fix(tracing): keep trace state on completed non-streamed runs

[BlueX888]

Summary

A non-streamed run that finishes normally produces a RunResult whose to_state() carries no trace. Resuming from that state with Runner.run(agent, state) opens a new trace instead of reattaching to the original one, so a single logical run ends up split across two traces.

Root cause: AgentRunner._run_impl (src/agents/run.py:584) has several RunResult construction sites, and the completion path at src/agents/run.py:2067-2070 was the only one that did not assign result._trace_state = run_state._trace_state. RunResult has no trace attribute, so the fallback in result_to_run_state (src/agents/result.py:166-169) resolves to None and to_state() writes a state without trace information. On resume, create_trace_for_run(..., trace_state=run_state._trace_state, reattach_resumed_trace=is_resumed_state) (src/agents/run.py:2508) then has nothing to reattach to and starts a fresh trace. The other construction sites (src/agents/run.py:1426, src/agents/run.py:1607, src/agents/run.py:2590) already copy it.

This is the remaining case of #2540, whose fix (#2547) covered the interrupted-run path but not the completed-run path. The symptom is not limited to human-in-the-loop: Runner.run(agent, input=...) followed by Runner.run(agent, result.to_state()) also starts two traces on the current tree.

Changes

• src/agents/run.py:2070 — carry run_state._trace_state onto the result in the completion path, matching the other RunResult construction sites in the module.
• tests/test_agent_tracing.py — add test_resumed_run_after_completed_turn_reuses_original_trace.

Test plan

The new test runs an agent with a needs_approval tool, approves the interruption, lets the run finish, then resumes from approved.to_state(). It asserts exactly one trace and one trace_start/trace_end pair.

With the fix reverted (test present, src/agents/run.py at the base revision) it fails:

    traces = fetch_traces()
>       assert len(traces) == 1
E       assert 2 == 1
E        +  where 2 = len([<agents.tracing.traces.TraceImpl object at 0x10c53ff50>, <agents.tracing.traces.TraceImpl object at 0x10c53fb60>])

tests/test_agent_tracing.py:450: AssertionError
=========================== short test summary info ============================
FAILED tests/test_agent_tracing.py::test_resumed_run_after_completed_turn_reuses_original_trace
============================== 1 failed in 0.28s ==============================

With the fix applied:

• uv run pytest tests/test_agent_tracing.py::test_resumed_run_after_completed_turn_reuses_original_trace -> 1 passed in 0.20s
• uv run pytest tests/test_agent_tracing.py -> 34 passed in 0.52s
• uv run pytest tests/test_agent_tracing.py tests/test_tracing.py tests/test_tracing_errors.py tests/test_tracing_errors_streamed.py tests/test_run.py tests/test_run_state.py tests/test_run_impl_resume_paths.py tests/test_responses_tracing.py -> 642 passed in 3.24s
• bash .agents/skills/code-change-verification/scripts/run.sh -> code-change-verification: all commands passed. (make format, make lint, make typecheck; 9637 passed, 29 skipped in 47.92s parallel, 77 passed, 4 skipped, 55 deselected in 1.70s serial)

Issue number

N/A — no issue was filed for this. Related: #2540, fixed by #2547.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

[karaaslanz]

There looks to be one remaining gap in the non-HITL path described in the PR body.

The new assignment is still inside if run_state is not None, so it only copies trace state when this completion itself came from a resumed RunState. For a fresh Runner.run(agent, input=...), run_state is None, this line is skipped, and I don't see another completion/finalization assignment that puts the current trace onto the resulting RunResult. result_to_run_state() then falls back from a missing _trace_state to getattr(result, "trace", None), but RunResult does not expose that trace surface.

The regression test currently starts with an approval interruption, so approved is already a resumed run and exercises exactly the branch this patch fixes. It does not cover the separate scenario called out in the description: Runner.run(agent, input=...) -> result.to_state() -> Runner.run(agent, state).

Could we add a direct fresh-completion regression for that path and either preserve the active trace state for fresh results as well, or narrow the stated behavior if continuing a fully completed fresh result is intentionally unsupported? That would keep the implementation/test aligned with the broader claim in the PR.