Skip to main content

​
At a glance

Config: LLM_PROVIDER plus a matching API key; integrations via opensre onboard or opensre integrations setup. See Environment variables. If something goes wrong: no integrations → thin evidence; LLM misconfigured → run fails at start; tool cap or stagnation breaker → early stop with a partial report (see stages below).

​
The six stages

1

See what's connected

OpenSRE checks which monitoring and infrastructure tools are connected (Grafana, Datadog, EKS, and others you set up). That set is what the investigation is allowed to query.
2

Decide if it's worth investigating

OpenSRE classifies the input: a real alert, or noise (greetings, “thanks,” replies in a closed thread). Real alerts continue — including informational or “all clear” notifications. Chit-chat stops here with no further work.
3

Sketch a plan

OpenSRE picks the tools most likely to explain this alert — matched by source (a Grafana alert starts with Grafana; an EKS alert starts with pod and cluster tools) and ranked by relevance — instead of calling every available tool.
4

Investigate

OpenSRE works through the plan: run a check, read the result, choose the next check. Findings stay in working memory so later steps build on earlier ones.Guardrails:
  • The same check is not run twice; a prior result is reused.
  • If the loop stops learning anything new, it exits and writes up what it has.
  • There is a hard limit on how long one investigation can run.
5

Diagnose

When there is enough evidence — or nothing useful left to check — OpenSRE produces a structured diagnosis: what happened, cause and effect, which claims have evidence vs. which are unconfirmed, remediation steps, and a confidence score.
6

Report

OpenSRE delivers the result as configured — terminal report, optional Slack or other chat message, GitLab comment, or JSON via --output. See Investigations overview for how to run one and what you get.

​
Long investigations and context

A long run can touch many tools and gather a lot of evidence. OpenSRE manages working memory so early findings are not crowded out by later ones. If the context grows past what it can usefully reason over, it trims the least useful parts first.
You do not configure tool selection, note-keeping, or these stop conditions — they run automatically.