Repository navigation
fix(gemini): 修复 Gemini 流式响应提前结束与静默截断问题 - #8980
universonic wants to merge 4 commits into
Conversation
30a8861 to
554cc17
Compare
…tions - Route /v1/messages mixed-scheduled Antigravity accounts through Claude Forward instead of ForwardGemini - Add async SSE keepalive heartbeats and idle timeout tracking across Gemini native, Messages, and Chat Completions streams - Defer SSE header/initial event commits to enable pre-output failover on empty streams, timeouts, and empty abnormal finish reasons - Abort truncated HTTP/1.1 chunked and HTTP/2 SSE streams after inner middleware completion so clients detect mid-stream drops - Enable HTTPUpstreamProfileLongStream on Gemini/Antigravity upstream requests and map content-filter finish reasons to refusal/content_filter
554cc17 to
38ea8aa
Compare
|
我在
对照中,追加 本地验证与反馈由 Codex 完成。 |
- Require a terminal Gemini finishReason (or promptFeedback.blockReason / in-band error) before completing /v1beta, /v1/messages, and /v1/chat/completions streams at EOF - Emit protocol-appropriate SSE error frames instead of fabricating end_turn/message_stop or finish_reason=stop/[DONE] on unfinished streams, and mark /v1/messages committed SSE streams for transport abort on failure - Add terminal-state matrix tests covering STOP completion, clean EOF without finishReason, and io.ErrUnexpectedEOF across Gemini and Antigravity entrypoints
|
感谢 @r266-tech 的细致验证与反馈!已在 c2fb138 中补齐“HTTP body 正常 修补改动
验证情况
|
|
已在 另用只含一个
旧 可否统一这几处的终态判定,让空的反馈和默认枚举保留首包前的 failover,同时让明确的拒绝本身可以完成 Chat 流?本地复验和反馈由 Codex 完成。 |
…andling - Treat both BLOCK_REASON_UNSPECIFIED and BLOCKED_REASON_UNSPECIFIED as non-blocking defaults across Gemini signal detection, Messages/Chat compat, and Antigravity transformers - Keep empty or unspecified promptFeedback frames from committing message_start so pre-output failover remains intact - Recognize refusal stop_reason deltas as meaningful output in Antigravity OpenAI compat streams so explicit prompt safety blocks complete with content_filter
… and candidate selection - Mark empty abnormal finishReason failovers RequestScopedTransient to avoid unscheduling healthy accounts, and return deterministic 4xx error envelopes directly - Preserve promptFeedback safety refusals in Antigravity non-streaming Claude collector - Emit response.failed terminal events and wire truncated stream abort on /v1/responses - Buffer and replay pre-candidate metadata frames on Antigravity /v1beta streams and align primary candidate selection on index==0
|
在
定向 |
|
补充说明最近两笔提交(28835b391、3984a9b3e)的新增改动,目前所有已知的流式终态、安全拦截与故障转移边界问题均已修复完成: 新增改动
验证状态
|
问题
通过基于 Vercel AI SDK 的客户端(如 Kilo / OpenCode 等,使用
@ai-sdk/google、@ai-sdk/anthropic、@ai-sdk/openai)调用 Gemini 模型时,频繁触发"Response ended unexpectedly and may be incomplete."(底层因流结束时缺少合法终止事件或收到未映射的finishReason被归为'other'/'unknown')。排查定位到以下服务端根因:/v1/messages混合调度协议错位:gateway_handler.go在 Gemini 分组混合调度命中开启mixed_scheduling的 Antigravity 账号时,误调用了原生ForwardGemini而非 Claude 协议入口Forward,把 Claude JSON 请求当作 Gemini 请求发送,并将原生 Gemini SSE 直接写回给期待 Anthropic SSE 的客户端。GeminiMessagesCompatService全流式路径缺失心跳与空闲超时:/v1beta原生流、/v1/messages与/v1/chat/completions兼容流均在主协程同步阻塞读行,未接入StreamKeepaliveInterval与StreamDataIntervalTimeout;长 CoT 思考期间易被前置代理/隧道(如 Cloudflare/Nginx)或客户端静默掐断。200 OK或写出message_start/ 首个 Chat role chunk,导致上游返回空 200 流、带内错误信封、首包前断流/超时、或无正文的异常终止(OTHER/MALFORMED_FUNCTION_CALL/MALFORMED_RESPONSE/UNEXPECTED_TOOL_CALL)时无法触发重试或 Failover。/v1beta流中途读取报错返回后,Gin 正常写出 HTTP/1.1 chunked 结束块0\r\n\r\n;且@ai-sdk/google的chunkSchema所有字段均为.nullish(),会将流内{"error":...}剥离为空对象{}忽略,导致客户端将截断流误判为正常 EOF(finishReason: 'other')而不触发传输层重试。HTTPUpstreamProfileLongStream(HTTP/2ReadIdleTimeoutPING 探测)未注入 Gemini / Antigravity 请求;兼容转换层未将 Gemini 内容安全拦截枚举映射为 Claude"refusal"/ OpenAI"content_filter"。改动
/v1/messages混合调度入口:Gemini 分组选中非 APIKey 的 Antigravity 账号时改为调用antigravityGatewayService.Forward。geminiSSEAsyncReader,为GeminiMessagesCompatService三条流式路径接入StreamKeepaliveInterval(/v1beta下对go-genai等客户端自动跳过注释心跳)与基于末次读取时间精确重置的StreamDataIntervalTimeout(同步更新 Antigravity 流式超时计时器)。200 OK与message_start/ role 首帧;首包前遇到空流、带内错误信封、连接错误、空闲超时或无正文的异常finishReason时返回UpstreamFailoverError(仅写出过心跳且未写出上游数据时标记SafeToFailoverAfterWrite: true);若 Keepalive 提交 SSE 头后下一账号返回非重试错误,按已提交的 SSE 协议写出标准错误事件而非裸 JSON。middleware.IsCommittedEventStream与MarkAbortTruncatedStream,仅对已建立为200 text/event-stream且中途异常中断的流标记中止,在Recovery()最外层待OpsErrorLoggerMiddleware与访问日志完成后触发http.ErrAbortHandler(HTTP/1.1 不写0\r\n\r\n、HTTP/2 发送RST_STREAM),不影响非流式/前置 JSON 错误响应。GeminiMessagesCompatService与AntigravityGatewayService上游请求注入HTTPUpstreamProfileLongStream;统一将全部 Gemini 内容安全拦截枚举映射为 Claude"refusal"及 OpenAI Responses / Chat Completions"content_filter"。验证
go build ./...全量编译通过。go vet -tags=unit与golangci-lint run检查修改涉及的handler、service、antigravity、apicompat、middleware包:0 issues。go test -race -tags unit ./internal/server/middleware/... ./internal/pkg/apicompat/... ./internal/pkg/antigravity/...go test -race -parallel 1 -tags unit ./internal/service -run 'TestGemini.*|TestAntigravity.*|TestHandle.*Streaming.*|TestStream.*' -count=1go test -race -parallel 1 -tags unit ./internal/handler -run 'TestWriteGoogleStreamingErrorAndAbort|TestGemini.*|TestStream.*' -count=1go test -tags unit ./internal/handler -count=1go-genai跳过策略、Keepalive 后首包前超时/空流/无正文异常终止的 Failover 与脏数据保护、已提交 SSE 流的错误帧格式化、HTTPUpstreamProfileLongStream传递、内容安全过滤映射,以及真实httptest.NewServer下截断 chunked 流中止(验证内层中间件正常收尾且普通 4xx JSON 错误不被误中止)。关联 Issue
POST /v1/messages未做 Anthropic→Gemini 协议转换)end_turn/[DONE],客户端丢失工具调用且无报错 #8505 (流式异常结束 /MALFORMED_FUNCTION_CALL等异常finishReason被当作正常end_turn返回;本 PR 处理首包无正文异常终止的自动 Failover、安全过滤refusal/content_filter映射与中途断流中止)@ai-sdk/google等客户端在 Gemini / Antigravity 流首字节前挂起及发送 Keepalive 后无法切号;本 PR 补齐首包前StreamDataIntervalTimeout换号及仅写出心跳时的SafeToFailoverAfterWrite放行)HTTPUpstreamProfileLongStream健康 PING 探测,并支持连接建立后、首包写出前断流的同号重试与 Failover)":\n\n"注释心跳对特定 SDK/CLI 客户端的兼容处理)