fix: serialize guardrail_response to JSON in OTEL traces

[yassin-berriai]

Summary

Guardrail OTEL spans previously set guardrail_response via safe_set_attribute, letting dict payloads reach the exporter as Python repr strings (e.g. "{'id': 'modr-7740', ...}"). Downstream log pipelines could not parse those as JSON, breaking metric creation from guardrail traces.

Serialize guardrail_response with safe_dumps before setting the span attribute — same treatment masked_entity_count already gets two lines above. Drop the attribute entirely when the value is None.

Resolves LIT-3233

Changes

• litellm/integrations/opentelemetry.py — JSON-serialize guardrail_response, skip on None.
• tests/test_litellm/integrations/test_opentelemetry.py — update both test_create_guardrail_span_with_valid_info assertions to expect safe_dumps("filtered_content").

Test plan

• uv run pytest tests/test_litellm/integrations/test_opentelemetry.py -v — 219 passed
• uv run ruff check litellm/ — clean
• uv run black --check on both files — clean
• uv run mypy litellm/integrations/opentelemetry.py --ignore-missing-imports — clean
• @greptileai review with confidence score >= 4/5

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[yassin-berriai]

@greptileai please review

[greptile-apps]

Greptile Summary

This PR fixes guardrail_response in OTEL guardrail spans being stored as a Python repr string instead of valid JSON, which broke downstream log pipeline parsing. It serializes the value with safe_dumps (matching the existing treatment of masked_entity_count) and skips the attribute entirely when the value is None.

• opentelemetry.py: Replaces safe_set_attribute with safe_dumps + direct set_attribute for guardrail_response, guarded by a None check.
• test_opentelemetry.py: Updates two existing assertions to expect JSON-serialized values; adds test_guardrail_response_dict_is_json_serialized and test_guardrail_response_none_is_skipped to cover the primary bug case and the None-skip behaviour.

Confidence Score: 5/5

Safe to merge — the change is a one-liner in a non-critical observability path, consistent with the existing masked_entity_count pattern, and covered by targeted new tests.

The fix is isolated to the guardrail OTEL span writer, follows an already-established pattern in the same method, introduces no new dependencies, and is backed by two new tests that directly exercise the corrected behaviour plus updated assertions on the existing ones.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/integrations/opentelemetry.py	Switches guardrail_response from safe_set_attribute to safe_dumps + direct set_attribute, matching the existing pattern for masked_entity_count; adds a None guard to omit the attribute when absent.
tests/test_litellm/integrations/test_opentelemetry.py	Updates two existing assertions to expect safe_dumps-serialized values; adds two new focused tests covering dict payloads and the None-skipping behaviour.

_{Reviews (3): Last reviewed commit: "test: cover dict-serialization and None-..." | Re-trigger Greptile}

[greptile-apps]

String values are double-encoded by safe_dumps. When guardrail_response is already a plain string (e.g. "filtered_content"), safe_dumps wraps it in JSON quotes producing '"filtered_content"'. Downstream consumers reading the attribute expecting a bare string will now get an extra layer of quoting. The primary bug was with dict/object payloads; strings didn't need to be re-serialized. Consider only applying safe_dumps when the value is not already a str.

Suggested change

	guardrail_response = guardrail_information.get("guardrail_response")
	if guardrail_response is not None:
	guardrail_span.set_attribute(
	"guardrail_response", safe_dumps(guardrail_response)
	)
	guardrail_response = guardrail_information.get("guardrail_response")
	if guardrail_response is not None:
	guardrail_span.set_attribute(
	"guardrail_response",
	guardrail_response if isinstance(guardrail_response, str) else safe_dumps(guardrail_response),
	)

[greptile-apps]

Greptile Summary

This PR fixes guardrail_response being written to OTEL spans as a Python repr string (via str(dict)) instead of valid JSON, by applying safe_dumps — the same treatment already used for masked_entity_count two lines above. It also drops the attribute entirely when the value is None, which is an improvement over the old path where None would have been serialized as the string "None".

• opentelemetry.py: Replaces safe_set_attribute(guardrail_response) with a None-guarded guardrail_span.set_attribute("guardrail_response", safe_dumps(guardrail_response)).
• test_opentelemetry.py: Two assertion sites updated from the bare string to safe_dumps("filtered_content"), correctly reflecting that all guardrail_response values — including plain strings — are now JSON-encoded before being stored in the span.

Confidence Score: 5/5

Safe to merge — a small, targeted serialization fix with no side effects on the broader request path.

The change is minimal and self-contained: one guard-and-serialize block replaces one safe_set_attribute call, matching the pattern already used for masked_entity_count in the same function. The None guard is strictly better than the previous behavior where None would surface as the string "None". Test assertions were updated correctly and no coverage was weakened.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/integrations/opentelemetry.py	Replaces safe_set_attribute (which produced Python repr strings via str()) with safe_dumps for guardrail_response, and skips the attribute when the value is None — consistent with the existing masked_entity_count pattern two lines above.
tests/test_litellm/integrations/test_opentelemetry.py	Two test assertions updated to expect safe_dumps("filtered_content") instead of the bare string, accurately reflecting the new serialization behavior; no test coverage weakened.

_{Reviews (2): Last reviewed commit: "fix: serialize guardrail_response to JSO..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[yassin-berriai]

@greptileai I addressed your feedback by adding two new tests: one verifying that dict payloads (the OpenAI moderation case from the report) are JSON-serialized, and one verifying that None skips the attribute entirely. The double-quoting of plain-string responses is intentional — the goal of the fix is that the attribute is always valid JSON so downstream pipelines can parse it uniformly. Please re-review.