This file documents how the fleet-engineering reference repository operates as a governed mini-fleet.
The goal: be the canonical, copyable collection of fleet patterns, starters, and tooling — and eat our own dogfood.
| Concern | This repo |
|---|---|
| Registry | patterns/registry.yaml + per-pattern docs |
| Identity | Maintainers own patterns; contributors via PR |
| Permissions | Public read; write via PR + CODEOWNERS (future) |
| Inbox | GitHub Issues + Discussions for human gates |
| Audit | fleet-audit workflow on every push/PR |
| Economics | No unattended token spend on this reference repo |
| Sovereign control | CI can block; no auto-merge of fleet policy changes |
For any automated action in this repo, we should answer:
Which agent/workflow did it, with what authority, against what task, evidenced by what?
Example:
- Which:
audit.ymlGitHub Action - Authority: repository
GITHUB_TOKEN, read-only on PRs - Task: score fleet readiness of changed paths
- Evidence: workflow run URL + posted comment / check result
- Registry:
patterns/registry.yaml - Manifest template:
templates/AGENT-MANIFEST.yaml - State:
FLEET-STATE.md(human-maintained catalog of what runs where)
fleet-auditCLI +.github/workflows/audit.yml- Scores projects for registry, identity docs, budget, kill switch
- Template:
templates/fleet-budget.md - This reference repo does not run paid agent loops
This repo does not replace loop-engineering. Loops live inside fleets.
Fleet = registry + identity + permissions + inbox + audit + economics + kill switch
Loop = schedule + state + sub-agents + verification (inside one agent system)
When a team runs 3+ loops or 5+ agents, graduate from loop-engineering checklists to fleet-engineering checklists.
See docs/multi-fleet.md for orgs with separate team fleets (platform, GTM, support).
To pause all fleet automation for this repo:
- Disable GitHub Actions workflows in repo settings, or
- Add
FLEET_PAUSE_ALL=trueto repo variables (honored by future automation), or - Remove
FLEET.mdactive pattern section and open an issue explaining why.
- Weekly: scan
FLEET-STATE.mdfor stale agent entries - Monthly: re-run
fleet-audit .on starters and update registry - Per release: verify accountability test still passes for all workflows