Ringtail
Concepts

The four layers

Get the root keys · map the actions · automate it · recover — the whole product, in order.

Ringtail is four layers. The first two exist only to make the third possible; the fourth makes sure the first three never dead-end.

1 — Get the root keys

The only place a human is genuinely needed. A setup wizard (open-urlpasteconfirm) plus local credential discovery plus the recipe fast-path. The goal is to shrink this to almost nothing: one consent per provider, ever — reused across every repo and every downstream automation.

2 — Map the actions

With a root grant in hand, the agent maps the repo-specific and cross-tool actions now possible: a Neon branch per environment, wire Infisical → Cloudflare Pages, set a Workers binding, point a domain, create the R2 bucket your code already references. The agent maps; the actions panel renders. This list is living — steer it in chat (see Cross-tool actions).

3 — Automatically make it happen

This is the point. With the root grant, everything downstream is auto: the agent orchestrates a chain of API calls and the work just happens, no human. Layers 1–2 exist only to make layer 3 possible.

The automation bias: default is auto-run; confirm is the exception, only for destructive. Safe actions (create a DB branch, set an env var, wire a binding, create a bucket) just happen. Only irreversible ones (domain transfer, NS swap, delete) hard-confirm. The cockpit should feel like watching the agent work — cells flipping green, resources appearing.

Orchestrate vs execute

The agent orchestrates (executeStep / executeAction); the daemon executes with the stored root creds and returns status, not values. Automation and "the agent never sees your secrets" coexist precisely because the agent triggers and the daemon does. See The guarantee.

4 — Recovery

Never a dead end. A wrong or expired key, an insufficient scope, or a failed action (API error, rate-limit, conflict) is a first-class state, not an exception. Ringtail catches it, explains it in plain language, and routes to the fix. The agent re-plans on failure — it authors a recovery wizard or action from the error. Every failure surfaces a cause and a next step. See Recovery.