Problem Statement

Users need a supported way to diagnose long agent response turnaround time during OpenClaw conversations running inside NemoClaw. NemoClaw now has opt-in onboard/setup profiling traces via #3769 and #4094, but that does not cover runtime conversation latency breakdowns.

OpenClaw already documents OTLP export through its diagnostics-otel plugin, so NemoClaw should make that path easy to use from the sandbox instead of reimplementing conversation tracing in NemoClaw.

Proposed Design

Add an opt-in OpenTelemetry diagnostics path for OpenClaw conversation traces in NemoClaw sandboxes:

• Bake or install clawhub:@openclaw/diagnostics-otel into the OpenClaw sandbox path.
• Provide documented NemoClaw configuration or environment variables for the OTLP/HTTP collector endpoint, including host, port, and path as needed.
• Add a NemoClaw network policy preset that permits export to the configured collector host and port with the minimum required method/path scope.
• Document a local visualization flow that starts with Jaeger because it is the shortest path to viewing spans.
• Keep Tempo/Grafana or other persistent observability stacks out of the initial scope unless team observability becomes a product requirement.
• Keep tracing disabled by default and make collector failures non-fatal for normal agent usage.

The goal is that a user can inspect spans for a conversation turn and identify where a long response spent time, similar to a Phoenix-style timeline or stack trace with time intervals.

Alternatives Considered

• Extend NemoClaw trace artifacts to cover conversation runtime events. This risks duplicating OpenClaw OTel support and would likely miss lower-level OpenClaw spans.
• Build a first-party trace visualization UI in NemoClaw. Useful later, but not necessary for the first diagnostic path if Jaeger can show the exported spans.
• Track only under Observability plugin support for external telemetry adapters #3915. That issue is broader and covers a backend-agnostic NemoClaw observability plugin API. This issue is narrower: enable OpenClaw existing OTel diagnostics inside NemoClaw sandboxes.

Related Work

• Broader NemoClaw observability plugin API: Observability plugin support for external telemetry adapters #3915
• Onboard/setup profiling traces: perf(onboard): operationalize profiling traces for latency diagnosis #3769, perf(onboard): operationalize profiling traces #4094
• OpenClaw OpenTelemetry export docs: https://docs.openclaw.ai/gateway/opentelemetry

Acceptance Criteria

• NemoClaw can enable clawhub:@openclaw/diagnostics-otel for OpenClaw in the sandbox without manual shell edits.
• Users can configure an OTLP/HTTP collector endpoint through documented NemoClaw configuration or environment variables.
• A network policy preset allows the configured collector endpoint without broadly opening outbound access.
• A conversation turn emits spans to an OTLP collector when tracing is enabled.
• Documentation shows a minimal local Jaeger flow for viewing conversation latency spans.
• Tracing remains opt-in and disabled by default.
• Failure to reach the collector does not break normal agent conversations.

Category

enhancement: feature

Checklist

• I searched existing issues and this is not a duplicate
• This is a design proposal, not a please build this request