[bot] Google GenAI video generation (generateVideos / Veo) traced with materially less detail than text generation

从 BraintrustApiClient.java 开始，重点查看 tagSpan 和四个传输重写，然后检查 BraintrustGenAITest.java 以及 Google GenAI 的 generateVideos 和 getVideosOperation API。为视频生成添加有针对性的覆盖，并验证请求详细信息、span 分类、可用指标以及长时间运行操作的轮询是否以预期的保真度得到体现。

Summary

The Google GenAI Java SDK (com.google.genai, v1.18.0, the version this module targets) exposes Models.generateVideos(...) (Veo video-generation models) and the associated long-running-operation polling API (client.operations.getVideosOperation(...)). Braintrust's Java GenAI instrumentation intercepts these calls only incidentally, through a generic HTTP-transport hook that was designed for generateContent. The result is spans that are silently mislabeled and largely empty for video generation — not a total absence of spans, but instrumentation with materially less detail/fidelity than the documented generate_content path.

What is missing

BraintrustApiClient (braintrust-sdk/instrumentation/genai_1_18_0/src/main/java/com/google/genai/BraintrustApiClient.java) intercepts every HTTP call made by the SDK by overriding the transport-level methods on ApiClient:

• request(String, String, String, Optional<HttpOptions>) — line 186
• request(String, String, byte[], Optional<HttpOptions>) — line 210
• asyncRequest(String, String, String, Optional<HttpOptions>) — line 239
• asyncRequest(String, String, byte[], Optional<HttpOptions>) — line 277

Because this hook is transport-level, a generateVideos call (and each operations.getVideosOperation poll of the resulting long-running operation) does produce a span — but tagSpan (lines 50–161) was written exclusively for the generateContent request/response shape:

• Line 156 unconditionally sets braintrust.span_attributes to {"type": "llm"} for every call this client makes, including video-generation calls, which are not LLM chat-completion calls.
• Lines 65–76 and 81–95 only promote model/systemInstruction/tools/toolConfig/safetySettings/cachedContent and generationConfig.{temperature,topP,topK,candidateCount,maxOutputTokens,stopSequences,responseMimeType,responseSchema} into braintrust.metadata/input_json. generateVideos requests use an entirely different shape (instances/parameters or prompt/image/video + a GenerateVideosConfig with fields like numberOfVideos, durationSeconds, aspectRatio, enhancePrompt, personGeneration) — none of these fields are extracted, so input_json for a video-generation span is effectively empty.
• Lines 127–149 only look for usageMetadata.{promptTokenCount,candidatesTokenCount,totalTokenCount,cachedContentTokenCount} to populate braintrust.metrics. Video-generation responses have no such field, so metrics are silently skipped (not wrong, just absent) while the span is still tagged type: llm.
• generateVideos returns a GenerateVideosOperation that is not done immediately; the caller polls client.operations.getVideosOperation(...) repeatedly until completion. Each poll produces its own disconnected span (named via URL-derived getOperation, e.g. get_videos_operation) rather than being correlated to the originating video-generation request — there is no long-running-operation span model, unlike the single request/response span Braintrust produces for generate_content.

There is no reference anywhere in this module (or its tests) to generateVideos, GenerateVideosOperation, GenerateVideosConfig, Veo, or getVideosOperation — confirmed via full-directory grep of braintrust-sdk/instrumentation/genai_1_18_0/.

Braintrust docs status

Checked https://www.braintrust.dev/docs/integrations/ai-providers/gemini (current as of this audit):

• Java section documents only content-generation spans (generate_content, sync/async) and generateContentStream, with request metadata (model, contents, generation config), response, and token metrics.
• No mention of video generation, generateVideos, or Veo models for any language (Java, Python, TypeScript, Go).
• Status: not_found — video generation tracing is undocumented for Java (and for every other language) in the current docs.

Upstream sources

• Google GenAI Java SDK repo: https://github.com/googleapis/java-genai — Models class exposes generateVideos(String model, GenerateVideosSource source, GenerateVideosConfig config) and generateVideos(String model, String prompt, Image image, Video video, GenerateVideosConfig config).
• SDK README usage example (video generation via Veo, polling operation.done() / client.operations.getVideosOperation): https://github.com/googleapis/java-genai/blob/main/README.md
• SDK CHANGELOG entries for video generation support (e.g. "Add backwards compatibility for generateVideos in Java SDK", "Add support for Veo frame interpolation and video extension"): https://github.com/googleapis/java-genai/blob/main/CHANGELOG.md
• Gemini API Veo docs (request/response/config shape — GenerateVideosConfig fields like numberOfVideos, durationSeconds, aspectRatio): https://ai.google.dev/gemini-api/docs/veo
• Java Veo 3 walkthrough (async polling pattern): https://glaforge.dev/posts/2025/09/10/generating-videos-in-java-with-veo3/

Local files inspected

• braintrust-sdk/instrumentation/genai_1_18_0/src/main/java/com/google/genai/BraintrustApiClient.java — full file read; tagSpan (lines 50–161) and the four transport overrides (lines 186, 210, 239, 277) are the only interception points, and are generateContent-shaped.
• braintrust-sdk/instrumentation/genai_1_18_0/src/test/java/dev/braintrust/instrumentation/genai/v1_18_0/BraintrustGenAITest.java — only tests sync generateContent; no video-generation test coverage.
• Full-directory grep of braintrust-sdk/instrumentation/genai_1_18_0/ for generateImages|GenerateImages|predictLongRunning|generate_videos|video — zero matches.