chore(release): backport #32542, #32655 to stable/1.91.x and cut 1.91.3

[yuneng-berri]

Relevant issues

Backports two already-merged litellm_internal_staging changes onto stable/1.91.x and cuts 1.91.3. The tip 1.91.2 is already published, so this line needs the patch bump

#32542 fixes a Responses-API bypass in the shared guardrail content helpers: every guardrail that reads litellm/proxy/guardrails/_content_utils.py saw zero text on the /v1/responses path, so AIM turned it into a 422 and every other content guardrail scanned an empty payload and let the request through unscanned. #32655 adds the GenAI-semconv gen_ai.client.operation.exception log event so a failed LLM call is recorded with exception.type / exception.message / exception.stacktrace at WARN, which OTel v2 never emitted before

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

What is included

In staging merge order, then the release bump:

• fix(guardrails): walk Responses-API text taxonomy in shared content helpers #32542 fix(guardrails): walk Responses-API text taxonomy in shared content helpers (cherry-picked from e84a19acd5)
• feat(otel): emit the gen_ai.client.operation.exception event on failed LLM calls #32655 feat(otel): emit the gen_ai.client.operation.exception event on failed LLM calls (cherry-picked from 99b4c5ed3e)
• bump: version 1.91.2 -> 1.91.3
• chore: refresh uv.lock for 1.91.3

Adaptation notes

#32542 applied verbatim (patch-id identical to the staging commit)

#32655 needed one import resolution in litellm/integrations/otel/logger.py. Staging imports from opentelemetry.context import Context, attach, get_current; this line dropped Context because it predates staging's identity-baggage and MCP-list-tools work that referenced it, so the pick keeps the line's from opentelemetry.context import attach, get_current and adds the fix's own new from opentelemetry.sdk._logs import LoggerProvider verbatim. Keeping Context would be an unused-import failure on this line. Nothing else diverges; the OTel dependency pins are identical between the line and staging (opentelemetry-sdk==1.28.0, opentelemetry-api==1.28.0), and OpenTelemetryV2Config.enable_events already exists on the line

Known noise on this line

None. The pre-pick baseline of the touched test files was 131 passed, 0 failures, so any red on those files is introduced by this PR and should be treated as a real signal

Screenshots / Proof of Fix

Verified end-to-end against a live proxy running this branch's source, hitting real provider APIs. The extraction proof uses a small custom pre_call guardrail that logs what _content_utils extracted (iter_message_text fragments, build_inspection_messages roles, walk_user_text visit count); the OTel proof runs the otel callback with LITELLM_OTEL_V2=true, LITELLM_OTEL_INTEGRATION_ENABLE_EVENTS=true, and the default console log exporter, then fails a call with a deliberately-bogus OpenAI key so OpenAI returns a genuine 401 that maps to AuthenticationError (nothing mocked)

#32542 - /v1/responses before vs after (same curl)

curl -sS http://localhost:4001/v1/responses -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" -d '{
  "model":"real-haiku",
  "input":[
    {"role":"user","content":[{"type":"input_text","text":"my aws key is AKIAEXAMPLE, help"}]},
    {"type":"function_call","call_id":"c1","name":"lookup","arguments":"{}"},
    {"type":"function_call_output","call_id":"c1","output":"tool saw AKIAEXAMPLE in logs"}
  ]
}'

Before (pre-pick, the bug - every text guardrail is a no-op on this path):

REPRO32542 call_type=aresponses fragments_seen=0 walk_visited=0 inspection_roles=[] fragments=[]

After (this branch - the Responses taxonomy is walked, and function_call_output surfaces as role: "tool"):

REPRO32542 call_type=aresponses fragments_seen=2 walk_visited=2 inspection_roles=['user', 'tool'] fragments=['my aws key is AKIAEXAMPLE, help', 'tool saw AKIAEXAMPLE in logs']

/v1/chat/completions was the control and reported fragments_seen=1 both before and after, so chat scanning is unchanged

#32655 - failed call before vs after (same curl)

curl -sS http://localhost:4001/v1/chat/completions -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" -d '{"model":"broken-openai","messages":[{"role":"user","content":"trigger a failure"}]}'

Both before and after, the call returns a real 401 AuthenticationError. Before, grepping the proxy log for the event returns nothing; after, exactly one record appears:

"event.name": "gen_ai.client.operation.exception"
"severity_number": "<SeverityNumber.WARN: 13>"
"exception.type": "AuthenticationError"
"exception.message": "litellm.AuthenticationError: AuthenticationError: OpenAIException - Incorrect API key provided: sk-proj-*************0000. ..."
"exception.stacktrace": "  File \".../litellm/main.py\", line 690, in acompletion ..."

The span-side error.type / error.message / litellm.provider.error.* keys are still present alongside it, so the existing error surface is untouched

Sanity and auth path (this branch)

A real anthropic/claude-haiku-4-5 completion returns pong; /key/generate plus a scoped completion with the generated key both succeed

Tests

Targeted suite on this line: 158 passed, 0 failures (test_content_utils.py, test_aim.py, test_otel_v2_components.py, test_otel_v2_logger.py), up from the 131-pass / 0-fail baseline by the tests these PRs add. As a regression check for existing callers of the modified helpers (build_inspection_messages, iter_message_text, walk_user_text, the SpanEmitter constructor), the full tests/test_litellm/proxy/guardrails/ and tests/test_litellm/integrations/otel/ trees pass 2371, 0 failures on this branch

Type

🆕 New Feature
🐛 Bug Fix

Changes

Two cherry-picks onto stable/1.91.x plus the 1.91.3 version bump and lock refresh. No dependency, schema, migration, auth-default, CI, or UI changes; the diff is code and tests only

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ yucheng-berri
✅ yuneng-berri
❌ yassin-berriai
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This backport cut (1.91.3) cherry-picks two already-merged fixes onto stable/1.91.x plus a version bump and lock refresh. Both changes are well-tested and the diff is limited to the intended files with no schema, auth, or dependency changes.

• Guardrail content helpers (_content_utils.py, aim.py): extends TEXT_PART_TYPES to recognise input_text / output_text Responses-API part types, rewrites _coerce_input_to_messages and walk_user_text to walk the full Responses-API item taxonomy (message, function_call, function_call_output), and adds an AIM-specific wrapper that coerces any role outside {system, user, assistant} to user before the AIM POST to avoid schema-validation errors.
• OTel V2 exception event (events.py, providers.py, emitter.py, logger.py, semconv.py): adds GenAIEventRecorder that emits the gen_ai.client.operation.exception log event at severity WARN on failed LLM-call spans, gated on enable_events (off by default); the existing span-side error.* attributes and exception span event are preserved for backwards compatibility.

Confidence Score: 4/5

Safe to merge; changes are well-scoped cherry-picks with comprehensive unit-test coverage and no modifications to auth, schema, or dependency versions.

Both cherry-picks apply cleanly and the import adaptation described in the PR notes (dropping the unused Context import) is correct. The guardrail content-helper rewrite is logically sound: function_call items (which carry no text content) are intentionally skipped in both the read and write paths; function_call_output is correctly routed to _rewrite_content for the walk path. The AIM role-coercion wrapper accesses m["role"] directly, which is safe because every code path in build_inspection_messages produces messages with an explicit role key. The OTel event plumbing uses opentelemetry.sdk._logs private-module imports, but the OTel dependency is pinned at 1.28.0 so API drift is bounded. Test additions are all unit-only with in-memory exporters; no existing test assertions were weakened.

No files require special attention. providers.py is the most complex addition (log exporter factory + provider resolution) but it closely mirrors the existing metrics pattern and is thoroughly tested.

Important Files Changed

Filename	Overview
litellm/proxy/guardrails/_content_utils.py	Extends TEXT_PART_TYPES to cover Responses-API part types and rewrites _coerce_input_to_messages / walk_user_text to walk the full Responses taxonomy (message items, function_call, function_call_output); logic is correct and well-tested
litellm/proxy/guardrails/guardrail_hooks/aim/aim.py	Adds _build_aim_inspection_messages to coerce non-standard roles to 'user' before the AIM POST; uses direct m['role'] access which is safe given build_inspection_messages always returns role-keyed dicts
litellm/integrations/otel/plumbing/events.py	New file; implements GenAIEventRecorder emitting gen_ai.client.operation.exception log events at WARN severity with the required exception.* trio; straightforward and correct
litellm/integrations/otel/plumbing/providers.py	Adds log exporter/provider factories (_otlp_logs_endpoint, build_log_exporter, build_logger_provider, resolve_logger_provider, get_event_logger) mirroring the existing metrics pattern; uses opentelemetry.sdk._logs private API which is pinned at 1.28.0
litellm/integrations/otel/emitter.py	Wires optional GenAIEventRecorder into SpanEmitter; event emission is correctly guarded on both event_recorder presence and SpanRole.LLM_CALL
litellm/integrations/otel/logger.py	Strengthens logger_provider type from Any to LoggerProvider and adds _init_events helper that gates event recorder construction on enable_events config; clean DI pattern
litellm/integrations/otel/model/semconv.py	Adds STACKTRACE constant to ExceptionEvent and new GenAIEvent class with OPERATION_EXCEPTION constant; pinned constant values match the semconv spec
tests/test_litellm/proxy/guardrails/test_content_utils.py	Adds comprehensive new tests for Responses-API content-part types and function_call_output taxonomy; existing test assertions unchanged (only whitespace reformatting)
tests/test_litellm/proxy/guardrails/guardrail_hooks/test_aim.py	New test file; covers role coercion in _build_aim_inspection_messages for tool, developer, and function_call_output inputs; unit tests only, no network calls
tests/test_litellm/integrations/otel/test_otel_v2_components.py	New tests for log exporter variants, logger-provider selection, and operation exception event emission; all use in-memory exporters, no network calls
tests/test_litellm/integrations/otel/test_otel_v2_logger.py	New integration tests verify enable_events wires through the full failure-callback path and that events are off by default; clean additions
pyproject.toml	Version bump 1.91.2 → 1.91.3 in both [project] and [tool.commitizen] blocks

_{Reviews (1): Last reviewed commit: "chore: refresh uv.lock for 1.91.3" | Re-trigger Greptile}