GitLab

injectClaudeCodePrompt 和 systemIncludesClaudeCodePrompt 的 type switch
无法匹配 json.RawMessage 类型（Go typed nil 陷阱），导致 ForwardAsResponses
和 ForwardAsChatCompletions 路径中用户 system prompt 被替换为仅 Claude Code
banner。新增 normalizeSystemParam 将 json.RawMessage 转为标准 Go 类型。
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

刷新失败不代表 access_token 无效，在后台定时刷新（不可重试错误 +
重试耗尽）和前端批量/单次刷新的失败路径中，均利用可能仍有效的
access_token 调用隐私设置。
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

复用已有的 set-privacy API 端点，Handler 通过 platform 分发到
ForceOpenAIPrivacy / ForceAntigravityPrivacy，前端 AccountActionMenu
扩展隐私按钮支持 OpenAI OAuth 账号。
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

新增功能：
- 新增 TLS 指纹 Profile CRUD 管理（Ent schema + 迁移 + Admin API + 前端管理界面）
- 支持账号绑定数据库中的自定义 TLS Profile，或随机选择（profile_id=-1）
- HTTPUpstream.DoWithTLS 接口从 bool 改为 *tlsfingerprint.Profile，支持按账号指定 Profile
- AccountUsageService 注入 TLSFingerprintProfileService，统一 usage 场景与网关的 Profile 解析逻辑

代码优化：
- 删除已被 TLSFingerprintProfileService 完全取代的 registry.go 死代码（418 行）
- 提取 3 个 dialer 的重复 TLS 握手逻辑为 performTLSHandshake() 共用函数
- 修复 GetTLSFingerprintProfileID 缺少 json.Number 处理的 bug
- gateway_service.Forward 中 ResolveTLSProfile 从重试循环内重复调用改为预解析局部变量
- 删除冗余的 buildClientHelloSpec() 单行 wrapper 和 int64(e.ID) 无效转换
- tls_fingerprint_profile_cache.go 日志从 log.Printf 改为 slog 结构化日志
- dialer_capture_test.go 添加 //go:build integration 标签，防止 CI 失败
- 去重 TestProfileExpectation 类型至共享 test_types_test.go
- 修复 9 个测试文件缺少 tlsfingerprint import 的编译错误
- 修复 error_policy_integration_test.go 中 handleError 回调签名被错误替换的问题

新增独立开关控制 API Key 账号的签名整流功能，支持配置自定义
匹配关键词以捕获不同格式的上游错误响应。

- 新增 apikey_signature_enabled 开关（默认关闭）
- 新增 apikey_signature_patterns 自定义关键词配置
- 内置签名检测规则对 API Key 账号同样生效
- 自定义关键词对完整响应体做不区分大小写匹配
- 重试二阶段检测仅做模式匹配，不重复校验开关
- Handler 层校验关键词数量（≤50）和长度（≤500）
- API 响应 nil patterns 统一序列化为空数组
- OAuth/SetupToken/Upstream/Bedrock 账号行为不变

- 新增 header_util.go，通过 setHeaderRaw/getHeaderRaw/addHeaderRaw 绕过
  Go 的 canonical-case 规范化，保持真实 Claude CLI 抓包的请求头大小写
  （如 "x-app" 而非 "X-App"，"X-Stainless-OS" 而非 "X-Stainless-Os"）
- 新增管理后台开关：指纹统一化（默认开启）和 metadata 透传（默认关闭），
  使用 atomic.Value + singleflight 缓存模式，60s TTL
- 调试日志从控制台 body 打印升级为文件级完整快照
  （按真实 wire 顺序输出 headers + 格式化 JSON body + 上下文元数据）
- 恢复 accept-encoding 到白名单，在 http_upstream.go 新增 decompressResponseBody
  处理 gzip/brotli/deflate 解压（Go 显式设置 Accept-Encoding 时不会自动解压）
- OAuth 服务 axios UA 从 1.8.4 更新至 1.13.6
- 测试断言改用 getHeaderRaw 适配 raw header 存储方式

- 401 token_invalidated / token_revoked: OAuth token 被永久作废，跳过临时不可调度逻辑，直接 SetError
- 402 deactivated_workspace: 解析 detail.code 字段，标记工作区已停用
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- gofmt: 修复 admin_service/antigravity_oauth_service/token_refresh_service 格式
- staticcheck S1009: 移除 SetUserSettingsResponse.IsSuccess 中冗余的 nil 检查
- unused: 将仅测试使用的 applyAntigravitySubscriptionResult 移至测试文件

Made-with: Cursor

复用已有 GetTier() 返回的 tier ID（free-tier / g1-pro-tier /
g1-ultra-tier），通过 TierIDToPlanType 映射为 Free / Pro / Ultra，
在 loadProjectIDWithRetry 中顺带提取并写入 credentials.plan_type；
前端增加 Abnormal 异常套餐红色标记。

Made-with: Cursor

新增 Antigravity OAuth 隐私设置能力，在账号创建、刷新、导入和后台
Token 刷新路径自动调用 setUserSettings + fetchUserInfo 关闭遥测；
持久化后同步内存 Extra，错误处理改为日志记录。

Made-with: Cursor

1. accounts/check 补全 plan_type：当 id_token 缺少 plan_type（如 Mobile RT），
   自动调用 accounts/check 端点获取订阅类型
2. orgID 精确匹配账号：从 JWT 提取 poid 匹配正确账号，避免 Go map
   遍历顺序随机导致 plan_type 不稳定
3. RT 刷新时设置隐私：调用 disableOpenAITraining 关闭训练数据共享，
   结果存入 extra.privacy_mode，后续跳过重复设置
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

New forwarding methods on GatewayService for Anthropic platform groups:

- ForwardAsResponses: accept Responses body → convert to Anthropic →
  forward to upstream → convert response back to Responses format.
  Supports both streaming (SSE event-by-event conversion) and buffered
  (accumulate then convert) response modes.
- ForwardAsChatCompletions: chain CC→Responses→Anthropic for request,
  Anthropic→Responses→CC for response. Streaming uses dual state machine
  chain with [DONE] marker.

Both methods reuse existing GatewayService infrastructure:
buildUpstreamRequest, Claude Code mimicry, cache control enforcement,
model mapping, and return UpstreamFailoverError for handler-level retry.

Empty text blocks inside tool_result.content were not being filtered,
causing upstream 400 errors: 'text content blocks must be non-empty'.

Changes:
- Add stripEmptyTextBlocksFromSlice helper for recursive content filtering
- FilterThinkingBlocksForRetry now recurses into tool_result nested content
- Add StripEmptyTextBlocks pre-filter on initial request path to avoid
  unnecessary 400+retry round-trips
- Add unit tests for nested empty text block scenarios

feat(ops): propagate endpoint/request-type context in handlers; add UpstreamURL to upstream error events

LegacyPrintf uses inferStdLogLevel() to infer log level from message
text. Any message containing the word "error" is classified as ERROR
level, causing the entire signature-retry recovery flow (which succeeds)
to produce spurious ERROR log entries.

Changes:
- Remove noisy [SignatureCheck] debug logs inside isThinkingBlockSignatureError
  that were logging every detected signature check as ERROR
- Change retry-start log to WARN level via [warn] prefix
- Change retry-success log to INFO level by removing "error" from message
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent

)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

1. Start() 中 wg.Add(1) 从 run() goroutine 内部移到 go s.run() 之前，
   防止 Stop().wg.Wait() 在 Add 之前返回导致孤儿 goroutine。
2. tryAcquireLeaderLock 返回的 release 闭包改用独立的
   context.Background()+5s 超时，避免捕获的 evaluateOnce ctx
   在 defer 执行时已过期导致锁释放失败（最长阻塞 90s TTL）。

The quota reset mechanism is lazy — quota_daily_used/quota_weekly_used
in the database are only reset on the next IncrementQuotaUsed call.
The scheduling layer (IsQuotaExceeded) correctly checks period expiry
before enforcing limits, so the account remains usable. However, the
API response mapper reads the raw DB value without checking expiry,
causing the frontend to display cumulative usage (e.g. 110%) even
after the reset period has passed.

Add IsDailyQuotaPeriodExpired/IsWeeklyQuotaPeriodExpired methods and
use them in the mapper to return used=0 when the period has expired.

Add maximum Claude Code version limit to complement the existing minimum
version check. Refactor the version cache from single-value to unified
bounds struct (min+max) with a single atomic.Value and singleflight group.

- Backend: new constant, struct field, cache refactor, validation (semver
  format + cross-validation max >= min), gateway enforcement, audit diff
- Frontend: settings UI input, TypeScript types, zh/en i18n
- Add CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 to all Claude Code
  tutorials on /keys page (unix/cmd/powershell/vscode settings.json)

shouldMarkCreditsExhausted was blocked by isURLLevelRateLimit check when
credit overages retry returned "Resource has been exhausted (e.g. check quota).",
causing credits to never be marked as exhausted. This led to an infinite loop
where each request injected credits, bypassed model rate limits, and failed again.

- Remove isURLLevelRateLimit guard from shouldMarkCreditsExhausted (only called
  for credit retry responses — if credits retry fails, mark exhausted)
- Add "resource has been exhausted" to creditsExhaustedKeywords
- Update tests to match corrected behavior

## Problem

When a proxy is unreachable, token refresh retries up to 4 times with
30s timeout each, causing requests to hang for ~2 minutes before
failing with a generic 502 error. The failed account is not marked,
so subsequent requests keep hitting it.

## Changes

### Proxy connection fast-fail
- Set TCP dial timeout to 5s and TLS handshake timeout to 5s on
  antigravity client, so proxy connectivity issues fail within 5s
  instead of 30s
- Reduce overall HTTP client timeout from 30s to 10s
- Export `IsConnectionError` for service-layer use
- Detect proxy connection errors in `RefreshToken` and return
  immediately with "proxy unavailable" error (no retries)

### Token refresh temp-unschedulable
- Add 8s context timeout for token refresh on request path
- Mark account as temp-unschedulable for 10min when refresh fails
  (both background `TokenRefreshService` and request-path
  `GetAccessToken`)
- Sync temp-unschedulable state to Redis cache for immediate
  scheduler effect
- Inject `TempUnschedCache` into `AntigravityTokenProvider`

### Account failover
- Return `UpstreamFailoverError` on `GetAccessToken` failure in
  `Forward`/`ForwardGemini` to trigger handler-level account switch
  instead of returning 502 directly

### Proxy probe alignment
- Apply same 5s dial/TLS timeout to shared `httpclient` pool
- Reduce proxy probe timeout from 30s to 10s

- 新增分组列：展示用户的专属/公开分组，支持 hover 查看详情
- 新增分组筛选：下拉选择或模糊搜索分组名过滤用户
- 专属分组替换：点击专属分组弹出操作菜单，选择目标分组后
  自动授予新分组权限、迁移绑定的 Key、移除旧分组权限
- 后端新增 POST /admin/users/:id/replace-group 端点，事务内
  完成分组替换并失效认证缓存

- 接入 gpt-5.4-mini/nano 模型识别与规范化，补充默认模型列表
- 增加 gpt-5.4-mini/nano 输入/缓存命中/输出价格与计费兜底逻辑
- 同步前端模型白名单与 OpenCode 配置
- 补充 service tier(priority/flex) 计费回归测试