Repository navigation
[Bug] gateway: message level output_config is stripped when the request carries the per-turn-control beta #8711
Description
When a /v1/messages request carries anthropic-beta: per-turn-control-2026-07-01, the gateway removes output_config from the messages that have one. The client's mid session effort change never reaches Anthropic.
The check is in stripAnthropicMessageOutputConfigUnlessBeta in backend/internal/service/gateway_request.go:
sub2api/backend/internal/service/gateway_request.go
Lines 1198 to 1201 in 530c5e3
It returns the body untouched only when the header contains claude.BetaMidConversationOutputConfig, which is mid-conversation-output-config-2026-07-01, defined at
per-turn-control-2026-07-01, so with only that beta the field is stripped, and a system message with an empty content is dropped completely.
To reproduce, send this to /v1/messages and look at the body the gateway forwards:
anthropic-beta: per-turn-control-2026-07-01
{
"model": "claude-opus-5-5",
"max_tokens": 256,
"messages": [
{"role": "user", "content": "Hello"},
{"role": "assistant", "content": "Hi, how can I help?"},
{"role": "system", "content": [], "output_config": {"effort": "low"}},
{"role": "user", "content": "Summarize that in one sentence."}
]
}The forwarded body has no output_config on the third message, and the whole message is gone because its content is empty.
Expected: the body is forwarded unchanged, the same as with mid-conversation-output-config-2026-07-01. Claude Code sends per-turn-control-2026-07-01 in its anthropic-beta list, for example in the header capture described in #8394, and a message level output_config on a system message is the shape Anthropic documents for changing effort partway through a conversation: https://platform.claude.com/docs/en/build-with-claude/effort-per-message
Activity
r266-tech commentedon Nov 12, 2026
核对了当前官方 per-message effort 文档：beta header、请求示例和可用性表列出的仍是
mid-conversation-output-config-2026-07-01。该页没有把per-turn-control-2026-07-01列为这个参数的别名。我用与
main@530c5e3a4相同的请求构造代码做了 9 个独立本地对照，使用本 issue 的四条消息和顶层effort: high：per-turn-control-2026-07-01mid-conversation-output-config-2026-07-01，或同时带两者per-turn-control-2026-07-01这确认了 API Key 路径上的本地差异，也给出了按当前文档可用的最小对照；尚未验证真实上游是否接受
per-turn-control单独授权这个字段。若已有该别名成功保留消息级 effort 的上游响应，请补充脱敏后的账号类型、是否透传、最终 beta header 及 HTTP 状态/错误，便于区分公开协议、账号专属行为或需要支持的别名。无需附密钥或完整会话。验证使用 Go 1.27.3/macOS arm64，未调用真实 Anthropic。AI 辅助源码核对和合成请求验证。
Anankke commentedon Nov 12, 2026
Tested on the real upstream through gateway that doesn't strip it with OAuth account not passthrough and claude-opus-5, sent on CC with
anthropic-beta: claude-code-20250219,per-turn-control-2026-07-01and a system message carryingoutput_configgot 200, so upstream accepts the field under that betaRow 3 in your table is the mimic path for clients that aren't Claude Code. Real Claude Code on OAuth isn't mimicked, so main drops the message there too
r266-tech commentedon Nov 12, 2026
你说得对：我原表格的 OAuth 行只验证了非 Claude Code 客户端的 mimic 分支，应明确标注为 “OAuth mimic”。它不能概括真实 Claude Code 的 OAuth 请求；此前的结论漏掉了这条路径。
已在
main@530c5e3a4的真实Forward入口补充对照，用 Claude Code UA 和合法metadata.user_id触发非 mimic 路径。结果确认：仅带claude-code-20250219,per-turn-control-2026-07-01时，真实 Claude Code OAuth、API Key 普通和 API Key 透传三条路径都会删除该控制消息；mimic 因最终 header 自带旧 beta 而保留。顶层 effort 均不受影响。修复已提交为 https://github.com/Wei-Shaw/sub2api/pull/8742（`bd89d3d94`）：按最终 header 中任一相关 beta 保留消息级
output_config；无相关 beta 或两者均被策略过滤时保持原行为。39 项相关 test/subtest、完整 unit（26,895 通过 / 41 跳过）、integration（16,143 通过 / 16 跳过）及 lint 均已通过。PR 中保留并披露了首轮计费内存断言失败、main 对照复现和后续复验结果。你的 OAuth + Claude Code +
claude-opus-5HTTP 200 是这里的真实上游接受证据；本账号完成的是本地合成请求验证，没有请求真实 Anthropic，也未验证模型实际 effort 效果或 API Key 的真实上游行为。当前官方文档仍列旧 beta，不能把上述实测写成官方文档已声明两者等价。AI 辅助实现、验证与回复。