chore: bump version to 3.3.24

Co-authored-by: Codex <noreply@openai.com>

.github/workflows/_quality.yml

@@ -33,6 +33,8 @@ jobs:
     timeout-minutes: 30
     env:
       CARGO_TERM_COLOR: always
+      # Rust 1.97 编译 AWS SDK 调试信息需要更大的 codegen 线程栈。
+      RUST_MIN_STACK: 16777216
       CPR_TEST_DATABASE_URL: postgres://codex_proxy:codex_proxy@127.0.0.1:5432/codex_proxy
       CPR_TEST_REDIS_URL: redis://127.0.0.1:6379
     services:

README.md

@@ -5,192 +5,161 @@
 
 # Codex Proxy RS
 
-**基于 Rust 的多 Provider 透明 AI 网关**
+集中管理 OpenAI、xAI 账号，为 Codex 和兼容客户端提供统一的访问地址与密钥。
 
 [![CI](https://github.com/zyycn/codex-proxy-rs/actions/workflows/ci.yml/badge.svg)](https://github.com/zyycn/codex-proxy-rs/actions/workflows/ci.yml)
 [![Release](https://img.shields.io/github/v/release/zyycn/codex-proxy-rs?display_name=tag&sort=semver&style=flat-square)](https://github.com/zyycn/codex-proxy-rs/releases)
 [![GHCR](https://img.shields.io/badge/GHCR-codex--proxy--rs-2496ED?logo=docker&logoColor=white&style=flat-square)](https://github.com/zyycn/codex-proxy-rs/pkgs/container/codex-proxy-rs)
 [![MIT](https://img.shields.io/badge/License-MIT-blue.svg?style=flat-square)](https://opensource.org/license/mit)
 
-[快速开始](#快速开始) · [客户端接入](#客户端接入) · [运行维护](#运行维护) · [部署](deploy/README.md) · [接口](docs/api.md) · [架构](docs/architecture.md)
+[快速开始](#快速开始) · [客户端接入](#客户端接入) · [常见问题](#常见问题) · [部署说明](deploy/README.md)
 
 </div>
 
-> [!NOTE]
-> 数据面支持 OpenAI Responses、Images 和模型目录协议，不提供 `/v1/chat/completions`。
-
 ## 功能
 
-| 领域 | 当前能力 |
-| --- | --- |
-| 数据面 | Responses JSON、SSE、WebSocket（含 Codex 子代理语义）、Images 生成/编辑与模型目录 |
-| Provider | 固定 `openai`、`xai` 两个 Provider，各自拥有 OAuth、credential、额度、目录与 transport |
-| 路由 | Client Key 限定账号分组；按模型能力编译 Provider 候选；支持全局精确模型映射、会话亲和和安全 fallback |
-| 透明边界 | OpenAI 保留未知请求字段及字段顺序，并原样转发 SSE、WebSocket 与 Images 业务字节；xAI 在 Provider 内转换协议 |
-| 账号 | OAuth、OpenAI AT/RT 与账号文件导入，credential/额度刷新、账号恢复及主动额度重置卡 |
-| 管理 | 账号与分组、Client Key、运行设置、用量/错误观测、S3/R2 备份、系统更新与回滚 |
-| 计量 | Token、图片、费用、延迟、账号与 Provider 归因，并记录上游响应确认的实际 `service_tier` |
+- 在网页中添加账号，查看额度、请求记录和费用，刷新或重新授权过期账号。
+- 用账号分组控制不同密钥可使用的账号，设置并发和请求频率。
+- 接入 Codex CLI、桌面端和支持 Responses API 的客户端。
+- 支持流式对话、WebSocket、图片生成与编辑、Codex 独立搜索。
+- 支持 S3/R2 数据库备份、同大版本在线更新和回滚。
 
-架构边界、状态所有权和请求不变量见 [架构文档](docs/architecture.md)。
+支持 OpenAI、xAI 账号，但图片与独立搜索目前仅通过 OpenAI 提供。模型和功能是否可用取决于上游账号权限。
+**不支持 `/v1/chat/completions`**；仅支持该接口的客户端不能直接接入。
 
 ## 快速开始
 
-需要 Docker Engine 与 Docker Compose Plugin。
+使用 Docker Compose 部署发布镜像 `ghcr.io/zyycn/codex-proxy-rs:latest`，同时启动 PostgreSQL 和 Redis。
+以下命令适用于 Linux amd64/arm64，需要 Docker Engine、Docker Compose Plugin、curl 和 OpenSSL。已有部署请先看
+[升级说明](deploy/README.md#镜像升级与源码构建)，不要覆盖原配置。
 
-### 1. 准备配置和目录
+### 1. 下载部署文件并配置
 
 ```bash
-git clone https://github.com/zyycn/codex-proxy-rs.git
+mkdir -p codex-proxy-rs/deploy
 cd codex-proxy-rs
 
+curl -fsSL https://raw.githubusercontent.com/zyycn/codex-proxy-rs/main/deploy/compose.yaml \
+  -o deploy/compose.yaml
+curl -fsSL https://raw.githubusercontent.com/zyycn/codex-proxy-rs/main/deploy/config.example.yaml \
+  -o deploy/config.example.yaml
+
 mkdir -p .runtime/data .runtime/logs
 install -d -m 0750 .runtime/postgres .runtime/redis
 cp deploy/config.example.yaml deploy/config.yaml
 sudo chown "$(id -u):10001" deploy/config.yaml
 chmod 0640 deploy/config.yaml
-```
-
-分别生成 PostgreSQL 与 Redis 密码：
-
-```bash
-openssl rand -hex 24
-openssl rand -hex 24
-```
-
-把两个结果分别写入 `deploy/config.yaml`，再设置管理员初始密码：
-
-| 配置 | 约束 |
-| --- | --- |
-| `store.database.password` | 48 位十六进制 |
-| `store.redis.password` | 48 位十六进制 |
-| `admin.default_password` | 至少 12 位，不能包含 `$` |
-
-PostgreSQL、Redis 和应用从同一份 `config.yaml` 读取基础设施密码，不需要额外导出环境变量。
-管理员密码只在首次创建管理员时生效。
-
-Linux 需要允许容器组写入运行目录：
-
-```bash
 sudo chown -R "$(id -u):10001" .runtime/data .runtime/logs
 chmod 0770 .runtime/data .runtime/logs
 ```
 
-### 2. 启动
+分别生成数据库和 Redis 密码：
+
+```bash
+openssl rand -hex 24
+openssl rand -hex 24
+```
+
+编辑 `deploy/config.yaml`，填好以下三项：
+
+| 配置项 | 填写内容 |
+| --- | --- |
+| `store.database.password` | 第一个生成的 48 位十六进制密码 |
+| `store.redis.password` | 第二个生成的 48 位十六进制密码 |
+| `admin.default_password` | 管理员初始密码，至少 12 位，不能包含 `$` |
+
+不需要创建 `.env`。管理员初始密码只在首次创建账号时使用。
+
+### 2. 启动服务
 
 ```bash
 docker compose -f deploy/compose.yaml config --quiet
 docker compose -f deploy/compose.yaml pull
-docker compose -f deploy/compose.yaml up -d --no-build
+docker compose -f deploy/compose.yaml up -d --no-build --wait
 curl -i http://127.0.0.1:8080/healthz
 ```
 
-`204 No Content` 表示应用、PostgreSQL 与 Redis 均可用。管理端地址为
-`http://127.0.0.1:8080`。
+健康检查返回 `204 No Content` 后，打开 `http://127.0.0.1:8080`，
+使用 `admin@cpr.local` 和刚设置的管理员密码登录。
 
-### 3. 初始化网关
+默认地址只能在服务器本机访问。从其他设备使用时，需要配置
+[HTTPS 反向代理](deploy/README.md#公网访问)。
 
-1. 使用 `admin@cpr.local` 与初始密码登录。
-2. 导入或授权 Provider 账号：OpenAI 支持 OAuth、逐行 AT/RT 与 OAuth JSON；
-   xAI 支持 OAuth 和账号 JSON。
-3. 创建账号分组并按需配置全局模型映射。未命中的模型名原样交给候选 Provider。
-4. 创建 `sk_...` Client Key，选择可用账号分组，并设置速率和并发限制；空分组范围表示全部账号。
+### 3. 添加账号和密钥
 
-OpenAI RT 导入会先向官方 token endpoint 换取 AT；仅 AT 的账号不具备自动续期能力，
-能否参与调度仍取决于 token 中可验证的身份与后续上游事实。账号导入和首次 OAuth 创建会尽力读取一次额度，
-额度上游暂时失败不会回滚已提交的账号。
+1. 在「账号」中添加 OpenAI 或 xAI 账号。OpenAI 支持 OAuth 授权、AT/RT 和账号 JSON；xAI 支持 OAuth 授权和账号 JSON。
+2. 按需创建账号分组，并把账号加入分组。
+3. 创建客户端密钥，选择可用分组。**不选分组表示可使用全部账号**。
+4. 点击该密钥的「使用密钥」，复制客户端配置。
 
-OpenAI 账号的主动重置卡只在用户打开相应弹窗或手工刷新时查询。消费成功后管理端重新查询卡片与额度；
-系统不会通过修改本地 `reset_at` 伪造重置。完整账号合同见 [账号接口](docs/api.md#5-账号)。
-
-> [!IMPORTANT]
-> xAI 使用 OAuth session，不支持把 xAI API Key 作为上游 credential。
-
-完整的权限、密码轮换、备份和升级规则见 [部署文档](deploy/README.md)。
+只导入 AT 的账号不能自动续期。xAI 暂不支持用 API Key 导入上游账号。
 
 ## 客户端接入
 
+### Codex CLI 和桌面端
+
+在「使用密钥」中按系统复制 `config.toml` 和 `auth.json`，也可以通过 CCSwitch 导入。
+已有文件先备份，再合并配置并重启 Codex。客户端使用代理密钥，OpenAI 账号登录状态由服务端管理。
+
+新配置已启用原生生图。可以直接提出需求，例如：
+
+> 帮我做一个咖啡店首页，生成一张咖啡主题配图并用到页面中。
+
+是否调用生图由 Codex 根据任务决定，也可以明确要求生成图片。需要支持生图的客户端、模型和 OpenAI 账号；
+普通对话可用不代表账号一定有生图额度。
+
+完整配置和旧版配置迁移见 [客户端配置](deploy/README.md#客户端配置)。
+
+### 其他客户端
+
+选择支持 **Responses API** 的客户端，填写：
+
 | 配置 | 值 |
 | --- | --- |
-| Base URL | `http://127.0.0.1:8080/v1` |
-| API Key | 管理端创建的 `sk_...` |
-| 鉴权 | `Authorization: Bearer <client-api-key>` |
+| Base URL | `http://127.0.0.1:8080/v1`，远程使用时换成服务器的 HTTPS 地址 |
+| API Key | 管理端创建的 `sk_...` 密钥 |
+
+可先检查密钥是否能读取模型列表：
 
 ```bash
 curl http://127.0.0.1:8080/v1/models \
   -H 'Authorization: Bearer <client-api-key>'
-
-curl http://127.0.0.1:8080/v1/responses \
-  -H 'Authorization: Bearer <client-api-key>' \
-  -H 'Content-Type: application/json' \
-  -d '{
-    "model": "<model-id>",
-    "input": "Reply with pong.",
-    "stream": false
-  }'
 ```
 
-<details>
-<summary>公开数据面路由</summary>
+接口列表与请求示例见 [接口文档](docs/api.md)。
 
-| 路由 | 用途 |
-| --- | --- |
-| `POST /v1/responses` | JSON 响应或 SSE Responses 流 |
-| `GET /v1/responses` | Responses WebSocket 升级 |
-| `POST /v1/images/generations` | OpenAI Images 生成 JSON 透明转发 |
-| `POST /v1/images/edits` | OpenAI Images 编辑 JSON 透明转发 |
-| `GET /v1/models` | 当前 Client Key 账号范围内的模型列表 |
-| `GET /v1/models/{model_id}` | OpenAI 兼容模型详情 |
+## 常见问题
 
-所有 `/v1/*` 路由都需要 Client Key。Images 是 OpenAI Provider 自有端点，不读取模型字段，也不参与
-文本模型映射。
+### 仍然要求登录，或无法生成图片
 
-</details>
+重新从管理端复制配置，确认修改的是客户端实际读取的用户配置目录，然后完全退出并重启客户端。
+保留仅含代理密钥的 `auth.json` 不会使新配置失效。详细检查步骤见
+[登录与生图排查](deploy/README.md#登录与生图排查)。
 
-完整的客户端与管理端路由、鉴权、请求字段和 mutation 语义见 [接口文档](docs/api.md)。
+### 没有可用账号或模型
 
-## 运行维护
+检查账号是否启用、凭据是否过期、额度是否耗尽，以及密钥绑定的分组是否包含可用账号。
+「恢复账号」只清除本地错误状态，不能恢复上游权限或增加额度。
+
+### 如何更新或查看日志
+
+同一大版本内拉取发布镜像并重建应用容器：
 
 ```bash
-# 同一大版本内升级发布镜像
 docker compose -f deploy/compose.yaml pull codex-proxy-rs
-docker compose -f deploy/compose.yaml up -d --no-build
-
-# 查看日志
-docker compose -f deploy/compose.yaml logs -f codex-proxy-rs
-
-# 从当前源码构建
-docker compose -f deploy/compose.yaml build codex-proxy-rs
-docker compose -f deploy/compose.yaml up -d
+docker compose -f deploy/compose.yaml up -d --no-build codex-proxy-rs
+docker compose -f deploy/compose.yaml logs --tail=100 -f codex-proxy-rs
 ```
 
-> [!IMPORTANT]
-> `.runtime/` 保存 PostgreSQL、Redis、日志、OpenAI 会话锚点和在线更新状态。删除该目录会永久清除
-> 本地运行状态。
-
 > [!WARNING]
-> 不支持跨大版本在线升级。升级到新的大版本时请使用全新的 `.runtime/` 数据目录，并重新导入或授权
-> Provider 账号与 Client Key。
-
-Compose 默认只绑定 `127.0.0.1`。公网接入应使用 HTTPS 反向代理，转发 WebSocket upgrade 与真实客户端
-IP；不要暴露 PostgreSQL 或 Redis。当前运行拓扑为单副本，不能通过复制应用容器扩容。
-
-## 维护与发布
-
-仓库版本只能从干净、已同步上游的发布分支发布：
-
-```bash
-release/publish <version>
-```
-
-脚本依赖已登录的 GitHub CLI（`gh auth login`），会更新 `release/version.yaml`、创建约定提交和带注释的
-`v<version>` tag，原子推送分支与 tag，再触发 Release 构建。源码提交、仓库发版和运行实例升级是三个
-独立状态；脚本不会登录服务器或改变任何运行实例。
+> 升级前先备份。跨大版本需使用新的数据目录重新部署，不支持直接沿用旧库在线升级。
+> `.runtime/` 包含数据库和运行数据，请勿删除；账号文件、密钥和日志也不要公开分享。
 
 ## 文档
 
-- [部署与运维](deploy/README.md)
-- [客户端与管理端接口](docs/api.md)
+- [客户端接入与生图](deploy/README.md#客户端配置)
+- [部署、备份与恢复](deploy/README.md)
+- [API 参考](docs/api.md)
 - [系统架构](docs/architecture.md)
-- [数据库迁移规则](backend/migrations/README.md)
-- [Release](https://github.com/zyycn/codex-proxy-rs/releases)
-- [容器镜像](https://github.com/zyycn/codex-proxy-rs/pkgs/container/codex-proxy-rs)
+- [管理端主题](docs/theme.md)
+- [数据库迁移](backend/migrations/README.md)

backend/Cargo.lock

@@ -154,9 +154,9 @@ dependencies = [
 
 [[package]]
 name = "aws-lc-rs"
-version = "1.17.1"
+version = "1.16.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4342d8937fc7e5dd9b1c60292261c0670c882a2cd1719cfc11b1af41731e32ad"
+checksum = "a054912289d18629dc78375ba2c3726a3afe3ff71b4edba9dedfca0e3446d1fc"
 dependencies = [
  "aws-lc-sys",
  "untrusted 0.7.1",
@@ -165,15 +165,14 @@ dependencies = [
 
 [[package]]
 name = "aws-lc-sys"
-version = "0.42.0"
+version = "0.39.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6d9ceb1da931507a12f4fccea479dccd00da1943e1b4ae72d8e502d707361444"
+checksum = "1fa7e52a4c5c547c741610a2c6f123f3881e409b714cd27e6798ef020c514f0a"
 dependencies = [
  "cc",
  "cmake",
  "dunce",
  "fs_extra",
- "pkg-config",
 ]
 
 [[package]]
@@ -195,7 +194,7 @@ dependencies = [
  "bytes-utils",
  "fastrand",
  "http 0.2.12",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 0.4.6",
  "http-body 1.0.1",
  "percent-encoding",
@@ -231,7 +230,7 @@ dependencies = [
  "hex",
  "hmac",
  "http 0.2.12",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "lru",
  "percent-encoding",
@@ -257,7 +256,7 @@ dependencies = [
  "hex",
  "hmac",
  "http 0.2.12",
- "http 1.4.2",
+ "http 1.4.0",
  "percent-encoding",
  "sha2 0.11.0",
  "time",
@@ -286,7 +285,7 @@ dependencies = [
  "bytes",
  "crc-fast",
  "hex",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
  "md-5",
@@ -320,7 +319,7 @@ dependencies = [
  "bytes-utils",
  "futures-core",
  "futures-util",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
  "percent-encoding",
@@ -339,7 +338,7 @@ dependencies = [
  "aws-smithy-runtime-api",
  "aws-smithy-types",
  "h2",
- "http 1.4.2",
+ "http 1.4.0",
  "hyper",
  "hyper-rustls",
  "hyper-util",
@@ -389,7 +388,7 @@ dependencies = [
  "bytes",
  "fastrand",
  "http 0.2.12",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 0.4.6",
  "http-body 1.0.1",
  "http-body-util",
@@ -410,7 +409,7 @@ dependencies = [
  "aws-smithy-types",
  "bytes",
  "http 0.2.12",
- "http 1.4.2",
+ "http 1.4.0",
  "pin-project-lite",
  "tokio",
  "tracing",
@@ -436,7 +435,7 @@ checksum = "7d56e0a4e53127a632224e43633b0fe045fa9e1e3cfc68b9830f1115e103f910"
 dependencies = [
  "aws-smithy-runtime-api",
  "aws-smithy-types",
- "http 1.4.2",
+ "http 1.4.0",
 ]
 
 [[package]]
@@ -450,7 +449,7 @@ dependencies = [
  "bytes-utils",
  "futures-core",
  "http 0.2.12",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 0.4.6",
  "http-body 1.0.1",
  "http-body-util",
@@ -503,7 +502,7 @@ dependencies = [
  "bytes",
  "form_urlencoded",
  "futures-util",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
  "hyper",
@@ -535,7 +534,7 @@ checksum = "08c78f31d7b1291f7ee735c1c6780ccde7785daae9a9206026862dab7d8792d1"
 dependencies = [
  "bytes",
  "futures-core",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
  "mime",
@@ -1478,12 +1477,14 @@ name = "gateway-core"
 version = "0.1.0"
 dependencies = [
  "async-trait",
+ "base64 0.23.1",
  "bytes",
  "futures",
  "futures-timer",
  "semver",
  "serde",
  "serde_json",
+ "sha2 0.11.0",
  "thiserror",
  "tracing",
  "uuid",
@@ -1533,6 +1534,7 @@ dependencies = [
 name = "gateway-store"
 version = "0.1.0"
 dependencies = [
+ "async-stream",
  "async-trait",
  "aws-credential-types",
  "aws-sdk-s3",
@@ -1638,7 +1640,7 @@ dependencies = [
  "fnv",
  "futures-core",
  "futures-sink",
- "http 1.4.2",
+ "http 1.4.0",
  "indexmap",
  "slab",
  "tokio",
@@ -1686,7 +1688,7 @@ dependencies = [
  "base64 0.22.1",
  "bytes",
  "headers-core",
- "http 1.4.2",
+ "http 1.4.0",
  "httpdate",
  "mime",
  "sha1 0.10.7",
@@ -1698,7 +1700,7 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "54b4a22553d4242c49fddb9ba998a99962b5cc6f22cb5a3482bec22522403ce4"
 dependencies = [
- "http 1.4.2",
+ "http 1.4.0",
 ]
 
 [[package]]
@@ -1760,9 +1762,9 @@ dependencies = [
 
 [[package]]
 name = "http"
-version = "1.4.2"
+version = "1.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6970f50e31d6fc17d3fa27329444bfa74e196cf62e95052a3f6fee181dba6425"
+checksum = "e3ba2a386d7f85a81f119ad7498ebe444d2e22c2af0b86b069416ace48b3311a"
 dependencies = [
  "bytes",
  "itoa",
@@ -1786,7 +1788,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1efedce1fb8e6913f23e0c92de8e62cd5b772a67e7b3946df930a62566c93184"
 dependencies = [
  "bytes",
- "http 1.4.2",
+ "http 1.4.0",
 ]
 
 [[package]]
@@ -1797,7 +1799,7 @@ checksum = "b021d93e26becf5dc7e1b75b1bed1fd93124b374ceb73f43d4d4eafec896a64a"
 dependencies = [
  "bytes",
  "futures-core",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "pin-project-lite",
 ]
@@ -1833,21 +1835,22 @@ dependencies = [
 
 [[package]]
 name = "hyper"
-version = "1.10.1"
+version = "1.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "55281c53a1894c864990125767da440a4e630446785086f52523b20033b74498"
+checksum = "2ab2d4f250c3d7b1c9fcdff1cece94ea4e2dfbec68614f7b87cb205f24ca9d11"
 dependencies = [
  "atomic-waker",
  "bytes",
  "futures-channel",
  "futures-core",
  "h2",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "httparse",
  "httpdate",
  "itoa",
  "pin-project-lite",
+ "pin-utils",
  "smallvec",
  "tokio",
  "want",
@@ -1859,7 +1862,7 @@ version = "0.27.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e3c93eb611681b207e1fe55d5a71ecf91572ec8a6705cdb6857f7d8d5242cf58"
 dependencies = [
- "http 1.4.2",
+ "http 1.4.0",
  "hyper",
  "hyper-util",
  "log",
@@ -1881,7 +1884,7 @@ dependencies = [
  "bytes",
  "futures-channel",
  "futures-util",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "hyper",
  "ipnet",
@@ -2593,6 +2596,7 @@ version = "0.0.0"
 dependencies = [
  "async-stream",
  "async-trait",
+ "aws-lc-rs",
  "base64 0.23.1",
  "bytes",
  "chrono",
@@ -2604,6 +2608,7 @@ dependencies = [
  "gateway-protocol",
  "getrandom 0.4.3",
  "hex",
+ "hyper",
  "hyper-rustls",
  "reqwest",
  "roxmltree",
@@ -2619,6 +2624,7 @@ dependencies = [
  "tempfile",
  "thiserror",
  "tokio",
+ "tokio-rustls",
  "tokio-tungstenite 0.28.0",
  "tokio-util",
  "tracing",
@@ -2910,7 +2916,7 @@ dependencies = [
  "futures-core",
  "futures-util",
  "h2",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
  "hyper",
@@ -3020,9 +3026,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-native-certs"
-version = "0.8.4"
+version = "0.8.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dab5152771c58876a2146916e53e35057e1a4dfa2b9df0f0305b07f611fdea4d"
+checksum = "612460d5f7bea540c490b2b6395d8e34a953e52b491accd6c86c8164c5932a63"
 dependencies = [
  "openssl-probe",
  "rustls-pki-types",
@@ -3032,9 +3038,9 @@ dependencies = [
 
 [[package]]
 name = "rustls-pki-types"
-version = "1.15.1"
+version = "1.14.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2f4925028c7eb5d1fcdaf196971378ed9d2c1c4efc7dc5d011256f76c99c0a96"
+checksum = "be040f8b0a225e40375822a563fa9524378b9d63112f53e19ffff34df5d33fdd"
 dependencies = [
  "web-time",
  "zeroize",
@@ -3698,7 +3704,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.4.3",
+ "getrandom 0.3.4",
  "once_cell",
  "rustix",
  "windows-sys 0.61.2",
@@ -3799,9 +3805,9 @@ checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
 
 [[package]]
 name = "tokio"
-version = "1.53.1"
+version = "1.52.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "202caea871b69668250d242070849eb495be178ed697a3e98aebce5bc81a0bed"
+checksum = "8fc7f01b389ac15039e4dc9531aa973a135d7a4135281b12d7c1bc79fd57fffe"
 dependencies = [
  "bytes",
  "libc",
@@ -3911,7 +3917,7 @@ dependencies = [
  "bitflags",
  "bytes",
  "futures-util",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "pin-project-lite",
  "tower",
@@ -3930,7 +3936,7 @@ dependencies = [
  "bytes",
  "futures-core",
  "futures-util",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body 1.0.1",
  "http-body-util",
  "http-range-header",
@@ -4062,7 +4068,7 @@ dependencies = [
  "data-encoding",
  "flate2",
  "headers",
- "http 1.4.2",
+ "http 1.4.0",
  "httparse",
  "log",
  "rand 0.9.4",
@@ -4081,7 +4087,7 @@ checksum = "6c01152af293afb9c7c2a57e4b559c5620b421f6d133261c60dd2d0cdb38e6b8"
 dependencies = [
  "bytes",
  "data-encoding",
- "http 1.4.2",
+ "http 1.4.0",
  "httparse",
  "log",
  "rand 0.9.4",
@@ -4331,9 +4337,9 @@ dependencies = [
 
 [[package]]
 name = "webpki-roots"
-version = "1.0.8"
+version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bf85cb06032201fa7c6f829d7db5a7e5aa45bcc0655327713065f6f0576731bf"
+checksum = "12bed680863276c63889429bfd6cab3b99943659923822de1c8a39c49e4d722c"
 dependencies = [
  "rustls-pki-types",
 ]
@@ -4535,7 +4541,7 @@ dependencies = [
  "base64 0.22.1",
  "deadpool",
  "futures",
- "http 1.4.2",
+ "http 1.4.0",
  "http-body-util",
  "hyper",
  "hyper-util",

backend/apps/gateway/Cargo.toml

@@ -26,7 +26,7 @@ tokio = { version = "1.52.3", default-features = false, features = ["rt-multi-th
 [dev-dependencies]
 config = { version = "0.15.23", default-features = false, features = ["yaml"] }
 serde_json = "1"
-syn = { version = "3.0.3", features = ["full"] }
+syn = { version = "3.0.3", features = ["full", "visit"] }
 tempfile = "3.27.0"
 
 [lints]

backend/apps/gateway/tests/architecture.rs

@@ -6,7 +6,7 @@ use std::{
     path::{Path, PathBuf},
 };
 
-use syn::Item;
+use syn::{Item, visit::Visit};
 
 /// workspace 成员冻结清单;新增 crate 必须同步扩展本文件的依赖规则。
 pub(super) const WORKSPACE_MEMBERS: &[&str] = &[
@@ -61,6 +61,82 @@ fn gateway_core_depends_on_no_http_db_redis_or_provider_crate() {
     );
 }
 
+#[test]
+fn core_value_owners_do_not_depend_on_execution_or_routing() {
+    let root = backend_root().join("crates/gateway-core/src");
+    for relative in super::rust_files(&root) {
+        let allowed: Option<&[&str]> = match relative.to_str().expect("source path") {
+            "validation.rs" => Some(&[]),
+            "identity.rs" => Some(&["validation"]),
+            "upstream.rs" => Some(&["validation"]),
+            "event.rs" => Some(&["metering", "operation", "upstream", "validation"]),
+            "account/store.rs" => Some(&["account", "error", "identity", "validation"]),
+            path if path.starts_with("policy/") => Some(&["account", "policy", "validation"]),
+            path if path.starts_with("account/") => Some(&["account", "identity", "validation"]),
+            _ => None,
+        };
+        let Some(allowed) = allowed else { continue };
+        let source = fs::read_to_string(root.join(&relative)).expect("read core source");
+        let syntax = syn::parse_file(&source).expect("parse core source");
+        let mut references = CrateReferences::default();
+        references.visit_file(&syntax);
+        for dependency in references.0 {
+            assert!(
+                allowed.contains(&dependency.as_str()),
+                "{} must not depend on crate::{dependency}",
+                relative.display()
+            );
+        }
+    }
+}
+
+#[derive(Default)]
+struct CrateReferences(BTreeSet<String>);
+
+impl<'ast> Visit<'ast> for CrateReferences {
+    fn visit_path(&mut self, path: &'ast syn::Path) {
+        let mut segments = path.segments.iter();
+        if segments
+            .next()
+            .is_some_and(|segment| segment.ident == "crate")
+            && let Some(owner) = segments.next()
+        {
+            self.0.insert(owner.ident.to_string());
+        }
+        syn::visit::visit_path(self, path);
+    }
+
+    fn visit_item_use(&mut self, item: &'ast syn::ItemUse) {
+        self.use_tree(&item.tree, false);
+    }
+}
+
+impl CrateReferences {
+    fn use_tree(&mut self, tree: &syn::UseTree, crate_root: bool) {
+        match tree {
+            syn::UseTree::Path(path) if crate_root => {
+                self.0.insert(path.ident.to_string());
+            }
+            syn::UseTree::Name(name) if crate_root => {
+                self.0.insert(name.ident.to_string());
+            }
+            syn::UseTree::Rename(rename) if crate_root => {
+                self.0.insert(rename.ident.to_string());
+            }
+            syn::UseTree::Path(path) => self.use_tree(&path.tree, path.ident == "crate"),
+            syn::UseTree::Group(group) => {
+                for item in &group.items {
+                    self.use_tree(item, crate_root);
+                }
+            }
+            syn::UseTree::Glob(_) if crate_root => {
+                self.0.insert("*".to_owned());
+            }
+            _ => {}
+        }
+    }
+}
+
 #[test]
 fn gateway_protocol_has_no_workspace_dependencies() {
     for name in dependency_names("crates/gateway-protocol") {
@@ -156,6 +232,46 @@ const ALLOWED_INTERNAL_EDGES: &[(&str, &str)] = &[
 #[test]
 fn workspace_internal_dependency_edges_match_frozen_dag() {
     let metadata = cargo_metadata_json();
+    let actual = workspace_runtime_dependency_edges(&metadata);
+    let mut expected: Vec<(String, String)> = ALLOWED_INTERNAL_EDGES
+        .iter()
+        .map(|(from, to)| ((*from).to_owned(), (*to).to_owned()))
+        .collect();
+    expected.sort();
+
+    assert_eq!(
+        actual, expected,
+        "workspace internal dependency edges diverged from the frozen DAG"
+    );
+}
+
+#[test]
+fn runtime_dependency_edges_exclude_dev_build_and_external_dependencies() {
+    let metadata = serde_json::json!({
+        "packages": [
+            {
+                "name": "gateway-api",
+                "dependencies": [
+                    { "name": "gateway-core", "kind": null },
+                    { "name": "gateway-store", "kind": "dev" },
+                    { "name": "gateway-host", "kind": "build" },
+                    { "name": "serde", "kind": null }
+                ]
+            },
+            {
+                "name": "external-package",
+                "dependencies": [{ "name": "gateway-core", "kind": null }]
+            }
+        ]
+    });
+
+    assert_eq!(
+        workspace_runtime_dependency_edges(&metadata),
+        vec![("gateway-api".to_owned(), "gateway-core".to_owned())]
+    );
+}
+
+fn workspace_runtime_dependency_edges(metadata: &serde_json::Value) -> Vec<(String, String)> {
     let packages = metadata["packages"]
         .as_array()
         .expect("cargo metadata packages");
@@ -173,31 +289,14 @@ fn workspace_internal_dependency_edges_match_frozen_dag() {
             if !internal_names.contains(dep_name) {
                 continue;
             }
-            let is_runtime_edge = dependency["dep_kinds"]
-                .as_array()
-                .map(|kinds| {
-                    kinds
-                        .iter()
-                        .any(|kind| kind["kind"].as_str() == Some("normal"))
-                })
-                .unwrap_or(true);
-            if is_runtime_edge {
+            if dependency.get("kind").expect("dependency kind").is_null() {
                 actual.push((name.to_owned(), dep_name.to_owned()));
             }
         }
     }
 
-    let mut expected: Vec<(String, String)> = ALLOWED_INTERNAL_EDGES
-        .iter()
-        .map(|(from, to)| ((*from).to_owned(), (*to).to_owned()))
-        .collect();
     actual.sort();
-    expected.sort();
-
-    assert_eq!(
-        actual, expected,
-        "workspace internal dependency edges diverged from the frozen DAG"
-    );
+    actual
 }
 
 #[test]

backend/apps/gateway/tests/bootstrap.rs

@@ -143,10 +143,7 @@ fn config_loader_should_reject_removed_fingerprint_section() {
 
 #[test]
 fn config_loader_should_reject_invalid_codex_cli_version() {
-    assert_rejected(valid_config().replace(
-        "codex_version: '0.151.0-alpha.7.2'",
-        "codex_version: 'latest'",
-    ));
+    assert_rejected(valid_config().replace("codex_version: '0.153.4'", "codex_version: 'latest'"));
 }
 
 #[test]
@@ -165,7 +162,7 @@ fn config_loader_should_reject_zero_server_port() {
 
 #[test]
 fn config_loader_should_reject_invalid_desktop_profile_fields() {
-    assert_rejected(valid_config().replace("desktop_build: '7377'", "desktop_build: 'build'"));
+    assert_rejected(valid_config().replace("desktop_build: '8109'", "desktop_build: 'build'"));
 }
 
 fn assert_rejected(config: String) {

backend/crates/gateway-admin/Cargo.toml

@@ -22,6 +22,7 @@ gateway-core = { path = "../gateway-core" }
 hex = "0.4.3"
 rand_core = { version = "0.6.4", features = ["getrandom"] }
 serde = { version = "1.0.228", features = ["derive"] }
+serde_json = { version = "1.0.150", features = ["arbitrary_precision", "preserve_order"] }
 secrecy = { version = "0.10.3", features = ["serde"] }
 subtle = "2.6.1"
 thiserror = "2.0.18"
@@ -32,7 +33,6 @@ chrono-tz = "0.10"
 tokio = { version = "1.52", features = ["rt", "time", "macros", "sync"] }
 
 [dev-dependencies]
-serde_json = { version = "1.0.150", features = ["arbitrary_precision", "preserve_order"] }
 sha2 = "0.11"
 tempfile = "3"
 tokio = { version = "1.52.3", default-features = false, features = ["macros", "rt"] }

backend/crates/gateway-admin/src/lib.rs

@@ -304,6 +304,7 @@ pub async fn initialize(
         client_distribution: Arc::new(DefaultClientDistributionService::new(client_distribution)),
         observability: Arc::new(DefaultObservabilityService::new(
             store.observability(),
+            store.accounts(),
             store.settings(),
             registry,
         )),

backend/crates/gateway-admin/src/model/auth.rs

@@ -4,6 +4,12 @@ use chrono::{DateTime, Utc};
 
 use super::{MutationActor, MutationContext, Revision};
 
+/// 会话审计使用的稳定管理员标识。
+#[must_use]
+pub fn admin_session_actor_ref(admin_user_id: &str) -> String {
+    format!("admin:{admin_user_id}")
+}
+
 /// 已认证的管理主体。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum AdminPrincipal {

backend/crates/gateway-admin/src/model/observability.rs

@@ -491,6 +491,8 @@ pub struct DashboardAccountUsage {
     ///
     /// `None` 表示上游未提供可比较的百分比，不应伪造为零。
     pub quota_used_percent: Option<f64>,
+    /// Provider 返回的代表窗口；缺失时保留未知语义。
+    pub quota_window: Option<super::provider_credentials::ProviderQuotaWindow>,
     pub models: Vec<DashboardAccountModelUsage>,
 }
 
@@ -702,6 +704,8 @@ pub struct UsageAttempt {
 /// 一条请求及其全部尝试。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UsageDetail {
+    pub trace: Option<serde_json::Value>,
+    pub related_requests: Vec<serde_json::Value>,
     pub request: UsageRecord,
     pub attempts: Vec<UsageAttempt>,
 }
@@ -783,7 +787,6 @@ pub struct OpsError {
     pub latency_ms: Option<u64>,
     pub message: String,
     pub raw_upstream_error: Option<String>,
-    pub occurrence_count: u32,
     pub client_ip: Option<String>,
     pub user_agent: Option<String>,
     pub reasoning_effort: Option<String>,

backend/crates/gateway-admin/src/model/provider_credentials.rs

@@ -10,6 +10,8 @@ use gateway_core::{
     account::{OpaqueProviderData, ProviderAccountId, ProviderAccountIdentity},
     routing::{ProviderKind, UpstreamModelId},
 };
+use serde::Deserialize;
+use serde_json::{Map, Value};
 use uuid::Uuid;
 
 use super::{
@@ -279,6 +281,101 @@ impl PendingAuthorizationMutation {
     }
 }
 
+impl PendingAuthorizationMutation {
+    /// 编码 v1 中立事务字段；版本号的位置和 Provider 私有外层文档由 adapter 保持。
+    #[must_use]
+    pub fn to_storage_v1(&self) -> Map<String, Value> {
+        let target = match &self.target {
+            AuthorizationMutationTarget::Create { name } => {
+                serde_json::json!({"kind": "create", "name": name})
+            }
+            AuthorizationMutationTarget::Reauthorize { account_id } => {
+                serde_json::json!({"kind": "reauthorize", "account_id": account_id.to_string()})
+            }
+        };
+        let owner = match self.owner_binding.owner() {
+            AuthorizationOwner::AdminSession { admin_user_id } => {
+                serde_json::json!({"kind": "admin_session", "admin_user_id": admin_user_id})
+            }
+            AuthorizationOwner::AdminApiKey => serde_json::json!({"kind": "admin_api_key"}),
+            AuthorizationOwner::System => serde_json::json!({"kind": "system"}),
+        };
+        Map::from_iter([
+            (
+                "provider_kind".to_owned(),
+                Value::String(self.provider_kind.as_str().to_owned()),
+            ),
+            ("target".to_owned(), target),
+            ("owner".to_owned(), owner),
+            (
+                "started_request_id".to_owned(),
+                Value::String(self.owner_binding.started_request_id().to_owned()),
+            ),
+        ])
+    }
+
+    /// 恢复 adapter 已确认版本为 v1 的中立事务字段。
+    ///
+    /// # Errors
+    ///
+    /// 文档字段、Provider kind 或账号身份非法时返回错误，不携带原始 pending 内容。
+    pub fn from_storage_v1(value: Value) -> Result<Self, AdminError> {
+        let invalid = || AdminError::invalid("invalid pending authorization mutation");
+        let document: StoredAuthorizationMutationV1 =
+            serde_json::from_value(value).map_err(|_| invalid())?;
+        let provider_kind = ProviderKind::new(document.provider_kind).map_err(|_| invalid())?;
+        let target = match document.target {
+            StoredAuthorizationTargetV1::Create { name } => {
+                AuthorizationMutationTarget::Create { name }
+            }
+            StoredAuthorizationTargetV1::Reauthorize { account_id } => {
+                AuthorizationMutationTarget::Reauthorize {
+                    account_id: ProviderAccountId::new(account_id).map_err(|_| invalid())?,
+                }
+            }
+        };
+        let owner = match document.owner {
+            StoredAuthorizationOwnerV1::AdminSession { admin_user_id } => {
+                AuthorizationOwner::AdminSession { admin_user_id }
+            }
+            StoredAuthorizationOwnerV1::AdminApiKey => AuthorizationOwner::AdminApiKey,
+            StoredAuthorizationOwnerV1::System => AuthorizationOwner::System,
+        };
+        Ok(Self::new(
+            provider_kind,
+            target,
+            AuthorizationOwnerBinding {
+                owner,
+                started_request_id: document.started_request_id,
+            },
+        ))
+    }
+}
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+struct StoredAuthorizationMutationV1 {
+    provider_kind: String,
+    target: StoredAuthorizationTargetV1,
+    owner: StoredAuthorizationOwnerV1,
+    started_request_id: String,
+}
+
+#[derive(Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+enum StoredAuthorizationTargetV1 {
+    Create { name: String },
+    Reauthorize { account_id: String },
+}
+
+#[derive(Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+enum StoredAuthorizationOwnerV1 {
+    AdminSession { admin_user_id: String },
+    AdminApiKey,
+    System,
+}
+
 /// 启动 Provider OAuth Authorization Code 流程。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct StartAuthorization {
@@ -776,6 +873,20 @@ impl ProviderQuota {
             .map(|(_, used_percent)| used_percent)
     }
 
+    /// 选择 Dashboard 展示的真实窗口；没有百分比时仍保留 Provider 的滚动窗口语义。
+    #[must_use]
+    pub fn representative_window(&self) -> Option<&ProviderQuotaWindow> {
+        self.representative_used_window()
+            .map(|(index, _)| &self.windows[index])
+            .or_else(|| {
+                self.windows
+                    .iter()
+                    .enumerate()
+                    .min_by_key(|(index, window)| (quota_usage_priority(window), *index))
+                    .map(|(_, window)| window)
+            })
+    }
+
     /// 将已确认的账号级额度耗尽事实投影到展示窗口。
     ///
     /// Provider 已指出具体触顶窗口时只归一化这些窗口；否则归一化 Dashboard

backend/crates/gateway-admin/src/ports/store.rs

@@ -6,6 +6,7 @@ use std::sync::Arc;
 
 use async_trait::async_trait;
 use chrono::{DateTime, Utc};
+use futures::stream::BoxStream;
 
 use super::backup::BackupStorePorts;
 use crate::model::{
@@ -289,6 +290,10 @@ pub trait AccountGroupStore: Send + Sync {
     ) -> AdminStoreResult<AccountGroupMutation>;
 }
 
+/// 逐条读取的已计算费用事实；消费结束或丢弃时释放查询资源。
+pub type UsageCalculatedBillingStream<'a> =
+    BoxStream<'a, AdminStoreResult<UsageCalculatedBillingFact>>;
+
 /// 用量、趋势、诊断与运维错误的只读能力。
 #[async_trait]
 pub trait ObservabilityStore: Send + Sync {
@@ -318,12 +323,13 @@ pub trait ObservabilityStore: Send + Sync {
         filter: UsageFilter,
     ) -> AdminStoreResult<Vec<RequestMetricPoint>>;
 
-    /// 返回可由 Provider 重新校验的已计算费用事实，用于恢复标准费用趋势。
-    async fn usage_calculated_billing_facts(
+    /// 流式返回可由 Provider 重新校验的已计算费用事实，不保证顺序。
+    /// 查询及解码错误由流返回；调用方应逐条聚合，避免收集整个区间。
+    fn usage_calculated_billing_facts(
         &self,
         range: TimeRange,
         filter: UsageFilter,
-    ) -> AdminStoreResult<Vec<UsageCalculatedBillingFact>>;
+    ) -> UsageCalculatedBillingStream<'_>;
 
     async fn list_usage_records(&self, query: UsageQuery) -> AdminStoreResult<UsagePage>;
 

backend/crates/gateway-admin/src/use_case/accounts.rs

@@ -269,6 +269,11 @@ impl DefaultAccountsService {
             })
             .await
             .map_err(|error| map_provider_error(error, "provider quota"))?;
+        let stored = if refresh_quota {
+            self.load_account(account_id).await?
+        } else {
+            stored
+        };
         self.attach_quota_local_usage(
             std::slice::from_ref(&stored),
             std::slice::from_mut(&mut quota),

backend/crates/gateway-admin/src/use_case/auth.rs

@@ -69,7 +69,7 @@ impl DefaultAuthService {
             id: format!("audit_{}", Uuid::now_v7().simple()),
             actor_kind: AuditActorKind::AdminSession,
             actor_admin_user_id: Some(self.default_admin_user_id.clone()),
-            actor_ref: format!("admin:{}", self.default_admin_user_id),
+            actor_ref: crate::model::auth::admin_session_actor_ref(&self.default_admin_user_id),
             request_id: None,
             action: action.to_owned(),
             entity_kind: "admin_session".to_owned(),
@@ -183,7 +183,7 @@ impl AuthService for DefaultAuthService {
         if let Some(session) = session {
             let mut event = self.auth_audit("admin.logout", Utc::now());
             event.actor_admin_user_id = Some(session.admin_user_id.clone());
-            event.actor_ref = format!("admin:{}", session.admin_user_id);
+            event.actor_ref = crate::model::auth::admin_session_actor_ref(&session.admin_user_id);
             event.entity_ref = session.admin_user_id;
             self.store
                 .append_audit_event(event)

backend/crates/gateway-admin/src/use_case/backup.rs

@@ -326,7 +326,9 @@ impl DefaultBackupService {
             crate::model::MutationActor::System => AuditActorKind::System,
         };
         let actor_ref = match &context.actor {
-            crate::model::MutationActor::AdminSession { admin_user_id } => admin_user_id.clone(),
+            crate::model::MutationActor::AdminSession { admin_user_id } => {
+                crate::model::auth::admin_session_actor_ref(admin_user_id)
+            }
             crate::model::MutationActor::AdminApiKey => "admin_api_key".to_owned(),
             crate::model::MutationActor::System => "system".to_owned(),
         };

backend/crates/gateway-admin/src/use_case/observability.rs

@@ -9,6 +9,7 @@ use std::{
 
 use async_trait::async_trait;
 use chrono::{DateTime, Duration, Timelike as _, Utc};
+use futures::TryStreamExt;
 use gateway_core::{account::ProviderAccountId, routing::ProviderKind};
 
 use crate::{
@@ -29,7 +30,7 @@ use crate::{
     },
     ports::{
         provider::{ProviderAdminErrorKind, ProviderAdminRegistry},
-        store::{ObservabilityStore, SettingsStore},
+        store::{AccountStore, ObservabilityStore, SettingsStore, UsageCalculatedBillingStream},
     },
 };
 
@@ -126,6 +127,7 @@ pub trait ObservabilityService: Send + Sync {
 
 pub(crate) struct DefaultObservabilityService {
     store: Arc<dyn ObservabilityStore>,
+    accounts: Arc<dyn AccountStore>,
     settings: Arc<dyn SettingsStore>,
     providers: ProviderAdminRegistry,
     dashboard_queries: DashboardQueryCache,
@@ -135,11 +137,13 @@ impl DefaultObservabilityService {
     #[must_use]
     pub(crate) fn new(
         store: Arc<dyn ObservabilityStore>,
+        accounts: Arc<dyn AccountStore>,
         settings: Arc<dyn SettingsStore>,
         providers: ProviderAdminRegistry,
     ) -> Self {
         Self {
             store,
+            accounts,
             settings,
             providers,
             dashboard_queries: DashboardQueryCache::default(),
@@ -301,13 +305,21 @@ impl ObservabilityService for DefaultObservabilityService {
         range: TimeRange,
         filter: UsageFilter,
     ) -> Result<UsageInsights, AdminError> {
-        let (overview, trend, billing_facts) = futures::try_join!(
-            self.store.usage_summary(range, filter.clone()),
-            self.store.usage_trend(range, filter.clone()),
-            self.store.usage_calculated_billing_facts(range, filter),
-        )
-        .map_err(|error| map_store_error(error, "usage insights"))?;
-        let standard_costs = recover_standard_costs(&self.providers, billing_facts)?;
+        let aggregates = async {
+            futures::try_join!(
+                self.store.usage_summary(range, filter.clone()),
+                self.store.usage_trend(range, filter.clone()),
+            )
+            .map_err(|error| map_store_error(error, "usage insights"))
+        };
+        let ((overview, trend), standard_costs) = futures::try_join!(
+            aggregates,
+            recover_standard_costs(
+                &self.providers,
+                self.store
+                    .usage_calculated_billing_facts(range, filter.clone()),
+            ),
+        )?;
         build_usage_insights(overview, trend, standard_costs)
     }
 
@@ -590,12 +602,16 @@ struct UsageCostScenarios {
     no_cache: ScenarioCosts,
 }
 
-fn recover_standard_costs(
+async fn recover_standard_costs(
     providers: &ProviderAdminRegistry,
-    facts: Vec<UsageCalculatedBillingFact>,
+    mut facts: UsageCalculatedBillingStream<'_>,
 ) -> Result<UsageCostScenarios, AdminError> {
     let mut scenarios = UsageCostScenarios::default();
-    for fact in facts {
+    while let Some(fact) = facts
+        .try_next()
+        .await
+        .map_err(|error| map_store_error(error, "usage billing facts"))?
+    {
         let Ok(provider_kind) = ProviderKind::new(fact.provider_kind.clone()) else {
             continue;
         };
@@ -708,8 +724,31 @@ fn usd_cost(costs: &[CurrencyCost]) -> Option<DecimalAmount> {
 
 impl DefaultObservabilityService {
     async fn enrich_dashboard_quotas(&self, accounts: &mut [DashboardAccountUsage]) {
+        if accounts.is_empty() {
+            return;
+        }
+        let now = Utc::now();
+        let ids = accounts
+            .iter()
+            .map(|account| account.account_id.clone())
+            .collect::<Vec<_>>();
+        let mut rolling_usage = self
+            .accounts
+            .load_account_usage(
+                TimeRange {
+                    start: now - Duration::hours(24),
+                    end: now,
+                },
+                &ids,
+            )
+            .await
+            .unwrap_or_default()
+            .into_iter()
+            .map(|usage| (usage.account_id.clone(), usage))
+            .collect::<HashMap<_, _>>();
         let providers = self.providers.clone();
         let quota_reads = accounts.iter().map(|account| {
+            let rolling_usage = rolling_usage.remove(&account.account_id);
             let account_id = ProviderAccountId::new(account.account_id.clone()).ok();
             let provider = ProviderKind::new(account.provider_kind.clone())
                 .ok()
@@ -722,16 +761,19 @@ impl DefaultObservabilityService {
                     .quota(ProviderQuotaRequest {
                         account_id,
                         refresh: false,
-                        rolling_usage: None,
+                        rolling_usage,
                     })
                     .await
                     .ok()
-                    .and_then(|quota| quota.representative_used_percent())
             }
         });
-        let quota_used = futures::future::join_all(quota_reads).await;
-        for (account, used_percent) in accounts.iter_mut().zip(quota_used) {
-            account.quota_used_percent = used_percent;
+        let quotas = futures::future::join_all(quota_reads).await;
+        for (account, quota) in accounts.iter_mut().zip(quotas) {
+            if let Some(mut quota) = quota {
+                quota.apply_limit_reached_display();
+                account.quota_used_percent = quota.representative_used_percent();
+                account.quota_window = quota.representative_window().cloned();
+            }
         }
     }
 

backend/crates/gateway-admin/tests/model/mod.rs

@@ -1,3 +1,4 @@
 //! 模型模块测试：镜像 src/model/ 中带逻辑的模块。
 
 mod backup;
+mod provider_credentials;

backend/crates/gateway-admin/tests/model/provider_credentials.rs

@@ -0,0 +1,134 @@
+use gateway_admin::model::provider_credentials::{
+    AuthorizationMutationTarget, AuthorizationOwnerBinding, PendingAuthorizationMutation,
+    ProviderQuota, ProviderQuotaWindow, ProviderQuotaWindowRole, QuotaLocalUsageAttribution,
+};
+use gateway_admin::model::{MutationActor, MutationContext};
+use gateway_core::account::ProviderAccountId;
+use gateway_core::routing::ProviderKind;
+use serde_json::{Value, json};
+
+#[test]
+fn dashboard_quota_should_preserve_unknown_and_actual_window_facts() {
+    let mut quota = ProviderQuota {
+        observed_at: None,
+        refresh_token_expires_at: None,
+        windows: Vec::new(),
+        limit_reached: false,
+        provider_data: None,
+    };
+    assert!(quota.representative_window().is_none());
+    assert!(quota.representative_used_percent().is_none());
+
+    let rolling = ProviderQuotaWindow {
+        key: "provider_rolling".to_owned(),
+        group: "requests".to_owned(),
+        label: "Provider rolling window".to_owned(),
+        limit_id: None,
+        limit_name: None,
+        role: Some(ProviderQuotaWindowRole::Primary),
+        local_usage_attribution: QuotaLocalUsageAttribution::AccountWide,
+        window_seconds: Some(86_400),
+        used_percent: None,
+        reset_at: None,
+        limit_reached: false,
+        local_usage: None,
+        provider_data: None,
+    };
+    quota.windows.push(rolling.clone());
+    assert_eq!(quota.representative_window(), Some(&rolling));
+    assert!(quota.representative_used_percent().is_none());
+
+    quota.windows[0].used_percent = Some(99.6);
+    quota.apply_limit_reached_display();
+    let actual = quota.representative_window().expect("actual window");
+    assert_eq!(actual.used_percent, Some(99.6));
+    assert!(
+        !actual.limit_reached,
+        "display rounding cannot exhaust quota"
+    );
+    assert_eq!(actual.window_seconds, Some(86_400));
+    assert_eq!(actual.reset_at, None);
+
+    quota.windows.insert(
+        0,
+        ProviderQuotaWindow {
+            key: "model_specific".to_owned(),
+            used_percent: Some(100.0),
+            local_usage_attribution: QuotaLocalUsageAttribution::Unavailable,
+            ..rolling
+        },
+    );
+    assert_eq!(
+        quota
+            .representative_window()
+            .map(|window| window.key.as_str()),
+        Some("provider_rolling")
+    );
+}
+
+#[test]
+fn pending_mutation_v1_should_round_trip_all_targets_and_owners() {
+    for provider in ["openai", "xai"] {
+        for target in [
+            AuthorizationMutationTarget::Create {
+                name: "account".to_owned(),
+            },
+            AuthorizationMutationTarget::Reauthorize {
+                account_id: ProviderAccountId::new("acct_1").expect("account id"),
+            },
+        ] {
+            for actor in [
+                MutationActor::AdminSession {
+                    admin_user_id: "admin-1".to_owned(),
+                },
+                MutationActor::AdminApiKey,
+                MutationActor::System,
+            ] {
+                let expected = PendingAuthorizationMutation::new(
+                    ProviderKind::new(provider).expect("provider"),
+                    target.clone(),
+                    AuthorizationOwnerBinding::from_context(&MutationContext {
+                        actor,
+                        request_id: "request-1".to_owned(),
+                    }),
+                );
+                assert_eq!(
+                    PendingAuthorizationMutation::from_storage_v1(Value::Object(
+                        expected.to_storage_v1()
+                    )),
+                    Ok(expected),
+                );
+            }
+        }
+    }
+}
+
+#[test]
+fn pending_mutation_v1_should_keep_existing_storage_fields_and_reject_invalid_documents() {
+    let legacy = json!({
+        "provider_kind": "openai",
+        "target": { "kind": "reauthorize", "account_id": "acct_1" },
+        "owner": { "kind": "admin_session", "admin_user_id": "admin-1" },
+        "started_request_id": "request-1"
+    });
+    let restored =
+        PendingAuthorizationMutation::from_storage_v1(legacy.clone()).expect("v1 document");
+    assert_eq!(Value::Object(restored.to_storage_v1()), legacy);
+    for (pointer, invalid) in [
+        ("/provider_kind", json!("")),
+        ("/target/account_id", json!("")),
+        ("/target/kind", json!("other")),
+        ("/owner/kind", json!("other")),
+        ("/started_request_id", json!(null)),
+    ] {
+        let mut document = legacy.clone();
+        *document.pointer_mut(pointer).expect("field") = invalid;
+        assert!(
+            PendingAuthorizationMutation::from_storage_v1(document).is_err(),
+            "{pointer}"
+        );
+    }
+    let mut document = legacy;
+    document["schema_version"] = json!(2);
+    assert!(PendingAuthorizationMutation::from_storage_v1(document).is_err());
+}

backend/crates/gateway-admin/tests/use_case/accounts.rs

@@ -72,6 +72,7 @@ pub(super) struct FakeProviderAdmin {
     import_account_ids: Mutex<Vec<String>>,
     quota_requests: Mutex<Vec<ProviderQuotaRequest>>,
     quota: Mutex<ProviderQuota>,
+    quota_refresh_account: Mutex<Option<(Arc<FakeAccountStore>, AccountRecord)>>,
     current_credential_revision: Mutex<Revision>,
     reset_credit_commands: Mutex<Vec<ConsumeProviderResetCredit>>,
 }
@@ -89,6 +90,7 @@ impl FakeProviderAdmin {
             import_account_ids: Mutex::new(vec!["acct_prepared".to_owned()]),
             quota_requests: Mutex::new(Vec::new()),
             quota: Mutex::new(empty_quota()),
+            quota_refresh_account: Mutex::new(None),
             current_credential_revision: Mutex::new(revision(1)),
             reset_credit_commands: Mutex::new(Vec::new()),
         })
@@ -346,6 +348,14 @@ impl ProviderAdmin for FakeProviderAdmin {
     ) -> Result<ProviderQuota, ProviderAdminError> {
         if request.refresh {
             self.record("provider.quota");
+            if let Some((store, account)) = self
+                .quota_refresh_account
+                .lock()
+                .expect("quota refresh account")
+                .take()
+            {
+                store.set_accounts(vec![account]);
+            }
         }
         self.quota_requests
             .lock()
@@ -1098,6 +1108,47 @@ async fn accounts_refresh_should_keep_guard_through_store_commit() {
     assert_eq!(store.audit_requests(), ["refresh-request"]);
 }
 
+#[tokio::test]
+async fn accounts_quota_refresh_should_return_the_updated_account_status() {
+    let events = events();
+    let provider = FakeProviderAdmin::new("openai", events.clone());
+    let mut account = account_record("openai");
+    account.quota = QuotaState::exhausted(
+        QuotaEvidence::UsageLimitReached,
+        std::time::SystemTime::now(),
+        None,
+    );
+    let store = FakeAccountStore::with_account(account.clone(), events.clone());
+    account.quota = QuotaState::allowed(std::time::SystemTime::now());
+    *provider
+        .quota_refresh_account
+        .lock()
+        .expect("quota refresh account") = Some((store.clone(), account));
+
+    let result = accounts_service(provider, store)
+        .await
+        .accounts()
+        .quota(
+            &ProviderAccountId::new("acct_test").expect("account ID"),
+            true,
+        )
+        .await
+        .expect("refresh quota");
+
+    assert_eq!(
+        result.account.quota.access(),
+        gateway_core::account::QuotaAccessState::Allowed
+    );
+    assert_eq!(
+        result.projection.status,
+        gateway_admin::model::accounts::AccountStatus::Normal
+    );
+    assert_eq!(
+        recorded(&events),
+        ["store.load_account", "provider.quota", "store.load_account"]
+    );
+}
+
 #[tokio::test]
 async fn accounts_recover_should_commit_facts_then_return_normal_account() {
     let events = events();

backend/crates/gateway-admin/tests/use_case/mod.rs

@@ -536,13 +536,14 @@ impl ObservabilityStore for UnavailableStore {
         Err(unavailable("usage trend"))
     }
 
-    async fn usage_calculated_billing_facts(
+    fn usage_calculated_billing_facts(
         &self,
         _: TimeRange,
         _: UsageFilter,
-    ) -> AdminStoreResult<Vec<gateway_admin::model::observability::UsageCalculatedBillingFact>>
-    {
-        Err(unavailable("usage billing facts"))
+    ) -> gateway_admin::ports::store::UsageCalculatedBillingStream<'_> {
+        Box::pin(futures::stream::once(async {
+            Err(unavailable("usage billing facts"))
+        }))
     }
 
     async fn list_usage_records(&self, _: UsageQuery) -> AdminStoreResult<UsagePage> {

backend/crates/gateway-admin/tests/use_case/observability.rs

@@ -2,7 +2,7 @@ use std::{
     str::FromStr as _,
     sync::{
         Arc, Mutex,
-        atomic::{AtomicUsize, Ordering},
+        atomic::{AtomicBool, AtomicUsize, Ordering},
     },
     time::Duration as StdDuration,
 };
@@ -537,10 +537,41 @@ async fn usage_records_should_enrich_provider_reported_totals_when_pricing_match
     ));
 }
 
+#[tokio::test]
+async fn usage_insights_should_reject_partial_costs_when_billing_stream_fails() {
+    let now = Utc::now();
+    let range = observation_range(now);
+    let store = Arc::new(FixtureObservabilityStore::new(range));
+    store.replace_calculated_billing_facts(vec![UsageCalculatedBillingFact {
+        bucket_start: quarter_hour_start(now),
+        provider_kind: "openai".to_owned(),
+        upstream_model_id: "gpt-5.5".to_owned(),
+        service_tier: None,
+        input_tokens: Some(800),
+        output_tokens: Some(200),
+        cached_tokens: Some(0),
+        cache_write_tokens: Some(0),
+        total: CurrencyCost {
+            currency: "USD".to_owned(),
+            amount: "1.25".parse().expect("cost"),
+        },
+    }]);
+    store.billing_stream_fails.store(true, Ordering::SeqCst);
+    let services = observability_services_with_calculated_billing(store).await;
+    assert!(
+        services
+            .observability()
+            .usage_insights(range, UsageFilter::default())
+            .await
+            .is_err()
+    );
+}
+
 struct FixtureObservabilityStore {
     trend: Mutex<Vec<RequestMetricPoint>>,
     overview: Mutex<UsageOverview>,
     calculated_billing_facts: Mutex<Vec<UsageCalculatedBillingFact>>,
+    billing_stream_fails: AtomicBool,
     diagnostics: Mutex<Vec<DiagnosticObservation>>,
     runtime_slots: Mutex<Option<DashboardRuntimeSlots>>,
     summary_observed_at: Mutex<Option<DateTime<Utc>>>,
@@ -561,6 +592,7 @@ impl FixtureObservabilityStore {
                 providers: Vec::new(),
             }),
             calculated_billing_facts: Mutex::new(Vec::new()),
+            billing_stream_fails: AtomicBool::new(false),
             diagnostics: Mutex::new(Vec::new()),
             runtime_slots: Mutex::new(None),
             summary_observed_at: Mutex::new(None),
@@ -663,16 +695,23 @@ impl ObservabilityStore for FixtureObservabilityStore {
         Ok(self.trend.lock().expect("trend").clone())
     }
 
-    async fn usage_calculated_billing_facts(
+    fn usage_calculated_billing_facts(
         &self,
         _: TimeRange,
         _: UsageFilter,
-    ) -> AdminStoreResult<Vec<UsageCalculatedBillingFact>> {
-        Ok(self
+    ) -> gateway_admin::ports::store::UsageCalculatedBillingStream<'_> {
+        let facts = self
             .calculated_billing_facts
             .lock()
             .expect("calculated billing facts")
-            .clone())
+            .clone();
+        let error = self
+            .billing_stream_fails
+            .load(Ordering::SeqCst)
+            .then(|| Err(super::unavailable("billing stream failed")));
+        Box::pin(futures::stream::iter(
+            facts.into_iter().map(Ok).chain(error),
+        ))
     }
 
     async fn list_usage_records(&self, query: UsageQuery) -> AdminStoreResult<UsagePage> {

backend/crates/gateway-api/src/admin/accounts/mod.rs

@@ -70,6 +70,7 @@ mod wire;
 
 pub use credentials::*;
 pub use handlers::{profile_avatar_response, router};
+pub(super) use presenter::quota_window_view;
 pub use wire::*;
 
 use credentials::{

backend/crates/gateway-api/src/admin/accounts/presenter.rs

@@ -140,7 +140,7 @@ pub(super) fn account_quota_view(
     )
 }
 
-pub(super) fn quota_window_view(window: ProviderQuotaWindow) -> AccountQuotaWindowView {
+pub(crate) fn quota_window_view(window: ProviderQuotaWindow) -> AccountQuotaWindowView {
     let ProviderQuotaWindow {
         key,
         group,
@@ -230,6 +230,8 @@ pub(super) fn account_usage_view(
         output_tokens_display: display_optional_tokens(usage.output_tokens),
         cached_tokens: usage.cached_tokens,
         cached_tokens_display: display_optional_tokens(usage.cached_tokens),
+        reasoning_tokens: usage.reasoning_tokens,
+        reasoning_tokens_display: display_optional_tokens(usage.reasoning_tokens),
         image_input_tokens: usage.image_input_tokens,
         image_input_tokens_display: display_optional_tokens(usage.image_input_tokens),
         image_output_tokens: usage.image_output_tokens,
@@ -347,6 +349,8 @@ pub(super) fn empty_account_usage() -> AccountUsageView {
         output_tokens_display: "—".to_owned(),
         cached_tokens: None,
         cached_tokens_display: "—".to_owned(),
+        reasoning_tokens: None,
+        reasoning_tokens_display: "—".to_owned(),
         image_input_tokens: None,
         image_input_tokens_display: "—".to_owned(),
         image_output_tokens: None,

backend/crates/gateway-api/src/admin/accounts/wire.rs

@@ -276,6 +276,8 @@ pub struct AccountUsageView {
     pub output_tokens_display: String,
     pub cached_tokens: Option<u64>,
     pub cached_tokens_display: String,
+    pub reasoning_tokens: Option<u64>,
+    pub reasoning_tokens_display: String,
     pub image_input_tokens: Option<u64>,
     pub image_input_tokens_display: String,
     pub image_output_tokens: Option<u64>,

backend/crates/gateway-api/src/admin/observability/presenter.rs

@@ -505,6 +505,8 @@ pub(crate) fn usage_attempt_view(attempt: domain::UsageAttempt) -> UsageAttemptV
 
 pub(crate) fn usage_detail_view(detail: domain::UsageDetail) -> UsageRecordDetailView {
     UsageRecordDetailView {
+        trace: detail.trace,
+        related_requests: detail.related_requests,
         request: usage_record_view(detail.request),
         attempts: detail
             .attempts
@@ -560,7 +562,6 @@ pub(crate) fn ops_error_view(error: domain::OpsError) -> OpsErrorView {
         failure_class: error.failure_kind,
         upstream_send_state: error.upstream_send_state,
         provider_error_code: error.provider_error_code,
-        occurrence_count: error.occurrence_count,
         response_id: error.client_response_id,
         upstream_request_id: error.upstream_request_id,
         latency_ms: error.latency_ms,
@@ -864,6 +865,28 @@ pub(crate) fn dashboard_view(
     } = observation;
     let mut account_usage_views = Vec::with_capacity(account_usage.len());
     for credential in account_usage {
+        let local_requests = credential
+            .quota_window
+            .as_ref()
+            .filter(|window| window.used_percent.is_none())
+            .and_then(|window| window.local_usage.as_ref())
+            .map(|usage| usage.request_count);
+        let (metric_label, metric_value) = match local_requests {
+            Some(count) => (
+                "次数".to_owned(),
+                if count > 0 {
+                    format_compact_number(count)
+                } else {
+                    "—".to_owned()
+                },
+            ),
+            None => (
+                "今日 Token".to_owned(),
+                credential
+                    .total_tokens
+                    .map_or_else(|| "—".to_owned(), format_compact_number),
+            ),
+        };
         account_usage_views.push(DashboardAccountUsageView {
             id: credential.account_id.clone(),
             provider: credential.provider_kind.clone(),
@@ -886,6 +909,11 @@ pub(crate) fn dashboard_view(
                 })
                 .collect(),
             quota_used_percent: credential.quota_used_percent,
+            usage_window: credential
+                .quota_window
+                .map(crate::admin::accounts::quota_window_view),
+            metric_label,
+            metric_value,
             last_used: relative_time(credential.last_used_at, range.end),
         });
     }

backend/crates/gateway-api/src/admin/observability/wire.rs

@@ -247,6 +247,8 @@ pub struct UsageAttemptView {
 #[derive(Serialize)]
 #[serde(rename_all = "camelCase")]
 pub struct UsageRecordDetailView {
+    pub trace: Option<serde_json::Value>,
+    pub related_requests: Vec<serde_json::Value>,
     #[serde(flatten)]
     pub request: UsageRecordView,
     pub attempts: Vec<UsageAttemptView>,
@@ -375,6 +377,9 @@ pub struct DashboardAccountUsageView {
     pub request_count: u64,
     pub request_buckets: Vec<DashboardAccountRequestBucketView>,
     pub quota_used_percent: Option<f64>,
+    pub usage_window: Option<crate::admin::accounts::AccountQuotaWindowView>,
+    pub metric_label: String,
+    pub metric_value: String,
     pub last_used: String,
 }
 
@@ -778,7 +783,6 @@ pub struct OpsErrorView {
     pub failure_class: String,
     pub upstream_send_state: Option<String>,
     pub provider_error_code: Option<String>,
-    pub occurrence_count: u32,
     pub response_id: Option<String>,
     pub upstream_request_id: Option<String>,
     pub latency_ms: Option<u64>,

backend/crates/gateway-api/src/lib.rs

@@ -180,13 +180,16 @@ pub fn initialize(
                 "request",
                 request_id = %request_id,
                 method = %request.method().as_str(),
-                uri = %request
-                    .uri()
-                    .path_and_query()
-                    .map(axum::http::uri::PathAndQuery::as_str)
-                    .unwrap_or_default(),
+                uri = %request.uri().path(),
             )
         }
+    })
+    .on_request(|_request: &Request<_>, _span: &tracing::Span| {
+        tracing::info!(target: "request_trace", stage = "http.received", "HTTP request received");
+    })
+    .on_response(|response: &axum::http::Response<_>, latency: Duration, _span: &tracing::Span| {
+        tracing::info!(target: "request_trace", stage = "http.response", status = response.status().as_u16(),
+            headers_ms = latency.as_millis(), "HTTP response headers ready");
     });
     let router = router
         .layer(PropagateRequestIdLayer::new(request_id_header.clone()))

backend/crates/gateway-api/src/openai/images/http.rs

@@ -10,14 +10,14 @@ use axum::{
 };
 use gateway_core::error::{GatewayError, GatewayErrorKind};
 use gateway_core::operation::{ImageRequest, ImageRequestKind, Operation, RawJsonPayload};
-use serde_json::{Map, Value};
+use serde_json::Value;
 
 use crate::ApiState;
 use crate::openai::{
     auth::{authenticate_client, client_access_error_response},
     error::gateway_error_response,
     provider_endpoint::collect_raw_json_response,
-    responses::request_client_context,
+    responses::{OpenAiRequestHeaders, request_client_context},
 };
 
 const OPENAI_PROTOCOL: &str = "openai";
@@ -95,7 +95,7 @@ fn image_operation(
     headers: &HeaderMap,
     kind: ImageRequestKind,
 ) -> Result<Operation, GatewayError> {
-    let mut context = Map::new();
+    let mut context = OpenAiRequestHeaders::from_headers(headers).session_context();
     if let Some(turn_id) = headers
         .get("x-codex-image-turn-id")
         .and_then(|value| value.to_str().ok())

backend/crates/gateway-api/src/openai/responses/http.rs

@@ -3,6 +3,7 @@
 use std::collections::VecDeque;
 use std::convert::Infallible;
 use std::net::{IpAddr, SocketAddr};
+use std::time::Duration;
 
 use axum::{
     body::{Body, Bytes},
@@ -13,9 +14,10 @@ use axum::{
     },
     response::{IntoResponse, Response},
 };
-use futures::stream;
-use gateway_core::engine::CommitRequirement;
+use futures::{StreamExt, stream};
+use gateway_core::diagnostics::TraceContext;
 use gateway_core::engine::execution::{ClientTransport, ExecutionSession, StartedExecution};
+use gateway_core::engine::{CommitRequirement, EngineError};
 use gateway_core::error::{GatewayError, GatewayErrorKind};
 use gateway_core::event::{ProviderEvent, ProviderResponseHeader};
 use gateway_core::lifecycle::ConnectionGuard;
@@ -35,10 +37,13 @@ use super::{
     request::decode_request_with_headers,
 };
 
+const SSE_KEEPALIVE_INTERVAL: Duration = Duration::from_secs(15);
+
 /// `POST /v1/responses`。
 pub(crate) async fn responses(
     State(state): State<ApiState>,
     connect_info: Option<Extension<ConnectInfo<SocketAddr>>>,
+    ingress_id: Option<Extension<tower_http::request_id::RequestId>>,
     headers: HeaderMap,
     body: Bytes,
 ) -> Response {
@@ -83,15 +88,25 @@ pub(crate) async fn responses(
         Ok(started) => started,
         Err(error) => return gateway_error_response(&error),
     };
+    let trace = started.session.trace();
+    trace.headers("client.request", serde_json::json!({
+        "ingressRequestId": ingress_id.as_ref().and_then(|Extension(id)| id.header_value().to_str().ok()),
+    }), headers.iter().map(|(name, value)| (name.as_str(), value.as_bytes())));
+    trace.capture("client.request.body", &body);
+    let request_id_header = HeaderValue::from_str(started.request_id.as_str()).ok();
     let StartedExecution {
         stream, session, ..
     } = started;
-    if stream {
+    let mut response = if stream {
         stream_execution_response(session, connection_guard).await
     } else {
         drop(connection_guard);
         collect_execution_response(session).await
+    };
+    if let Some(id) = request_id_header {
+        response.headers_mut().insert("x-gateway-request-id", id);
     }
+    response
 }
 
 /// 从 socket 与标准转发头提取旧 Usage 页面使用的诊断事实。
@@ -152,7 +167,8 @@ pub async fn collect_execution_response(session: Box<dyn ExecutionSession>) -> R
     let events = match session.collect_uncommitted().await {
         Ok(events) => events,
         Err(error) => {
-            let response = engine_error_response(&error);
+            let response_headers = session.response_headers().to_vec();
+            let response = engine_error_response_with_headers(&error, &response_headers);
             return execution.record_response_status(response).await;
         }
     };
@@ -168,6 +184,11 @@ pub async fn collect_execution_response(session: Box<dyn ExecutionSession>) -> R
         }
     };
     let response_headers = session.response_headers().to_vec();
+    session.trace().record(
+        "downstream.encoded",
+        serde_json::json!({"transport": "http_json", "bytes": encoded.len()}),
+    );
+    session.trace().dump("downstream.body", &encoded);
     let response = json_body_response(encoded, &response_headers);
 
     let Some(session) = execution.session_mut() else {
@@ -178,7 +199,7 @@ pub async fn collect_execution_response(session: Box<dyn ExecutionSession>) -> R
         .await
     {
         execution.cancel_and_finalize().await;
-        let response = engine_error_response(&error);
+        let response = engine_error_response_with_headers(&error, &response_headers);
         return execution.record_response_status(response).await;
     }
     if !session.is_finalized() {
@@ -246,6 +267,13 @@ fn apply_response_headers(
     response
 }
 
+fn engine_error_response_with_headers(
+    error: &EngineError,
+    response_headers: &[ProviderResponseHeader],
+) -> Response {
+    apply_response_headers(engine_error_response(error), response_headers)
+}
+
 /// 编码首个 SSE frame 后提交下游，再持续驱动同一执行会话。
 pub async fn stream_execution_response(
     session: Box<dyn ExecutionSession>,
@@ -267,7 +295,8 @@ pub async fn stream_execution_response(
             return response;
         }
         Err(error) => {
-            let response = engine_error_response(&error);
+            let response_headers = session.response_headers().to_vec();
+            let response = engine_error_response_with_headers(&error, &response_headers);
             return execution.record_response_status(response).await;
         }
     };
@@ -299,7 +328,7 @@ pub async fn stream_execution_response(
         .await
     {
         execution.cancel_and_finalize().await;
-        let response = engine_error_response(&error);
+        let response = engine_error_response_with_headers(&error, &response_headers);
         return execution.record_response_status(response).await;
     }
     let Some(session) = execution.into_session() else {
@@ -309,9 +338,11 @@ pub async fn stream_execution_response(
     if state.encoder.is_completed() {
         state.finish_completed(Vec::new()).await;
     }
-    let body = Body::from_stream(stream::unfold(state, |mut state| async move {
+    let output = Box::pin(stream::unfold(state, |mut state| async move {
         loop {
             if let Some(chunk) = state.pending.pop_front() {
+                state.trace.dump("downstream.chunk", &chunk);
+                state.handed_off_bytes += chunk.len() as u64;
                 return Some((Ok::<Bytes, Infallible>(chunk), state));
             }
             if state.output_finished {
@@ -320,19 +351,35 @@ pub async fn stream_execution_response(
             state.advance().await;
         }
     }));
+    // 保活只轮询已固定的输出 stream；不能取消并重建 advance/next_event future，
+    // 否则一次心跳就可能丢失正在等待的上游事件或执行终态清理。
+    let body = Body::from_stream(stream::unfold(output, |mut output| async move {
+        let chunk = tokio::select! {
+            biased;
+            chunk = output.next() => chunk?,
+            () = tokio::time::sleep(SSE_KEEPALIVE_INTERVAL) => {
+                Ok(Bytes::from_static(b": keep-alive\n\n"))
+            }
+        };
+        Some((chunk, output))
+    }));
     event_stream_response(body, &response_headers)
 }
 
 fn event_stream_response(body: Body, response_headers: &[ProviderResponseHeader]) -> Response {
-    let mut response = Response::new(body);
+    let mut response = apply_response_headers(Response::new(body), response_headers);
     *response.status_mut() = StatusCode::OK;
     response
         .headers_mut()
         .insert(CONTENT_TYPE, HeaderValue::from_static("text/event-stream"));
+    response.headers_mut().insert(
+        CACHE_CONTROL,
+        HeaderValue::from_static("no-cache, no-transform"),
+    );
     response
         .headers_mut()
-        .insert(CACHE_CONTROL, HeaderValue::from_static("no-cache"));
-    apply_response_headers(response, response_headers)
+        .insert("x-accel-buffering", HeaderValue::from_static("no"));
+    response
 }
 
 fn internal_gateway_response(message: &'static str) -> Response {
@@ -394,12 +441,18 @@ impl Drop for PendingExecution {
         if session.is_finalized() {
             return;
         }
+        session.trace().record(
+            "downstream.cancelled",
+            serde_json::json!({"reason": "response_guard_dropped"}),
+        );
         session.cancel();
         detach_finalize(session);
     }
 }
 
 struct ResponsesStreamState {
+    trace: TraceContext,
+    handed_off_bytes: u64,
     session: Option<Box<dyn ExecutionSession>>,
     encoder: OpenAiResponsesEncoder,
     pending: VecDeque<Bytes>,
@@ -416,6 +469,8 @@ impl ResponsesStreamState {
         connection_guard: Option<Box<dyn ConnectionGuard>>,
     ) -> Self {
         Self {
+            trace: session.trace(),
+            handed_off_bytes: 0,
             session: Some(session),
             encoder,
             pending: initial_frames.into_iter().collect(),
@@ -571,6 +626,11 @@ impl ResponsesStreamState {
 
 impl Drop for ResponsesStreamState {
     fn drop(&mut self) {
+        self.trace.record("downstream.body.closed", serde_json::json!({
+            "handedOffBytes": self.handed_off_bytes,
+            "pendingChunks": self.pending.len(),
+            "outputFinished": self.output_finished, "executionTerminal": self.execution_terminal,
+        }));
         if self.execution_terminal {
             return;
         }

backend/crates/gateway-api/src/openai/responses/request.rs

@@ -16,6 +16,7 @@ const OPENAI_PROTOCOL: &str = "openai";
 const OPENAI_SUBAGENT_KEY: &str = "x-openai-subagent";
 const CODEX_TURN_METADATA_KEY: &str = "x-codex-turn-metadata";
 const PASSTHROUGH_HEADERS_CONTEXT_KEY: &str = "opaque_request_headers";
+const DOWNSTREAM_WEBSOCKET_CONNECTION_ID_CONTEXT_KEY: &str = "downstream_websocket_connection_id";
 
 /// Responses 请求进入共享解码内核时的下游传输来源。
 #[derive(Clone, Copy)]
@@ -35,6 +36,7 @@ pub struct OpenAiRequestHeaders {
     include_timing_metrics: Option<String>,
     codex_window_id: Option<String>,
     parent_thread_id: Option<String>,
+    downstream_websocket_connection_id: Option<String>,
 
     conversation_id: Option<String>,
     session_id: Option<String>,
@@ -60,6 +62,7 @@ impl OpenAiRequestHeaders {
             include_timing_metrics: header_string(headers, "x-responsesapi-include-timing-metrics"),
             codex_window_id: header_string(headers, "x-codex-window-id"),
             parent_thread_id: header_string(headers, "x-codex-parent-thread-id"),
+            downstream_websocket_connection_id: None,
             conversation_id: header_string(headers, "conversation-id")
                 .or_else(|| header_string(headers, "conversation_id")),
             session_id: header_string(headers, "session-id")
@@ -74,12 +77,31 @@ impl OpenAiRequestHeaders {
         }
     }
 
+    /// 绑定代理为当前下游 WebSocket 分配的连接身份。
+    #[must_use]
+    pub(super) fn with_downstream_websocket_connection_id(
+        mut self,
+        connection_id: impl Into<String>,
+    ) -> Self {
+        self.downstream_websocket_connection_id = Some(connection_id.into());
+        self
+    }
+
     fn apply_subagent(&self, body: &mut Map<String, Value>) {
         if let Some(subagent) = self.subagent.as_deref() {
             inject_subagent_metadata(body, subagent);
         }
     }
 
+    /// 独立端点只消费会话身份与 turn metadata，不携带 Responses 的连接状态。
+    pub(crate) fn session_context(&self) -> Map<String, Value> {
+        let mut context = Map::new();
+        insert_protocol_context(&mut context, "session_id", self.session_id.as_ref());
+        insert_protocol_context(&mut context, "thread_id", self.thread_id.as_ref());
+        insert_protocol_context(&mut context, "turn_metadata", self.turn_metadata.as_ref());
+        context
+    }
+
     fn protocol_context(
         &self,
         use_websocket: Option<bool>,
@@ -109,6 +131,11 @@ impl OpenAiRequestHeaders {
             "parent_thread_id",
             self.parent_thread_id.as_ref(),
         );
+        insert_protocol_context(
+            &mut context,
+            DOWNSTREAM_WEBSOCKET_CONNECTION_ID_CONTEXT_KEY,
+            self.downstream_websocket_connection_id.as_ref(),
+        );
         insert_protocol_context(
             &mut context,
             "conversation_id",
@@ -465,6 +492,8 @@ fn passthrough_header_name(name: &str, connection_headers: &[String]) -> bool {
             // 下游鉴权和账号 cookie 绝不能成为上游账号身份。
             | "authorization"
             | "x-api-key"
+            // Codex 的服务端托管认证标记只用于客户端能力判断，不代表上游身份。
+            | "x-openai-actor-authorization"
             | "cookie"
             | "cookie2"
             | "chatgpt-account-id"

backend/crates/gateway-api/src/openai/responses/websocket/connection.rs

@@ -195,6 +195,7 @@ pub struct ResponsesWebSocketConnection {
     expired: Arc<AtomicBool>,
     commands: Option<mpsc::Sender<ConnectionCommand>>,
     incoming: mpsc::Receiver<ConnectionEvent>,
+    exited: oneshot::Receiver<PumpExitReason>,
     pump_task: Option<JoinHandle<()>>,
     stats: Arc<ConnectionStats>,
     config: ConnectionConfig,
@@ -231,7 +232,16 @@ impl ResponsesWebSocketConnection {
 
     /// 等待下一个需要业务层处理的客户端事件。
     pub async fn next_event(&mut self) -> Option<ConnectionEvent> {
-        let event = self.incoming.recv().await;
+        if let Some(reason) = self.exit_reason {
+            return Some(ConnectionEvent::Exited(reason));
+        }
+        let event = tokio::select! {
+            biased;
+            reason = &mut self.exited => {
+                Some(ConnectionEvent::Exited(reason.unwrap_or(PumpExitReason::PumpStopped)))
+            }
+            event = self.incoming.recv() => event,
+        };
         if let Some(ConnectionEvent::Exited(reason)) = event.as_ref() {
             self.exit_reason.get_or_insert(*reason);
         } else if event.is_none() {
@@ -240,6 +250,18 @@ impl ResponsesWebSocketConnection {
         event
     }
 
+    /// 等待连接退出，不消费留给后续串行请求的业务帧。
+    pub async fn wait_for_exit(&mut self) -> PumpExitReason {
+        if let Some(reason) = self.exit_reason {
+            return reason;
+        }
+        let reason = (&mut self.exited)
+            .await
+            .unwrap_or(PumpExitReason::PumpStopped);
+        self.exit_reason = Some(reason);
+        reason
+    }
+
     /// 串行写入文本帧，并等待 pump 确认 transport 写入结果。
     ///
     /// # Errors
@@ -425,10 +447,12 @@ where
     let stats = Arc::new(ConnectionStats::default());
     let (command_tx, command_rx) = mpsc::channel(OUTBOUND_COMMAND_BUFFER);
     let (incoming_tx, incoming_rx) = mpsc::channel(INBOUND_EVENT_BUFFER);
+    let (exit_tx, exit_rx) = oneshot::channel();
     let pump_task = tokio::spawn(run_pump(
         socket,
         command_rx,
         incoming_tx,
+        exit_tx,
         Arc::clone(&connection_id),
         cancellation,
         opened_at,
@@ -442,6 +466,7 @@ where
         expired,
         commands: Some(command_tx),
         incoming: incoming_rx,
+        exited: exit_rx,
         pump_task: Some(pump_task),
         stats,
         config,
@@ -454,6 +479,7 @@ async fn run_pump<S, E>(
     mut socket: S,
     mut commands: mpsc::Receiver<ConnectionCommand>,
     incoming: mpsc::Sender<ConnectionEvent>,
+    exited: oneshot::Sender<PumpExitReason>,
     connection_id: Arc<str>,
     cancellation: CancellationToken,
     opened_at: Instant,
@@ -474,8 +500,13 @@ async fn run_pump<S, E>(
             () = &mut deadline, if !deadline_elapsed => {
                 deadline_elapsed = true;
                 expired.store(true, Ordering::Release);
-                if let Err(reason) = emit_incoming(&incoming, ConnectionEvent::Expired) {
-                    break reason;
+                // 满队列本身已能唤醒空闲协调层；到期标志仍阻止启动下一轮。
+                // 不能因无法追加 Expired 通知而中断正在收尾的响应。
+                match incoming.try_send(ConnectionEvent::Expired) {
+                    Ok(()) | Err(mpsc::error::TrySendError::Full(_)) => {}
+                    Err(mpsc::error::TrySendError::Closed(_)) => {
+                        break PumpExitReason::CoordinatorDropped;
+                    }
                 }
             }
             command = commands.recv() => {
@@ -550,7 +581,8 @@ async fn run_pump<S, E>(
             },
         }
     };
-    let _ = incoming.try_send(ConnectionEvent::Exited(reason));
+    // 退出不能排在待执行请求之后，也不能因业务队列已满而丢失。
+    let _ = exited.send(reason);
     tracing::debug!(
         websocket_connection_id = %connection_id,
         pump_exit_reason = reason.as_str(),

backend/crates/gateway-api/src/openai/responses/websocket/forward.rs

@@ -13,12 +13,10 @@ use crate::openai::error::{gateway_error_contract, gateway_error_from_engine};
 
 use super::{
     super::{DecodedResponsesRequest, OpenAiResponsesEncoder, PendingExecution, ProtocolErrorBody},
-    connection::{ConnectionEvent, FramePhase, ResponsesWebSocketConnection, WriteContext},
+    connection::{FramePhase, ResponsesWebSocketConnection, WriteContext},
     protocol::{error_event, response_metadata_event},
 };
 
-const ACTIVE_RESPONSE_VIOLATION: &str = "Only one response.create may be active per connection";
-
 #[derive(Clone, Copy, PartialEq, Eq)]
 pub(super) enum ForwardOutcome {
     Continue,
@@ -66,29 +64,20 @@ pub(super) async fn forward_execution(
         );
         return send_gateway_error(connection, &error, &request_id).await;
     }
-    let first = loop {
-        match next_active_input(connection, &mut execution).await {
-            ActiveInput::Event(Ok(Some(event))) => break event,
-            ActiveInput::Event(Ok(None)) => {
-                let error = GatewayError::new(
-                    GatewayErrorKind::Internal,
-                    "gateway response ended before its first event",
-                );
-                return send_gateway_error(connection, &error, &request_id).await;
-            }
-            ActiveInput::Event(Err(error)) => {
-                let error = gateway_error_from_engine(&error);
-                return send_gateway_error(connection, &error, &request_id).await;
-            }
-            ActiveInput::Control => continue,
-            ActiveInput::Disconnect => return ForwardOutcome::Disconnect,
-            ActiveInput::ProtocolViolation => {
-                connection
-                    .close_policy(ACTIVE_RESPONSE_VIOLATION, Some(&request_id))
-                    .await;
-                return ForwardOutcome::Disconnect;
-            }
+    let first = match next_active_input(connection, &mut execution).await {
+        ActiveInput::Event(Ok(Some(event))) => event,
+        ActiveInput::Event(Ok(None)) => {
+            let error = GatewayError::new(
+                GatewayErrorKind::Internal,
+                "gateway response ended before its first event",
+            );
+            return send_gateway_error(connection, &error, &request_id).await;
         }
+        ActiveInput::Event(Err(error)) => {
+            return send_initial_engine_error(connection, &mut execution, &error, &request_id)
+                .await;
+        }
+        ActiveInput::Disconnect => return ForwardOutcome::Disconnect,
     };
     let requirement = first.commit_requirement();
     let mut first = first.into_provider_events();
@@ -259,18 +248,35 @@ pub(super) async fn forward_execution(
                 let error = gateway_error_from_engine(&error);
                 return send_gateway_error(connection, &error, &request_id).await;
             }
-            ActiveInput::Control => {}
             ActiveInput::Disconnect => return ForwardOutcome::Disconnect,
-            ActiveInput::ProtocolViolation => {
-                connection
-                    .close_policy(ACTIVE_RESPONSE_VIOLATION, Some(&request_id))
-                    .await;
-                return ForwardOutcome::Disconnect;
-            }
         }
     }
 }
 
+async fn send_initial_engine_error(
+    connection: &mut ResponsesWebSocketConnection,
+    execution: &mut PendingExecution,
+    error: &EngineError,
+    request_id: &Arc<str>,
+) -> ForwardOutcome {
+    let response_headers = execution
+        .session_mut()
+        .map(|session| session.response_headers().to_vec())
+        .unwrap_or_default();
+    if !response_headers.is_empty()
+        && !send_metadata(
+            connection,
+            request_id,
+            response_metadata_event(request_id, &response_headers),
+        )
+        .await
+    {
+        return ForwardOutcome::Disconnect;
+    }
+    let error = gateway_error_from_engine(error);
+    send_gateway_error(connection, &error, request_id).await
+}
+
 fn commit_connection_replay(
     replay: &mut ConnectionReplaySnapshot,
     encoder: &OpenAiResponsesEncoder,
@@ -286,50 +292,39 @@ async fn confirm_completed_execution(
     execution: &mut PendingExecution,
     request_id: &Arc<str>,
 ) -> Result<(), ForwardOutcome> {
-    loop {
-        match next_active_input(connection, execution).await {
-            ActiveInput::Event(Ok(None))
-                if execution
-                    .session_mut()
-                    .is_some_and(|session| session.is_finalized()) =>
-            {
-                return Ok(());
-            }
-            ActiveInput::Event(Ok(None)) => {
-                let error = GatewayError::new(
-                    GatewayErrorKind::Internal,
-                    "gateway response was not finalized after its terminal event",
-                );
-                return Err(send_gateway_error(connection, &error, request_id).await);
-            }
-            ActiveInput::Event(Ok(Some(_))) => {
-                let error = GatewayError::new(
-                    GatewayErrorKind::Internal,
-                    "gateway response continued after its terminal event",
-                );
-                return Err(send_gateway_error(connection, &error, request_id).await);
-            }
-            ActiveInput::Event(Err(error)) => {
-                let error = gateway_error_from_engine(&error);
-                return Err(send_gateway_error(connection, &error, request_id).await);
-            }
-            ActiveInput::Control => {}
-            ActiveInput::Disconnect => return Err(ForwardOutcome::Disconnect),
-            ActiveInput::ProtocolViolation => {
-                connection
-                    .close_policy(ACTIVE_RESPONSE_VIOLATION, Some(request_id))
-                    .await;
-                return Err(ForwardOutcome::Disconnect);
-            }
+    match next_active_input(connection, execution).await {
+        ActiveInput::Event(Ok(None))
+            if execution
+                .session_mut()
+                .is_some_and(|session| session.is_finalized()) =>
+        {
+            Ok(())
         }
+        ActiveInput::Event(Ok(None)) => {
+            let error = GatewayError::new(
+                GatewayErrorKind::Internal,
+                "gateway response was not finalized after its terminal event",
+            );
+            Err(send_gateway_error(connection, &error, request_id).await)
+        }
+        ActiveInput::Event(Ok(Some(_))) => {
+            let error = GatewayError::new(
+                GatewayErrorKind::Internal,
+                "gateway response continued after its terminal event",
+            );
+            Err(send_gateway_error(connection, &error, request_id).await)
+        }
+        ActiveInput::Event(Err(error)) => {
+            let error = gateway_error_from_engine(&error);
+            Err(send_gateway_error(connection, &error, request_id).await)
+        }
+        ActiveInput::Disconnect => Err(ForwardOutcome::Disconnect),
     }
 }
 
 enum ActiveInput {
     Event(Result<Option<CoordinatedEvent>, EngineError>),
-    Control,
     Disconnect,
-    ProtocolViolation,
 }
 
 async fn next_active_input(
@@ -339,15 +334,17 @@ async fn next_active_input(
     let Some(session) = execution.session_mut() else {
         return ActiveInput::Disconnect;
     };
+    // 与 Codex stream_request 的连接锁一致：本轮结束前不消费下一条请求。
+    // pump 继续接收有界业务帧和处理 Ping/Pong；退出通过独立通知取消本轮。
     tokio::select! {
-        event = session.next_event() => ActiveInput::Event(event),
-        message = connection.next_event() => match message {
-            Some(ConnectionEvent::Expired) => ActiveInput::Control,
-            Some(ConnectionEvent::Text(_) | ConnectionEvent::Binary) => {
-                ActiveInput::ProtocolViolation
-            }
-            Some(ConnectionEvent::Exited(_)) | None => ActiveInput::Disconnect,
+        biased;
+        _ = connection.wait_for_exit() => {
+            session.trace().record("downstream.cancelled", serde_json::json!({
+                "reason": "websocket_connection_exit", "connectionId": connection.id(),
+            }));
+            ActiveInput::Disconnect
         },
+        event = session.next_event() => ActiveInput::Event(event),
     }
 }
 
@@ -459,6 +456,7 @@ fn log_terminal_write_success(
     provider_terminal_to_write: std::time::Duration,
 ) {
     tracing::info!(
+        target: "request_trace", stage = "downstream.terminal.written",
         websocket_connection_id = connection.id(),
         request_id = %request_id,
         provider_terminal_to_terminal_write_ms = provider_terminal_to_write.as_millis(),
@@ -473,6 +471,7 @@ fn log_terminal_write_failure(
     provider_terminal_to_write: std::time::Duration,
 ) {
     tracing::info!(
+        target: "request_trace", stage = "downstream.terminal.write_failed",
         websocket_connection_id = connection.id(),
         request_id = %request_id,
         provider_terminal_to_terminal_write_ms = provider_terminal_to_write.as_millis(),

backend/crates/gateway-api/src/openai/responses/websocket/mod.rs

@@ -19,6 +19,7 @@ use axum::{
     response::{IntoResponse, Response},
 };
 use gateway_core::{
+    diagnostics::{TraceContext, diagnostic_json},
     engine::execution::{AuthenticatedClient, ClientTransport},
     lifecycle::{ConnectionGuard, ConnectionLifecycle},
 };
@@ -60,12 +61,14 @@ pub(crate) async fn responses_websocket(
         &headers,
         connect_info.map(|Extension(ConnectInfo(address))| address),
     );
+    let request_headers = OpenAiRequestHeaders::from_headers(&headers);
     ResponsesWebSocketAdapter::new(service).upgrade_with_client_context(
         websocket,
         client,
         client_ip,
         user_agent,
-        OpenAiRequestHeaders::from_headers(&headers),
+        request_headers,
+        headers,
     )
 }
 
@@ -89,15 +92,24 @@ impl ResponsesWebSocketAdapter {
         client_ip: Option<IpAddr>,
         user_agent: Option<String>,
         request_headers: OpenAiRequestHeaders,
+        raw_headers: HeaderMap,
     ) -> Response {
         let connection_guard = match self.service.try_register_connection() {
             Ok(guard) => guard,
             Err(_) => return runtime_unavailable_response().into_response(),
         };
+        let connection_id = self.service.next_request_id().replacen("req_", "ws_", 1);
+        TraceContext::new(&connection_id).headers(
+            "client.connection.headers",
+            serde_json::json!({"transport": "websocket", "method": "GET", "path": "/v1/responses"}),
+            raw_headers
+                .iter()
+                .map(|(name, value)| (name.as_str(), value.as_bytes())),
+        );
         let session = ResponsesWebSocketSession {
             service: self.service.clone(),
             client,
-            connection_id: self.service.next_request_id().replacen("req_", "ws_", 1),
+            connection_id,
             client_ip,
             user_agent,
             request_headers,
@@ -143,6 +155,8 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
     );
     let _connection_guard = connection_guard;
     let cancellation = lifecycle.cancellation();
+    let request_headers =
+        request_headers.with_downstream_websocket_connection_id(connection_id.clone());
     let mut connection = ResponsesWebSocketConnection::new(socket, connection_id, cancellation);
     let mut request_count = 0_u64;
     let mut replay = ConnectionReplaySnapshot::default();
@@ -172,11 +186,12 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
         let decoded = match decode_response_create_with_context(&payload, &request_headers) {
             Ok(decoded) => decoded.with_client_context(client_ip, user_agent.clone()),
             Err(error) => {
-                tracing::info!(
-                    websocket_connection_id = connection.id(),
-                    request_id = %correlation_id,
-                    error = %error,
-                    "Responses WebSocket request rejected"
+                trace_rejected_request(
+                    &correlation_id,
+                    connection.id(),
+                    &payload,
+                    "decode",
+                    &error.to_string(),
                 );
                 if send_protocol_error(
                     &mut connection,
@@ -195,6 +210,13 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
         let decoded = replay.prepare(decoded);
         // deadline 与 Text 可能同时就绪；在任何上游执行开始前再次封住该竞争窗口。
         if connection.is_expired() {
+            trace_rejected_request(
+                &correlation_id,
+                connection.id(),
+                &payload,
+                "connection_expired",
+                "",
+            );
             expire_connection(&mut connection).await;
             break;
         }
@@ -209,6 +231,13 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
         {
             Ok(started) => started,
             Err(error) => {
+                trace_rejected_request(
+                    &correlation_id,
+                    connection.id(),
+                    &payload,
+                    "start",
+                    &error.to_string(),
+                );
                 if send_gateway_error(&mut connection, &error, &correlation_id).await
                     == ForwardOutcome::Disconnect
                 {
@@ -217,6 +246,14 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
                 continue;
             }
         };
+        started.session.trace().record("client.connection", serde_json::json!({
+            "transport": "websocket", "connectionId": connection.id(), "correlationId": correlation_id,
+            "requestIndex": request_count,
+        }));
+        started
+            .session
+            .trace()
+            .capture("client.request.body", payload.as_bytes());
 
         if forward_execution(&mut connection, started, &mut replay).await
             == ForwardOutcome::Disconnect
@@ -232,6 +269,28 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
     connection.log_summary(request_count);
 }
 
+// 尚未建立模型请求的拒绝也保留原文入口，用返回给客户端的 correlation ID 检索。
+fn trace_rejected_request(
+    correlation_id: &str,
+    connection_id: &str,
+    payload: &str,
+    phase: &'static str,
+    error: &str,
+) {
+    let trace = TraceContext::new(correlation_id);
+    trace.record(
+        "client.connection",
+        serde_json::json!({"connectionId": connection_id}),
+    );
+    trace.capture("client.request.body", payload.as_bytes());
+    trace.record(
+        "client.request.rejected",
+        serde_json::json!({
+            "phase": phase, "detail": diagnostic_json(&serde_json::json!({"message": error})),
+        }),
+    );
+}
+
 async fn expire_connection(connection: &mut ResponsesWebSocketConnection) {
     if connection
         .send_text(

backend/crates/gateway-api/src/openai/search/http.rs

@@ -10,18 +10,16 @@ use axum::{
 };
 use gateway_core::error::{GatewayError, GatewayErrorKind};
 use gateway_core::operation::{Operation, RawJsonPayload, StandaloneSearchRequest};
-use serde_json::{Map, Value};
 
 use crate::ApiState;
 use crate::openai::{
     auth::{authenticate_client, client_access_error_response},
     error::gateway_error_response,
     provider_endpoint::collect_raw_json_response,
-    responses::request_client_context,
+    responses::{OpenAiRequestHeaders, request_client_context},
 };
 
 const OPENAI_PROTOCOL: &str = "openai";
-const TURN_METADATA_CONTEXT_KEY: &str = "turn_metadata";
 
 /// `POST /v1/alpha/search`。
 pub(crate) async fn standalone_search(
@@ -54,16 +52,7 @@ pub(crate) async fn standalone_search(
 }
 
 fn search_operation(body: Bytes, headers: &HeaderMap) -> Result<Operation, GatewayError> {
-    let mut context = Map::new();
-    if let Some(turn_metadata) = headers
-        .get("x-codex-turn-metadata")
-        .and_then(|value| value.to_str().ok())
-    {
-        context.insert(
-            TURN_METADATA_CONTEXT_KEY.to_owned(),
-            Value::String(turn_metadata.to_owned()),
-        );
-    }
+    let context = OpenAiRequestHeaders::from_headers(headers).session_context();
     let payload = RawJsonPayload::new(OPENAI_PROTOCOL, body)
         .map_err(|_| {
             GatewayError::new(

backend/crates/gateway-api/tests/admin/accounts/mod.rs

@@ -348,6 +348,8 @@ mod response {
             output_tokens_display: "-".to_owned(),
             cached_tokens: None,
             cached_tokens_display: "-".to_owned(),
+            reasoning_tokens: None,
+            reasoning_tokens_display: "-".to_owned(),
             image_input_tokens: None,
             image_input_tokens_display: "-".to_owned(),
             image_output_tokens: None,
@@ -374,6 +376,8 @@ mod response {
         let value = serde_json::to_value(view).expect("serialize account usage");
         assert!(value["inputTokens"].is_null());
         assert!(value["totalTokens"].is_null());
+        assert!(value["reasoningTokens"].is_null());
+        assert_eq!(value["reasoningTokensDisplay"], "-");
         assert_eq!(value["createdTokensDisplay"], "-");
     }
 

backend/crates/gateway-api/tests/admin/mod.rs

@@ -875,13 +875,14 @@ impl ObservabilityStore for UnusedStore {
         Err(unavailable("usage trend"))
     }
 
-    async fn usage_calculated_billing_facts(
+    fn usage_calculated_billing_facts(
         &self,
         _: TimeRange,
         _: UsageFilter,
-    ) -> AdminStoreResult<Vec<gateway_admin::model::observability::UsageCalculatedBillingFact>>
-    {
-        Err(unavailable("usage billing facts"))
+    ) -> gateway_admin::ports::store::UsageCalculatedBillingStream<'_> {
+        Box::pin(futures::stream::once(async {
+            Err(unavailable("usage billing facts"))
+        }))
     }
 
     async fn list_usage_records(&self, query: UsageQuery) -> AdminStoreResult<UsagePage> {

backend/crates/gateway-api/tests/admin/observability/response.rs

@@ -71,6 +71,9 @@ fn dashboard_account_usage_should_keep_daily_request_timeline() {
             request_count: 3,
         }],
         quota_used_percent: None,
+        usage_window: None,
+        metric_label: "次数".to_owned(),
+        metric_value: "3".to_owned(),
         last_used: "刚刚".to_owned(),
     })
     .expect("dashboard account usage");
@@ -426,6 +429,8 @@ async fn usage_detail_should_keep_attempt_snapshot_contract() {
         .lock()
         .expect("usage detail")
         .replace(UsageDetail {
+            trace: None,
+            related_requests: Vec::new(),
             request: usage_record_with_account(
                 "req_detail",
                 "acct_snap_a",
@@ -563,7 +568,6 @@ async fn ops_errors_should_keep_account_label_and_authentication_contract() {
             raw_upstream_error: Some(
                 r#"{"error":{"code":"upstream","message":"raw upstream marker"}}"#.to_owned(),
             ),
-            occurrence_count: 1,
             client_ip: Some("203.0.113.8".to_owned()),
             user_agent: Some("codex-cli/0.144.0".to_owned()),
             reasoning_effort: Some("medium".to_owned()),

backend/crates/gateway-api/tests/openai/auth.rs

@@ -13,6 +13,32 @@ fn bearer_client_api_key_should_reject_missing_authorization() {
     );
 }
 
+#[test]
+fn actor_authorization_marker_should_not_replace_client_api_key() {
+    let mut headers = HeaderMap::new();
+    headers.insert(
+        "x-openai-actor-authorization",
+        HeaderValue::from_static("proxy-managed"),
+    );
+
+    assert_eq!(
+        bearer_client_api_key(&headers),
+        Err(ClientApiKeyAuthError::MissingAuthorization)
+    );
+}
+
+#[test]
+fn actor_authorization_marker_should_preserve_existing_bearer_authentication() {
+    let mut headers = HeaderMap::new();
+    headers.insert(
+        "x-openai-actor-authorization",
+        HeaderValue::from_static("proxy-managed"),
+    );
+    headers.insert(AUTHORIZATION, HeaderValue::from_static("Bearer sk_client"));
+
+    assert_eq!(bearer_client_api_key(&headers), Ok("sk_client"));
+}
+
 #[test]
 fn bearer_client_api_key_should_reject_non_utf8_authorization() {
     let mut headers = HeaderMap::new();

backend/crates/gateway-api/tests/openai/images/mod.rs

@@ -260,6 +260,8 @@ async fn image_routes_should_not_decode_bodies_and_should_preserve_both_directio
                     .header(AUTHORIZATION, "Bearer sk_images_test")
                     .header("content-type", "application/json")
                     .header("x-codex-image-turn-id", "turn_image_route")
+                    .header("session-id", "root-image-session")
+                    .header("thread-id", "child-image-thread")
                     .body(Body::from(body.to_vec()))
                     .expect("image request"),
             )
@@ -286,7 +288,7 @@ async fn image_routes_should_not_decode_bodies_and_should_preserve_both_directio
         assert_eq!(captured.body.as_ref(), *body);
         assert_eq!(
             captured.context,
-            json!({"image_turn_id": "turn_image_route"})
+            json!({"image_turn_id": "turn_image_route", "session_id": "root-image-session", "thread_id": "child-image-thread"})
         );
     }
     assert_eq!(

backend/crates/gateway-api/tests/openai/responses/http.rs

@@ -4,6 +4,7 @@ use std::sync::{
     Arc, Mutex,
     atomic::{AtomicBool, Ordering},
 };
+use std::time::Duration;
 
 use axum::{
     body::{Body, to_bytes},
@@ -14,6 +15,7 @@ use axum::{
     },
 };
 use bytes::Bytes;
+use futures::StreamExt;
 use futures::future::{BoxFuture, pending};
 use gateway_core::engine::execution::{
     AuthenticatedClient, ClientAuthenticationError, ExecutionService, ExecutionSession,
@@ -79,6 +81,7 @@ impl Trace {
 
 enum NextStep {
     Event(CoordinatedEvent),
+    DelayedEvent(Duration, CoordinatedEvent),
     Error(EngineError),
     FinalizeCancelled,
     FinalizeSuccess,
@@ -280,6 +283,12 @@ impl ExecutionSession for FakeSession {
                     self.trace.push("next_event");
                     Ok(Some(event))
                 }
+                NextStep::DelayedEvent(delay, event) => {
+                    self.trace.push("wait_event");
+                    tokio::time::sleep(delay).await;
+                    self.trace.push("next_event");
+                    Ok(Some(event))
+                }
                 NextStep::Error(error) => {
                     self.trace.push("next_error");
                     self.finalized = true;
@@ -709,6 +718,154 @@ async fn streaming_encodes_first_frame_before_commit_and_http_delivery() {
     std::mem::forget(response);
 }
 
+#[tokio::test(start_paused = true)]
+async fn streaming_should_keep_alive_during_silent_compaction_without_restarting_execution() {
+    let trace = Arc::new(Trace::default());
+    let session = FakeSession::streaming(
+        Arc::clone(&trace),
+        vec![
+            NextStep::Event(delivery(started(), CommitRequirement::CommitBeforeDelivery)),
+            NextStep::DelayedEvent(
+                Duration::from_secs(240),
+                delivery(completed(), CommitRequirement::AlreadyCommitted),
+            ),
+            NextStep::FinalizeSuccess,
+        ],
+    );
+    let response = stream_execution_response(Box::new(session), None).await;
+    let mut body = response.into_body().into_data_stream();
+    let first = body.next().await.expect("first frame").expect("body bytes");
+    assert!(String::from_utf8_lossy(&first).contains("response.created"));
+    let start = tokio::time::Instant::now();
+
+    for heartbeat in 1..16 {
+        let chunk = tokio::time::timeout(Duration::from_secs(20), body.next())
+            .await
+            .expect("silent upstream must not leave the downstream idle")
+            .expect("heartbeat frame")
+            .expect("heartbeat bytes");
+        assert_eq!(chunk, Bytes::from_static(b": keep-alive\n\n"));
+        assert_eq!(start.elapsed(), Duration::from_secs(15 * heartbeat));
+    }
+
+    let terminal = body
+        .next()
+        .await
+        .expect("terminal frame")
+        .expect("body bytes");
+    assert!(String::from_utf8_lossy(&terminal).contains("response.completed"));
+    assert_eq!(start.elapsed(), Duration::from_secs(240));
+    assert_eq!(
+        body.next().await.expect("done frame").expect("body bytes"),
+        Bytes::from_static(b"data: [DONE]\n\n")
+    );
+    assert!(body.next().await.is_none());
+    assert_eq!(
+        trace.snapshot(),
+        vec![
+            "next_event",
+            "commit",
+            "wait_event",
+            "next_event",
+            "next_end"
+        ]
+    );
+    assert!(!trace.is_cancelled());
+}
+
+#[tokio::test(start_paused = true)]
+async fn dropping_stream_during_keepalive_should_cancel_and_finalize_execution() {
+    let trace = Arc::new(Trace::default());
+    let session = FakeSession::streaming(
+        Arc::clone(&trace),
+        vec![
+            NextStep::Event(delivery(started(), CommitRequirement::CommitBeforeDelivery)),
+            NextStep::DelayedEvent(
+                Duration::from_secs(240),
+                delivery(completed(), CommitRequirement::AlreadyCommitted),
+            ),
+            NextStep::FinalizeCancelled,
+        ],
+    );
+    let response = stream_execution_response(Box::new(session), None).await;
+    let mut body = response.into_body().into_data_stream();
+    let _ = body.next().await.expect("first frame").expect("body bytes");
+    let heartbeat = tokio::time::timeout(Duration::from_secs(20), body.next())
+        .await
+        .expect("heartbeat before client disconnect")
+        .expect("heartbeat frame")
+        .expect("heartbeat bytes");
+    assert_eq!(heartbeat, Bytes::from_static(b": keep-alive\n\n"));
+
+    drop(body);
+    tokio::task::yield_now().await;
+
+    assert!(trace.is_cancelled());
+    assert_eq!(
+        trace.snapshot(),
+        vec!["next_event", "commit", "wait_event", "cancel_finalize"]
+    );
+}
+
+#[tokio::test(start_paused = true)]
+async fn streaming_keepalive_should_wait_for_first_event_before_committing_http_status() {
+    let trace = Arc::new(Trace::default());
+    let session = FakeSession::streaming(
+        Arc::clone(&trace),
+        vec![
+            NextStep::DelayedEvent(
+                Duration::from_secs(45),
+                delivery(started(), CommitRequirement::CommitBeforeDelivery),
+            ),
+            NextStep::FinalizeCancelled,
+        ],
+    );
+    let response = tokio::spawn(stream_execution_response(Box::new(session), None));
+    tokio::task::yield_now().await;
+    tokio::time::advance(Duration::from_secs(30)).await;
+
+    assert!(!response.is_finished());
+    assert!(trace.client_statuses().is_empty());
+    let response = response.await.expect("response task");
+    assert_eq!(response.status(), StatusCode::OK);
+    assert_eq!(trace.snapshot(), vec!["wait_event", "next_event", "commit"]);
+}
+
+#[tokio::test]
+async fn streaming_response_should_disable_proxy_buffering_and_cache_transformation() {
+    let session = FakeSession::streaming(
+        Arc::new(Trace::default()),
+        vec![
+            NextStep::Event(delivery(started(), CommitRequirement::CommitBeforeDelivery)),
+            NextStep::FinalizeCancelled,
+        ],
+    )
+    .with_response_headers(vec![
+        ProviderResponseHeader::new("x-accel-buffering", Bytes::from_static(b"yes")),
+        ProviderResponseHeader::new("cache-control", Bytes::from_static(b"public, max-age=60")),
+    ]);
+
+    let response = stream_execution_response(Box::new(session), None).await;
+
+    assert_eq!(response.headers()["x-accel-buffering"], "no");
+    assert_eq!(
+        response.headers()["cache-control"],
+        "no-cache, no-transform"
+    );
+    assert_eq!(
+        response
+            .headers()
+            .get_all("x-accel-buffering")
+            .iter()
+            .count(),
+        1
+    );
+    assert_eq!(
+        response.headers().get_all("cache-control").iter().count(),
+        1
+    );
+}
+
 #[tokio::test]
 async fn streaming_response_should_append_ordinary_headers_and_skip_unrepresentable_or_sensitive_ones()
  {
@@ -893,6 +1050,41 @@ async fn streaming_rate_limit_before_first_frame_should_persist_the_returned_429
     assert_eq!(trace.client_statuses(), vec![429]);
 }
 
+#[tokio::test]
+async fn streaming_failure_before_first_frame_should_return_safe_observed_headers() {
+    let trace = Arc::new(Trace::default());
+    let session = FakeSession::streaming(
+        Arc::clone(&trace),
+        vec![NextStep::Error(EngineError::Provider(ProviderError::new(
+            ProviderErrorKind::Unavailable,
+            UpstreamSendState::Ambiguous,
+        )))],
+    )
+    .with_response_headers(vec![
+        ProviderResponseHeader::new(
+            "x-codex-turn-state",
+            Bytes::from_static(b"turn-state-before-close"),
+        ),
+        ProviderResponseHeader::new(
+            "authorization",
+            Bytes::from_static(b"must-not-cross-boundary"),
+        ),
+    ]);
+
+    let response = stream_execution_response(Box::new(session), None).await;
+
+    assert_eq!(response.status(), StatusCode::BAD_GATEWAY);
+    assert_eq!(
+        response
+            .headers()
+            .get("x-codex-turn-state")
+            .and_then(|value| value.to_str().ok()),
+        Some("turn-state-before-close")
+    );
+    assert!(response.headers().get("authorization").is_none());
+    assert_eq!(trace.client_statuses(), vec![502]);
+}
+
 #[tokio::test]
 async fn streaming_upstream_http_failure_before_first_frame_should_preserve_raw_response() {
     let trace = Arc::new(Trace::default());
@@ -979,11 +1171,29 @@ async fn buffered_upstream_http_failure_should_preserve_raw_response() {
             raw_body.clone(),
         ));
     let session = FakeSession::buffered(Arc::clone(&trace), vec![started(), completed()])
-        .with_collect_error(EngineError::Provider(error));
+        .with_collect_error(EngineError::Provider(error))
+        .with_response_headers(vec![
+            ProviderResponseHeader::new(
+                "x-codex-turn-state",
+                Bytes::from_static(b"turn-state-before-close"),
+            ),
+            ProviderResponseHeader::new(
+                "authorization",
+                Bytes::from_static(b"must-not-cross-boundary"),
+            ),
+        ]);
 
     let response = collect_execution_response(Box::new(session)).await;
 
     assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
+    assert_eq!(
+        response
+            .headers()
+            .get("x-codex-turn-state")
+            .and_then(|value| value.to_str().ok()),
+        Some("turn-state-before-close")
+    );
+    assert!(response.headers().get("authorization").is_none());
     assert_eq!(
         to_bytes(response.into_body(), usize::MAX)
             .await

backend/crates/gateway-api/tests/openai/responses/mod.rs

@@ -281,6 +281,10 @@ fn decoder_should_preserve_ordinary_request_headers_as_opaque_multivalues() {
     headers.insert("content-length", HeaderValue::from_static("999"));
     headers.insert("authorization", HeaderValue::from_static("Bearer client"));
     headers.insert("x-api-key", HeaderValue::from_static("client-key"));
+    headers.insert(
+        "x-openai-actor-authorization",
+        HeaderValue::from_static("proxy-managed"),
+    );
     headers.insert("cookie", HeaderValue::from_static("client=cookie"));
     headers.insert(
         "chatgpt-account-id",
@@ -348,6 +352,7 @@ fn decoder_should_preserve_ordinary_request_headers_as_opaque_multivalues() {
         "content-length",
         "authorization",
         "x-api-key",
+        "x-openai-actor-authorization",
         "cookie",
         "chatgpt-account-id",
         "chatgpt-project-id",

backend/crates/gateway-api/tests/openai/responses/websocket/connection.rs

@@ -2,8 +2,8 @@ use std::{
     fmt,
     pin::Pin,
     sync::{
-        Arc,
-        atomic::{AtomicBool, Ordering},
+        Arc, Mutex,
+        atomic::{AtomicBool, AtomicUsize, Ordering},
     },
     task::{Context, Poll},
     time::{Duration, SystemTime},
@@ -25,6 +25,7 @@ use gateway_core::{
     error::GatewayError,
     event::{GatewayEvent, ProtocolWireEvent, ProviderEvent, ProviderResponseHeader, ResponseMeta},
     lifecycle::CancellationToken,
+    operation::Operation,
     routing::PublicModelId,
 };
 use serde_json::{Value, json};
@@ -32,7 +33,7 @@ use tokio::sync::mpsc::{
     UnboundedReceiver, UnboundedSender, error::TryRecvError, unbounded_channel,
 };
 use tokio_tungstenite::{
-    connect_async,
+    MaybeTlsStream, WebSocketStream, connect_async,
     tungstenite::{
         Message as ClientMessage, client::IntoClientRequest, protocol::frame::coding::CloseCode,
     },
@@ -44,27 +45,33 @@ use crate::openai::{api_router, authenticated_client};
 #[derive(Default)]
 struct CrossingLimitTrace {
     release_terminal: tokio::sync::Notify,
-    started_emitted: AtomicBool,
-    terminal_emitted: AtomicBool,
+    starts: AtomicUsize,
+    inputs: Mutex<Vec<Value>>,
+    cancelled: AtomicBool,
+    cancellation_observed: tokio::sync::Notify,
     finalized: AtomicBool,
 }
 
 struct CrossingLimitSession {
     trace: Arc<CrossingLimitTrace>,
+    started_emitted: bool,
+    terminal_emitted: bool,
+    finalized: bool,
 }
 
 impl ExecutionSession for CrossingLimitSession {
     fn next_event(&mut self) -> BoxFuture<'_, Result<Option<CoordinatedEvent>, EngineError>> {
         Box::pin(async move {
-            if !self.trace.started_emitted.swap(true, Ordering::AcqRel) {
+            if !self.started_emitted {
+                self.started_emitted = true;
                 return Ok(Some(crossing_limit_started_batch()));
             }
-            if !self.trace.terminal_emitted.load(Ordering::Acquire) {
+            if !self.terminal_emitted {
                 self.trace.release_terminal.notified().await;
-                if !self.trace.terminal_emitted.swap(true, Ordering::AcqRel) {
-                    return Ok(Some(crossing_limit_terminal_batch()));
-                }
+                self.terminal_emitted = true;
+                return Ok(Some(crossing_limit_terminal_batch()));
             }
+            self.finalized = true;
             self.trace.finalized.store(true, Ordering::Release);
             Ok(None)
         })
@@ -87,10 +94,13 @@ impl ExecutionSession for CrossingLimitSession {
     }
 
     fn is_finalized(&self) -> bool {
-        self.trace.finalized.load(Ordering::Acquire)
+        self.finalized
     }
 
-    fn cancel(&self) {}
+    fn cancel(&self) {
+        self.trace.cancelled.store(true, Ordering::Release);
+        self.trace.cancellation_observed.notify_one();
+    }
 
     fn detach_finalize(self: Box<Self>) -> BoxFuture<'static, ()> {
         Box::pin(async move {
@@ -127,6 +137,14 @@ impl ExecutionService for CrossingLimitExecution {
         request: StartExecution,
     ) -> BoxFuture<'_, Result<StartedExecution, GatewayError>> {
         Box::pin(async move {
+            self.trace.starts.fetch_add(1, Ordering::AcqRel);
+            if let Operation::Generate(generate) = &request.operation {
+                self.trace
+                    .inputs
+                    .lock()
+                    .expect("input trace lock")
+                    .push(generate.protocol_payload().body()["input"].clone());
+            }
             Ok(StartedExecution {
                 request_id: gateway_core::engine::ModelRequestId::new("req_ws_crossing_limit")
                     .expect("request id"),
@@ -134,6 +152,9 @@ impl ExecutionService for CrossingLimitExecution {
                 stream: request.metadata.stream,
                 session: Box::new(CrossingLimitSession {
                     trace: Arc::clone(&self.trace),
+                    started_emitted: false,
+                    terminal_emitted: false,
+                    finalized: false,
                 }),
             })
         })
@@ -192,7 +213,7 @@ fn crossing_limit_terminal_batch() -> CoordinatedEvent {
 }
 
 use gateway_api::openai::responses::websocket::connection::{
-    ConnectionConfig, ConnectionEvent, ConnectionWriteError, FramePhase,
+    ConnectionConfig, ConnectionEvent, ConnectionWriteError, FramePhase, PumpExitReason,
     ResponsesWebSocketConnection, WriteContext, spawn_connection,
 };
 
@@ -255,12 +276,14 @@ struct PumpHarness {
     incoming: UnboundedSender<Result<Message, TestSocketError>>,
     written: UnboundedReceiver<Message>,
     dropped: Arc<AtomicBool>,
+    cancellation: CancellationToken,
 }
 
 fn test_connection(stall_writes: bool) -> PumpHarness {
     let (incoming_tx, incoming_rx) = unbounded_channel();
     let (written_tx, written_rx) = unbounded_channel();
     let dropped = Arc::new(AtomicBool::new(false));
+    let cancellation = CancellationToken::new();
     let socket = TestSocket {
         incoming: incoming_rx,
         written: written_tx,
@@ -270,7 +293,7 @@ fn test_connection(stall_writes: bool) -> PumpHarness {
     let connection = spawn_connection(
         socket,
         Arc::from("ws_test"),
-        CancellationToken::new(),
+        cancellation.clone(),
         ConnectionConfig::PRODUCTION,
     );
     PumpHarness {
@@ -278,6 +301,7 @@ fn test_connection(stall_writes: bool) -> PumpHarness {
         incoming: incoming_tx,
         written: written_rx,
         dropped,
+        cancellation,
     }
 }
 
@@ -316,6 +340,110 @@ async fn pump_consumes_pong_without_forwarding_a_business_event() {
     assert!(result.is_err(), "Pong unexpectedly reached the coordinator");
 }
 
+#[tokio::test(start_paused = true)]
+async fn waiting_for_exit_should_preserve_queued_business_frames() {
+    let PumpHarness {
+        mut connection,
+        incoming,
+        ..
+    } = test_connection(false);
+    for payload in ["first", "second"] {
+        incoming
+            .send(Ok(Message::Text(payload.into())))
+            .expect("queue text");
+    }
+
+    assert!(
+        tokio::time::timeout(Duration::from_secs(1), connection.wait_for_exit())
+            .await
+            .is_err()
+    );
+    for expected in ["first", "second"] {
+        assert!(
+            matches!(connection.next_event().await, Some(ConnectionEvent::Text(payload)) if payload == expected)
+        );
+    }
+}
+
+#[tokio::test]
+async fn client_close_should_take_priority_over_queued_business_frames() {
+    let PumpHarness {
+        mut connection,
+        incoming,
+        ..
+    } = test_connection(false);
+    incoming
+        .send(Ok(Message::Text("queued".into())))
+        .expect("queue text");
+    incoming
+        .send(Ok(Message::Close(None)))
+        .expect("close socket");
+
+    assert_eq!(
+        tokio::time::timeout(Duration::from_secs(1), connection.wait_for_exit())
+            .await
+            .expect("observe close"),
+        PumpExitReason::ClientClose
+    );
+    assert!(matches!(
+        connection.next_event().await,
+        Some(ConnectionEvent::Exited(PumpExitReason::ClientClose))
+    ));
+}
+
+#[tokio::test]
+async fn full_business_queue_should_not_block_lifecycle_shutdown() {
+    let PumpHarness {
+        mut connection,
+        incoming,
+        mut written,
+        cancellation,
+        ..
+    } = test_connection(false);
+    for _ in 0..32 {
+        incoming
+            .send(Ok(Message::Text("queued".into())))
+            .expect("queue text");
+    }
+    incoming
+        .send(Ok(Message::Ping(vec![1].into())))
+        .expect("send Ping after queue fills");
+    assert!(matches!(written.recv().await, Some(Message::Pong(_))));
+    cancellation.cancel();
+
+    assert_eq!(
+        tokio::time::timeout(Duration::from_secs(1), connection.wait_for_exit())
+            .await
+            .expect("observe shutdown"),
+        PumpExitReason::LifecycleShutdown
+    );
+}
+
+#[tokio::test]
+async fn inbound_overload_should_close_without_executing_queued_frames() {
+    let PumpHarness {
+        mut connection,
+        incoming,
+        ..
+    } = test_connection(false);
+    for _ in 0..33 {
+        incoming
+            .send(Ok(Message::Text("queued".into())))
+            .expect("queue text");
+    }
+
+    assert_eq!(
+        tokio::time::timeout(Duration::from_secs(1), connection.wait_for_exit())
+            .await
+            .expect("observe overload"),
+        PumpExitReason::InboundOverload
+    );
+    assert!(matches!(
+        connection.next_event().await,
+        Some(ConnectionEvent::Exited(PumpExitReason::InboundOverload))
+    ));
+}
+
 #[tokio::test(start_paused = true)]
 async fn pump_does_not_emit_an_active_ping() {
     let PumpHarness {
@@ -371,6 +499,7 @@ async fn stalled_write_times_out_at_three_hundred_seconds_and_aborts_the_pump()
         incoming: _incoming,
         written: _written,
         dropped,
+        ..
     } = test_connection(true);
     let request_id = Arc::<str>::from("req_test");
 
@@ -413,6 +542,7 @@ async fn dropping_the_connection_aborts_and_drops_the_socket_owner() {
         incoming: _incoming,
         written: _written,
         dropped,
+        ..
     } = test_connection(false);
 
     drop(connection);
@@ -428,6 +558,7 @@ async fn idle_connection_reaches_the_official_limit_without_starting_an_executio
         client: authenticated_client("sk_ws_atomic"),
         trace: Arc::clone(&trace),
         response_headers: Vec::new(),
+        fail_before_first_event: false,
     });
     let app = api_router(execution).await;
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
@@ -539,6 +670,27 @@ async fn active_response_crossing_the_limit_writes_terminal_before_the_limit_err
         }
     }
 
+    for _ in 0..32 {
+        socket
+            .send(ClientMessage::Text(
+                json!({"type": "response.create", "model": "model-a", "input": "queued after limit"})
+                    .to_string()
+                    .into(),
+            ))
+            .await
+            .expect("fill request queue before expiry");
+    }
+    socket
+        .send(ClientMessage::Ping(vec![1].into()))
+        .await
+        .expect("send Ping after queued request");
+    let pong = tokio::time::timeout(Duration::from_secs(1), socket.next())
+        .await
+        .expect("Pong timeout")
+        .expect("connection remains open")
+        .expect("valid frame");
+    assert!(matches!(pong, ClientMessage::Pong(_)));
+
     tokio::time::pause();
     tokio::time::advance(Duration::from_secs(60 * 60)).await;
     tokio::task::yield_now().await;
@@ -567,15 +719,21 @@ async fn active_response_crossing_the_limit_writes_terminal_before_the_limit_err
 
     assert_eq!(event_types, ["response.completed", "error"]);
     assert!(trace.finalized.load(Ordering::Acquire));
+    assert_eq!(trace.starts.load(Ordering::Acquire), 1);
     server.abort();
 }
 
-#[tokio::test]
-async fn second_text_frame_during_an_active_response_keeps_the_single_response_contract() {
+type TestClientSocket = WebSocketStream<MaybeTlsStream<tokio::net::TcpStream>>;
+
+async fn start_active_response() -> (
+    Arc<CrossingLimitTrace>,
+    TestClientSocket,
+    tokio::task::JoinHandle<()>,
+) {
     let trace = Arc::new(CrossingLimitTrace::default());
     let execution = Arc::new(CrossingLimitExecution {
         client: authenticated_client("sk_ws_crossing_limit"),
-        trace,
+        trace: Arc::clone(&trace),
     });
     let app = api_router(execution).await;
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
@@ -607,7 +765,7 @@ async fn second_text_frame_during_an_active_response_keeps_the_single_response_c
         .into(),
     );
     socket
-        .send(response_create.clone())
+        .send(response_create)
         .await
         .expect("send first response.create");
     loop {
@@ -625,19 +783,169 @@ async fn second_text_frame_during_an_active_response_keeps_the_single_response_c
         }
     }
 
+    (trace, socket, server)
+}
+
+#[tokio::test]
+async fn requests_received_during_an_active_response_should_execute_in_order() {
+    let (trace, mut socket, server) = start_active_response().await;
+    for input in ["second", "third"] {
+        socket
+            .send(ClientMessage::Text(
+                json!({"type": "response.create", "model": "model-a", "input": input})
+                    .to_string()
+                    .into(),
+            ))
+            .await
+            .expect("send queued response.create");
+    }
     socket
-        .send(response_create)
+        .send(ClientMessage::Ping(vec![1, 2, 3].into()))
         .await
-        .expect("send overlapping response.create");
+        .expect("send Ping after queued requests");
+    let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
+        .await
+        .expect("Pong timeout while response is active")
+        .expect("WebSocket remains open")
+        .expect("valid WebSocket frame");
+    assert!(matches!(message, ClientMessage::Pong(_)));
+    assert_eq!(trace.starts.load(Ordering::Acquire), 1);
+
+    for completed_count in 1..=3 {
+        trace.release_terminal.notify_one();
+        let expected_types = if completed_count < 3 {
+            vec![
+                "response.completed",
+                "response.metadata",
+                "response.created",
+            ]
+        } else {
+            vec!["response.completed"]
+        };
+        let mut event_types = Vec::new();
+        for _ in &expected_types {
+            let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
+                .await
+                .expect("serial response timeout")
+                .expect("WebSocket remains open")
+                .expect("valid WebSocket frame");
+            let ClientMessage::Text(text) = message else {
+                panic!("unexpected frame during serial execution: {message:?}");
+            };
+            let event: Value = serde_json::from_str(&text).expect("response event JSON");
+            event_types.push(event["type"].as_str().expect("event type").to_owned());
+        }
+        assert_eq!(event_types, expected_types);
+        assert_eq!(
+            trace.starts.load(Ordering::Acquire),
+            (completed_count + 1).min(3)
+        );
+    }
+    assert_eq!(
+        *trace.inputs.lock().expect("input trace lock"),
+        vec![json!("hello"), json!("second"), json!("third")]
+    );
+    assert!(!trace.cancelled.load(Ordering::Acquire));
+    socket.close(None).await.expect("close WebSocket");
+    server.abort();
+}
+
+#[tokio::test]
+async fn queued_unknown_message_should_be_rejected_before_the_next_valid_request() {
+    let (trace, mut socket, server) = start_active_response().await;
+    for payload in [
+        json!({"type": "response.future"}),
+        json!({"type": "response.create", "model": "model-a", "input": "next"}),
+    ] {
+        socket
+            .send(ClientMessage::Text(payload.to_string().into()))
+            .await
+            .expect("queue frame");
+    }
+    trace.release_terminal.notify_one();
+
+    let mut event_types = Vec::new();
+    for _ in 0..4 {
+        let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
+            .await
+            .expect("response timeout")
+            .expect("connection remains open")
+            .expect("valid frame");
+        let ClientMessage::Text(text) = message else {
+            panic!("unexpected frame: {message:?}");
+        };
+        let event: Value = serde_json::from_str(&text).expect("event JSON");
+        event_types.push(event["type"].as_str().expect("event type").to_owned());
+    }
+    assert_eq!(
+        event_types,
+        [
+            "response.completed",
+            "error",
+            "response.metadata",
+            "response.created"
+        ]
+    );
+    assert_eq!(trace.starts.load(Ordering::Acquire), 2);
+    assert_eq!(
+        *trace.inputs.lock().expect("input trace lock"),
+        vec![json!("hello"), json!("next")]
+    );
+    socket.close(None).await.expect("close WebSocket");
+    server.abort();
+}
+
+#[tokio::test]
+async fn queued_binary_frame_should_be_rejected_after_the_active_response_finishes() {
+    let (trace, mut socket, server) = start_active_response().await;
+    socket
+        .send(ClientMessage::Binary(b"{}".to_vec().into()))
+        .await
+        .expect("queue binary frame");
+    trace.release_terminal.notify_one();
+
+    let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
+        .await
+        .expect("terminal timeout")
+        .expect("connection remains open")
+        .expect("valid frame");
+    let ClientMessage::Text(text) = message else {
+        panic!("active response did not finish before binary rejection: {message:?}");
+    };
+    let event: Value = serde_json::from_str(&text).expect("event JSON");
+    assert_eq!(event["type"], "response.completed");
     let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
         .await
         .expect("policy close timeout")
-        .expect("WebSocket returns a policy close")
-        .expect("valid WebSocket frame");
-    let ClientMessage::Close(Some(frame)) = message else {
-        panic!("overlapping response.create must close the connection");
-    };
+        .expect("connection closes")
+        .expect("valid frame");
+    assert!(
+        matches!(message, ClientMessage::Close(Some(frame)) if frame.code == CloseCode::Policy)
+    );
+    assert_eq!(trace.starts.load(Ordering::Acquire), 1);
+    server.abort();
+}
 
-    assert_eq!(frame.code, CloseCode::Policy);
+#[tokio::test]
+async fn client_close_should_cancel_active_execution_without_starting_queued_requests() {
+    let (trace, mut socket, server) = start_active_response().await;
+    socket
+        .send(ClientMessage::Text(
+            json!({"type": "response.create", "model": "model-a", "input": "queued"})
+                .to_string()
+                .into(),
+        ))
+        .await
+        .expect("queue request");
+    socket.close(None).await.expect("close WebSocket");
+
+    tokio::time::timeout(
+        Duration::from_secs(2),
+        trace.cancellation_observed.notified(),
+    )
+    .await
+    .expect("cancel active execution promptly");
+    assert_eq!(trace.starts.load(Ordering::Acquire), 1);
+    assert!(trace.cancelled.load(Ordering::Acquire));
     server.abort();
 }

backend/crates/gateway-api/tests/openai/responses/websocket/mod.rs

@@ -2,7 +2,7 @@ mod connection;
 mod protocol;
 
 use std::sync::{
-    Arc,
+    Arc, Mutex,
     atomic::{AtomicBool, AtomicUsize, Ordering},
 };
 use std::time::{Duration, SystemTime};
@@ -280,17 +280,26 @@ struct AtomicFailureTrace {
     next_calls: AtomicUsize,
     committed: AtomicBool,
     finalized: AtomicBool,
+    downstream_websocket_connection_ids: Mutex<Vec<String>>,
 }
 
 struct AtomicFailureSession {
     trace: Arc<AtomicFailureTrace>,
     response_headers: Vec<ProviderResponseHeader>,
+    fail_before_first_event: bool,
 }
 
 impl ExecutionSession for AtomicFailureSession {
     fn next_event(&mut self) -> BoxFuture<'_, Result<Option<CoordinatedEvent>, EngineError>> {
         Box::pin(async move {
-            match self.trace.next_calls.fetch_add(1, Ordering::AcqRel) {
+            let next_call = self.trace.next_calls.fetch_add(1, Ordering::AcqRel);
+            if self.fail_before_first_event && next_call == 0 {
+                return Err(EngineError::Provider(ProviderError::new(
+                    ProviderErrorKind::Unavailable,
+                    UpstreamSendState::Ambiguous,
+                )));
+            }
+            match next_call {
                 0 => Ok(Some(atomic_failure_batch())),
                 1 => {
                     self.trace.finalized.store(true, Ordering::Release);
@@ -340,6 +349,7 @@ struct AtomicFailureExecution {
     client: AuthenticatedClient,
     trace: Arc<AtomicFailureTrace>,
     response_headers: Vec<ProviderResponseHeader>,
+    fail_before_first_event: bool,
 }
 
 impl ExecutionService for AtomicFailureExecution {
@@ -365,6 +375,19 @@ impl ExecutionService for AtomicFailureExecution {
         request: StartExecution,
     ) -> BoxFuture<'_, Result<StartedExecution, GatewayError>> {
         self.trace.starts.fetch_add(1, Ordering::AcqRel);
+        if let Operation::Generate(generate) = &request.operation
+            && let Some(connection_id) = generate
+                .protocol_payload()
+                .context()
+                .get("downstream_websocket_connection_id")
+                .and_then(Value::as_str)
+        {
+            self.trace
+                .downstream_websocket_connection_ids
+                .lock()
+                .expect("downstream connection trace lock")
+                .push(connection_id.to_owned());
+        }
         Box::pin(async move {
             Ok(StartedExecution {
                 request_id: gateway_core::engine::ModelRequestId::new("req_ws_atomic")
@@ -374,6 +397,7 @@ impl ExecutionService for AtomicFailureExecution {
                 session: Box::new(AtomicFailureSession {
                     trace: Arc::clone(&self.trace),
                     response_headers: self.response_headers.clone(),
+                    fail_before_first_event: self.fail_before_first_event,
                 }),
             })
         })
@@ -439,6 +463,7 @@ async fn websocket_atomic_upstream_failure_batch_should_be_forwarded_once() {
         client: authenticated_client("sk_ws_atomic"),
         trace: Arc::clone(&trace),
         response_headers: Vec::new(),
+        fail_before_first_event: false,
     });
     let app = api_router(execution).await;
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
@@ -510,6 +535,14 @@ async fn websocket_atomic_upstream_failure_batch_should_be_forwarded_once() {
     assert!(trace.committed.load(Ordering::Acquire));
     assert_eq!(trace.next_calls.load(Ordering::Acquire), 2);
     assert!(trace.finalized.load(Ordering::Acquire));
+    {
+        let downstream_connection_ids = trace
+            .downstream_websocket_connection_ids
+            .lock()
+            .expect("downstream connection trace lock");
+        assert_eq!(downstream_connection_ids.len(), 1);
+        assert!(downstream_connection_ids[0].starts_with("ws_"));
+    }
 
     socket.close(None).await.expect("close WebSocket");
     server.abort();
@@ -540,6 +573,7 @@ async fn websocket_response_metadata_should_preserve_ordinary_headers_without_bl
             ProviderResponseHeader::new("content-type", Bytes::from_static(b"application/private")),
             ProviderResponseHeader::new("x-request-id", Bytes::from_static(b"req_upstream")),
         ],
+        fail_before_first_event: false,
     });
     let app = api_router(execution).await;
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
@@ -624,6 +658,91 @@ async fn websocket_response_metadata_should_preserve_ordinary_headers_without_bl
     server.abort();
 }
 
+#[tokio::test]
+async fn websocket_failure_before_first_event_should_return_observed_metadata_then_error() {
+    let trace = Arc::new(AtomicFailureTrace::default());
+    let execution = Arc::new(AtomicFailureExecution {
+        client: authenticated_client("sk_ws_atomic"),
+        trace: Arc::clone(&trace),
+        response_headers: vec![
+            ProviderResponseHeader::new(
+                "x-codex-turn-state",
+                Bytes::from_static(b"turn-state-before-close"),
+            ),
+            ProviderResponseHeader::new(
+                "authorization",
+                Bytes::from_static(b"must-not-cross-boundary"),
+            ),
+        ],
+        fail_before_first_event: true,
+    });
+    let app = api_router(execution).await;
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("bind WebSocket test server");
+    let address = listener.local_addr().expect("test server address");
+    let server = tokio::spawn(async move {
+        axum::serve(listener, app)
+            .await
+            .expect("serve WebSocket test router");
+    });
+    let mut request = format!("ws://{address}/v1/responses")
+        .into_client_request()
+        .expect("WebSocket request");
+    request.headers_mut().insert(
+        AUTHORIZATION,
+        "Bearer sk_ws_atomic".parse().expect("authorization"),
+    );
+    let (mut socket, response) = connect_async(request).await.expect("upgrade WebSocket");
+    assert_eq!(response.status(), StatusCode::SWITCHING_PROTOCOLS);
+    socket
+        .send(ClientMessage::Text(
+            json!({
+                "type": "response.create",
+                "model": "model-a",
+                "input": "hello"
+            })
+            .to_string()
+            .into(),
+        ))
+        .await
+        .expect("send response.create");
+
+    let mut messages = Vec::<Value>::new();
+    while messages.len() < 2 {
+        let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
+            .await
+            .expect("WebSocket response timeout")
+            .expect("WebSocket remains open")
+            .expect("valid WebSocket frame");
+        let ClientMessage::Text(text) = message else {
+            continue;
+        };
+        messages.push(serde_json::from_str(&text).expect("JSON WebSocket event"));
+    }
+
+    assert_eq!(
+        messages
+            .iter()
+            .filter_map(|message| message.get("type").and_then(Value::as_str))
+            .collect::<Vec<_>>(),
+        vec!["response.metadata", "error"]
+    );
+    let headers = messages[0]
+        .get("headers")
+        .and_then(Value::as_object)
+        .expect("response metadata headers");
+    assert_eq!(
+        headers.get("x-codex-turn-state"),
+        Some(&json!("turn-state-before-close"))
+    );
+    assert!(!headers.contains_key("authorization"));
+    assert!(!trace.committed.load(Ordering::Acquire));
+
+    socket.close(None).await.expect("close WebSocket");
+    server.abort();
+}
+
 #[tokio::test]
 async fn websocket_response_create_should_not_have_a_private_16_mib_frame_limit() {
     let trace = Arc::new(AtomicFailureTrace::default());
@@ -631,6 +750,7 @@ async fn websocket_response_create_should_not_have_a_private_16_mib_frame_limit(
         client: authenticated_client("sk_ws_atomic"),
         trace,
         response_headers: Vec::new(),
+        fail_before_first_event: false,
     });
     let app = api_router(execution).await;
     let listener = tokio::net::TcpListener::bind("127.0.0.1:0")

backend/crates/gateway-api/tests/openai/search/mod.rs

@@ -162,6 +162,8 @@ async fn search_route_should_preserve_request_and_success_response_bytes() {
                 .header(AUTHORIZATION, "Bearer sk_search_test")
                 .header("content-type", "application/json")
                 .header("x-codex-turn-metadata", turn_metadata)
+                .header("session-id", "root-session")
+                .header("thread-id", "child-thread")
                 .body(Body::from(request_body.to_vec()))
                 .expect("search request"),
         )
@@ -184,7 +186,10 @@ async fn search_route_should_preserve_request_and_success_response_bytes() {
     assert_eq!(captured[0].endpoint, "/v1/alpha/search");
     assert_eq!(captured[0].transport, ClientTransport::HttpJson);
     assert_eq!(captured[0].body.as_ref(), request_body);
-    assert_eq!(captured[0].context, json!({"turn_metadata": turn_metadata}));
+    assert_eq!(
+        captured[0].context,
+        json!({"turn_metadata": turn_metadata, "session_id": "root-session", "thread_id": "child-thread"})
+    );
     assert_eq!(
         *execution
             .committed_statuses

backend/crates/gateway-core/Cargo.toml

@@ -13,11 +13,13 @@ path = "tests/main.rs"
 
 [dependencies]
 async-trait = "0.1.89"
+base64 = "0.23.0"
 bytes = "1.11.1"
 futures = "0.3.32"
 futures-timer = "3.0.3"
 serde = { version = "1.0.228", features = ["derive"] }
 serde_json = { version = "1.0.150", features = ["arbitrary_precision", "preserve_order"] }
+sha2 = "0.11.0"
 semver = "1.0.27"
 thiserror = "2.0.18"
 tracing = "0.1.44"

backend/crates/gateway-core/src/account/mod.rs

@@ -2,6 +2,7 @@
 
 mod error;
 mod model;
+pub mod scope;
 mod selection;
 mod store;
 

backend/crates/gateway-core/src/account/model.rs

@@ -6,8 +6,8 @@ use std::time::{Duration, SystemTime};
 
 use serde_json::{Map, Value};
 
-use crate::error::{IdentifierError, validate_text};
-use crate::routing::ProviderKind;
+use crate::identity::ProviderKind;
+use crate::validation::{IdentifierError, validate_text};
 
 use super::CredentialError;
 

backend/crates/gateway-core/src/account/scope.rs

@@ -0,0 +1,291 @@
+//! 请求认证时冻结的账号范围与目录，不依赖路由选择器。
+
+use super::ProviderAccountId;
+use crate::identity::ProviderKind;
+use crate::validation::{IdentifierError, RoutingError};
+use std::{
+    collections::{BTreeMap, BTreeSet},
+    fmt,
+    sync::Arc,
+};
+
+/// `account_groups.id` 的核心值对象。
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct AccountGroupId(String);
+
+impl AccountGroupId {
+    /// 校验并创建账号分组 ID。
+    pub fn new(value: impl Into<String>) -> Result<Self, IdentifierError> {
+        let value = value.into();
+        let Some(suffix) = value.strip_prefix("grp_") else {
+            return Err(IdentifierError::MissingPrefix { expected: "grp_" });
+        };
+        if suffix.len() != 32
+            || !suffix
+                .bytes()
+                .all(|byte| byte.is_ascii_digit() || (b'a'..=b'f').contains(&byte))
+        {
+            return Err(IdentifierError::InvalidFormat);
+        }
+        Ok(Self(value))
+    }
+
+    #[must_use]
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+}
+
+impl fmt::Display for AccountGroupId {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter.write_str(&self.0)
+    }
+}
+
+/// 快照中一个账号的 Provider 与分组归属。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct RuntimeAccount {
+    provider_kind: ProviderKind,
+    group_ids: Arc<BTreeSet<AccountGroupId>>,
+}
+
+impl RuntimeAccount {
+    #[must_use]
+    pub fn new(provider_kind: ProviderKind, group_ids: BTreeSet<AccountGroupId>) -> Self {
+        Self {
+            provider_kind,
+            group_ids: Arc::new(group_ids),
+        }
+    }
+
+    #[must_use]
+    pub const fn provider_kind(&self) -> &ProviderKind {
+        &self.provider_kind
+    }
+
+    #[must_use]
+    pub fn group_ids(&self) -> &BTreeSet<AccountGroupId> {
+        &self.group_ids
+    }
+}
+
+/// 全快照共享的账号、Provider 与分组反向索引。
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct RuntimeAccountDirectory {
+    accounts: BTreeMap<ProviderAccountId, RuntimeAccount>,
+    providers_with_accounts: BTreeSet<ProviderKind>,
+    providers_by_group: BTreeMap<AccountGroupId, BTreeSet<ProviderKind>>,
+}
+
+impl RuntimeAccountDirectory {
+    #[must_use]
+    pub fn new(accounts: BTreeMap<ProviderAccountId, RuntimeAccount>) -> Self {
+        let mut providers_with_accounts = BTreeSet::new();
+        let mut providers_by_group = BTreeMap::<AccountGroupId, BTreeSet<ProviderKind>>::new();
+        for account in accounts.values() {
+            providers_with_accounts.insert(account.provider_kind.clone());
+            for group_id in account.group_ids.iter() {
+                providers_by_group
+                    .entry(group_id.clone())
+                    .or_default()
+                    .insert(account.provider_kind.clone());
+            }
+        }
+        Self {
+            accounts,
+            providers_with_accounts,
+            providers_by_group,
+        }
+    }
+
+    #[must_use]
+    pub fn account(&self, account_id: &ProviderAccountId) -> Option<&RuntimeAccount> {
+        self.accounts.get(account_id)
+    }
+
+    #[must_use]
+    pub fn providers_with_accounts(&self) -> &BTreeSet<ProviderKind> {
+        &self.providers_with_accounts
+    }
+
+    #[must_use]
+    pub fn providers_for_groups<'a>(
+        &self,
+        group_ids: impl IntoIterator<Item = &'a AccountGroupId>,
+    ) -> BTreeSet<ProviderKind> {
+        group_ids
+            .into_iter()
+            .filter_map(|group_id| self.providers_by_group.get(group_id))
+            .flatten()
+            .cloned()
+            .collect()
+    }
+}
+
+/// 历史请求保存的账号范围种类。
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum AccountRoutingScopeKind {
+    All,
+    Groups,
+}
+
+impl AccountRoutingScopeKind {
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Self::All => "all",
+            Self::Groups => "groups",
+        }
+    }
+}
+
+/// 请求开始时冻结的分组 ID 与名称。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct RoutingGroupSnapshot {
+    id: AccountGroupId,
+    name: String,
+}
+
+impl RoutingGroupSnapshot {
+    #[must_use]
+    pub fn new(id: AccountGroupId, name: String) -> Self {
+        Self { id, name }
+    }
+
+    #[must_use]
+    pub const fn id(&self) -> &AccountGroupId {
+        &self.id
+    }
+
+    #[must_use]
+    pub fn name(&self) -> &str {
+        &self.name
+    }
+}
+
+/// 请求历史所需的完整、稳定账号范围快照。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct AccountRoutingSnapshot {
+    kind: AccountRoutingScopeKind,
+    groups: Arc<[RoutingGroupSnapshot]>,
+}
+
+impl AccountRoutingSnapshot {
+    #[must_use]
+    pub fn all() -> Self {
+        Self {
+            kind: AccountRoutingScopeKind::All,
+            groups: Arc::from([]),
+        }
+    }
+
+    #[must_use]
+    pub fn groups(groups: Vec<RoutingGroupSnapshot>) -> Self {
+        Self {
+            kind: AccountRoutingScopeKind::Groups,
+            groups: Arc::from(groups),
+        }
+    }
+
+    #[must_use]
+    pub const fn kind(&self) -> AccountRoutingScopeKind {
+        self.kind
+    }
+
+    #[must_use]
+    pub fn groups_snapshot(&self) -> &[RoutingGroupSnapshot] {
+        &self.groups
+    }
+}
+
+/// Key 持久 binding 编译出的账号权限。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum ClientRoutingScope {
+    AllAccounts,
+    Restricted {
+        bound_groups: Arc<[RoutingGroupSnapshot]>,
+        enabled_group_ids: Arc<BTreeSet<AccountGroupId>>,
+        provider_kinds: Arc<BTreeSet<ProviderKind>>,
+    },
+}
+
+impl ClientRoutingScope {
+    #[must_use]
+    pub fn all_accounts() -> Self {
+        Self::AllAccounts
+    }
+
+    pub fn restricted(
+        bound_groups: Vec<RoutingGroupSnapshot>,
+        enabled_group_ids: BTreeSet<AccountGroupId>,
+        provider_kinds: BTreeSet<ProviderKind>,
+    ) -> Result<Self, RoutingError> {
+        if bound_groups.is_empty() {
+            return Err(RoutingError::InvalidAccountScope);
+        }
+        Ok(Self::Restricted {
+            bound_groups: Arc::from(bound_groups),
+            enabled_group_ids: Arc::new(enabled_group_ids),
+            provider_kinds: Arc::new(provider_kinds),
+        })
+    }
+}
+
+/// 一次认证随 RuntimeSnapshot 冻结的账号目录与 Key 权限。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct FrozenAccountScope {
+    directory: Arc<RuntimeAccountDirectory>,
+    client_scope: ClientRoutingScope,
+}
+
+impl FrozenAccountScope {
+    #[must_use]
+    pub const fn new(
+        directory: Arc<RuntimeAccountDirectory>,
+        client_scope: ClientRoutingScope,
+    ) -> Self {
+        Self {
+            directory,
+            client_scope,
+        }
+    }
+
+    #[must_use]
+    pub fn allows(&self, account_id: &ProviderAccountId) -> bool {
+        let Some(account) = self.directory.account(account_id) else {
+            return false;
+        };
+        match &self.client_scope {
+            ClientRoutingScope::AllAccounts => true,
+            ClientRoutingScope::Restricted {
+                enabled_group_ids, ..
+            } => account
+                .group_ids()
+                .iter()
+                .any(|group_id| enabled_group_ids.contains(group_id)),
+        }
+    }
+
+    #[must_use]
+    pub fn provider_kinds(&self) -> &BTreeSet<ProviderKind> {
+        match &self.client_scope {
+            ClientRoutingScope::AllAccounts => self.directory.providers_with_accounts(),
+            ClientRoutingScope::Restricted { provider_kinds, .. } => provider_kinds,
+        }
+    }
+
+    #[must_use]
+    pub fn routing_snapshot(&self) -> AccountRoutingSnapshot {
+        match &self.client_scope {
+            ClientRoutingScope::AllAccounts => AccountRoutingSnapshot::all(),
+            ClientRoutingScope::Restricted { bound_groups, .. } => {
+                AccountRoutingSnapshot::groups(bound_groups.to_vec())
+            }
+        }
+    }
+
+    #[must_use]
+    pub const fn directory(&self) -> &Arc<RuntimeAccountDirectory> {
+        &self.directory
+    }
+}

backend/crates/gateway-core/src/account/selection.rs

@@ -7,7 +7,7 @@ use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::{Mutex, RwLock};
 use std::time::{Duration, Instant, SystemTime};
 
-use crate::routing::ProviderKind;
+use crate::identity::ProviderKind;
 
 use super::{AccountStatus, ProviderAccount, ProviderAccountId};
 
@@ -87,6 +87,7 @@ pub struct AccountRuntimeSignals {
     pub in_flight: u32,
     pub last_started_at: Option<SystemTime>,
     pub quota_reset_at: Option<SystemTime>,
+    /// Provider 归一化的剩余额度基点：0 耗尽，10_000 全部可用。
     pub quota_remaining_rank: Option<u64>,
     pub rate_limited_until: Option<SystemTime>,
     pub failure_rate_basis_points: Option<u16>,
@@ -323,6 +324,7 @@ pub struct AccountQuotaSignals {
 }
 
 impl AccountQuotaSignals {
+    /// `remaining_rank` 使用剩余额度基点，范围为 0..=10_000。
     #[must_use]
     pub const fn new(reset_at: Option<SystemTime>, remaining_rank: Option<u64>) -> Self {
         Self {
@@ -386,9 +388,11 @@ pub struct AccountSelectionContext {
     pub now: SystemTime,
     pub excluded_accounts: BTreeSet<ProviderAccountId>,
     pub preferred_account: Option<ProviderAccountId>,
+    /// 优先账号可调度时，是否优先于其它账号的持久权重。
+    pub preferred_account_overrides_weight: bool,
     pub round_robin_cursor: u64,
     pub eligibility: AccountEligibilityPolicy,
-    pub account_scope: Option<std::sync::Arc<crate::routing::FrozenAccountScope>>,
+    pub account_scope: Option<std::sync::Arc<crate::account::scope::FrozenAccountScope>>,
 }
 
 /// 选择账号时是否执行本地调度资格投影。
@@ -516,7 +520,9 @@ impl AccountSelector {
             {
                 Some(candidate) => match self.scheduling_blocker(candidate, context) {
                     Some(blocker) => PreferredAccountSelection::Blocked(blocker),
-                    None if candidate.account.weight() < highest_weight => {
+                    None if !context.preferred_account_overrides_weight
+                        && candidate.account.weight() < highest_weight =>
+                    {
                         PreferredAccountSelection::Blocked(AccountSchedulingBlocker::LowerWeight)
                     }
                     None => {
@@ -636,6 +642,10 @@ const SMART_LOAD_WEIGHT: f64 = 1.0;
 const SMART_QUOTA_WEIGHT: f64 = 0.8;
 const SMART_FAILURE_WEIGHT: f64 = 1.0;
 const SMART_LATENCY_WEIGHT: f64 = 0.5;
+// 容忍 5 个百分点的单项负载/失败率差异，避免微小信号波动独占新会话。
+const SMART_SCORE_TOLERANCE: f64 = 0.05;
+// 首输出 10 秒时延迟得分减半；固定尺度不随其他候选账号变化。
+const SMART_LATENCY_HALF_SCORE_MS: f64 = 10_000.0;
 
 fn capacity_utilization(candidate: &AccountCandidate, default_concurrency: NonZeroU32) -> f64 {
     f64::from(candidate.signals.in_flight)
@@ -647,94 +657,32 @@ fn capacity_utilization(candidate: &AccountCandidate, default_concurrency: NonZe
         )
 }
 
-struct SmartNormalization {
-    max_capacity_utilization: f64,
-    min_quota: Option<u64>,
-    max_quota: Option<u64>,
-    min_latency_ms: Option<u64>,
-    max_latency_ms: Option<u64>,
-}
-
-impl SmartNormalization {
-    fn from_candidates(candidates: &[&AccountCandidate], default_concurrency: NonZeroU32) -> Self {
-        let max_capacity_utilization = candidates
-            .iter()
-            .map(|candidate| capacity_utilization(candidate, default_concurrency))
-            .reduce(f64::max)
-            .unwrap_or_default();
-        let min_quota = candidates
-            .iter()
-            .filter_map(|candidate| candidate.signals.quota_remaining_rank)
-            .min();
-        let max_quota = candidates
-            .iter()
-            .filter_map(|candidate| candidate.signals.quota_remaining_rank)
-            .max();
-        let min_latency_ms = candidates
-            .iter()
-            .filter_map(|candidate| candidate.signals.first_output_latency_ms)
-            .filter(|latency| *latency > 0)
-            .min();
-        let max_latency_ms = candidates
-            .iter()
-            .filter_map(|candidate| candidate.signals.first_output_latency_ms)
-            .filter(|latency| *latency > 0)
-            .max();
-        Self {
-            max_capacity_utilization,
-            min_quota,
-            max_quota,
-            min_latency_ms,
-            max_latency_ms,
-        }
-    }
-}
-
 fn select_smart_candidate<'a>(
     candidates: &[&'a AccountCandidate],
     default_concurrency: NonZeroU32,
     cursor: u64,
 ) -> Option<&'a AccountCandidate> {
-    let normalization = SmartNormalization::from_candidates(candidates, default_concurrency);
     let mut ranked = candidates
         .iter()
-        .map(|candidate| {
-            (
-                *candidate,
-                smart_score(candidate, default_concurrency, &normalization),
-            )
-        })
+        .map(|candidate| (*candidate, smart_score(candidate, default_concurrency)))
         .collect::<Vec<_>>();
-    ranked.sort_by(|(left, left_score), (right, right_score)| {
-        right_score
-            .total_cmp(left_score)
-            .then_with(|| left.account.id().cmp(right.account.id()))
-    });
-    let best_score = ranked.first()?.1.to_bits();
-    let tied = ranked
+    let best_score = ranked
         .iter()
-        .take_while(|(_, score)| score.to_bits() == best_score)
-        .count();
-    let index = cursor as usize % tied;
+        .map(|(_, score)| *score)
+        .max_by(f64::total_cmp)?;
+    ranked.retain(|(_, score)| best_score - score <= SMART_SCORE_TOLERANCE);
+    // 轮换顺序保持稳定，避免分数轻微交错与 cursor 同步后仍反复命中同一账号。
+    ranked.sort_unstable_by(|(left, _), (right, _)| left.account.id().cmp(right.account.id()));
+    let index = (cursor % ranked.len() as u64) as usize;
     Some(ranked[index].0)
 }
 
-fn smart_score(
-    candidate: &AccountCandidate,
-    default_concurrency: NonZeroU32,
-    normalization: &SmartNormalization,
-) -> f64 {
-    let load = lower_ratio_is_better(
-        capacity_utilization(candidate, default_concurrency),
-        normalization.max_capacity_utilization,
-    );
-    let quota = candidate.signals.quota_remaining_rank.map_or(0.5, |quota| {
-        higher_is_better(
-            quota,
-            normalization.min_quota.unwrap_or(quota),
-            normalization.max_quota.unwrap_or(quota),
-        )
-    });
+fn smart_score(candidate: &AccountCandidate, default_concurrency: NonZeroU32) -> f64 {
+    let load = 1.0 - capacity_utilization(candidate, default_concurrency).clamp(0.0, 1.0);
+    let quota = candidate
+        .signals
+        .quota_remaining_rank
+        .map_or(0.5, |quota| quota.min(10_000) as f64 / 10_000.0);
     let failure = 1.0
         - f64::from(
             candidate
@@ -748,11 +696,7 @@ fn smart_score(
         .first_output_latency_ms
         .filter(|latency| *latency > 0)
         .map_or(1.0, |latency| {
-            lower_is_better(
-                latency,
-                normalization.min_latency_ms.unwrap_or(latency),
-                normalization.max_latency_ms.unwrap_or(latency),
-            )
+            SMART_LATENCY_HALF_SCORE_MS / (SMART_LATENCY_HALF_SCORE_MS + latency as f64)
         });
 
     SMART_LOAD_WEIGHT * load
@@ -760,24 +704,3 @@ fn smart_score(
         + SMART_FAILURE_WEIGHT * failure
         + SMART_LATENCY_WEIGHT * latency
 }
-
-fn lower_ratio_is_better(value: f64, maximum: f64) -> f64 {
-    if maximum <= 0.0 {
-        return 1.0;
-    }
-    1.0 - (value / maximum).clamp(0.0, 1.0)
-}
-
-fn lower_is_better(value: u64, minimum: u64, maximum: u64) -> f64 {
-    if maximum <= minimum {
-        return 1.0;
-    }
-    1.0 - (value.saturating_sub(minimum) as f64 / (maximum - minimum) as f64).clamp(0.0, 1.0)
-}
-
-fn higher_is_better(value: u64, minimum: u64, maximum: u64) -> f64 {
-    if maximum <= minimum {
-        return 1.0;
-    }
-    (value.saturating_sub(minimum) as f64 / (maximum - minimum) as f64).clamp(0.0, 1.0)
-}

backend/crates/gateway-core/src/account/store.rs

@@ -3,7 +3,7 @@
 use async_trait::async_trait;
 
 use crate::error::StoreError;
-use crate::routing::ProviderKind;
+use crate::identity::ProviderKind;
 
 use super::{
     AccountStateChange, CredentialCasOutcome, CredentialCasUpdate, CredentialRevision,

backend/crates/gateway-core/src/diagnostics/capture.rs

@@ -0,0 +1,157 @@
+//! 诊断内容的统一筛选边界。未知字符串只保留长度和摘要。
+
+use serde_json::{Map, Value, json};
+use sha2::{Digest, Sha256};
+
+/// 原始字节的长度与 SHA-256，不复制或保留正文。
+#[must_use]
+pub fn body_fingerprint(bytes: &[u8]) -> Value {
+    let digest = Sha256::digest(bytes);
+    let hex: String = digest
+        .iter()
+        .flat_map(|byte| {
+            let alphabet = b"0123456789abcdef";
+            [
+                char::from(alphabet[usize::from(byte >> 4)]),
+                char::from(alphabet[usize::from(byte & 15)]),
+            ]
+        })
+        .collect();
+    json!({"bytes": bytes.len(), "sha256": hex})
+}
+
+/// 保留可诊断的协议结构；凭据永不进入普通日志或数据库。
+#[must_use]
+pub fn diagnostic_json(value: &Value) -> Value {
+    capture(value, "", 0, &mut 96)
+}
+
+fn capture(value: &Value, key: &str, depth: usize, budget: &mut usize) -> Value {
+    if *budget == 0 || depth > 6 {
+        return json!({"omitted": true});
+    }
+    *budget -= 1;
+    let key = key.to_ascii_lowercase();
+    if secret_key(&key) {
+        return Value::String("<redacted>".to_owned());
+    }
+    match value {
+        Value::Object(object) => {
+            let mut result = Map::new();
+            for (name, value) in object.iter().take(48) {
+                if *budget == 0 {
+                    break;
+                }
+                result.insert(bounded(name, 96), capture(value, name, depth + 1, budget));
+            }
+            if result.len() < object.len() {
+                result.insert(
+                    "_omittedFields".to_owned(),
+                    json!(object.len() - result.len()),
+                );
+            }
+            Value::Object(result)
+        }
+        Value::Array(items) => json!({
+            "length": items.len(),
+            "sample": items.iter().take(4)
+                .map(|item| capture(item, key.as_str(), depth + 1, budget))
+                .collect::<Vec<_>>()
+        }),
+        Value::String(text)
+            if safe_header(&key) || safe_value_key(&key) && protocol_label(text) =>
+        {
+            Value::String(bounded(text, 256))
+        }
+        Value::String(text) => body_fingerprint(text.as_bytes()),
+        other => other.clone(),
+    }
+}
+
+/// 请求/响应头保持多值；安全 trace ID 原样保存，未知值只保存摘要。
+#[must_use]
+pub fn diagnostic_headers<'a>(headers: impl IntoIterator<Item = (&'a str, &'a str)>) -> Value {
+    let mut result = Map::new();
+    for (name, value) in headers.into_iter().take(64) {
+        let name = bounded(&name.to_ascii_lowercase(), 96);
+        let value = if secret_key(&name) {
+            Value::String("<redacted>".to_owned())
+        } else if safe_header(&name) {
+            Value::String(bounded(value, 256))
+        } else {
+            body_fingerprint(value.as_bytes())
+        };
+        let values = result
+            .entry(name)
+            .or_insert_with(|| Value::Array(Vec::new()));
+        if let Value::Array(values) = values {
+            values.push(value);
+        }
+    }
+    Value::Object(result)
+}
+
+fn secret_key(key: &str) -> bool {
+    key.contains("authorization")
+        || key.contains("cookie")
+        || key.contains("api_key")
+        || key.contains("api-key")
+        || key.contains("token") && !key.ends_with("tokens")
+        || key.contains("password")
+        || key.contains("secret")
+        || key.contains("attestation")
+        || matches!(key, "x-oai-is" | "x-oai-is-update")
+}
+
+fn safe_header(key: &str) -> bool {
+    matches!(
+        key,
+        "x-oai-request-id"
+            | "x-client-request-id"
+            | "x-request-id"
+            | "request-id"
+            | "openai-request-id"
+            | "x-openai-request-id"
+            | "cf-ray"
+            | "traceparent"
+            | "tracestate"
+            | "content-type"
+            | "content-length"
+            | "content-encoding"
+            | "retry-after"
+            | "openai-processing-ms"
+            | "x-processing-ms"
+            | "date"
+            | "server"
+            | "sec-websocket-extensions"
+            | "x-codex-allowed"
+            | "x-codex-limit-reached"
+            | "x-codex-active-limit"
+            | "x-models-etag"
+    )
+}
+
+fn safe_value_key(key: &str) -> bool {
+    safe_header(key)
+        || matches!(
+            key,
+            "type" | "event" | "status" | "code" | "model" | "role" | "object" | "service_tier"
+        )
+}
+
+pub(super) fn bounded(value: &str, limit: usize) -> String {
+    value
+        .chars()
+        .take(limit)
+        .map(|c| if c.is_control() { ' ' } else { c })
+        .collect()
+}
+
+/// Event names/codes are protocol labels, never arbitrary upstream text.
+pub(super) fn protocol_label(value: &str) -> bool {
+    !value.is_empty()
+        && value.len() <= 128
+        && value
+            .bytes()
+            .all(|byte| byte.is_ascii_alphanumeric() || matches!(byte, b'.' | b'_' | b'-' | b':'))
+}

backend/crates/gateway-core/src/diagnostics/mod.rs

@@ -0,0 +1,9 @@
+//! 有界、请求局部的诊断证据；不参与路由、重试或客户端协议决策。
+
+mod capture;
+mod stream;
+mod trace;
+
+pub use capture::{body_fingerprint, diagnostic_headers, diagnostic_json};
+pub use stream::{StreamCapture, StreamFormat};
+pub use trace::TraceContext;

backend/crates/gateway-core/src/diagnostics/stream.rs

@@ -0,0 +1,165 @@
+//! 仅用于观测的有界流分帧；超过限额时明确记缺口，不影响实际协议解析。
+
+use serde_json::json;
+
+use super::TraceContext;
+
+const MAX_FRAME_BYTES: usize = 128 * 1024;
+
+#[derive(Clone, Copy)]
+pub enum StreamFormat {
+    Sse,
+    JsonLines,
+}
+
+/// 同一 exchange 持有一个 observer，跨 HTTP chunk 拼接完整事件。
+pub struct StreamCapture {
+    trace: TraceContext,
+    format: StreamFormat,
+    line: Vec<u8>,
+    data: Vec<u8>,
+    event_name: Option<String>,
+    previous_cr: bool,
+    first_line: bool,
+    skipping_line: bool,
+    skipping_event: bool,
+    finished: bool,
+}
+
+impl StreamCapture {
+    #[must_use]
+    pub fn new(trace: TraceContext, format: StreamFormat) -> Self {
+        Self {
+            trace,
+            format,
+            line: Vec::new(),
+            data: Vec::new(),
+            event_name: None,
+            previous_cr: false,
+            first_line: true,
+            skipping_line: false,
+            skipping_event: false,
+            finished: false,
+        }
+    }
+
+    pub fn push(&mut self, bytes: &[u8]) {
+        self.trace.dump("upstream.chunk", bytes);
+        for byte in bytes {
+            if self.previous_cr && *byte == b'\n' {
+                self.previous_cr = false;
+                continue;
+            }
+            self.previous_cr = *byte == b'\r';
+            if matches!(*byte, b'\n' | b'\r') {
+                if !self.skipping_line {
+                    self.line();
+                }
+                self.line.clear();
+                self.skipping_line = false;
+            } else if !self.skipping_line {
+                if self.line.len() < MAX_FRAME_BYTES {
+                    self.line.push(*byte);
+                } else {
+                    self.trace.record(
+                        "capture.gap",
+                        json!({"reason": "line_size_limit", "limitBytes": MAX_FRAME_BYTES}),
+                    );
+                    self.line.clear();
+                    self.skipping_line = true;
+                    self.skipping_event = matches!(self.format, StreamFormat::Sse);
+                    self.data.clear();
+                }
+            }
+        }
+    }
+
+    pub fn finish(&mut self) {
+        if self.finished {
+            return;
+        }
+        let partial_frame = !self.line.is_empty()
+            || !self.data.is_empty()
+            || self.skipping_event
+            || self.skipping_line;
+        if !self.line.is_empty() && !self.skipping_line {
+            self.line();
+        }
+        if !self.data.is_empty() && !self.skipping_event {
+            self.trace.capture_named_event(
+                "upstream.event",
+                &self.data,
+                self.event_name.as_deref(),
+            );
+        }
+        self.trace
+            .record("upstream.eof", json!({"partialFrame": partial_frame}));
+        self.finished = true;
+    }
+
+    fn line(&mut self) {
+        let line = if self.first_line {
+            self.first_line = false;
+            self.line
+                .strip_prefix(b"\xef\xbb\xbf")
+                .unwrap_or(&self.line)
+        } else {
+            &self.line
+        };
+        if matches!(self.format, StreamFormat::JsonLines) {
+            if !line.is_empty() {
+                self.trace.capture_event("upstream.event", line);
+            }
+            return;
+        }
+        if line.is_empty() {
+            if !self.data.is_empty() && !self.skipping_event {
+                self.trace.capture_named_event(
+                    "upstream.event",
+                    &self.data,
+                    self.event_name.as_deref(),
+                );
+            }
+            self.data.clear();
+            self.event_name = None;
+            self.skipping_event = false;
+        } else if let Some(name) = line.strip_prefix(b"event:") {
+            self.event_name = Some(super::capture::bounded(
+                String::from_utf8_lossy(name).trim(),
+                128,
+            ));
+        } else if let Some(data) = line.strip_prefix(b"data:") {
+            let data = data.strip_prefix(b" ").unwrap_or(data);
+            if self.skipping_event {
+                return;
+            }
+            if self.data.len() + data.len() + 1 > MAX_FRAME_BYTES {
+                self.trace.record(
+                    "capture.gap",
+                    json!({"reason": "event_size_limit", "limitBytes": MAX_FRAME_BYTES}),
+                );
+                self.data.clear();
+                self.skipping_event = true;
+            } else {
+                if !self.data.is_empty() {
+                    self.data.push(b'\n');
+                }
+                self.data.extend_from_slice(data);
+            }
+        }
+    }
+}
+
+impl Drop for StreamCapture {
+    fn drop(&mut self) {
+        if !self.finished {
+            self.trace.record(
+                "upstream.reader.dropped",
+                json!({
+                    "partialBytes": self.line.len() + self.data.len(),
+                    "captureIncomplete": self.skipping_event || self.skipping_line,
+                }),
+            );
+        }
+    }
+}

backend/crates/gateway-core/src/diagnostics/trace.rs

@@ -0,0 +1,338 @@
+//! 首部与最近事件同时保留，流式 delta 合并，所有淘汰均有计数。
+
+use std::{
+    collections::VecDeque,
+    sync::{Arc, Mutex},
+    time::{Instant, SystemTime, UNIX_EPOCH},
+};
+
+use base64::{Engine as _, engine::general_purpose::STANDARD};
+use serde::Serialize;
+use serde_json::{Value, json};
+
+use super::capture::bounded;
+use super::{body_fingerprint, diagnostic_headers, diagnostic_json};
+
+const MAX_EVENTS: usize = 128;
+const HEAD_EVENTS: usize = 8;
+const MAX_DATA_BYTES: usize = 4096;
+const MAX_BUFFER_BYTES: usize = 64 * 1024;
+
+/// 显式传给异步任务的关联上下文；默认值禁用捕获，避免全局请求映射。
+#[derive(Clone, Default)]
+pub struct TraceContext {
+    state: Option<Arc<Mutex<TraceState>>>,
+    attempt_index: u32,
+    exchange_id: Option<u64>,
+}
+
+impl std::fmt::Debug for TraceContext {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("TraceContext")
+            .field("attempt_index", &self.attempt_index)
+            .field("exchange_id", &self.exchange_id)
+            .finish_non_exhaustive()
+    }
+}
+
+struct TraceState {
+    request_id: String,
+    started_at_ms: u128,
+    started: Instant,
+    sequence: u64,
+    exchanges: u64,
+    wire_frames: u64,
+    wire_bytes: u64,
+    dropped_events: u64,
+    bytes: usize,
+    events: VecDeque<TraceEvent>,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct TraceEvent {
+    sequence: u64,
+    last_sequence: u64,
+    elapsed_ms: u64,
+    last_elapsed_ms: u64,
+    attempt_index: u32,
+    exchange_id: Option<u64>,
+    stage: &'static str,
+    count: u64,
+    data: Value,
+    #[serde(skip)]
+    bytes: usize,
+}
+
+impl TraceContext {
+    /// 一个模型请求只创建一次，所有 attempt 共享同一有界时间线。
+    #[must_use]
+    pub fn new(request_id: &str) -> Self {
+        Self {
+            state: Some(Arc::new(Mutex::new(TraceState {
+                request_id: bounded(request_id, 128),
+                started_at_ms: SystemTime::now()
+                    .duration_since(UNIX_EPOCH)
+                    .unwrap_or_default()
+                    .as_millis(),
+                started: Instant::now(),
+                sequence: 0,
+                exchanges: 0,
+                wire_frames: 0,
+                wire_bytes: 0,
+                dropped_events: 0,
+                bytes: 0,
+                events: VecDeque::new(),
+            }))),
+            ..Self::default()
+        }
+    }
+
+    #[must_use]
+    pub fn attempt(&self, attempt_index: u32) -> Self {
+        Self {
+            state: self.state.clone(),
+            attempt_index,
+            exchange_id: None,
+        }
+    }
+
+    /// 一个 attempt 中的每次实际 transport exchange 也有独立编号。
+    #[must_use]
+    pub fn exchange(&self, transport: &'static str) -> Self {
+        let mut context = self.clone();
+        if let Some(state) = &self.state {
+            let mut state = state
+                .lock()
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+            state.exchanges += 1;
+            context.exchange_id = Some(state.exchanges);
+        }
+        context.record("upstream.exchange.started", json!({"transport": transport}));
+        context
+    }
+
+    /// details 只允许调用方构造的诊断事实；不接收原始请求/响应正文。
+    pub fn record(&self, stage: &'static str, data: Value) {
+        self.push(stage, data, false);
+    }
+
+    /// 同一个头部边界只采集一次：默认脱敏，显式开启 dump 时另存完整字节。
+    /// facts 必须为调用方构造的安全对象；原始头值只进入独立 dump。
+    pub fn headers<'a>(
+        &self,
+        stage: &'static str,
+        mut facts: Value,
+        headers: impl IntoIterator<Item = (&'a str, &'a [u8])>,
+    ) {
+        if self.state.is_none() {
+            return;
+        }
+        let Some(facts_object) = facts.as_object_mut() else {
+            return;
+        };
+        let headers: Vec<_> = headers.into_iter().collect();
+        let summary = diagnostic_headers(
+            headers
+                .iter()
+                .map(|(name, value)| (*name, std::str::from_utf8(value).unwrap_or("<binary>"))),
+        );
+        facts_object.insert("headers".to_owned(), summary);
+        self.record(stage, facts);
+        if tracing::enabled!(target: "request_dump", tracing::Level::INFO) {
+            let values: Vec<_> = headers
+                .iter()
+                .map(|(name, value)| json!({"name": name, "valueBase64": STANDARD.encode(value)}))
+                .collect();
+            if let Ok(bytes) = serde_json::to_vec(&json!({"headers": values})) {
+                self.dump(stage, &bytes);
+            }
+        }
+    }
+
+    /// 入站 JSON 先记录再解析业务语义，包括未知事件和 metadata。
+    /// 原文仅在现有 request_dump 开关打开时写入其独立文件。
+    pub fn capture(&self, stage: &'static str, bytes: &[u8]) {
+        self.dump(stage, bytes);
+        self.capture_event(stage, bytes);
+    }
+
+    /// StreamCapture 已转储原始 chunk 时，只提取完整事件的安全摘要。
+    pub fn capture_event(&self, stage: &'static str, bytes: &[u8]) {
+        self.capture_named_event(stage, bytes, None);
+    }
+
+    pub fn capture_named_event(&self, stage: &'static str, bytes: &[u8], name: Option<&str>) {
+        if self.state.is_none() {
+            return;
+        }
+        let fingerprint = body_fingerprint(bytes);
+        if bytes.len() > 128 * 1024 {
+            self.record("capture.gap", json!({"reason": "json_inspection_size_limit", "stage": stage, "body": fingerprint}));
+            return;
+        }
+        let parsed = serde_json::from_slice::<Value>(bytes).ok();
+        let event_type = parsed
+            .as_ref()
+            .and_then(|v| v.get("type"))
+            .and_then(Value::as_str)
+            .or(name)
+            .filter(|kind| super::capture::protocol_label(kind));
+        let delta = event_type.is_some_and(|kind| kind.ends_with(".delta"));
+        let data = json!({
+            "body": fingerprint,
+            "eventType": event_type.map(|kind| bounded(kind, 128)),
+            "jsonValid": parsed.is_some(),
+            "metadata": parsed.as_ref().filter(|_| !delta).map(diagnostic_json),
+        });
+        self.push(stage, data, delta);
+    }
+
+    /// 只记录业务解析边界；metadata 已由 transport 在解析前捕获。
+    pub fn wire_event(&self, protocol: &str, event_type: Option<&str>) {
+        let event_type = event_type.filter(|kind| super::capture::protocol_label(kind));
+        let delta = event_type.is_some_and(|kind| kind.ends_with(".delta"));
+        self.push(
+            "provider.event",
+            json!({
+                "protocol": protocol,
+                "eventType": event_type.map(|kind| bounded(kind, 128)),
+            }),
+            delta,
+        );
+    }
+
+    /// 完整报文捕获与摘要使用相同 request / attempt / exchange ID。
+    pub fn dump(&self, stage: &'static str, bytes: &[u8]) {
+        if !tracing::enabled!(target: "request_dump", tracing::Level::INFO) {
+            return;
+        }
+        if let Some(state) = &self.state {
+            let mut state = state
+                .lock()
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+            state.wire_frames += 1;
+            state.wire_bytes = state.wire_bytes.saturating_add(bytes.len() as u64);
+            let wire_sequence = state.wire_frames;
+            let chunk_count = bytes.len().div_ceil(32 * 1024).max(1);
+            // Each log entry stays bounded even for images or very large WebSocket frames.
+            for chunk_index in 0..chunk_count {
+                let start = chunk_index * 32 * 1024;
+                let chunk = &bytes[start..bytes.len().min(start + 32 * 1024)];
+                tracing::info!(target: "request_dump", request_id = %state.request_id,
+                    attempt_index = self.attempt_index, exchange_id = self.exchange_id,
+                    stage, wire_sequence, chunk_index, chunk_count,
+                    body_bytes = bytes.len(), body_base64 = %STANDARD.encode(chunk),
+                    contains_sensitive_data = true, "diagnostic wire capture");
+            }
+        }
+    }
+
+    /// 与请求终态原子持久化；旧记录没有此字段，不伪造历史时间线。
+    #[must_use]
+    pub fn snapshot(&self) -> Option<Value> {
+        let state = self
+            .state
+            .as_ref()?
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        Some(json!({
+            "schemaVersion": 1, "requestId": state.request_id,
+            "startedAtUnixMs": state.started_at_ms, "totalEvents": state.sequence,
+            "droppedEvents": state.dropped_events, "maxEvents": MAX_EVENTS,
+            "captureMode": "sanitized", "events": state.events,
+            "wireDumpEnabled": tracing::enabled!(target: "request_dump", tracing::Level::INFO),
+            "wireFrames": state.wire_frames, "wireBytes": state.wire_bytes,
+            "maxBufferBytes": MAX_BUFFER_BYTES, "headEvents": HEAD_EVENTS,
+            "snapshotBoundary": "execution_finalization",
+        }))
+    }
+
+    fn push(&self, stage: &'static str, mut data: Value, coalesce: bool) {
+        let Some(state) = &self.state else { return };
+        let mut encoded = serde_json::to_vec(&data).unwrap_or_default();
+        if encoded.len() > MAX_DATA_BYTES {
+            data = json!({"truncated": true, "summary": body_fingerprint(&encoded),
+                "eventType": data.get("eventType"), "body": data.get("body"),
+                "jsonValid": data.get("jsonValid"),
+            });
+            encoded = serde_json::to_vec(&data).unwrap_or_default();
+        }
+        let mut state = state
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        state.sequence += 1;
+        let sequence = state.sequence;
+        let elapsed_ms = u64::try_from(state.started.elapsed().as_millis()).unwrap_or(u64::MAX);
+        if coalesce
+            && let Some(last) = state.events.iter_mut().rev().find(|event| {
+                event.stage == stage
+                    && event.attempt_index == self.attempt_index
+                    && event.exchange_id == self.exchange_id
+            })
+            && last.stage == stage
+            && last.attempt_index == self.attempt_index
+            && last.exchange_id == self.exchange_id
+            && last.data.get("eventType") == data.get("eventType")
+        {
+            last.count += 1;
+            last.last_sequence = sequence;
+            last.last_elapsed_ms = elapsed_ms;
+            return;
+        }
+        // 普通日志只输出有界安全事实；进程意外退出时仍可按 ID 检索已发生阶段。
+        tracing::info!(target: "request_trace", request_id = %state.request_id,
+            attempt_index = self.attempt_index, exchange_id = self.exchange_id,
+            sequence, elapsed_ms, stage, data = %data, "request trace");
+        let bytes = encoded.len() + 256;
+        state.bytes += bytes;
+        state.events.push_back(TraceEvent {
+            sequence,
+            last_sequence: sequence,
+            elapsed_ms,
+            last_elapsed_ms: elapsed_ms,
+            attempt_index: self.attempt_index,
+            exchange_id: self.exchange_id,
+            stage,
+            count: 1,
+            data,
+            bytes,
+        });
+        while state.events.len() > MAX_EVENTS || state.bytes > MAX_BUFFER_BYTES {
+            // Preserve failures and retry decisions ahead of routine middle-of-stream events.
+            // Head/tail and important stages compete under the same hard memory limit.
+            let index = (HEAD_EVENTS..state.events.len().saturating_sub(8))
+                .find(|index| {
+                    !matches!(
+                        state.events[*index].stage,
+                        "attempt.started"
+                            | "account.selected"
+                            | "attempt.failed"
+                            | "retry.decided"
+                            | "upstream.close"
+                            | "upstream.read.failed"
+                            | "capture.gap"
+                            | "request.finished"
+                    )
+                })
+                .unwrap_or(if state.events.len() > HEAD_EVENTS {
+                    HEAD_EVENTS
+                } else {
+                    0
+                });
+            if let Some(event) = state.events.remove(index) {
+                state.bytes = state.bytes.saturating_sub(event.bytes);
+                state.dropped_events += event.count;
+            }
+        }
+    }
+}
+
+impl Drop for TraceState {
+    fn drop(&mut self) {
+        tracing::info!(target: "request_trace", request_id = %self.request_id,
+            stage = "request.trace.closed", total_events = self.sequence,
+            dropped_events = self.dropped_events, wire_frames = self.wire_frames,
+            wire_bytes = self.wire_bytes, "request diagnostic capture closed");
+    }
+}

backend/crates/gateway-core/src/engine/continuation.rs

@@ -8,9 +8,9 @@ use std::fmt;
 use futures::future::BoxFuture;
 
 use crate::account::ProviderAccountId;
+use crate::identity::ProviderKind;
 use crate::operation::ProviderSessionState;
 use crate::policy::ClientApiKeyId;
-use crate::routing::ProviderKind;
 
 /// 客户端或 Provider 传递的 opaque response handle。
 ///

backend/crates/gateway-core/src/engine/coordinator.rs

@@ -1,21 +1,26 @@
 //! 唯一的账号重试、发送与下游 commit barrier owner。
 
+use crate::diagnostics::TraceContext;
+use serde_json::json;
+
 use std::collections::BTreeSet;
 use std::future::Future;
 use std::num::NonZeroU32;
 use std::sync::Arc;
 use std::time::{Duration, Instant, SystemTime};
 
+use super::observation::ResponseObservation;
 use crate::engine::continuation::{
     ContinuationBinding, NativeContinuationPin, NativeContinuationScope, PreviousResponseId,
 };
+
 use crate::engine::provider::{Provider, ProviderCallMetadata, ProviderRequest, ProviderStream};
 use crate::engine::{
     AccountAttemptContext, AttemptContext, AttemptRecord, AttemptTransport, AttemptTrigger,
     CommitRequirement, ContinuationAttempt, CoordinatedEvent, EngineError, ExecutionOutcome,
     ExecutionStore, GatewayEngine, IntermediateFailure, ModelRequestFailureObservation,
-    ModelRequestFinalization, ModelRequestId, ModelRequestTimings, NewModelRequest,
-    ProviderAccountStateOwner, ProviderAttemptOutcome, RequestAttemptContext, UpstreamSendState,
+    ModelRequestFinalization, ModelRequestId, NewModelRequest, ProviderAccountStateOwner,
+    ProviderAttemptOutcome, RequestAttemptContext, UpstreamSendState,
 };
 use crate::error::{
     ContinuationRecoveryDisposition, GatewayError, GatewayErrorKind, ProviderError,
@@ -25,7 +30,6 @@ use crate::event::{
     GatewayEvent, ProviderEvent, ProviderResponseHeader, ProviderResponseObservation,
 };
 use crate::lifecycle::CancellationToken;
-use crate::metering::{CostEstimate, CostSource, Usage};
 use crate::operation::{Operation, ProviderSessionState};
 use crate::routing::RoutingPlan;
 use futures::future::Fuse;
@@ -143,12 +147,26 @@ where
                     .position(|candidate| candidate.provider() == pin.provider())
                     .ok_or(EngineError::ContinuationPinMismatch)
             })?;
+        let trace = TraceContext::new(request_id.as_str());
+        trace.record(
+            "request.routed",
+            json!({
+                "endpoint": request.endpoint, "protocol": request.protocol,
+                "clientTransport": request.client_transport,
+                "model": request.requested_model.as_ref().map(|model| model.as_str()),
+                "configRevision": format!("{:?}", request.config_revision),
+                "admissionMs": request.admission_decision_ms,
+                "continuation": format!("{continuation_attempt:?}"),
+                "candidateCount": plan.candidates().len(),
+            }),
+        );
         let image_generation_requested = operation.image_generation_requested();
         let mut session = ResponseExecutionSession {
             engine: Arc::clone(&self.engine),
             request_id,
             client_api_key_ref,
-            timing_started_at,
+            observation: ResponseObservation::new(timing_started_at),
+            trace,
             deadline,
             deadline_timer: Delay::new(
                 deadline
@@ -179,12 +197,7 @@ where
             delivery_pending: false,
             upstream_complete: false,
             finalized: false,
-            usage: Usage::new(),
             image_generation_requested,
-            cost: CostEstimate::unavailable(),
-            timings: ModelRequestTimings::default(),
-            client_response_id: None,
-            upstream_response_id: None,
             last_retryable_failure: None,
             last_retryable_failure_events: Vec::new(),
             provider_attempt_outcomes: Vec::new(),
@@ -239,7 +252,8 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     engine: Arc<GatewayEngine<S>>,
     request_id: ModelRequestId,
     client_api_key_ref: crate::policy::ClientApiKeyId,
-    timing_started_at: Instant,
+    observation: ResponseObservation,
+    trace: TraceContext,
     deadline: SystemTime,
     /// 会话级 deadline 计时器；deadline 固定，帧循环内复用而非逐事件新建。
     deadline_timer: Fuse<Delay>,
@@ -274,12 +288,7 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     delivery_pending: bool,
     upstream_complete: bool,
     finalized: bool,
-    usage: Usage,
     image_generation_requested: bool,
-    cost: CostEstimate,
-    timings: ModelRequestTimings,
-    client_response_id: Option<String>,
-    upstream_response_id: Option<String>,
     /// 最近一次为无感恢复而被丢弃的原始上游错误；只在后续空选路时成为终态。
     last_retryable_failure: Option<ProviderError>,
     /// 与 `last_retryable_failure` 同属一个 attempt 的原始失败批次。
@@ -298,6 +307,11 @@ impl<S: ?Sized> ResponseExecutionSession<S>
 where
     S: ExecutionStore,
 {
+    /// 当前请求共享的诊断上下文。
+    pub fn trace(&self) -> TraceContext {
+        self.trace.clone()
+    }
+
     /// 读取下一条 canonical event；首条未提交事件会携带 commit 要求。
     ///
     /// # Errors
@@ -405,6 +419,10 @@ where
             return Err(EngineError::InvalidDeliveryState);
         }
         let committed_at = SystemTime::now();
+        self.trace.record(
+            "downstream.committed",
+            json!({"status": client_status_code}),
+        );
         if self.request_persisted {
             best_effort_store_write(
                 "mark_downstream_committed",
@@ -498,7 +516,7 @@ where
             return None;
         }
         let account = current.metadata.provider_account_id().clone();
-        let response_id = self.upstream_response_id.as_deref()?;
+        let response_id = self.observation.upstream_response_id.as_deref()?;
         let previous_response_id = PreviousResponseId::new(response_id.to_owned());
         let upstream_response_id = PreviousResponseId::new(response_id.to_owned());
         Some(
@@ -570,11 +588,19 @@ where
                     return Err(EngineError::Deadline);
                 }
                 PollBoundary::Item(Some(Ok(mut event))) => {
+                    if let Some(wire) = event.wire_event() {
+                        self.trace
+                            .attempt(self.attempts)
+                            .wire_event(wire.protocol(), wire.event_type());
+                    }
+                    if let Some(state) = event.session_update() {
+                        self.observe_session_update(state);
+                    }
                     if let Some(observation) = event.take_observation() {
                         self.observe_response(observation);
                     }
                     for fact in event.canonical_facts() {
-                        self.observe_response_identity(fact);
+                        self.observation.observe_identity(fact);
                     }
                     for fact in event.canonical_facts() {
                         self.observe_event(fact).await;
@@ -686,7 +712,8 @@ where
         }));
         let context = AttemptContext::new(
             RequestAttemptContext::new(self.request_id.clone(), self.client_api_key_ref.clone())
-                .with_timing_started_at(self.timing_started_at),
+                .with_timing_started_at(self.observation.timing_started_at)
+                .with_trace(self.trace.clone()),
             next_attempt,
             self.deadline,
             self.plan.account_selection_policy(),
@@ -709,6 +736,18 @@ where
             .ok_or_else(|| EngineError::ProviderNotRegistered {
                 provider: candidate.provider().as_str().to_owned(),
             })?;
+        let attempt_trace = self.trace.attempt(next_attempt.get());
+        attempt_trace.record(
+            "attempt.started",
+            json!({
+                "provider": candidate.provider().as_str(),
+                "model": candidate.upstream_model().map(|model| model.as_str()),
+                "transportPolicy": format!("{attempt_transport:?}"),
+                "continuation": format!("{:?}", self.continuation_attempt),
+                "pinnedAccount": pinned_account.as_ref().map(|id| id.as_str()),
+                "transportRecovery": is_transport_recovery,
+            }),
+        );
         let provider_request = ProviderRequest::new(self.operation.clone(), candidate.clone());
         let stream = match poll_provider(
             provider,
@@ -733,6 +772,7 @@ where
             ProviderBoundary::Result(result) => match *result {
                 Ok(stream) => stream,
                 Err(error) => {
+                    record_trace_error(&attempt_trace, &error);
                     if self.prepare_unavailable_native_continuation_replay(&error) {
                         return Ok(Some(PullOutcome::AttemptDiscarded));
                     }
@@ -804,6 +844,11 @@ where
         }
 
         let metadata = stream.metadata().clone();
+        attempt_trace.record("account.selected", json!({
+            "provider": metadata.provider().as_str(), "accountId": metadata.provider_account_id().as_str(),
+            "transport": metadata.transport().as_str(),
+            "selectionMs": metadata.selection_observation().map(|o| o.account_selection_wait_ms()),
+        }));
         let selection_observation = metadata.selection_observation();
         let capacity = selection_observation.and_then(|observation| observation.capacity());
         if !self.account_selection.is_diagnostic()
@@ -948,30 +993,24 @@ where
     }
 
     async fn observe_event(&mut self, event: &GatewayEvent) {
-        self.observe_event_facts(event);
+        self.observation.observe_event(event);
         self.mark_send_observed().await;
     }
 
-    fn observe_event_facts(&mut self, event: &GatewayEvent) {
-        let elapsed = elapsed_ms(self.timing_started_at);
-        observe_event_timing(&mut self.timings, event, elapsed);
-        if let GatewayEvent::Usage(observed) = event {
-            self.usage.merge(observed);
-        }
-        if let GatewayEvent::CalculatedCost(observed) = event
-            && self.cost.source() != CostSource::ProviderReported
-        {
-            self.cost = observed.into_estimate();
-        }
-        if let GatewayEvent::ProviderCost(observed) = event {
-            self.cost = observed.into_estimate();
-        }
-    }
-
     async fn observe_wire_event(&mut self) {
         self.mark_send_observed().await;
     }
 
+    fn observe_session_update(&mut self, state: &ProviderSessionState) {
+        let Some(current) = self.current.as_mut() else {
+            return;
+        };
+        if state.provider() != current.metadata.provider().as_str() {
+            return;
+        }
+        self.operation.set_provider_session_state(state.clone());
+    }
+
     async fn mark_send_observed(&mut self) {
         let Some(current) = self.current.as_mut() else {
             return;
@@ -1003,44 +1042,10 @@ where
         {
             return;
         }
-        let observed = observation.timings();
-        if let Some(value) = observed.transport_decision_wait_ms {
-            self.timings.transport_decision_wait_ms = Some(value);
-        }
-        if let Some(value) = observed.connect_ms {
-            self.timings.connect_ms = Some(value);
-        }
-        if let Some(value) = observed.headers_ms {
-            self.timings.headers_ms = Some(value);
-        }
-        if let Some(value) = observed.first_event_ms {
-            self.timings.first_event_ms = Some(value);
-        }
-        if let Some(value) = observed.first_reasoning_ms {
-            self.timings.first_reasoning_ms = Some(value);
-        }
-        if let Some(value) = observed.first_text_ms {
-            self.timings.first_text_ms = Some(value);
-        }
-        if let Some(value) = observed.first_token_ms {
-            self.timings.first_token_ms = Some(value);
-        }
-        if let Some(value) = observed.provider_processing_ms {
-            self.timings.provider_processing_ms = Some(value);
-        }
+        self.observation.observe_response(&observation);
         current.response_observation = Some(observation);
     }
 
-    fn observe_response_identity(&mut self, event: &GatewayEvent) {
-        let metadata = match event {
-            GatewayEvent::Started(metadata) | GatewayEvent::Completed(metadata) => metadata,
-            _ => return,
-        };
-        let response_id = metadata.response_id().to_owned();
-        self.client_response_id = Some(response_id.clone());
-        self.upstream_response_id = Some(response_id);
-    }
-
     /// 失败可重试时要求丢弃本 attempt；预算耗尽时可返回最后一批原始失败事件。
     async fn handle_stream_error(
         &mut self,
@@ -1048,6 +1053,7 @@ where
     ) -> Result<StreamErrorOutcome, EngineError> {
         // 原始 wire 只活在 request-local 决策状态；clone、attempt 记录与持久化终态
         // 均只接触已剥离的稳定错误字段。
+        record_trace_error(&self.trace.attempt(self.attempts), &error);
         let mut atomic_client_events = error.take_atomic_client_events();
         let current = self.current.take().ok_or(EngineError::NoActiveAttempt)?;
         self.record_provider_failure(current.metadata.provider().clone(), error.kind());
@@ -1128,6 +1134,14 @@ where
             || account_rotation_retry
             || transport_recovery.is_some();
 
+        self.trace.attempt(current.index.get()).record("retry.decided", json!({
+            "retryable": retryable, "continuationRetry": continuation_retry,
+            "sameAccountRetry": same_account_retry, "accountRotationRetry": account_rotation_retry,
+            "ordinaryRetry": ordinary_retry, "transportRecovery": transport_recovery.is_some(),
+            "delayMs": transport_recovery.map(|(_, delay)| duration_ms(delay)),
+            "downstreamCommitted": self.downstream_committed_at.is_some(),
+            "sendState": format!("{attempt_send_state:?}"),
+        }));
         if retryable {
             // 普通 clone 只保留稳定事实；原始 wire/HTTP response 由 request-local
             // 所有权保留到下一次 attempt 成功，或最终空选路时返回客户端。
@@ -1203,8 +1217,8 @@ where
                 self.observe_response(observation);
             }
             for fact in event.canonical_facts() {
-                self.observe_response_identity(fact);
-                self.observe_event_facts(fact);
+                self.observation.observe_identity(fact);
+                self.observation.observe_event(fact);
             }
         }
     }
@@ -1303,18 +1317,7 @@ where
     }
 
     fn reset_uncommitted_observations(&mut self) {
-        self.usage = Usage::new();
-        self.cost = CostEstimate::unavailable();
-        self.client_response_id = None;
-        self.upstream_response_id = None;
-        self.timings.transport_decision_wait_ms = None;
-        self.timings.connect_ms = None;
-        self.timings.headers_ms = None;
-        self.timings.first_event_ms = None;
-        self.timings.first_reasoning_ms = None;
-        self.timings.first_text_ms = None;
-        self.timings.first_token_ms = None;
-        self.timings.provider_processing_ms = None;
+        self.observation.reset_for_attempt();
         self.upstream_complete = false;
     }
 
@@ -1323,7 +1326,7 @@ where
             return Ok(());
         }
         let completed_at = SystemTime::now();
-        self.timings.latency_ms = Some(elapsed_ms(self.timing_started_at));
+        self.observation.finish();
         let upstream_request_id = self.current.as_ref().and_then(|current| {
             current
                 .response_observation
@@ -1349,6 +1352,10 @@ where
             self.current_transport_observation();
         let service_tier = self.current_service_tier();
         let provider_metadata_json = self.current_provider_metadata_json();
+        self.trace.record(
+            "request.finished",
+            json!({"outcome": "succeeded", "attempts": self.attempts}),
+        );
         if self.request_persisted {
             best_effort_store_write(
                 "finalize_model_request",
@@ -1363,23 +1370,24 @@ where
                         downstream_committed_at: self.downstream_committed_at,
                         client_status_code: self.client_status_code,
                         upstream_status_code,
-                        client_response_id: self.client_response_id.clone(),
+                        client_response_id: self.observation.client_response_id.clone(),
                         upstream_request_id,
-                        upstream_response_id: self.upstream_response_id.clone(),
+                        upstream_response_id: self.observation.upstream_response_id.clone(),
                         upstream_transport,
                         http_version,
                         websocket_pool,
                         service_tier,
                         provider_metadata_json,
+                        diagnostic_trace_json: self.trace.snapshot().map(|value| value.to_string()),
                         error: None,
                         provider_error_code: None,
                         raw_upstream_error: None,
                         failure_observation: ModelRequestFailureObservation::default(),
                         retry_after_ms: None,
-                        usage: self.usage.clone(),
+                        usage: self.observation.usage.clone(),
                         image_generation_succeeded: self.image_generation_succeeded(),
-                        cost: self.cost.clone(),
-                        timings: self.timings.clone(),
+                        cost: self.observation.cost.clone(),
+                        timings: self.observation.timings.clone(),
                         completed_at,
                     }),
             )
@@ -1489,12 +1497,20 @@ where
         if self.finalized {
             return Ok(());
         }
+        self.trace.record(
+            "request.finished",
+            json!({
+                "outcome": format!("{:?}", finalization.outcome),
+                "errorKind": finalization.error.kind().as_str(),
+                "sendState": format!("{:?}", finalization.send_state), "attempts": self.attempts,
+            }),
+        );
         if !self.request_persisted {
             self.finalized = true;
             return Ok(());
         }
         let completed_at = SystemTime::now();
-        self.timings.latency_ms = Some(elapsed_ms(self.timing_started_at));
+        self.observation.finish();
         let upstream_request_id = self.current.as_ref().and_then(|current| {
             current
                 .response_observation
@@ -1535,23 +1551,24 @@ where
                     upstream_status_code: finalization
                         .upstream_status_code
                         .or(observed_status_code),
-                    client_response_id: self.client_response_id.clone(),
+                    client_response_id: self.observation.client_response_id.clone(),
                     upstream_request_id,
-                    upstream_response_id: self.upstream_response_id.clone(),
+                    upstream_response_id: self.observation.upstream_response_id.clone(),
                     upstream_transport,
                     http_version,
                     websocket_pool,
                     service_tier,
                     provider_metadata_json,
+                    diagnostic_trace_json: self.trace.snapshot().map(|value| value.to_string()),
                     error: Some(finalization.error),
                     provider_error_code: finalization.provider_error_code,
                     raw_upstream_error: finalization.raw_upstream_error,
                     failure_observation: finalization.observation,
                     retry_after_ms: finalization.retry_after_ms,
-                    usage: self.usage.clone(),
+                    usage: self.observation.usage.clone(),
                     image_generation_succeeded: self.image_generation_succeeded(),
-                    cost: self.cost.clone(),
-                    timings: self.timings.clone(),
+                    cost: self.observation.cost.clone(),
+                    timings: self.observation.timings.clone(),
                     completed_at,
                 }),
         )
@@ -1599,7 +1616,11 @@ where
             if matches!(self.operation, Operation::GenerateImage(_)) {
                 self.upstream_complete
             } else {
-                self.usage.image_output_tokens.unwrap_or_default() > 0
+                self.observation
+                    .usage
+                    .image_output_tokens
+                    .unwrap_or_default()
+                    > 0
             }
         })
     }
@@ -1637,7 +1658,7 @@ where
 
     fn record_provider_failure(
         &mut self,
-        provider_kind: crate::routing::ProviderKind,
+        provider_kind: crate::identity::ProviderKind,
         error_kind: ProviderErrorKind,
     ) {
         self.provider_attempt_outcomes
@@ -1826,31 +1847,18 @@ const fn escalate_send_state(a: UpstreamSendState, b: UpstreamSendState) -> Upst
     }
 }
 
-fn observe_event_timing(timings: &mut ModelRequestTimings, event: &GatewayEvent, elapsed_ms: u64) {
-    timings.first_event_ms.get_or_insert(elapsed_ms);
-    match event {
-        GatewayEvent::ReasoningDelta(_) => {
-            timings.first_reasoning_ms.get_or_insert(elapsed_ms);
-            timings.first_token_ms.get_or_insert(elapsed_ms);
-        }
-        GatewayEvent::TextDelta(_) => {
-            timings.first_text_ms.get_or_insert(elapsed_ms);
-            timings.first_token_ms.get_or_insert(elapsed_ms);
-        }
-        // `response.output_item.added` 会先投影一个空参数的 tool delta；它只是结构帧，
-        // 不能抢在真实工具参数之前成为首个可消费 token。
-        GatewayEvent::ToolCallDelta(delta) if !delta.arguments_delta.is_empty() => {
-            timings.first_token_ms.get_or_insert(elapsed_ms);
-        }
-        GatewayEvent::CalculatedCost(_) | GatewayEvent::ProviderCost(_) => {}
-        _ => {}
-    }
-}
-
-fn elapsed_ms(started_at: Instant) -> u64 {
-    duration_ms(started_at.elapsed())
-}
-
 fn duration_ms(duration: Duration) -> u64 {
     u64::try_from(duration.as_millis()).unwrap_or(u64::MAX)
 }
+
+fn record_trace_error(trace: &TraceContext, error: &ProviderError) {
+    trace.record("attempt.failed", json!({
+        "kind": error.kind().as_str(), "sendState": format!("{:?}", error.send_state()),
+        "upstreamStatus": error.upstream_status(),
+        "upstreamCode": error.upstream_code().map(|code| code.as_str()),
+        "rawError": error.raw_upstream_error().map(|raw| {
+            let value = serde_json::from_str(raw.as_str()).unwrap_or_else(|_| json!(raw.as_str()));
+            crate::diagnostics::diagnostic_json(&value)
+        }),
+    }));
+}

backend/crates/gateway-core/src/engine/execution.rs

@@ -31,12 +31,12 @@ use crate::engine::{
 };
 use crate::error::{GatewayError, GatewayErrorKind, ProviderErrorKind, StoreError};
 use crate::event::{GatewayEvent, ProviderEvent, ProviderResponseHeader};
+use crate::identity::ProviderKind;
 use crate::lifecycle::CancellationToken;
 use crate::operation::{Operation, ProviderSessionState};
 use crate::policy::{ClientApiKeyId, ClientPolicy};
 use crate::routing::{
-    ProviderKind, PublicModelId, PublicModelProfile, RoutingContext, RuntimeSnapshot,
-    UpstreamModelId,
+    PublicModelId, PublicModelProfile, RoutingContext, RuntimeSnapshot, UpstreamModelId,
 };
 use crate::runtime::RuntimeSnapshotHandle;
 
@@ -155,6 +155,9 @@ pub struct StartedExecution {
 }
 
 pub trait ExecutionSession: Send {
+    fn trace(&self) -> crate::diagnostics::TraceContext {
+        crate::diagnostics::TraceContext::default()
+    }
     fn next_event(&mut self) -> BoxFuture<'_, Result<Option<CoordinatedEvent>, EngineError>>;
     fn collect_uncommitted(&mut self) -> BoxFuture<'_, Result<Vec<ProviderEvent>, EngineError>>;
     fn response_headers(&self) -> &[ProviderResponseHeader];
@@ -1035,6 +1038,12 @@ impl Drop for DefaultExecutionSession {
 }
 
 impl ExecutionSession for DefaultExecutionSession {
+    fn trace(&self) -> crate::diagnostics::TraceContext {
+        self.core
+            .as_ref()
+            .map(ResponseExecutionSession::trace)
+            .unwrap_or_default()
+    }
     fn next_event(&mut self) -> BoxFuture<'_, Result<Option<CoordinatedEvent>, EngineError>> {
         Box::pin(async move {
             let result = self.core_mut()?.next_event().await;
@@ -1202,11 +1211,11 @@ fn new_request_id() -> Result<ModelRequestId, GatewayError> {
         .map_err(|_| GatewayError::new(GatewayErrorKind::Internal, "failed to allocate request ID"))
 }
 
-fn map_routing_error(error: crate::error::RoutingError) -> GatewayError {
+fn map_routing_error(error: crate::validation::RoutingError) -> GatewayError {
     match error {
-        crate::error::RoutingError::NoCapableProvider { .. }
-        | crate::error::RoutingError::NoCapableProviderEndpoint { .. }
-        | crate::error::RoutingError::EmptyAccountScope => GatewayError::new(
+        crate::validation::RoutingError::NoCapableProvider { .. }
+        | crate::validation::RoutingError::NoCapableProviderEndpoint { .. }
+        | crate::validation::RoutingError::EmptyAccountScope => GatewayError::new(
             GatewayErrorKind::NoAvailableProvider,
             "no provider can execute this request",
         ),

backend/crates/gateway-core/src/engine/mod.rs

@@ -4,6 +4,7 @@ pub mod admission;
 pub mod continuation;
 pub mod coordinator;
 pub mod execution;
+mod observation;
 pub mod probe;
 pub mod provider;
 
@@ -22,17 +23,18 @@ use thiserror::Error;
 use crate::account::{AccountSelectionPolicy, ProviderAccountId};
 use crate::engine::continuation::{ContinuationBinding, NativeContinuationPin};
 use crate::error::{
-    GatewayError, IdentifierError, ProviderConnectionObservation, ProviderError, ProviderErrorKind,
-    StoreError, validate_text,
+    GatewayError, ProviderConnectionObservation, ProviderError, ProviderErrorKind, StoreError,
 };
 use crate::event::ProviderEvent;
+use crate::identity::ProviderKind;
 use crate::lifecycle::CancellationToken;
 use crate::metering::{CostEstimate, Usage};
 use crate::operation::OperationKind;
 use crate::operation::ProviderSessionState;
 use crate::policy::ClientApiKeyId;
-use crate::routing::{ConfigRevision, ProviderKind, PublicModelId, UpstreamModelId};
+use crate::routing::{ConfigRevision, PublicModelId, UpstreamModelId};
 use crate::upstream::UpstreamSendState;
+use crate::validation::{IdentifierError, validate_text};
 
 /// `model_requests.id`。
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
@@ -153,7 +155,7 @@ pub struct AccountAttemptContext {
     state_owner: Option<ProviderAccountStateOwner>,
     credential_recovery_attempted: bool,
     diagnostic_required_account: bool,
-    account_scope: Option<Arc<crate::routing::FrozenAccountScope>>,
+    account_scope: Option<Arc<crate::account::scope::FrozenAccountScope>>,
 }
 
 impl AccountAttemptContext {
@@ -194,7 +196,10 @@ impl AccountAttemptContext {
 
     /// 附着普通请求认证时冻结的账号范围。
     #[must_use]
-    pub fn with_account_scope(mut self, scope: Arc<crate::routing::FrozenAccountScope>) -> Self {
+    pub fn with_account_scope(
+        mut self,
+        scope: Arc<crate::account::scope::FrozenAccountScope>,
+    ) -> Self {
         self.account_scope = Some(scope);
         self
     }
@@ -233,7 +238,7 @@ impl AccountAttemptContext {
     }
 
     #[must_use]
-    pub const fn account_scope(&self) -> Option<&Arc<crate::routing::FrozenAccountScope>> {
+    pub const fn account_scope(&self) -> Option<&Arc<crate::account::scope::FrozenAccountScope>> {
         self.account_scope.as_ref()
     }
 }
@@ -311,6 +316,7 @@ pub struct RequestAttemptContext {
     request_id: ModelRequestId,
     client_api_key_ref: ClientApiKeyId,
     timing_started_at: Instant,
+    trace: crate::diagnostics::TraceContext,
 }
 
 impl RequestAttemptContext {
@@ -320,9 +326,16 @@ impl RequestAttemptContext {
             request_id,
             client_api_key_ref,
             timing_started_at: Instant::now(),
+            trace: crate::diagnostics::TraceContext::default(),
         }
     }
 
+    #[must_use]
+    pub fn with_trace(mut self, trace: crate::diagnostics::TraceContext) -> Self {
+        self.trace = trace;
+        self
+    }
+
     /// 覆盖本次请求的单调计时原点。
     #[must_use]
     pub fn with_timing_started_at(mut self, timing_started_at: Instant) -> Self {
@@ -362,6 +375,12 @@ pub struct AttemptContext {
 }
 
 impl AttemptContext {
+    /// 当前 attempt 的诊断关联；克隆后可传给后台 transport 任务。
+    #[must_use]
+    pub fn trace(&self) -> crate::diagnostics::TraceContext {
+        self.request.trace.attempt(self.attempt_index.get())
+    }
+
     #[must_use]
     pub const fn new(
         request: RequestAttemptContext,
@@ -469,7 +488,7 @@ impl AttemptContext {
 
     /// 普通请求认证时冻结的账号范围；管理端诊断为 `None`。
     #[must_use]
-    pub const fn account_scope(&self) -> Option<&Arc<crate::routing::FrozenAccountScope>> {
+    pub const fn account_scope(&self) -> Option<&Arc<crate::account::scope::FrozenAccountScope>> {
         self.account.account_scope()
     }
 
@@ -607,6 +626,8 @@ pub struct ModelRequestFinalization {
     pub service_tier: Option<String>,
     /// Provider 已筛选的专有观测 JSON；Core 不解释字段。
     pub provider_metadata_json: Option<String>,
+    /// 请求全程的有界诊断快照，跨 Provider 与重试保留。
+    pub diagnostic_trace_json: Option<String>,
     pub error: Option<GatewayError>,
     pub provider_error_code: Option<String>,
     /// Provider 返回的原始错误正文或 WebSocket close/error frame。

backend/crates/gateway-core/src/engine/observation.rs

@@ -0,0 +1,127 @@
+//! 单次响应的计量、时间与响应身份事实；重试丢弃时统一清理。
+
+use super::ModelRequestTimings;
+use crate::event::GatewayEvent;
+use crate::event::ProviderResponseObservation;
+use crate::metering::{CostEstimate, CostSource, Usage};
+use std::time::Instant;
+
+pub(super) struct ResponseObservation {
+    pub(super) timing_started_at: Instant,
+    pub(super) usage: Usage,
+    pub(super) cost: CostEstimate,
+    pub(super) timings: ModelRequestTimings,
+    pub(super) client_response_id: Option<String>,
+    pub(super) upstream_response_id: Option<String>,
+}
+
+impl ResponseObservation {
+    pub(super) fn new(timing_started_at: Instant) -> Self {
+        Self {
+            timing_started_at,
+            usage: Usage::new(),
+            cost: CostEstimate::unavailable(),
+            timings: ModelRequestTimings::default(),
+            client_response_id: None,
+            upstream_response_id: None,
+        }
+    }
+
+    fn elapsed_ms(&self) -> u64 {
+        u64::try_from(self.timing_started_at.elapsed().as_millis()).unwrap_or(u64::MAX)
+    }
+
+    pub(super) fn finish(&mut self) {
+        self.timings.latency_ms = Some(self.elapsed_ms());
+    }
+
+    pub(super) fn observe_event(&mut self, event: &GatewayEvent) {
+        let elapsed = self.elapsed_ms();
+        observe_event_timing(&mut self.timings, event, elapsed);
+        if let GatewayEvent::Usage(observed) = event {
+            self.usage.merge(observed);
+        }
+        if let GatewayEvent::CalculatedCost(observed) = event
+            && self.cost.source() != CostSource::ProviderReported
+        {
+            self.cost = observed.into_estimate();
+        }
+        if let GatewayEvent::ProviderCost(observed) = event {
+            self.cost = observed.into_estimate();
+        }
+    }
+
+    pub(super) fn observe_identity(&mut self, event: &GatewayEvent) {
+        let metadata = match event {
+            GatewayEvent::Started(metadata) | GatewayEvent::Completed(metadata) => metadata,
+            _ => return,
+        };
+        let response_id = metadata.response_id().to_owned();
+        self.client_response_id = Some(response_id.clone());
+        self.upstream_response_id = Some(response_id);
+    }
+
+    pub(super) fn reset_for_attempt(&mut self) {
+        self.usage = Usage::new();
+        self.cost = CostEstimate::unavailable();
+        self.client_response_id = None;
+        self.upstream_response_id = None;
+        self.timings.transport_decision_wait_ms = None;
+        self.timings.connect_ms = None;
+        self.timings.headers_ms = None;
+        self.timings.first_event_ms = None;
+        self.timings.first_reasoning_ms = None;
+        self.timings.first_text_ms = None;
+        self.timings.first_token_ms = None;
+        self.timings.provider_processing_ms = None;
+    }
+
+    pub(super) fn observe_response(&mut self, observation: &ProviderResponseObservation) {
+        let observed = observation.timings();
+        if let Some(value) = observed.transport_decision_wait_ms {
+            self.timings.transport_decision_wait_ms = Some(value);
+        }
+        if let Some(value) = observed.connect_ms {
+            self.timings.connect_ms = Some(value);
+        }
+        if let Some(value) = observed.headers_ms {
+            self.timings.headers_ms = Some(value);
+        }
+        if let Some(value) = observed.first_event_ms {
+            self.timings.first_event_ms = Some(value);
+        }
+        if let Some(value) = observed.first_reasoning_ms {
+            self.timings.first_reasoning_ms = Some(value);
+        }
+        if let Some(value) = observed.first_text_ms {
+            self.timings.first_text_ms = Some(value);
+        }
+        if let Some(value) = observed.first_token_ms {
+            self.timings.first_token_ms = Some(value);
+        }
+        if let Some(value) = observed.provider_processing_ms {
+            self.timings.provider_processing_ms = Some(value);
+        }
+    }
+}
+
+fn observe_event_timing(timings: &mut ModelRequestTimings, event: &GatewayEvent, elapsed_ms: u64) {
+    timings.first_event_ms.get_or_insert(elapsed_ms);
+    match event {
+        GatewayEvent::ReasoningDelta(_) => {
+            timings.first_reasoning_ms.get_or_insert(elapsed_ms);
+            timings.first_token_ms.get_or_insert(elapsed_ms);
+        }
+        GatewayEvent::TextDelta(_) => {
+            timings.first_text_ms.get_or_insert(elapsed_ms);
+            timings.first_token_ms.get_or_insert(elapsed_ms);
+        }
+        // `response.output_item.added` 会先投影一个空参数的 tool delta；它只是结构帧，
+        // 不能抢在真实工具参数之前成为首个可消费 token。
+        GatewayEvent::ToolCallDelta(delta) if !delta.arguments_delta.is_empty() => {
+            timings.first_token_ms.get_or_insert(elapsed_ms);
+        }
+        GatewayEvent::CalculatedCost(_) | GatewayEvent::ProviderCost(_) => {}
+        _ => {}
+    }
+}

backend/crates/gateway-core/src/engine/probe.rs

@@ -6,7 +6,8 @@ use bytes::Bytes;
 use futures::future::BoxFuture;
 
 use crate::error::{ClientVisibleUpstreamResponse, GatewayError, GatewayErrorKind};
-use crate::routing::{ProviderKind, UpstreamModelId};
+use crate::identity::ProviderKind;
+use crate::routing::UpstreamModelId;
 use crate::{account::ProviderAccountId, operation::Operation, upstream::UpstreamSendState};
 
 #[derive(Debug, Clone, PartialEq)]

backend/crates/gateway-core/src/engine/provider.rs

@@ -15,13 +15,13 @@ use crate::account::{
     AccountAttemptFeedback, AccountCapacitySnapshot, AccountFeedbackStats, ProviderAccountId,
 };
 use crate::engine::AttemptContext;
-use crate::error::{OpaqueUpstreamValue, PreDeliveryRetry, ProviderError, ProviderErrorKind};
+use crate::error::{PreDeliveryRetry, ProviderError, ProviderErrorKind};
 use crate::event::{EventSequenceValidator, ProviderEvent};
+use crate::identity::ProviderKind;
 use crate::operation::Operation;
 use crate::policy::ClientApiKeyId;
-use crate::routing::{
-    ModelCapabilities, ModelPresentation, ProviderCandidate, ProviderKind, UpstreamModelId,
-};
+use crate::routing::{ModelCapabilities, ModelPresentation, ProviderCandidate, UpstreamModelId};
+use crate::upstream::OpaqueUpstreamValue;
 use crate::upstream::{UpstreamSendState, UpstreamTransport};
 
 /// Box 只出现在 Provider Registry 的统一 event envelope 边界。

backend/crates/gateway-core/src/error.rs

@@ -8,126 +8,11 @@ use bytes::Bytes;
 use thiserror::Error;
 
 use crate::event::{ProviderEvent, ProviderResponseHeader};
+pub use crate::upstream::OpaqueUpstreamValue;
 use crate::upstream::UpstreamSendState;
-
-/// 应用层标识不满足约束。
-#[derive(Debug, Clone, PartialEq, Eq, Error)]
-pub enum IdentifierError {
-    /// 标识为空。
-    #[error("identifier must not be empty")]
-    Empty,
-    /// 标识超过核心允许的字节数。
-    #[error("identifier exceeds {max_bytes} bytes")]
-    TooLong {
-        /// 最大字节数。
-        max_bytes: usize,
-    },
-    /// 标识使用了保留的系统前缀。
-    #[error("identifier uses the reserved system prefix")]
-    ReservedPrefix,
-    /// 标识包含控制字符。
-    #[error("identifier contains control characters")]
-    ControlCharacter,
-    /// 标识缺少规定的语义前缀。
-    #[error("identifier must start with `{expected}`")]
-    MissingPrefix {
-        /// 规定前缀。
-        expected: &'static str,
-    },
-    /// 标识不满足该领域值对象的固定格式。
-    #[error("identifier has an invalid format")]
-    InvalidFormat,
-}
-
-/// Operation 构造或校验失败。
-#[derive(Debug, Clone, PartialEq, Eq, Error)]
-pub enum OperationError {
-    /// 必填文本为空。
-    #[error("`{field}` must not be empty")]
-    EmptyField {
-        /// 字段名。
-        field: &'static str,
-    },
-    /// 数量字段为零。
-    #[error("`{field}` must be greater than zero")]
-    ZeroValue {
-        /// 字段名。
-        field: &'static str,
-    },
-    /// JSON 字段必须是 object。
-    #[error("`{field}` must be a JSON object")]
-    JsonObjectRequired {
-        /// 字段名。
-        field: &'static str,
-    },
-}
-
-/// 路由快照或 Route Plan 不满足不变量。
-#[derive(Debug, Clone, PartialEq, Eq, Error)]
-pub enum RoutingError {
-    /// 动态 Provider/model 标识无法构造。
-    #[error("routing identifier is invalid")]
-    InvalidIdentifier,
-    /// 配置 revision 必须为正数。
-    #[error("config revision must be greater than zero")]
-    InvalidRevision,
-    /// Restricted scope 必须保留至少一个持久化分组 binding。
-    #[error("account routing scope is invalid")]
-    InvalidAccountScope,
-    /// Key 的账号范围中当前没有任何可参与路由的账号。
-    #[error("account routing scope is empty")]
-    EmptyAccountScope,
-    /// 快照中存在重复实体。
-    #[error("duplicate {entity} `{id}`")]
-    DuplicateEntity {
-        /// 实体类型。
-        entity: &'static str,
-        /// 实体 ID。
-        id: String,
-    },
-    /// 实体引用不存在。
-    #[error("{entity} `{id}` was not found")]
-    NotFound {
-        /// 实体类型。
-        entity: &'static str,
-        /// 实体 ID。
-        id: String,
-    },
-    /// 固定平台内没有可执行本次请求的 Provider。
-    #[error("no provider can execute model `{model}`")]
-    NoCapableProvider {
-        /// 客户端提交的模型名称。
-        model: String,
-    },
-    /// 固定 Provider 的原生端点当前不可执行。
-    #[error("provider endpoint `{provider}` is unavailable")]
-    NoCapableProviderEndpoint {
-        /// adapter 已绑定的 Provider。
-        provider: String,
-    },
-}
-
-/// 调用方策略不满足约束或拒绝请求。
-#[derive(Debug, Clone, PartialEq, Eq, Error)]
-pub enum PolicyError {
-    /// 请求超过调用方策略。
-    #[error("request was denied by caller policy: {reason}")]
-    Denied {
-        /// 稳定拒绝原因。
-        reason: &'static str,
-    },
-}
-
-/// 用量或价格估算不满足事实约束。
-#[derive(Debug, Clone, PartialEq, Eq, Error)]
-pub enum MeteringError {
-    /// 十进制定点值格式无效或超过 `numeric(20, 10)`。
-    #[error("decimal value must fit unsigned numeric(20, 10)")]
-    InvalidDecimal,
-    /// 货币代码无效。
-    #[error("currency must be a three-letter uppercase ASCII code")]
-    InvalidCurrency,
-}
+pub use crate::validation::{
+    IdentifierError, MeteringError, OperationError, PolicyError, RoutingError,
+};
 
 /// 跨 Provider 稳定的上游失败分类。
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
@@ -231,33 +116,6 @@ pub enum PreDeliveryRetry {
     SameAccountTransportFallback,
 }
 
-/// Adapter 已明确分类为非 bearer 的不透明上游值。
-///
-/// Core 不解释或校验内容，只通过自定义 [`Debug`] 避免它意外进入日志。协议 adapter
-/// 可在原客户端响应中读取原值。
-#[derive(Clone, PartialEq, Eq, Hash)]
-pub struct OpaqueUpstreamValue(String);
-
-impl OpaqueUpstreamValue {
-    /// 原样保存上游值。
-    #[must_use]
-    pub fn new(value: impl Into<String>) -> Self {
-        Self(value.into())
-    }
-
-    /// 返回已经由 adapter 安全分类的原值。
-    #[must_use]
-    pub fn as_str(&self) -> &str {
-        &self.0
-    }
-}
-
-impl fmt::Debug for OpaqueUpstreamValue {
-    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
-        formatter.write_str("OpaqueUpstreamValue(<redacted-from-Debug>)")
-    }
-}
-
 /// 已由 Provider 从结构化上游错误中提取、仅供原客户端协议展示的错误详情。
 ///
 /// 它不是诊断日志或持久化错误信息：[`Debug`] 会隐藏全部值，调用方应只在
@@ -623,13 +481,21 @@ impl ProviderError {
             .get_or_insert_with(|| Box::new(ProviderErrorFailureObservation::default()))
     }
 
-    /// 标记 Provider 已证明本次拒绝没有执行生成，可在下游提交前重放。
+    /// 标记 Provider 已证明本次失败可在下游提交前安全重放。
+    ///
+    /// 证明可以来自明确的“未执行”拒绝，也可以来自同 attempt 已交付、可恢复
+    /// 同一业务请求的状态检查点；后者仍须由 Core 的同账号恢复门禁单独校验。
     #[must_use]
     pub const fn with_replay_safe(mut self) -> Self {
         self.replay_safe = true;
         self
     }
 
+    /// 原地标记 Provider 已证明本次失败可从已交付的状态检查点安全恢复。
+    pub const fn set_replay_safe(&mut self) {
+        self.replay_safe = true;
+    }
+
     /// 允许 Core 仅在客户端尚未收到任何事件时执行一次受预算约束的换号恢复。
     ///
     /// 该标记不证明上游未执行，也不等价于 [`Self::with_replay_safe`]；它只表达
@@ -1234,29 +1100,3 @@ impl StoreError {
         self.kind
     }
 }
-
-pub(crate) fn validate_text(
-    value: &str,
-    max_bytes: usize,
-    reject_reserved_prefix: bool,
-    required_prefix: Option<&'static str>,
-) -> Result<(), IdentifierError> {
-    if value.is_empty() {
-        return Err(IdentifierError::Empty);
-    }
-    if value.len() > max_bytes {
-        return Err(IdentifierError::TooLong { max_bytes });
-    }
-    if value.chars().any(char::is_control) {
-        return Err(IdentifierError::ControlCharacter);
-    }
-    if reject_reserved_prefix && value.starts_with("__") {
-        return Err(IdentifierError::ReservedPrefix);
-    }
-    if let Some(prefix) = required_prefix
-        && !value.starts_with(prefix)
-    {
-        return Err(IdentifierError::MissingPrefix { expected: prefix });
-    }
-    Ok(())
-}

backend/crates/gateway-core/src/event.rs

@@ -7,10 +7,11 @@ use bytes::Bytes;
 use serde_json::Value;
 use thiserror::Error;
 
-use crate::error::{IdentifierError, OpaqueUpstreamValue, validate_text};
 use crate::metering::{CalculatedCost, ProviderReportedCost, Usage};
 use crate::operation::ProviderSessionState;
+use crate::upstream::OpaqueUpstreamValue;
 use crate::upstream::UpstreamTransport;
+use crate::validation::{IdentifierError, validate_text};
 
 /// 一次响应的稳定元数据。
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -763,7 +764,8 @@ impl ProviderEvent {
         }
     }
 
-    /// 把本事件标记为 Provider 连接内状态的提交边界。
+    /// 附加 Provider 私有状态检查点；Core 可将其用于同请求恢复，协议连接可在
+    /// terminal 事件上持有它供下一轮使用。
     pub fn attach_session_update(&mut self, state: ProviderSessionState) {
         self.session_update = Some(Box::new(state));
     }

backend/crates/gateway-core/src/identity.rs

@@ -0,0 +1,28 @@
+//! 编译期 Provider 身份；账号与路由共同使用，不依赖路由计划。
+
+use crate::validation::{IdentifierError, validate_text};
+use std::fmt;
+
+/// 编译进二进制的 Provider adapter slug。
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct ProviderKind(String);
+
+impl ProviderKind {
+    /// 校验 Provider slug。
+    pub fn new(value: impl Into<String>) -> Result<Self, IdentifierError> {
+        let value = value.into();
+        validate_text(&value, 64, true, None)?;
+        Ok(Self(value))
+    }
+
+    #[must_use]
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+}
+
+impl fmt::Display for ProviderKind {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter.write_str(&self.0)
+    }
+}

backend/crates/gateway-core/src/lib.rs

@@ -4,10 +4,12 @@
 //! 具体客户端协议和具体 Provider 都通过外层 adapter 接入。
 
 pub mod account;
+pub mod diagnostics;
 pub mod engine;
 pub mod error;
 pub mod event;
 pub mod health;
+pub mod identity;
 pub mod lifecycle;
 pub mod metering;
 pub mod operation;
@@ -17,6 +19,7 @@ pub mod routing;
 pub mod runtime;
 pub mod task;
 pub mod upstream;
+pub mod validation;
 
 use std::sync::Arc;
 use std::time::SystemTime;

backend/crates/gateway-core/src/metering/mod.rs

@@ -3,7 +3,7 @@
 use std::fmt;
 use std::str::FromStr;
 
-use crate::error::MeteringError;
+use crate::validation::MeteringError;
 
 const DECIMAL_SCALE: u128 = 10_000_000_000;
 const MAX_SCALED_DECIMAL: u128 = 99_999_999_999_999_999_999;

backend/crates/gateway-core/src/operation/mod.rs

@@ -12,7 +12,7 @@ use bytes::Bytes;
 use serde::{Deserialize, Serialize};
 use serde_json::{Map, Value};
 
-use crate::error::{OperationError, validate_text};
+use crate::validation::{OperationError, validate_text};
 
 /// 网关支持的稳定 operation 分类。
 #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]

backend/crates/gateway-core/src/policy/mod.rs

@@ -11,8 +11,8 @@ pub use client_version::{
 use std::fmt;
 use std::sync::Arc;
 
-use crate::error::{IdentifierError, PolicyError, validate_text};
-use crate::routing::FrozenAccountScope;
+use crate::account::scope::FrozenAccountScope;
+use crate::validation::{IdentifierError, PolicyError, validate_text};
 
 /// `client_api_keys.id` 的核心值对象。
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]

backend/crates/gateway-core/src/provider_ports/mod.rs

@@ -12,9 +12,10 @@ use crate::account::{
     AccountFeedbackStats, AccountRuntimeSignals, CredentialRevision, CredentialState,
     OpaqueProviderData, ProviderAccountId, ProviderAccountStore,
 };
-use crate::error::{IdentifierError, validate_text};
+use crate::identity::ProviderKind;
 use crate::policy::ClientApiKeyId;
-use crate::routing::{ProviderKind, UpstreamModelId};
+use crate::routing::UpstreamModelId;
+use crate::validation::{IdentifierError, validate_text};
 
 const MAX_PENDING_FLOW_TTL: Duration = Duration::from_secs(30 * 60);
 

backend/crates/gateway-core/src/routing/mod.rs

@@ -2,6 +2,11 @@
 
 pub mod snapshot;
 
+pub use crate::account::scope::{
+    AccountGroupId, AccountRoutingScopeKind, AccountRoutingSnapshot, ClientRoutingScope,
+    FrozenAccountScope, RoutingGroupSnapshot, RuntimeAccount, RuntimeAccountDirectory,
+};
+pub use crate::identity::ProviderKind;
 pub use snapshot::RuntimeSnapshot;
 
 use std::collections::{BTreeMap, BTreeSet};
@@ -9,317 +14,12 @@ use std::fmt;
 use std::num::{NonZeroU32, NonZeroU64};
 use std::sync::Arc;
 
-use crate::account::{AccountSelectionPolicy, ProviderAccountId};
-use crate::error::{IdentifierError, RoutingError, validate_text};
+use crate::account::AccountSelectionPolicy;
 use crate::operation::{CapabilityRequirements, Feature, OperationKind};
+use crate::validation::{IdentifierError, RoutingError, validate_text};
 
 const MAX_REQUEST_ATTEMPTS: u32 = 32;
 
-/// `account_groups.id` 的核心值对象。
-#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
-pub struct AccountGroupId(String);
-
-impl AccountGroupId {
-    /// 校验并创建账号分组 ID。
-    pub fn new(value: impl Into<String>) -> Result<Self, IdentifierError> {
-        let value = value.into();
-        let Some(suffix) = value.strip_prefix("grp_") else {
-            return Err(IdentifierError::MissingPrefix { expected: "grp_" });
-        };
-        if suffix.len() != 32
-            || !suffix
-                .bytes()
-                .all(|byte| byte.is_ascii_digit() || (b'a'..=b'f').contains(&byte))
-        {
-            return Err(IdentifierError::InvalidFormat);
-        }
-        Ok(Self(value))
-    }
-
-    #[must_use]
-    pub fn as_str(&self) -> &str {
-        &self.0
-    }
-}
-
-impl fmt::Display for AccountGroupId {
-    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
-        formatter.write_str(&self.0)
-    }
-}
-
-/// 快照中一个账号的 Provider 与分组归属。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct RuntimeAccount {
-    provider_kind: ProviderKind,
-    group_ids: Arc<BTreeSet<AccountGroupId>>,
-}
-
-impl RuntimeAccount {
-    #[must_use]
-    pub fn new(provider_kind: ProviderKind, group_ids: BTreeSet<AccountGroupId>) -> Self {
-        Self {
-            provider_kind,
-            group_ids: Arc::new(group_ids),
-        }
-    }
-
-    #[must_use]
-    pub const fn provider_kind(&self) -> &ProviderKind {
-        &self.provider_kind
-    }
-
-    #[must_use]
-    pub fn group_ids(&self) -> &BTreeSet<AccountGroupId> {
-        &self.group_ids
-    }
-}
-
-/// 全快照共享的账号、Provider 与分组反向索引。
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-pub struct RuntimeAccountDirectory {
-    accounts: BTreeMap<ProviderAccountId, RuntimeAccount>,
-    providers_with_accounts: BTreeSet<ProviderKind>,
-    providers_by_group: BTreeMap<AccountGroupId, BTreeSet<ProviderKind>>,
-}
-
-impl RuntimeAccountDirectory {
-    #[must_use]
-    pub fn new(accounts: BTreeMap<ProviderAccountId, RuntimeAccount>) -> Self {
-        let mut providers_with_accounts = BTreeSet::new();
-        let mut providers_by_group = BTreeMap::<AccountGroupId, BTreeSet<ProviderKind>>::new();
-        for account in accounts.values() {
-            providers_with_accounts.insert(account.provider_kind.clone());
-            for group_id in account.group_ids.iter() {
-                providers_by_group
-                    .entry(group_id.clone())
-                    .or_default()
-                    .insert(account.provider_kind.clone());
-            }
-        }
-        Self {
-            accounts,
-            providers_with_accounts,
-            providers_by_group,
-        }
-    }
-
-    #[must_use]
-    pub fn account(&self, account_id: &ProviderAccountId) -> Option<&RuntimeAccount> {
-        self.accounts.get(account_id)
-    }
-
-    #[must_use]
-    pub fn providers_with_accounts(&self) -> &BTreeSet<ProviderKind> {
-        &self.providers_with_accounts
-    }
-
-    #[must_use]
-    pub fn providers_for_groups<'a>(
-        &self,
-        group_ids: impl IntoIterator<Item = &'a AccountGroupId>,
-    ) -> BTreeSet<ProviderKind> {
-        group_ids
-            .into_iter()
-            .filter_map(|group_id| self.providers_by_group.get(group_id))
-            .flatten()
-            .cloned()
-            .collect()
-    }
-}
-
-/// 历史请求保存的账号范围种类。
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum AccountRoutingScopeKind {
-    All,
-    Groups,
-}
-
-impl AccountRoutingScopeKind {
-    #[must_use]
-    pub const fn as_str(self) -> &'static str {
-        match self {
-            Self::All => "all",
-            Self::Groups => "groups",
-        }
-    }
-}
-
-/// 请求开始时冻结的分组 ID 与名称。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct RoutingGroupSnapshot {
-    id: AccountGroupId,
-    name: String,
-}
-
-impl RoutingGroupSnapshot {
-    #[must_use]
-    pub fn new(id: AccountGroupId, name: String) -> Self {
-        Self { id, name }
-    }
-
-    #[must_use]
-    pub const fn id(&self) -> &AccountGroupId {
-        &self.id
-    }
-
-    #[must_use]
-    pub fn name(&self) -> &str {
-        &self.name
-    }
-}
-
-/// 请求历史所需的完整、稳定账号范围快照。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct AccountRoutingSnapshot {
-    kind: AccountRoutingScopeKind,
-    groups: Arc<[RoutingGroupSnapshot]>,
-}
-
-impl AccountRoutingSnapshot {
-    #[must_use]
-    pub fn all() -> Self {
-        Self {
-            kind: AccountRoutingScopeKind::All,
-            groups: Arc::from([]),
-        }
-    }
-
-    #[must_use]
-    pub fn groups(groups: Vec<RoutingGroupSnapshot>) -> Self {
-        Self {
-            kind: AccountRoutingScopeKind::Groups,
-            groups: Arc::from(groups),
-        }
-    }
-
-    #[must_use]
-    pub const fn kind(&self) -> AccountRoutingScopeKind {
-        self.kind
-    }
-
-    #[must_use]
-    pub fn groups_snapshot(&self) -> &[RoutingGroupSnapshot] {
-        &self.groups
-    }
-}
-
-/// Key 持久 binding 编译出的账号权限。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub enum ClientRoutingScope {
-    AllAccounts,
-    Restricted {
-        bound_groups: Arc<[RoutingGroupSnapshot]>,
-        enabled_group_ids: Arc<BTreeSet<AccountGroupId>>,
-        provider_kinds: Arc<BTreeSet<ProviderKind>>,
-    },
-}
-
-impl ClientRoutingScope {
-    #[must_use]
-    pub fn all_accounts() -> Self {
-        Self::AllAccounts
-    }
-
-    pub fn restricted(
-        bound_groups: Vec<RoutingGroupSnapshot>,
-        enabled_group_ids: BTreeSet<AccountGroupId>,
-        provider_kinds: BTreeSet<ProviderKind>,
-    ) -> Result<Self, RoutingError> {
-        if bound_groups.is_empty() {
-            return Err(RoutingError::InvalidAccountScope);
-        }
-        Ok(Self::Restricted {
-            bound_groups: Arc::from(bound_groups),
-            enabled_group_ids: Arc::new(enabled_group_ids),
-            provider_kinds: Arc::new(provider_kinds),
-        })
-    }
-}
-
-/// 一次认证随 RuntimeSnapshot 冻结的账号目录与 Key 权限。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct FrozenAccountScope {
-    directory: Arc<RuntimeAccountDirectory>,
-    client_scope: ClientRoutingScope,
-}
-
-impl FrozenAccountScope {
-    #[must_use]
-    pub const fn new(
-        directory: Arc<RuntimeAccountDirectory>,
-        client_scope: ClientRoutingScope,
-    ) -> Self {
-        Self {
-            directory,
-            client_scope,
-        }
-    }
-
-    #[must_use]
-    pub fn allows(&self, account_id: &ProviderAccountId) -> bool {
-        let Some(account) = self.directory.account(account_id) else {
-            return false;
-        };
-        match &self.client_scope {
-            ClientRoutingScope::AllAccounts => true,
-            ClientRoutingScope::Restricted {
-                enabled_group_ids, ..
-            } => account
-                .group_ids()
-                .iter()
-                .any(|group_id| enabled_group_ids.contains(group_id)),
-        }
-    }
-
-    #[must_use]
-    pub fn provider_kinds(&self) -> &BTreeSet<ProviderKind> {
-        match &self.client_scope {
-            ClientRoutingScope::AllAccounts => self.directory.providers_with_accounts(),
-            ClientRoutingScope::Restricted { provider_kinds, .. } => provider_kinds,
-        }
-    }
-
-    #[must_use]
-    pub fn routing_snapshot(&self) -> AccountRoutingSnapshot {
-        match &self.client_scope {
-            ClientRoutingScope::AllAccounts => AccountRoutingSnapshot::all(),
-            ClientRoutingScope::Restricted { bound_groups, .. } => {
-                AccountRoutingSnapshot::groups(bound_groups.to_vec())
-            }
-        }
-    }
-
-    #[must_use]
-    pub const fn directory(&self) -> &Arc<RuntimeAccountDirectory> {
-        &self.directory
-    }
-}
-
-/// 编译进二进制的 Provider adapter slug。
-#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
-pub struct ProviderKind(String);
-
-impl ProviderKind {
-    /// 校验 Provider slug。
-    pub fn new(value: impl Into<String>) -> Result<Self, IdentifierError> {
-        let value = value.into();
-        validate_text(&value, 64, true, None)?;
-        Ok(Self(value))
-    }
-
-    #[must_use]
-    pub fn as_str(&self) -> &str {
-        &self.0
-    }
-}
-
-impl fmt::Display for ProviderKind {
-    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
-        formatter.write_str(&self.0)
-    }
-}
-
 /// 客户端请求中的模型名称。
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
 pub struct PublicModelId(String);

backend/crates/gateway-core/src/routing/snapshot.rs

@@ -9,12 +9,12 @@ use futures::future::BoxFuture;
 
 use crate::account::{AccountSelectionPolicy, ProviderAccountId, RotationStrategy};
 use crate::engine::provider::{ProviderCatalogGeneration, ProviderRegistry};
-use crate::error::RoutingError;
 use crate::operation::Operation;
 use crate::policy::{
     ClientApiKeyId, ClientPolicy, CodexClientMinVersions, CodexClientVersion,
     PlaintextClientApiKey, RateLimits,
 };
+use crate::validation::RoutingError;
 
 use super::{
     AccountGroupId, ClientRoutingScope, ConfigRevision, FrozenAccountScope, ModelCapabilities,

backend/crates/gateway-core/src/runtime/mod.rs

@@ -11,11 +11,12 @@ use futures_timer::Delay;
 
 use crate::engine::provider::ProviderCatalogGeneration;
 use crate::health::{HealthProbe, HealthState};
+use crate::identity::ProviderKind;
 use crate::lifecycle::CancellationToken;
+use crate::routing::ConfigRevision;
 use crate::routing::snapshot::{
     RuntimeSnapshot, RuntimeSnapshotCompileError, RuntimeSnapshotCompiler, SnapshotStorePort,
 };
-use crate::routing::{ConfigRevision, ProviderKind};
 use crate::task::{
     DaemonRestartPolicy, DaemonTask, ScheduledTask, WorkerContribution, WorkerCycleContext,
     WorkerDefinitionError, WorkerId, WorkerKind, WorkerRegistration, WorkerRunnable,

backend/crates/gateway-core/src/upstream.rs

@@ -1,6 +1,8 @@
 //! 上游调用跨 Engine、Event、Error 与 Provider 共享的中立边界事实。
 
-use crate::error::{IdentifierError, validate_text};
+use std::fmt;
+
+use crate::validation::{IdentifierError, validate_text};
 
 /// 上游是否可能已经收到业务 payload。
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
@@ -43,3 +45,30 @@ impl UpstreamTransport {
         &self.0
     }
 }
+
+/// Adapter 已明确分类为非 bearer 的不透明上游值。
+///
+/// Core 不解释或校验内容，只通过自定义 [`Debug`] 避免它意外进入日志。协议 adapter
+/// 可在原客户端响应中读取原值。
+#[derive(Clone, PartialEq, Eq, Hash)]
+pub struct OpaqueUpstreamValue(String);
+
+impl OpaqueUpstreamValue {
+    /// 原样保存上游值。
+    #[must_use]
+    pub fn new(value: impl Into<String>) -> Self {
+        Self(value.into())
+    }
+
+    /// 返回已经由 adapter 安全分类的原值。
+    #[must_use]
+    pub fn as_str(&self) -> &str {
+        &self.0
+    }
+}
+
+impl fmt::Debug for OpaqueUpstreamValue {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter.write_str("OpaqueUpstreamValue(<redacted-from-Debug>)")
+    }
+}

backend/crates/gateway-core/src/validation.rs

@@ -0,0 +1,148 @@
+//! 领域值对象的纯校验错误与文本约束，不依赖执行事件或路由状态。
+
+use thiserror::Error;
+
+/// 应用层标识不满足约束。
+#[derive(Debug, Clone, PartialEq, Eq, Error)]
+pub enum IdentifierError {
+    /// 标识为空。
+    #[error("identifier must not be empty")]
+    Empty,
+    /// 标识超过核心允许的字节数。
+    #[error("identifier exceeds {max_bytes} bytes")]
+    TooLong {
+        /// 最大字节数。
+        max_bytes: usize,
+    },
+    /// 标识使用了保留的系统前缀。
+    #[error("identifier uses the reserved system prefix")]
+    ReservedPrefix,
+    /// 标识包含控制字符。
+    #[error("identifier contains control characters")]
+    ControlCharacter,
+    /// 标识缺少规定的语义前缀。
+    #[error("identifier must start with `{expected}`")]
+    MissingPrefix {
+        /// 规定前缀。
+        expected: &'static str,
+    },
+    /// 标识不满足该领域值对象的固定格式。
+    #[error("identifier has an invalid format")]
+    InvalidFormat,
+}
+
+/// Operation 构造或校验失败。
+#[derive(Debug, Clone, PartialEq, Eq, Error)]
+pub enum OperationError {
+    /// 必填文本为空。
+    #[error("`{field}` must not be empty")]
+    EmptyField {
+        /// 字段名。
+        field: &'static str,
+    },
+    /// 数量字段为零。
+    #[error("`{field}` must be greater than zero")]
+    ZeroValue {
+        /// 字段名。
+        field: &'static str,
+    },
+    /// JSON 字段必须是 object。
+    #[error("`{field}` must be a JSON object")]
+    JsonObjectRequired {
+        /// 字段名。
+        field: &'static str,
+    },
+}
+
+/// 路由快照或 Route Plan 不满足不变量。
+#[derive(Debug, Clone, PartialEq, Eq, Error)]
+pub enum RoutingError {
+    /// 动态 Provider/model 标识无法构造。
+    #[error("routing identifier is invalid")]
+    InvalidIdentifier,
+    /// 配置 revision 必须为正数。
+    #[error("config revision must be greater than zero")]
+    InvalidRevision,
+    /// Restricted scope 必须保留至少一个持久化分组 binding。
+    #[error("account routing scope is invalid")]
+    InvalidAccountScope,
+    /// Key 的账号范围中当前没有任何可参与路由的账号。
+    #[error("account routing scope is empty")]
+    EmptyAccountScope,
+    /// 快照中存在重复实体。
+    #[error("duplicate {entity} `{id}`")]
+    DuplicateEntity {
+        /// 实体类型。
+        entity: &'static str,
+        /// 实体 ID。
+        id: String,
+    },
+    /// 实体引用不存在。
+    #[error("{entity} `{id}` was not found")]
+    NotFound {
+        /// 实体类型。
+        entity: &'static str,
+        /// 实体 ID。
+        id: String,
+    },
+    /// 固定平台内没有可执行本次请求的 Provider。
+    #[error("no provider can execute model `{model}`")]
+    NoCapableProvider {
+        /// 客户端提交的模型名称。
+        model: String,
+    },
+    /// 固定 Provider 的原生端点当前不可执行。
+    #[error("provider endpoint `{provider}` is unavailable")]
+    NoCapableProviderEndpoint {
+        /// adapter 已绑定的 Provider。
+        provider: String,
+    },
+}
+
+/// 调用方策略不满足约束或拒绝请求。
+#[derive(Debug, Clone, PartialEq, Eq, Error)]
+pub enum PolicyError {
+    /// 请求超过调用方策略。
+    #[error("request was denied by caller policy: {reason}")]
+    Denied {
+        /// 稳定拒绝原因。
+        reason: &'static str,
+    },
+}
+
+/// 用量或价格估算不满足事实约束。
+#[derive(Debug, Clone, PartialEq, Eq, Error)]
+pub enum MeteringError {
+    /// 十进制定点值格式无效或超过 `numeric(20, 10)`。
+    #[error("decimal value must fit unsigned numeric(20, 10)")]
+    InvalidDecimal,
+    /// 货币代码无效。
+    #[error("currency must be a three-letter uppercase ASCII code")]
+    InvalidCurrency,
+}
+
+pub(crate) fn validate_text(
+    value: &str,
+    max_bytes: usize,
+    reject_reserved_prefix: bool,
+    required_prefix: Option<&'static str>,
+) -> Result<(), IdentifierError> {
+    if value.is_empty() {
+        return Err(IdentifierError::Empty);
+    }
+    if value.len() > max_bytes {
+        return Err(IdentifierError::TooLong { max_bytes });
+    }
+    if value.chars().any(char::is_control) {
+        return Err(IdentifierError::ControlCharacter);
+    }
+    if reject_reserved_prefix && value.starts_with("__") {
+        return Err(IdentifierError::ReservedPrefix);
+    }
+    if let Some(prefix) = required_prefix
+        && !value.starts_with(prefix)
+    {
+        return Err(IdentifierError::MissingPrefix { expected: prefix });
+    }
+    Ok(())
+}

backend/crates/gateway-core/tests/account/mod.rs

@@ -100,6 +100,7 @@ fn context(strategy: RotationStrategy) -> AccountSelectionContext {
         now: SystemTime::now(),
         excluded_accounts: BTreeSet::new(),
         preferred_account: None,
+        preferred_account_overrides_weight: false,
         round_robin_cursor: 0,
         eligibility: AccountEligibilityPolicy::Enforce,
         account_scope: None,
@@ -417,6 +418,84 @@ fn selector_should_not_let_soft_affinity_bypass_a_higher_weight_account() {
     );
 }
 
+#[test]
+fn selector_should_prioritize_affinity_over_weight_when_requested() {
+    let candidates = [
+        weighted_candidate("acct_preferred", 1, 0),
+        weighted_candidate("acct_higher", 100, 0),
+    ];
+    for strategy in [
+        RotationStrategy::Smart,
+        RotationStrategy::QuotaResetPriority,
+        RotationStrategy::RoundRobin,
+        RotationStrategy::Sticky,
+    ] {
+        let mut selection = context(strategy);
+        selection.preferred_account = Some(candidates[0].account.id().clone());
+        selection.preferred_account_overrides_weight = true;
+        let selected = AccountSelector
+            .select(&candidates, &selection)
+            .expect("preferred account");
+        assert_eq!(
+            (
+                selected.candidate().account.id().as_str(),
+                selected.preferred()
+            ),
+            ("acct_preferred", PreferredAccountSelection::Hit),
+            "{strategy:?}"
+        );
+    }
+}
+
+#[test]
+fn highest_priority_affinity_should_still_obey_existing_scheduling_constraints() {
+    for blocker in [
+        AccountSchedulingBlocker::Excluded,
+        AccountSchedulingBlocker::ConcurrencyLimit,
+        AccountSchedulingBlocker::RequestInterval,
+        AccountSchedulingBlocker::LocalAvailability,
+    ] {
+        let mut candidates = [
+            weighted_candidate("acct_preferred", 1, 0),
+            weighted_candidate("acct_higher", 100, 0),
+        ];
+        let mut selection = context(RotationStrategy::Smart);
+        selection.preferred_account = Some(candidates[0].account.id().clone());
+        selection.preferred_account_overrides_weight = true;
+        match blocker {
+            AccountSchedulingBlocker::Excluded => {
+                selection
+                    .excluded_accounts
+                    .insert(candidates[0].account.id().clone());
+            }
+            AccountSchedulingBlocker::ConcurrencyLimit => candidates[0].signals.in_flight = 3,
+            AccountSchedulingBlocker::RequestInterval => {
+                candidates[0].signals.last_started_at = Some(selection.now);
+                selection.policy = AccountSelectionPolicy::new(
+                    RotationStrategy::Smart,
+                    NonZeroU32::new(3).expect("limit"),
+                    Duration::from_secs(1),
+                );
+            }
+            AccountSchedulingBlocker::LocalAvailability => {
+                candidates[0].signals.rate_limited_until =
+                    Some(selection.now + Duration::from_secs(60))
+            }
+            _ => unreachable!("test cases"),
+        }
+        let selected = AccountSelector
+            .select(&candidates, &selection)
+            .expect("fallback account");
+        assert_eq!(
+            (
+                selected.candidate().account.id().as_str(),
+                selected.preferred()
+            ),
+            ("acct_higher", PreferredAccountSelection::Blocked(blocker))
+        );
+    }
+}
+
 #[test]
 fn selector_should_fall_back_when_a_higher_weight_account_reaches_its_override() {
     let mut saturated = weighted_candidate("acct_high_weight", 100, 1);
@@ -618,8 +697,8 @@ fn account_feedback_should_decay_failure_rate_after_success() {
 #[test]
 fn smart_selector_should_use_provider_quota_rank_after_load_is_equal() {
     let candidates = vec![
-        candidate("acct_low_quota", 0, Some(20)),
-        candidate("acct_high_quota", 0, Some(80)),
+        candidate("acct_low_quota", 0, Some(2_000)),
+        candidate("acct_high_quota", 0, Some(8_000)),
     ];
     let selected = AccountSelector
         .select(&candidates, &context(RotationStrategy::Smart))
@@ -664,10 +743,10 @@ fn quota_reset_selector_should_use_capacity_utilization_as_load_tiebreaker() {
 }
 
 #[test]
-fn smart_selector_should_prefer_known_quota_over_unknown_after_load_is_equal() {
+fn smart_selector_should_prefer_sufficient_known_quota_over_unknown_after_load_is_equal() {
     let candidates = vec![
         candidate("acct_unknown", 0, None),
-        candidate("acct_known", 0, Some(1)),
+        candidate("acct_known", 0, Some(8_000)),
     ];
     let selected = AccountSelector
         .select(&candidates, &context(RotationStrategy::Smart))

backend/crates/gateway-core/tests/account/selection.rs

@@ -1,8 +1,13 @@
 use std::time::{Duration, Instant};
 
-use gateway_core::account::{AccountAttemptFeedback, AccountFeedbackStats, ProviderAccountId};
+use gateway_core::account::{
+    AccountAttemptFeedback, AccountCandidate, AccountFeedbackStats, AccountSelector,
+    ProviderAccountId, RotationStrategy,
+};
 use gateway_core::routing::ProviderKind;
 
+use super::{candidate, candidate_with_concurrency, context};
+
 const FAILURE_RATE_HALF_LIFE: Duration = Duration::from_secs(15 * 60);
 
 fn feedback_subject() -> (AccountFeedbackStats, ProviderKind, ProviderAccountId) {
@@ -96,3 +101,139 @@ fn concurrent_account_failures_should_not_lose_samples() {
 
     assert_eq!(failure_rate, Some(8_322));
 }
+fn smart_selection_ids(candidates: &[AccountCandidate]) -> Vec<&str> {
+    let mut selection = context(RotationStrategy::Smart);
+    (0..20)
+        .map(|cursor| {
+            selection.round_robin_cursor = cursor;
+            AccountSelector
+                .select(candidates, &selection)
+                .expect("candidate available")
+                .candidate()
+                .account
+                .id()
+                .as_str()
+        })
+        .collect()
+}
+
+#[test]
+fn smart_selector_should_rotate_despite_small_signal_differences() {
+    for signal in ["quota", "latency", "load", "failure"] {
+        let mut candidates = [
+            candidate_with_concurrency("acct_a", 0, 100),
+            candidate_with_concurrency("acct_b", 0, 100),
+        ];
+        match signal {
+            "quota" => {
+                candidates[0].signals.quota_remaining_rank = Some(600);
+                candidates[1].signals.quota_remaining_rank = Some(601);
+            }
+            "latency" => {
+                candidates[0].signals.first_output_latency_ms = Some(2_500);
+                candidates[1].signals.first_output_latency_ms = Some(2_501);
+            }
+            "load" => candidates[0].signals.in_flight = 1,
+            "failure" => candidates[0].signals.failure_rate_basis_points = Some(1),
+            _ => unreachable!(),
+        }
+        let selected = smart_selection_ids(&candidates);
+
+        assert_eq!(selected, ["acct_a", "acct_b"].repeat(10), "{signal}");
+    }
+}
+
+#[test]
+fn smart_selector_should_keep_rotation_order_when_nearby_scores_cross() {
+    let mut candidates = [
+        candidate("acct_a", 0, Some(8_000)),
+        candidate("acct_b", 0, Some(8_001)),
+        candidate("acct_worse", 0, Some(2_000)),
+    ];
+    let mut selection = context(RotationStrategy::Smart);
+    let mut selected = Vec::new();
+    for cursor in 0..20 {
+        selection.round_robin_cursor = cursor;
+        for candidate in &mut candidates {
+            match candidate.account.id().as_str() {
+                "acct_a" => candidate.signals.quota_remaining_rank = Some(8_000 + cursor % 2),
+                "acct_b" => candidate.signals.quota_remaining_rank = Some(8_001 - cursor % 2),
+                _ => {}
+            }
+        }
+        candidates.reverse();
+        selected.push(
+            AccountSelector
+                .select(&candidates, &selection)
+                .expect("candidate available")
+                .candidate()
+                .account
+                .id()
+                .as_str()
+                .to_owned(),
+        );
+    }
+
+    assert_eq!(selected, ["acct_a", "acct_b"].repeat(10));
+}
+
+#[test]
+fn smart_selector_should_only_rotate_among_candidates_close_to_the_best() {
+    let candidates = [
+        candidate("acct_a", 0, Some(10_000)),
+        candidate("acct_b", 0, Some(9_500)),
+        candidate("acct_c", 0, Some(9_000)),
+    ];
+
+    assert_eq!(
+        smart_selection_ids(&candidates),
+        ["acct_a", "acct_b"].repeat(10)
+    );
+}
+
+#[test]
+fn smart_selector_should_preserve_material_signal_advantages() {
+    for signal in ["quota", "latency", "load", "failure"] {
+        let mut candidates = [
+            candidate_with_concurrency("acct_a", 0, 10),
+            candidate_with_concurrency("acct_b", 0, 10),
+        ];
+        match signal {
+            "quota" => {
+                candidates[0].signals.quota_remaining_rank = Some(600);
+                candidates[1].signals.quota_remaining_rank = Some(2_600);
+            }
+            "latency" => {
+                candidates[0].signals.first_output_latency_ms = Some(20_000);
+                candidates[1].signals.first_output_latency_ms = Some(2_500);
+            }
+            "load" => candidates[0].signals.in_flight = 1,
+            "failure" => candidates[0].signals.failure_rate_basis_points = Some(2_000),
+            _ => unreachable!(),
+        }
+
+        assert_eq!(smart_selection_ids(&candidates), ["acct_b"; 20], "{signal}");
+    }
+}
+
+#[test]
+fn smart_selector_should_balance_actual_load_against_remaining_quota() {
+    let mut candidates = [
+        candidate_with_concurrency("acct_a", 0, 10),
+        candidate_with_concurrency("acct_b", 1, 10),
+    ];
+    candidates[0].signals.quota_remaining_rank = Some(600);
+    candidates[1].signals.quota_remaining_rank = Some(2_600);
+
+    assert_eq!(smart_selection_ids(&candidates), ["acct_b"; 20]);
+}
+
+#[test]
+fn smart_selector_should_treat_unknown_quota_as_neutral() {
+    let candidates = [
+        candidate("acct_known_low", 0, Some(600)),
+        candidate("acct_unknown", 0, None),
+    ];
+
+    assert_eq!(smart_selection_ids(&candidates), ["acct_unknown"; 20]);
+}

backend/crates/gateway-core/tests/diagnostics/capture.rs

@@ -0,0 +1,49 @@
+use gateway_core::diagnostics::{diagnostic_headers, diagnostic_json};
+use serde_json::json;
+
+#[test]
+fn credentials_and_content_never_enter_diagnostics_but_trace_ids_survive() {
+    let headers = diagnostic_headers([
+        ("Authorization", "Bearer never-persist-this"),
+        ("Set-Cookie", "session=never-persist-this"),
+        ("X-OAI-Request-ID", "upstream-123"),
+        ("x-codex-turn-state", "opaque-never-persist-this"),
+        ("cf-ray", "ray-one"),
+        ("cf-ray", "ray-two"),
+    ]);
+    assert_eq!(headers["x-oai-request-id"][0], "upstream-123");
+    assert_eq!(headers["cf-ray"].as_array().unwrap().len(), 2);
+    assert!(!headers.to_string().contains("never-persist-this"));
+    let content = diagnostic_json(&json!({
+        "type": "codex.response.metadata", "headers": headers,
+        "access_token": "never-persist-this", "refresh_token": "never-persist-this",
+        "response": {"output": [{"text": "private model output"}]},
+        "input": "private user prompt", "unknown": "private extension",
+    }));
+    assert_eq!(content["type"], "codex.response.metadata");
+    let text = content.to_string();
+    for secret in [
+        "never-persist-this",
+        "private model output",
+        "private user prompt",
+        "private extension",
+    ] {
+        assert!(!text.contains(secret), "leaked {secret}");
+    }
+}
+
+#[test]
+fn arbitrary_event_type_text_is_fingerprinted_instead_of_logged() {
+    let trace = gateway_core::diagnostics::TraceContext::new("req_type");
+    trace.capture(
+        "upstream.event",
+        br#"{"type":"secret@example.com","message":"private-content"}"#,
+    );
+    let snapshot = trace.snapshot().unwrap();
+    assert_eq!(
+        snapshot["events"][0]["data"]["eventType"],
+        serde_json::Value::Null
+    );
+    assert!(!snapshot.to_string().contains("secret@example.com"));
+    assert!(!snapshot.to_string().contains("private-content"));
+}

backend/crates/gateway-core/tests/diagnostics/mod.rs

@@ -0,0 +1,3 @@
+mod capture;
+mod stream;
+mod trace;

backend/crates/gateway-core/tests/diagnostics/stream.rs

@@ -0,0 +1,70 @@
+use gateway_core::diagnostics::{StreamCapture, StreamFormat, TraceContext};
+
+#[test]
+fn sse_capture_reassembles_every_chunk_boundary_and_preserves_unknown_events() {
+    let bytes = "event: future.metadata\r\ndata: {\"headers\":{\"x-request-id\":\"up-1\"},\"unknown\":\"中文\"}\r\n\r\ndata: {\"type\":\"response.completed\"}\n\n".as_bytes();
+    for boundary in 0..=bytes.len() {
+        let trace = TraceContext::new("req_sse");
+        let mut capture = StreamCapture::new(trace.clone(), StreamFormat::Sse);
+        capture.push(&bytes[..boundary]);
+        capture.push(&bytes[boundary..]);
+        capture.finish();
+        let snapshot = trace.snapshot().unwrap();
+        let events = snapshot["events"].as_array().unwrap();
+        assert_eq!(events.len(), 3, "boundary {boundary}");
+        assert_eq!(events[0]["data"]["eventType"], "future.metadata");
+        assert_eq!(
+            events[0]["data"]["metadata"]["headers"]["x-request-id"],
+            "up-1"
+        );
+        assert_eq!(events[1]["data"]["eventType"], "response.completed");
+        assert_eq!(events[2]["data"]["partialFrame"], false);
+    }
+}
+
+#[test]
+fn oversize_frames_report_a_gap_and_resume_at_the_next_frame() {
+    let trace = TraceContext::new("req_sse");
+    let mut capture = StreamCapture::new(trace.clone(), StreamFormat::Sse);
+    capture.push(b"data: ");
+    capture.push(&vec![b'x'; 130 * 1024]);
+    capture.push(b"\n\ndata: {\"type\":\"response.completed\"}\n\n");
+    capture.finish();
+    let snapshot = trace.snapshot().unwrap();
+    let events = snapshot["events"].as_array().unwrap();
+    assert_eq!(events[0]["stage"], "capture.gap");
+    assert_eq!(events[1]["data"]["eventType"], "response.completed");
+}
+
+#[test]
+fn dropping_a_partial_stream_is_distinct_from_upstream_eof() {
+    let trace = TraceContext::new("req_partial");
+    {
+        let mut capture = StreamCapture::new(trace.clone(), StreamFormat::Sse);
+        capture.push(b"data: {\"type\":");
+    }
+    let snapshot = trace.snapshot().unwrap();
+    assert_eq!(snapshot["events"][0]["stage"], "upstream.reader.dropped");
+    assert!(
+        snapshot["events"][0]["data"]["partialBytes"]
+            .as_u64()
+            .unwrap()
+            > 0
+    );
+}
+
+#[test]
+fn sse_accepts_bom_and_bare_carriage_return_even_when_split_into_single_bytes() {
+    let trace = TraceContext::new("req_cr");
+    let mut capture = StreamCapture::new(trace.clone(), StreamFormat::Sse);
+    for byte in b"\xef\xbb\xbfdata: {\"type\":\"codex.response.metadata\"}\r\r" {
+        capture.push(&[*byte]);
+    }
+    capture.finish();
+    let snapshot = trace.snapshot().unwrap();
+    assert_eq!(
+        snapshot["events"][0]["data"]["eventType"],
+        "codex.response.metadata"
+    );
+    assert_eq!(snapshot["events"][1]["data"]["partialFrame"], false);
+}

backend/crates/gateway-core/tests/diagnostics/trace.rs

@@ -0,0 +1,142 @@
+use gateway_core::diagnostics::TraceContext;
+use serde_json::json;
+
+#[test]
+fn bounded_history_preserves_start_failure_and_final_result() {
+    let trace = TraceContext::new("req_trace");
+    trace.record("request.started", json!({}));
+    for index in 0..400 {
+        trace.record(
+            "test.event",
+            json!({"index": index, "detail": "x".repeat(2000)}),
+        );
+    }
+    trace
+        .attempt(1)
+        .record("attempt.failed", json!({"code": "websocket_close_1000"}));
+    trace
+        .attempt(2)
+        .record("request.finished", json!({"outcome": "succeeded"}));
+    let snapshot = trace.snapshot().unwrap();
+    let events = snapshot["events"].as_array().unwrap();
+    assert_eq!(events[0]["stage"], "request.started");
+    assert_eq!(events[events.len() - 2]["attemptIndex"], 1);
+    assert_eq!(events.last().unwrap()["attemptIndex"], 2);
+    assert!(snapshot["droppedEvents"].as_u64().unwrap() > 0);
+    assert!(events.len() <= 128);
+    assert!(snapshot.to_string().len() < 64 * 1024);
+    let kept: u64 = events
+        .iter()
+        .map(|event| event["count"].as_u64().unwrap())
+        .sum();
+    assert_eq!(
+        kept + snapshot["droppedEvents"].as_u64().unwrap(),
+        snapshot["totalEvents"]
+    );
+}
+
+#[test]
+fn shared_attempts_and_exchanges_coalesce_delta_without_losing_metadata() {
+    let trace = TraceContext::new("req_trace");
+    let ws = trace.attempt(1).exchange("websocket");
+    ws.capture(
+        "upstream.event",
+        br#"{"type":"codex.response.metadata","headers":{"x-request-id":"up-1"}}"#,
+    );
+    trace
+        .attempt(1)
+        .wire_event("openai", Some("codex.response.metadata"));
+    for _ in 0..1000 {
+        ws.capture(
+            "upstream.event",
+            br#"{"type":"response.output_text.delta","delta":"secret"}"#,
+        );
+        trace
+            .attempt(1)
+            .wire_event("openai", Some("response.output_text.delta"));
+    }
+    let http = trace.attempt(1).exchange("http_sse");
+    http.capture("upstream.event", br#"{"type":"response.completed"}"#);
+    let snapshot = trace.snapshot().unwrap();
+    let events = snapshot["events"].as_array().unwrap();
+    assert_eq!(snapshot["droppedEvents"], 0);
+    assert_eq!(
+        events
+            .iter()
+            .find(|e| e["data"]["eventType"] == "response.output_text.delta")
+            .unwrap()["count"],
+        1000
+    );
+    assert_eq!(events.last().unwrap()["exchangeId"], 2);
+    assert!(
+        events
+            .iter()
+            .any(|e| e["data"]["metadata"]["headers"]["x-request-id"] == "up-1")
+    );
+    assert!(!snapshot.to_string().contains("secret"));
+    assert!(
+        events
+            .iter()
+            .filter(|event| event["stage"] == "provider.event")
+            .all(|event| event["data"].get("metadata").is_none())
+    );
+}
+
+#[test]
+fn header_capture_keeps_duplicates_but_excludes_credentials_from_timeline() {
+    let trace = TraceContext::new("req_headers");
+    trace.headers(
+        "upstream.response.headers",
+        json!({"status": 101}),
+        [
+            ("authorization", b"Bearer SECRET".as_slice()),
+            ("x-request-id", b"upstream-one".as_slice()),
+            ("x-request-id", b"upstream-two".as_slice()),
+            ("set-cookie", b"PRIVATE_COOKIE".as_slice()),
+        ],
+    );
+    let snapshot = trace.snapshot().unwrap();
+    assert_eq!(
+        snapshot["events"][0]["data"]["headers"]["x-request-id"],
+        json!(["upstream-one", "upstream-two"])
+    );
+    assert!(!snapshot.to_string().contains("SECRET"));
+    assert!(!snapshot.to_string().contains("PRIVATE_COOKIE"));
+}
+
+#[test]
+fn disabled_context_does_not_allocate_a_history() {
+    let trace = TraceContext::default().attempt(2).exchange("http");
+    trace.capture("upstream.event", b"not json");
+    assert!(trace.snapshot().is_none());
+}
+
+#[test]
+fn a_long_successful_retry_keeps_the_earlier_failure_ahead_of_routine_events() {
+    let trace = TraceContext::new("req_retry");
+    for _ in 0..12 {
+        trace.record("setup", json!({}));
+    }
+    trace
+        .attempt(1)
+        .record("attempt.failed", json!({"kind": "unavailable"}));
+    trace
+        .attempt(1)
+        .record("retry.decided", json!({"retryable": true}));
+    for index in 0..500 {
+        trace
+            .attempt(2)
+            .record("upstream.event", json!({"index": index}));
+    }
+    trace.record("request.finished", json!({"outcome": "succeeded"}));
+    let snapshot = trace.snapshot().unwrap();
+    let events = snapshot["events"].as_array().unwrap();
+    assert!(
+        events
+            .iter()
+            .any(|event| event["stage"] == "attempt.failed" && event["attemptIndex"] == 1)
+    );
+    assert!(events.iter().any(|event| event["stage"] == "retry.decided"));
+    assert_eq!(events.last().unwrap()["stage"], "request.finished");
+    assert!(events.len() <= 128);
+}

backend/crates/gateway-core/tests/engine/coordinator.rs

@@ -49,6 +49,7 @@ use serde_json::{Map, Value, json};
 
 #[derive(Debug, Clone, PartialEq, Eq)]
 struct FinalState {
+    diagnostic_trace_json: Option<String>,
     outcome: ExecutionOutcome,
     send_state: UpstreamSendState,
     attempt_count: u32,
@@ -207,6 +208,7 @@ impl ExecutionStore for FakeStore {
             .expect("store lock")
             .finalizations
             .push(FinalState {
+                diagnostic_trace_json: finalization.diagnostic_trace_json,
                 outcome: finalization.outcome,
                 send_state: finalization.send_state,
                 attempt_count: finalization.attempt_count,
@@ -270,6 +272,7 @@ enum Script {
 struct ScriptedProvider {
     scripts: Mutex<VecDeque<Script>>,
     contexts: Mutex<Vec<AttemptContext>>,
+    operations: Mutex<Vec<Operation>>,
 }
 
 impl ScriptedProvider {
@@ -277,6 +280,7 @@ impl ScriptedProvider {
         Self {
             scripts: Mutex::new(scripts.into()),
             contexts: Mutex::new(Vec::new()),
+            operations: Mutex::new(Vec::new()),
         }
     }
 }
@@ -302,6 +306,10 @@ impl Provider for ScriptedProvider {
         request: ProviderRequest,
         context: AttemptContext,
     ) -> Result<ProviderStream, ProviderError> {
+        self.operations
+            .lock()
+            .expect("operations lock")
+            .push(request.operation().clone());
         self.contexts.lock().expect("contexts lock").push(context);
         let script = self
             .scripts
@@ -1308,6 +1316,21 @@ fn discarded_attempt_observation_does_not_leak_into_retry_result() {
 
     let state = store.state.lock().expect("store lock");
     let finalization = &state.finalizations[0];
+    let trace: Value =
+        serde_json::from_str(finalization.diagnostic_trace_json.as_deref().unwrap()).unwrap();
+    let events = trace["events"].as_array().unwrap();
+    assert!(
+        events
+            .iter()
+            .any(|event| event["stage"] == "attempt.failed" && event["attemptIndex"] == 1)
+    );
+    assert!(
+        events
+            .iter()
+            .any(|event| event["stage"] == "attempt.started" && event["attemptIndex"] == 2)
+    );
+    assert_eq!(events.last().unwrap()["stage"], "request.finished");
+
     assert_eq!(finalization.upstream_transport.as_deref(), Some("http_sse"));
     assert_eq!(finalization.http_version.as_deref(), Some("HTTP/2"));
     assert_eq!(
@@ -2838,14 +2861,25 @@ fn ambiguous_send_state_stops_retry() {
 fn ambiguous_pre_delivery_retry_marker_does_not_rotate_account() {
     let operation = generate_operation();
     let route_plan = plan(&operation);
+    let checkpoint = ProviderSessionState::new(
+        "openai",
+        Map::from_iter([("checkpoint".to_owned(), json!("opaque-state"))]),
+    )
+    .expect("provider checkpoint");
+    let mut checkpoint_event = ProviderEvent::observation(ProviderResponseObservation::new(
+        UpstreamTransport::new("websocket").expect("transport"),
+    ));
+    checkpoint_event.attach_session_update(checkpoint);
     let (coordinator, store, provider) = coordinator(vec![
-        Script::Stream {
+        Script::ObservedStream {
             account_id: "acct_first",
-            items: vec![Err(ProviderError::new(
-                ProviderErrorKind::Transport,
-                UpstreamSendState::Ambiguous,
-            )
-            .with_pre_delivery_retry())],
+            items: vec![
+                Ok(checkpoint_event),
+                Err(
+                    ProviderError::new(ProviderErrorKind::Transport, UpstreamSendState::Ambiguous)
+                        .with_pre_delivery_retry(),
+                ),
+            ],
         },
         Script::Stream {
             account_id: "acct_second",
@@ -2874,6 +2908,57 @@ fn ambiguous_pre_delivery_retry_marker_does_not_rotate_account() {
     assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Failed);
 }
 
+#[test]
+fn ambiguous_transport_retry_is_rejected_even_with_provider_session_state() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let checkpoint = ProviderSessionState::new(
+        "openai",
+        Map::from_iter([("checkpoint".to_owned(), json!("opaque-state"))]),
+    )
+    .expect("provider checkpoint");
+    let mut checkpoint_event = ProviderEvent::observation(ProviderResponseObservation::new(
+        UpstreamTransport::new("websocket").expect("transport"),
+    ));
+    checkpoint_event.attach_session_update(checkpoint);
+    let retry_index = NonZeroU32::new(1).expect("retry index");
+    let (coordinator, store, provider) = coordinator(vec![
+        Script::ObservedStream {
+            account_id: "acct_first",
+            items: vec![
+                Ok(checkpoint_event),
+                Err(
+                    ProviderError::new(ProviderErrorKind::Transport, UpstreamSendState::Ambiguous)
+                        .with_replay_safe()
+                        .with_pre_delivery_transport_retry(retry_index, Duration::ZERO),
+                ),
+            ],
+        },
+        Script::Stream {
+            account_id: "acct_first",
+            items: complete_stream(None),
+        },
+    ]);
+
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        None,
+        None,
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+    let error =
+        block_on(session.collect_uncommitted()).expect_err("ambiguous send cannot be replayed");
+    assert!(matches!(error, EngineError::Provider(_)));
+    assert_eq!(provider.contexts.lock().expect("contexts lock").len(), 1);
+    let state = store.state.lock().expect("store lock");
+    assert_eq!(state.intermediate_failures, 0);
+    assert_eq!(state.finalizations[0].attempt_count, 1);
+    assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Failed);
+}
+
 #[test]
 fn pre_delivery_transport_fallback_retries_the_same_account_with_fallback_transport() {
     let operation = generate_operation();

backend/crates/gateway-core/tests/main.rs

@@ -1,4 +1,5 @@
 mod account;
+mod diagnostics;
 mod engine;
 mod error;
 mod event;

backend/crates/gateway-host/src/config.rs

@@ -118,6 +118,9 @@ pub struct LoggingConfig {
     pub level: String,
     pub stdout: bool,
     pub file: FileLoggingConfig,
+    /// 将完整请求原文写入独立诊断日志；默认关闭，因为内容包含凭据。
+    #[serde(default)]
+    pub request_dump: bool,
 }
 
 impl LoggingConfig {

backend/crates/gateway-host/src/logging.rs

@@ -6,16 +6,19 @@ use std::io::{self, Write};
 use std::path::{Path, PathBuf};
 
 use chrono::Utc;
-use tracing_appender::non_blocking::{NonBlockingBuilder, WorkerGuard};
+use tracing_appender::non_blocking::{NonBlocking, NonBlockingBuilder, WorkerGuard};
 use tracing_subscriber::EnvFilter;
 use tracing_subscriber::{Layer as _, layer::SubscriberExt as _, util::SubscriberInitExt as _};
 
-use crate::config::LoggingConfig;
+use crate::config::{FileLoggingConfig, LoggingConfig};
 
 const LOG_FILE_PREFIX: &str = "codex-proxy-rs";
 const OAUTH_RECOVERY_LOG_FILE_PREFIX: &str = "codex-proxy-rs-oath";
 /// OAuth 恢复记录含原始 AT/RT，文件日志开启时必须落到独立文件。
 const OAUTH_RECOVERY_LOG_TARGET: &str = "oauth_recovery";
+const REQUEST_DUMP_LOG_FILE_PREFIX: &str = "codex-proxy-rs-request-dump";
+/// 请求转储含完整凭据与正文，只允许在显式开启时写入独立文件。
+const REQUEST_DUMP_LOG_TARGET: &str = "request_dump";
 
 /// non-blocking 日志 writer 的进程级守卫。
 pub struct LogGuard {
@@ -39,6 +42,7 @@ pub fn initialize_logging(config: &LoggingConfig) -> Result<LogGuard, LogError>
     let directive = env::var("RUST_LOG").unwrap_or_else(|_| config.level.clone());
     let file_filter = application_file_filter(&directive)?;
     let recovery_file_filter = oauth_recovery_file_filter();
+    let request_dump_file_filter = request_dump_file_filter();
     let stdout_filter =
         EnvFilter::try_new(stdout_filter_directive(&directive, config.file.enabled))
             .map_err(|_| LogError::InvalidFilter)?;
@@ -52,37 +56,30 @@ pub fn initialize_logging(config: &LoggingConfig) -> Result<LogGuard, LogError>
         writer
     });
     let (file_writer, recovery_file_writer) = if config.file.enabled {
-        let maximum_bytes = config
-            .file
-            .max_file_size_mb
-            .checked_mul(1024 * 1024)
-            .ok_or(LogError::SizeOverflow)?;
-        let file_writer = RotatingLogWriter::open(
-            config.file.directory.clone(),
-            LOG_FILE_PREFIX,
-            maximum_bytes,
-            config.file.retention_days,
-            config.file.max_files,
-        )?;
-        let recovery_file_writer = RotatingLogWriter::open(
-            config.file.directory.clone(),
+        let (file_writer, file_guard) =
+            rotating_file_writer(&config.file, LOG_FILE_PREFIX, "gateway-log-file")?;
+        let (recovery_file_writer, recovery_file_guard) = rotating_file_writer(
+            &config.file,
             OAUTH_RECOVERY_LOG_FILE_PREFIX,
-            maximum_bytes,
-            config.file.retention_days,
-            config.file.max_files,
+            "gateway-log-oauth-recovery-file",
         )?;
-        let (file_writer, file_guard) = NonBlockingBuilder::default()
-            .thread_name("gateway-log-file")
-            .finish(file_writer);
-        let (recovery_file_writer, recovery_file_guard) = NonBlockingBuilder::default()
-            .thread_name("gateway-log-oauth-recovery-file")
-            .finish(recovery_file_writer);
         guards.push(file_guard);
         guards.push(recovery_file_guard);
         (Some(file_writer), Some(recovery_file_writer))
     } else {
         (None, None)
     };
+    let request_dump_file_writer = if config.request_dump {
+        let (writer, guard) = rotating_file_writer(
+            &config.file,
+            REQUEST_DUMP_LOG_FILE_PREFIX,
+            "gateway-log-request-dump-file",
+        )?;
+        guards.push(guard);
+        Some(writer)
+    } else {
+        None
+    };
 
     let stdout_layer = stdout_writer.map(|writer| {
         tracing_subscriber::fmt::layer()
@@ -118,10 +115,24 @@ pub fn initialize_logging(config: &LoggingConfig) -> Result<LogGuard, LogError>
             .with_span_list(true)
             .with_filter(recovery_file_filter)
     });
+    let request_dump_file_layer = request_dump_file_writer.map(|writer| {
+        tracing_subscriber::fmt::layer()
+            .json()
+            .with_writer(writer)
+            .with_target(true)
+            .with_file(true)
+            .with_line_number(true)
+            .with_thread_ids(true)
+            .with_thread_names(true)
+            .with_current_span(true)
+            .with_span_list(true)
+            .with_filter(request_dump_file_filter)
+    });
     tracing_subscriber::registry()
         .with(stdout_layer)
         .with(file_layer)
         .with(recovery_file_layer)
+        .with(request_dump_file_layer)
         .try_init()
         .map_err(|_| LogError::AlreadyInitialized)?;
     Ok(LogGuard { _writers: guards })
@@ -131,8 +142,12 @@ fn application_file_filter(directive: &str) -> Result<EnvFilter, LogError> {
     let recovery_directive = format!("{OAUTH_RECOVERY_LOG_TARGET}=off")
         .parse()
         .expect("static OAuth recovery log directive is valid");
+    let request_dump_directive = format!("{REQUEST_DUMP_LOG_TARGET}=off")
+        .parse()
+        .expect("static request dump log directive is valid");
     EnvFilter::try_new(directive)
         .map(|filter| filter.add_directive(recovery_directive))
+        .map(|filter| filter.add_directive(request_dump_directive))
         .map_err(|_| LogError::InvalidFilter)
 }
 
@@ -142,14 +157,41 @@ fn oauth_recovery_file_filter() -> EnvFilter {
         .expect("static OAuth recovery log filter is valid")
 }
 
+fn request_dump_file_filter() -> EnvFilter {
+    format!("off,{REQUEST_DUMP_LOG_TARGET}=info")
+        .parse()
+        .expect("static request dump log filter is valid")
+}
+
 fn stdout_filter_directive(directive: &str, persistent_log_enabled: bool) -> String {
     if !persistent_log_enabled {
-        directive.to_owned()
+        format!("{directive},{REQUEST_DUMP_LOG_TARGET}=off")
     } else {
         "off,gateway_startup=info".to_owned()
     }
 }
 
+fn rotating_file_writer(
+    config: &FileLoggingConfig,
+    file_prefix: &'static str,
+    thread_name: &'static str,
+) -> Result<(NonBlocking, WorkerGuard), LogError> {
+    let maximum_bytes = config
+        .max_file_size_mb
+        .checked_mul(1024 * 1024)
+        .ok_or(LogError::SizeOverflow)?;
+    let writer = RotatingLogWriter::open(
+        config.directory.clone(),
+        file_prefix,
+        maximum_bytes,
+        config.retention_days,
+        config.max_files,
+    )?;
+    Ok(NonBlockingBuilder::default()
+        .thread_name(thread_name)
+        .finish(writer))
+}
+
 struct RotatingLogWriter {
     directory: PathBuf,
     file_prefix: &'static str,

backend/crates/gateway-host/tests/config.rs

@@ -103,6 +103,7 @@ fn valid_config() -> HostConfig {
                 max_file_size_mb: 100,
                 max_files: 30,
             },
+            request_dump: false,
         },
         system_update,
         drain_timeout_seconds: 30,

backend/crates/gateway-host/tests/logging.rs

@@ -8,13 +8,18 @@ use gateway_host::system_update::SystemUpdateConfig;
 
 const LOG_DIRECTORY_ENV: &str = "CPR_LOGGING_TEST_DIRECTORY";
 const CHILD_PROCESS_ENV: &str = "CPR_LOGGING_TEST_CHILD";
+const REQUEST_DUMP_ENABLED_ENV: &str = "CPR_LOGGING_TEST_REQUEST_DUMP_ENABLED";
 const APPLICATION_LOG_FILE_PREFIX: &str = "codex-proxy-rs.";
 const OAUTH_RECOVERY_LOG_FILE_PREFIX: &str = "codex-proxy-rs-oath.";
+const REQUEST_DUMP_LOG_FILE_PREFIX: &str = "codex-proxy-rs-request-dump.";
 const APPLICATION_LOG_TARGET: &str = "logging_test_application";
 const APPLICATION_LOG_MARKER: &str = "application-file-filter-test";
 const OAUTH_RECOVERY_LOG_TARGET: &str = "oauth_recovery";
 const OAUTH_RECOVERY_LOG_MARKER: &str = "oauth-recovery-file-filter-test";
 const OAUTH_RECOVERY_PROVIDER: &str = "openai";
+const REQUEST_DUMP_LOG_TARGET: &str = "request_dump";
+const REQUEST_DUMP_LOG_MARKER: &str = "request-dump-file-filter-test";
+const REQUEST_DUMP_SECRET: &str = "unredacted-authorization-value";
 
 #[test]
 fn logging_requires_at_least_one_sink() {
@@ -34,6 +39,7 @@ fn logging_requires_at_least_one_sink() {
                 max_file_size_mb: 100,
                 max_files: 30,
             },
+            request_dump: false,
         },
         system_update: SystemUpdateConfig::default(),
         drain_timeout_seconds: 30,
@@ -48,11 +54,12 @@ fn logging_requires_at_least_one_sink() {
 }
 
 #[test]
-fn oauth_recovery_file_logging_is_separate_and_overrides_global_log_level() {
+fn sensitive_file_logging_is_separate_and_overrides_global_log_level() {
     if env::var_os(CHILD_PROCESS_ENV).is_some() {
-        write_oauth_recovery_log(PathBuf::from(
-            env::var_os(LOG_DIRECTORY_ENV).expect("child log directory"),
-        ));
+        write_sensitive_logs(
+            PathBuf::from(env::var_os(LOG_DIRECTORY_ENV).expect("child log directory")),
+            env::var(REQUEST_DUMP_ENABLED_ENV).is_ok_and(|value| value == "true"),
+        );
         return;
     }
 
@@ -60,14 +67,15 @@ fn oauth_recovery_file_logging_is_separate_and_overrides_global_log_level() {
     let output = Command::new(env::current_exe().expect("current test executable"))
         .args([
             "--exact",
-            "logging::oauth_recovery_file_logging_is_separate_and_overrides_global_log_level",
+            "logging::sensitive_file_logging_is_separate_and_overrides_global_log_level",
             "--nocapture",
         ])
         .env(CHILD_PROCESS_ENV, "1")
         .env(LOG_DIRECTORY_ENV, directory.path())
+        .env(REQUEST_DUMP_ENABLED_ENV, "true")
         .env(
             "RUST_LOG",
-            "off,logging_test_application=info,oauth_recovery=off",
+            "off,logging_test_application=info,oauth_recovery=off,request_dump=trace",
         )
         .output()
         .expect("run logging child process");
@@ -84,6 +92,8 @@ fn oauth_recovery_file_logging_is_separate_and_overrides_global_log_level() {
     assert!(application_log.contains(APPLICATION_LOG_MARKER));
     assert!(!application_log.contains(&json_target_field(OAUTH_RECOVERY_LOG_TARGET)));
     assert!(!application_log.contains(OAUTH_RECOVERY_LOG_MARKER));
+    assert!(!application_log.contains(&json_target_field(REQUEST_DUMP_LOG_TARGET)));
+    assert!(!application_log.contains(REQUEST_DUMP_SECRET));
 
     let recovery_log = read_log_file_set(directory.path(), OAUTH_RECOVERY_LOG_FILE_PREFIX);
     assert!(recovery_log.contains(&json_target_field(OAUTH_RECOVERY_LOG_TARGET)));
@@ -91,9 +101,49 @@ fn oauth_recovery_file_logging_is_separate_and_overrides_global_log_level() {
     assert!(!recovery_log.contains(&json_target_field(APPLICATION_LOG_TARGET)));
     assert!(!recovery_log.contains(APPLICATION_LOG_MARKER));
     assert!(recovery_log.contains(&format!(r#""provider":"{OAUTH_RECOVERY_PROVIDER}""#)));
+    assert!(!recovery_log.contains(&json_target_field(REQUEST_DUMP_LOG_TARGET)));
+    assert!(!recovery_log.contains(REQUEST_DUMP_SECRET));
+
+    let request_dump_log = read_log_file_set(directory.path(), REQUEST_DUMP_LOG_FILE_PREFIX);
+    assert!(request_dump_log.contains(&json_target_field(REQUEST_DUMP_LOG_TARGET)));
+    assert!(request_dump_log.contains(REQUEST_DUMP_LOG_MARKER));
+    assert!(request_dump_log.contains(REQUEST_DUMP_SECRET));
+    assert!(!request_dump_log.contains(&json_target_field(APPLICATION_LOG_TARGET)));
+    assert!(!request_dump_log.contains(APPLICATION_LOG_MARKER));
+    assert!(!request_dump_log.contains(&json_target_field(OAUTH_RECOVERY_LOG_TARGET)));
+    assert!(!request_dump_log.contains(OAUTH_RECOVERY_LOG_MARKER));
+
+    let disabled_directory = tempfile::tempdir().expect("create disabled log directory");
+    let output = Command::new(env::current_exe().expect("current test executable"))
+        .args([
+            "--exact",
+            "logging::sensitive_file_logging_is_separate_and_overrides_global_log_level",
+            "--nocapture",
+        ])
+        .env(CHILD_PROCESS_ENV, "1")
+        .env(LOG_DIRECTORY_ENV, disabled_directory.path())
+        .env(REQUEST_DUMP_ENABLED_ENV, "false")
+        .env(
+            "RUST_LOG",
+            "off,logging_test_application=info,oauth_recovery=off,request_dump=trace",
+        )
+        .output()
+        .expect("run logging child process with request dump disabled");
+    assert!(
+        output.status.success(),
+        "logging child process failed:\nstdout:\n{}\nstderr:\n{}",
+        String::from_utf8_lossy(&output.stdout),
+        String::from_utf8_lossy(&output.stderr),
+    );
+    assert!(!log_file_set_exists(
+        disabled_directory.path(),
+        REQUEST_DUMP_LOG_FILE_PREFIX
+    ));
+    let application_log = read_log_file_set(disabled_directory.path(), APPLICATION_LOG_FILE_PREFIX);
+    assert!(!application_log.contains(REQUEST_DUMP_SECRET));
 }
 
-fn write_oauth_recovery_log(directory: PathBuf) {
+fn write_sensitive_logs(directory: PathBuf, request_dump: bool) {
     let config = HostConfig {
         listen: ListenConfig {
             host: "127.0.0.1".to_owned(),
@@ -110,6 +160,7 @@ fn write_oauth_recovery_log(directory: PathBuf) {
                 max_file_size_mb: 1,
                 max_files: 1,
             },
+            request_dump,
         },
         system_update: SystemUpdateConfig::default(),
         drain_timeout_seconds: 30,
@@ -134,6 +185,12 @@ fn write_oauth_recovery_log(directory: PathBuf) {
         marker = OAUTH_RECOVERY_LOG_MARKER,
         "OAuth recovery test record"
     );
+    tracing::info!(
+        target: REQUEST_DUMP_LOG_TARGET,
+        marker = REQUEST_DUMP_LOG_MARKER,
+        authorization = REQUEST_DUMP_SECRET,
+        "request dump test record"
+    );
     drop(bundle);
 }
 
@@ -156,6 +213,18 @@ fn read_log_file_set(directory: &Path, file_prefix: &str) -> String {
         .collect()
 }
 
+fn log_file_set_exists(directory: &Path, file_prefix: &str) -> bool {
+    fs::read_dir(directory)
+        .expect("read log directory")
+        .filter_map(Result::ok)
+        .any(|entry| {
+            entry
+                .file_name()
+                .to_str()
+                .is_some_and(|name| name.starts_with(file_prefix))
+        })
+}
+
 fn json_target_field(target: &str) -> String {
     format!(r#""target":"{target}""#)
 }

backend/crates/gateway-protocol/src/openai/codex.rs

@@ -4,6 +4,41 @@ const MULTI_AGENT_MODE_OPEN_TAG: &str = "<multi_agent_mode>";
 const MULTI_AGENT_MODE_CLOSE_TAG: &str = "</multi_agent_mode>";
 const PROACTIVE_MULTI_AGENT_MODE_PREFIX: &str = "Proactive multi-agent delegation is active.";
 
+/// 各 Codex 端点共用的显式根会话身份。连接边界的 session_id 优先，
+/// 正文及官方 metadata 只做回退；不从 turn、window 或请求内容猜测归属。
+#[must_use]
+pub fn codex_session_id(body: &Map<String, Value>, context: &Map<String, Value>) -> Option<String> {
+    codex_identity_field(body, context, "session_id")
+}
+
+/// 各 Codex 端点共用的显式线程身份，与根会话采用相同的来源优先级。
+#[must_use]
+pub fn codex_thread_id(body: &Map<String, Value>, context: &Map<String, Value>) -> Option<String> {
+    codex_identity_field(body, context, "thread_id")
+}
+
+fn codex_identity_field(
+    body: &Map<String, Value>,
+    context: &Map<String, Value>,
+    field: &str,
+) -> Option<String> {
+    non_empty_string(context.get(field))
+        .or_else(|| non_empty_string(body.get(field)))
+        .or_else(|| client_metadata_string(body, field))
+        .map(str::to_owned)
+        .or_else(|| {
+            let metadata = non_empty_string(context.get("turn_metadata"))
+                .or_else(|| request_turn_metadata(body))?;
+            let metadata: Value = serde_json::from_str(metadata).ok()?;
+            non_empty_string(metadata.get(field)).map(str::to_owned)
+        })
+        .or_else(|| {
+            body.get("context")
+                .and_then(|context| non_empty_string(context.get(field)))
+                .map(str::to_owned)
+        })
+}
+
 /// 从 OpenAI Responses 请求中提取的稳定 Codex 请求语义。
 #[derive(Debug, Clone, Default, PartialEq, Eq)]
 pub struct CodexResponsesRequestSemantics {

backend/crates/gateway-protocol/src/openai/events.rs

@@ -98,6 +98,37 @@ pub fn extract_usage(body: &Value) -> Option<TokenUsage> {
     })
 }
 
+/// 校验计价所需的输入、输出和缓存读取字段，拒绝缺失、非法或不一致的数值。
+///
+/// 用量提取允许缺失字段补零；计价必须额外确认这些零值有明确的上游事实。
+#[must_use]
+pub fn billable_usage_is_complete(response: &Value, usage: TokenUsage) -> bool {
+    let Some(raw) = response.get("usage").filter(|value| value.is_object()) else {
+        return false;
+    };
+    let input = raw
+        .get("input_tokens")
+        .or_else(|| raw.get("prompt_tokens"))
+        .and_then(Value::as_u64);
+    let output = raw
+        .get("output_tokens")
+        .or_else(|| raw.get("completion_tokens"))
+        .and_then(Value::as_u64);
+    let cached = raw
+        .pointer("/input_tokens_details/cached_tokens")
+        .or_else(|| raw.pointer("/prompt_tokens_details/cached_tokens"))
+        .or_else(|| raw.get("cached_tokens"));
+    let cached = match cached {
+        Some(value) => value.as_u64(),
+        None => Some(0),
+    };
+
+    input == Some(usage.input_tokens)
+        && output == Some(usage.output_tokens)
+        && cached == Some(usage.cached_tokens)
+        && usage.cached_tokens <= usage.input_tokens
+}
+
 /// 从完整 SSE 文本中提取最终可见用量。
 ///
 /// 如果存在 `response.completed` 的 usage，则优先返回它；否则回退到最后一条

backend/crates/gateway-protocol/src/openai/mod.rs

@@ -18,5 +18,5 @@ pub mod sse;
 
 pub use codex::{
     CodexResponsesRequestSemantics, codex_responses_request_semantics,
-    codex_responses_request_semantics_with_turn_metadata,
+    codex_responses_request_semantics_with_turn_metadata, codex_session_id, codex_thread_id,
 };

backend/crates/gateway-protocol/tests/openai/codex.rs

@@ -1,6 +1,101 @@
-use gateway_protocol::openai::codex_responses_request_semantics;
+use gateway_protocol::openai::{
+    codex_responses_request_semantics, codex_session_id, codex_thread_id,
+};
 use serde_json::{Map, Value, json};
 
+#[test]
+fn codex_session_should_prefer_the_connection_identity_over_body_metadata() {
+    let body = object(
+        json!({"session_id":"body-session", "client_metadata":{"session_id":"metadata-session"}}),
+    );
+    let context = object(
+        json!({"session_id":" root-session ", "turn_metadata":"{\"session_id\":\"turn-session\"}"}),
+    );
+    assert_eq!(
+        codex_session_id(&body, &context).as_deref(),
+        Some("root-session")
+    );
+}
+
+#[test]
+fn codex_session_should_accept_explicit_metadata_and_ignore_unrelated_ids() {
+    for (body, context) in [
+        (json!({"session_id":"root"}), json!({})),
+        (json!({"client_metadata":{"session_id":"root"}}), json!({})),
+        (
+            json!({"client_metadata":{"x-codex-turn-metadata":"{\"session_id\":\"root\"}"}}),
+            json!({}),
+        ),
+        (
+            json!({}),
+            json!({"turn_metadata":"{\"session_id\":\"root\"}"}),
+        ),
+        (json!({"context":{"session_id":"root"}}), json!({})),
+    ] {
+        assert_eq!(
+            codex_session_id(&object(body), &object(context)).as_deref(),
+            Some("root")
+        );
+    }
+    for body in [
+        json!({"session_id":" ", "thread_id":"child", "turn_id":"turn", "id":"request"}),
+        json!({"client_metadata":{"x-codex-turn-metadata":"invalid JSON"}}),
+    ] {
+        assert!(codex_session_id(&object(body), &Map::new()).is_none());
+    }
+}
+
+#[test]
+fn codex_thread_should_use_explicit_identity_and_never_guess_from_other_ids() {
+    for (body, context, expected) in [
+        (
+            json!({"thread_id":"body", "client_metadata":{"thread_id":"metadata"}}),
+            json!({"thread_id":" child "}),
+            Some("child"),
+        ),
+        (
+            json!({"thread_id":"body", "client_metadata":{"thread_id":"metadata"}}),
+            json!({}),
+            Some("body"),
+        ),
+        (
+            json!({"client_metadata":{"thread_id":"metadata"}}),
+            json!({}),
+            Some("metadata"),
+        ),
+        (
+            json!({"client_metadata":{"x-codex-turn-metadata":"{\"thread_id\":\"metadata-thread\"}"}}),
+            json!({}),
+            Some("metadata-thread"),
+        ),
+        (
+            json!({"turnMetadata":"{\"thread_id\":\"body-turn\"}"}),
+            json!({"turn_metadata":"{\"thread_id\":\"header-turn\"}"}),
+            Some("header-turn"),
+        ),
+        (
+            json!({"context":{"thread_id":"context-thread"}}),
+            json!({}),
+            Some("context-thread"),
+        ),
+        (
+            json!({"thread_id":" ", "session_id":"root", "turn_id":"turn", "agent_name":"agent"}),
+            json!({}),
+            None,
+        ),
+        (
+            json!({"client_metadata":{"x-codex-turn-metadata":"invalid JSON"}}),
+            json!({}),
+            None,
+        ),
+    ] {
+        assert_eq!(
+            codex_thread_id(&object(body), &object(context)).as_deref(),
+            expected
+        );
+    }
+}
+
 #[test]
 fn codex_semantics_should_use_protocol_context_before_body_metadata() {
     let body = json!({

backend/crates/gateway-protocol/tests/openai/events.rs

@@ -1,10 +1,55 @@
 use gateway_protocol::openai::events::{
-    RateLimitKeySource, RateLimitWindow, TokenUsage, extract_sse_usage, extract_usage,
-    is_codex_quota_header_name, is_rate_limit_header_name, parse_rate_limit_headers,
+    RateLimitKeySource, RateLimitWindow, TokenUsage, billable_usage_is_complete, extract_sse_usage,
+    extract_usage, is_codex_quota_header_name, is_rate_limit_header_name, parse_rate_limit_headers,
     parse_rate_limits_event, rate_limits_to_header_pairs, retry_after_seconds_from_body,
 };
 use serde_json::json;
 
+#[test]
+fn billable_usage_should_require_valid_input_output_and_cache_facts() {
+    for (raw, expected) in [
+        (json!({}), false),
+        (json!({"input_tokens": 0}), false),
+        (json!({"input_tokens": 0, "output_tokens": 0}), true),
+        (
+            json!({"prompt_tokens": 8, "completion_tokens": 4, "cached_tokens": 2}),
+            true,
+        ),
+        (json!({"input_tokens": -1, "output_tokens": 1}), false),
+        (
+            json!({"input_tokens": "8", "prompt_tokens": 8, "output_tokens": 1}),
+            false,
+        ),
+        (json!({"input_tokens": 8, "output_tokens": 1.5}), false),
+        (
+            json!({"input_tokens": 8, "output_tokens": 1, "cached_tokens": 9}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 8, "output_tokens": 1, "cached_tokens": null}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 8, "output_tokens": 1, "input_tokens_details": {"cached_tokens": "2"}, "cached_tokens": 2}),
+            false,
+        ),
+        (
+            json!({"input_tokens": 8, "prompt_tokens": 9, "output_tokens": 1}),
+            true,
+        ),
+    ] {
+        let body = json!({"usage": raw});
+        let usage = extract_usage(&body).unwrap_or_default();
+        assert_eq!(billable_usage_is_complete(&body, usage), expected, "{body}");
+    }
+    let body = json!({"usage": {"input_tokens": 8, "output_tokens": 4}});
+    assert!(!billable_usage_is_complete(&body, TokenUsage::default()));
+    assert!(!billable_usage_is_complete(
+        &json!({}),
+        TokenUsage::default()
+    ));
+}
+
 #[test]
 fn extract_usage_should_read_codex_usage_shape() {
     let body = json!({

backend/crates/gateway-store/Cargo.toml

@@ -12,6 +12,7 @@ name = "main"
 path = "tests/main.rs"
 
 [dependencies]
+async-stream = "0.3.6"
 async-trait = "0.1.89"
 aws-credential-types = "1.2"
 aws-sdk-s3 = { version = "1.140", default-features = false, features = ["rt-tokio", "default-https-client"] }

backend/crates/gateway-store/src/postgres/backup.rs

@@ -59,12 +59,25 @@ impl BackupRepository for PgBackupRepository {
             .map_err(|_| store_unavailable("begin backup storage update"))?;
         let result = async {
             let current = lock_settings_in_transaction(&mut transaction).await?;
+            let mut changed = storage_changed_fields(&current, &command);
+            if changed.is_empty() {
+                let revision: i64 =
+                    sqlx::query_scalar("select config_revision from runtime_settings where id = 1")
+                        .fetch_one(&mut *transaction)
+                        .await
+                        .map_err(|_| store_unavailable("load unchanged backup revision"))?;
+                let revision = AdminRevision::new(
+                    u64::try_from(revision).map_err(|_| store_invalid("config revision"))?,
+                )
+                .map_err(|_| store_invalid("config revision"))?;
+                return Ok((current, revision));
+            }
             ensure_storage_identity_stable(&current, &command, &mut transaction).await?;
 
             let secret = command
                 .secret_access_key
                 .as_ref()
-                .map(|secret| secret.expose_secret().to_owned());
+                .map(|secret| secret.expose_secret());
             sqlx::query(
                 "update backup_settings
                     set endpoint = $1,
@@ -76,6 +89,8 @@ impl BackupRepository for PgBackupRepository {
                         force_path_style = $7,
                         storage_revision = storage_revision + 1,
                         last_verified_at = null,
+                        schedule_enabled = false,
+                        next_run_at = null,
                         updated_at = now()
                   where id = 1",
             )
@@ -83,7 +98,7 @@ impl BackupRepository for PgBackupRepository {
             .bind(&command.region)
             .bind(&command.bucket)
             .bind(&command.access_key_id)
-            .bind(secret.as_deref())
+            .bind(secret)
             .bind(&command.prefix)
             .bind(command.force_path_style)
             .execute(&mut *transaction)
@@ -94,7 +109,13 @@ impl BackupRepository for PgBackupRepository {
                 .await
                 .map_err(map_admin_error)?;
             let revision = admin_revision(store_revision)?;
-            let changed = storage_changed_fields(&current, &command);
+            if current.last_verified_at.is_some() {
+                changed.push("last_verified_at".to_owned());
+            }
+            if current.schedule_enabled {
+                changed.push("schedule_enabled".to_owned());
+                changed.push("next_run_at".to_owned());
+            }
             let audit = crate::mutation_audit(
                 context,
                 "backup.s3_config_updated",
@@ -719,7 +740,13 @@ fn storage_changed_fields(
     if current.access_key_id.as_deref() != Some(command.access_key_id.as_str()) {
         fields.push("access_key_id".to_owned());
     }
-    if command.secret_access_key.is_some() {
+    if command.secret_access_key.as_ref().is_some_and(|secret| {
+        current
+            .secret_access_key
+            .as_ref()
+            .map(|value| value.expose_secret())
+            != Some(secret.expose_secret())
+    }) {
         fields.push("secret_access_key".to_owned());
     }
     if current.prefix.as_deref() != Some(command.prefix.as_str()) {

backend/crates/gateway-store/src/postgres/execution.rs

@@ -283,6 +283,7 @@ pub struct ModelRequestFinalization {
     pub websocket_pool: Option<String>,
     pub service_tier: Option<String>,
     pub provider_metadata_json: Option<Value>,
+    pub diagnostic_trace_json: Option<Value>,
     pub error_kind: Option<String>,
     pub provider_error_code: Option<String>,
     pub error_message: Option<String>,
@@ -353,6 +354,15 @@ impl ModelRequestFinalization {
         {
             return Err(invalid("provider observation must be a JSON object"));
         }
+        if self
+            .diagnostic_trace_json
+            .as_ref()
+            .is_some_and(|trace| !trace.is_object() || trace.to_string().len() > 64 * 1024)
+        {
+            return Err(invalid(
+                "diagnostic trace must be a JSON object within 64 KiB",
+            ));
+        }
         validate_optional_stable_reason(
             self.continuation_unavailable_reason.as_deref(),
             "continuation unavailable reason",
@@ -725,10 +735,11 @@ impl ModelRequestRepository for PgExecutionStore {
                  upstream_connection_id = $44,
                  upstream_connection_exit_reason = $45,
                  upstream_connection_age_ms = $46,
-                 upstream_connection_idle_ms = $47
+                 upstream_connection_idle_ms = $47, diagnostic_trace_json = $48
              where id = $1 and outcome = 'running'
              returning id, client_api_key_ref, continuation_affinity_hash,
-                       continuation_requested, outcome, started_at, completed_at
+                       continuation_requested, provider_kind, upstream_transport,
+                       outcome, started_at, completed_at
            ), recovery_target as (
              select prior.id
              from model_requests prior
@@ -775,9 +786,47 @@ impl ModelRequestRepository for PgExecutionStore {
              where prior.id = target.id
                and prior.recovered_at is null
              returning prior.id
+           ), session_transport_recovery_target as (
+             select prior.id
+             from model_requests prior
+             cross join finalized current
+             where current.outcome = 'succeeded'
+               and current.provider_kind = 'openai'
+               and current.upstream_transport = 'http_sse'
+               and current.continuation_affinity_hash is not null
+               and prior.client_api_key_ref = current.client_api_key_ref
+               and prior.continuation_affinity_hash = current.continuation_affinity_hash
+               and prior.provider_kind = current.provider_kind
+               and prior.outcome = 'failed'
+               and prior.error_kind = 'upstream_unavailable'
+               and prior.upstream_transport = 'websocket'
+               and prior.upstream_send_state = 'ambiguous'
+               and prior.downstream_committed_at is null
+               and prior.recovered_at is null
+               and current.started_at >= prior.completed_at
+               and current.started_at <= prior.completed_at + interval '30 seconds'
+           ), session_transport_recovery_update as (
+             update model_requests prior
+             set recovery_attempt_count = prior.recovery_attempt_count + 1,
+                 recovery_request_id = current.id,
+                 recovered_at = current.completed_at,
+                 recovery_retry_delay_ms = greatest(
+                   0,
+                   floor(extract(epoch from (current.started_at - prior.completed_at)) * 1000)
+                 )::bigint,
+                 recovery_total_latency_ms = greatest(
+                   0,
+                   floor(extract(epoch from (current.completed_at - prior.completed_at)) * 1000)
+                 )::bigint
+             from finalized current
+             join session_transport_recovery_target target on true
+             where prior.id = target.id
+               and prior.recovered_at is null
+             returning prior.id
            )
            select (select count(*) from finalized)::bigint
-                + (select count(*) * 0 from recovery_update)::bigint",
+                + (select count(*) * 0 from recovery_update)::bigint
+                + (select count(*) * 0 from session_transport_recovery_update)::bigint",
         )
         .bind(&finalization.model_request_id)
         .bind(finalization.outcome.as_str())
@@ -877,6 +926,7 @@ impl ModelRequestRepository for PgExecutionStore {
             finalization.upstream_connection_idle_ms,
             "upstream_connection_idle_ms",
         )?)
+        .bind(finalization.diagnostic_trace_json.map(sqlx::types::Json))
         .fetch_one(&self.pool)
         .await
         .map_err(|_| postgres_unavailable("finalize model request"))?;
@@ -1039,7 +1089,6 @@ impl ExecutionStore for PgExecutionStore {
                     || "intermediate upstream failure".to_owned(),
                     |diagnostic| diagnostic.as_str().to_owned(),
                 ),
-                occurrence_count: 1,
                 created_at: Utc::now(),
             },
         )
@@ -1084,7 +1133,6 @@ impl ExecutionStore for PgExecutionStore {
                     || "account connection test failed".to_owned(),
                     |diagnostic| diagnostic.as_str().to_owned(),
                 ),
-                occurrence_count: 1,
                 created_at: Utc::now(),
             },
         )
@@ -1176,6 +1224,12 @@ impl ExecutionStore for PgExecutionStore {
                 websocket_pool: finalization.websocket_pool,
                 service_tier: finalization.service_tier,
                 provider_metadata_json,
+                diagnostic_trace_json: finalization
+                    .diagnostic_trace_json
+                    .as_deref()
+                    .map(serde_json::from_str)
+                    .transpose()
+                    .map_err(|_| CoreStoreError::new(CoreStoreErrorKind::InvalidData))?,
                 error_kind,
                 provider_error_code: finalization.provider_error_code,
                 error_message,

backend/crates/gateway-store/src/postgres/execution_buffer.rs

@@ -583,6 +583,7 @@ impl ExecutionObservationWrite {
                     finalization.websocket_pool.as_deref(),
                     finalization.service_tier.as_deref(),
                     finalization.provider_metadata_json.as_deref(),
+                    finalization.diagnostic_trace_json.as_deref(),
                     finalization.provider_error_code.as_deref(),
                 ])
                 .saturating_add(error_bytes)

backend/crates/gateway-store/src/postgres/observability/admin_adapter.rs

@@ -1,5 +1,7 @@
 //! Pg 观测 adapter：实现 `ObservabilityRepository` 与 `AdminObservabilityStore`。
 
+use futures::{StreamExt, TryStreamExt, stream::BoxStream};
+use gateway_admin::ports::store::UsageCalculatedBillingStream;
 use gateway_core::provider_ports::ProviderCooldownPort;
 use std::sync::Arc;
 
@@ -185,17 +187,15 @@ impl ObservabilityRepository for PgObservabilityRepository {
             .await
     }
 
-    async fn usage_calculated_billing_facts(
+    fn usage_calculated_billing_facts(
         &self,
         range: ObservabilityRange,
         filter: UsageRecordFilter,
-    ) -> StoreResult<Vec<CalculatedUsageBillingFact>> {
-        self.query_budget
-            .run(
-                "load calculated usage billing facts",
-                calculated_usage_billing_facts(&self.pool, range, &filter),
-            )
-            .await
+    ) -> BoxStream<'_, StoreResult<CalculatedUsageBillingFact>> {
+        self.query_budget.run_stream(
+            "load calculated usage billing facts",
+            calculated_usage_billing_facts(&self.pool, range, filter),
+        )
     }
 
     async fn provider_account_usage(
@@ -387,18 +387,20 @@ impl AdminObservabilityStore for PgAdminObservabilityStore {
             .collect()
     }
 
-    async fn usage_calculated_billing_facts(
+    fn usage_calculated_billing_facts(
         &self,
         range: admin_observability::TimeRange,
         filter: admin_observability::UsageFilter,
-    ) -> AdminStoreResult<Vec<admin_observability::UsageCalculatedBillingFact>> {
+    ) -> UsageCalculatedBillingStream<'_> {
+        let range = match store_range(range) {
+            Ok(range) => range,
+            Err(error) => return Box::pin(futures::stream::once(async { Err(error) })),
+        };
         self.repository
-            .usage_calculated_billing_facts(store_range(range)?, store_usage_filter(filter))
-            .await
-            .map_err(observability_error)?
-            .into_iter()
-            .map(admin_calculated_usage_billing_fact)
-            .collect()
+            .usage_calculated_billing_facts(range, store_usage_filter(filter))
+            .map_err(observability_error)
+            .map(|fact| fact.and_then(admin_calculated_usage_billing_fact))
+            .boxed()
     }
 
     async fn list_usage_records(

backend/crates/gateway-store/src/postgres/observability/mapping.rs

@@ -319,6 +319,7 @@ pub(crate) fn admin_dashboard_account_usage(
             )
             .collect(),
         quota_used_percent: None,
+        quota_window: None,
         models: usage
             .models
             .into_iter()
@@ -583,6 +584,8 @@ pub(crate) fn admin_usage_detail(
     detail: UsageRecordDetail,
 ) -> AdminStoreResult<admin_observability::UsageDetail> {
     Ok(admin_observability::UsageDetail {
+        trace: detail.trace,
+        related_requests: detail.related_requests,
         request: admin_usage_record(detail.request)?,
         attempts: detail
             .attempts
@@ -721,7 +724,6 @@ pub(crate) fn admin_ops_error(error: OpsErrorRecord) -> admin_observability::Ops
         latency_ms: error.latency_ms,
         message: error.message,
         raw_upstream_error: error.raw_upstream_error,
-        occurrence_count: error.occurrence_count,
         client_ip: error.client_ip,
         user_agent: error.user_agent,
         reasoning_effort: error.reasoning_effort,
@@ -912,7 +914,6 @@ pub(crate) fn ops_error_from_row(row: &sqlx::postgres::PgRow) -> StoreResult<Ops
         latency_ms: optional_unsigned(row, "latency_ms")?,
         message: get(row, "message")?,
         raw_upstream_error: get(row, "raw_upstream_error")?,
-        occurrence_count: to_u32(get(row, "occurrence_count")?)?,
         client_ip: get(row, "client_ip")?,
         user_agent: get(row, "user_agent")?,
         reasoning_effort: get(row, "reasoning_effort")?,

backend/crates/gateway-store/src/postgres/observability/model.rs

@@ -1,6 +1,7 @@
 //! 查询模型、校验与观测端口契约。
 
 use super::*;
+use futures::stream::BoxStream;
 
 pub(crate) const MAX_PAGE_SIZE: u16 = 100;
 pub(crate) const MAX_FILTER_BYTES: usize = 256;
@@ -644,6 +645,8 @@ pub struct UsageAttemptObservation {
 
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UsageRecordDetail {
+    pub trace: Option<serde_json::Value>,
+    pub related_requests: Vec<serde_json::Value>,
     pub request: UsageRecord,
     pub attempts: Vec<UsageAttemptObservation>,
 }
@@ -714,7 +717,6 @@ pub struct OpsErrorRecord {
     pub latency_ms: Option<u64>,
     pub message: String,
     pub raw_upstream_error: Option<String>,
-    pub occurrence_count: u32,
     pub client_ip: Option<String>,
     pub user_agent: Option<String>,
     pub reasoning_effort: Option<String>,
@@ -762,11 +764,11 @@ pub trait ObservabilityRepository: Send + Sync {
         range: ObservabilityRange,
         filter: UsageRecordFilter,
     ) -> StoreResult<Vec<RequestMetricPoint>>;
-    async fn usage_calculated_billing_facts(
+    fn usage_calculated_billing_facts(
         &self,
         range: ObservabilityRange,
         filter: UsageRecordFilter,
-    ) -> StoreResult<Vec<CalculatedUsageBillingFact>>;
+    ) -> BoxStream<'_, StoreResult<CalculatedUsageBillingFact>>;
     async fn provider_account_usage(
         &self,
         query: ProviderAccountUsageQuery,

backend/crates/gateway-store/src/postgres/observability/queries/dashboard.rs

@@ -283,42 +283,44 @@ async fn request_metric_series_inner(
     fill_metric_gaps(range, granularity, points)
 }
 
-pub(crate) async fn calculated_usage_billing_facts(
+pub(crate) fn calculated_usage_billing_facts(
     pool: &PgPool,
     range: ObservabilityRange,
-    filter: &UsageRecordFilter,
-) -> StoreResult<Vec<CalculatedUsageBillingFact>> {
-    filter.validate()?;
-    let granularity = granularity_for(range);
-    let mut query = QueryBuilder::<Postgres>::new("select date_bin(");
-    query.push_bind(granularity.sql_interval());
-    query.push(
-        "::interval, mr.started_at, timestamptz '1970-01-01 00:00:00+00') as bucket_start,
-                mr.provider_kind, mr.upstream_model_id, mr.service_tier,
-                mr.input_tokens, mr.output_tokens, mr.cached_tokens, mr.cache_write_tokens,
-                mr.cost_currency, mr.cost_amount::text as amount
-         from model_requests mr where mr.started_at >= ",
-    );
-    query.push_bind(range.start);
-    query.push(" and mr.started_at < ");
-    query.push_bind(range.end);
-    query.push(
-        " and mr.cost_source = 'calculated'
-           and mr.cost_amount is not null and mr.cost_currency is not null
-           and nullif(mr.provider_kind, '') is not null
-           and nullif(mr.upstream_model_id, '') is not null",
-    );
-    push_completed_usage_fact_filter(&mut query, "mr");
-    push_usage_filter(&mut query, filter, "mr");
-    query.push(" order by bucket_start, mr.id");
-    query
-        .build()
-        .fetch_all(pool)
-        .await
-        .map_err(|_| postgres_unavailable("load calculated usage billing facts"))?
-        .iter()
-        .map(calculated_usage_billing_fact_from_row)
-        .collect()
+    filter: UsageRecordFilter,
+) -> futures::stream::BoxStream<'_, StoreResult<CalculatedUsageBillingFact>> {
+    use futures::TryStreamExt;
+
+    Box::pin(async_stream::try_stream! {
+        filter.validate()?;
+        let granularity = granularity_for(range);
+        let mut query = QueryBuilder::<Postgres>::new("select date_bin(");
+        query.push_bind(granularity.sql_interval());
+        query.push(
+            "::interval, mr.started_at, timestamptz '1970-01-01 00:00:00+00') as bucket_start,
+                    mr.provider_kind, mr.upstream_model_id, mr.service_tier,
+                    mr.input_tokens, mr.output_tokens, mr.cached_tokens, mr.cache_write_tokens,
+                    mr.cost_currency, mr.cost_amount::text as amount
+             from model_requests mr where mr.started_at >= ",
+        );
+        query.push_bind(range.start);
+        query.push(" and mr.started_at < ");
+        query.push_bind(range.end);
+        query.push(
+            " and mr.cost_source = 'calculated'
+               and mr.cost_amount is not null and mr.cost_currency is not null
+               and nullif(mr.provider_kind, '') is not null
+               and nullif(mr.upstream_model_id, '') is not null",
+        );
+        push_completed_usage_fact_filter(&mut query, "mr");
+        push_usage_filter(&mut query, &filter, "mr");
+        // 费用逐条累加与行顺序无关，避免全量物化和不必要的数据库排序。
+        let mut rows = query.build().fetch(pool);
+        while let Some(row) = rows.try_next().await
+            .map_err(|_| postgres_unavailable("load calculated usage billing facts"))?
+        {
+            yield calculated_usage_billing_fact_from_row(&row)?;
+        }
+    })
 }
 
 pub(crate) async fn request_costs_by_bucket(
@@ -363,22 +365,22 @@ pub(crate) async fn attempt_metrics(
     filter: &UsageRecordFilter,
 ) -> StoreResult<AttemptMetrics> {
     filter.validate()?;
-    let fact = completed_usage_fact_predicate("sr");
+    let completed_usage = completed_usage_fact_predicate("mr");
     // PERF: `model_requests` 包含宽请求元数据。这里只物化两个聚合族会复用的列，
     // 避免生产基数下 `mr.*` 与多次标量子查询把 CTE 反复写入临时文件。
-    let mut query = QueryBuilder::<Postgres>::new(
+    let mut query = QueryBuilder::<Postgres>::new(format!(
         "with selected_requests as materialized (
            select mr.id, mr.attempt_count, mr.outcome, mr.error_kind,
                   mr.upstream_status_code, mr.client_status_code,
-                  mr.downstream_committed_at, mr.client_transport, mr.cost_source
-           from model_requests mr where mr.started_at >= ",
-    );
+                  mr.cost_source, ({completed_usage}) as is_completed_usage
+           from model_requests mr where mr.started_at >= "
+    ));
     query.push_bind(range.start);
     query.push(" and mr.started_at < ");
     query.push_bind(range.end);
     push_unrecovered_request_filter(&mut query, "mr");
     push_usage_filter(&mut query, filter, "mr");
-    query.push(format!(
+    query.push(
         "), request_aggregate as (
            select coalesce(sum(sr.attempt_count), 0)::bigint as attempt_count,
                   count(*) filter (
@@ -413,30 +415,30 @@ pub(crate) async fn attempt_metrics(
                                    sr.client_status_code) between 500 and 599
                   )::bigint as request_provider_5xx_count,
                   count(*) filter (
-                    where {fact} and sr.cost_source = 'provider_reported'
+                    where sr.is_completed_usage and sr.cost_source = 'provider_reported'
                   )::bigint as provider_reported_count,
                   count(*) filter (
-                    where {fact} and sr.cost_source = 'calculated'
+                    where sr.is_completed_usage and sr.cost_source = 'calculated'
                   )::bigint as calculated_count,
                   count(*) filter (
-                    where {fact} and sr.cost_source = 'unavailable'
+                    where sr.is_completed_usage and sr.cost_source = 'unavailable'
                   )::bigint as unavailable_count
            from selected_requests sr
          ), ops_aggregate as (
-           select coalesce(sum(coalesce(oe.occurrence_count, 1)), 0)::bigint
+           select count(*)::bigint
                     as ops_failure_count,
-                  coalesce(sum(coalesce(oe.occurrence_count, 1)) filter (
+                  count(*) filter (
                     where oe.failure_kind in ('rate_limited', 'quota_exhausted')
                        or oe.status_code = 429
-                  ), 0)::bigint as ops_rate_limited_count,
-                  coalesce(sum(coalesce(oe.occurrence_count, 1)) filter (
+                  )::bigint as ops_rate_limited_count,
+                  count(*) filter (
                     where oe.failure_kind in
                             ('authentication', 'authorization', 'invalid_credential')
                        or oe.status_code in (401, 403)
-                  ), 0)::bigint as ops_auth_failure_count,
-                  coalesce(sum(coalesce(oe.occurrence_count, 1)) filter (
+                  )::bigint as ops_auth_failure_count,
+                  count(*) filter (
                     where oe.status_code between 500 and 599
-                  ), 0)::bigint as ops_provider_5xx_count
+                  )::bigint as ops_provider_5xx_count
            from ops_events oe
            join selected_requests sr on sr.id = oe.model_request_id
          )
@@ -455,8 +457,8 @@ pub(crate) async fn attempt_metrics(
                 requests.calculated_count,
                 requests.unavailable_count
          from request_aggregate requests
-         cross join ops_aggregate ops"
-    ));
+         cross join ops_aggregate ops",
+    );
     let row = query
         .build()
         .fetch_one(pool)

backend/crates/gateway-store/src/postgres/observability/queries/ops.rs

@@ -20,7 +20,6 @@ const REQUEST_ERROR_SELECT: &str = "select 'model_request'::text as source,
        mr.latency_ms,
        coalesce(mr.error_message, mr.error_kind, 'request failed') as message,
        mr.raw_upstream_error,
-       1::integer as occurrence_count,
        host(mr.client_ip) as client_ip, mr.user_agent,
        mr.reasoning_effort, mr.reasoning_preset, mr.request_kind,
        mr.subagent_kind, mr.compact,
@@ -48,7 +47,7 @@ const OPS_EVENT_SELECT: &str = "select 'ops_event'::text as source,
        oe.upstream_send_state,
        null::integer as client_status_code, oe.status_code as upstream_status_code,
        oe.provider_error_code, mr.client_response_id, oe.upstream_request_id,
-       oe.latency_ms, oe.message, oe.raw_upstream_error, oe.occurrence_count,
+       oe.latency_ms, oe.message, oe.raw_upstream_error,
        host(mr.client_ip) as client_ip, mr.user_agent,
        mr.reasoning_effort, mr.reasoning_preset, mr.request_kind,
        mr.subagent_kind, mr.compact,
@@ -128,7 +127,6 @@ fn push_request_error_predicates(
     range: ObservabilityRange,
     filter: &OpsErrorFilter,
 ) {
-    statement.push(" and mr.recovered_at is null");
     push_range(statement, "mr.completed_at", range);
     for (column, value) in [
         ("mr.client_api_key_ref", &filter.client_api_key_ref),
@@ -171,7 +169,6 @@ fn push_ops_event_predicates(
     range: ObservabilityRange,
     filter: &OpsErrorFilter,
 ) {
-    statement.push(" and (mr.id is null or mr.recovered_at is null)");
     push_range(statement, "oe.created_at", range);
     for (column, value) in [
         ("mr.client_api_key_ref", &filter.client_api_key_ref),