| name | operational-observability-review |
| description | Design operational observability with protected operator evidence and intentional audience-safe customer/public projections. Use when an end-to-end service operating and diagnostic-access contract—logs, metrics, traces, health/readiness, SLOs, alerts, and actions—is the primary artifact. Do not use for product analytics, one feature instrumentation edit, an active incident, recurring data-product quality, or agent tracing alone. |
Operational Observability Review
Produce one Operational Observability Contract that lets an authorized
operator explain service behavior, detect material degradation, diagnose causes,
act safely, and verify recovery without exposing raw internal evidence through
customer or public contracts.
Atomic boundary
Own service/user journeys to signals, telemetry semantics and correlation,
collection and processing, SLIs/SLOs, alert-to-action routing, operator views,
diagnostic access, privacy/security, retention, cardinality/cost, fault
validation, and blind spots. Keep logs, metrics, traces, profiles, synthetics,
and events together when they answer the same operating questions; do not split
them into medium-specific Skills.
Read references/operational-observability-contract.md
for signal, audience, SLO, alert, validation, and public-boundary patterns.
Workflow
- Name the exact service boundary, owners, users/tenants, critical journeys,
dependencies, deployment units, failure domains, data sensitivity, and
operator decisions. Do not begin with a list of telemetry products.
- For each material success, degradation, failure, retry, recovery, and
saturation path, state the question an operator must answer and the minimum
correlated signals required to answer it.
- Define stable semantic fields, units, clocks, resource/service identity,
version, environment, tenant/actor scope where safe, trace and request
context, error taxonomy, causality, sampling, and missing-signal behavior.
- Map service objectives to SLIs, SLOs or explicit operating thresholds,
measurement windows, exclusions, data-quality requirements, burn or breach
conditions, and action owners. Page only on conditions that require timely
action; route the rest to review, automation, or backlog.
- Design operator journeys from alert or report through correlated evidence,
hypothesis, safe action, runbook, rollout/rollback, recovery verification,
and incident handoff. A dashboard without a decision or action consumer is
not an observability contract.
- Classify every output by audience and purpose: protected operator,
privileged support/security, authorized tenant/customer, public status, or
external protocol. Raw internal telemetry remains protected. Customer or
public state is a separate allowlisted product/status/protocol projection.
- Bound collection and processing: redaction before export, access, tenancy,
residency, retention/deletion, high-cardinality controls, tail/head sampling,
ingestion failure, buffering/backpressure, cost budgets, provider failure,
and observability-of-observability.
- Prove the contract with schema/semantic checks, representative traffic,
synthetic journeys, fault injection, missing/corrupt signal cases,
alert/action tests, public negative-leakage tests, and recovery readback.
Composition
engineering-standard owns binding implementation predicates for durable
product code and should compose when code, contracts, or runtime paths change.
agent-native-standard owns model/tool/delegation execution traces and
agent-specific operator views.
incident-standard owns an active incident's severity, mitigation, recovery,
communication, and postmortem lifecycle.
product-analytics-instrumentation-review owns product behavior events and
decision metrics, not service health or operator diagnostics.
data-quality-observability-review owns dataset and pipeline reliability.
security-threat-modeling owns a material new or changed diagnostic trust
boundary or exposure threat model.
Guardrails
- Observability means authorized inspectability, not public disclosure.
- Never place raw traces, logs, topology, migration/cutover state, internal
identifiers, stack details, control knobs, secrets, or cross-tenant evidence
in a public/customer response merely to prove that the system is observable.
- Legitimate customer-visible operational facts are allowed only when they are
an intentional product, status, support, incident, audit, or protocol
contract with a named audience, authorization scope, allowlisted minimum
schema, stable semantics, and negative leakage tests.
- Prefer opaque correlation identifiers in public errors and resolve them to
protected evidence inside the operator boundary.
- Do not collect hidden chain-of-thought, unrestricted payloads, secrets, or
personal data as a debugging shortcut.
- Telemetry is evidence about system behavior; it is not payment, entitlement,
permission, workflow, or domain-state authority.
- Health/readiness, successful rollout, and customer outcome are separate
claims. One cannot silently stand in for another.
- Missing telemetry is unknown, not healthy. Instrumentation failure must be
observable without creating an infinite self-monitoring hierarchy.
Output contract
Return one Operational Observability Contract containing:
- service boundary, critical journeys, dependencies, failure domains, owners,
and operating questions;
- selected applicable signal map across logs, metrics, traces, profiles,
synthetics, events, and authoritative non-telemetry state, with a reasoned
disposition for families not needed by the actual failure model;
- semantic field/context/correlation/error/sampling contract;
- SLI/SLO/threshold, alert, action, escalation, and recovery-verification map;
- operator, support/security, tenant/customer, public-status, and protocol
audience projections with access and minimum-disclosure rules;
- collection pipeline, privacy/security, retention, residency, cardinality,
cost, capacity, and failure behavior;
- dashboards, queries, runbooks, automation, and incident handoffs; and
- validation fixtures, fault cases, negative leakage tests, blind spots,
rollout/readback, and ownership.
Complete only when material operating questions can be answered from correlated
authorized evidence, alerts lead to bounded actions, failure and recovery are
testable, and no raw internal process state is part of an unintended public or
customer contract.