[aikins01]

Adds @telemetry-dev/openai-agents, a trace processor for the OpenAI Agents SDK (@openai/agents 0.21 and later). The SDK has its own tracing model and no OpenTelemetry export, so the processor maps each Agents trace to OTel spans through @telemetry-dev/sdk: agents, turns, model calls, tool calls, handoffs, guardrails, and the sandbox.* spans of sandbox agents. It's registered with addTraceProcessor, which keeps the OpenAI Traces exporter, or setTraceProcessors to replace it.

Only model calls (chat spans) carry token usage. The task, turn, and agent spans repeat the totals of their calls, so sending them too would count each call two or three times. Chat Completions calls keep their usage only in the stored completion, so the processor reads it from there.

Sandbox spans get sandbox.operation and sandbox.backend. The exec_command and write_stdin tool spans get process.exit.code, read from the header line Process exited with code N in the tool output, since the JS SDK puts no exit code on its spans. A non-zero exit is recorded but doesn't mark the span as an error, because commands like grep exit 1 without failing. instrumentSandboxClient writes the provider's sandbox id (E2B's sandboxId, for example) onto the session span, and the run's command spans copy it as sandbox.id, so the provider's own sandbox events can be joined to the trace. The command and working directory of sandbox.exec are input content: they're sent only with captureInput on and go through mask.

The trace lifecycle is where I'd look most closely. The Agents SDK ends a trace only when its run returns, fires span starts and ends without awaiting them, and calls each processor in turn, so a slow processor registered first delivers span ends after the trace end. The root span therefore ends once the trace and every one of its spans have ended. A run that throws gets its root ended after 60 seconds without new spans, or at the next forceFlush. Finished traces are kept for that long so late spans still join their root. shutdown, which the SDK also calls on SIGTERM, ends open spans but keeps their state, so a server that keeps draining requests still gets connected traces.

Before the first release, @telemetry-dev/openai-agents needs to exist on npm with its GitHub Actions trusted publisher set, as RELEASING.md describes. Otherwise the release's npm publish fails. The lockfile adds the package's dependencies; pnpm also moves the shared ws peer from 8.21.3 to 8.22.1 (the Agents SDK needs 8.22) and re-resolves @types/node under protobufjs.

The package tests run real agents against the local sandbox, with the OpenAI HTTP calls faked: vp test run in packages/openai-agents passes 19 tests, and each new assertion fails when its check is removed.