Skip to main content
Incidents

IncidentSutra AI

Turn an exported incident into a UTC timeline and competing hypotheses.

Offline, read-only incident investigation CLI. npm [email protected] published 2026-11-15. The README says confidence stays undetermined and it refuses to confirm a root cause.

  • Offline
  • 20 skills
  • Node 20+
GitHub version
0.2.1
Licence
Apache-2.0
Language
JavaScript
Weekly downloads
490
Overall downloads
490
GitHub stars
4

Live npm weekly downloads, npm overall downloads, and the latest npm version, plus GitHub stars and the package.json version, fetched 2026-11-17 19:07 UTC. Cached 10 minutes. Weekly downloads cover 2026-11-10 to 2026-11-16. Overall downloads cover 2026-10-01 to 2026-11-17. npm does not publish an earlier total. Stars and downloads show public use. They are not a quality score.

Use it when

  • 1

    You have exported logs, metrics, alerts, or deployment records and want an auditable timeline.

  • 2

    You want several unconfirmed hypotheses, each with contradictions.

  • 3

    You do not want a tool that restarts, rolls back, or pages anyone.

What it does, and what it does not

Read this before the guide. A limit here is a limit of the tool.

Does

  • Parses JSON, JSONL, CSV, and OTLP log JSON. Timestamps must be RFC3339 with an explicit offset.
  • Sorts by instant and keeps nanoseconds. The README says it will not guess a time zone or treat wall-clock order as causal order.
  • Offers unconfirmed hypotheses: change, dependency, saturation, datastore, latent defect, and unknown. Each has support, contradictions, and what would disconfirm it.
  • Writes incident-report.html, .md, .json, postmortem-draft.md, hypothesis-register.md, normalized-events.jsonl, incident-timeline.json, and provenance.json.
  • 20 skills. The README says 13 are backed by CLI behaviour and 7 are guidance only. Copying files is tested. IDE auto-discovery is not.
  • The README records 192 tests passing and 0 failing on Node 24 / Windows, and a 26-case synthetic benchmark. It says those labels are author-written and are not real-world accuracy.

Does not

  • It does not say a deployment caused an outage because it happened first. It does not invent impact numbers or a root-cause score.
  • It does not restart, roll back, page, post to a status page, or touch production. Diagnostics are listed as not run.
  • Redaction is best effort. The README says it is not DLP and does not catch IPv6, personal names, street addresses, unprefixed secrets, or non-US IDs.
  • Not implemented, per the README: live connectors for Prometheus, Loki, Grafana, Sentry, and OTel Collector; LLM inference; calibrated confidence; a multi-tenant server; signed provenance.
  • Limits stated in the README: 2 MB per file, 5,000 events, 100 files. Symlinks are never followed. 0 runtime dependencies.
  • The self-audit is not a third-party certification. The README says no real production telemetry has been tested.

Claims above are from the project pages and are attributed as stated there. They are not independently verified here.

How a run moves

From the IncidentSutra AI README, npm 0.2.0, published 2026-11-15. It reads an export. It does not connect to Prometheus, Loki, or Datadog.

  1. 1. Input

    An export

    JSON, JSONL, CSV, or OTLP log JSON. Each timestamp needs RFC3339 with Z or an offset.

  2. 2. Timeline

    Sort by instant

    Nanoseconds are kept. Wall-clock order is not treated as causal order.

  3. 3. Hypotheses

    Several, unconfirmed

    Each one lists support, contradictions, and what would disconfirm it. Confidence stays undetermined.

  4. 4. Report

    Then stop

    HTML, Markdown, JSON, and a postmortem draft. Nothing is restarted or rolled back.

How to read the result

Result It means It does not mean
A hypothesis One unconfirmed explanation, with contradictions. Not a confirmed root cause, and not a probability score.
Unknown A metric, trace, or change record was missing. The README says missing data is never fabricated.
Redacted output Common secret and personal-data patterns were masked, best effort. Not DLP. Sanitize the export before you run the tool.

Wording is from the project page or README and is attributed as stated there. It is not independently verified here.

Commands and notes

  • npx [email protected] analyze --input fixtures/example-incident --out report --synthetic the README 30-second demo. The fixture path is in that repository.
  • incidentsutra verify --report report/incident-report.json
  • incidentsutra audit --out audit automated self-audit.
  • incidentsutra skills install --host claude --project . also cursor, codex, gemini, copilot, agents.
  • Exit codes in the README: 0 ok, 1 usage or runtime error, 2 verification, audit, or benchmark failed.

Guide

How to use the result. GitHub is the source repository. npm is the install.

What the report contains

From the IncidentSutra AI README, npm 0.2.0, published 2026-11-15. Offline. Node.js 20 or newer.

  • JSON, JSONL, CSV, and OTLP log JSON. Timestamps must be RFC3339 with an explicit offset. The tool will not guess a time zone.
  • A UTC timeline that keeps nanoseconds. Wall-clock order is not treated as causal order.
  • Unconfirmed hypotheses: change, dependency, saturation, datastore, latent defect, and unknown. Each has support, contradictions, and what would disconfirm it. Confidence stays undetermined.
  • Files: incident-report.html, .md, .json, postmortem-draft.md, hypothesis-register.md, normalized-events.jsonl, incident-timeline.json, and provenance.json.

What it refuses

  • It does not say a deployment caused an outage because it happened first. It does not invent impact numbers or a root-cause score.
  • It does not restart, roll back, page, or post to a status page. Diagnostics are listed as not run.
  • Redaction is best effort, not DLP. Limits: 2 MB per file, 5,000 events, 100 files. Symlinks are never followed.
  • Not implemented: live Prometheus, Loki, Grafana, Sentry, and OTel Collector connectors; LLM inference; calibrated confidence.
  • 20 skills: 13 backed by CLI behaviour and 7 guidance only. The README records 192 tests passing on Node 24 / Windows and a 26-case synthetic benchmark. It says those labels are author-written and are not real-world accuracy.

Questions

Does IncidentSutra find the root cause?

No. The README says it builds evidence and competing explanations and refuses to confirm any. Confidence stays undetermined.

Can it connect to Prometheus or Datadog?

Not in this version. The README says you export sanitized data. Live connectors are listed as not implemented.

WhatsApp