GitLab

- Add sanitizeExtraBaseRPM to BulkUpdate handler (was missing)
- Add WindowCost scheduling checks to legacy non-sticky selection
  paths (4 sites), matching existing sticky + load-aware coverage
- Export ParseExtraInt from service package, remove duplicate
  parseExtraIntForValidation from admin handler

- Move IncrementRPM after Forward success to prevent phantom RPM
  consumption during account switch retries
- Add base_rpm input sanitization (clamp to 0-10000) in Create/Update
- Add WindowCost scheduling checks to legacy path sticky sessions
  (4 check sites + 4 prefetch sites), fixing pre-existing gap
- Clean up rpm_strategy/rpm_sticky_buffer when disabling RPM in
  BulkEditModal (JSONB merge cannot delete keys, use empty values)
- Add json.Number test cases to TestGetBaseRPM/TestGetRPMStickyBuffer
- Document TOCTOU race as accepted soft-limit design trade-off

Ensures isAccountSchedulableForRPM calls within the routing segment
hit the prefetch cache instead of querying Redis individually.

- Use TxPipeline (MULTI/EXEC) instead of Pipeline for atomic INCR+EXPIRE
- Filter negative values in GetBaseRPM(), update test expectation
- Add RPM batch query (GetRPMBatch) to account List API
- Add warn logs for RPM increment failures in gateway handler
- Reset enableRpmLimit on BulkEditAccountModal close
- Use union type 'tiered' | 'sticky_exempt' for rpmStrategy refs
- Add design decision comments for rdb.Time() RTT trade-off

- Add migration 060 to update model_mapping for all antigravity accounts
- Remove gemini-3-pro-image and gemini-3-pro-image-preview mappings
- Add gemini-3.1-flash-image and gemini-3.1-flash-image-preview mappings
- Update frontend usage window to show GImage for new model
- Update isImageGenerationModel to support new model

PR #635 returned HTTP 200 with {"input_tokens": 0} when upstream doesn't
support count_tokens (404). This caused Claude Code CLI to trust the zero
value, believing context uses 0 tokens, so auto-compression never triggers.

Fix: return 404 with proper error body so CLI falls back to its local
tokenizer for accurate estimation. Return nil (not error) to avoid
polluting ops error metrics with expected 404s.

Affected paths:
- Passthrough APIKey accounts: upstream 404 now passed through as 404
- Antigravity accounts: same fix (was also returning fake 200)

第三方 Anthropic 中转站通常不支持 /v1/messages/count_tokens 端点，
上游返回 404 时降级返回 {input_tokens: 0}，客户端 fallback 到本地估算。

- 仅匹配 404 状态码，语义明确：端点不存在
- 其他错误 (400/429/500) 保留原始处理链和 ops 遥测
- 无需解析错误消息内容，不依赖字符串匹配
- 新增 table-driven 测试覆盖 fallback 和 non-fallback 路径

Remove the special case that bypassed model-supported checks for Gemini
API Key accounts, allowing model_mapping to filter requests properly.
Add tests for multiplatform model filtering behavior.

- 新增 Anthropic API Key 自动透传开关与后端透传分支（仅替换认证）

- 账号编辑页新增自动透传开关，默认关闭

- 优化透传性能：SSE usage 解析 gjson 快路径、减少请求体重复拷贝、优化流式写回与非流式 usage 解析

- 补充单元测试与 benchmark，确保 Claude OAuth 路径不受影响

- 在 failover 场景透传上游响应头并识别 Cloudflare challenge/cf-ray

- 统一 Sora 任务请求的 UA 与代理使用，sentinel 与业务请求保持一致

- 修复流式错误事件 JSON 转义问题并补充相关单元测试

- Account-level cache TTL override: rewrite Anthropic cache_creation
  token classification (5m↔

1h) in streaming/non-streaming responses
- New DB field cache_ttl_overridden in usage_log for billing tracking
- Migration 055_add_cache_ttl_overridden
- Frontend: CacheTTL override toggle in account create/edit modals
- Ent schema regenerated for new usage_log fields
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- 仅在 delta 中 5m/1h 值大于0时覆盖 usage 明细
- 新增回归测试覆盖 delta 默认 0 不应覆盖 message_start 非零值
- 迁移 054 在删除 legacy 字段前追加一次回填，避免升级实例丢失历史写入
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Anthropic API 的 cache_creation 对象区分了 ephemeral_5m 和 ephemeral_1h
两种缓存创建 token，1h 单价远高于 5m（如 claude-3-5-haiku: 5m=$1/MTok,
1h=$6/MTok）。此前系统统一按 5m 单价计费，导致计费偏低。

后端：
- pricing_service: 加载 LiteLLM 的 cache_creation_input_token_cost_above_1hr
- billing_service: GetModelPricing 启用分类计费（安全守卫 1h>5m），
  CalculateCost 按 5m/1h 分别计费，无明细时回退到 5m 单价
- gateway_service: parseSSEUsage/handleNonStreamingResponse 用 gjson
  提取嵌套 cache_creation 对象的 ephemeral_5m/1h_input_tokens
- antigravity_gateway_service: extractSSEUsage/extractClaudeUsage 同步提取
- usage_log: 修复 GORM column tag 确保写入正确的数据库列
- 新增迁移 054: 删除 GORM 自动生成的重复列

前端：
- 使用记录 tooltip 展示 5m/1h 缓存创建明细（带彩色 badge 区分）
- 表格单元格缓存写入数值旁显示 1h 标识

- 将高密度服务与处理器日志迁移到新日志系统（LegacyPrintf/结构化日志）
- 增加 stdlog bridge 与兼容测试，保留旧日志捕获能力
- 将 OpenAI 断流告警改为结构化 Warn 并改造对应测试为 sink 捕获
- 补齐后端相关文件 logger 引用并通过全量 go test

✨ feat(antigravity): 支持 thinking adaptive 类型并适配 Opus 4.6 动态预算
🧪 test(gateway): 增加 thinking 模式解析与签名块过滤的边界用例测试

- gofmt: 修复 error_passthrough_service.go 格式问题
- errcheck: 修复 error_passthrough_runtime_test.go 类型断言未检查
- staticcheck: if-else 改为 switch (gateway_service.go)
- test: 修复两个测试用例错误使用 MODEL_CAPACITY_EXHAUSTED 导致走错路径

For retryable transient errors (Google 400 "invalid project resource name"
and empty stream responses), retry on the same account up to 2 times
(with 500ms delay) before switching to another account.

- Add RetryableOnSameAccount field to UpstreamFailoverError
- Add same-account retry loop in both Gemini and Claude/OpenAI handler paths
- Move temp-unschedule from service layer to handler layer (only after
  all same-account retries exhausted)
- Reduce temp-unschedule cooldown from 30 minutes to 1 minute

单账号 antigravity 分组收到 503 (MODEL_CAPACITY_EXHAUSTED) 时，
原逻辑会设置 ~29s 模型限流标记。由于只有一个账号无法切换，
后续所有新请求在预检查时命中限流 → 几毫秒内直接返回 503，
导致约 30 秒的雪崩窗口。

修复：在 Handler 入口处检查分组是否只有单个 antigravity 账号，
如果是则提前设置 SingleAccountRetry context 标记，让 Service 层
首次 503 就走原地重试逻辑（不设限流标记），避免污染后续请求。

Merge functional changes from develop branch:
- Remove AntigravityQuotaScope system (claude/gemini_text/gemini_image)
- Replace with per-model rate limiting using resolveAntigravityModelKey
- Remove model load statistics (IncrModelCallCount/GetModelLoadBatch)
- Simplify account selection to unified priority→load→LRU algorithm
- Remove SetAntigravityQuotaScopeLimit from AccountRepository
- Clean up scope-related UI indicators and API fields

Add post-sort shuffle for accounts with identical (priority, loadRate,
lastUsedAt) to break deterministic ordering when concurrent requests
read the same scheduler snapshot. Applies to both Antigravity and
OpenAI scheduling paths, plus the sortAccountsByPriorityAndLastUsed
helper.

Keeps upstream CallCount/ModelLoadInfo scheduling intact; shuffle is
additive and only randomises within equivalent-rank groups.

ParseGatewayRequest only parsed Anthropic format (system/messages),
ignoring Gemini native format (systemInstruction/contents). This caused
GenerateSessionHash to produce identical hashes for all Gemini sessions.

Add protocol parameter to ParseGatewayRequest to branch between
Anthropic and Gemini parsing. Update GenerateSessionHash message
traversal to extract text from both formats.

Mix SessionContext (ClientIP, UserAgent, APIKeyID) into
GenerateSessionHash 3rd-level fallback to differentiate requests
from different users sending identical content.

Also switch hashContent from SHA256-truncated to XXHash64 for
better performance, and optimize Trie Lua script to match from
longest prefix first.

The previous fallback (step 3) in GenerateSessionHash hashed system +
all messages together, producing a different hash each round as the
conversation grew ([a] -> [a,b] -> [a,b,c]). This made fallback sticky
sessions ineffective for multi-turn conversations.

Implement per-message Trie digest chain matching (reusing Gemini's Trie
infrastructure) so that the previous round's chain is always a prefix
of the current round's chain, enabling reliable session affinity.

SSE 热路径中 replaceModelInSSELine 和 replaceModelInResponseBody 原来
使用 json.Unmarshal/Marshal 对每个事件做全量反序列化再序列化，现改为
gjson.Get/sjson.Set 精确字段操作，消除 O(n) 中间 map 分配，保持 JSON
字段顺序不变。涉及 OpenAIGatewayService 和 GatewayService 两个服务。

新增 23 个单元测试覆盖：顶层/嵌套 model 替换、不匹配跳过、空行/[DONE]/
非法 JSON 等边界情况。

Fixes: P1-08
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>