Arize AI Openinference Openinference

Python Code Reviewer for OpenInference Instrumentors

Review a Python OpenInference instrumentation package against the project's established patterns and conventions. This is a checklist-driven review — go through each section, report findings with file paths and line numbers, and surface issues organized by severity.

Workflow

Step 1: Identify the package to review

• Ask the user which instrumentor to review if not already clear from context
• The package lives under python/instrumentation/openinference-instrumentation-<name>/
• Read the key files: __init__.py, _wrappers.py (or equivalent), pyproject.toml, and the full tests/ directory

Step 2: Pull the instrumented library source and use it as ground truth

OpenInference instrumentors work by monkey-patching functions in the library they instrument. All correctness judgments — whether wrappers target the right methods, handle the right signatures, process the right data structures, and cover the right edge cases — must be verified against the actual library source code. Do NOT make assumptions about how the instrumented library works.

Note: The tox env name <pkg> and the library's Python import path <library> often differ. For example, google_genai is the tox env name but the library installs as google/genai/ in site-packages. Check test-requirements.txt or pyproject.toml to find the actual library package name.

1. Set up the tox environment to install the pinned library version. Look up the tox envlist in python/tox.ini to find the correct env name (use the highest Python version available, e.g., py314, py313):

   cd python && uvx --with tox-uv tox run -e <pyVER>-ci-<pkg> -- --co -q
   

   (-- --co -q tells pytest to collect without running, which triggers the install.) If the .tox env already exists, skip this step. If tox setup fails (missing Python version, dependency conflicts), fall back to pip install <library> in a temporary venv to unblock the review.

2. Locate the installed library source at:

   python/.tox/<pyVER>-ci-<pkg>/lib/python<X.Y>/site-packages/<library>/
   

3. Reference the library source throughout the review. Before flagging any finding, verify it against the actual code:

   • Are the monkey-patched methods/classes correct? Check they exist and have the expected signatures.
   • Are parameter types handled correctly? Read the real type annotations and defaults.
   • Are edge cases real? Check whether a supposed edge case can actually occur given the library's actual types, validation, and control flow.
   • Are attribute extractions correct? Verify field names, nesting, and optional vs. required fields against the library's actual data classes.

4. Calibrate severity based on what the library actually does:

   • A bug affecting types/paths the library actually uses → High or Critical
   • An edge case for a type that can't actually appear at runtime → Low
   • A missing handler for a type in the library's Union that is common → higher severity
   • A missing handler for a rare/internal type → lower severity

Step 3: Run all review sections below

Step 4: Present findings organized by severity:

• Critical: Will cause incorrect behavior or CI failure
• High: Missing required convention or test coverage gap
• Medium: Deviates from established patterns but functional
• Low: Style or minor improvement suggestions

Section 1: Test Setup and CI Config

The tox.ini install pattern matters because a broken pattern silently installs the wrong version of the library, making the "pinned version" test target useless.

1.1 tox.ini install pattern

Read python/tox.ini and find the commands_pre entries for this package.

Correct pattern (google_adk style — 4 steps, substitute <pkg> with the actual package name):

<pkg>: uv pip uninstall -r test-requirements.txt
<pkg>: uv pip install --reinstall-package openinference-instrumentation-<pkg> .
<pkg>: python -c 'import openinference.instrumentation.<pkg>'
<pkg>: uv pip install -r test-requirements.txt

Broken pattern (causes under-resolution — the pinned version test may silently test the wrong version):

<pkg>: uv pip install --reinstall {toxinidir}/instrumentation/openinference-instrumentation-<pkg>[test]

Flag the broken pattern as Critical — it defeats the purpose of version-pinned testing.

1.2 test-requirements.txt

Check that test-requirements.txt exists in the package root and contains:

• A pinned version of the library being instrumented (e.g., openai==2.8.0)
• opentelemetry-sdk
• pytest-recording (if using VCR cassettes)
• Any other test utilities needed (pytest-asyncio, respx, responses, etc.)

If test-requirements.txt is missing entirely, flag as Critical (the correct tox pattern depends on it).

1.3 Latest test target

Verify that the tox envlist has both pinned and -latest variants:

py3{10,14}-ci-{pkg,pkg-latest}

And that the -latest variant upgrades the library:

pkg-latest: uv pip install -U <library-name>

Section 2: Testing Patterns

2.1 conftest.py fixtures

Read tests/conftest.py and verify these fixtures exist:

Required fixtures:

• in_memory_span_exporter — returns InMemorySpanExporter()
• tracer_provider — creates TracerProvider with SimpleSpanProcessor wired to the exporter
• instrument (autouse) — calls Instrumentor().instrument(tracer_provider=...), clears exporter, yields, then calls .uninstrument() and clears again

Scope considerations:

• Session-scoped instrumentor fixtures are fine when the instrumentor is stateless
• Function-scoped exporter+provider is safer for test isolation but session-scoped works if tests clear the exporter properly

VCR config fixture (if using cassettes):

@pytest.fixture(scope="session")
def vcr_config() -> dict[str, Any]:
    return {
        "before_record_request": _strip_request_headers,
        "before_record_response": _strip_response_headers,
        "decode_compressed_response": True,
        "record_mode": "once",
    }

With helper functions that strip sensitive headers from recorded cassettes.

2.2 pytest-recording / VCR cassettes

If the instrumentor calls external APIs (LLM providers, embedding services, etc.):

• Tests should use @pytest.mark.vcr decorator
• Cassettes should live in tests/cassettes/ (pytest-recording default)
• Cassette YAML files should have headers stripped (no API keys recorded)
• pytest-recording should be in test-requirements.txt

If tests use mocking instead of VCR, that's acceptable but note it as a pattern difference.

2.3 Exhaustive attribute assertions (pop-style)

This is the most important testing pattern. Tests should verify ALL span attributes, not just spot-check a few. The pattern prevents regressions where unexpected attributes appear or expected ones disappear silently.

Correct pattern:

attributes = dict(span.attributes or {})
assert attributes.pop(OPENINFERENCE_SPAN_KIND) == OpenInferenceSpanKindValues.CHAIN.value
assert attributes.pop(INPUT_VALUE)
assert attributes.pop(INPUT_MIME_TYPE) == JSON
assert attributes.pop(OUTPUT_VALUE)
assert attributes.pop(OUTPUT_MIME_TYPE) == JSON
# ... pop all remaining attributes ...
assert not attributes  # Nothing unexpected left

What to flag:

• Tests that only check a few attributes without the final assert not attributes — High
• Tests that use span.attributes[KEY] or span.attributes.get(KEY) instead of pop — Medium (functional but doesn't catch unexpected extras)
• Missing assert not attributes at the end — High

2.4 Context attribute propagation tests

There should be at least one test that uses using_attributes() context manager and verifies that context attributes appear on spans:

with using_attributes(
    session_id="test-session",
    user_id="test-user",
    metadata={"key": "value"},
    tags=["tag-1", "tag-2"],
    prompt_template="template {var}",
    prompt_template_version="v1.0",
    prompt_template_variables={"var": "value"},
):
    # run instrumented code

Then verify these attributes appear on the spans via pop assertions.

Section 3: OpenInference Semantic Conventions

Check which conventions apply based on the type of library being instrumented. Not every instrumentor needs every attribute — match the conventions to what the library actually does.

3.1 Always required

Every span must have:

• OPENINFERENCE_SPAN_KIND — set to the appropriate kind enum value
• INPUT_VALUE + INPUT_MIME_TYPE — what went into the operation
• OUTPUT_VALUE + OUTPUT_MIME_TYPE — what came out

MIME types should be application/json for structured data (dicts, Pydantic models) and text/plain for strings. Flag if MIME type is missing when value is set — High.

When setting input/output attributes, the instrumentor should use:

from openinference.instrumentation import get_input_attributes, get_output_attributes
span.set_attributes(dict(get_input_attributes(val, mime_type=OpenInferenceMimeTypeValues.JSON)))

3.2 LLM libraries (OpenAI, Anthropic, Mistral, LiteLLM, etc.)

These should set:

• LLM_MODEL_NAME — the model identifier
• LLM_PROVIDER — the provider name (e.g., "openai", "anthropic")
• LLM_INVOCATION_PARAMETERS — JSON of parameters like temperature, max_tokens
• LLM_INPUT_MESSAGES — array of message objects with role and content
• LLM_OUTPUT_MESSAGES — array of response message objects
• LLM_TOKEN_COUNT_PROMPT / LLM_TOKEN_COUNT_COMPLETION / LLM_TOKEN_COUNT_TOTAL — token usage

Span kind should be LLM.

3.3 Embedding libraries

• EMBEDDING_MODEL_NAME
• EMBEDDING_EMBEDDINGS — the embedding vectors (unless masked by TraceConfig)
• EMBEDDING_TEXT — input text

Span kind should be EMBEDDING.

3.4 Tool/function calling

When the library supports tool use or function calling:

• TOOL_NAME — name of the tool/function
• TOOL_DESCRIPTION — description (if available)
• TOOL_PARAMETERS — JSON schema of parameters

Span kind should be TOOL.

3.5 Agent/orchestration frameworks (CrewAI, LangChain, DSPy, etc.)

These typically produce multiple span kinds in a hierarchy:

• CHAIN for orchestration/workflow spans
• AGENT for agent execution spans
• TOOL for tool invocations
• LLM for underlying model calls

3.6 Retrieval libraries

• RETRIEVAL_DOCUMENTS — array of retrieved documents
• Document attributes: DOCUMENT_ID, DOCUMENT_CONTENT, DOCUMENT_SCORE, DOCUMENT_METADATA

Span kind should be RETRIEVER.

Section 4: Span Hierarchy

4.1 Parent-child relationships

For instrumentors that create multiple spans, verify:

• Spans nest correctly (child span's parent.span_id matches parent span's context.span_id)
• All spans from a single operation share the same trace_id
• No orphaned root spans that should be children

Tests should verify hierarchy explicitly:

trace_ids = {span.context.trace_id for span in spans}
assert len(trace_ids) == 1  # All in one trace

assert child_span.parent.span_id == parent_span.context.span_id

Flag missing hierarchy tests as High for multi-span instrumentors.

4.2 Correct span kinds in hierarchy

Common correct hierarchies:

• CHAIN -> LLM (simple chain with LLM call)
• CHAIN -> AGENT -> TOOL (agent framework)
• CHAIN -> AGENT -> LLM (agent making LLM calls)
• CHAIN -> RETRIEVER -> EMBEDDING (RAG pipeline)
• CHAIN -> CHAIN -> LLM (nested chains)

4.3 Thread/async context propagation

If the instrumented library uses threads or async:

• Verify that OTel context is properly propagated across thread boundaries (using contextvars.copy_context() if needed)
• For async code, ensure spans created in async functions are properly parented
• Flag if the library is known to use ThreadPoolExecutor or similar and the instrumentor doesn't handle context propagation — Critical

4.4 Suppress tracing support

Every wrapper should check suppression at the start. Either pattern is acceptable:

# Pattern 1: private key (common in this repo)
if context_api.get_value(context_api._SUPPRESS_INSTRUMENTATION_KEY):
    return wrapped(*args, **kwargs)

# Pattern 2: public API
from opentelemetry.context import suppress_instrumentation
if suppress_instrumentation():
    return wrapped(*args, **kwargs)

Missing suppression check — High.

4.5 TraceConfig masking support

The instrumentor should accept and respect TraceConfig:

• Pass it to OITracer or use it to mask attributes before setting them
• At minimum, hide_inputs and hide_outputs should work

Missing TraceConfig support — Medium (functional but incomplete).

Presenting Results

Organize findings into a table:

Then list what's working well — positive findings help the user understand what doesn't need to change.

Finally, ask the user what they'd like to do:

• "Fix the issues" — generate patches
• "Run the tests" — execute tox run -e test-<pkg>
• "Just reviewing" — done

Source: Arize-ai/openinference — distributed by TomeVault.