fix(client): release the global OTel tracer provider on shutdown

[XieX]

Companion to launchdarkly/js-ai-sdk#103, which fixes the same lifecycle bugs in JS. Two of them exist here, and one has a Python-specific counterpart.

1. An init/shutdown/init cycle exported nothing

shutdown() dropped its provider handle but left OpenTelemetry's global tracer provider registered. That global is once-guarded — a second set_tracer_provider logs Overriding of current TracerProvider is not allowed and keeps the provider already in place — so after a re-init, every span routed to the provider that had just been shut down. Observed across two cycles: the global service.name stayed cycle1 and now correctly becomes cycle2.

Releasing it means resetting both the global slot and the Once that guards it; clearing the slot alone leaves the guard tripped, so the next set is a silent no-op. Both are private, since opentelemetry-python has no public way to unset them, so _release_otel_globals's docstring explains the reach. The propagator needs no reset — set_global_textmap is a plain assignment.

The release only happens when our set_tracer_provider actually took effect (trace.get_tracer_provider() is provider). Bugbot caught that the first version gated it on having built a provider, which would have wiped a host app's already-registered provider; the JS PR had the same flaw. A refused set now logs a warning.

2. A failed telemetry setup left a half-initialized client behind

_client is not None is the idempotency guard, but _client was assigned before _setup_telemetry ran. A malformed OTEL_EXPORTER_OTLP_TIMEOUT makes setup raise, and _client stayed set: the next init_client() returned the client as a silent success with no telemetry, and on the SDK-key path its connection was never closed. _client is now assigned only after setup succeeds. On the SDK-key path the client we built is closed on failure; a BYOC client is left open, since the caller owns it. JS's counterpart was a cached failed-init promise.

Not a bug here: BYOC idempotency

JS checked the singleton below the pre-initialized-client branch and re-registered OTel on every initClient(client). _resolve_client checks it first — 1 _setup_telemetry call across three inits — and two tests pin that ordering.

Verification

• 8 new tests; each one covering a bug here was confirmed to fail against the code it fixes. The end-to-end cycle and env-var tests drive the real _setup_telemetry, with the OTLP exporter stubbed or failing before any network use.
• Full suite 1356 pass. ruff check, ruff format, and mypy are clean.

🤖 Generated with Claude Code

---

Note

Overview
Fixes init/shutdown/init so telemetry works on the second cycle: shutdown() and test reset now release the process-global OTel tracer provider (via _release_otel_globals, clearing OTel’s private slot and Once guard) only when _owns_otel_globals is true—i.e. this SDK’s set_tracer_provider actually won. If another library registered first, shutdown still shuts down the provider we built but does not clear the host’s global, and a warning is logged when LD’s telemetry config is not in effect.

Init hardening: the singleton _client is set only after _setup_telemetry succeeds (BYOC and SDK paths). Failed telemetry no longer leaves a “successful” half-init; SDK-path failures close the new LDClient before re-raising. BYOC clients are never closed on failed init.

Docs in agents.md describe shutdown’s OTel release. Tests add restore_otel_globals, coverage for repeat-init idempotency, failed telemetry, full two-cycle service.name, and foreign-provider safety.

^{Reviewed by Cursor Bugbot for commit d1e89e6. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d5ee277. Configure here.}