Skip to content

fix(gateway): fail over on account-scoped Antigravity and Codex rejections - #8781

Open
Henglie wants to merge 1 commit into
Wei-Shaw:mainfrom
Henglie:fix/antigravity-codex-failover
Open

Henglie wants to merge 1 commit into
Wei-Shaw:mainfrom
Henglie:fix/antigravity-codex-failover

Conversation

@Henglie

@Henglie Henglie commented Nov 14, 2026

Copy link
Copy Markdown
Contributor

问题

  • Antigravity 账号不支持请求的模型时,上游返回 404 Requested entity was not found.,网关把它当普通 404 直接回写客户端,不换号。
  • OpenAI Codex gpt-6 账号触发 Invalid prompt ... potentially violating our usage policy 误报时,同样直接失败,不换号。

改动

  • Antigravity 三个入口(native Gemini / Claude / compat)识别精确 404 Requested entity was not found.,返回 UpstreamFailoverError,RetryableOnSameAccount=false,跳过账号冷却/停用副作用。
  • OpenAI 仅对 OAuth-like 账号的 400/404 invalid_prompt + prompt was flagged as potentially violating our usage policy 识别为请求级换号;普通 400/404 与 API Key 内容拒绝保持终止。
  • 不修改通用 shouldFailoverUpstreamError 策略,避免把真实客户端错误也当可重试。

验证

  • go test -tags=unit ./internal/service -run 'Test(Antigravity...|OpenAICodex...)' -count=1 → PASS
  • go test -tags=unit ./internal/service -count=1 -skip '^TestOllamaProbeCallback_StaleLongDoesNotOverrideNewShort$' → PASS
  • go vet -tags=unit ./internal/service、gofmt -l → PASS
  • 回退证明:仅回退 6 个源码文件、保留新测试时,Antigravity 各入口断言 UpstreamFailoverError 失败,OpenAI shouldFailover... 期望 true 实际 false;恢复后全部通过。

未验证

  • 未对真实 Antigravity / OpenAI 上游做端到端复现。
  • 未覆盖 HTTP 200 SSE response.failed 事件内的 Codex invalid_prompt(本次只处理 HTTP 400/404 响应分类)。
  • 未跑全仓 go test ./...。

修复 #8679

…tions

Antigravity returns 404 "Requested entity was not found." when an account's
model mapping points at a model the account cannot serve. Treat only that
exact 404 as an account/model availability signal so the handler can select
another account, without cooling or disabling the current one.

Codex gpt-6 accounts can also reject every prompt with a false-positive
invalid_prompt usage-policy error (openai/codex#52846). Classify that exact
marker for OpenAI OAuth-like accounts as request-level failover while keeping
ordinary 400/404 responses and API-key content rejections terminal.

Both paths skip account-health side effects and never retry the same account,
since it cannot produce a different result.
riofutab added a commit to riofutab/sub2api that referenced this pull request Nov 17, 2026
- 带入上游 11-16 已合的 55 个 PR,同步基准更新为 9cd1a1e;用量汇总改为非阻塞建表(Wei-Shaw#8693),被动用量换窗清理与 Fable 限额恢复采用上游实现(Wei-Shaw#8723 Wei-Shaw#8779)。
- Codex 独立上下文改用上游 Wei-Shaw#8769 的 user 文本归一化,撤回 fork 的 Wei-Shaw#8680 并补回 proposed_plan、guardian、cloud_threads 等类型;Gemini 思考等级改用 Wei-Shaw#8673,撤回 fork 的 Wei-Shaw#8619。
- 合入 24 个开放 PR,Wei-Shaw#8781 与 Wei-Shaw#8705 部分合入;Wei-Shaw#8856、Wei-Shaw#8858 与已有实现重复,未合入。
- 修复转义的 1h TTL 被快速路径漏判,并按 fork 的构造参数与模型目录调整上游测试。
- 后端全量 unit、golangci-lint、前端 vitest、迁移空库回放与 CI 集成测试通过;go.sum 未整理。
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant