docs: add OrcaReplay to the external tracing processors list

[xizhuomengcontin]

What

One line appended to External tracing processors list in docs/tracing.md.

Disclosure: I maintain OrcaReplay. Apache-2.0, free, no paid tier, no account, no hosted service.

It is a tracing processor, which is what this list is for

orcareplay-openai-agents installs an OrcaTracingProcessor through the SDK's own API:

from agents import add_trace_processor
add_trace_processor(processor)

It implements on_trace_start / on_trace_end / on_span_start / on_span_end and writes agent, handoff and guardrail spans to disk. It is not a subclass of agents.tracing.TracingProcessor only because that would force an SDK import at definition time; it satisfies the protocol.

Two details that seemed worth getting right rather than assuming:

• add_trace_processor is used, not set_trace_processors, so a user who later registers their own still gets theirs — ours is appended, not substituted. A user who deliberately calls set_trace_processors replaces ours, which is their explicit choice.
• Installing it does not disable the OpenAI traces dashboard.

pip install orcareplay-openai-agents

What it is for, stated plainly

The rest of OrcaReplay records model traffic at a proxy and replays it later with the provider unreachable, so a past run reproduces without a key. That proxy sees HTTP and nothing else — which agent was speaking, which handoff fired, which guardrail tripped are not on the wire. This package supplies exactly that structure, so the recording carries the run's shape alongside its bytes.

Where it sits relative to the rest of the list: the other entries send traces to a backend for viewing and analysis. This one writes them next to a recording that can be served back, so the same session happens again offline. Complementary rather than competing — several people will reasonably run one of the above and this.

Limits I would rather state than have found

• A replay is not a determinism result. It proves the recorded run reproduces, not that the model is stable.
• egress=blocked means model-provider egress only. Recorded tool calls still execute for real on replay; it is not a sandbox.
• This package is not required for recording an Agents SDK run. A run recorded without it is complete in the HTTP sense; the package adds the structural layer.

Placement

Appended to the end of the list, which is where the most recent entries have gone, rather than inserted by perceived importance. Happy to move it, reword it, or have it declined — it is your list.