5524 Commits
Author SHA1 Message Date
wucm667 ece1ecccca fix(gateway): normalize streamed Anthropic usage 2026-09-21 12:58:56 +08:00
wucm667 b13200d7a1 fix(gateway): forward received chat stream usage 2026-09-21 12:58:56 +08:00
wucm667 de98bdff1e fix(images): preserve insufficient balance failures 2026-09-21 10:23:10 +08:00
aofee d1ba57977d feat: add native TypeSafe Jev System One support 2026-09-20 20:20:16 +08:00
xlplbo 0d02000592 fix(openai): 非高级调度按 previous_response 路由到持有响应的账号 2026-09-20 19:28:13 +08:00
xlplbo 8d88b66668 fix(openai): 补齐非高级调度选号决策标签 2026-09-20 19:27:53 +08:00
xlplbo 763443abf2 fix(openai): 改写 Codex turn metadata 时保留非 ASCII 转义 2026-09-20 19:27:47 +08:00
HuntercodeT fb58dfbd93 test(antigravity): cover the identity rewrite end-to-end through the transformer
Adds a case with the real block layout (attribution block, identity sentence,
task instructions) and asserts that user content mentioning Claude Code and
claude-sonnet-4-6 is forwarded unchanged.
2026-09-20 16:48:04 +08:00
haruka b8d52fad34 feat(billing): support configurable reasoning effort multipliers 2026-09-20 16:41:49 +08:00
shaw d2e319b2a1 test: 修复 WebSocket 同线程抢占测试的关闭时序竞态 2026-09-20 16:37:07 +08:00
HuntercodeT dd292e3b00 fix(antigravity): neutralize Claude Agent SDK identity in system prompts
Stripping the x-anthropic-billing-header line is not enough. Claude Code's
Agent SDK prompts (session title generation, and other SDK-driven calls) open
a separate system block with

    You are a Claude agent, built on Anthropic's Claude Agent SDK.

and the Google upstream still rejects those requests with 429 RESOURCE_EXHAUSTED.

Measured on Sub2API 0.2.7, which already contains the attribution fix, using an isolated single-account group, waiting for the per-model cooldown
to expire before every request so that a previous failure could not turn later
results into 503:

  original request replayed                      429 (8/8)
  same request, generic system prompt            200 (8/8)
  attribution also removed client-side           429 (3/3)
  attribution removed + identity neutralized     200 (3/3)
  only 'Claude' replaced in system               200 (2/2)
  only 'Anthropic' removed from system           200 (2/2)
  output_config / max_tokens / non-stream        no effect

The rewrite is anchored at the start of a system block, so user instructions that
merely mention Claude or Anthropic, and model ids such as claude-sonnet-4-6, are
left untouched. Scoped to the Antigravity transformer, as with the existing
attribution handling.
2026-09-20 16:34:26 +08:00
Wesley Liddick acdf3c54da Merge pull request #7401 from StarryKira/codex/codex-credits-referrals-upstream
feat(openai): display Codex credits and support referral invitations
2026-09-20 16:00:28 +08:00
haruka fc96132e72 fix(openai): harden referral outcomes and isolate HTTP transport 2026-09-20 15:47:53 +08:00
haruka 8f6bbb59c7 test: stabilize Gemini keepalive test 2026-09-20 15:04:15 +08:00
shaw e009ea3036 fix(upstream): 恢复 OpenAI HTTP/2 保活容错期限 2026-09-20 14:46:25 +08:00
a408665321 63c079d873 fix(account): preserve settings on OAuth reauth 2026-09-20 14:39:36 +08:00
haruka 5090ffe05e feat(openai): show Codex credits and manage referral invitations 2026-09-20 14:38:29 +08:00
Wesley Liddick bbdcfbac0a Merge pull request #7304 from wucm667/fix/issue-7298-response-model-alias
fix(gateway): restore public response model aliases
2026-09-20 14:07:16 +08:00
Wesley Liddick 7403a01177 Merge pull request #7314 from wucm667/fix/issue-7306-keyword-reminder-bypass
fix(moderation): prevent reminder tags from bypassing keyword checks
2026-09-20 14:07:04 +08:00
Wesley Liddick 53b4bbe736 Merge pull request #7349 from wucm667/fix/issue-7348-preserve-client-version
fix(settings): preserve client version on invalid user agent
2026-09-20 14:06:48 +08:00
Wesley Liddick 8bb48622a3 Merge pull request #7362 from wucm667/fix/issue-7358-preserve-account-errors
fix(accounts): preserve refresh errors during usage queries
2026-09-20 14:06:36 +08:00
Wesley Liddick f28adb6ddb Merge pull request #7387 from Brisbanehuang/codex/upstream-content-audit-typesafe
feat: 内容审计接入 TypeSafe AI,支持双引擎切换与独立配置
2026-09-20 13:47:14 +08:00
shaw 29eead917f feat(pluginapi): HostService 账号目录返回结构化只读元数据
纯宿主侧改动,向后兼容。

- 契约(pkg/pluginapi/v1/plugin.proto,已重新生成):ListAccounts 增补
  repeated AccountInfo accounts(保留 account_ids 向后兼容),新增 metadata_json
  只读快照字段。HostServiceAPIVersion 1→2(InitHostServices 运行时协商,加法式,
  旧调用方忽略新字段即可)。
- 权限边界:可见范围由清单能力经 pluginCapabilityAccountScopeGrants 推导,
  ListAccounts 与 ResolveOutboundIdentity 均以该范围为权威边界。
- metadata_json 采用 denylist:剥离 Credentials 与含反向引用的关联对象
  Groups/AccountGroups(避免 encoding/json 循环引用),其余字段自动透传,
  数据模型新增非机密字段无需改宿主。

测试:范围/元数据/无凭据泄漏用例、denylist 反射护栏 + 循环引用安全断言、纵深过滤用例;
build/vet/gofmt、-tags unit 与 -race 全绿。
2026-09-20 11:26:19 +08:00
Achord Chan 1411b7eb7e fix(gateway): recognize Baseten reasoning budget errors 2026-09-20 11:02:06 +08:00
Achord Chan ef51690b5a fix(antigravity): preserve string const constraints in tool schemas 2026-09-20 10:46:09 +08:00
Brisbanehuang 1bfe0d3728 fix: sync draft audit key status and threshold defaults 2026-09-19 20:52:31 -04:00
Brisbanehuang 783a0a9830 feat(content-audit): add independent TypeSafe engine profiles 2026-09-19 20:52:31 -04:00
wucm667 b252821c59 fix(accounts): preserve refresh errors during usage queries 2026-09-19 17:22:39 +08:00
wucm667 68aa26477f feat(simple): make startup default group creation optional 2026-09-19 16:47:17 +08:00
Long LiandCursor 7e530c9d4e fix(openai): alias DeepSeek Responses input_image with url
Codex sends images as input_image.image_url. DeepSeek's live /responses
decoder requires a url field and returns 422 missing field url, including
on openai-platform accounts whose base_url host is api.deepseek.com.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-19 14:55:02 +09:00
Wesley Liddick aea725f2ea Merge pull request #7247 from Qiaogun/codex/seedance-native-api
feat: 支持 Seedance / Ark 原生视频任务 API
2026-09-19 12:11:17 +08:00
Wesley Liddick 74431ffad8 Merge pull request #7258 from HuntercodeT/fix/antigravity-gemini-bare-model-thinking-variant
fix(antigravity): Gemini 原生请求的裸模型名按 thinkingConfig 解析到 -low/-medium/-high 变体
2026-09-19 12:01:22 +08:00
Wesley Liddick 0c9e83b59c Merge pull request #7278 from HuntercodeT/fix/antigravity-gemini-sse-comment-keepalive
fix(antigravity): Gemini 原生流对 go-genai / python-genai 客户端不发 SSE 注释心跳
2026-09-19 12:00:18 +08:00
Wesley Liddick 1b45d21328 Merge pull request #7340 from ye4241/fix/kimi-cn-quota-403-rate-limit-pause
fix(ratelimit): CN coding-plan 账号配额耗尽的 403 按限流暂停调度,而非永久禁用
2026-09-19 11:59:21 +08:00
Wesley Liddick d5cea617a5 Merge pull request #7345 from laoxiechuzheng/fix/anthropic-root-union-schema
fix(apicompat): 摊平 Responses→Anthropic 工具 schema 顶层联合,修复 Codex 内置工具导致的 400
2026-09-19 11:59:00 +08:00
Wesley Liddick 4275047d16 Merge pull request #7313 from baryon/contrib/deepseek-chat-reasoning-content
fix(openai): pass DeepSeek thinking-mode reasoning_content on chat fallback
2026-09-19 11:57:55 +08:00
shaw 6477143533 test(plugin): check type assertions to satisfy errcheck lint 2026-09-19 11:47:14 +08:00
shaw c63bd14a02 feat(plugin): generic host services + read-only status bridge channel
Extend the local plugin system's host<->plugin contract with generic,
capability-neutral facilities, so transport plugins can be built on top
without embedding plugin-specific logic in the host.

Host services (plugin dials a broker-provided HostService gRPC):
- per-plugin namespaced KV store (key/value size and TTL bounded)
- account directory: ListAccounts + ResolveOutboundIdentity, restricted to
  the accounts a plugin's declared capability already covers
- InitHostServices lifecycle hook; plugins that don't implement it run
  unchanged and are simply offered no host services

Read-only runtime status channel:
- HealthResponse gains an opaque, plugin-defined status_json produced
  passively (no config apply, no upstream call, no side effects)
- GET /admin/plugins/:id/status (ungated) and the plugin.status UI-bridge
  verb map to Health, so a status UI can poll live state without a step-up
  prompt or a "test" side effect
- config.test no longer force-shows a host success toast; the plugin UI owns
  result presentation, so plugins that read status do not spam toasts

Document the UI-bridge verb taxonomy (read vs mutate/test gating) and the
status contract under pkg/pluginapi.
2026-09-19 11:25:56 +08:00
wucm667 2d37088bd9 fix(settings): preserve client version on invalid user agent 2026-09-19 10:36:51 +08:00
laoxiechuzheng a8a1a3a958 fix(apicompat): flatten root union tool schemas for Anthropic
Codex 会把 codex_app 等内置工具的 parameters 根节点声明成对象分支的 oneOf/anyOf,Anthropic 拒绝工具 input_schema 顶层的 oneOf/anyOf/allOf,导致 /v1/responses -> /v1/messages 直接 400。

Responses->Anthropic 转换时摊平顶层联合:属性取各分支并集,同名属性用嵌套联合保留两侧约束,required 在 oneOf/anyOf 下取分支交集、allOf 下取并集;嵌套联合保持不变。
2026-09-19 07:34:21 +08:00
Rain 4ccdead2a2 fix(proxies): skip inactive fallback targets 2026-09-19 00:51:01 +08:00
Rain f7280213f7 fix(proxies): skip expiry writes from stale snapshots 2026-09-19 00:51:01 +08:00
ye4241andClaude Opus 4.8 db8692d679 fix(ratelimit): pause CN coding-plan accounts on quota-exhausted 403
Kimi reports Coding Plan window exhaustion (5h/weekly) as a 403 with
error.type=access_terminated_error instead of 429. It fell into the
generic handleOpenAI403 escalation, so three 403s within 180 minutes
permanently disabled accounts that recover by themselves at the window
reset.

Detect the quota-exhausted 403 for CN coding-plan accounts and cool it
down exactly like the 429 path: SetRateLimited until the earliest future
window reset from the quota snapshot, falling back to a short
temp-unschedulable window when the snapshot is missing. Non-coding-plan
accounts keep the generic 403 escalation since the rolling-window
recovery mechanism does not apply to them.

Also extract cooldownCNProviderToQuotaSnapshotReset and
setCNProviderTempUnschedulable so the 429 (OpenCodeGo / Coding Plan)
and 403 branches stop carrying three copies of the same sequence.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-09-19 00:00:57 +08:00
superman2003 37238c0985 fix(openai): honor error.status when classifying in-stream failures
OpenAI 兼容上游(含二级中转)常把流内 error / response.failed 事件的状态码写在
error.status 上,而不是 OpenAI 使用的 error.status_code。SSE 路径只读
status_code,于是这些上游的 401/403/429/529 全部被降级成通用 502:认证失败不
再判定为凭据故障、429 不再标记限流与配额窗口、error 帧丢掉 401/429/529 的
failover 快路径,运维日志也记成 502。WS 路径的 openAIWSPayloadTransientStatus
早已同时读这两种拼写,SSE 路径一直没跟上。

把状态码路径集合抽成共享常量,让 SSE 路径与 WS 路径对齐。5xx 仍按通用上游故障
处理,不改动既有的账号状态语义。
2026-09-18 23:51:08 +08:00
wucm667 7b4de8b6ae fix(moderation): prevent reminder tags from bypassing keyword checks 2026-09-18 14:54:18 +08:00
Long LiandCursor bcc73f8d40 fix(openai): pass DeepSeek thinking-mode reasoning_content on chat fallback
Codex histories often replay encrypted-only reasoning items. After a
Responses→Chat Completions fallback, a cache miss left assistant
messages without reasoning_content and DeepSeek rejected the turn
with 400. Fill empty assistant fields with a single-space placeholder
for DeepSeek hosts, including openai-platform accounts whose base_url
points at api.deepseek.com. Existing plaintext and cache hits are
left unchanged.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-18 15:42:22 +09:00
Yoga Sakti 4805f069e4 fix(scheduler): keep account RPM config in the scheduler projection
`filterSchedulerExtra` 的白名单没有 `base_rpm` / `rpm_strategy` /
`rpm_sticky_buffer`,写进 `sched:meta:<id>` 的精简投影因此丢掉整套 RPM 配置。
候选过滤走 `ListSchedulableAccounts`,缓存命中时读的正是这份投影:
`isAccountSchedulableForRPM` 拿到的账号上 `GetBaseRPM()` 返回 0,`baseRPM <= 0`
直接放行,已配置限流的账号在选号阶段被当作未开启 RPM,单个自然分钟内被超额调度。
缓存未命中时回退数据库读到完整 extra,所以现象是间歇性的。

选中的账号随后由 `hydrateSelectedAccount` 从 `sched:acc:<id>` 重新水合,
RPM 计数与用户消息队列限速读的是完整账号,不受影响;缺的只是选号阶段这道门。

修法沿用投影里既有的处理方式(见 `openai_passthrough` 一组注释):把判定依赖的
键补进白名单,而不是在判定侧另开一条读取完整账号的路径。

回归测试覆盖序列化往返后的三个键:`tiered` 子用例证明缓冲被裁掉时红区会被误判成
黄区,`sticky_exempt` 子用例证明策略被裁掉时会退回三区判定。

Refs #7225
2026-09-18 13:37:58 +07:00
HuntercodeT cfd2fc779b fix(codex-models): 非 GPT 模型的 Codex 提示词模板去掉 GPT 身份声明,避免 Antigravity 429
模型目录对所有模型都用 openai.CodexBaseInstructionsForModel 填 instructions_template,
gemini / claude 等因此下发 "You are Codex, a coding agent based on GPT-5."。Codex 把它作为
instructions 发出,Antigravity 对「Codex 提示词 + GPT-5 身份」返回 429 RESOURCE_EXHAUSTED,
并让账号进入限流冷却。

OpenAI GPT 模型保持原模板;其他模型族(含 gpt-oss)去掉 "based on GPT-N" /
"You are GPT-N" 身份声明,提示词其余部分不变。

Refs #7307
2026-09-18 14:21:21 +08:00
wucm667 0920e06a42 fix(gateway): restore public response model aliases 2026-09-18 13:31:58 +08:00
little-greenbean 3fdd54ca12 fix(upstream): cancel response attempt before closing body 2026-09-18 13:30:35 +08:00