Flows

A flow is a directed graph of steps. Each step runs an agent or an integration action, passes its output downstream, and can branch on conditions.

Triggers

Anatomy of a flow node

Every flow node is one of:

• Agent step — runs an agent with a templated task prompt. Variables from upstream nodes are interpolated.
• Integration action — calls a provider action directly, like gw_gmail_list_messages or slack_post_message. Skips the model and goes straight to the API.
• Conditional — branches on an expression.
• Output — final step that posts to a Slack channel, sends an email, or writes a webhook.

When you wire an integration action, set the integrationAction field to the action name as-is, for example gw_gsc_search_analytics. Do not strip or add a provider prefix; the executor resolves tools by action name directly.

Cron-triggered flows in detail

Cron flows live in crons/{id} with the schema:

interface CronJob {
  name: string;
  schedule: string;        // "0 9 * * *"
  timezone?: string;        // IANA, defaults to UTC
  enabled: boolean;
  agentId?: string;         // pin a specific agent…
  routing?: "auto";         // …or let the router pick
  task: string;             // the prompt
  outputs: OutputChannel[]; // where results go (Slack, email, webhook, SMS)
}

Cloud Scheduler does not fire on localhost. To test locally, hit the manual trigger endpoint or use the Run button in the flow builder.

Idempotency and timestamps

Cron lastRunAt is a Firestore Timestamp object at read time, not a string. Always convert with toISOOrFallback() from @/lib/firebase/timestamp — a previous bug treated it as a string, returned undefined, and made every cron fire every minute.