Skip to main content

Card review and human input

An agent has two separate ways to interrupt itself and wait for you. Both can show up at once on the same card while it sits in Review, each in its own labelled section.

What is a Card review?​

Labelled "Card review", with a "Decision needed" pill and a note clarifying what this is: "This is a card workflow decision. The agent is not waiting on a HITL question." This is the card's own gate, raised by the agent's run itself, in three flavours:

  • Approval: the agent finished a chunk of work and wants your sign-off. It shows a summary (markdown) and, if applicable, a list of file changes (path, added and removed line counts, and a kind tag: added, modified, removed, or renamed). Three actions:
    • Accept: lets the workflow continue.
    • Request changes: opens a required feedback box ("What needs to change?"); submitting sends the card back to In Progress for another attempt (this is what populates the "Request changes history" panel on the card). This is explicitly not a cancel; it just asks for rework.
    • Reject (see below): always available alongside the other two, as a terminal way out.
  • Questions (and its twin, "form"): the agent needs clarification before it can continue. Every question renders as its own field: free text, single-choice (radio buttons), or multi-choice (checkboxes), each optionally offering an "Other" choice that reveals a follow-up text box. Required questions carry a red asterisk and are validated before submitting; Submit answers sends everything back at once.
  • Confirmation: the agent is about to do something costly or hard to undo (deploy, merge, delete, spend) and wants an explicit go/no-go. If the agent declared specific options, each is its own one-click button; otherwise you get a plain Confirm / Cancel pair.

Without issue:approve, every type still shows you what is being asked, but the action buttons are replaced with: "Read-only · approval access is required to decide this review."

What is Human input?​

Labelled "Human input", also with an "Action needed" pill, directly below Card review when both are present. This is a related but separate interruption system. It is not tied to the card's own run engine the way Card review is, but to a more general project-wide human-in-the-loop interaction list that happens to be filtered down to this card. Two kinds:

  • Approval: a description and a row of buttons (their labels and styling, including a "danger" red style, come from the agent's request); clicking one submits your choice. Without issue:approve: "Read-only · approval access is required to respond."
  • Form: the same question-rendering as Card review's Questions type (text, single, multi, boolean, with "Other" support), submitted with Submit answers. Without issue:approve: "Read-only · approval access is required to answer."

Once you respond to either kind, the card immediately turns into a read-only "resolved" view showing who answered ("You," or the responder's name) and what was answered. This stays visible for the rest of that session, since the list underneath only asks for pending interactions and will not show it again after a reload.

Stop and Reject​

Independent of whichever checkpoint is showing, two direct actions are available wherever one appears (and Stop is also offered as its own banner on the card):

  • Stop: halts the active run immediately, with no agreement from the agent needed. What it says depends on what is actually running:
    • For a normal build run: "Stop this process," confirmed with a Stop process button. Any subtask that was mid-work goes back to Todo, and the card is forced into Review with a fresh Card review, where you then choose Accept (resume), Request changes, or Reject. An optional note explaining why you stopped it is recorded there.
    • For a run driven from a chat message: "Stop this chat run," confirmed with a Stop chat run button, with its own copy: "Stops this active conversation right away... The card itself does not move columns or create an approval checkpoint." Stopping a chat run just ends that conversation turn; it does not touch the card's column or raise a review.
  • Reject: the terminal option, available anywhere a checkpoint can be resolved. It cancels the active run and any pending checkpoint outright, with no rework attempt. The card moves straight to Done, unlocked, and is marked rejected (surfaced afterward as the red "Issue rejected" note on the detail page, and a "Rejected" chip on the board card). An optional reason is recorded. This cannot be undone.
On this page