docs: add OrcaReplay to external tracing processors list

[xizhuomengcontin]

Summary

Adds OrcaReplay to the existing external tracing processors list — one line, docs/tracing.md.

The published orcareplay-openai-agents implements the SDK's own TracingProcessor surface — on_trace_start, on_trace_end, on_span_start, on_span_end, shutdown, force_flush — and registers through set_trace_processors(), or add_trace_processor() when the SDK's own exporter is to be kept alongside it.

It exports bounded, metadata-only spans, and only the three types its reader consumes:

span	fields kept
AgentSpanData	name, handoffs, tools, output_type
HandoffSpanData	from_agent, to_agent
GuardrailSpanData	name, triggered

Everything else is dropped, including every other span type. Raw prompts, model output, tool arguments, tool results, mcp_data, CustomSpanData payloads, transcription and speech audio, and SpanError.data never reach the file — an allow-list rather than a deny-list, so a field a future SDK version adds stays out without a code change. ResponseSpanData and GenerationSpanData are dropped for a second reason as well: OrcaReplay records model traffic at a proxy and already holds those exchanges byte for byte, so a second copy would be a lossier duplicate of the same conversation.

What the tracing data is for: OrcaReplay records a run and replays it offline, and the three span types above are the part of a run a proxy cannot reconstruct. A handoff reaches the wire as a transfer_to_<agent> function tool, which names the destination and never the source; a guardrail that passes need make no request at all. Verified against openai-agents 0.20.0 and 0.22.2.

The linked page is the integration's own README: install, the interface it implements, what it writes, and how to turn it off.

Test plan

Docs-only, +1/-0 in docs/tracing.md. No code paths are touched, and the entry follows the existing format and position of the list.

The link target was checked after the package was published: pip install orcareplay-openai-agents on a clean interpreter, then a two-agent run with a handoff and a guardrail, which produced agent.start ×2, agent.handoff and agent.guardrail through the processor — on 0.22.2, two minor versions past the one it was written against.

I did not run .agents/skills/code-change-verification/scripts/run.sh: it drives the repository's make targets and needs the full uv dev environment, which is unrelated to a one-line markdown addition. Happy to run it if you would rather I did.

Issue number

n/a

Checks

• I've added new tests, if relevant — docs-only
• I've run .agents/skills/code-change-verification/scripts/run.sh — see Test plan
• I've confirmed all verification steps pass — for the change made: the diff is one markdown line and the linked page resolves
• If using Codex, I've run /review before submitting this PR — not using Codex