[bot] Google GenAI Batch API (`ai.batches.create()`) not instrumented

[braintrust-bot]

Summary

The @google/genai SDK's Batch API (ai.batches.create() / client.batches.*) is a documented, GA generative-execution surface for asynchronous, high-throughput generateContent (and embeddings) processing at a 50% cost discount. It is not instrumented anywhere in this repository — no wrapper support, no channel definitions, no plugin, no auto-instrumentation config entry, and no e2e coverage.

What instrumentation is missing

Channels (js/src/instrumentation/plugins/google-genai-channels.ts)

googleGenAIChannels (lines 20–57) defines only four channels: generateContent, generateContentStream, embedContent, and interactionsCreate. There is no batches-related channel.

Wrapper (js/src/wrappers/google-genai.ts)

wrapModels() (lines 118–133) proxies only generateContent, generateContentStream, and embedContent on instance.models. The instance.batches namespace (ai.batches.create(), .get(), .list(), .cancel()) is not proxied at all and passes through completely untraced.

Plugin (js/src/instrumentation/plugins/google-genai-plugin.ts)

Only subscribes to the four existing channels above — no batch handler exists.

Vendored types (js/src/vendor-sdk-types/google-genai.ts)

No type exists for the batches namespace or its request/response/job-status shapes.

Auto-instrumentation config (js/src/auto-instrumentations/configs/google-genai.ts)

No batch-related entry.

e2e coverage (e2e/scenarios/google-genai-instrumentation/)

No batch-related scenario, cassette, or span-tree snapshot among the existing google-genai-v1/v2 (pinned + latest) scenarios.

Upstream operations not instrumented

• ai.batches.create({ model, src, config }) — submit a batch job (file-based or inline requests) for generateContent
• ai.batches.create_embeddings(...) — submit a batch job for embeddings
• ai.batches.get(name) — poll job status (JOB_STATE_SUCCEEDED, etc.)
• ai.batches.list() — list batch jobs
• ai.batches.cancel(name) — cancel a batch job

Braintrust docs status

not_found. The Braintrust Gemini integration guide (https://www.braintrust.dev/docs/integrations/ai-providers/gemini) documents generateContent/generateContentStream, embedContent, grounding metadata, generate_images (Python only), and interactions/Live API (Python only) — but does not mention batch processing anywhere.

Upstream references

• Gemini API Batch Mode guide: https://ai.google.dev/gemini-api/docs/batch-mode
• Gemini API Batch API reference: https://ai.google.dev/gemini-api/docs/batch-api
• @google/genai npm package: https://www.npmjs.com/package/@google/genai
• SDK source: https://github.com/googleapis/js-genai

Braintrust docs sources checked

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

Local files inspected

• js/src/instrumentation/plugins/google-genai-channels.ts (full file — only 4 channels defined, no batches)
• js/src/instrumentation/plugins/google-genai-plugin.ts (subscribes only to the 4 existing channels)
• js/src/wrappers/google-genai.ts (full file — wrapModels() proxies only generateContent/generateContentStream/embedContent; instance.batches untouched)
• js/src/vendor-sdk-types/google-genai.ts (no batches type)
• js/src/auto-instrumentations/configs/google-genai.ts (no batches entry)
• e2e/scenarios/google-genai-instrumentation/ (full directory listing — no batch scenario/cassette/snapshot)
• Repo-wide case-insensitive grep for batch under js/src — zero matches related to Google GenAI