← Overview · Index · Next: Principles
System map
At product level, Ratchet is four cooperating ideas:
flowchart TB
subgraph human [Human intent]
UI[Composer · goal capture]
Q[Mission queue]
end
subgraph factory [Build-and-verify loop]
H[Build → deploy gate → live test]
end
subgraph outside [Outside the loop]
GH[Product git remote]
LIVE[Live app]
V[Credentials boundary]
end
UI --> Q --> H
V -.->|brokered access only| H
H -->|commit + push| GH
GH -->|host deploy| LIVE
H -->|check deployed version| LIVEASCII fallback:
Human goal → Composer → Queue → Build-and-verify loop
├→ push product repo
├→ wait for live version signal
└→ test the live app onlyGallery: diagrams.md
Trust boundaries (product ideas)
| Zone | Who | Product rule |
|---|---|---|
| Human + Composer | Operator | Captures goals, owns the queue of missions |
| Builder workspace | Coding agent | May edit the product repo; never holds cloud secrets |
| Tester workspace | Tester agent | Judges the live app only — not the builder’s local tree |
| Credentials boundary | Harness-side broker | Short-lived, named actions — tokens never enter agent env |
| Product live | Public users | Exposes an honest version signal the loop can wait on |
Secrets stay out of agent prompts and builder/tester environments. That boundary is the architecture; how any one install implements it is private.
End-to-end product flow
flowchart LR
subgraph intent [1 Intent]
UI[Goal capture]
Q[Queue]
end
subgraph work [2 Work]
BL[Build]
DG[Deploy gate]
TE[Live test]
end
subgraph done [3 Finish]
EX[Streak of passes]
end
UI --> Q --> BL --> DG --> TE
TE -->|FAIL| BL
TE -->|PASS streak| EX1. Intent
A human describes product work. The control plane turns that into one or more focused missions scoped to a product — not one mega-mission for a multi-part goal.
2. Build → deploy gate → test
- Build — a coding agent changes the product repo and pushes the deploy branch.
- Deploy gate — the loop waits until the live app reports the same version the builder just pushed.
- Test — a tester agent exercises the live site and returns pass or fail with actionable feedback.
3. Finish
The run ends only after a streak of consecutive passes. A single fail resets the streak. Durable notes stay with the campaign so the next iteration knows what still fails.
Pluggable roles (concept)
The loop has three replaceable roles:
| Role | Purpose |
|---|---|
| Builder | Change product code and prove real git work happened |
| Deploy gate | Wait until live version matches what was pushed |
| Tester | Grade the live URL only; emit structured pass/fail |
Early development can simulate any role; production-minded installs use real agents against a real live URL. Mixing simulated and real roles is a rollout technique, not a host recipe.
Where product state conceptually lives
No install paths here — only the ideas:
| State | Product idea |
|---|---|
| Queue | Ordered missions per product |
| Run workspace | One isolated workspace per mission attempt |
| Product shell | Repo + live URL + version URL bound together |
| Credentials store | Encrypted secrets outside agent workspaces |
Continue → Principles