IncidentSutra AI

Observe precisely. Reason cautiously. Recover responsibly.

An offline, read-only CLI that turns exported logs, metrics, alerts and deployments into an auditable UTC timeline, competing root-cause hypotheses with counter-evidence, and a blameless postmortem draft, as HTML, Markdown and JSON. Ships 20 portable agent skills.

Quick start

npx incidentsutra analyze --input ./export --out ./report
npx incidentsutra verify  --report ./report/incident-report.json
npx incidentsutra audit   --out ./audit

Node.js 20 or newer. No API key, no network, no dependencies. Open report/incident-report.html in a browser.

What it does and refuses to do

Auditable timeline

RFC3339 with explicit offsets only, sorted by instant, nanoseconds kept. Every fact has an evidence ID, source file, record and SHA-256.

Competing hypotheses

Change, dependency, saturation, datastore, latent defect, unknown. Each lists what must be true, what contradicts it and what would disprove it.

Unknown stays unknown

Missing metrics, traces and change records are reported as gaps, never filled in. Confidence stays undetermined.

Never touches production

No network client, no process spawn, no eval in product code. Diagnostics are listed read-only and marked not run.

Three report formats

Single-file HTML (no scripts, strict CSP, light and dark), Markdown and JSON, plus a blameless postmortem draft.

Privacy first

Best-effort masking of common secrets and personal data with counts. Published gaps. Not DLP: sanitize upstream.

Proof, reproducible

EvidenceResultReproduce
Automated tests192 pass, 0 failnpm test
Self-audit20 checks, audit reportnpx incidentsutra audit
Synthetic benchmark26 case families, 36 expectationsnpx incidentsutra benchmark
Redaction corpus14 synthetic secret classes maskedaudit A06
Prompt-injection corpus6 hostile payloads change nothingaudit A08
Tamper detection6 of 6 mutations caughtaudit A10
Dependencies0 runtime, 0 devnpm ls
CIUbuntu and Windows (Node 20, 22, 24), macOS (Node 24)Actions

The audit is an automated self-audit, not a third-party certification. Benchmark labels are synthetic and author-written, so they show internal consistency, not real-world accuracy. Read the full audit write-up, including defects found and fixed.

20 agent skills

Install project-locally for your AI coding tool; existing files are never overwritten.

npx incidentsutra skills install --host claude --project .

Claude CodeCursorCodexGemini CLIGitHub Copilot

Roles are instructions, not 20 running models: timeline builder, change detector, hypothesis tester, evidence ledger, privacy guardian, postmortem writer and more. Copying is tested; native activation in each IDE is not. See the agent catalogue.

Honest limits

Documentation