Repository navigation
Description
Deck should let people run repeatable work without having to supervise every agent or remember every step. A flow is a named, editable description of how work moves from intake to a checked outcome. Software development is a useful first test, not the boundary of the feature. The same mechanism should make sense for research or preparing a weekly digest.
Put the process in the flow
Use code for steps with known rules: checking outputs, running tests, waiting for a result and deciding whether a recorded check passed. Call an agent for work that needs judgement, investigation or writing. Name the role rather than tying the flow to a particular model or runtime. A flow can describe a straight sequence, parallel work or a bounded retry, but it should not depend on an agent remembering a rule from an earlier conversation. Define how to check the result before work starts. Evidence should show both that the work passes its checks and that it meets the original request.
Make it legible without a terminal
Make flows easy to configure and compose. Often an agent will prepare the definition; the person using it should still be able to see what will run, change the pattern when needed and understand where approval is required. Use plain stage names. Show each active task, the pattern it is following, its current stage and why it is waiting. Draw progress from recorded run events, not an agent's summary or a raw transcript. Start with a plain run view; a graphical editor is not a prerequisite for a readable flow.
Bring a genuine choice to the person as a decision card that names the agent and stage, explains the options and pauses the run. Use action cards for things to do that do not hold up the run, and short events for progress that needs no reply. Keep routine work low touch. More autonomy is something to earn and configure, not a default permission to deploy or carry out protected actions. Uncertainty about an earlier effect is a reason to pause, not to repeat it blindly.
Keep tickets, inputs and evidence distinct
If a task comes from an external issue tracker, that ticket remains the source of truth. Bring selected work onto Deck's board, track the run there and reconcile supported updates without duplicating them. People without an external tracker can use native board tasks. Review comments and build results are separate inputs that can inform later work. Verification asks whether this task is ready; model evals ask how reliably different models handle the same representative tasks. They can share a runner without sharing a pass condition.
Build in slices
Issue 225 tracks the runtime groundwork: issue 153 for the terminal runner and issue 154 for structured agent sessions. Prove those contracts independently with fixtures before relying on real adapters. Issue 155 owns the flow itself. Start with one manually started local flow, a bounded agent step, deterministic verification, a decision or safe stop, restart recovery and a plain run view. Tracker behaviour is exercised with fixtures; the first slice does not push changes, write to a live tracker, send or deploy. Issue 205 adds basic review and build inputs; issue 206 expands verification and issue 207 covers model evals. In every slice, ask whether someone can tell what will happen next, what already happened and whether the flow needs them.
Activity