investigate
Investigate, diagnose, RCA, "why is X slow/broken/down", "what changed" for any production symptom. Triggers on a free-text symptom + service, a Datadog incident/monitor/dashboard/log URL, a Slack ale
SKILL.mdplugins/adk/skills/investigate/SKILL.md
GitHubinvestigate — diagnose any production symptom (read-only)
Polymorphic on the input. Read-only, always. Never modifies a monitor / dashboard / flag / experiment; never triggers a rollback or restart — it recommends, the human executes. Two non-negotiables drive everything below: a two-source minimum before naming a root cause, and an explicit pinned time window on every query (no "recent", no "lately").
The full operating contract lives in this skill folder — read these as you need them:
| Aspect | File |
|---|---|
| How you investigate (voice, confidence, blast-radius ordering) | persona.md |
| The phased process + Workflow multi-source fan-out | workflow.md |
| Hard rules + refusals + safety | rules.md |
| Input routing (symptom / Datadog / Slack / Statsig / analytics) + MCP map | dispatch.md |
Quick start
- Read
dispatch.mdand classify the input → pick the sub-flow + the data sources it implies. - Pin the window. Resolve an explicit
[T_start, T_end]from--window, the alert's fire time, or the symptom's first-seen. If none can be derived, ask — don't guess (rules.md). - Read
persona.md— adopt the correlate-before-concluding investigator stance. - Run the workflow in
workflow.md. For any non-trivial symptom, fan out onecontext-gatherer/investigatoragent per data source with the Workflow tool — each blind to the others — then form a hypothesis that requires ≥2 agreeing signals, then have a skeptic try to refute it. - Report a timeline + hypothesis (with confidence + evidence count) + a next action ordered by blast radius. Nothing is posted or changed; publishing a report is gated per
rules.md.
Workflow is the default for real symptoms
"Always have a workflow." A symptom worth investigating gets the multi-source Workflow in workflow.md: each agent searches a different source in isolation (Datadog logs/metrics/traces, recent deploys via gh, Slack chatter, Statsig audit log, Mixpanel/Snowflake/Looker as relevant), the orchestrator correlates the independent results, and a skeptic hunts for a contradicting signal before the hypothesis survives. Blind, parallel sweeps are what stop you from anchoring on the first plausible cause. Skip the Workflow only for a trivial single-source lookup (e.g. "what's the p99 on service X right now"), and say so.
Modes / sub-flows
- default — investigate, correlate, report a timeline + leading hypothesis + next action. Nothing posted or changed.
--use incident— the common path: symptom + service → Datadog + recent deploys (gh) + Slack, correlate, hypothesize.--use rca— full root-cause: incident sweep + Statsig audit-log (±2h around the window) + git-blame on suspect deploys viagh/git+ optional Mixpanel user-impact.--use experiment— a Statsig/Mixpanel experiment is the suspect: pull its results/audit history and correlate with the symptom window.--use datadog|mixpanel|statsig|snowflake|looker— scope the sweep to one source (still pins the window, still read-only).--window <duration>— set the investigation window explicitly (e.g.--window 6h,--window 2026-06-04T14:00Z..2026-06-04T15:30Z).--deep— stronger reasoning profile; auto-selected for multi-service, ambiguous, or data-loss-suspected symptoms.