feat(boto3): Instrument Bedrock Runtime Converse calls as gen_ai spans - #8186

Description

This is the first of 3 stacked PRs that add AWS Bedrock Runtime gen_ai instrumentation to Boto3Integration, so Bedrock calls appear in AI Agents monitoring. Spans, ops, and attributes match the JS SDK's aws-sdk Bedrock Runtime extension (packages/server-utils/src/integrations/aws-sdk/services/bedrock-runtime.ts).

This PR covers Converse, using the existing _services extension registry:

• span chat {modelId}, op gen_ai.chat, gen_ai.provider.name = aws.bedrock
• request model, max tokens, temperature, top_p, stop sequences (new SPANDATA.GEN_AI_REQUEST_STOP_SEQUENCES), finish reasons, and token usage
• adds a get_span_name() hook to _ServiceExtension

Decisions the diff doesn't show:

• Origin auto.ai.boto3, not the generic auto.http.boto3. It follows the trace-origin spec (auto.<category>.<integration>). JS uses auto.ai.aws_sdk for the same spans. Other bedrock-runtime operations keep the generic HTTP span.
• Prompt/response capture honors the data_collection option. When data_collection is set, data_collection.gen_ai.inputs gates system instructions, input messages, and tool definitions, and data_collection.gen_ai.outputs gates output messages (text, reasoning, tool calls). Each is gated independently. Without data_collection, both require send_default_pii and the new Boto3Integration(include_prompts=True) option, which defaults to True. JS capture is in progress in feat(server-utils): Record Bedrock prompts, tools, and responses sentry-javascript#26118.
• Tool definitions are gated on inputs in both modes. This differs from Anthropic and OpenAI, which always record tool definitions when data_collection isn't set (a legacy branch kept until send_default_pii is deprecated).
• Input messages are truncated like Anthropic and OpenAI. When should_truncate_gen_ai_input() is true (gen_ai spans sent in transactions), truncate_and_annotate_messages reduces the history to the last message. The text, tool call arguments, and tool results of that message's parts are then capped at MAX_SINGLE_MESSAGE_CONTENT_CHARS in total, with the same ... suffix, in the Bedrock code: the shared helper only caps a top-level content, not parts. The original-length _meta annotation doesn't apply to gen_ai.input.messages yet, and teaching the shared helper about parts is a follow-up.
• Only the input messages are added after the span starts, by a new _ServiceExtension.get_span_start_attributes() hook, because truncation needs the started span. ignore_spans doesn't see them, and spans it drops skip that work. Max tokens, temperature, top_p, stop sequences, system instructions, and tool definitions stay request attributes.
• gen_ai.usage.input_tokens includes cache read/write tokens. Bedrock's inputTokens excludes them, so this differs from JS only for prompt-cached requests.
• Media blocks (image, document, video) aren't recorded.
• Like the rest of boto3, this never starts a transaction. Calls without a parent span aren't recorded.

Stack: this PR → InvokeModel → streaming (ConverseStream + InvokeModelWithResponseStream).

Heads-up: major/3.0 has diverged boto3 service-extension work (_services/*, _utils.py), so porting this there will need conflict resolution.

Changelog note

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

Issues

• Refs AWS SDK Agent Support #5470 (PY-2085). bedrock-agent-runtime (InvokeAgent) is a follow-up.

via @dcramer.

--

View Session [Sentry]