Summary

• Extends the promote parameter on @trace() / @trace_async() to resolve keys from function outputs (dict, Pydantic model, or dataclass), not just inputs
• Keys are resolved from inputs first; unresolved keys fall back to output field extraction
• Fixes context loss bug in sync generator ASGI path by running promote inside captured_context.run()
• Adds 10 new tests and an example demonstrating all supported output types and nested traces

Closes OPEN-9557

How it works

@tracer.trace(
    promote={
        "user_query": "agent_input_query",       # resolved from input
        "tool_call_count": "agent_tool_calls",   # resolved from output
        "tool_names": "agent_tools",             # resolved from output
    }
)
def run_agent(user_query: str) -> AgentResult:
    ...

Promoted fields become top-level columns in the trace data, enabling Openlayer tests against child step data (e.g. agent_tool_calls < 10).

Test plan

• Promote from dict output
• Promote from Pydantic model output
• Promote from dataclass output
• Mixed input + output promotion
• List form (no aliasing) and dict form (with aliasing)
• Nested @trace() calls — child promotes to parent trace
• Async function promotion
• Missing key warning (key in neither inputs nor outputs)
• None output safety
• Existing promote-input-only tests pass (regression)

🤖 Generated with Claude Code