Workers AI integration drops output of non-streaming chat responses in the OpenAI format (`choices`)

未关闭 适合新手
  • 哪里不对？
#26,120 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
78/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
typescript
领域
observability

调研方向

首先阅读 packages/server-utils/src/ai/workers-ai/utils.ts 及其 addResponseAttributes 函数，然后比较 ai/workers-ai/streaming.ts 中现有的 Workers AI 流处理与 ai/openai/utils.ts 中的 OpenAI 处理。更新非流式路径，从 choices 中捕获输出消息、响应文本和结束原因；确认新格式的响应会记录这些属性，同时保留旧的响应格式。

由索引模型根据 Issue 内容生成。

描述

Bug Cloudflare Workers Spans
Is there an existing issue for this?
  • I have checked for existing issues
How do you use Sentry?

Sentry SaaS (sentry.io)

Which SDK are you using?

@sentry/cloudflare

SDK Version

11.10.0 (@sentry/server-utils 11.10.0). The same code is on develop.

Framework Version

Cloudflare Workers, Workers AI binding

Steps to Reproduce

  1. Instrument the Workers AI binding (the default @sentry/cloudflare setup).

  2. Call a model that answers in the OpenAI chat-completion format, without stream: true. For example, @cf/google/gemma-4-26b-a4b-it:

    const result = await env.AI.run("@cf/google/gemma-4-26b-a4b-it", {
  messages: [{ role: "user", content: "Say hi" }],
  max_tokens: 50,
});

  3. Open the gen_ai.chat span in Sentry.

Expected Result

The span has gen_ai.output.messages, gen_ai.response.text and gen_ai.response.finish_reasons, the same as for a stream: true call to the same model.

Actual Result

The span has the token usage, but no output and no finish reason.

I sent the same prompt to @cf/google/gemma-4-26b-a4b-it twice through the Workers AI REST API, once without and once with stream: true. Then I passed the real responses through instrumentWorkersAiClient (11.10.0). The gen_ai.* span attributes were:

--- no stream
gen_ai.operation.name = "chat"
gen_ai.usage.input_tokens = 19
gen_ai.usage.output_tokens = 4
gen_ai.usage.total_tokens = 23

--- stream: true
gen_ai.operation.name = "chat"
gen_ai.output.messages = "[{\"role\":\"assistant\",\"parts\":[{\"type\":\"text\",\"content\":\"Hello there!\"}]}]"
gen_ai.response.finish_reasons = "[\"stop\"]"
gen_ai.response.streaming = true
gen_ai.response.text = "Hello there!"
gen_ai.usage.input_tokens = 19
gen_ai.usage.output_tokens = 4
gen_ai.usage.total_tokens = 23

The non-streaming response was:

{"id":"...","object":"chat.completion","created":0,"model":"@cf/google/gemma-4-26b-a4b-it","choices":[{"index":0,"message":{"role":"assistant","content":"Hello there!"},"finish_reason":"stop","logprobs":null}],"usage":{"prompt_tokens":19,"completion_tokens":4,"total_tokens":23}}
Cause

Workers AI returns two response formats:

  • older models: { response: string, tool_calls?, usage }
  • newer models, such as gemma-4: the OpenAI chat-completion format with choices[].message

The streaming path supports both. instrumentWorkersAiStream in ai/workers-ai/streaming.ts reads parsed.response and also parsed.choices[].delta.content, delta.tool_calls and finish_reason.

The non-streaming path supports only the old format. addResponseAttributes in ai/workers-ai/utils.ts reads only response.response and response.tool_calls:

https://github.com/getsentry/sentry-javascript/blob/develop/packages/server-utils/src/ai/workers-ai/utils.ts

A possible fix: when response.choices is an array, read choices[].message.content, choices[].message.tool_calls and choices[].finish_reason, the same way as the OpenAI integration (ai/openai/utils.ts) does.

Additional Context

Workaround: we wrap the call in our own gen_ai.chat span that sets gen_ai.output.messages from choices, and we call env.AI.run inside Sentry.suppressTracing() so that the SDK span does not duplicate it.

主要语言
TypeScript
星标
8.8k
派生
1.9k
平均合并
1 天 20 小时
30 天内合并 PR
682

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

getsentry/sentry-javascript 的其他 Issue

查看 getsentry/sentry-javascript 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。