fix(otel): record full error message on standard exception event in otel v2

[yassin-berriai]

Relevant issues

Error messages emitted by the OpenTelemetry v2 integration are truncated downstream. Reported via EAIP-2334; the root cause analysis there is that the backend (Elasticsearch) applies a dynamic index field-type mapping that converts the error message field from string to keyword, and keyword defaults to ignore_above: 1024, so anything past 1024 chars is dropped

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• 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

Screenshots / Proof of Fix

Ran the proxy with otel v2 + the console exporter (LITELLM_OTEL_V2=true OTEL_EXPORTER=console), pointing a model at a dead endpoint so the call fails and the v2 failure path emits the LLM-call span.

Config used:

model_list:
  - model_name: broken-model
    litellm_params:
      model: openai/gpt-4o-mini
      api_key: sk-fake-key
      api_base: http://127.0.0.1:9/v1
litellm_settings:
  callbacks: ["otel"]
general_settings:
  master_key: sk-1234

Request:

curl -s http://127.0.0.1:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"model":"broken-model","messages":[{"role":"user","content":"hi"}]}'

Before the fix, the emitted chat broken-model span carried the message only in the status description; its event list was empty, so the only place the message lived was a field a backend dynamic-maps to a truncating keyword:

"status": {
    "status_code": "ERROR",
    "description": "litellm.InternalServerError: InternalServerError: OpenAIException - Connection error."
},
"error.type": "InternalServerError"
"events": []

After the fix, the same span carries the full message on the standard exception event under exception.message (recognized semconv field, mapped as full text), while error.type stays a low-cardinality attribute:

"error.type": "InternalServerError"
"events": [
    {
        "name": "exception",
        "attributes": {
            "exception.type": "InternalServerError",
            "exception.message": "litellm.InternalServerError: InternalServerError: OpenAIException - Connection error."
        }
    }
]

Type

🐛 Bug Fix

Changes

The v2 span engine (emitter.py) previously stamped only error.type and put the message into the span status description; it never recorded the standard OTel exception event that v1 emits via record_exception. A non-semconv field name (e.g. error_message) falls into the backend's default dynamic template, which maps strings to a keyword capped at ignore_above: 1024 and truncates them.

finish_span now also records the standard exception span event carrying exception.type and the full exception.message, so the message rides a recognized semantic-convention field that backends map as full text rather than a truncated keyword. Added ExceptionEvent semconv constants and a regression test asserting a 5000-char message survives intact on the event (and that success spans record no exception event)

---

Generated by 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.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[yassin-berriai]

@greptileai

[greptile-apps]

Greptile Summary

This PR fixes truncation of OTel v2 error messages by adding a standard exception span event carrying exception.message and exception.type on the error path in emitter.py. Previously, the message lived only in the span status description and as a bare string attribute; backends like Elasticsearch dynamic-map such attributes as keyword fields capped at 1024 characters.

• semconv.py gains an ExceptionEvent constant class with the standard exception.* OTel attribute keys (NAME, TYPE, MESSAGE).
• emitter.py calls span.add_event() with those semconv keys when a span error is present, so the full message rides a recognized field that backends map as full text.
• Two new in-memory-exporter regression tests assert a 5000-char message survives intact on the event and that success spans emit no exception event.

Confidence Score: 5/5

Additive-only change to the OTel v2 error path; no existing attributes or status fields are removed, and the new event is skipped entirely on the success path.

The change is narrowly scoped to finish_span's error branch: it adds one span.add_event() call using standard OTel semconv constants and leaves every existing attribute and status field intact. The two new tests exercise both the error and success code paths with an in-memory exporter and make no real network calls.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/integrations/otel/emitter.py	Adds a standard OTel exception span event carrying exception.type and exception.message on the error path; refactors two repeated expressions into named variables. No existing behaviour changed.
litellm/integrations/otel/model/semconv.py	Adds ExceptionEvent semconv constant class (NAME, TYPE, MESSAGE), matching standard OTel exception.* attribute keys. Clean additive change.
tests/test_litellm/integrations/otel/test_otel_v2_components.py	Adds two new regression tests and two helper functions. Tests use the in-memory exporter only (no real network calls), covering both the long-message survival path and the no-event-on-success path.

_{Reviews (2): Last reviewed commit: "fix(otel): record full error message on ..." | Re-trigger Greptile}