Skip to main content

Overview

Type /init in the input box of any AI SRE session to switch the agent into an operational onboarding interviewer. It walks you through building operational knowledge from scratch. It scans your Flashduty incidents and notification channels and asks you questions. It records what you describe, such as service topology, runbooks, and cluster access, into knowledge files. When needed, it helps you connect external tools (MCP). /init is the starting point for knowledge. AI SRE’s diagnostic quality depends on how much real knowledge it can read about your systems. The more complete and accurate your knowledge is, the faster and more reliably the agent pinpoints root causes. /init is the guided flow that builds that knowledge from zero — and once it’s built, every session loads it automatically.
/init asks for your item-by-item confirmation before it writes or installs anything, and never echoes credentials in plain text. See Safety and consent.

When to use /init vs. plain natural language

/init is for structured setup from scratch / systematic backfilling; small, one-off edits don’t need it.
/init and one-off natural-language edits complement each other. Use /init to build the foundation. Then maintain it during day-to-day troubleshooting by telling the agent to “record this lesson in the knowledge.” Both write to the same knowledge.

How to run

1

Type /init in a session

Type /init in the input box of any AI SRE session and send it. No parameters are required.
2

Confirm the scope

The agent first confirms the scope it writes to. If the session is bound to a team, it writes to that team’s team level, which loads only in that team’s sessions. Otherwise, it writes to the account level, which all sessions in the account can see. Once confirmed, the whole session uses that scope. To switch to a different team’s scope, reopen /init from the target team.
3

Scan and check

The agent reads your channels, incidents from the last 30 days, teams, and members, infers the integration types you use and your most frequent incident labels, and plays them back in one paragraph: “I see N channels, M incidents in the last 30 days… Is that right? What’s missing?” It waits for you to confirm or correct, and does not write any files at this point. For a brand-new account with nothing to scan, it collects everything through the interview instead.
4

Import from local AI coding tools (optional)

This step needs your explicit consent and never runs automatically. If you already have knowledge and memory in local tools such as Claude Code, Codex, Cursor, Windsurf, Copilot, or Gemini CLI, you can import it in one pass. When a self-hosted Runner on your own machine can reach those files, the agent reads their config and memory files directly. Otherwise the agent hands you two universal prompts to run locally, and you paste or upload the results back. Operational and service knowledge goes into knowledge files, personal preferences and reusable habits go into agent memory, and duplicate or generic items are dropped. Everything imported is secret-redacted before any preview or write.
5

Interview by topic and confirm each item

The agent asks about the topics in the table below one at a time and turns your answers into draft knowledge files. Before writing each item, it lists the files it will create or update (each with a 3–5 line summary). It writes only after you confirm, then links the new files into the DUTY.md table of contents. After each item it asks whether to continue to the next or stop here.
The agent fills datasource gaps through Monitors first. If the account already uses Monitors (it has datasources or alert rules), the agent gives you a link to the Monitors add-datasource form. The form opens with the right type preselected (Prometheus, Loki, VictoriaLogs, Tencent Cloud CLS, SLS, MySQL, PostgreSQL, Oracle, ClickHouse or Elasticsearch). Once the datasource is enabled, AI SRE can query it, with nothing to install. Only when the account does not use Monitors at all does it register the tools you already run as a direct MCP connection (see MCP). /init itself never creates or edits a Monitors datasource.
Native kubectl is the primary way to access a cluster, with k8s-MCP as a fallback (never configure both for the same cluster). This path is BYOC-only: place a read-only kubeconfig for each cluster on the Runner (<workspace>/.kube/<name>.config); clusters.md records only the path, never the token.
6

Close, pause, or rerun anytime

When you say you’re done, the agent gives a short summary of the files it created or updated and the MCP servers it registered. You can say “skip this,” “go back to the previous item,” or “stop here” at any time. Typing /init again at any later point continues backfilling from the existing knowledge.
/init writes knowledge, may install MCP, and touches credentials, so consent and least privilege are hard constraints:
  • Per-item explicit consent: any write or install requires your explicit consent. Before writing each item, the agent lists the files it will create or update, each with a 3–5 line summary. It proceeds only after a clear yes, not a vague “ok.”
  • Credentials are never echoed: when you provide a token, password, or AK/SK, the agent records it only as <recorded (length=N)> and never reprints the plaintext.
  • Least-privilege credentials: whenever you provide or generate a credential (kubeconfig, cloud AK/SK, database account, API token), the agent asks you to make it read-only / least privilege. When the boundary is machine-verifiable, it runs a read-only boundary check before recording it.
  • Minimality: the agent only writes content without which the AI would make worse decisions during an incident; nice-to-have details are left out so the knowledge doesn’t bloat.

What it produces

The result of /init is a Knowledge base that every subsequent session loads automatically:
  • DUTY.md — the table-of-contents entry point for the knowledge, holding only a one-line introduction and a list of @filename links pointing to each topic file;
  • Topic files — services.md, topology.md, observability.md, runbooks/<topic>.md, common-failures.md, clusters.md, and so on, where all the substantive content lives;
  • MCP registrations (optional) — if external tools were connected during the interview, they’re recorded in tools.md and the MCP server registration is completed.
/init’s primary output is knowledge, not skills. It does not save content as a Skill unless you explicitly ask. It also does not automatically install Agents — those resources currently need to be added manually in the console.

/init and /insight: a best-practice pair

/init and /insight are the two ends of the operational-knowledge “build → refine” loop:

/init — build the foundation

Build the knowledge from scratch. The agent captures services, topology, runbooks, and cluster access as a coherent whole in one pass, so it understands your systems from the first session.

/insight — keep refining

Review the last 30 days of sessions to surface repeatedly pasted context, missing runbooks, and wrong data sources, telling you what to add to the knowledge next.
The recommended rhythm: use /init to build a solid foundation, then run a few real investigations. Next, use /insight to review them and fold the friction it finds back into the knowledge. Rerun /init for a systematic cleanup when needed. The more complete your knowledge, the more accurate and useful AI SRE.

Manage Knowledge

Where /init’s output lands — learn the DUTY.md structure, file constraints, and how to edit and maintain it manually.

Usage Insights

Use /insight to review sessions, surface operational friction, and guide ongoing backfilling of the knowledge.

MCP (External Tools)

The external tools /init can help you connect during the interview — learn how to connect and manage MCP.