Instrucciones de origen · Vista previa de solo lectura
name
maintain-observability
description
Maintain or extend NeMo Relay observability surfaces across ATIF and typed OpenTelemetry projections
author
NVIDIA Corporation and Affiliates
license
Apache-2.0
Maintain Observability Surfaces
Companion Guidance
Use karpathy-guidelines alongside this skill for implementation or review
work. Keep changes scoped, surface assumptions, and define focused validation
before editing.
Use this skill when changing event fields, exporter behavior, subscriber config,
or binding parity for ATIF or the full, gen_ai, and openinference
OpenTelemetry projections.
Surfaces To Keep In Sync
Core event model and emitted fields
crates/core/src/observability/atif.rs
crates/core/src/observability/otel.rs
crates/core/src/observability/openinference.rs
FFI and binding-native wrappers where the config or lifecycle is exposed
Python, Go, and Node.js config objects and subscriber/exporter methods
Observability config version 3, where one opentelemetry section contains
typed endpoints and OpenInference has no standalone public surface
Docs under docs/about-nemo-relay/concepts/subscribers.mdx and
docs/configure-plugins/observability/
Design Checklist
Is this an event-model change, exporter-config change, or lifecycle change?
Do all bindings expose the same logical knobs and semantics?
Does every OpenTelemetry endpoint require a type and nonblank destination?
Does each endpoint resolve header_env values at activation and reject
missing, blank, or duplicate headers?
Do layered ATOF sink, ATIF storage, and OpenTelemetry endpoint lists
concatenate with higher-precedence entries first?
Are OpenTelemetry and OpenInference dependencies unconditional rather
than Cargo feature-gated?
Does gen_ai avoid nemo_relay.*, project sanitized LLM instructions
and messages into the standard content attributes, and emit minimal spans
for scopes without GenAI semantics so their parentage is preserved?
Does enable_full_payloads preserve complete sanitized LLM request input
and annotations while leaving credential removal and sanitizers active?
Does Relay derive compliant trace and span IDs consistently across typed
OpenTelemetry endpoints while preserving lifecycle parentage?
Are mark events, start/end events, and orphan cases still handled correctly?
Does a sanitized tool result annotation remain opaque under
category_profile.tool_result_annotation, ATIF observation-result
extra.tool_result_annotation, and the single
nemo_relay.tool.result.annotation attribute in full and
openinference, while gen_ai omits it?
Do examples and docs use each exporter's documented flush/deregister
order before shutdown?
Are span or trajectory fields still derived from the intended event data?
Validation
Run the affected Rust crate tests plus just test-rust if event
fields changed.
Run just test-python, just test-go, and just test-node when
binding-native config or lifecycle changed.
