operations

Plugin operations from marketplace greenblacked-ai in greenblacked/AI. Commands read from the repo's manifest — 2 days ago.

Skills (19)

• alert-design

  Write, review or delete alerting rules so every page is user-visible, urgent and actionable — classify each signal as page, ticket or dashboard-only, alert on SLO burn rather than on the mechanism, write multiwindow multi-burn-rate Prometheus rules, give every alert an owner, a runbook link and a ti

• backup-recovery-design

  Design or review backup and disaster-recovery architecture for production services, databases, object stores and control planes. Turns business impact into workload-specific RTO, RPO and retention targets; maps recoverable state and dependencies; chooses consistent backup, replication, immutability,

• capacity-planning

  Work out whether a system will survive an expected load — a launch, a migration, a seasonal peak — and what to change if it will not: build a demand model with a peak-to-mean ratio before opening a load-test tool, find which resource saturates first and name it, size headroom from the latency tail y

• ci-triage

  Triage a red CI pipeline in a fixed order: check whether main is already broken before blaming the PR, locate the failing step from check-run annotations and --log-failed instead of whole logs, classify into exactly one of real / flake / infra / config-permission / dependency-drift, confirm with one

• cost-review

  Investigate a cloud bill that has grown, or reduce spend deliberately, without trading away reliability — attribute the spend before touching anything, read the top movers period over period rather than the top spenders, classify each finding as idle, overprovisioned, wrong purchase model, data tran

• disaster-recovery-exercise

  Rehearse restoration of a business service from recovery copies and measure whether the declared RTO, RPO and integrity criteria actually hold. Use when planning or reviewing a restore drill, proving cross-region recovery from backups, measuring the full dependency-ordered restore path, validating r

• docker-compose

  Compose several containers into one working environment for local development and CI: gate every depends_on edge on a healthcheck condition, since service_started means started, not ready; design the check and its start_period; service discovery by name, container port versus published port; named v

• event-driven-reliability

  Design or review reliable event consumers, queue workers and replayable asynchronous workflows. Use for at-least-once delivery, duplicate processing, idempotency keys, durable deduplication, retry and dead-letter policy, poison messages, replay safety, ordering, backpressure, consumer crash windows,

• game-day

  Plan and run a deliberate reliability exercise — a chaos experiment, a live failover or an unannounced on-call rehearsal — and convert what it finds into owned, dated work: a falsifiable hypothesis, measured steady state, bounded blast radius with tested abort, roles, human and system timings, and a

• incident-response

  Declare, command and mitigate a live production incident on any surface — a managed database failover, a third-party outage, a VM fleet, serverless, a CDN or DNS provider — mitigation first, diagnosis second. Assign an Incident Commander who is not a resolver, set severity once without debating it,

• instrumentation

  Add telemetry to a service so the next incident is diagnosable: start from the questions you will need answered at 3am, choose a metric, a span or a log per question, instrument with OpenTelemetry and the semantic conventions, keep high-cardinality identity off metric labels and on spans, propagate

• k8s-triage

  Triage a broken Kubernetes workload, mitigation before diagnosis — check for a deploy and roll back, scope the blast radius, capture evidence before mutating, pattern-match the symptom, then escalate: rollback, scale-out, load shedding, failover, drain, restore. Use when a pod, deployment, service o

• k8s-upgrade

  Upgrade a Kubernetes cluster one minor at a time — control plane, add-ons, node pools — after finding removed or deprecated APIs from live traffic, not git manifests: the apiserver_requested_deprecated_apis metric and the audit log, with kubent or pluto as a second opinion. Fix every PodDisruptionBu

• k8s-workloads

  Specify a Kubernetes workload so it holds under pressure: measure the real CPU and memory profile before writing any number, then set requests and limits knowing CPU limits throttle while memory limits kill, pick the QoS class on purpose, keep liveness, readiness and startup probes distinct, size th

• llm-cost

  Control what an LLM feature or agent pipeline costs without degrading it: fetch today's prices and count tokens for the exact model, not from memory, attribute token spend to a prompt path before changing anything, judge caching on reads per write, not hit rate, since a write costs a premium, route

• routeros-config

  Change, upgrade or recover a MikroTik RouterOS device without locking yourself out of it. Establish the version and a second way in, export the config as text, make the change in safe mode so a dropped session reverts it, verify with a connection opened after it, then release. Covers the changes tha

• runbook

  Write, review or repair the operational document someone follows at 3am — the thing on the other end of an alert's runbook link: what the symptom means in user terms, how to confirm it is real rather than a monitoring artefact, what to capture before mutating anything, the mitigations ordered least-

• slo-design

  Design or review service level objectives before alert rules exist: map the user journey to measurable good and valid events, choose availability, latency, freshness or correctness indicators, set a target and rolling compliance window from observed performance and user needs, calculate an error bud

• terraform-state

  Run a Terraform or OpenTofu state operation already decided: import a single resource, move or rename one after a refactor, drop one from state without destroying it, reconcile drift, recover a lost or corrupted state file, clear a stuck lock, or change the backend. Backs up first; prefers reviewabl

Slash commands (2)

• /operations:ci-fail

  Classify a failing CI run as a real failure, a flake, an infrastructure fault, a config or permission problem, or dependency drift, working from annotations rather than whole logs.

• /operations:oncall-handover

  Draft an on-call handover from what actually happened — pages, deploys, open incidents and anything left mid-flight — so the next person starts informed rather than surprised.

Subagents (7)

• ci-log-reader

  Read a failing CI run and return a classification with the log line that decided it, so the logs themselves never enter the caller's context. Use when a build, workflow, job or check is red and someone needs to know which of the five triage classes it falls into before deciding what to do about it.

• cost-analyst

  Read a cloud billing export — a Cost and Usage Report, a BigQuery billing export, a CSV of hundreds of thousands of line items — and return the top movers period over period with the cause of each, so the export never enters the caller's context. Use when a bill grew and nobody knows why, when a mon

• incident-scribe

  Turn raw triage notes, command output and chat scrollback into a blameless postmortem draft, keeping the source material out of the caller's context. Use once service is restored and an incident needs writing up, or when scattered evidence needs assembling into a timeline someone can review.

• k8s-evidence-reader

  Read supplied Kubernetes snapshots, describe output, events, object exports, rollout records, and current or previous container logs, then return a bounded evidence report without contacting the cluster. Use when a captured incident bundle is too large or mixed to inspect inline and the caller needs

• llm-trace-reader

  Read an exported LLM usage or trace file — a provider usage export, OpenTelemetry GenAI spans as JSON, a LiteLLM, Langfuse or Helicone log export — and return cost and latency hotspots by feature, prompt path, model and user, with cache reads per write, retry and error counts that keep spend-cap 429

• page-history-reader

  Read a quarter or more of pager or alert history — PagerDuty incidents and log entries, Opsgenie alerts, Grafana OnCall or Cloud IRM alert groups, a Prometheus ALERTS range query, or a plain CSV — and return per rule: fires, human actions separate from auto-resolve, off-hours share, time to acknowle

• telemetry-reader

  Read traces, spans and structured logs for a slow or failing request path and return the critical path with per-span durations, whose latency it is, and the p50/p95/p99 shape — or an explicit statement that the telemetry cannot answer the question and what to instrument next. Use when an endpoint is

Similar plugins