[OTel] Stop stashing the session ID in gen_ai.agent.name

[krisztianfekete]

In utils/log_enrichment.py we write the session ID into the semconv attribute gen_ai.agent.name, then read it back later as the agent's display name.
That's a semconv misuse and it means a real agent name would get clobbered. Use a namespaced agentevals.session.id instead.

[AaronProbha18]

hey @krisztianfekete
I'd like to work on this issue.

Plan

The current enrichment flow conflates two different concepts by storing the session identifier in gen_ai.agent.name, which causes real agent names from telemetry to be replaced. The goal is to separate those concerns by introducing a dedicated agentevals.session.id attribute while leaving the OpenTelemetry semantic-convention field available for actual agent metadata.

I'll start by tracing where the session ID is written and where gen_ai.agent.name is interpreted downstream to confirm whether any compatibility handling is needed. From there, I'll add the new attribute, update the enrichment pipeline to emit it consistently across all enrichment paths, and adjust metadata extraction so session IDs and agent names are sourced independently.

Changes

• src/agentevals/utils/log_enrichment.py will be updated so enrichment writes the session identifier to agentevals.session.id instead of gen_ai.agent.name, ensuring both per-span and broadcast enrichment paths remain consistent.
• src/agentevals/trace_attrs.py will define the new telemetry attribute as a centralized constant so all producers and consumers reference the same key.
• src/agentevals/trace_metrics.py will separate session ID extraction from agent metadata extraction, preserving gen_ai.agent.name as the canonical agent display name while reading session information from the new custom attribute.
• tests/test_extraction.py will be updated to verify that session IDs are extracted from the new attribute, genuine agent names remain intact, and existing extraction behavior outside this change is preserved. I'll also add enrichment-focused coverage if there are dedicated tests for that path.

[krisztianfekete]

Thanks for your interest in this one!
A few comments before you open a PR:

Keep the scope tight to the sessionid change. New agentevals.session.id constant in trace_attrs.py, stop writing the session id into gen_ai.agent.name. Please leave converter.py's use of gen_ai.agent.name alone.

trace_metrics.py needs no new reader. Once enrichment stops polluting gen_ai.agent.name, extract_agent_identity / extract_trace_metadata already pick up the real agent name (or fall back correctly).

Note that

• sessionid is appended in both _inject_messages and the no-logs branch of _enrich_per_span. Worth factoring the append into one small helper so they can't drift probably.
• this enrichment also runs on OTLP live sessions, not just the custom SDK path, so real OTLP agent names are being overwritten too. Good to cover that in the test, especially in context of Retire the custom WebSocket ingestion #170