feat(server-utils): Record Bedrock prompts, tools, and responses - #26118

Why: Bedrock gen_ai spans carried only the model, params, and usage, while the other AI integrations also record prompts and responses. This PR records gen_ai.input.messages, gen_ai.system_instructions, gen_ai.tool.definitions, and gen_ai.output.messages for Converse, ConverseStream, InvokeModel, and InvokeModelWithResponseStream. It also sets the deprecated gen_ai.response.text and gen_ai.response.tool_calls, as endStreamSpan does.

Capture follows dataCollection.genAI, through resolveAIRecordingOptions like the other providers. Inputs and outputs are recorded by default (dataCollection.genAI defaults to on), so Bedrock prompts and responses become recorded by default once this ships. This needs an Important Changes note in the release. dataCollection: { genAI: { inputs: false } } turns off prompts, system instructions, and tool definitions, and outputs: false turns off responses. awsIntegration({ recordInputs, recordOutputs }) takes precedence over both and applies only to Bedrock spans. The setting is read per call, so it does not matter whether the client exists when the integration is built. Nothing else in dataCollection applies here. No other AI provider reads userInfo, and Bedrock records no user fields.

Converse and Nova messages are mapped to {role, parts}. Claude InvokeModel messages are recorded as sent, as the Anthropic integration does, except that image and document blocks (also inside tool_result) are replaced with blob parts so their base64 data is never recorded. Titan, Llama, Mistral, and Cohere prompts become a single user text part.

Bugs found along the way and fixed here:

• InvokeModel with a non-JSON body or no modelId threw and dropped the span.
• Prompt-cache read/write tokens were missing from input tokens.
• Claude stream output tokens were read from the message_start placeholder.
• Mistral and Cohere stream estimates used only the last chunk.
• A failed stream left the span status ok, and a missing stream body left the span open.

The stream tests hand-encode AWS eventstream frames, so nock can mock streaming after all.

How to review: Start with bedrock-runtime-content.ts, which has the mapping, then read the response hooks in bedrock-runtime.ts. scenario-stream.mjs is mostly frame-encoding boilerplate. The integration tests run the plain and streaming scenarios against the default, inputs: false, outputs: false, and an integration option overriding dataCollection in each direction.

Unsure about:

• The options are top-level recordInputs/recordOutputs on awsIntegration. They only affect Bedrock, but nesting them under bedrock felt heavier.
• gen_ai.response.finish_reasons is now a JSON string ('["end_turn"]'), matching the conventions type and the other integrations. It used to be an array attribute.
• Images, documents, and video are recorded as blob parts with a modality (and MIME type when known) and no bytes. Cache points and unknown Converse blocks are dropped. Streams only buffer response content when outputs are recorded.

via @sergical.

--

View Session [Sentry]