← Help overview

All help topics

Recipes

How recipes are structured, how triggers work, and how ARIA executes them.

Recipes are ARIA’s visible automation model. Legacy Skills remain only as compatibility bridges.

A recipe is a curated JSON manifest with:

  • triggers and description
  • connection references
  • ordered steps
  • optional LLM transforms
  • guardrail and confirmation behavior

Capability families

  • SSH
  • SFTP
  • SMB
  • RSS
  • Discord
  • Webhook
  • HTTP API
  • SMTP / IMAP
  • MQTT
  • LLM transform steps

How recipes relate to Agentic Actions

Recipes are explicit, reviewable workflows. Agentic Action Flow is the runtime architecture around them:

  1. enrich context
  2. build or select a bounded draft
  3. apply policy/guardrails
  4. execute or ask/block

Recipes can therefore stay controlled while still benefiting from LLM understanding.

Recurring workflows

Repeated comparable actions can be offered as an inactive recipe draft. A user can also explicitly preserve a workflow through recipe_remember. Neither path bypasses review: a recipe is stored only after a deliberate user action and remains subject to the existing guardrails and confirmations.

The former start/stop chat learning mode has been removed. Those command-like messages are now handled as ordinary chat messages.

  • eligible: at least 5 learning-evidence points.
  • Side-effect actions: at least 5 learning-evidence points for review_ready; they still remain review-only and must not bypass confirmation or policy.
  • Multi-target observations: stay context-only. Create an explicit reviewed recipe for the target set instead of promoting the observation directly.

Evidence is weighted:

  • New successful pattern: +1.0.
  • Same pattern with another user wording: +0.75.
  • Same pattern with another target/scope: +0.75.
  • Same pattern with another safe action shape: +0.5.
  • Exact repeat of the same target, action and wording: +0.25.
  • Risky action deviation: +0; kept for review only.

The UI may also show a maturity hint such as "keep observing" or "strong evidence". The actual lifecycle state is still the stored promotion_state (observed, review_ready, eligible, promoted).

Samples

Bundled samples are templates. Adjust refs, hosts, URLs, Discord targets, and guardrails for your environment.

Current sample directions:

  • read-only SSH health and disk checks
  • RSS to chat or RSS to Discord digests
  • SFTP read/config-preview examples
  • SMB read/list examples

Useful references:

Source: docs/wiki/Recipes.md