← All proposals

Context: AgentInbox

Net-new componentOpen in Storybook

proposals/AgentInbox/decisions.md73 lines · View on GitHub

Decisions: AgentInbox

Append as you go, newest last.

Started from a reference prototype, not ported from it

  • Decision: The designer's single-file reference (board of task cards, Needs you / Running / Done) is design intent only. The sketch is rebuilt on PDS parts: Panel (lane sunken, card default), Badge (lane count), SegmentedButton (Board/List), Button (Approve, Reject, Retry, Open) and the local Icon and Spinner.
  • Why: The reference uses raw hex and px, its own --aui-* variables, injected CSS and hardcoded text. Every one of those has a PDS equivalent.
  • Port impact: none yet.

Running time is a string from the consumer

  • Decision: time is formatted by the consumer, like ThinkingState; the inbox runs no clock. The reference ticked a shared one-second clock.
  • Why: The consumer owns the timer (settled for Thinking State). Keeps the first sketch small.
  • Port impact: none.

Left out of the first sketch

  • Decision: The glide animation between lanes (FLIP), the page heading and breadcrumb bar, and the highlight prop of the reference. Steps are in; sources and tool counts are optional props.
  • Why: Ideation sketch. Glide is the most code for the least value.
  • Port impact: none.

Status icons reuse Event Log Item icons

  • Decision: running is the Spinner, done is check, failed is diamondExclamation, waiting is circle. Colors come from the info, success, critical and warning status tokens.
  • Why: Same set as Event Log Item. Waiting and failed share a lane, so the shape alone does not separate them. See open questions.
  • Port impact: swap in the real Icon and Spinner from the package.

Status is a PDS StatusBadge

  • Decision: The state word in the card footer is the PDS StatusBadge with hasStatusIndicator: statusType working (spinner) for running, warning for waiting, success for done, critical for failed. The elapsed time sits after it. Badge is only used for the lane count. labelStrings.statusIndicator is the screen-reader text.
  • Why: The designer asked for the PDS status badge with the dot. A first try used the plain Badge by mistake.
  • Port impact: none; both are package components. The leading status icon in the title row now repeats the StatusBadge; decide whether to drop it.

Steps are an Event Log

  • Decision: The open card's steps are the local EventLog (hasRail) with EventLogItems, not a plain list and not a Chain of Thought. AgentTaskStep gets an optional status; without it, earlier steps are done and the last follows the task (error when failed, done otherwise, including running and waiting).
  • Why: The steps are logged actions with times, which is what Event Log Item is for, and it gives the steps a status icon. Chain of Thought would be a collapse inside the collapsing card, and is for reasoning in progress.
  • Port impact: the port needs Event Log and Event Log Item together with this component. The steps list CSS was removed.

Running cards end on a Thinking State

  • Decision: A running card's Event Log ends with the package ThinkingState (size='xs', label from labelStrings.thinking, default "Working"). All steps are then done; the spinner row is gone.
  • Why: One live marker instead of a spinner on the last step. The Thinking State is the existing agent "working" signal. It is not an Event Log Item, so it breaks the rail under the last step.
  • Port impact: none beyond the Event Log dependency. The designer chose to keep the generic "Working" label (2026-11-17); a per-task activity text was considered and not added.

Task subtitle names the feature

  • Decision: AgentTask.subtitle (optional string) shows under the title in pds-copy-2xs, secondary color (for example Autopilot, Lifecycle Intelligence). The status icon and chevron stay centered on the title's first line.
  • Why: The designer wants each task to say which feature it belongs to.
  • Port impact: none. The designer chose to keep the smallest size (pds-copy-2xs, 10px) over xs (2026-11-17); the engineer may want to check its legibility. Named subtitle as in Timeline Item; feature would be more specific. Rename if the engineer prefers.

Sample content: 12 agent scenarios

  • Decision: The stories use 12 tasks across Autopilot, Advanced Edge, Insights and Site dashboard (3 each, mixed statuses). The Autopilot ones come from the agentic-fleet-management-v1 prototype in design-prototyping-sandbox (the Ironwood Athletics composer fix, the failed run). Advanced Edge, Insights and Site dashboard are invented, based on that repo's edge and insights mocks; numbers are illustrative.
  • Why: Realistic copy shows how long titles, descriptions and steps behave, and how many cards a lane holds.
  • Port impact: none; story data only.

Task sites sit in the metadata row

  • Decision: AgentTask.sites (array of site names) shows in the same row as the subtitle, after a dot: one site shows its name ("Autopilot · ACME Ironwood Athletics"), more than one shows the count ("Advanced Edge · 3 sites"). labelStrings.sites is the word "sites". Sample titles no longer repeat the site name.
  • Why: The designer wants to see which site the agent is working on without reading the title.
  • Port impact: The count is built as ${count} ${labelStrings.sites}, which does not work for every language (plurals, word order). The port should use the toolkit's translation and plural handling. An array was chosen over a count so the open card can list the sites later; nothing lists them yet.

Open questions

  • Board or a single list sorted by urgency as the main model? (designer)
  • Should failed tasks share the Needs you lane or get their own signal? (designer)
  • Waiting and failed need clearly different icons, and the title icon may now be redundant with the Badge. (designer)
  • Approve and Reject as plain Buttons (now) or Suggestion Card actions? (designer)
  • Container breakpoint for the forced list: 640px is a guess. (designer)
  • Which screen needs it, and when? Priority is not set. (designer)