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:
- an auditable UTC timeline where every line points to a source file, record and SHA-256;
- several competing hypotheses, each with what must be true, what supports it, what contradicts it and what would disprove it;
- an explicit list of unknowns (missing metrics, traces, change records) and checks that were not run;
- a blameless postmortem draft with human-only sections left open;
- reports as HTML, Markdown and JSON, and an automated audit report of the tool itself.
What makes it different
| Typical tooling | IncidentSutra AI |
|---|---|
| Names one probable cause | Keeps several unconfirmed causes and refuses to confirm any |
| Hides gaps | Reports gaps as unknown |
| Needs cloud and keys | Offline, zero dependencies, no API key |
| Can act on production | Cannot: no network or process execution in product code |
| Claims accuracy | Publishes limits; benchmark labelled synthetic and self-authored |
Business value
- Faster, safer triage: a defensible timeline in seconds from exports you already have.
- Lower risk: nothing leaves the machine; no production credentials are involved.
- Better learning: postmortems start from evidence, not recollection, and stay blameless.
- Auditability: report integrity digest, source re-hash (
verify --input), provenance manifest. - Cost: free, Apache-2.0, no vendor lock-in.
Trust summary
| Question | Answer | Where to verify |
|---|---|---|
| Does it run? | Yes: automated tests on Windows, Linux, macOS in CI | test results |
| Does it leak secrets? | 14 synthetic secret classes masked; known gaps published | 06, audit A06 |
| Can hostile logs steer it? | 6 injection payloads change nothing | audit A08 |
| Can it touch production? | No code path exists | audit A09 |
| Is it independently certified? | No. Self-audit only | 07 |
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.