Introduction to Lumis SDK
An experimental, open-source Python SDK that investigates operational incidents: deterministic checks first, an optional bounded model, every explanation tested against evidence, and a report for a person to review.
What Lumis does
When something breaks in a complex system, the alert tells you where it hurts, rarely what broke. Lumis helps answer the second question. Given an incident (which services are affected and when), it gathers evidence from your existing telemetry, tests explanations against that evidence, and returns a structured report: what is supported, what is contradicted, and what is still unknown.
It works in a fixed order. Known failure patterns you describe are checked first, with no model involved. Only if they are not enough, and only if you enable it, a tool-using model investigates within strict limits. Whatever the model proposes is then checked mechanically against facts Lumis collected itself. Lumis never acts on your systems; a person reads the report and decides.
Where the project is
| Status | |
|---|---|
| Release | lumis-sdk 0.1.0 on PyPI (5 October 2026) |
| Python | 3.11, 3.12 and 3.13 |
| License | Apache-2.0 for this SDK. The SDK is the open-source proof of concept of Lumis' investigation core; other Lumis products and services are separate and not necessarily open source. |
| Evaluation | The GridCast reference estate: 15 injected failures, one model (DeepSeek v4 pro) |
| Actions on your systems | None. Lumis reads and reports. |
What it is not
- Not a monitoring or alerting system. It reads the telemetry you already have.
- Not an automatic fixer. There is no remediation executor; suggestions are text for a person.
- Not a source of confirmed root causes. A supported explanation is supported by evidence, not proven.
- Not a hosted product. It is a library and a CLI you run yourself.
Where to go next
- Quickstart: install and run an offline investigation in a minute.
- Your first real project: one service, Prometheus and one check.
- How Lumis works: the full flow and the vocabulary, in plain language.
- Evaluation: what we measured on GridCast, including what went wrong.
For AI assistants: llms.txt is a compact index of these pages and llms-full.txt contains all of them as one Markdown file.