inject-incidentslisted

Inject synthetic incidents (instantaneous spike, sustained magnitude change, data outage, value ramp) into a time-series dataset and evaluate analytics agents with `rockfish.agentfuel` — incident math and ground-truth generation run locally, no remote service. Covers the whole agentfuel toolkit; typed incident configs with boolean SQL test cases verified via DataFusion, Claude-backed natural-language prompts, variations, and answer grading, plus the pandas-based suite generators (analytics aggregation questions, scenario Q&A templates, stateful event/state analysis). Trigger on phrases like "inject an incident", "incident injection", "inject an anomaly", "add a spike to this series", "simulate an outage", "scenario injection", "evaluate my analytics agent", "generate a test suite from my data", "ground-truth test cases from an anomaly", "event/state questions from an event log", "agentfuel", or mentions of `rockfish.agentfuel` or the removed `rockfish.labs.scenarios`, which this skill replaces.
Rockfish-Data/rockfish-skills · ★ 0 · AI & Automation · score 72

Install: claude install-skill Rockfish-Data/rockfish-skills

# Inject incidents Use `rockfish.agentfuel` to inject typed, reproducible incidents into a time-series dataset and turn each incident into an evaluation suite for an analytics agent: SQL test cases with known boolean answers, verified locally, optionally phrased as natural-language questions and LLM-graded. The same package also generates deterministic question/answer suites straight from data — analytics aggregations, scenario Q&A, and event/state analysis — covered under [Test-suite generators](#test-suite-generators) below. ## When to use this skill Use when the user has (or will create) a baseline time-series dataset and wants to: - Inject an **instantaneous spike** (set the measurement to `absolute_magnitude` at one timestamp). - Inject a **sustained magnitude change** (add `delta_magnitude` across a time range). - Inject a **data outage** (hold the measurement at `absolute_magnitude` across a time range). - Inject a **value ramp** (linearly ramp the measurement across a time range). - Restrict the incident to rows matching metadata filters (e.g. only `site == "store"`). - Generate **ground-truth boolean test cases** (SQL + expected answer) from an incident and **verify them locally**. - Phrase test cases as **natural-language questions**, generate intent-preserving **variations**, and **grade** an agent's answers. - Generate an **analytics question suite** from any tabular dataset (aggregations × filters × time windows × group-bys, with reproducible ground truths).