Render MCP list-tools spans instead of leaving them blank

[laminar-coding-agent]

Problem

_apply_mcp_span_data is the only handler in openai_agents/span_data.py that never calls set_lmnr_span_io. It writes openai.agents.mcp.server and openai.agents.mcp.result and stops there — but those custom attributes are not rendered as span I/O, so an agents.mcp_tools span arrives with lmnr.span.input / lmnr.span.output unset and shows up completely empty in the trace view.

Found while reviewing a chi-bench collection run: 15 agents.mcp_tools spans, every one blank, with all 30 tool names sitting unread in openai.agents.mcp.result.

input len: 0 | output len: 0
openai.agents.mcp.server = chi_bench
openai.agents.mcp.result = ["chart.get_patient_chart_by_patient","chart.list_candidate_orders", ...]

Fix

Set span I/O alongside the attributes — server in, tool list out. Nothing else changes; the existing attributes stay for anyone querying them.

Testing

New tests/test_instrumentations/test_openai_agents/test_mcp_span_data.py (2 tests) covers the populated case and the no-result case. The full tests/test_instrumentations/test_openai_agents/ suite passes (9 passed).

Also exercised end-to-end by a live chi-bench run exporting to production Laminar.

Notes

CLAUDE.md gains an openai_agents section recording the invariant (every _apply_* handler sets span I/O), plus two things that look like bugs and aren't: the first turn's mcp_tools span parenting to the trace root is upstream SDK ordering, and one mcp_tools span per turn is expected even with cache_tools_list=True (the repeats are ~0.3ms cache hits).

🤖 Generated with Claude Code

---

Note

Low Risk
Small instrumentation change that only adds trace-view I/O attributes for MCP list-tools spans; no auth, export, or runtime behavior changes.

Overview
agents.mcp_tools spans were empty in the trace UI because _apply_mcp_span_data only set openai.agents.mcp.* attributes, which the frontend does not use for I/O.

The handler now also calls set_lmnr_span_io: MCP server name as input ({"server": ...}) and the tool list as output. Existing openai.agents.mcp.server / openai.agents.mcp.result attributes are unchanged for queries.

CLAUDE.md documents the openai_agents invariant (every _apply_*_span_data handler should set span I/O) and notes two upstream SDK behaviors (first-turn MCP span parenting, one MCP span per turn even with cached tool lists).

Tests: test_mcp_span_data.py covers populated and no-result MCP list-tools spans.

^{Reviewed by Cursor Bugbot for commit c1573e8. Bugbot is set up for automated code reviews on this repo. Configure here.}