Architecture decisions
Why stromboli is built the way it is. Each record explains one decision: its context, what we decided and the consequences (0001).
To record a new decision, copy template.md to the next number, open it as Proposed in a PR, and add it to the list below. Don’t rewrite an accepted record: supersede it with a new one.
| # | Decision | Status |
|---|---|---|
| 0001 | Record architecture decisions | Accepted |
| 0002 | Self-hosted, one instance per organisation | Accepted |
| 0003 | The service catalog is owned by stromboli | Accepted |
| 0004 | Postgres as the only storage | Accepted |
| 0005 | A single Go binary, API first with ConnectRPC | Accepted |
| 0006 | Suggest, don’t execute | Accepted |
| 0007 | Query data on demand, store only the evidence | Accepted |
| 0008 | Log backends: Loki first | Accepted |
| 0009 | Mitigate first, investigate second | Accepted |
| 0010 | Investigation flow: deterministic brief, then a single agent | Accepted |
| 0011 | MCP first, else the provider’s API | Accepted |
| 0012 | Metrics through the Prometheus HTTP API | Accepted |
| 0013 | Approved machine actions, Fly.io only | Accepted |
| 0014 | Approve actions from Slack | Accepted |
| 0015 | The CLI follows clig.dev | Accepted |
| 0016 | Any LLM provider, one model per task | Accepted |
| 0017 | Multi-role investigation | Proposed |
| 0018 | Memory notes with a lifecycle | Proposed |
| 0019 | Evaluation by replaying recorded incidents | Proposed |
| 0020 | MIT license | Accepted |
| 0021 | stromboli as an MCP server | Proposed |
| 0022 | OAuth for remote MCP servers | Proposed |
Open questions
Not decided yet; each becomes an ADR once it is.
- Outputs beyond Slack: Discord, Teams, the openstatus dashboard, openstatus incident notes and status report drafts?
- UI authentication in v1: local users only, or OIDC from the start?
- Telemetry: anonymous, opt-out usage stats to learn which connectors to build next?