Problem

NemoClaw is starting to receive requests for concrete observability integrations, such as built-in Prometheus metrics endpoints. We do not want core to accumulate one-off exporter implementations for every monitoring backend.

Instead, NemoClaw should expose a stable observability plugin surface that lets external adapters consume core lifecycle and operation telemetry without requiring each integration to be merged into the main repo.

Desired direction

Create a generic observability plugin extension point where NemoClaw core can emit stable telemetry events and out-of-tree plugins can translate those events into backend-specific integrations.

Examples of adapters that should be possible outside core:

• Prometheus metrics endpoint/exporter
• OpenTelemetry exporter
• Datadog/New Relic/etc. adapters
• Internal CI or fleet-monitoring adapters

Initial design goals

• Keep NemoClaw core backend-agnostic.
• Provide stable event names and payload shapes for key lifecycle/operation events.
• Avoid adding exporter-specific dependencies to core.
• Support opt-in registration/configuration for observability plugins.
• Document how a plugin can subscribe to events and expose its own exporter/service.
• Preserve zero/near-zero overhead when no observability plugin is enabled.

Candidate event families

• Blueprint plan/apply execution start/finish/failure/duration
• Endpoint/API key validation start/finish/failure/duration
• Sandbox lifecycle operations start/finish/failure/duration
• Plugin registration/start/stop events
• Future policy/security decision events if needed

Relationship to PR #2494 / issue #233

PR #2494 implemented a concrete Prometheus-compatible /metrics endpoint in core. Rather than merging that specific integration, this issue tracks the architectural path that would let Prometheus be implemented as an external observability adapter.

Issue #233 captured the original built-in Prometheus proposal and is superseded by this more generic plugin-support direction.

Acceptance criteria

• A documented observability plugin API exists.
• Core emits a minimal set of stable telemetry events.
• Plugins can subscribe to those events without patching core execution paths.
• At least one example or skeleton adapter demonstrates the intended pattern.
• Documentation explains how a Prometheus adapter could be implemented out-of-tree.