Render mcp_list_tools spans instead of leaving them blank

[laminar-coding-agent]

Why

On any Agents-SDK run with an MCP server, every agents.mcp_tools span renders empty in the trace view. A 15-turn chi-bench run produces 15 blank TOOL spans, which reads as broken instrumentation.

The data was never missing. _apply_mcp_span_data wrote it to openai.agents.mcp.server / openai.agents.mcp.result, and the trace view reads lmnr.span.input / lmnr.span.output. It was the only _apply_*_span_data handler in the file that never called set_lmnr_span_io.

What changed

_apply_mcp_span_data now also sets span I/O — the server name as input, the returned tool list as output. The openai.agents.mcp.* attributes are unchanged, so anything querying them in SQL keeps working.

Verification

New tests/test_openai_agents_span_data.py (3 tests, hermetic — no SDK or OTEL provider needed):

3 passed, 5 warnings in 7.31s

Covers the populated case plus the two partial ones (no server, no result), asserting that a None field records no attribute rather than a null payload.

Confirmed against a production trace before the fix — all 15 spans had input len: 0 | output len: 0 while carrying:

openai.agents.mcp.server = chi_bench
openai.agents.mcp.result = ["chart.get_patient_chart_by_patient", "chart.search_patients", ...]  (30 tools)

Also documented in CLAUDE.md

Three related non-obvious behaviours found while diagnosing this, none of them bugs:

• mcp_tools spans are emitted once per turn even with cache_tools_list=True — the span wraps server.list_tools(), which checks the cache internally. Measured: first span 11ms (real round trip), the other 14 ~0.3ms.
• The first turn's mcp_tools span parents to the trace root, not the agent span — agents/run.py calls get_all_tools() before it starts agent_span(...). One sibling-of-the-agent span per run is upstream ordering, not a parenting bug.
• A custom openai.agents.* attribute is not visible in the trace view at all, which is the general form of the bug fixed here.

🤖 Generated with Claude Code

---

Note

Low Risk
Localized instrumentation change with unit tests; no auth, export, or API surface changes.

Overview
Fixes blank mcp_tools / mcp_list_tools spans in the Laminar trace UI. Those spans already had openai.agents.mcp.server and .result, but the view only shows lmnr.span.input / lmnr.span.output.

_apply_mcp_span_data now calls set_lmnr_span_io: server name as input (when present) and the tool list as output. The custom openai.agents.mcp.* attributes are unchanged for SQL queries.

Adds hermetic tests/test_openai_agents_span_data.py (full case, no server, no result). CLAUDE.md documents OpenAI Agents tracing pitfalls (trace view vs custom attrs, per-turn mcp_tools spans, first-turn parenting, tool defs on Responses only).

^{Reviewed by Cursor Bugbot for commit fdafd3c. Bugbot is set up for automated code reviews on this repo. Configure here.}