FAQ

Basics

What is IncidentSutra AI? An offline CLI plus 20 agent skills that turn exported incident data into an evidence-backed timeline, competing hypotheses and a postmortem draft.

Does it find the root cause? No. It lists plausible, falsifiable explanations with counter-evidence. Every status stays unconfirmed and confidence undetermined.

Does it need an API key, account or internet? No.

Why is confidence never a percentage? No calibrated scoring exists. Inventing a number would be false precision.

What file formats? JSON, JSONL, NDJSON, CSV and OTLP log JSON.

Trust

Is it audited? It has an automated self-audit (20 checks) and a reviewer-style write-up. It has not had a third-party audit.

Where is the proof? 07 audit, test results, evidence/.

Can I reproduce your numbers? Yes: npm test, npx incidentsutra audit, npx incidentsutra benchmark.

How accurate is it? Unknown on real incidents. The benchmark is synthetic and self-authored, so no accuracy figure is published.

Safety and privacy

Can it change production? No. Product code has no network client or process execution (audit A09).

Does it send my data anywhere? No.

Is redaction reliable? It is a best-effort safety net with published gaps. Sanitize upstream.

What if a log tells the AI to run commands? Log text is data. Tests with hostile payloads show decisions and statuses do not change. A separate LLM reading the report could still be swayed; the skills instruct it not to follow embedded instructions.

Is the report digest a signature? No. Anyone can recompute it. It catches accidents.

Usage

Why was my record rejected? Open ingestion-errors.json. The usual cause is a timestamp without an offset.

Why no hypothesis for my dependency outage? Classification is keyword based. Check the gaps and the unknown-cause hypothesis; open an issue with a sanitized sample.

What does --synthetic do? Adds a banner so demo reports are never mistaken for a real incident.

Which Node versions? 20, 22 and 24.

Will it work in Cursor, Codex, Gemini CLI or Copilot? Skills are copied to each tool's project folder (tested). Whether the tool discovers them depends on its version and is untested here.

Project

Licence? Apache-2.0.

Who maintains it? Vatsal Shah.

How do I report a vulnerability? GitHub private advisory; see SECURITY.md.

What is planned? Read-only Prometheus or Loki adapter, OTel ingest, calibrated scoring, accessible console, team server. See ROADMAP.md. None is built.