Fix microsoft.gen_ai.main_agent.* propagation and self-promotion issues

[hectorhdzg]

Fixes three issues identified during MAF Python and LangChain Python testing:

Issue 1: on_end self-promotion never fires (MAF + LC Python)

Symptom: Root invoke_agent spans never get microsoft.gen_ai.main_agent.* attributes. They have gen_ai.agent.name but self-promotion to microsoft.gen_ai.main_agent.name never happens.

Root cause: The OTel SDK's SpanProcessor.on_end() receives a ReadableSpan, not a Span. ReadableSpan does NOT have set_attribute(). The old code had:

if not hasattr(span, "set_attribute"):
    return  # ALWAYS returns, ReadableSpan never has set_attribute

This guard silently aborted self-promotion for every span.

Fix: Write directly to span._attributes (the internal BoundedAttributes dict) instead of calling set_attribute(). Mutations are visible to downstream SpanExporters since they read from the same object.

Issue 2: on_start inheritance misses late-set parent attributes (MAF Python)

Symptom: Some chat, execute_tool, and HTTP spans inside a trace whose invoke_agent ancestor has gen_ai.agent.name end up with no microsoft.gen_ai.main_agent.name.

Root cause: on_start reads parent.attributes at child-creation time. If the parent's gen_ai.agent.* attributes are set AFTER the child span is created (timing issue in some MAF SDK flows), on_start sees an empty parent and propagates nothing. No fallback existed.

Fix: Store a reference to the parent Span during on_start. During on_end, if the span still lacks microsoft.gen_ai.main_agent.* and is not an invoke_agent span, re-read from the stored parent. By on_end time the parent attributes have been set, so the fallback succeeds. References are cleaned up in on_end and shutdown().

Issue 3: Nested LangChain agent gen_ai.agent.name overwritten by parent (LC Python)

Symptom: A nested invoke_agent span for a sub-agent (e.g. "Data agent") shows the parent agent's name (e.g. "Main agent") in gen_ai.agent.name, while gen_ai.agent.id and gen_ai.agent.description correctly show the sub-agent's values.

Root cause: _resolve_agent_name() always checked self._agent_config.get("agent_name") first - the shared top-level config. For nested sub-agents, this override clobbered the sub-agent's own name from run metadata. agent_id/description were correct because extract_agent_metadata(run) via set_attributes() ran after config values, but agent_name from config won because _resolve_agent_name checked it first.

Fix: Added use_config parameter to _resolve_agent_name(). Nested agents (detected via _find_agent_ancestor()) pass use_config=False so they derive identity solely from run metadata. Config-based agent_id, agent_description, and agent_version are also skipped for nested agents.

[github-actions]

Performance comparison

Threshold: regressions >15.0% on gating scenarios fail the build. Higher ops/s is better; positive Δ means the PR is slower.

Scenario	Gating	Baseline (ops/s)	Candidate (ops/s)	Δ %	Status
azure_monitor_log	yes	26,990.6	26,688.0	+1.13%	✅
azure_monitor_span	yes	160,205.1	158,957.2	+0.79%	✅
otel_log	no	32,725.7	32,941.3	-0.65%	✅
otel_span	no	34,395.0	34,924.7	-1.52%	✅

[Copilot]

Pull request overview

This PR updates the GenAI main-agent attribution flow so microsoft.gen_ai.main_agent.* attributes are reliably populated across SDK spans (including timing edge-cases) and LangChain nested-agent runs (avoiding top-level config overriding nested agent identity).

Changes:

• Fix GenAIMainAgentSpanProcessor.on_end() to self-promote by mutating the underlying span attributes mapping and add an on-end fallback propagation path for late-set parent attributes.
• Expand real-SDK tests to cover on-end fallback propagation and root-span self-promotion behavior.
• Adjust LangChain tracer agent-name resolution so nested agents don’t inherit identity from shared top-level _agent_config.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
src/microsoft/opentelemetry/_genai/main_agent/_processor.py	Adds parent tracking for on-end fallback propagation and switches on-end enrichment to mutate internal attributes.
src/microsoft/opentelemetry/_genai/_langchain/_tracer.py	Prevents nested agents from using top-level config overrides for agent identity fields.
tests/genai/main_agent/test_span_processor.py	Updates unit tests to validate on-end enrichment via internal attributes mutation.
tests/genai/main_agent/test_sdk_propagation.py	Adds/updates real-SDK regression tests for timing recovery and self-promotion scenarios.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[JacksonWeber]

LGTM