[opentelemetry-instrumentation-genai-bedrock] Instrument invoke_inline_agent method

[DylanRussell]

Description

Monkey patch invoke_inline_agent methods so they generate an invoke remote agent span (RemoteAgentInvocation).

This API lets you create an agent on the fly which exists / retains conversation history for a specified number of seconds.

Type of change

Please delete options that are not relevant.

• New feature (non-breaking change which adds functionality)

How has this been tested?

Unit tests

Checklist

• Followed the style guidelines of this project
• Changelog updated if the change requires an entry
• Unit tests added
• Documentation updated

[opentelemetry-pr-dashboard]

Pull request dashboard status

Waiting on reviewers · refreshed 2026-09-19 03:22 UTC

Review the latest changes.

Status above doesn't look right?

• Just replied or pushed? Anything around or after the refresh time above may not be picked up yet — give it a few minutes.
• Anything look wrong? Report it with what you expected; it helps us improve the dashboard.

[Copilot]

🔵 Needs a closer look

Unresolved correctness, stream-lifecycle, compatibility, and scope or disclosure findings remain.

Pull request overview

This PR expands Bedrock GenAI instrumentation to async clients, embeddings, agents, retrieval/RAG APIs, and stream lifecycle handling.

Changes:

• Adds sync/async patching, extractors, and stream wrappers.
• Adds unit, conformance, cassette, and integration coverage.
• Updates dependencies, documentation, shared utilities, and changelogs.

File summaries

File	Reviewed change
util/opentelemetry-util-genai/tests/test_stream.py	Tests abandoned-stream finalization.
util/opentelemetry-util-genai/src/opentelemetry/util/genai/stream.py	Adds abandoned-stream finalization; the finalizer must also swallow BaseException cancellation errors.
util/opentelemetry-util-genai/.changelog/723.fixed	Documents the stream finalization fix.
instrumentation/opentelemetry-instrumentation-genai-llama-index/tests/requirements.oldest.txt	Adds a workflow test dependency; unrelated scope concern remains.
instrumentation/opentelemetry-instrumentation-genai-llama-index/tests/requirements.latest.txt	Adds a workflow test dependency; unrelated scope concern remains.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_invoke_model.py	Tests embedding instrumentation.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_conformance.py	Registers new conformance scenarios.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_async.py	Tests async operations; compatibility concerns remain for the oldest supported environment.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/test_agent.py	Tests agent and retrieval APIs.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/requirements.oldest.txt	Adds the oldest async test dependency.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/requirements.latest.txt	Adds the latest async test dependency.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/conformance/retrieve.py	Adds retrieval conformance coverage.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/conformance/retrieve_and_generate.py	Adds RAG conformance coverage.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/conformance/retrieve_and_generate_streaming.py	Adds streaming RAG conformance coverage.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/conformance/invoke_agent.py	Adds agent conformance coverage.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/conformance/embedding.py	Adds embedding conformance coverage.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/cassettes/test_retrieve_conformance.yaml	Adds a retrieval cassette; its synthetic origin requires disclosure, tracking, or re-recording.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/cassettes/test_retrieve_and_generate_stream_conformance.yaml	Adds a streaming RAG cassette.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/cassettes/test_retrieve_and_generate_conformance.yaml	Adds a RAG cassette.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/cassettes/test_invoke_model_titan_embedding_conformance.yaml	Adds an embedding cassette.
instrumentation/opentelemetry-instrumentation-genai-bedrock/tests/cassettes/test_invoke_agent_conformance.yaml	Adds an agent cassette.
instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/stream.py	Adds Bedrock stream wrappers; unresolved lifecycle, buffering, close-error, abandonment, and UTF-8 handling findings remain.
instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/patch.py	Patches sync/async APIs; unresolved embedding-body, error-telemetry, alias/version, and scope findings remain.
instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py	Adds telemetry extraction; alias/version mapping, retrieval typing, and function-schema normalization require correction.
instrumentation/opentelemetry-instrumentation-genai-bedrock/README.rst	Documents expanded support; the documented scope does not match the PR description.
instrumentation/opentelemetry-instrumentation-genai-bedrock/pyproject.toml	Updates dependency configuration.
instrumentation/opentelemetry-instrumentation-genai-bedrock/.changelog/736.added	Documents inline-agent support.
instrumentation/opentelemetry-instrumentation-genai-bedrock/.changelog/733.added	Documents RAG support.
instrumentation/opentelemetry-instrumentation-genai-bedrock/.changelog/723.added	Documents agent and retrieval support.
instrumentation/opentelemetry-instrumentation-genai-bedrock/.changelog/722.added	Documents embedding support.
instrumentation/opentelemetry-instrumentation-genai-bedrock/.changelog/721.added	Documents async support.

Review details

Suppressed comments (8)

instrumentation/opentelemetry-instrumentation-genai-bedrock/README.rst:19

• The PR description is limited to instrumenting invoke_inline_agent, but these documented changes also add embeddings, invoke_agent, retrieval/RAG operations, async client support, abandoned-stream behavior, and unrelated llama-index requirements. Please split the unrelated work or update the PR scope/description before merging.

* Synchronous and asynchronous embeddings via the InvokeModel API (``client.invoke_model`` with embedding models such as Amazon Titan and Cohere)
* Synchronous and asynchronous remote agent invocation via the Bedrock Agent Runtime API (``client.invoke_agent``)
* Synchronous and asynchronous remote inline agent invocation via the Bedrock Agent Runtime API (``client.invoke_inline_agent``)
* Synchronous and asynchronous Knowledge Base retrieval via the Bedrock Agent Runtime API (``client.retrieve``)
* Synchronous and asynchronous Knowledge Base RetrieveAndGenerate via the Bedrock Agent Runtime API (``client.retrieve_and_generate``)
* Synchronous and asynchronous streaming Knowledge Base RetrieveAndGenerateStream via the Bedrock Agent Runtime API (``client.retrieve_and_generate_stream``)

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py:951

• This repeats the incorrect mapping of Bedrock's alias ID to gen_ai.agent.version. Since invoke_agent does not expose the version selected by the alias, recording the alias here mislabels the semantic attribute; omit it unless the actual resolved version is available.

    # version the alias resolved to. The alias is the closest available identifier of
    # which agent revision served the request.
    agent_alias_id = api_params.get("agentAliasId")
    if agent_alias_id:
        invocation.agent_version = str(agent_alias_id)

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py:1243

• The retrieval extractor constructs the structured retrieval-document attribute as list[dict[str, Any]] and stores it directly on the invocation. Structured GenAI attributes must use a util-genai model; add the shared retrieval-document type in util-genai first and populate that type here so the schema cannot drift.

    docs: list[dict[str, Any]] = []
    for item in results:
        if not _is_dict(item):
            continue
        doc: dict[str, Any] = {}

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/extractors.py:1012

• Bedrock's functionSchema.functions[].parameters is a name-to-ParameterDetail map, whereas FunctionToolDefinition.parameters represents the JSON Schema object used by gen_ai.tool.definitions. Passing this map through emits an invalid shape such as { "city": { "type": "string" } } instead of an object schema with properties; normalize the map, including descriptions and required fields, before storing it.

                        tool_defs.append(
                            FunctionToolDefinition(
                                name=name,
                                description=desc,
                                parameters=params if _is_dict(params) else {},

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/patch.py:406

• agentAliasId is an alias identifier, not the resolved agent version; the comment explicitly says the API does not return that version. Assigning it to agent_version emits a misleading gen_ai.agent.version attribute. Leave the version unset (or use a provider-specific alias attribute), and remove the duplicate assignment in extract_invoke_agent_request.

    # InvokeAgent addresses an agent by alias rather than by version, and an alias
    # resolves to a version server-side.
    raw_alias_id = api_params.get("agentAliasId")
    if raw_alias_id:
        invocation.agent_version = str(raw_alias_id)

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/patch.py:207

• Reading the StreamingBody here eagerly consumes the embedding response before the caller receives it. That adds I/O and latency and changes when body-read failures occur, violating the SDK's lazy return contract; use a lazy body wrapper and finalize when the caller consumes or closes it instead.

        if hasattr(body_stream, "read"):
            raw_bytes = body_stream.read()
            response["body"] = StreamingBody(
                io.BytesIO(raw_bytes), len(raw_bytes)
            )

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/patch.py:731

• The PR title and description scope the change to invoke_inline_agent, but this branch also adds embeddings, async clients, invoke_agent, retrieval, retrieve-and-generate, and a global stream finalizer. These are separate user-visible features with separate changelog fragments; split the work or update the PR scope and review description before merging.

        if service_name not in (BEDROCK_RUNTIME, BEDROCK_AGENT_RUNTIME):
            return wrapped(*args, **kwargs)

        operation_name = args[0] if args else kwargs.get("operation_name")
        raw_params = args[1] if len(args) > 1 else kwargs.get("api_params")

instrumentation/opentelemetry-instrumentation-genai-bedrock/src/opentelemetry/instrumentation/genai/bedrock/patch.py:215

• A failure while reading the synchronous embedding response body occurs after the only try block has ended, so the exception is re-raised without calling invocation.fail; the embedding span can remain unfinished and no error telemetry is recorded. Include body reading and extraction in the telemetry error boundary and re-raise the original exception.

        raw_bytes = b""
        body_stream = response.get("body")
        if hasattr(body_stream, "read"):
            raw_bytes = body_stream.read()
            response["body"] = StreamingBody(
                io.BytesIO(raw_bytes), len(raw_bytes)
            )

        extract_embedding_response(
            response,
            raw_bytes,
            invocation,
        )
        invocation.stop()
        return response

• Files reviewed: 5/5 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.