[bot] Google GenAI generateImages (Imagen) not instrumented in Go SDK

[braintrust-bot]

What instrumentation is missing

The Google GenAI Go SDK (google.golang.org/genai v1.41.0) exposes an image generation API via client.Models.GenerateImages() (Imagen models). The genai HTTP router in trace/contrib/genai/tracegenai.go only matches four path patterns:

func containsGenerateContent(path string) bool {
    return strings.Contains(path, ":generateContent") ||
        strings.Contains(path, "/generateContent") ||
        strings.Contains(path, ":streamGenerateContent") ||
        strings.Contains(path, "/streamGenerateContent")
}

func containsEmbedContent(path string) bool {
    return strings.Contains(path, ":embedContent") ||
        strings.Contains(path, "/embedContent") ||
        strings.Contains(path, ":batchEmbedContents") ||
        strings.Contains(path, "/batchEmbedContents")
}

The Imagen API uses the :predict method suffix:

POST https://generativelanguage.googleapis.com/v1beta/models/imagen-4.0-generate-001:predict

This path contains neither :generateContent nor :embedContent variants, so genaiRouter returns nil and all GenerateImages calls are untraced.

Go SDK usage

result, err := client.Models.GenerateImages(ctx,
    "imagen-4.0-generate-001",
    "A photorealistic image of a cat",
    &genai.GenerateImagesConfig{NumberOfImages: 1},
)

What could be traced

A dedicated generateImagesTracer (similar to the existing generateContentTracer and embedContentTracer) matching :predict paths on Imagen model IDs could capture:

• braintrust.input_json: {"model": "imagen-4.0-generate-001", "prompt": "...", "config": {...}}
• braintrust.output_json: image output summary — e.g. {"images_count": 1, "mime_type": "image/png"} (raw pixel bytes omitted, matching the embeddingsOutputSummary convention of not logging raw vectors)
• braintrust.metadata: model, numberOfImages, aspectRatio, safetyFilterLevel, outputOptions

Note: the :predict path is also shared by Imagen's EditImage and UpscaleImage methods; the model ID in the path (e.g. imagen-4.0-generate-001 vs imagegeneration@006) distinguishes them.

Braintrust docs status

supported (Python only) — the Braintrust Gemini integration docs include this in the operations table:

"generate_images and async generate_images | Image generation calls, with generated image output captured in the span output. Python only."

This confirms generate_images tracing is an intended Braintrust capability for the GenAI SDK; Go is currently the only unsupported runtime.

Source: https://www.braintrust.dev/docs/integrations/ai-providers/gemini

Upstream sources

• Gemini API — Imagen overview: https://ai.google.dev/gemini-api/docs/imagen
• Gemini API — Image generation guide: https://ai.google.dev/gemini-api/docs/image-generation
• google.golang.org/genai pkg.go.dev — Models.GenerateImages: https://pkg.go.dev/google.golang.org/genai#Models.GenerateImages
• googleapis/go-genai GitHub: https://github.com/googleapis/go-genai

Local repo files inspected

• trace/contrib/genai/tracegenai.go — genaiRouter and path-matching helpers (containsGenerateContent, containsEmbedContent): no :predict case
• trace/contrib/genai/generatecontent.go — reference pattern for request parsing and response instrumentation
• trace/contrib/genai/embedcontent.go — reference pattern for output summary (shape-only, no raw vectors)
• trace/contrib/genai/go.mod — google.golang.org/genai v1.41.0 (Models.GenerateImages available in this version)
• trace/contrib/genai/tracegenai_test.go — no tests for :predict / Imagen paths
• examples/internal/genai/main.go — no GenerateImages example