0006. Suggest, don’t execute

  • Status: Accepted
  • Date: 2026-11-02

Context

An agent that can change production can make an incident worse, and teams will not grant it credentials they cannot trust. OpenAI’s SevBot and Google’s ProdAgent stay read-only by default; actions go through allow-listed tools with policy checks and an audit trail.

Decision

  • stromboli suggests and never acts on its own. Every connector uses read-only credentials, and every tool exposed to the model is read-only.
  • Mitigations are text: the exact command, how to verify it, how to undo it and its risk. A human applies them.
  • The only exception is opt-in, allow-listed machine actions that a named human approves (0013).
  • Redaction (tokens, passwords, emails) runs before anything reaches the model.
  • Every brief, tool call and model turn is stored, so an investigation can be audited and replayed.
  • stromboli does not write code or open PRs. It hands context to an engineer or a coding agent.

Consequences

  • Users can connect stromboli with the same read-only tokens they would give a new on-call engineer.
  • Time to mitigation still depends on a human applying the fix; 0013 narrows that gap without removing the human.

Edit this page on GitHub ↗