feat(boto3): Instrument Bedrock Runtime InvokeModel calls as gen_ai spans - #8187

Description

This PR (2 of 3, stacked on #8186) records InvokeModel as a generate_content {modelId} span with op gen_ai.generate_content (new OP.GEN_AI_GENERATE_CONTENT). It matches the JS SDK.

The request and response bodies have a model-specific format. Like JS, this parses them per model family: Anthropic Claude (including inference profile IDs like us.anthropic...), Amazon Nova and Titan, Meta Llama, Cohere Command and Command R, and Mistral. Prompts and responses use the same gating as Converse: data_collection.gen_ai.inputs/outputs independently, else send_default_pii with include_prompts.

Decisions the diff doesn't show:

• The SDK doesn't read the response body itself or replace the StreamingBody object. Instead, body.read, body.close, and body._raw_stream.close are patched on that instance, the same approach the S3 instrumentation on master uses. _patch_streaming_body() is extracted from it for this. The bytes are recorded as the caller reads them (read(), read(amt), iter_chunks()), and the span finishes when the body is fully read, closed, or a read fails. A new _ServiceExtension.instrument_response() hook lets an extension take over finishing the span.
• Token usage falls back to Bedrock's x-amzn-bedrock-input/output-token-count response headers when the body has none (Cohere, Mistral, unknown models, early close). This replaces JS's chars / 6 estimates.
• The request body is parsed once, in get_span_start_attributes(), for its parameters and input messages, so its parameters aren't visible to ignore_spans and spans it drops skip the parse.
• Request bodies passed as file objects aren't read. Request and response bodies over 1 MB (MAX_INVOKE_MODEL_BODY_BYTES) aren't parsed, e.g. requests with images or image generation results; response usage still comes from the headers. Malformed JSON bodies are skipped quietly.
• Media blocks (Claude image and document blocks, including images in tool_result, and Nova images) aren't recorded.
• Unknown model families still get the span, op, model, and header usage.

Changelog note

InvokeModel spans change from http.client / auto.http.boto3 to gen_ai.generate_content / auto.ai.boto3, and their name changes from aws.bedrock-runtime.InvokeModel to generate_content <modelId>, which affects ignore_spans rules that match on the name. Boto3Integration is auto-enabled, so users with send_default_pii=True start sending InvokeModel prompts and responses unless they set Boto3Integration(include_prompts=False) or configure data_collection.

Issues

• Refs AWS SDK Agent Support #5470. Follow-ups, both matching JS today: amazon.titan-embed models match the Titan text family, so they get a generate_content span rather than an embeddings span, and newer Mistral chat formats (messages/choices) aren't parsed.

via @dcramer.

--

View Session [Sentry]