[bot] AWS Bedrock Runtime: `start_async_invoke()` not instrumented

[braintrust-bot]

Summary

The AWS Bedrock Runtime start_async_invoke() API is not instrumented. Calls to client.start_async_invoke(...) produce zero Braintrust tracing. This method submits an asynchronous model invocation job — output is written to an S3 bucket — and is the Bedrock analog to batch inference APIs that are already instrumented for Anthropic (messages.batches.create()) and tracked as gaps for OpenAI (#295), Mistral (#272), and Google GenAI (#332).

The Bedrock Runtime integration was completed in #279 (closed 2026-06-16) and instruments the four synchronous execution surfaces. start_async_invoke was not included.

What is missing

Bedrock Runtime method	Description	Instrumented?
client.converse(...)	Synchronous unified chat	Yes
client.converse_stream(...)	Streaming unified chat	Yes
client.invoke_model(...)	Synchronous raw model invocation	Yes
client.invoke_model_with_response_stream(...)	Streaming raw invocation	Yes
client.start_async_invoke(...)	Asynchronous model invocation job	No

API details

• Endpoint: POST /async-invoke
• Purpose: Submit a long-running async model invocation. The caller supplies modelId, modelInput (provider-specific JSON), and outputDataConfig (S3 output location). The API returns an invocationArn immediately; output is written to S3 when complete.
• Poll for results: Status checked via client.get_async_invoke(invocationArn=...) or client.list_async_invokes().
• Use case: Large-scale batch generation workloads where latency is not time-sensitive (similar to Anthropic's Message Batches API or OpenAI's Batch API).

Minimum instrumentation for start_async_invoke

At minimum, start_async_invoke should create a span capturing:

Span field	Content
input	modelInput (parsed JSON body), modelId
output	invocationArn, submission HTTP status
metadata	provider: "bedrock", model (from modelId), S3 output location, clientRequestToken if set
metrics	Submission latency

Span type should be SpanTypeAttribute.TASK, matching the Anthropic batch pattern.

Availability

start_async_invoke is in both boto3 matrix pins:

• boto3==1.43.29 (latest)
• boto3==1.34.116 (floor)

It is in the bedrock-runtime service (not bedrock-agent-runtime), so it is reachable via the same wrap_bedrock_client(client) helper that wraps the existing four methods.

Braintrust docs status

not_found — The AWS Bedrock integration page documents tracing for converse and invoke_model. No mention of async invocation.

Upstream sources

• AWS Bedrock Runtime API reference: StartAsyncInvoke — https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_StartAsyncInvoke.html
• boto3 Bedrock Runtime client docs: start_async_invoke method in the bedrock-runtime service
• The method is in the GA (non-preview) service, available in both matrix pins

Local files inspected

• py/src/braintrust/integrations/bedrock_runtime/patchers.py — defines ConversePatcher, ConverseStreamPatcher, InvokeModelPatcher, InvokeModelWithResponseStreamPatcher, composed into BedrockRuntimeClientMethodsPatcher; zero references to start_async_invoke
• py/src/braintrust/integrations/bedrock_runtime/tracing.py — _converse_wrapper, _converse_stream_wrapper, _invoke_model_wrapper, _invoke_model_stream_wrapper only; no async invoke wrappers
• py/src/braintrust/integrations/bedrock_runtime/integration.py — registers BedrockRuntimeClientMethodsPatcher; no async invoke registration
• py/src/braintrust/integrations/anthropic/tracing.py — Anthropic messages.batches.create() IS instrumented as a task-type span — reference implementation for this pattern
• py/pyproject.toml — [tool.braintrust.matrix.boto3] pins: latest = "boto3==1.43.29", "1.34.116" = "boto3==1.34.116"

Relationship to existing issues

• Add AWS Bedrock Runtime (boto3) integration for Converse and InvokeModel instrumentation #279 (closed) completed the Bedrock Runtime integration for converse, converse_stream, invoke_model, invoke_model_with_response_stream; start_async_invoke was not in scope
• [bot] OpenAI: Batch API (client.batches.create()) not instrumented #295 tracks the analogous OpenAI Batch API gap (client.batches.create())
• [BOT ISSUE] Mistral: Batch Jobs API (client.batch.jobs.create()) not instrumented #272 tracks the analogous Mistral Batch Jobs API gap (client.batch.jobs.create())
• [bot] Google GenAI: instrument Batch API (client.batches.create / create_embeddings) #332 tracks the analogous Google GenAI Batch API gap (client.batches.create())