latent-sre/save-toolkit
Application-engineering and site-reliability agents and reusable skills.
Create, repair, or security-review LLM-facing prompts, agents, skills, tool descriptions, graders, bounded Loop Engineering for evaluation/verification, and agent roster/delegation graphs. Triggers: 'write me an agent/skill/prompt', 'my skill fires too often', 'the output is the wrong shape', 'is this agent safe / prompt injection'. Not for source-code dependency, knowledge, or GraphRAG graphs, implementing a graph runtime, or a team service's workflow graph (principal-engineer).
Akamai edge work in three lanes — triage (edge vs origin, Reference # error strings, cache status, WAF denials, DataStream 2), delivery config (Property Manager versions, staging-first activation, fast fallback), and mPulse RUM (network-side vs app-side slowdowns). Triggers: 'is it the CDN or the origin', 'an Akamai error page with Reference #[n].[hex].[epoch].[hex]', 'is the WAF blocking real users', 'mPulse shows slow pages'. Not for backend log queries (obs-analysis) or a firing alert (incident-investigation).
Design, build or change HTTP APIs, webhook receivers, and HTTP clients. Covers consumer contracts, integrations, failure behavior, and durable work accepted by an API. Triggers: 'add an endpoint', 'build an API', 'wrap X behind an API', 'integrate an upstream API'. Not for standalone workers, schedulers or persistence changes (software-engineer), UI work (frontend-craft), or live database operations (database-reliability).
Review, design, troubleshoot, and optimize GitHub Actions workflows: fast feedback, caching, test matrices, reusable jobs, reproducible artifacts, and secure delivery. Includes advice-only workflow reviews when the caller asks for findings without edits. Triggers: 'set up CI', 'speed up this pipeline', 'why is this workflow failing', 'harden the pipeline'. Not for application failures after deployment (pcf-ops, gcp-ops), runtime bugs (root-cause), or a live outage, including one a deploy caused (incident-investigation).
Diagnose and improve data-layer reliability: slow queries, lock contention, replication lag, connection pools, schema migrations, and recovery evidence. Triggers: 'this query is slow', 'plan this schema migration', 'the connection pool is exhausted'. Not for app-side triage (pcf-ops), burn alerts (obs-alerting), or persistence implementation (software-engineer).
Build or change web applications and SRE interfaces in React or Vue, from one page to a complete app integrated with its backend. Triggers: 'build a UI for', 'add a page/form/table/console', 'make this look less generic', 'wire the UI to the API'. Not for backend implementation (api-craft) or Grafana dashboard design and operations (grafana).
Investigate application-side GCP failures during the migration — Cloud Run services and revisions, gcloud logging reads, what-changed correlation against revision deploys, and the project-vs-platform boundary. Triggers: 'the Cloud Run service is 503ing', 'read the GCP logs', 'container failed to listen on PORT', 'roll back the Cloud Run service to the previous revision'. Related owners: stack-profile (boundary and runtime decisions), obs-analysis (log-query dialects), pcf-ops (the PCF side while both runtimes coexist).
Write, refactor, modernize, or test Go in services, CLIs, libraries, and workers. Load before editing Go, including module upgrades and concurrent code. Triggers: 'build this in Go', 'refactor this Go package', 'modernize this Go module', 'test this Go worker'. Not for runtime-only operations; use incident-investigation for incident troubleshooting. Compose with api-craft for HTTP and operator-cli for command contracts.
Design, explain, create, and edit Grafana dashboards; inspect alert state and notification paths; operate Grafana-managed alert rules and temporary silences. Triggers: 'design a dashboard', 'edit this Grafana dashboard', 'create a Grafana alert', 'silence this alert'. Alert/SLO design uses obs-alerting; application UI dashboards use frontend-craft; active diagnosis uses incident-investigation.
Use before coaching a human through an incident: interpreting evidence, choosing a next check, correcting an assessment, comparing mitigation, checking recovery, or handing over. This includes synthetic incidents and drills in the advisor role, even with supplied evidence or a required response format. Also explains operational signals outside an incident. Triggers: 'what should I check next', 'update the assessment from this evidence', 'walk me through this incident'. Not for any bounded lookup or investigation owned by sre-assistant, whether assigned directly by a human, delegated, or synthetic; that lane uses root-cause and shared references. Not for running incident command, stakeholder communications, or writing postmortems (scribe).
Design alerting that pages on symptoms — SLIs/SLOs and multi-window burn rates, Splunk saved-search alerts, Moogsoft correlation, and ThousandEyes synthetics. Triggers: 'define an SLO', 'this alert is too noisy', 'what should page', 'design a synthetic check'. Not for signal queries or interpretation (obs-analysis) or dashboards (grafana).
Query and interpret logs, metrics, and traces: signal meaning, request correlation, and bounded comparisons. Triggers: 'search the logs', 'write a metric query', 'trace this request', 'explain this telemetry'. Live incident decisions use incident-investigation; alert/SLO design uses obs-alerting, Grafana operations use grafana, and instrumentation/export uses obs-pipeline.
What ships telemetry where — instrument a service with OTel and route metrics, traces, and structured logs through Alloy/collectors to Loki, Mimir, and Tempo. Triggers: 'instrument this service', 'add telemetry', 'logs are not reaching Loki', 'configure the Alloy exporter'. Not for reading signals (obs-analysis) or Grafana datasource/dashboard configuration (grafana).
Apply repository templates and evidence rules to resolved-incident retrospective intake and postmortem drafting, runbook creation or edits, and durable service or alert knowledge. Also covers Confluence runbook imports and ownership or dependency lookup from existing cards. Triggers: 'draft the retro for the outage', 'update the runbook from this incident', 'capture durable operational lessons', 'which team owns payments and how do I page them'. Direct operational-document writing belongs to scribe. Active incidents use incident-investigation, alert design uses observability-engineer, and fleet prompt failures use agent-authoring.
Build or repair an operator-facing CLI that works interactively and in automation, including partial failures, machine output, dry runs, and interruption. Triggers: 'build an operator CLI', 'make this command safe for automation', 'fix this CLI contract'. Not for running an existing operational command (production-change-gate; during a live incident, incident-investigation), a web UI (frontend-craft), or the HTTP API behind a CLI (api-craft).
Plan human-approved VMware TAS/PCF application deploys, blue-green cutovers, scaling, and rollback verification. Triggers: 'deploy this app to PCF', 'design a blue-green deploy', 'scale this PCF app'. Not for readiness (production-change-gate) or incident mitigation advice (incident-investigation).
Investigate application-side PCF/TAS failures with cf app, events, logs, and routes, and distinguish app faults from platform-wide symptoms. Triggers: 'the app is crashing', 'why is my app 502-ing', 'exit code 137', 'X-Cf-RouterError'. Not for widespread Diego/Gorouter failures, which go to the platform team with evidence.
Gate the path from code to production with one checklist per question: is this change ready to merge, is this build ready to ship, and may this exact action run against production. Records a human decision and never executes. Triggers: 'is this ready to merge', 'is this build ready to ship', 'can I run this cf command in prod', 'authorize this production change'. Not for code review itself (reviewer) or for incident mitigation advice (incident-investigation).
Build, implement, refactor, rewrite, and modernize production-quality Python: services, CLIs, libraries, scripts, tests, packaging, and tooling. Use before writing or changing Python, choosing, upgrading, or replacing Python dependencies, restructuring or modernizing a legacy module, or setting up uv, pyproject.toml, Ruff, pytest, and type checking. Triggers: 'write this in Python', 'refactor this Python', 'modernize this module', 'replace this custom code with a library'. Not for operating a running service or changing another language.
Analyze failure propagation and design service resilience using source, configuration, operational evidence, and explicit assumptions. Triggers: 'what happens if this dependency slows down', 'find resilience weaknesses', 'design a safer recovery path'. Covers capacity under failure, partial completion, degraded behavior, recovery, and existing service readiness. Active incidents use incident-investigation; change verdicts belong to reviewer.
Use when diagnosing a bug, test failure, or unexpected behavior before permanent remediation, and especially after a fix attempt has already failed, or when guessing has started ("maybe it's X, let me try changing it"). Triggers: 'debug this failure', 'why did this test fail', 'the fix did not work'. Human incident coaching uses incident-investigation; bounded sre-assistant causal assignments use this method whether direct, delegated, or synthetic. A bounded evidence lookup does not require a full diagnosis.
What this team runs today: the current stack, which stacks it authors versus only supports, the stay-in-lane rule, and the platform boundary. Load before editing code in a support-only stack (Java/JVM), before recommending any runtime, tool, or infrastructure change, and when choosing between observability backends. Triggers: "what's our stack", "should we use X for this", "can we move this to Kubernetes / the cloud", "which backend do I query".
Analyze recurring operational work and design measurable elimination, simplification, self-service or automation improvements. Triggers: 'reduce these repeated manual interventions', 'which toil should we automate', 'is this automation worth maintaining'. Uses frequency, effort, interruption, rework and maintenance evidence. Implementation belongs to software-engineer or the application owner; active incidents use incident-investigation.