Executive summary

The problem

During an outage, teams build a story fast: "the 14:02 deploy broke it". Common failure modes are over-trusting deploy timing, mixing time zones, reading missing data as "all clear", and writing postmortems from memory. The cost is longer incidents, misdirected fixes and blame.

What IncidentSutra AI does

IncidentSutra AI is a command-line tool and a set of 20 agent skills. You give it exported, sanitized logs, metrics, alerts and deployment records. It returns:

  1. an auditable UTC timeline where every line points to a source file, record and SHA-256;
  2. several competing hypotheses, each with what must be true, what supports it, what contradicts it and what would disprove it;
  3. an explicit list of unknowns (missing metrics, traces, change records) and checks that were not run;
  4. a blameless postmortem draft with human-only sections left open;
  5. reports as HTML, Markdown and JSON, and an automated audit report of the tool itself.

What makes it different

Typical toolingIncidentSutra AI
Names one probable causeKeeps several unconfirmed causes and refuses to confirm any
Hides gapsReports gaps as unknown
Needs cloud and keysOffline, zero dependencies, no API key
Can act on productionCannot: no network or process execution in product code
Claims accuracyPublishes limits; benchmark labelled synthetic and self-authored

Business value

Trust summary

QuestionAnswerWhere to verify
Does it run?Yes: automated tests on Windows, Linux, macOS in CItest results
Does it leak secrets?14 synthetic secret classes masked; known gaps published06, audit A06
Can hostile logs steer it?6 injection payloads change nothingaudit A08
Can it touch production?No code path existsaudit A09
Is it independently certified?No. Self-audit only07

Status and limits

Shipped: file-based analysis, reports, verification, audit, benchmark, loopback API, skill installer. Planned: live connectors (Prometheus, Loki, Grafana, Sentry, OTel), calibrated scoring, team server, signed provenance, verified IDE activation. It is not a production incident controller and not an automatic root-cause engine.

Decision guidance

Adopt it as an evidence companion for on-call and postmortems. Do not adopt it as an autonomous responder, a compliance control or a DLP product.