maxBodyCapture {
+		s += fmt.Sprintf("\n...(truncated, total %d bytes)", totalBytes)
+	}
+	return s
+}
+
+// withRecording wraps a handler so that each request's headers, body,
+// response status, and response body (truncated) are captured into the
+// server's RecentRecorder. It is applied to the /v1/ API routes so that
+// both successful and error responses are recorded.
+func (s *Server) withRecording(next http.HandlerFunc) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		start := time.Now()
+
+		// Read and restore the request body so the downstream handler
+		// still sees the full content.
+		var reqBody []byte
+		if r.Body != nil {
+			reqBody, _ = io.ReadAll(r.Body)
+			r.Body = io.NopCloser(bytes.NewReader(reqBody))
+		}
+
+		recW := newRecordingResponseWriter(w)
+		next(recW, r)
+
+		s.recorder.Record(RecentEntry{
+			Timestamp:      start,
+			Method:         r.Method,
+			Path:           r.URL.Path,
+			Status:         recW.statusCode,
+			DurationMs:     time.Since(start).Milliseconds(),
+			RequestHeaders: captureRequestHeaders(r),
+			RequestBody:    truncateBody(reqBody),
+			ResponseBody:   formatResponseBody(&recW.body, recW.totalBytes),
+		})
+	}
+}
+
+// getRecent handles GET /api/recent — returns the most recent API requests
+// as JSON, newest first.
+func (s *Server) getRecent(w http.ResponseWriter, r *http.Request) {
+	writeJSON(w, http.StatusOK, map[string]any{
+		"ok":       true,
+		"requests": s.recorder.Entries(),
+	})
+}

internal/server/recorder_test.go

@@ -0,0 +1,148 @@
+package server
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// TestRecentRecorderRingBuffer verifies that the recorder keeps at most
+// maxRecentEntries and evicts the oldest when full.
+func TestRecentRecorderRingBuffer(t *testing.T) {
+	rec := NewRecentRecorder()
+	for i := 0; i < maxRecentEntries+5; i++ {
+		rec.Record(RecentEntry{Method: "POST", Path: "/v1/test", Status: 200})
+	}
+	entries := rec.Entries()
+	if len(entries) != maxRecentEntries {
+		t.Fatalf("got %d entries, want %d", len(entries), maxRecentEntries)
+	}
+}
+
+// TestRecentRecorderNewestFirst verifies entries are returned newest-first.
+func TestRecentRecorderNewestFirst(t *testing.T) {
+	rec := NewRecentRecorder()
+	rec.Record(RecentEntry{Path: "/first"})
+	rec.Record(RecentEntry{Path: "/second"})
+	rec.Record(RecentEntry{Path: "/third"})
+	entries := rec.Entries()
+	if len(entries) != 3 {
+		t.Fatalf("got %d entries", len(entries))
+	}
+	if entries[0].Path != "/third" {
+		t.Fatalf("first entry = %q, want /third", entries[0].Path)
+	}
+	if entries[2].Path != "/first" {
+		t.Fatalf("last entry = %q, want /first", entries[2].Path)
+	}
+}
+
+// TestWithRecordingCapturesRequestResponse verifies the middleware captures
+// request headers, request body, response status, and response body.
+func TestWithRecordingCapturesRequestResponse(t *testing.T) {
+	rec := NewRecentRecorder()
+	s := &Server{recorder: rec}
+
+	handler := s.withRecording(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		w.Write([]byte(`{"result":"ok"}`))
+	})
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{"model":"GLM-4.7","messages":[]}`))
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer sk-secret-key-12345")
+	req.Header.Set("User-Agent", "test-client/1.0")
+	w := httptest.NewRecorder()
+	handler(w, req)
+
+	entries := rec.Entries()
+	if len(entries) != 1 {
+		t.Fatalf("got %d entries, want 1", len(entries))
+	}
+	e := entries[0]
+	if e.Method != "POST" {
+		t.Fatalf("method = %q", e.Method)
+	}
+	if e.Path != "/v1/chat/completions" {
+		t.Fatalf("path = %q", e.Path)
+	}
+	if e.Status != 200 {
+		t.Fatalf("status = %d", e.Status)
+	}
+	if e.RequestBody != `{"model":"GLM-4.7","messages":[]}` {
+		t.Fatalf("request body = %q", e.RequestBody)
+	}
+	if e.ResponseBody != `{"result":"ok"}` {
+		t.Fatalf("response body = %q", e.ResponseBody)
+	}
+	// Authorization must be masked
+	auth, ok := e.RequestHeaders["Authorization"]
+	if !ok || !strings.Contains(auth, "****") {
+		t.Fatalf("authorization not masked: %q", auth)
+	}
+	if strings.Contains(auth, "sk-secret-key-12345") {
+		t.Fatal("authorization leaked raw key")
+	}
+	if e.RequestHeaders["Content-Type"] != "application/json" {
+		t.Fatalf("content-type = %v", e.RequestHeaders["Content-Type"])
+	}
+	if e.RequestHeaders["User-Agent"] != "test-client/1.0" {
+		t.Fatalf("user-agent = %v", e.RequestHeaders["User-Agent"])
+	}
+}
+
+// TestWithRecordingCapturesErrors verifies error responses are recorded.
+func TestWithRecordingCapturesErrors(t *testing.T) {
+	rec := NewRecentRecorder()
+	s := &Server{recorder: rec}
+
+	handler := s.withRecording(func(w http.ResponseWriter, r *http.Request) {
+		writeOpenAIError(w, http.StatusUnauthorized, "invalid api key", "auth_error", "invalid_api_key")
+	})
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(`{"input":"hi"}`))
+	req.Header.Set("Authorization", "Bearer wrong-key")
+	w := httptest.NewRecorder()
+	handler(w, req)
+
+	entries := rec.Entries()
+	if len(entries) != 1 {
+		t.Fatalf("got %d entries, want 1", len(entries))
+	}
+	e := entries[0]
+	if e.Status != 401 {
+		t.Fatalf("status = %d, want 401", e.Status)
+	}
+	if !strings.Contains(e.ResponseBody, "invalid api key") {
+		t.Fatalf("response body = %q", e.ResponseBody)
+	}
+}
+
+// TestGetRecentAPI verifies GET /api/recent returns recorded entries as JSON.
+func TestGetRecentAPI(t *testing.T) {
+	rec := NewRecentRecorder()
+	rec.Record(RecentEntry{Method: "POST", Path: "/v1/chat/completions", Status: 200})
+	rec.Record(RecentEntry{Method: "POST", Path: "/v1/responses", Status: 500})
+	s := &Server{recorder: rec}
+
+	req := httptest.NewRequest(http.MethodGet, "/api/recent", nil)
+	w := httptest.NewRecorder()
+	s.getRecent(w, req)
+
+	body, _ := io.ReadAll(w.Body)
+	if !strings.Contains(string(body), `"ok":true`) {
+		t.Fatalf("response missing ok:true: %s", string(body))
+	}
+	if !strings.Contains(string(body), "/v1/responses") {
+		t.Fatalf("response missing /v1/responses: %s", string(body))
+	}
+	// newest first
+	idxResponses := strings.Index(string(body), "/v1/responses")
+	idxChat := strings.Index(string(body), "/v1/chat/completions")
+	if idxResponses < 0 || idxChat < 0 || idxResponses > idxChat {
+		t.Fatalf("entries not newest-first: %s", string(body))
+	}
+}

internal/server/responses.go

@@ -0,0 +1,563 @@
+package server
+
+import (
+	"bufio"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"time"
+
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/openai"
+)
+
+// responses handles POST /v1/responses — the OpenAI Responses API. It
+// translates the request to a chat-completions request for the upstream
+// gateway, then translates the response back into the Responses format.
+func (s *Server) responses(w http.ResponseWriter, r *http.Request) {
+	creds, err := s.currentCredentials()
+	if err != nil {
+		writeOpenAIError(w, http.StatusUnauthorized, "zhanlu credentials are not configured; open /login first", "auth_error", "missing_credentials")
+		return
+	}
+
+	body, err := io.ReadAll(r.Body)
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, err.Error(), "invalid_request_error", "bad_body")
+		return
+	}
+
+	chatReq, err := openai.ParseResponsesRequest(body)
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, err.Error(), "invalid_request_error", "bad_json")
+		return
+	}
+	meta := openai.ParseResponsesMeta(body)
+	if chatReq.Model == "" {
+		chatReq.Model = "zhanlu/auto"
+	}
+
+	start := time.Now()
+	clientWantsStream := chatReq.Stream
+	chatReq.Stream = true // upstream always streams
+	upstreamBody, err := chatReq.MarshalForUpstream()
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, err.Error(), "invalid_request_error", "bad_body")
+		return
+	}
+
+	resp, ok := s.callUpstream(w, r, creds, chatReq.Model, upstreamBody, clientWantsStream, start)
+	if !ok {
+		return
+	}
+	defer resp.Body.Close()
+
+	if clientWantsStream {
+		usage, status := s.proxyResponsesStream(w, resp, chatReq.Model, meta)
+		s.record(chatReq.Model, true, usage, status, start)
+		return
+	}
+	usage, status := s.aggregateResponsesStream(w, resp, chatReq.Model, meta)
+	s.record(chatReq.Model, false, usage, status, start)
+}
+
+// ---------------------------------------------------------------- non-streaming
+
+// aggregateResponsesStream reads the upstream SSE, aggregates it (same logic
+// as aggregateStream), then emits a Responses API JSON object.
+func (s *Server) aggregateResponsesStream(w http.ResponseWriter, resp *http.Response, model string, meta openai.ResponsesMeta) (any, string) {
+	var content, reasoning, id string
+	var usage any
+	finishReason := "stop"
+	type toolCall struct {
+		ID       string `json:"id"`
+		Type     string `json:"type"`
+		Function struct {
+			Name      string `json:"name"`
+			Arguments string `json:"arguments"`
+		} `json:"function"`
+	}
+	toolCalls := map[int]*toolCall{}
+	err := forEachSSEChunk(resp.Body, func(chunk []byte) error {
+		var event chatStreamChunk
+		if err := json.Unmarshal(chunk, &event); err != nil {
+			return err
+		}
+		if event.ID != "" {
+			id = event.ID
+		}
+		if event.Usage != nil {
+			usage = event.Usage
+		}
+		if len(event.Choices) > 0 {
+			content += event.Choices[0].Delta.Content
+			reasoning += event.Choices[0].Delta.ReasoningContent + event.Choices[0].Delta.Reasoning
+			for _, part := range event.Choices[0].Delta.ToolCalls {
+				call := toolCalls[part.Index]
+				if call == nil {
+					call = &toolCall{Type: "function"}
+					toolCalls[part.Index] = call
+				}
+				if part.ID != "" {
+					call.ID = part.ID
+				}
+				if part.Type != "" {
+					call.Type = part.Type
+				}
+				call.Function.Name += part.Function.Name
+				call.Function.Arguments += part.Function.Arguments
+			}
+			if event.Choices[0].FinishReason != nil {
+				finishReason = *event.Choices[0].FinishReason
+			}
+		}
+		return nil
+	})
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadGateway, err.Error(), "upstream_error", "zhanlu_stream_error")
+		return nil, "upstream_error"
+	}
+
+	_ = id
+	// Map the upstream finish_reason onto the Responses API status: a
+	// length/content_filter cutoff means the response is incomplete, not
+	// completed. Default ("stop"/"tool_calls") stays "completed".
+	status := "completed"
+	switch finishReason {
+	case "length", "content_filter":
+		status = "incomplete"
+	}
+	respID := "resp_" + randomRequestID()
+	output := []any{}
+
+	if reasoning != "" {
+		output = append(output, map[string]any{
+			"type": "reasoning", "id": "rs_" + randomRequestID(), "status": "completed",
+			"content": []any{},
+			"summary": []map[string]any{{"type": "summary_text", "text": reasoning}},
+		})
+	}
+
+	// Message output item — included when there is text content.
+	if content != "" {
+		output = append(output, map[string]any{
+			"type": "message", "id": "msg_" + randomRequestID(), "status": "completed",
+			"role": "assistant",
+			"content": []map[string]any{{
+				"type": "output_text", "text": content,
+				"annotations": []any{}, "logprobs": []any{},
+			}},
+		})
+	}
+
+	if len(toolCalls) > 0 {
+		indices := make([]int, 0, len(toolCalls))
+		for i := range toolCalls {
+			indices = append(indices, i)
+		}
+		sort.Ints(indices)
+		ordered := make([]*toolCall, 0, len(toolCalls))
+		for _, i := range indices {
+			ordered = append(ordered, toolCalls[i])
+		}
+		for _, tc := range ordered {
+			output = append(output, map[string]any{
+				"type": "function_call", "id": "fc_" + randomRequestID(),
+				"call_id": tc.ID, "name": tc.Function.Name,
+				"arguments": tc.Function.Arguments, "status": "completed",
+			})
+		}
+	}
+
+	// If there was no content and no tool calls, still emit an empty message
+	// so the response always has at least one output item.
+	if len(output) == 0 {
+		output = append(output, map[string]any{
+			"type": "message", "id": "msg_" + randomRequestID(), "status": "completed",
+			"role": "assistant", "content": []map[string]any{},
+		})
+	}
+
+	result := map[string]any{
+		"id": respID, "object": "response", "created_at": time.Now().Unix(),
+		"model": model, "status": status, "output": output,
+		"parallel_tool_calls": meta.ParallelToolCalls,
+		"tool_choice":         meta.ToolChoice,
+		"tools":               meta.Tools,
+	}
+	if usage != nil {
+		result["usage"] = openai.UsageToResponses(usage)
+	}
+	writeJSON(w, http.StatusOK, result)
+	return usage, "success"
+}
+
+// ------------------------------------------------------------------- streaming
+
+// responsesStreamState tracks the lifecycle of output items while converting
+// Chat Completions SSE into Responses API SSE events.
+type responsesStreamState struct {
+	w       http.ResponseWriter
+	flusher http.Flusher
+	respID  string
+	msgID   string
+	rsID    string
+	meta    openai.ResponsesMeta
+
+	outputIdx     int
+	reasoningOn   bool
+	reasonPartOn  bool
+	messageOn     bool
+	partOn        bool
+	fullContent   string
+	fullReasoning string
+	usage         any
+	status        string
+	finishReason  string
+
+	// tool call tracking
+	tools     map[int]*streamToolCall
+	toolOrder []int
+}
+
+type streamToolCall struct {
+	itemID  string
+	callID  string
+	name    string
+	args    string
+	started bool
+	doneIdx int // assigned output index when started
+}
+
+func newResponsesStreamState(w http.ResponseWriter, model string, meta openai.ResponsesMeta) *responsesStreamState {
+	flusher, _ := w.(http.Flusher)
+	w.Header().Set("Content-Type", "text/event-stream; charset=utf-8")
+	w.Header().Set("Cache-Control", "no-cache")
+	w.Header().Set("Connection", "keep-alive")
+	w.Header().Set("X-Accel-Buffering", "no")
+	w.WriteHeader(http.StatusOK)
+	st := &responsesStreamState{
+		w:       w,
+		flusher: flusher,
+		respID:  "resp_" + randomRequestID(),
+		msgID:   "msg_" + randomRequestID(),
+		rsID:    "rs_" + randomRequestID(),
+		meta:    meta,
+		tools:   map[int]*streamToolCall{},
+		status:  "success",
+	}
+	baseResp := func(st2 string) map[string]any {
+		return map[string]any{
+			"id": st.respID, "object": "response", "created_at": time.Now().Unix(),
+			"model": model, "status": st2, "output": []any{},
+			"parallel_tool_calls": meta.ParallelToolCalls,
+			"tool_choice":         meta.ToolChoice,
+			"tools":               meta.Tools,
+		}
+	}
+	st.emit("response.created", map[string]any{"type": "response.created", "response": baseResp("in_progress")})
+	st.emit("response.in_progress", map[string]any{"type": "response.in_progress", "response": baseResp("in_progress")})
+	return st
+}
+
+func (st *responsesStreamState) emit(event string, data any) {
+	b, _ := json.Marshal(data)
+	fmt.Fprintf(st.w, "event: %s\ndata: %s\n\n", event, b)
+	if st.flusher != nil {
+		st.flusher.Flush()
+	}
+}
+
+// closeReasoning closes the reasoning output item if it is open.
+func (st *responsesStreamState) closeReasoning() {
+	if !st.reasoningOn {
+		return
+	}
+	if st.reasonPartOn {
+		st.emit("response.reasoning_summary_text.done", map[string]any{
+			"type": "response.reasoning_summary_text.done", "item_id": st.rsID,
+			"output_index": st.outputIdx, "summary_index": 0, "text": st.fullReasoning,
+		})
+		st.emit("response.reasoning_summary_part.done", map[string]any{
+			"type": "response.reasoning_summary_part.done", "item_id": st.rsID,
+			"output_index": st.outputIdx, "summary_index": 0,
+			"part": map[string]any{"type": "summary_text", "text": st.fullReasoning},
+		})
+		st.reasonPartOn = false
+	}
+	st.emit("response.output_item.done", map[string]any{
+		"type": "response.output_item.done", "output_index": st.outputIdx,
+		"item": map[string]any{
+			"type": "reasoning", "id": st.rsID, "status": "completed",
+			"content": []any{},
+			"summary": []map[string]any{{"type": "summary_text", "text": st.fullReasoning}},
+		},
+	})
+	st.reasoningOn = false
+	st.outputIdx++
+}
+
+// closeMessage closes the message output item if it is open.
+func (st *responsesStreamState) closeMessage() {
+	if !st.messageOn {
+		return
+	}
+	if st.partOn {
+		st.emit("response.output_text.done", map[string]any{
+			"type": "response.output_text.done", "item_id": st.msgID,
+			"output_index": st.outputIdx, "content_index": 0, "text": st.fullContent,
+		})
+		st.emit("response.content_part.done", map[string]any{
+			"type": "response.content_part.done", "item_id": st.msgID,
+			"output_index": st.outputIdx, "content_index": 0,
+			"part": map[string]any{"type": "output_text", "text": st.fullContent, "annotations": []any{}, "logprobs": []any{}},
+		})
+		st.partOn = false
+	}
+	st.emit("response.output_item.done", map[string]any{
+		"type": "response.output_item.done", "output_index": st.outputIdx,
+		"item": map[string]any{
+			"type": "message", "id": st.msgID, "status": "completed", "role": "assistant",
+			"content": []map[string]any{{"type": "output_text", "text": st.fullContent, "annotations": []any{}, "logprobs": []any{}}},
+		},
+	})
+	st.messageOn = false
+	st.outputIdx++
+}
+
+// handleReasoning processes a reasoning content delta.
+func (st *responsesStreamState) handleReasoning(delta string) {
+	if !st.reasoningOn {
+		st.reasoningOn = true
+		st.emit("response.output_item.added", map[string]any{
+			"type": "response.output_item.added", "output_index": st.outputIdx,
+			"item": map[string]any{
+				"type": "reasoning", "id": st.rsID, "status": "in_progress",
+				"content": []any{}, "summary": []any{},
+			},
+		})
+		st.emit("response.reasoning_summary_part.added", map[string]any{
+			"type": "response.reasoning_summary_part.added", "item_id": st.rsID,
+			"output_index": st.outputIdx, "summary_index": 0,
+			"part": map[string]any{"type": "summary_text", "text": ""},
+		})
+		st.reasonPartOn = true
+	}
+	st.fullReasoning += delta
+	st.emit("response.reasoning_summary_text.delta", map[string]any{
+		"type": "response.reasoning_summary_text.delta", "item_id": st.rsID,
+		"output_index": st.outputIdx, "summary_index": 0, "delta": delta,
+	})
+}
+
+// handleContent processes a text content delta.
+func (st *responsesStreamState) handleContent(delta string) {
+	// Close reasoning if open — text content comes after reasoning.
+	st.closeReasoning()
+	if !st.messageOn {
+		st.messageOn = true
+		st.emit("response.output_item.added", map[string]any{
+			"type": "response.output_item.added", "output_index": st.outputIdx,
+			"item": map[string]any{
+				"type": "message", "id": st.msgID, "status": "in_progress", "role": "assistant", "content": []any{},
+			},
+		})
+		st.emit("response.content_part.added", map[string]any{
+			"type": "response.content_part.added", "item_id": st.msgID,
+			"output_index": st.outputIdx, "content_index": 0,
+			"part": map[string]any{"type": "output_text", "text": "", "annotations": []any{}, "logprobs": []any{}},
+		})
+		st.partOn = true
+	}
+	st.fullContent += delta
+	st.emit("response.output_text.delta", map[string]any{
+		"type": "response.output_text.delta", "item_id": st.msgID,
+		"output_index": st.outputIdx, "content_index": 0, "delta": delta,
+	})
+}
+
+// handleToolCall processes a tool call delta from the chat completion stream.
+func (st *responsesStreamState) handleToolCall(index int, id, name, args string) {
+	// Close reasoning/message if open — tool calls are separate output items.
+	st.closeReasoning()
+	st.closeMessage()
+
+	call := st.tools[index]
+	if call == nil {
+		call = &streamToolCall{itemID: "fc_" + randomRequestID()}
+		st.tools[index] = call
+		st.toolOrder = append(st.toolOrder, index)
+	}
+	if id != "" {
+		call.callID = id
+	}
+	if name != "" {
+		call.name += name
+	}
+	if !call.started {
+		call.started = true
+		call.doneIdx = st.outputIdx
+		st.emit("response.output_item.added", map[string]any{
+			"type": "response.output_item.added", "output_index": st.outputIdx,
+			"item": map[string]any{
+				"type": "function_call", "id": call.itemID, "call_id": call.callID,
+				"name": call.name, "arguments": "", "status": "in_progress",
+			},
+		})
+		st.outputIdx++
+	}
+	if args != "" {
+		call.args += args
+		st.emit("response.function_call_arguments.delta", map[string]any{
+			"type": "response.function_call_arguments.delta", "item_id": call.itemID,
+			"output_index": call.doneIdx, "delta": args,
+		})
+	}
+}
+
+// finish closes all open items and sends the response.completed event.
+func (st *responsesStreamState) finish(model string) {
+	st.closeReasoning()
+	st.closeMessage()
+	for _, idx := range st.toolOrder {
+		call := st.tools[idx]
+		st.emit("response.function_call_arguments.done", map[string]any{
+			"type": "response.function_call_arguments.done", "item_id": call.itemID,
+			"output_index": call.doneIdx, "arguments": call.args,
+		})
+		st.emit("response.output_item.done", map[string]any{
+			"type": "response.output_item.done", "output_index": call.doneIdx,
+			"item": map[string]any{
+				"type": "function_call", "id": call.itemID, "call_id": call.callID,
+				"name": call.name, "arguments": call.args, "status": "completed",
+			},
+		})
+	}
+
+	// Build the final output array for the completed event.
+	finalOutput := []any{}
+	if st.fullReasoning != "" {
+		finalOutput = append(finalOutput, map[string]any{
+			"type": "reasoning", "id": st.rsID, "status": "completed",
+			"content": []any{},
+			"summary": []map[string]any{{"type": "summary_text", "text": st.fullReasoning}},
+		})
+	}
+	if st.fullContent != "" {
+		finalOutput = append(finalOutput, map[string]any{
+			"type": "message", "id": st.msgID, "status": "completed", "role": "assistant",
+			"content": []map[string]any{{"type": "output_text", "text": st.fullContent, "annotations": []any{}, "logprobs": []any{}}},
+		})
+	}
+	for _, idx := range st.toolOrder {
+		call := st.tools[idx]
+		finalOutput = append(finalOutput, map[string]any{
+			"type": "function_call", "id": call.itemID, "call_id": call.callID,
+			"name": call.name, "arguments": call.args, "status": "completed",
+		})
+	}
+	// Ensure at least one output item exists.
+	if len(finalOutput) == 0 {
+		finalOutput = append(finalOutput, map[string]any{
+			"type": "message", "id": st.msgID, "status": "completed", "role": "assistant",
+			"content": []map[string]any{},
+		})
+	}
+
+	// Map the upstream finish_reason onto the Responses API status, mirroring
+	// the non-streaming path: a length/content_filter cutoff is "incomplete".
+	respStatus := "completed"
+	switch st.finishReason {
+	case "length", "content_filter":
+		respStatus = "incomplete"
+	}
+	completedResp := map[string]any{
+		"id": st.respID, "object": "response", "created_at": time.Now().Unix(),
+		"model": model, "status": respStatus, "output": finalOutput,
+		"parallel_tool_calls": st.meta.ParallelToolCalls,
+		"tool_choice":         st.meta.ToolChoice,
+		"tools":               st.meta.Tools,
+	}
+	if st.usage != nil {
+		completedResp["usage"] = openai.UsageToResponses(st.usage)
+	}
+	st.emit("response.completed", map[string]any{"type": "response.completed", "response": completedResp})
+}
+
+// proxyResponsesStream reads the upstream Chat Completions SSE and emits
+// Responses API SSE events via a responsesStreamState state machine.
+func (s *Server) proxyResponsesStream(w http.ResponseWriter, resp *http.Response, model string, meta openai.ResponsesMeta) (any, string) {
+	st := newResponsesStreamState(w, model, meta)
+
+	reader := bufio.NewReader(resp.Body)
+	for {
+		line, err := reader.ReadString('\n')
+		if line != "" {
+			payload := sseDataPayload(line)
+			if payload != "" && payload != "[DONE]" {
+				var evt struct {
+					State        string `json:"state"`
+					ErrorMessage string `json:"errorMessage"`
+					Usage        any    `json:"usage"`
+					Choices      []struct {
+						Delta struct {
+							Content          string `json:"content"`
+							ReasoningContent string `json:"reasoning_content"`
+							Reasoning        string `json:"reasoning"`
+							ToolCalls        []struct {
+								Index    int    `json:"index"`
+								ID       string `json:"id"`
+								Function struct {
+									Name      string `json:"name"`
+									Arguments string `json:"arguments"`
+								} `json:"function"`
+							} `json:"tool_calls"`
+						} `json:"delta"`
+						FinishReason *string `json:"finish_reason"`
+					} `json:"choices"`
+				}
+				if jErr := json.Unmarshal([]byte(payload), &evt); jErr == nil {
+					if evt.State == "ERROR" {
+						st.status = "upstream_error"
+					}
+					if evt.Usage != nil {
+						st.usage = evt.Usage
+					}
+					if len(evt.Choices) > 0 {
+						delta := evt.Choices[0].Delta
+						if evt.Choices[0].FinishReason != nil {
+							st.finishReason = *evt.Choices[0].FinishReason
+						}
+						reasoningDelta := delta.ReasoningContent + delta.Reasoning
+						if reasoningDelta != "" {
+							st.handleReasoning(reasoningDelta)
+						}
+						if delta.Content != "" {
+							st.handleContent(delta.Content)
+						}
+						for _, tc := range delta.ToolCalls {
+							st.handleToolCall(tc.Index, tc.ID, tc.Function.Name, tc.Function.Arguments)
+						}
+					}
+				}
+			}
+		}
+		if err != nil {
+			if err != io.EOF {
+				b, _ := json.Marshal(map[string]any{
+					"message": err.Error(), "type": "upstream_error", "code": "zhanlu_stream_error",
+				})
+				fmt.Fprintf(w, "data: %s\n\n", b)
+				if st.flusher != nil {
+					st.flusher.Flush()
+				}
+				st.status = "upstream_error"
+			}
+			break
+		}
+	}
+
+	st.finish(model)
+	return st.usage, st.status
+}

internal/server/responses_test.go

@@ -0,0 +1,304 @@
+package server
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/auth"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/config"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/store"
+)
+
+// TestResponsesNonStreaming verifies POST /v1/responses with stream:false
+// returns a properly formatted Responses API JSON object.
+func TestResponsesNonStreaming(t *testing.T) {
+	upstream, proxy, st := setupTestServer(t)
+	defer upstream.Close()
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		AccessKey: "AK", SecretKey: "SK", Token: "TOKEN",
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	if err := st.SaveCredentials(creds); err != nil {
+		t.Fatal(err)
+	}
+
+	body := `{"model":"GLM-4.7","input":"hi","stream":false}`
+	resp, err := http.Post(proxy.URL+"/v1/responses", "application/json", strings.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != 200 {
+		b, _ := io.ReadAll(resp.Body)
+		t.Fatalf("status %d: %s", resp.StatusCode, string(b))
+	}
+	var result map[string]any
+	if err := json.NewDecoder(resp.Body).Decode(&result); err != nil {
+		t.Fatal(err)
+	}
+	if result["object"] != "response" {
+		t.Fatalf("object = %v", result["object"])
+	}
+	if result["status"] != "completed" {
+		t.Fatalf("status = %v", result["status"])
+	}
+	output, ok := result["output"].([]any)
+	if !ok || len(output) == 0 {
+		t.Fatalf("output = %v", result["output"])
+	}
+	msg, ok := output[0].(map[string]any)
+	if !ok {
+		t.Fatalf("output[0] type = %T", output[0])
+	}
+	if msg["type"] != "message" {
+		t.Fatalf("output[0] type = %v", msg["type"])
+	}
+	if msg["role"] != "assistant" {
+		t.Fatalf("output[0] role = %v", msg["role"])
+	}
+	content, ok := msg["content"].([]any)
+	if !ok || len(content) == 0 {
+		t.Fatalf("content = %v", msg["content"])
+	}
+	part, ok := content[0].(map[string]any)
+	if !ok {
+		t.Fatalf("content[0] type = %T", content[0])
+	}
+	if part["type"] != "output_text" {
+		t.Fatalf("content[0] type = %v", part["type"])
+	}
+	if part["text"] != "hello" {
+		t.Fatalf("content[0] text = %v", part["text"])
+	}
+	// Verify required SDK fields are present.
+	if _, ok := result["parallel_tool_calls"]; !ok {
+		t.Fatal("missing parallel_tool_calls in response")
+	}
+	if _, ok := result["tool_choice"]; !ok {
+		t.Fatal("missing tool_choice in response")
+	}
+	if _, ok := result["tools"]; !ok {
+		t.Fatal("missing tools in response")
+	}
+	// Verify output_text has logprobs field.
+	if _, ok := part["logprobs"]; !ok {
+		t.Fatal("missing logprobs in output_text content part")
+	}
+	usage, ok := result["usage"].(map[string]any)
+	if !ok {
+		t.Fatalf("usage = %v", result["usage"])
+	}
+	if totNum(usage["input_tokens"]) != 10 {
+		t.Fatalf("input_tokens = %v", usage["input_tokens"])
+	}
+	if totNum(usage["output_tokens"]) != 20 {
+		t.Fatalf("output_tokens = %v", usage["output_tokens"])
+	}
+	if totNum(usage["total_tokens"]) != 30 {
+		t.Fatalf("total_tokens = %v", usage["total_tokens"])
+	}
+}
+
+// TestResponsesStreaming verifies POST /v1/responses with stream:true
+// emits proper Responses API SSE events.
+func TestResponsesStreaming(t *testing.T) {
+	upstream, proxy, st := setupTestServer(t)
+	defer upstream.Close()
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		AccessKey: "AK", SecretKey: "SK", Token: "TOKEN",
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	if err := st.SaveCredentials(creds); err != nil {
+		t.Fatal(err)
+	}
+
+	body := `{"model":"GLM-4.7","input":"hi","stream":true}`
+	resp, err := http.Post(proxy.URL+"/v1/responses", "application/json", strings.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != 200 {
+		b, _ := io.ReadAll(resp.Body)
+		t.Fatalf("status %d: %s", resp.StatusCode, string(b))
+	}
+	raw, _ := io.ReadAll(resp.Body)
+	s := string(raw)
+
+	// Check key lifecycle events are present.
+	checks := []string{
+		"event: response.created",
+		"event: response.in_progress",
+		"event: response.output_item.added",
+		"event: response.content_part.added",
+		"event: response.output_text.delta",
+		"event: response.output_text.done",
+		"event: response.content_part.done",
+		"event: response.output_item.done",
+		"event: response.completed",
+	}
+	for _, c := range checks {
+		if !strings.Contains(s, c) {
+			t.Fatalf("missing %q in SSE body:\n%s", c, s)
+		}
+	}
+
+	// Verify the text delta contains "hello".
+	if !strings.Contains(s, `"delta":"hello"`) {
+		t.Fatalf("text delta missing 'hello' in SSE body:\n%s", s)
+	}
+
+	// Verify required SDK fields are present in completed event.
+	if !strings.Contains(s, `"parallel_tool_calls"`) {
+		t.Fatalf("missing parallel_tool_calls in SSE body:\n%s", s)
+	}
+	if !strings.Contains(s, `"tool_choice"`) {
+		t.Fatalf("missing tool_choice in SSE body:\n%s", s)
+	}
+	// Verify logprobs field in output_text.
+	if !strings.Contains(s, `"logprobs"`) {
+		t.Fatalf("missing logprobs in SSE body:\n%s", s)
+	}
+
+	// Verify usage in the completed event has input_tokens.
+	if !strings.Contains(s, `"input_tokens":`) {
+		t.Fatalf("usage missing input_tokens in SSE body:\n%s", s)
+	}
+}
+
+// TestResponsesStringInput verifies the proxy correctly translates a string
+// "input" to a chat-completions "messages" array before forwarding upstream.
+func TestResponsesStringInput(t *testing.T) {
+	var capturedBody []byte
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/chat/completions" {
+			capturedBody, _ = io.ReadAll(r.Body)
+			w.Header().Set("Content-Type", "text/event-stream")
+			_, _ = w.Write([]byte("data: {\"id\":\"x\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"ok\"},\"finish_reason\":null}]}\n\ndata: {\"id\":\"x\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\ndata: [DONE]\n\n"))
+			return
+		}
+		http.NotFound(w, r)
+	}))
+	defer upstream.Close()
+
+	st, err := store.Open(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+
+	cfg := config.Config{
+		MobileLoginBaseURL: upstream.URL,
+		MobileModelBaseURL: upstream.URL,
+		UpstreamPath:       "/chat/completions",
+		DBPath:             filepath.Join(t.TempDir(), "zhanlu.db"),
+		PublicKeyPEM:       defaultTestPublicKey,
+		PhonePublicKeyPEM:  defaultTestPublicKey,
+		SM2PrivateKey:      testSM2Key,
+		PluginVersion:      "1.4.2",
+	}
+	h := New(cfg, st)
+	proxy := httptest.NewServer(h)
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	_ = st.SaveCredentials(creds)
+
+	body := `{"model":"GLM-4.7","input":"hello world","stream":false}`
+	resp, err := http.Post(proxy.URL+"/v1/responses", "application/json", strings.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	io.Copy(io.Discard, resp.Body)
+
+	if len(capturedBody) == 0 {
+		t.Fatal("no upstream request body captured")
+	}
+	var upstreamReq map[string]any
+	if err := json.Unmarshal(capturedBody, &upstreamReq); err != nil {
+		t.Fatal(err)
+	}
+	messages, ok := upstreamReq["messages"].([]any)
+	if !ok || len(messages) != 1 {
+		t.Fatalf("messages = %v", upstreamReq["messages"])
+	}
+	msg, _ := messages[0].(map[string]any)
+	if msg["role"] != "user" || msg["content"] != "hello world" {
+		t.Fatalf("upstream message = %v", msg)
+	}
+}
+
+// TestResponsesInstructions verifies instructions are prepended as a system
+// message in the upstream request.
+func TestResponsesInstructions(t *testing.T) {
+	var capturedBody []byte
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/chat/completions" {
+			capturedBody, _ = io.ReadAll(r.Body)
+			w.Header().Set("Content-Type", "text/event-stream")
+			_, _ = w.Write([]byte("data: {\"id\":\"x\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"ok\"},\"finish_reason\":null}]}\n\ndata: {\"id\":\"x\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\ndata: [DONE]\n\n"))
+			return
+		}
+		http.NotFound(w, r)
+	}))
+	defer upstream.Close()
+
+	st, err := store.Open(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer st.Close()
+
+	cfg := config.Config{
+		MobileLoginBaseURL: upstream.URL,
+		MobileModelBaseURL: upstream.URL,
+		UpstreamPath:       "/chat/completions",
+		DBPath:             filepath.Join(t.TempDir(), "zhanlu.db"),
+		PublicKeyPEM:       defaultTestPublicKey,
+		PhonePublicKeyPEM:  defaultTestPublicKey,
+		SM2PrivateKey:      testSM2Key,
+		PluginVersion:      "1.4.2",
+	}
+	h := New(cfg, st)
+	proxy := httptest.NewServer(h)
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	_ = st.SaveCredentials(creds)
+
+	body := `{"model":"GLM-4.7","instructions":"be concise","input":"hello","stream":false}`
+	resp, err := http.Post(proxy.URL+"/v1/responses", "application/json", strings.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	io.Copy(io.Discard, resp.Body)
+
+	if len(capturedBody) == 0 {
+		t.Fatal("no upstream request body captured")
+	}
+	var upstreamReq map[string]any
+	_ = json.Unmarshal(capturedBody, &upstreamReq)
+	messages, _ := upstreamReq["messages"].([]any)
+	if len(messages) != 2 {
+		t.Fatalf("messages = %d items", len(messages))
+	}
+	sys, _ := messages[0].(map[string]any)
+	if sys["role"] != "system" || sys["content"] != "be concise" {
+		t.Fatalf("system message = %v", sys)
+	}
+}

internal/server/server_test.go

@@ -15,12 +15,16 @@ import (
 
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/auth"
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/config"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/stats"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/store"
 )
 
 const testSM2Key = "8d6ee90b3c4d299ae5abd655dbc3547c110ae8aeff1de18b0df241f215f90748"
 
-// setupTestServer spins up a mock Zhanlu upstream and a proxy server wired to it.
-func setupTestServer(t *testing.T) (*httptest.Server, *httptest.Server, string) {
+// setupTestServer spins up a mock Zhanlu upstream and a proxy server wired to
+// it. The proxy is backed by a temp SQLite store so credentials persist in the
+// db during the test instead of a JSON file.
+func setupTestServer(t *testing.T) (*httptest.Server, *httptest.Server, *store.Store) {
 	t.Helper()
 	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		switch r.URL.Path {
@@ -47,7 +51,7 @@ func setupTestServer(t *testing.T) (*httptest.Server, *httptest.Server, string)
 				return
 			}
 			w.Header().Set("Content-Type", "text/event-stream")
-			_, _ = fmt.Fprint(w, "data: {\"id\":\"chatcmpl-x\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"hello\"},\"finish_reason\":null}],\"usage\":{\"prompt_tokens\":1}}\n\n")
+			_, _ = fmt.Fprint(w, "data: {\"id\":\"chatcmpl-x\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"hello\"},\"finish_reason\":null}],\"usage\":{\"prompt_tokens\":10,\"completion_tokens\":20,\"total_tokens\":30,\"prompt_tokens_details\":{\"cached_tokens\":4}}}\n\n")
 			_, _ = fmt.Fprint(w, "data: {\"id\":\"chatcmpl-x\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n")
 			_, _ = fmt.Fprint(w, "data: [DONE]\n\n")
 		case "/gateway/v1/model/info":
@@ -59,28 +63,32 @@ func setupTestServer(t *testing.T) (*httptest.Server, *httptest.Server, string)
 		}
 	}))
 
-	credsFile := filepath.Join(t.TempDir(), "credentials.json")
+	st, err := store.Open(filepath.Join(t.TempDir(), "stats.db"))
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	t.Cleanup(func() { st.Close() })
 	cfg := config.Config{
 		ListenAddr:         ":0",
 		MobileLoginBaseURL: upstream.URL,
 		MobileModelBaseURL: upstream.URL,
 		UpstreamPath:       "/chat/completions",
-		CredentialsPath:    credsFile,
+		DBPath:             filepath.Join(t.TempDir(), "zhanlu.db"),
 		TokenDecryptKey:    "3jw7woww2rvhla6k",
 		PublicKeyPEM:       defaultTestPublicKey,
 		PhonePublicKeyPEM:  defaultTestPublicKey,
 		SM2PrivateKey:      testSM2Key,
 		PluginVersion:      "1.4.2",
 	}
-	h := New(cfg)
+	h := New(cfg, st)
 	proxy := httptest.NewServer(h)
-	return upstream, proxy, credsFile
+	return upstream, proxy, st
 }
 
 // TestPhoneLoginAndChat exercises the full v1.4.2 flow: SMS login, profile
 // fetch, SM2 API-key provisioning, then OpenAI-compatible chat and models.
 func TestPhoneLoginAndChat(t *testing.T) {
-	upstream, proxy, credsFile := setupTestServer(t)
+	upstream, proxy, st := setupTestServer(t)
 	defer upstream.Close()
 	defer proxy.Close()
 
@@ -107,8 +115,8 @@ func TestPhoneLoginAndChat(t *testing.T) {
 		t.Fatalf("login failed: %v", loginResp)
 	}
 
-	// 3. credentials file should contain the provisioned api key
-	creds, err := auth.LoadCredentials(credsFile)
+	// 3. credentials store should contain the provisioned api key
+	creds, err := st.LoadCredentials()
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -154,11 +162,11 @@ func TestPhoneLoginAndChat(t *testing.T) {
 }
 
 func TestStreamingChat(t *testing.T) {
-	upstream, proxy, credsFile := setupTestServer(t)
+	upstream, proxy, st := setupTestServer(t)
 	defer upstream.Close()
 	defer proxy.Close()
 
-	// Seed credentials directly with the api key
+	// Seed credentials directly with the api key into the store
 	creds := auth.Credentials{
 		AccessKey:    "AK",
 		SecretKey:    "SK",
@@ -167,7 +175,7 @@ func TestStreamingChat(t *testing.T) {
 		ModelBaseURL: upstream.URL,
 		Email:        "[email protected]",
 	}
-	if err := auth.SaveCredentials(credsFile, creds); err != nil {
+	if err := st.SaveCredentials(creds); err != nil {
 		t.Fatal(err)
 	}
 
@@ -188,6 +196,223 @@ func TestStreamingChat(t *testing.T) {
 	}
 }
 
+// TestStats verifies both streaming and non-streaming chat paths record token
+// usage into the store and that GET /api/stats aggregates them correctly.
+func TestStats(t *testing.T) {
+	upstream, proxy, st := setupTestServer(t)
+	defer upstream.Close()
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		AccessKey: "AK", SecretKey: "SK", Token: "TOKEN",
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	if err := st.SaveCredentials(creds); err != nil {
+		t.Fatal(err)
+	}
+
+	// non-streaming chat
+	resp, err := http.Post(proxy.URL+"/v1/chat/completions", "application/json", strings.NewReader(`{"model":"GLM-4.7","messages":[{"role":"user","content":"hi"}],"stream":false}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	io.Copy(io.Discard, resp.Body)
+	resp.Body.Close()
+
+	// streaming chat
+	resp, err = http.Post(proxy.URL+"/v1/chat/completions", "application/json", strings.NewReader(`{"model":"GLM-4.7","messages":[{"role":"user","content":"hi"}],"stream":true}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	io.Copy(io.Discard, resp.Body)
+	resp.Body.Close()
+
+	// query stats
+	resp, err = http.Get(proxy.URL + "/api/stats")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	var statsResp map[string]any
+	if err := json.NewDecoder(resp.Body).Decode(&statsResp); err != nil {
+		t.Fatal(err)
+	}
+	if enabled, _ := statsResp["enabled"].(bool); !enabled {
+		t.Fatalf("stats not enabled: %v", statsResp)
+	}
+	s, _ := statsResp["stats"].(map[string]any)
+	if s == nil {
+		t.Fatalf("no stats object: %v", statsResp)
+	}
+	totals, _ := s["totals"].(map[string]any)
+	if totals == nil {
+		t.Fatalf("no totals: %v", s)
+	}
+	if got := totNum(totals["requests"]); got != 2 {
+		t.Fatalf("requests = %v, want 2", totals["requests"])
+	}
+	if got := totNum(totals["prompt_tokens"]); got != 20 {
+		t.Fatalf("prompt_tokens = %v, want 20", totals["prompt_tokens"])
+	}
+	if got := totNum(totals["completion_tokens"]); got != 40 {
+		t.Fatalf("completion_tokens = %v, want 40", totals["completion_tokens"])
+	}
+	if got := totNum(totals["total_tokens"]); got != 60 {
+		t.Fatalf("total_tokens = %v, want 60", totals["total_tokens"])
+	}
+	if got := totNum(totals["cached_tokens"]); got != 8 {
+		t.Fatalf("cached_tokens = %v, want 8", totals["cached_tokens"])
+	}
+	if rate, _ := totals["cache_rate"].(float64); rate < 0.39 || rate > 0.41 {
+		t.Fatalf("cache_rate = %v, want ~0.4", totals["cache_rate"])
+	}
+	perModel, _ := s["per_model"].([]any)
+	if len(perModel) != 1 {
+		t.Fatalf("per_model = %v, want 1 entry", perModel)
+	}
+}
+
+// TestAdminModels verifies the admin /api/models endpoint returns the live
+// model list advertised by the upstream gateway model-info endpoint.
+func TestAdminModels(t *testing.T) {
+	upstream, proxy, st := setupTestServer(t)
+	defer upstream.Close()
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		AccessKey: "AK", SecretKey: "SK", Token: "TOKEN",
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	if err := st.SaveCredentials(creds); err != nil {
+		t.Fatal(err)
+	}
+
+	resp, err := http.Get(proxy.URL + "/api/models")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	var data map[string]any
+	if err := json.NewDecoder(resp.Body).Decode(&data); err != nil {
+		t.Fatal(err)
+	}
+	if ok, _ := data["ok"].(bool); !ok {
+		t.Fatalf("models endpoint not ok: %v", data)
+	}
+	models, _ := data["models"].([]any)
+	if len(models) != 2 {
+		t.Fatalf("models = %v, want 2", models)
+	}
+}
+
+// TestModelTest verifies the admin /api/models/test endpoint probes a model and
+// reports availability plus latency against the mock streaming upstream.
+func TestModelTest(t *testing.T) {
+	upstream, proxy, st := setupTestServer(t)
+	defer upstream.Close()
+	defer proxy.Close()
+
+	creds := auth.Credentials{
+		AccessKey: "AK", SecretKey: "SK", Token: "TOKEN",
+		APIKey: "sk-test-456", ModelBaseURL: upstream.URL, Email: "[email protected]",
+	}
+	if err := st.SaveCredentials(creds); err != nil {
+		t.Fatal(err)
+	}
+
+	body, _ := json.Marshal(map[string]string{"model": "GLM-4.7"})
+	resp, err := http.Post(proxy.URL+"/api/models/test", "application/json", bytes.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	var data map[string]any
+	if err := json.NewDecoder(resp.Body).Decode(&data); err != nil {
+		t.Fatal(err)
+	}
+	if ok, _ := data["ok"].(bool); !ok {
+		t.Fatalf("test endpoint not ok: %v", data)
+	}
+	if available, _ := data["available"].(bool); !available {
+		t.Fatalf("model should be available: %v", data)
+	}
+	if _, ok := data["ttft_ms"]; !ok {
+		t.Fatalf("ttft_ms should be present: %v", data)
+	}
+}
+
+// TestModelTestNoCredentials verifies the test endpoint reports unavailable
+// gracefully when no credentials are configured, instead of erroring.
+func TestModelTestNoCredentials(t *testing.T) {
+	upstream, proxy, _ := setupTestServer(t)
+	defer upstream.Close()
+	defer proxy.Close()
+
+	body, _ := json.Marshal(map[string]string{"model": "GLM-4.7"})
+	resp, err := http.Post(proxy.URL+"/api/models/test", "application/json", bytes.NewReader(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer resp.Body.Close()
+	var data map[string]any
+	if err := json.NewDecoder(resp.Body).Decode(&data); err != nil {
+		t.Fatal(err)
+	}
+	if ok, _ := data["ok"].(bool); !ok {
+		t.Fatalf("test endpoint should stay ok: %v", data)
+	}
+	if available, _ := data["available"].(bool); available {
+		t.Fatalf("model should not be available without creds: %v", data)
+	}
+}
+
+// TestAdminRendersModelRangeFilter verifies the admin page renders without
+// panic for both empty and populated summaries, and that the by-model panel
+// carries the 1d/7d/all range filter controls and a client-renderable tbody.
+func TestAdminRendersModelRangeFilter(t *testing.T) {
+	populated := &stats.Summary{
+		PerModel: []stats.ModelStat{
+			{Model: "GLM-4.7", Requests: 3, PromptTokens: 10, CompletionTokens: 20, TotalTokens: 30, CachedTokens: 4},
+		},
+	}
+	cases := []struct {
+		name    string
+		summary *stats.Summary
+	}{
+		{"empty", &stats.Summary{}},
+		{"populated", populated},
+	}
+	for _, tc := range cases {
+		var buf bytes.Buffer
+		if err := adminTemplate.Execute(&buf, map[string]any{
+			"Enabled":         true,
+			"Stats":           tc.summary,
+			"DBPath":          "zhanlu.db",
+			"PasswordEnabled": false,
+		}); err != nil {
+			t.Fatalf("%s: render admin: %v", tc.name, err)
+		}
+		body := buf.String()
+		for _, want := range []string{
+			`id="model-range"`,
+			`data-range="1d"`,
+			`data-range="7d"`,
+			`data-range="all"`,
+			`id="model-tbody"`,
+		} {
+			if !strings.Contains(body, want) {
+				t.Fatalf("%s: admin output missing %q", tc.name, want)
+			}
+		}
+	}
+}
+
+// totNum extracts an int from a JSON-decoded numeric value (float64).
+func totNum(v any) int {
+	f, _ := v.(float64)
+	return int(f)
+}
+
 func okValue(m map[string]any) bool {
 	ok, _ := m["ok"].(bool)
 	return ok

internal/server/templates.go

@@ -0,0 +1,49 @@
+package server
+
+import (
+	"embed"
+	"fmt"
+	"html/template"
+)
+
+// templateFS holds the rendered admin/login HTML pages. Keeping them as
+// separate files (rather than inline raw-string literals in server.go) gives
+// them real syntax highlighting and keeps server.go focused on handlers.
+//
+//go:embed templates/*.html
+var templateFS embed.FS
+
+var (
+	loginTemplate       = mustParseTemplate("login.html", "login")
+	loginResultTemplate = mustParseTemplate("login_result.html", "login-result")
+	adminTemplate       = mustParseTemplateFuncs("admin.html", "admin", template.FuncMap{
+		"pct": func(f float64) string { return fmt.Sprintf("%.1f%%", f*100) },
+		"rate": func(cached, prompt int64) string {
+			if prompt <= 0 {
+				return "0%"
+			}
+			return fmt.Sprintf("%.1f%%", float64(cached)/float64(prompt)*100)
+		},
+		"human": humanNum,
+	})
+)
+
+// mustParseTemplate reads a single embedded template file and parses it,
+// panicking on error (a malformed template is a build-time mistake).
+func mustParseTemplate(filename, name string) *template.Template {
+	data, err := templateFS.ReadFile("templates/" + filename)
+	if err != nil {
+		panic("embed template " + filename + ": " + err.Error())
+	}
+	return template.Must(template.New(name).Parse(string(data)))
+}
+
+// mustParseTemplateFuncs is mustParseTemplate with a FuncMap registered before
+// parsing, so the template body may reference the custom functions.
+func mustParseTemplateFuncs(filename, name string, funcs template.FuncMap) *template.Template {
+	data, err := templateFS.ReadFile("templates/" + filename)
+	if err != nil {
+		panic("embed template " + filename + ": " + err.Error())
+	}
+	return template.Must(template.New(name).Funcs(funcs).Parse(string(data)))
+}

internal/server/templates/admin.html

@@ -0,0 +1,609 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <title>湛卢代理管理后台</title>
+  <style>
+    :root{
+      color-scheme:light;
+      font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",system-ui,"PingFang SC","Microsoft YaHei",sans-serif;
+      --bg:#f6f7f9; --surface:#ffffff; --border:#e6e8ec; --border-strong:#d4d8df;
+      --ink:#111827; --body:#4b5563; --muted:#9ca3af;
+      --accent:#2563eb; --accent-hover:#1d4ed8; --accent-soft:#eff4ff;
+      --ok:#059669; --err:#dc2626; --stream:#2563eb; --nonstream:#9ca3af;
+      --radius:16px; --radius-sm:10px;
+    }
+    *{box-sizing:border-box}
+    body{margin:0;min-height:100vh;padding:32px 20px 64px;color:var(--ink);background:var(--bg);-webkit-font-smoothing:antialiased}
+    @media(prefers-reduced-motion:no-preference){body{animation:fade .4s ease both}}
+    @keyframes fade{from{opacity:0;transform:translateY(4px)}to{opacity:1;transform:none}}
+    .wrap{max-width:1080px;margin:0 auto;display:grid;gap:20px}
+    header.top{display:flex;align-items:center;justify-content:space-between;gap:16px;flex-wrap:wrap}
+    .brand{display:flex;align-items:center;gap:10px}
+    .dot{width:10px;height:10px;border-radius:3px;background:var(--accent);flex:none}
+    .eyebrow{font-size:12px;font-weight:600;letter-spacing:.12em;text-transform:uppercase;color:var(--muted)}
+    h1{margin:6px 0 0;font-size:clamp(24px,3vw,30px);font-weight:700;letter-spacing:-.02em}
+    .top-actions{display:flex;gap:10px;flex-wrap:wrap}
+    .btn{display:inline-flex;align-items:center;justify-content:center;gap:8px;border:1px solid transparent;border-radius:var(--radius-sm);padding:9px 16px;font:inherit;font-weight:600;font-size:13.5px;text-decoration:none;cursor:pointer;transition:background .15s ease,border-color .15s ease,box-shadow .15s ease,transform .05s ease}
+    .btn:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
+    .btn-primary{background:var(--accent);color:#fff}
+    .btn-primary:hover{background:var(--accent-hover)}
+    .btn-primary:active{transform:translateY(1px)}
+    .btn-ghost{background:var(--surface);border-color:var(--border-strong);color:var(--ink)}
+    .btn-ghost:hover{border-color:var(--accent);color:var(--accent);background:var(--accent-soft)}
+    .btn-ghost:active{transform:translateY(1px)}
+    .tabs{display:flex;gap:2px;border-bottom:1px solid var(--border);margin-bottom:24px}
+    .tab{padding:10px 18px;border:0;background:none;font:inherit;font-weight:600;font-size:14px;color:var(--muted);cursor:pointer;border-bottom:2px solid transparent;margin-bottom:-1px;transition:color .15s ease,border-color .15s ease;border-radius:8px 8px 0 0}
+    .tab:hover{color:var(--ink)}
+    .tab[aria-selected="true"]{color:var(--accent);border-bottom-color:var(--accent)}
+    .tab:focus-visible{outline:2px solid var(--accent);outline-offset:-2px}
+    .tabpanel{display:none;min-width:0}
+    .tabpanel.active{display:block;min-width:0}
+    .sub-actions{display:flex;justify-content:flex-end;gap:10px;margin-bottom:16px}
+    .panel-head{display:flex;align-items:center;justify-content:space-between;gap:12px;margin-bottom:16px;flex-wrap:wrap}
+    .panel-head h2{margin:0}
+    .seg{display:inline-flex;border:1px solid var(--border-strong);border-radius:999px;padding:2px;background:var(--bg)}
+    .seg-btn{border:0;background:none;font:inherit;font-weight:600;font-size:12.5px;color:var(--body);padding:6px 14px;border-radius:999px;cursor:pointer;transition:background .15s ease,color .15s ease}
+    .seg-btn:hover{color:var(--ink)}
+    .seg-btn[aria-current="true"]{background:var(--accent);color:#fff}
+    .seg-btn:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
+    .panel{background:var(--surface);border:1px solid var(--border);border-radius:var(--radius);box-shadow:0 1px 2px rgba(16,24,40,.04);padding:24px;min-width:0}
+    .panel.disabled{background:#eef0f3}
+    h2{margin:0 0 16px;font-size:16px;font-weight:600;letter-spacing:-.01em}
+    .grid4{display:grid;grid-template-columns:repeat(auto-fit,minmax(168px,1fr));gap:12px}
+    .stat{border:1px solid var(--border);border-radius:12px;padding:16px;background:var(--bg)}
+    .stat .label{font-size:11.5px;letter-spacing:.04em;color:var(--muted);margin-bottom:8px}
+    .stat .val{font-size:26px;font-weight:700;letter-spacing:-.02em;font-variant-numeric:tabular-nums}
+    .stat .sub{font-size:12px;color:var(--body);margin-top:4px;font-variant-numeric:tabular-nums}
+    table{width:100%;border-collapse:collapse;font-size:13px}
+    th,td{text-align:left;padding:10px 12px;border-bottom:1px solid var(--border);white-space:nowrap}
+    tbody tr:last-child td{border-bottom:0}
+    th{color:var(--muted);font-weight:600;font-size:11.5px;letter-spacing:.04em;text-transform:uppercase}
+    td.num,th.num{text-align:right;font-variant-numeric:tabular-nums}
+    .badge{display:inline-block;padding:2px 9px;border-radius:999px;font-size:11.5px;font-weight:600;border:1px solid transparent}
+    .badge.ok{background:#e7f6ef;color:var(--ok);border-color:#c3e8d6}
+    .badge.err{background:#fdecec;color:var(--err);border-color:#f7d3d3}
+    .badge.stream{background:var(--accent-soft);color:var(--stream);border-color:#dbe6fb}
+    .badge.nonstream{background:#eef0f3;color:var(--nonstream);border-color:#dde1e6}
+    .barrow{display:grid;grid-template-columns:92px 1fr 72px;align-items:center;gap:12px;padding:5px 0}
+    .barrow .day{font-size:12.5px;color:var(--body);font-variant-numeric:tabular-nums}
+    .barrow .track{height:10px;border-radius:6px;background:#eef0f3;overflow:hidden}
+    .barrow .bar{height:100%;border-radius:6px;background:var(--accent);min-width:2px;width:0;transition:width .4s ease}
+    .barrow .amt{font-size:12.5px;color:var(--muted);text-align:right;font-variant-numeric:tabular-nums}
+    .muted{color:var(--muted);font-size:13px;line-height:1.6}
+    .scroll{overflow-x:auto;min-width:0}
+    .empty{color:var(--muted);font-size:13px;padding:8px 0}
+    .pager{display:flex;align-items:center;justify-content:center;gap:6px;margin-top:16px;flex-wrap:wrap}
+    .page-btn{min-width:32px;height:32px;padding:0 8px;border:1px solid var(--border-strong);border-radius:8px;background:var(--surface);color:var(--body);font:inherit;font-size:13px;font-weight:600;cursor:pointer;display:inline-flex;align-items:center;justify-content:center;transition:border-color .15s ease,color .15s ease,background .15s ease;font-variant-numeric:tabular-nums}
+    .page-btn:hover:not(:disabled):not(.dots){border-color:var(--accent);color:var(--accent);background:var(--accent-soft)}
+    .page-btn:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
+    .page-btn[aria-current="true"]{background:var(--accent);border-color:var(--accent);color:#fff}
+    .page-btn:disabled{opacity:.4;cursor:not-allowed}
+    .page-btn.dots{border:0;background:none;cursor:default;color:var(--muted);min-width:auto;padding:0 2px}
+    .page-info{font-size:12.5px;color:var(--muted);margin-left:8px;font-variant-numeric:tabular-nums}
+    .login-card{background:var(--surface);border:1px solid var(--border);border-radius:var(--radius);box-shadow:0 1px 2px rgba(16,24,40,.04);padding:32px;max-width:460px;margin:0 auto}
+    .lede{margin:0 0 24px;color:var(--body);font-size:14px;line-height:1.6}
+    form{display:grid;gap:18px}
+    label{display:grid;gap:7px;font-size:13px;font-weight:500;color:var(--body)}
+    input{width:100%;border:1px solid var(--border-strong);border-radius:var(--radius-sm);padding:11px 13px;background:var(--surface);color:var(--ink);outline:none;font:inherit;font-size:14px;transition:border-color .15s ease,box-shadow .15s ease}
+    input::placeholder{color:var(--muted)}
+    input:hover{border-color:#bdc2cc}
+    input:focus{border-color:var(--accent);box-shadow:0 0 0 3px var(--accent-soft)}
+    .row{display:grid;grid-template-columns:1fr auto;gap:10px;align-items:stretch}
+    .btn-primary.full{width:100%}
+    .status{min-height:20px;font-size:13px;line-height:1.6;color:var(--muted);display:flex;align-items:flex-start;gap:8px;margin-top:4px}
+    .status::before{content:"";flex:none;width:7px;height:7px;border-radius:50%;margin-top:6px;background:currentColor}
+    .status[data-state="ok"]{color:var(--ok)}
+    .status[data-state="err"]{color:var(--err)}
+    .status[data-state="busy"]{color:var(--accent)}
+    .status[data-state="busy"]::before{animation:pulse 1.1s ease-in-out infinite}
+    @media(prefers-reduced-motion:reduce){.status[data-state="busy"]::before{animation:none}}
+    .footer{margin-top:22px;padding-top:18px;border-top:1px solid var(--border);font-size:12px;color:var(--muted);line-height:1.7}
+    .footer code{font-family:"SF Mono",ui-monospace,Consolas,monospace;font-size:12px;color:var(--body);word-break:break-all}
+    .recent-card{border:1px solid var(--border);border-radius:12px;padding:16px;margin-bottom:12px;background:var(--bg)}
+    .recent-head{display:flex;align-items:center;gap:8px;flex-wrap:wrap;margin-bottom:4px}
+    .recent-time{font-size:12px;color:var(--muted);font-variant-numeric:tabular-nums}
+    .recent-method{font-size:11.5px;font-weight:700;color:var(--accent);background:var(--accent-soft);padding:2px 8px;border-radius:999px}
+    .recent-path{font-size:13px;color:var(--ink);font-family:"SF Mono",ui-monospace,Consolas,monospace;flex:1;min-width:0;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+    .recent-duration{font-size:12px;color:var(--muted);font-variant-numeric:tabular-nums}
+    .recent-card details{margin-top:6px}
+    .recent-card summary{cursor:pointer;font-size:12.5px;font-weight:600;color:var(--body);padding:4px 0;user-select:none}
+    .recent-card summary:hover{color:var(--accent)}
+    .recent-card pre{background:#1e1e2e;color:#cdd6f4;padding:12px;border-radius:8px;font-size:12px;overflow-x:auto;max-height:400px;overflow-y:auto;font-family:"SF Mono",ui-monospace,Consolas,monospace;line-height:1.5;margin:8px 0 0;white-space:pre-wrap;word-break:break-all}
+  </style>
+</head>
+<body>
+  <div class="wrap">
+    <header class="top">
+      <div>
+        <div class="brand"><span class="dot"></span><span class="eyebrow">Zhanlu Proxy · 管理后台</span></div>
+        <h1>湛卢代理管理</h1>
+      </div>
+      <div class="top-actions">{{if .PasswordEnabled}}<button class="btn btn-ghost" id="logout-button" type="button">退出登录</button>{{end}}</div>
+    </header>
+
+    <div class="tabs" role="tablist">
+      <button class="tab" role="tab" data-tab="stats" aria-selected="true">Token 统计</button>
+      <button class="tab" role="tab" data-tab="models" aria-selected="false">可用模型</button>
+      <button class="tab" role="tab" data-tab="recent" aria-selected="false">最近请求</button>
+      <button class="tab" role="tab" data-tab="login" aria-selected="false">凭据登录</button>
+    </div>
+
+    <section class="tabpanel active" data-tab="stats" role="tabpanel">
+      <div class="sub-actions">
+        <button class="btn btn-ghost" id="refresh">刷新</button>
+        <button class="btn btn-primary" id="reset">重置统计</button>
+      </div>
+      {{if not .Enabled}}<div class="panel disabled"><p class="muted">统计已关闭（ZHANLU_STATS_DISABLED=true）。</p></div>{{end}}
+      <div class="panel">
+        <h2>总览</h2>
+        <div class="grid4">
+          <div class="stat"><div class="label">请求总数</div><div class="val">{{.Stats.Totals.Requests}}</div><div class="sub">成功 {{.Stats.Totals.SuccessRequests}} · 失败 {{.Stats.Totals.ErrorRequests}}</div></div>
+          <div class="stat"><div class="label">Prompt Tokens</div><div class="val">{{human .Stats.Totals.PromptTokens}}</div><div class="sub">缓存 {{human .Stats.Totals.CachedTokens}}</div></div>
+          <div class="stat"><div class="label">Completion Tokens</div><div class="val">{{human .Stats.Totals.CompletionTokens}}</div><div class="sub">含思考 {{human .Stats.Totals.ReasoningTokens}}</div></div>
+          <div class="stat"><div class="label">Total Tokens</div><div class="val">{{human .Stats.Totals.TotalTokens}}</div></div>
+          <div class="stat"><div class="label">缓存命中率</div><div class="val">{{pct .Stats.Totals.CacheRate}}</div><div class="sub">缓存 {{human .Stats.Totals.CachedTokens}} / Prompt {{human .Stats.Totals.PromptTokens}}</div></div>
+        </div>
+      </div>
+      <div class="panel" style="margin-top:20px">
+        <div class="panel-head">
+          <h2>按模型</h2>
+          <div class="seg" id="model-range" role="group" aria-label="按模型统计时间范围">
+            <button class="seg-btn" type="button" data-range="1d">1 天</button>
+            <button class="seg-btn" type="button" data-range="7d">7 天</button>
+            <button class="seg-btn" type="button" data-range="all" aria-current="true">全部</button>
+          </div>
+        </div>
+        <div class="scroll" id="model-scroll"><table>
+          <thead><tr><th>模型</th><th class="num">请求数</th><th class="num">Prompt</th><th class="num">Comp</th><th class="num">Total</th><th class="num">缓存</th><th class="num">命中率</th></tr></thead>
+          <tbody id="model-tbody">
+          {{range .Stats.PerModel}}<tr><td>{{.Model}}</td><td class="num">{{.Requests}}</td><td class="num">{{human .PromptTokens}}</td><td class="num">{{human .CompletionTokens}}</td><td class="num">{{human .TotalTokens}}</td><td class="num">{{human .CachedTokens}}</td><td class="num">{{rate .CachedTokens .PromptTokens}}</td></tr>{{end}}
+          </tbody>
+        </table></div>
+        <p class="empty" id="model-empty" style="display:none">暂无数据</p>
+      </div>
+      <div class="panel" style="margin-top:20px">
+        <h2>按日</h2>
+        {{if .Stats.Daily}}
+        <div id="daily">
+          {{range .Stats.Daily}}<div class="barrow"><div class="day">{{.Day}}</div><div class="track"><div class="bar" data-token="{{.TotalTokens}}"></div></div><div class="amt">{{human .TotalTokens}}</div></div>{{end}}
+        </div>
+        {{else}}<p class="empty">暂无数据</p>{{end}}
+      </div>
+      <div class="panel" style="margin-top:20px">
+        <h2>最近请求</h2>
+        {{if .Stats.Recent}}
+        <div class="scroll"><table id="recent">
+          <thead><tr><th>时间</th><th>模型</th><th>模式</th><th class="num">Prompt</th><th class="num">Comp</th><th class="num">Total</th><th class="num">缓存</th><th>状态</th><th class="num">耗时</th></tr></thead>
+          <tbody>
+          {{range .Stats.Recent}}<tr><td>{{.Ts.Format "01-02 15:04:05"}}</td><td>{{.Model}}</td><td>{{if .Stream}}<span class="badge stream">流式</span>{{else}}<span class="badge nonstream">非流式</span>{{end}}</td><td class="num">{{human .PromptTokens}}</td><td class="num">{{human .CompletionTokens}}</td><td class="num">{{human .TotalTokens}}</td><td class="num">{{human .CachedTokens}}</td><td>{{if eq .Status "success"}}<span class="badge ok">成功</span>{{else}}<span class="badge err">失败</span>{{end}}</td><td class="num">{{.LatencyMs}}ms</td></tr>{{end}}
+          </tbody>
+        </table></div>
+        <div class="pager" id="recent-pager"></div>
+        {{else}}<p class="empty">暂无数据</p>{{end}}
+      </div>
+    </section>
+
+    <section class="tabpanel" data-tab="models" role="tabpanel">
+      <div class="sub-actions">
+        <button class="btn btn-ghost" id="models-refresh" type="button">刷新</button>
+        <button class="btn btn-primary" id="models-test-all" type="button">全部测试</button>
+      </div>
+      <div class="panel">
+        <h2>可用模型</h2>
+        <p class="muted" id="models-status" style="margin:0 0 16px">点击刷新获取当前上游接口返回的模型列表。</p>
+        <div class="scroll">
+          <table id="models-table">
+            <thead><tr><th>模型</th><th>状态</th><th class="num">首字延时</th><th class="num">总耗时</th><th>操作</th></tr></thead>
+            <tbody></tbody>
+          </table>
+        </div>
+      </div>
+    </section>
+
+    <section class="tabpanel" data-tab="recent" role="tabpanel">
+      <div class="sub-actions">
+        <button class="btn btn-ghost" id="recent-refresh" type="button">刷新</button>
+      </div>
+      <div class="panel">
+        <h2>最近 API 请求</h2>
+        <p class="muted" style="margin:0 0 16px">记录最近 10 条 /v1/ 请求的请求头和返回结果（含报错）。</p>
+        <div id="recent-list"></div>
+      </div>
+    </section>
+
+    <section class="tabpanel" data-tab="login" role="tabpanel">
+      <div class="login-card">
+        <p class="lede">输入手机号获取验证码，按插件默认的移动云登录接口换取凭据和模型 API Key。凭据保存到本地数据库，后续 OpenAI 兼容接口自动使用。</p>
+        <form id="phone-form">
+          <label>手机号<input name="telephone" inputmode="numeric" autocomplete="tel" placeholder="请输入 11 位手机号" required></label>
+          <label>验证码
+            <div class="row">
+              <input name="code" inputmode="numeric" autocomplete="one-time-code" placeholder="6 位验证码" required>
+              <button class="btn btn-ghost" id="code-button" type="button">获取验证码</button>
+            </div>
+          </label>
+          <button class="btn btn-primary full" type="submit">登录并保存凭据</button>
+        </form>
+        <div class="status" id="status" data-state="busy">正在检查登录状态...</div>
+      </div>
+    </section>
+  </div>
+  <script>
+    (function () {
+      var tabs = document.querySelectorAll('.tab');
+      var panels = document.querySelectorAll('.tabpanel');
+      function activate(name) {
+        tabs.forEach(function (t) { t.setAttribute('aria-selected', t.dataset.tab === name ? 'true' : 'false'); });
+        panels.forEach(function (p) { p.classList.toggle('active', p.dataset.tab === name); });
+      }
+      tabs.forEach(function (tab) {
+        tab.addEventListener('click', function () {
+          var name = tab.dataset.tab;
+          activate(name);
+          if (history.replaceState) history.replaceState(null, '', '#' + name);
+        });
+      });
+      var hash = location.hash.replace('#', '');
+      if (hash === 'models') activate('models');
+      else if (hash === 'recent') activate('recent');
+      else if (hash === 'login') activate('login');
+
+      var rows = document.querySelectorAll('#daily .bar');
+      var max = 1;
+      for (var i = 0; i < rows.length; i++) {
+        var t = parseInt(rows[i].getAttribute('data-token') || '0', 10);
+        if (t > max) max = t;
+      }
+      for (var j = 0; j < rows.length; j++) {
+        var v = parseInt(rows[j].getAttribute('data-token') || '0', 10);
+        rows[j].style.width = Math.max(2, Math.round(v * 100 / max)) + '%';
+      }
+      var refresh = document.getElementById('refresh');
+      if (refresh) refresh.addEventListener('click', function () { location.reload(); });
+      var reset = document.getElementById('reset');
+      if (reset) reset.addEventListener('click', function () {
+        if (!confirm('确定清空所有统计数据？')) return;
+        fetch('/api/stats/reset', { method: 'POST' }).then(function () { location.reload(); });
+      });
+
+      // by-model time-range filter (1d / 7d / all) — fetches /api/stats with a
+      // since cutoff and re-renders only the per-model table client-side.
+      (function () {
+        var seg = document.getElementById('model-range');
+        var tbody = document.getElementById('model-tbody');
+        var scroll = document.getElementById('model-scroll');
+        var empty = document.getElementById('model-empty');
+        if (!seg || !tbody) return;
+        function trimZero(s) { return s.indexOf('.') >= 0 ? s.replace(/\.0$/, '') : s; }
+        function human(v) {
+          var f = Number(v) || 0;
+          if (f < 1000) return String(Math.round(f));
+          if (f < 1e6) return trimZero((f / 1e3).toFixed(1)) + 'K';
+          if (f < 1e9) return trimZero((f / 1e6).toFixed(1)) + 'M';
+          return trimZero((f / 1e9).toFixed(1)) + 'B';
+        }
+        function rate(cached, prompt) {
+          if (prompt <= 0) return '0%';
+          return (cached / prompt * 100).toFixed(1) + '%';
+        }
+        function escapeHtml(s) {
+          return String(s).replace(/[&<>"']/g, function (c) {
+            return { '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' }[c];
+          });
+        }
+        function showEmpty(yes) {
+          scroll.style.display = yes ? 'none' : '';
+          empty.style.display = yes ? '' : 'none';
+        }
+        function render(rows) {
+          if (!rows || rows.length === 0) { tbody.innerHTML = ''; showEmpty(true); return; }
+          showEmpty(false);
+          var html = '';
+          for (var i = 0; i < rows.length; i++) {
+            var r = rows[i];
+            html += '<tr><td>' + escapeHtml(r.model) + '</td>' +
+              '<td class="num">' + (r.requests || 0) + '</td>' +
+              '<td class="num">' + human(r.prompt_tokens) + '</td>' +
+              '<td class="num">' + human(r.completion_tokens) + '</td>' +
+              '<td class="num">' + human(r.total_tokens) + '</td>' +
+              '<td class="num">' + human(r.cached_tokens) + '</td>' +
+              '<td class="num">' + rate(r.cached_tokens, r.prompt_tokens) + '</td></tr>';
+          }
+          tbody.innerHTML = html;
+        }
+        // initial empty check (server-rendered "all" may have no rows)
+        if (!tbody.querySelector('tr')) showEmpty(true);
+        function sinceParam(range) {
+          if (range === 'all') return '';
+          var days = range === '1d' ? 1 : 7;
+          return 'since=' + encodeURIComponent(new Date(Date.now() - days * 86400000).toISOString());
+        }
+        seg.addEventListener('click', async function (ev) {
+          var btn = ev.target.closest('.seg-btn');
+          if (!btn) return;
+          seg.querySelectorAll('.seg-btn').forEach(function (b) {
+            b.setAttribute('aria-current', b === btn ? 'true' : 'false');
+          });
+          var range = btn.dataset.range;
+          var param = sinceParam(range);
+          var url = '/api/stats' + (param ? '?' + param : '');
+          showEmpty(false);
+          tbody.innerHTML = '<tr><td colspan="7" class="muted" style="text-align:center">加载中…</td></tr>';
+          try {
+            var res = await fetch(url);
+            var data = await res.json();
+            if (!data.stats) { render([]); return; }
+            render(data.stats.per_model || []);
+          } catch (e) {
+            tbody.innerHTML = '<tr><td colspan="7" class="muted" style="text-align:center">加载失败</td></tr>';
+          }
+        });
+      })();
+
+      // paginate the recent-requests table (data is already rendered server-side)
+      (function () {
+        var tbody = document.querySelector('#recent tbody');
+        var pager = document.getElementById('recent-pager');
+        if (!tbody || !pager) return;
+        var rows = tbody.querySelectorAll('tr');
+        if (rows.length === 0) { pager.style.display = 'none'; return; }
+        var pageSize = 10;
+        var totalPages = Math.max(1, Math.ceil(rows.length / pageSize));
+        var page = 1;
+        function pageList(c, t) {
+          var p = [];
+          if (t <= 7) { for (var k = 1; k <= t; k++) p.push(k); return p; }
+          p.push(1);
+          if (c > 3) p.push('…');
+          var s = Math.max(2, c - 1), e = Math.min(t - 1, c + 1);
+          for (var m = s; m <= e; m++) p.push(m);
+          if (c < t - 2) p.push('…');
+          p.push(t);
+          return p;
+        }
+        function render() {
+          var start = (page - 1) * pageSize;
+          for (var i = 0; i < rows.length; i++) {
+            rows[i].style.display = (i >= start && i < start + pageSize) ? '' : 'none';
+          }
+          var html = '<button class="page-btn" data-act="prev"' + (page === 1 ? ' disabled' : '') + '>‹</button>';
+          var list = pageList(page, totalPages);
+          for (var n = 0; n < list.length; n++) {
+            var item = list[n];
+            if (item === '…') html += '<span class="page-btn dots">…</span>';
+            else html += '<button class="page-btn"' + (item === page ? ' aria-current="true"' : '') + ' data-page="' + item + '">' + item + '</button>';
+          }
+          html += '<button class="page-btn" data-act="next"' + (page === totalPages ? ' disabled' : '') + '>›</button>';
+          html += '<span class="page-info">第 ' + page + ' / ' + totalPages + ' 页 · 共 ' + rows.length + ' 条</span>';
+          pager.innerHTML = html;
+        }
+        pager.addEventListener('click', function (ev) {
+          var btn = ev.target.closest('.page-btn');
+          if (!btn || btn.disabled || btn.classList.contains('dots')) return;
+          if (btn.dataset.act === 'prev' && page > 1) page--;
+          else if (btn.dataset.act === 'next' && page < totalPages) page++;
+          else if (btn.dataset.page) page = parseInt(btn.dataset.page, 10);
+          render();
+        });
+        render();
+      })();
+
+      var logout = document.getElementById('logout-button');
+      if (logout) logout.addEventListener('click', async function () {
+        await fetch('/api/logout', { method: 'POST' });
+        window.location.href = '/login';
+      });
+    })();
+
+    (function () {
+      var statusEl = document.getElementById('status');
+      var form = document.getElementById('phone-form');
+      var codeButton = document.getElementById('code-button');
+      var secret = '';
+      var countdown = 0;
+      var countdownTimer = null;
+      function setStatus(text, state) { statusEl.textContent = text; statusEl.dataset.state = state || ''; }
+      function startCountdown() {
+        countdown = 60;
+        codeButton.disabled = true;
+        countdownTimer && clearInterval(countdownTimer);
+        countdownTimer = setInterval(function () {
+          if (countdown <= 0) {
+            clearInterval(countdownTimer);
+            codeButton.disabled = false;
+            codeButton.textContent = '获取验证码';
+            return;
+          }
+          codeButton.textContent = countdown + 's';
+          countdown--;
+        }, 1000);
+      }
+      if (form) {
+        fetch('/api/credentials').then(function (r) { return r.json(); }).then(function (data) {
+          statusEl.textContent = data.configured ? ('已登录：' + (data.access_key || '')) : '当前未登录';
+          statusEl.dataset.state = data.configured ? 'ok' : '';
+        });
+      }
+      if (codeButton) codeButton.addEventListener('click', async function () {
+        var telephone = form.telephone.value.trim();
+        if (!/^1[3-9]\d{9}$/.test(telephone)) { setStatus('请输入有效的 11 位手机号', 'err'); return; }
+        codeButton.disabled = true;
+        setStatus('正在发送验证码...', 'busy');
+        var res = await fetch('/api/auth/code', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ telephone: telephone }) });
+        var data = await res.json();
+        if (!res.ok || !data.ok) { codeButton.disabled = false; setStatus(data.error || '验证码发送失败', 'err'); return; }
+        secret = data.secret;
+        setStatus('验证码已发送', 'ok');
+        startCountdown();
+      });
+      if (form) form.addEventListener('submit', async function (event) {
+        event.preventDefault();
+        var telephone = form.telephone.value.trim();
+        var code = form.code.value.trim();
+        if (!secret) { setStatus('请先获取验证码', 'err'); return; }
+        setStatus('正在登录并保存凭据...', 'busy');
+        var res = await fetch('/api/auth/login', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ telephone: telephone, code: code, secret: secret }) });
+        var data = await res.json();
+        if (!res.ok || !data.ok) { setStatus(data.error || '登录失败', 'err'); return; }
+        setStatus('登录成功，已保存凭据：' + (data.access_key || '') + '；数据库：' + (data.path || ''), 'ok');
+      });
+    })();
+
+    (function () {
+      var refreshBtn = document.getElementById('models-refresh');
+      var testAllBtn = document.getElementById('models-test-all');
+      var statusEl = document.getElementById('models-status');
+      var tbody = document.querySelector('#models-table tbody');
+      if (!refreshBtn || !tbody) return;
+      var loaded = false;
+
+      function setStatus(text, state) { statusEl.textContent = text; statusEl.dataset.state = state || ''; }
+      function escapeHtml(s) {
+        return String(s).replace(/[&<>"']/g, function (c) {
+          return {'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c];
+        });
+      }
+      function renderRow(model) {
+        var tr = document.createElement('tr');
+        tr.dataset.model = model;
+        tr.innerHTML = '<td>' + escapeHtml(model) + '</td>' +
+          '<td class="m-status"><span class="muted">未测试</span></td>' +
+          '<td class="num m-ttft">—</td>' +
+          '<td class="num m-total">—</td>' +
+          '<td><button class="btn btn-ghost m-test" type="button">测试</button></td>';
+        return tr;
+      }
+      async function loadModels() {
+        setStatus('正在获取模型列表...', 'busy');
+        refreshBtn.disabled = true;
+        try {
+          var res = await fetch('/api/models');
+          var data = await res.json();
+          if (!res.ok || !data.ok) { setStatus(data.error || '获取模型列表失败', 'err'); return; }
+          var models = data.models || [];
+          if (models.length === 0) { setStatus('上游未返回任何模型', 'err'); tbody.innerHTML = ''; return; }
+          tbody.innerHTML = '';
+          for (var i = 0; i < models.length; i++) tbody.appendChild(renderRow(models[i]));
+          setStatus('共 ' + models.length + ' 个模型，点击测试检查可用性与延时', '');
+        } catch (e) {
+          setStatus('获取模型列表失败：' + e.message, 'err');
+        } finally {
+          refreshBtn.disabled = false;
+        }
+      }
+      async function testModel(model, row) {
+        var statusCell = row.querySelector('.m-status');
+        var ttftCell = row.querySelector('.m-ttft');
+        var totalCell = row.querySelector('.m-total');
+        var btn = row.querySelector('.m-test');
+        statusCell.innerHTML = '<span class="badge" style="background:var(--accent-soft);color:var(--accent);border-color:#dbe6fb">测试中...</span>';
+        ttftCell.textContent = '—';
+        totalCell.textContent = '—';
+        btn.disabled = true;
+        try {
+          var res = await fetch('/api/models/test', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ model: model }) });
+          var data = await res.json();
+          if (data.available) {
+            statusCell.innerHTML = '<span class="badge ok">可用</span>';
+            ttftCell.textContent = data.ttft_ms != null ? data.ttft_ms + 'ms' : '—';
+            totalCell.textContent = data.total_ms != null ? data.total_ms + 'ms' : '—';
+          } else {
+            statusCell.innerHTML = '<span class="badge err">不可用</span>';
+            statusCell.title = data.error || '';
+            totalCell.textContent = data.total_ms != null ? data.total_ms + 'ms' : '—';
+          }
+        } catch (e) {
+          statusCell.innerHTML = '<span class="badge err">请求错误</span>';
+          statusCell.title = e.message;
+        } finally {
+          btn.disabled = false;
+        }
+      }
+      refreshBtn.addEventListener('click', loadModels);
+      if (testAllBtn) testAllBtn.addEventListener('click', async function () {
+        var rows = tbody.querySelectorAll('tr');
+        for (var i = 0; i < rows.length; i++) {
+          await testModel(rows[i].dataset.model, rows[i]);
+        }
+      });
+      tbody.addEventListener('click', function (ev) {
+        var btn = ev.target.closest('.m-test');
+        if (!btn) return;
+        var row = btn.closest('tr');
+        if (row && row.dataset.model) testModel(row.dataset.model, row);
+      });
+      // lazy-load the model list the first time the tab becomes active
+      function loadIfActive() {
+        if (loaded) return;
+        var panel = document.querySelector('.tabpanel[data-tab="models"]');
+        if (panel && panel.classList.contains('active')) { loaded = true; loadModels(); }
+      }
+      var modelsTab = document.querySelector('.tab[data-tab="models"]');
+      if (modelsTab) modelsTab.addEventListener('click', loadIfActive);
+      loadIfActive();
+    })();
+
+    (function () {
+      var recentList = document.getElementById('recent-list');
+      if (!recentList) return;
+      var loaded = false;
+
+      function escapeHtml(s) {
+        return String(s).replace(/[&<>"']/g, function (c) {
+          return { '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' }[c];
+        });
+      }
+      function formatTime(ts) {
+        var d = new Date(ts);
+        return d.toLocaleString('zh-CN', { hour12: false });
+      }
+      async function loadRecent() {
+        recentList.innerHTML = '<p class="muted">加载中…</p>';
+        try {
+          var res = await fetch('/api/recent');
+          var data = await res.json();
+          if (!data.ok || !data.requests || data.requests.length === 0) {
+            recentList.innerHTML = '<p class="empty">暂无数据</p>';
+            return;
+          }
+          var html = '';
+          for (var i = 0; i < data.requests.length; i++) {
+            var r = data.requests[i];
+            var statusClass = r.status >= 200 && r.status < 300 ? 'ok' : 'err';
+            var headers = r.request_headers || {};
+            var headerStr = Object.keys(headers).map(function (k) {
+              return k + ': ' + headers[k];
+            }).join('\n');
+            html += '<div class="recent-card">' +
+              '<div class="recent-head">' +
+                '<span class="recent-time">' + escapeHtml(formatTime(r.timestamp)) + '</span>' +
+                '<span class="recent-method">' + escapeHtml(r.method) + '</span>' +
+                '<span class="recent-path">' + escapeHtml(r.path) + '</span>' +
+                '<span class="badge ' + statusClass + '">' + r.status + '</span>' +
+                '<span class="recent-duration">' + r.duration_ms + 'ms</span>' +
+              '</div>' +
+              (headerStr ? '<details><summary>请求头</summary><pre>' + escapeHtml(headerStr) + '</pre></details>' : '') +
+              '<details><summary>请求体</summary><pre>' + escapeHtml(r.request_body || '(空)') + '</pre></details>' +
+              '<details><summary>响应体</summary><pre>' + escapeHtml(r.response_body || '(空)') + '</pre></details>' +
+            '</div>';
+          }
+          recentList.innerHTML = html;
+        } catch (e) {
+          recentList.innerHTML = '<p class="empty">加载失败: ' + escapeHtml(e.message) + '</p>';
+        }
+      }
+      var recentTab = document.querySelector('.tab[data-tab="recent"]');
+      if (recentTab) recentTab.addEventListener('click', function () {
+        if (!loaded) { loaded = true; loadRecent(); }
+      });
+      var refreshBtn = document.getElementById('recent-refresh');
+      if (refreshBtn) refreshBtn.addEventListener('click', loadRecent);
+      // auto-load if the tab is active on page load
+      var panel = document.querySelector('.tabpanel[data-tab="recent"]');
+      if (panel && panel.classList.contains('active')) { loaded = true; loadRecent(); }
+    })();
+  </script>
+</body>
+</html>

internal/server/templates/login.html

@@ -0,0 +1,81 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <title>湛卢代理登录</title>
+  <style>
+    :root{
+      color-scheme:light;
+      font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",system-ui,"PingFang SC","Microsoft YaHei",sans-serif;
+      --bg:#f6f7f9; --surface:#ffffff; --border:#e6e8ec; --border-strong:#d4d8df;
+      --ink:#111827; --body:#4b5563; --muted:#9ca3af;
+      --accent:#2563eb; --accent-hover:#1d4ed8; --accent-soft:#eff4ff;
+      --ok:#059669; --err:#dc2626;
+      --radius:16px; --radius-sm:10px;
+    }
+    *{box-sizing:border-box}
+    body{margin:0;min-height:100vh;display:grid;place-items:center;padding:32px 20px;color:var(--ink);background:var(--bg);-webkit-font-smoothing:antialiased}
+    @media(prefers-reduced-motion:no-preference){body{animation:fade .4s ease both}}
+    @keyframes fade{from{opacity:0;transform:translateY(4px)}to{opacity:1;transform:none}}
+    .card{width:min(440px,100%);background:var(--surface);border:1px solid var(--border);border-radius:var(--radius);box-shadow:0 1px 2px rgba(16,24,40,.04),0 12px 32px -12px rgba(16,24,40,.12);padding:40px 36px}
+    .brand{display:flex;align-items:center;gap:10px;margin-bottom:8px}
+    .dot{width:10px;height:10px;border-radius:3px;background:var(--accent);flex:none}
+    .eyebrow{font-size:12px;font-weight:600;letter-spacing:.12em;text-transform:uppercase;color:var(--muted)}
+    h1{margin:0 0 10px;font-size:24px;font-weight:700;letter-spacing:-.02em}
+    .lede{margin:0 0 28px;color:var(--body);font-size:14px;line-height:1.6}
+    .lede code{font-family:"SF Mono",ui-monospace,Consolas,monospace;font-size:12.5px;color:var(--ink);background:var(--bg);padding:1px 6px;border-radius:5px;border:1px solid var(--border)}
+    form{display:grid;gap:18px}
+    label{display:grid;gap:7px;font-size:13px;font-weight:500;color:var(--body)}
+    input{width:100%;border:1px solid var(--border-strong);border-radius:var(--radius-sm);padding:11px 13px;background:var(--surface);color:var(--ink);outline:none;font:inherit;font-size:14px;transition:border-color .15s ease,box-shadow .15s ease}
+    input::placeholder{color:var(--muted)}
+    input:hover{border-color:#bdc2cc}
+    input:focus{border-color:var(--accent);box-shadow:0 0 0 3px var(--accent-soft)}
+    .btn{display:inline-flex;align-items:center;justify-content:center;gap:8px;border:1px solid transparent;border-radius:var(--radius-sm);padding:11px 18px;font:inherit;font-weight:600;font-size:14px;cursor:pointer;text-decoration:none;color:#fff;transition:background .15s ease,border-color .15s ease,box-shadow .15s ease,transform .05s ease}
+    .btn:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
+    .btn-primary{background:var(--accent);width:100%}
+    .btn-primary:hover{background:var(--accent-hover)}
+    .btn-primary:active{transform:translateY(1px)}
+    .btn-primary:disabled{background:#c2c9d4;cursor:not-allowed}
+    .status{min-height:20px;font-size:13px;line-height:1.6;color:var(--muted);display:flex;align-items:flex-start;gap:8px;margin-top:4px}
+    .status::before{content:"";flex:none;width:7px;height:7px;border-radius:50%;margin-top:6px;background:currentColor}
+    .status[data-state="ok"]{color:var(--ok)}
+    .status[data-state="err"]{color:var(--err)}
+    .status[data-state="busy"]{color:var(--accent)}
+    .status[data-state="busy"]::before{animation:pulse 1.1s ease-in-out infinite}
+    @keyframes pulse{0%,100%{opacity:1}50%{opacity:.35}}
+    @media(prefers-reduced-motion:reduce){.status[data-state="busy"]::before{animation:none}}
+    .footer{margin-top:24px;padding-top:18px;border-top:1px solid var(--border);font-size:12px;color:var(--muted);line-height:1.7}
+    .footer code{font-family:"SF Mono",ui-monospace,Consolas,monospace;font-size:12px;color:var(--body);word-break:break-all}
+  </style>
+</head>
+<body>
+  <main class="card">
+    <div class="brand"><span class="dot"></span><span class="eyebrow">Zhanlu Proxy</span></div>
+    <h1>湛卢代理登录</h1>
+    <p class="lede">请输入服务环境变量 <code>ZHANLU_LOGIN_PASSWORD</code> 配置的管理密码，验证后进入管理后台。</p>
+    <form id="password-form">
+      <label>登录密码<input name="password" type="password" autocomplete="current-password" placeholder="请输入服务访问密码" required></label>
+      <button class="btn btn-primary" type="submit">进入管理后台</button>
+    </form>
+    <div class="status" id="status" data-state="busy">需要登录后才能管理湛卢凭据。</div>
+  </main>
+  <script>
+    const statusEl = document.getElementById('status');
+    const passwordForm = document.getElementById('password-form');
+    function setStatus(text, state) { statusEl.textContent = text; statusEl.dataset.state = state || ''; }
+    if (passwordForm) {
+      passwordForm.addEventListener('submit', async (event) => {
+        event.preventDefault();
+        const password = passwordForm.password.value;
+        if (!password) { setStatus('请输入登录密码', 'err'); return; }
+        setStatus('正在登录...', 'busy');
+        const res = await fetch('/api/login', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ password }) });
+        const data = await res.json();
+        if (!res.ok || !data.ok) { setStatus(data.error || '登录失败', 'err'); return; }
+        window.location.href = '/admin';
+      });
+    }
+  </script>
+</body>
+</html>

internal/server/templates/login_result.html

@@ -0,0 +1,41 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <title>湛卢登录结果</title>
+  <style>
+    :root{
+      color-scheme:light;
+      font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",system-ui,"PingFang SC","Microsoft YaHei",sans-serif;
+      --bg:#f6f7f9; --surface:#ffffff; --border:#e6e8ec;
+      --ink:#111827; --body:#4b5563; --muted:#9ca3af;
+      --accent:#2563eb; --accent-hover:#1d4ed8; --ok:#059669; --err:#dc2626;
+    }
+    *{box-sizing:border-box}
+    body{margin:0;min-height:100vh;display:grid;place-items:center;padding:32px 20px;color:var(--ink);background:var(--bg);-webkit-font-smoothing:antialiased}
+    @media(prefers-reduced-motion:no-preference){body{animation:fade .4s ease both}}
+    @keyframes fade{from{opacity:0;transform:translateY(4px)}to{opacity:1;transform:none}}
+    .card{width:min(440px,100%);text-align:center;background:var(--surface);border:1px solid var(--border);border-radius:16px;box-shadow:0 1px 2px rgba(16,24,40,.04),0 12px 32px -12px rgba(16,24,40,.12);padding:44px 36px}
+    .mark{width:56px;height:56px;margin:0 auto 20px;border-radius:50%;display:grid;place-items:center}
+    .mark svg{width:26px;height:26px}
+    .mark.ok{background:#e7f6ef;border:1px solid #c3e8d6}
+    .mark.err{background:#fdecec;border:1px solid #f7d3d3}
+    h1{margin:0;font-size:22px;font-weight:700;letter-spacing:-.02em}
+    p{margin:14px 0 28px;color:var(--body);line-height:1.7;font-size:14px;word-break:break-word}
+    .btn{display:inline-flex;align-items:center;justify-content:center;border:1px solid transparent;border-radius:10px;padding:11px 24px;font:inherit;font-weight:600;font-size:14px;text-decoration:none;color:#fff;background:var(--accent);cursor:pointer;transition:background .15s ease}
+    .btn:hover{background:var(--accent-hover)}
+    .btn:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
+  </style>
+</head>
+<body>
+  <main class="card">
+    <div class="mark {{if .Success}}ok{{else}}err{{end}}">
+      {{if .Success}}<svg viewBox="0 0 24 24" fill="none" stroke="#059669" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"><path d="M20 6 9 17l-5-5"/></svg>{{else}}<svg viewBox="0 0 24 24" fill="none" stroke="#dc2626" stroke-width="2.4" stroke-linecap="round"><path d="M18 6 6 18M6 6l12 12"/></svg>{{end}}
+    </div>
+    {{if .Success}}<h1>登录成功</h1>{{else}}<h1>登录失败</h1>{{end}}
+    <p>{{.Message}}</p>
+    <a class="btn" href="/admin">返回管理后台</a>
+  </main>
+</body>
+</html>

internal/stats/stats.go

@@ -0,0 +1,83 @@
+// Package stats defines the types and helpers for OpenAI-compatible token
+// usage statistics recorded by the zhanlu proxy. The concrete SQLite-backed
+// recorder lives in internal/store; this package is dependency-free so it can
+// be referenced by both store and server without import cycles.
+package stats
+
+import "time"
+
+// Record is a single chat-completion usage observation.
+type Record struct {
+	Ts               time.Time `json:"ts"`
+	Model            string    `json:"model"`
+	Stream           bool      `json:"stream"`
+	PromptTokens     int       `json:"prompt_tokens"`
+	CompletionTokens int       `json:"completion_tokens"`
+	TotalTokens      int       `json:"total_tokens"`
+	ReasoningTokens  int       `json:"reasoning_tokens"`
+	CachedTokens     int       `json:"cached_tokens"`
+	Status           string    `json:"status"` // "success" | "upstream_error"
+	LatencyMs        int64     `json:"latency_ms"`
+}
+
+// Query filters the recorded stats. Zero-value time fields mean unbounded on
+// that end; empty Model means all models. Limit caps the recent-records list
+// (0 = default). Stream filters by streaming mode (nil = both).
+type Query struct {
+	Since  time.Time
+	Until  time.Time
+	Model  string
+	Limit  int
+	Stream *bool
+}
+
+// Totals aggregates request counts and token sums over a filtered set.
+type Totals struct {
+	Requests         int     `json:"requests"`
+	SuccessRequests  int     `json:"success_requests"`
+	ErrorRequests    int     `json:"error_requests"`
+	PromptTokens     int64   `json:"prompt_tokens"`
+	CompletionTokens int64   `json:"completion_tokens"`
+	TotalTokens      int64   `json:"total_tokens"`
+	ReasoningTokens  int64   `json:"reasoning_tokens"`
+	CachedTokens     int64   `json:"cached_tokens"`
+	CacheRate        float64 `json:"cache_rate"` // cached_tokens / prompt_tokens, 0..1
+}
+
+// ModelStat is a per-model aggregation row.
+type ModelStat struct {
+	Model            string `json:"model"`
+	Requests         int    `json:"requests"`
+	PromptTokens     int64  `json:"prompt_tokens"`
+	CompletionTokens int64  `json:"completion_tokens"`
+	TotalTokens      int64  `json:"total_tokens"`
+	CachedTokens     int64  `json:"cached_tokens"`
+}
+
+// DayStat is a per-day aggregation row (server-local time, YYYY-MM-DD).
+type DayStat struct {
+	Day              string `json:"day"`
+	Requests         int    `json:"requests"`
+	PromptTokens     int64  `json:"prompt_tokens"`
+	CompletionTokens int64  `json:"completion_tokens"`
+	TotalTokens      int64  `json:"total_tokens"`
+	CachedTokens     int64  `json:"cached_tokens"`
+}
+
+// Summary is the full result returned by a Recorder's Stats query.
+type Summary struct {
+	Totals   Totals      `json:"totals"`
+	PerModel []ModelStat `json:"per_model"`
+	Daily    []DayStat   `json:"daily"`
+	Recent   []Record    `json:"recent"`
+}
+
+// Recorder persists and queries usage statistics. The concrete implementation
+// lives in internal/store; the interface is declared here so server code can
+// depend on the contract and tests can inject fakes.
+type Recorder interface {
+	Record(r Record) error
+	Stats(q Query) (*Summary, error)
+	Reset() error
+	Close() error
+}

internal/stats/usage.go

@@ -0,0 +1,78 @@
+package stats
+
+import (
+	"encoding/json"
+	"time"
+)
+
+// Usage is the subset of the OpenAI chat-completion usage object the recorder
+// persists. Numbers arrive from JSON unmarshal as float64.
+type Usage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
+	ReasoningTokens  int `json:"reasoning_tokens"`
+	CachedTokens     int `json:"cached_tokens"`
+
+	// PromptTokensDetails.CachedTokens is emitted by providers that support
+	// prompt caching (OpenAI/DeepSeek/Zhipu litellm gateways). Some upstreams
+	// put cached_tokens at the top level instead.
+	PromptTokensDetails struct {
+		CachedTokens int `json:"cached_tokens"`
+	} `json:"prompt_tokens_details"`
+
+	// CompletionTokensDetails.ReasoningTokens is emitted by reasoning models;
+	// some upstreams put reasoning_tokens at top level instead.
+	CompletionTokensDetails struct {
+		ReasoningTokens int `json:"reasoning_tokens"`
+	} `json:"completion_tokens_details"`
+}
+
+// ExtractUsage decodes a raw usage value (as produced by encoding/json into an
+// any) into token counts. It accepts both full usage maps and raw JSON bytes.
+// Missing fields default to 0; a nil v yields zero usage.
+func ExtractUsage(v any) Usage {
+	var u Usage
+	if v == nil {
+		return u
+	}
+	switch t := v.(type) {
+	case []byte:
+		_ = json.Unmarshal(t, &u)
+	case json.RawMessage:
+		_ = json.Unmarshal(t, &u)
+	case map[string]any:
+		// Re-marshal + unmarshal is the simplest robust path for nested
+		// *_tokens_details; usage payloads are tiny.
+		if b, err := json.Marshal(t); err == nil {
+			_ = json.Unmarshal(b, &u)
+		}
+	}
+	if u.ReasoningTokens == 0 {
+		u.ReasoningTokens = u.CompletionTokensDetails.ReasoningTokens
+	}
+	if u.CachedTokens == 0 {
+		u.CachedTokens = u.PromptTokensDetails.CachedTokens
+	}
+	return u
+}
+
+// RecordFromUsage builds a Record from a captured usage value plus context.
+func RecordFromUsage(model string, stream bool, usage any, status string, start time.Time) Record {
+	u := ExtractUsage(usage)
+	if u.TotalTokens == 0 && (u.PromptTokens != 0 || u.CompletionTokens != 0) {
+		u.TotalTokens = u.PromptTokens + u.CompletionTokens
+	}
+	return Record{
+		Ts:               time.Now(),
+		Model:            model,
+		Stream:           stream,
+		PromptTokens:     u.PromptTokens,
+		CompletionTokens: u.CompletionTokens,
+		TotalTokens:      u.TotalTokens,
+		ReasoningTokens:  u.ReasoningTokens,
+		CachedTokens:     u.CachedTokens,
+		Status:           status,
+		LatencyMs:        time.Since(start).Milliseconds(),
+	}
+}

internal/store/credentials.go

@@ -0,0 +1,45 @@
+package store
+
+import (
+	"fmt"
+	"time"
+
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/auth"
+)
+
+// LoadCredentials reads the persisted credentials. The credentials row always
+// exists after Open; an empty row (nothing saved yet) yields a zero-value
+// Credentials with a nil error — callers check Validate()/HasAPIKey().
+func (s *Store) LoadCredentials() (auth.Credentials, error) {
+	var c auth.Credentials
+	var savedAt int64
+	err := s.db.QueryRow(`SELECT access_key, secret_key, token, api_key, model_base_url, email, organization, team, base_url, saved_at FROM credentials WHERE id = 1`).
+		Scan(&c.AccessKey, &c.SecretKey, &c.Token, &c.APIKey, &c.ModelBaseURL, &c.Email, &c.Organization, &c.Team, &c.BaseURL, &savedAt)
+	if err != nil {
+		return auth.Credentials{}, fmt.Errorf("load credentials: %w", err)
+	}
+	c.SavedAt = time.Unix(savedAt, 0).Local()
+	return c, nil
+}
+
+// SaveCredentials upserts the credentials into the single row. It only writes
+// when the credentials validate or already carry an API key, so partial /
+// env-only creds are not persisted.
+func (s *Store) SaveCredentials(c auth.Credentials) error {
+	if c.Validate() != nil && !c.HasAPIKey() {
+		return fmt.Errorf("save credentials: %w", c.Validate())
+	}
+	c.SavedAt = time.Now()
+	_, err := s.db.Exec(`INSERT INTO credentials (id, access_key, secret_key, token, api_key, model_base_url, email, organization, team, base_url, saved_at)
+VALUES (1, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+ON CONFLICT(id) DO UPDATE SET
+  access_key=excluded.access_key, secret_key=excluded.secret_key, token=excluded.token,
+  api_key=excluded.api_key, model_base_url=excluded.model_base_url, email=excluded.email,
+  organization=excluded.organization, team=excluded.team, base_url=excluded.base_url,
+  saved_at=excluded.saved_at`,
+		c.AccessKey, c.SecretKey, c.Token, c.APIKey, c.ModelBaseURL, c.Email, c.Organization, c.Team, c.BaseURL, c.SavedAt.Unix())
+	if err != nil {
+		return fmt.Errorf("save credentials: %w", err)
+	}
+	return nil
+}

internal/store/stats_recorder.go

@@ -0,0 +1,187 @@
+package store
+
+import (
+	"fmt"
+	"time"
+
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/stats"
+)
+
+// Compile-time guard: *Store satisfies stats.Recorder.
+var _ stats.Recorder = (*Store)(nil)
+
+const defaultRecentLimit = 200
+
+// Record appends a single usage observation.
+func (s *Store) Record(rec stats.Record) error {
+	if rec.Status == "" {
+		rec.Status = "success"
+	}
+	stream := 0
+	if rec.Stream {
+		stream = 1
+	}
+	_, err := s.db.Exec(`INSERT INTO requests (ts, model, stream, prompt_tokens, completion_tokens, total_tokens, reasoning_tokens, cached_tokens, status, latency_ms) VALUES (?,?,?,?,?,?,?,?,?,?)`,
+		rec.Ts.Unix(), rec.Model, stream, rec.PromptTokens, rec.CompletionTokens, rec.TotalTokens, rec.ReasoningTokens, rec.CachedTokens, rec.Status, rec.LatencyMs)
+	return err
+}
+
+// Stats computes the aggregate summary for the given query.
+func (s *Store) Stats(q stats.Query) (*stats.Summary, error) {
+	q = normalizeQuery(q)
+	where, args := whereClause(q)
+
+	sum := &stats.Summary{}
+	if err := s.scanTotals(sum, where, args); err != nil {
+		return nil, err
+	}
+	if sum.Totals.PromptTokens > 0 {
+		sum.Totals.CacheRate = float64(sum.Totals.CachedTokens) / float64(sum.Totals.PromptTokens)
+	}
+	if err := s.scanPerModel(sum, where, args); err != nil {
+		return nil, err
+	}
+	if err := s.scanDaily(sum, where, args); err != nil {
+		return nil, err
+	}
+	if err := s.scanRecent(sum, q, where, args); err != nil {
+		return nil, err
+	}
+	return sum, nil
+}
+
+// Reset deletes all recorded usage statistics (the credentials row is kept).
+func (s *Store) Reset() error {
+	_, err := s.db.Exec(`DELETE FROM requests`)
+	return err
+}
+
+func normalizeQuery(q stats.Query) stats.Query {
+	if q.Limit <= 0 {
+		q.Limit = defaultRecentLimit
+	}
+	if q.Limit > 5000 {
+		q.Limit = 5000
+	}
+	return q
+}
+
+func whereClause(q stats.Query) (string, []any) {
+	var conds []string
+	var args []any
+	if !q.Since.IsZero() {
+		conds = append(conds, "ts >= ?")
+		args = append(args, q.Since.Unix())
+	}
+	if !q.Until.IsZero() {
+		conds = append(conds, "ts <= ?")
+		args = append(args, q.Until.Unix())
+	}
+	if q.Model != "" {
+		conds = append(conds, "model = ?")
+		args = append(args, q.Model)
+	}
+	if q.Stream != nil {
+		conds = append(conds, "stream = ?")
+		args = append(args, boolToInt(*q.Stream))
+	}
+	if len(conds) == 0 {
+		return "", args
+	}
+	out := ""
+	for i, p := range conds {
+		if i > 0 {
+			out += " AND "
+		}
+		out += p
+	}
+	return " WHERE " + out, args
+}
+
+func boolToInt(b bool) int {
+	if b {
+		return 1
+	}
+	return 0
+}
+
+func (s *Store) scanTotals(sum *stats.Summary, where string, args []any) error {
+	q := `SELECT COUNT(*),
+  COALESCE(SUM(CASE WHEN status='success' THEN 1 ELSE 0 END),0),
+  COALESCE(SUM(CASE WHEN status!='success' THEN 1 ELSE 0 END),0),
+  COALESCE(SUM(prompt_tokens),0),
+  COALESCE(SUM(completion_tokens),0),
+  COALESCE(SUM(total_tokens),0),
+  COALESCE(SUM(reasoning_tokens),0),
+  COALESCE(SUM(cached_tokens),0) FROM requests` + where
+	row := s.db.QueryRow(q, args...)
+	var success, failures int64
+	err := row.Scan(&sum.Totals.Requests, &success, &failures,
+		&sum.Totals.PromptTokens, &sum.Totals.CompletionTokens,
+		&sum.Totals.TotalTokens, &sum.Totals.ReasoningTokens, &sum.Totals.CachedTokens)
+	if err != nil {
+		return fmt.Errorf("scan totals: %w", err)
+	}
+	sum.Totals.SuccessRequests = int(success)
+	sum.Totals.ErrorRequests = int(failures)
+	return nil
+}
+
+func (s *Store) scanPerModel(sum *stats.Summary, where string, args []any) error {
+	q := `SELECT model, COUNT(*), COALESCE(SUM(prompt_tokens),0), COALESCE(SUM(completion_tokens),0), COALESCE(SUM(total_tokens),0), COALESCE(SUM(cached_tokens),0) FROM requests` + where + " GROUP BY model ORDER BY SUM(total_tokens) DESC"
+	rows, err := s.db.Query(q, args...)
+	if err != nil {
+		return fmt.Errorf("scan per-model: %w", err)
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var m stats.ModelStat
+		if err := rows.Scan(&m.Model, &m.Requests, &m.PromptTokens, &m.CompletionTokens, &m.TotalTokens, &m.CachedTokens); err != nil {
+			return err
+		}
+		sum.PerModel = append(sum.PerModel, m)
+	}
+	return rows.Err()
+}
+
+func (s *Store) scanDaily(sum *stats.Summary, where string, args []any) error {
+	q := `SELECT date(ts,'unixepoch','localtime') AS day, COUNT(*), COALESCE(SUM(prompt_tokens),0), COALESCE(SUM(completion_tokens),0), COALESCE(SUM(total_tokens),0), COALESCE(SUM(cached_tokens),0) FROM requests` + where + " GROUP BY day ORDER BY day ASC"
+	rows, err := s.db.Query(q, args...)
+	if err != nil {
+		return fmt.Errorf("scan daily: %w", err)
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var d stats.DayStat
+		if err := rows.Scan(&d.Day, &d.Requests, &d.PromptTokens, &d.CompletionTokens, &d.TotalTokens, &d.CachedTokens); err != nil {
+			return err
+		}
+		sum.Daily = append(sum.Daily, d)
+	}
+	return rows.Err()
+}
+
+func (s *Store) scanRecent(sum *stats.Summary, q stats.Query, where string, args []any) error {
+	limit := q.Limit
+	if limit <= 0 {
+		limit = defaultRecentLimit
+	}
+	query := `SELECT ts, model, stream, prompt_tokens, completion_tokens, total_tokens, reasoning_tokens, cached_tokens, status, latency_ms FROM requests` + where + " ORDER BY id DESC LIMIT ?"
+	rows, err := s.db.Query(query, append(args, limit)...)
+	if err != nil {
+		return fmt.Errorf("scan recent: %w", err)
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var rec stats.Record
+		var ts int64
+		var stream int
+		if err := rows.Scan(&ts, &rec.Model, &stream, &rec.PromptTokens, &rec.CompletionTokens, &rec.TotalTokens, &rec.ReasoningTokens, &rec.CachedTokens, &rec.Status, &rec.LatencyMs); err != nil {
+			return err
+		}
+		rec.Ts = time.Unix(ts, 0).Local()
+		rec.Stream = stream == 1
+		sum.Recent = append(sum.Recent, rec)
+	}
+	return rows.Err()
+}

internal/store/store.go

@@ -0,0 +1,89 @@
+// Package store owns the single embedded SQLite database backing the zhanlu
+// proxy: token-usage records (the requests table) and the persisted login
+// credentials (the credentials table, single-tenant single row). It implements
+// stats.Recorder for usage tracking and exposes Load/Save for credentials.
+package store
+
+import (
+	"database/sql"
+	"fmt"
+	"strings"
+
+	_ "modernc.org/sqlite"
+)
+
+// Store is the single owner of the proxy's SQLite database handle.
+type Store struct {
+	db *sql.DB
+}
+
+// Open opens (or creates) the database at path and ensures both tables exist.
+// SQLite is opened with WAL journaling and a busy timeout so concurrent reads
+// (stats queries) and writes (request records, credential saves) do not
+// collide.
+func Open(path string) (*Store, error) {
+	db, err := sql.Open("sqlite", path+"?_pragma=busy_timeout(5000)&_pragma=journal_mode(WAL)&_pragma=synchronous(NORMAL)&_pragma=foreign_keys(ON)")
+	if err != nil {
+		return nil, fmt.Errorf("open db %q: %w", path, err)
+	}
+	if err := ensureSchema(db); err != nil {
+		db.Close()
+		return nil, err
+	}
+	return &Store{db: db}, nil
+}
+
+// Close releases the database handle.
+func (s *Store) Close() error {
+	return s.db.Close()
+}
+
+func ensureSchema(db *sql.DB) error {
+	stmts := []string{
+		`CREATE TABLE IF NOT EXISTS requests (
+  id INTEGER PRIMARY KEY AUTOINCREMENT,
+  ts INTEGER NOT NULL,
+  model TEXT NOT NULL,
+  stream INTEGER NOT NULL DEFAULT 0,
+  prompt_tokens INTEGER NOT NULL DEFAULT 0,
+  completion_tokens INTEGER NOT NULL DEFAULT 0,
+  total_tokens INTEGER NOT NULL DEFAULT 0,
+  reasoning_tokens INTEGER NOT NULL DEFAULT 0,
+  cached_tokens INTEGER NOT NULL DEFAULT 0,
+  status TEXT NOT NULL DEFAULT 'success',
+  latency_ms INTEGER NOT NULL DEFAULT 0
+)`,
+		`CREATE TABLE IF NOT EXISTS credentials (
+  id INTEGER PRIMARY KEY CHECK (id = 1),
+  access_key TEXT NOT NULL DEFAULT '',
+  secret_key TEXT NOT NULL DEFAULT '',
+  token TEXT NOT NULL DEFAULT '',
+  api_key TEXT NOT NULL DEFAULT '',
+  model_base_url TEXT NOT NULL DEFAULT '',
+  email TEXT NOT NULL DEFAULT '',
+  organization TEXT NOT NULL DEFAULT '',
+  team TEXT NOT NULL DEFAULT '',
+  base_url TEXT NOT NULL DEFAULT '',
+  saved_at INTEGER NOT NULL DEFAULT 0
+)`,
+		`CREATE INDEX IF NOT EXISTS idx_requests_ts ON requests(ts)`,
+		`CREATE INDEX IF NOT EXISTS idx_requests_model ON requests(model)`,
+		// Ensure the single credentials row exists so UPSERTs and SELECTs always
+		// have a target.
+		`INSERT INTO credentials (id) VALUES (1) ON CONFLICT(id) DO NOTHING`,
+	}
+	for _, q := range stmts {
+		if _, err := db.Exec(q); err != nil {
+			return fmt.Errorf("schema: %w", err)
+		}
+	}
+	// Add cached_tokens to databases created before this column existed. SQLite
+	// returns "duplicate column name" when it already exists; that is expected
+	// and ignored.
+	if _, err := db.Exec(`ALTER TABLE requests ADD COLUMN cached_tokens INTEGER NOT NULL DEFAULT 0`); err != nil {
+		if !strings.Contains(err.Error(), "duplicate column") {
+			return fmt.Errorf("migrate cached_tokens: %w", err)
+		}
+	}
+	return nil
+}

internal/store/store_test.go

@@ -0,0 +1,120 @@
+package store
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/auth"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/stats"
+)
+
+func newTestStore(t *testing.T) *Store {
+	t.Helper()
+	st, err := Open(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatalf("open store: %v", err)
+	}
+	t.Cleanup(func() { st.Close() })
+	return st
+}
+
+func TestCredentialsRoundTrip(t *testing.T) {
+	st := newTestStore(t)
+	in := auth.Credentials{
+		AccessKey: "AK", SecretKey: "SK", Token: "TOK",
+		APIKey: "sk-1", ModelBaseURL: "https://up.example", Email: "[email protected]",
+		Organization: "org", Team: "team",
+	}
+	if err := st.SaveCredentials(in); err != nil {
+		t.Fatalf("save: %v", err)
+	}
+	got, err := st.LoadCredentials()
+	if err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	if got.APIKey != "sk-1" || got.Email != "[email protected]" || got.Organization != "org" {
+		t.Fatalf("round-trip mismatch: %+v", got)
+	}
+	if got.SavedAt.IsZero() {
+		t.Fatalf("saved_at not set")
+	}
+}
+
+func TestEmptyLoadReturnsZero(t *testing.T) {
+	st := newTestStore(t)
+	got, err := st.LoadCredentials()
+	if err != nil {
+		t.Fatalf("load on empty store: %v", err)
+	}
+	if got.HasAPIKey() {
+		t.Fatalf("expected no api key on empty store, got %+v", got)
+	}
+}
+
+func TestRecordStatsReset(t *testing.T) {
+	st := newTestStore(t)
+	now := time.Now()
+	recs := []stats.Record{
+		{Ts: now, Model: "glm-4.7", Stream: false, PromptTokens: 10, CompletionTokens: 20, TotalTokens: 30, CachedTokens: 4, Status: "success", LatencyMs: 5},
+		{Ts: now, Model: "glm-4.7", Stream: true, PromptTokens: 5, CompletionTokens: 5, TotalTokens: 10, Status: "success", LatencyMs: 8},
+		{Ts: now, Model: "minimax", Stream: false, PromptTokens: 1, CompletionTokens: 1, TotalTokens: 2, Status: "upstream_error", LatencyMs: 3},
+	}
+	for _, r := range recs {
+		if err := st.Record(r); err != nil {
+			t.Fatalf("record: %v", err)
+		}
+	}
+
+	sum, err := st.Stats(stats.Query{})
+	if err != nil {
+		t.Fatalf("stats: %v", err)
+	}
+	if sum.Totals.Requests != 3 {
+		t.Fatalf("requests = %d, want 3", sum.Totals.Requests)
+	}
+	if sum.Totals.SuccessRequests != 2 || sum.Totals.ErrorRequests != 1 {
+		t.Fatalf("success/error = %d/%d, want 2/1", sum.Totals.SuccessRequests, sum.Totals.ErrorRequests)
+	}
+	if sum.Totals.TotalTokens != 42 {
+		t.Fatalf("total tokens = %d, want 42", sum.Totals.TotalTokens)
+	}
+	if sum.Totals.CachedTokens != 4 {
+		t.Fatalf("cached tokens = %d, want 4", sum.Totals.CachedTokens)
+	}
+	// prompt total = 10+5+1 = 16, cached = 4 => 0.25
+	if sum.Totals.CacheRate < 0.24 || sum.Totals.CacheRate > 0.26 {
+		t.Fatalf("cache rate = %v, want ~0.25", sum.Totals.CacheRate)
+	}
+	if len(sum.PerModel) != 2 {
+		t.Fatalf("per-model entries = %d, want 2", len(sum.PerModel))
+	}
+	// glm-4.7 should lead on total tokens (40 vs 2)
+	if sum.PerModel[0].Model != "glm-4.7" || sum.PerModel[0].TotalTokens != 40 || sum.PerModel[0].CachedTokens != 4 {
+		t.Fatalf("top model = %+v, want glm-4.7/40/4 cached", sum.PerModel[0])
+	}
+	if len(sum.Recent) != 3 {
+		t.Fatalf("recent entries = %d, want 3", len(sum.Recent))
+	}
+	// most recent first (id desc) => minimax record
+	if sum.Recent[0].Model != "minimax" {
+		t.Fatalf("most recent = %+v, want minimax", sum.Recent[0])
+	}
+
+	// model filter
+	sumF, _ := st.Stats(stats.Query{Model: "minimax"})
+	if sumF.Totals.Requests != 1 || sumF.Totals.TotalTokens != 2 {
+		t.Fatalf("filtered stats = %+v, want 1/2", sumF.Totals)
+	}
+
+	if err := st.Reset(); err != nil {
+		t.Fatalf("reset: %v", err)
+	}
+	sum2, _ := st.Stats(stats.Query{})
+	if sum2.Totals.Requests != 0 {
+		t.Fatalf("after reset requests = %d, want 0", sum2.Totals.Requests)
+	}
+	// credentials must survive a stats reset
+	creds, _ := st.LoadCredentials()
+	_ = creds
+}

internal/util/util.go

@@ -0,0 +1,36 @@
+// Package util holds small shared helpers used across internal packages to
+// avoid divergent same-named copies.
+package util
+
+import (
+	"crypto/rand"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// FirstNonEmpty returns the first trimmed-non-empty value, or "" when none of
+// the values are non-empty. Callers that need a specific fallback for the
+// all-empty case should apply it explicitly at the call site.
+func FirstNonEmpty(values ...string) string {
+	for _, v := range values {
+		if strings.TrimSpace(v) != "" {
+			return v
+		}
+	}
+	return ""
+}
+
+// RandomRequestID returns a random v4 UUID string, matching the format the
+// Zhanlu plugin sends in the `request` header of every gateway call
+// (crypto.randomUUID()). Extracted so the login and phone-code paths share
+// one implementation.
+func RandomRequestID() string {
+	b := make([]byte, 16)
+	if _, err := rand.Read(b); err != nil {
+		return fmt.Sprintf("%d", time.Now().UnixNano())
+	}
+	b[6] = (b[6] & 0x0f) | 0x40
+	b[8] = (b[8] & 0x3f) | 0x80
+	return fmt.Sprintf("%x-%x-%x-%x-%x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:])
+}

internal/zhanlu/client.go

@@ -13,6 +13,7 @@ import (
 
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/auth"
 	"git.misaka.ren/M1saka/zhanlu_proxy/internal/sign"
+	"git.misaka.ren/M1saka/zhanlu_proxy/internal/util"
 )
 
 type Client struct {
@@ -93,8 +94,8 @@ func (c *Client) ProvisionAPIKey(ctx context.Context, email, organization, team
 	if strings.TrimSpace(email) == "" {
 		return "", fmt.Errorf("profile email is required to provision an API key")
 	}
-	org := firstNonEmpty(organization, "未配置")
-	tm := firstNonEmpty(team, "未配置")
+	org := util.FirstNonEmpty(organization, "未配置")
+	tm := util.FirstNonEmpty(team, "未配置")
 	// Field order matters: the SM2 signature covers the exact JSON body bytes,
 	// matching the plugin's JSON.stringify({email, organization, team}).
 	body, err := json.Marshal(struct {
@@ -195,7 +196,7 @@ func (c *Client) Models(ctx context.Context, apiKey string) ([]string, error) {
 	seen := map[string]bool{}
 	models := make([]string, 0, len(out.Data))
 	for _, m := range out.Data {
-		id := firstNonEmpty(m.ModelName, m.ID, m.ModelInfo.ID)
+		id := util.FirstNonEmpty(m.ModelName, m.ID, m.ModelInfo.ID)
 		if id == "" || seen[id] {
 			continue
 		}
@@ -209,7 +210,7 @@ func (c *Client) setPluginHeaders(req *http.Request) {
 	req.Header.Set("Content-Type", "application/json")
 	req.Header.Set("plugin_type", "zhanlu_ide")
 	req.Header.Set("plugin_version", c.PluginVersion)
-	req.Header.Set("request", randomRequestID())
+	req.Header.Set("request", util.RandomRequestID())
 }
 
 func decodeJSON(resp *http.Response, out any) error {
@@ -237,15 +238,6 @@ func findString(m map[string]any, keys ...string) string {
 	return ""
 }
 
-func firstNonEmpty(values ...string) string {
-	for _, v := range values {
-		if strings.TrimSpace(v) != "" {
-			return v
-		}
-	}
-	return ""
-}
-
 const alnum = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789"
 
 func randomAlnum(n int) string {
@@ -256,13 +248,3 @@ func randomAlnum(n int) string {
 	}
 	return string(b)
 }
-
-func randomRequestID() string {
-	b := make([]byte, 16)
-	if _, err := rand.Read(b); err != nil {
-		return fmt.Sprintf("%d", time.Now().UnixNano())
-	}
-	b[6] = (b[6] & 0x0f) | 0x40
-	b[8] = (b[8] & 0x3f) | 0x80
-	return fmt.Sprintf("%x-%x-%x-%x-%x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:])
-}

scripts/test-instance.sh

@@ -0,0 +1,96 @@
+#!/usr/bin/env bash
+# 实例测试流程固化脚本
+#
+#   ./scripts/test-instance.sh start    # 编译新二进制 + 以当前目录 zhanlu.db 启动
+#   ./scripts/test-instance.sh stop     # 停止进程 + 清理 *-shm/*-wal
+#   ./scripts/test-instance.sh status   # 查看运行状态
+#
+# 可用环境变量覆盖默认值:
+#   ZHANLU_DB_FILE       数据库路径        (默认 <项目根>/zhanlu.db)
+#   ZHANLU_LISTEN_ADDR   监听地址          (默认 127.0.0.1:8080)
+set -euo pipefail
+
+# 项目根目录(脚本位于 scripts/ 下)
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+WORKDIR="$ROOT/tmp"                       # 二进制/pid/log 都放这里(已 gitignore)
+BIN="$WORKDIR/zhanlu-proxy-test"
+PIDFILE="$WORKDIR/test-instance.pid"
+LOGFILE="$WORKDIR/test-instance.log"
+DB="${ZHANLU_DB_FILE:-$ROOT/zhanlu.db}"
+ADDR="${ZHANLU_LISTEN_ADDR:-127.0.0.1:8080}"
+
+start() {
+  mkdir -p "$WORKDIR"
+  if [[ -f "$PIDFILE" ]] && kill -0 "$(cat "$PIDFILE")" 2>/dev/null; then
+    echo "已有实例在运行 (pid $(cat "$PIDFILE"))，请先执行: $0 stop" >&2
+    exit 1
+  fi
+  echo "==> 编译二进制..."
+  (cd "$ROOT" && go build -o "$BIN" ./cmd/zhanlu-proxy)
+  echo "==> 启动 (db=$DB  addr=$ADDR)"
+  ZHANLU_DB_FILE="$DB" ZHANLU_LISTEN_ADDR="$ADDR" \
+    nohup "$BIN" > "$LOGFILE" 2>&1 &
+  echo $! > "$PIDFILE"
+  sleep 1
+  if ! kill -0 "$(cat "$PIDFILE")" 2>/dev/null; then
+    echo "启动失败，日志:" >&2
+    cat "$LOGFILE" >&2
+    rm -f "$PIDFILE"
+    exit 1
+  fi
+  port="${ADDR##*:}"
+  echo "==> 已启动 pid=$(cat "$PIDFILE")"
+  echo "    管理后台: http://127.0.0.1:$port/admin"
+  echo "    日志:     $LOGFILE"
+  echo "    测试完成后执行: $0 stop"
+}
+
+stop() {
+  pid=""
+  [[ -f "$PIDFILE" ]] && pid="$(cat "$PIDFILE")"
+  if [[ -n "$pid" ]] && kill -0 "$pid" 2>/dev/null; then
+    kill "$pid" 2>/dev/null || true
+    # 优雅等待退出(给 SQLite checkpoint WAL 的时间)
+    for _ in $(seq 1 40); do
+      kill -0 "$pid" 2>/dev/null || break
+      sleep 0.25
+    done
+    if kill -0 "$pid" 2>/dev/null; then
+      echo "==> 优雅退出超时，强制结束 pid=$pid"
+      kill -9 "$pid" 2>/dev/null || true
+    else
+      echo "==> 已停止进程 pid=$pid"
+    fi
+  else
+    echo "==> 无运行中的进程"
+  fi
+  rm -f "$PIDFILE"
+  # 清理 wal/shm(与 DB 同目录同名)
+  removed=0
+  for ext in db-shm db-wal; do
+    f="${DB%.*}.$ext"
+    if [[ -f "$f" ]]; then
+      rm -f "$f"
+      echo "==> 已清理 $f"
+      removed=1
+    fi
+  done
+  [[ $removed -eq 0 ]] && echo "==> 无 shm/wal 需清理"
+  rm -f "$BIN"
+  echo "==> 完成"
+}
+
+status() {
+  if [[ -f "$PIDFILE" ]] && kill -0 "$(cat "$PIDFILE")" 2>/dev/null; then
+    echo "运行中 pid=$(cat "$PIDFILE")  addr=$ADDR  db=$DB"
+  else
+    echo "未运行"
+  fi
+}
+
+case "${1:-}" in
+  start)  start ;;
+  stop)   stop ;;
+  status) status ;;
+  *) echo "用法: $0 {start|stop|status}" >&2; exit 1 ;;
+esac