docs / the deck
the deck
what the deck is
the deck is your operational center. it sits between the roster and relay — the place where you read what's happening across your operation at a glance.
it has two layers: outcome cards at the top, and the execution view below.
outcome cards
outcome cards summarize the current workflow run. four cards, always visible:
total runs
how many agent executions have completed in this session.
interrupts
how many trust gates have fired. a high interrupt count means your agents are doing consequential work and runshift is catching it.
changes made
outputs approved and written. the delta between what agents proposed and what you let through.
total cost
sum of all execution costs in this session, to four decimal places. this is real spend — tokens billed at live model rates.
agent cards
each agent in your workflow has a card in the roster panel showing:
model
which AI model is powering the agent (Claude, GPT-4o, Gemini).
status
idle, running, gate, or done. status updates in real time via Supabase Realtime.
cost
what this agent has spent in the current session. per-agent cost lets you see which models are doing the most work.
when a gate fires for an agent, its card turns amber. when it completes, it returns to its default state.
reading the execution view
below the outcome cards is the execution view for the active workflow. this shows the current run in progress — agent output streaming live, gate state, and the sequence of decisions made.
think of it as the flight recorder for your operation. while relay shows you the conversation layer, the deck shows you the execution layer.
agent skills
native agents can be equipped with skills that extend what they can do before their model call runs. skills are set in the wizard and visible on each agent card.
web search
searches the web and injects the top results into the agent's context before the model runs. powered by Tavily. each search costs $0.004 and is tracked separately.
read url
extracts the first URL from the agent's input, fetches the page, and injects the text content into context. useful for agents that process links or monitor pages.
email me
delivers the agent's output to your inbox on gate approval. configure the delivery address and subject when building the agent. no additional gate is required — the email sends when you approve the run.
read github
pulls recent commits and changed files from a connected repository and injects the diff context before the model runs. requires a GitHub integration key in settings.
scheduling
agents can run on a schedule without being manually triggered. set a schedule in the wizard using a cron expression, or tell relay in plain language.
change rover's schedule to weekdays at 9am
scheduled runs fire through the same gate and audit system as manual runs. if a gate fires on a scheduled run, it waits for your approval the same way.
next