Repository navigation
release: v0.4.20 同步上游、更新无限画布并新增分组强制 Fast 计费 - #59
Merged
Merged
Conversation
/keys 页面要等密钥列表和每个 key 的用量统计都返回,才撤掉表格骨架屏。 用量统计(POST /usage/dashboard/api-keys-usage)按 api_key_id 汇总近 30 天的 usage_logs.actual_cost。数据量大时,要逐行回表读取几十万到上百万行,接口耗时 十几到三十几秒,甚至超时,页面就一直停在骨架屏。 - 前端:列表返回就渲染表格,用量列单独显示加载占位;统计失败时回落为 $0.0000。 - 后端:新增覆盖索引 (api_key_id, created_at) INCLUDE (actual_cost),使该统计可以 index-only scan,不再回表。迁移采用 CONCURRENTLY 建索引;与其他 usage_logs 索引迁移一样,重试前会先清理中断构建留下的 INVALID 索引。 Refs #5858 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
异步审计 worker 每 500ms 轮询一次数据库认领任务,本进程刚入队的任务 平均要等约 250ms 才被处理。入队成功和租约回收后改为立即通知本进程的 worker 尝试认领;worker 启动时先认领一次遗留任务。500ms 轮询保留不变, 继续兜底到期重试和其他副本入队的任务。
各 tab 表单很长,保存按钮只在滚到最底部时可见。改为 sticky 浮标贴在视口右下角, 样式沿用已有 tab 导航条的毛玻璃胶囊语言,深色模式覆盖放在非 scoped 块内。
Problem: SSE streams decide first-token time with the admin openai_ttft_mode setting (semantic: first non-lifecycle event; visible: first visible output), but WebSocket passthrough turns always wait for the first ".delta" event and ignore the setting. On reasoning models the reasoning item is announced when reasoning starts while the first delta only arrives after it ends, so in semantic mode a WS turn reports the reasoning time as first-token latency and an HTTP request does not. The same user's WS turns look seconds slower than their HTTP requests, and the gap grows with reasoning tokens while the HTTP numbers stay flat. Fix: let RelayOptions take the first-token rule and have the passthrough adapter pass the configured TTFT mode through the same check the SSE path uses (openAIStreamDataStartsTTFT). codex.* frames are WebSocket-only out-of-band events and never count; terminal events still never count, as before. Without a rule the relay keeps the delta-based behaviour. Test: TestOpenAIWSPassthroughFirstTokenFollowsTTFTMode replays two reasoning turns on one passthrough connection in both modes and expects each turn's first token at the event the SSE rule picks. TestRelay_IsFirstTokenEventDecidesEveryTurnFirstToken checks the relay applies the injected rule per turn and still skips terminal events.
部分上游在未请求 includeThoughts 时仍返回 thought=true 的 part,非流式转换把它当正文输出,导致 chat/messages 的 content 混入思考文字。转换时跳过思考 part 的文字,与流式路径的分流口径一致;用量仍取自 usageMetadata,计费不变。
…e-calls fix(apicompat): preserve sealed function call metadata
…envelope fix(apicompat): skip foreign CC Switch reasoning envelopes
…n-prefix fix(openai): stabilize Responses session seeds across tool loops
…licy fix(gateway): honor account error policies in compatibility paths
fix(gemini): 兼容入口非流式响应不再把思考摘要混入正文
…l-result-images fix(antigravity): keep parallel tool functionResponses contiguous before media
fix(antigravity): isolate Pro Argon quota cooldowns
…code fix(payment): configure EasyPay refund success codes per instance
…error fix(scheduling): release acquired slot when account hydration fails
…rotation fix(auth): consume refresh tokens atomically during rotation
…oning-mapping fix(openai): apply mappings to synthesized reasoning effort
perf(keys): API 密钥列表不再等待 30 天用量统计,并为其加覆盖索引
fix(accounts): retain rows and quota values while refreshing
…er-6-decimals fix(payment): support up to 6 decimals for balance recharge multiplier
feat(settings): 系统设置保存按钮改为底部吸附浮标
fix(openai): normalize Codex Desktop page context before tool validation
…tatus fix(openai): preserve native Codex image error statuses
…-invalid-utf8 fix(openai): reject invalid UTF-8 Responses bodies
fix(securityaudit): 审计任务入队后立即唤醒本进程 worker
feat: add Claude Fable 5.2 and Mythos 5.2 support
Merge upstream 530c5e3 into sync-upstream. Preserve ModuRelay branding, billing recovery, session binding, routing, and TTFT fields while integrating atomic refresh tokens and cache TTLs. Prior local verification passed backend unit/race and integration checks, frontend tests, typecheck, lint, build, and the browser matrix. Environment-dependent checks remain skipped; production is unverified.
- 开启后 OpenAI Token 请求统一按 priority 计费,忽略上游返回档位 - 默认关闭,关闭后保持原有计费逻辑 - 与免费 Fast 互斥,独立于强制 Fast 请求开关 - 完善配置持久化、认证缓存及计费回归测试
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
验证:最新提交的 CI、Security Scan 共 7 项检查全部通过,已完成用户验收。