Auto-instrumentation emits less detailed Claude Agent SDK traces than wrapper instrumentation #1593
Description
Summary
Claude Agent SDK auto-instrumentation currently captures a shallower trace shape than the wrapper path.
What differs
The wrapper path emits richer semantic spans for tool calls and nested subagent work, while the auto-instrumented path mainly captures the top-level query(...) / agent span and related LLM activity.
Why this matters
Users get materially different trace detail depending on whether they use wrappers or auto-instrumentation. It also prevents full parity in the shared e2e trace contract.
Repro
The aligned e2e coverage in:
e2e/scenarios/wrap-claude-agent-sdk-tracese2e/scenarios/claude-agent-sdk-auto-instrumentation-node-hook
shows that wrapper-only assertions are still needed for tool spans, nested subagent spans, and failure-tool details.
Expected
Auto-instrumentation should emit trace detail that is closer to the wrapper path, especially for tool execution, subagent nesting, and failure spans.