Three product directions, each drawn across the same three screens with identical content: Sessions, Chat, Approval. The directions differ in hierarchy, container strategy, status communication and interaction character — not in colour.
List-first and container-free. Hairlines and type weight carry the hierarchy instead of cards, so the content sits directly on the surface and reads at a glance. One primary action per screen, large targets, and status expressed as a small semantic mark inside the row it belongs to. Approval arrives as a sheet over the work it interrupts, so context is never lost.
Built for supervision. A persistent state strip, a semantic rail on every row and left-aligned monospace metadata make machine, agent, branch and run state scannable in one pass. The chat is a timeline with a gutter, so tool work and model output are structurally distinct. Approval escalates to a full-screen takeover with a visible queue — the decision owns the device until it is made.
Replaces the near-black surface with warm paper, on the argument that long-running agent work is read more than it is operated. Elevation is spent once — on the session that is live — and everything else sits as a plain row. The transcript is set for reading, with tool work quieted into the margin. Approval leads with the plain-language consequence before the command, so the decision is understood before it is executed.
The moments that define cc-pocket are supervisory: knowing which machine and branch an agent is on, noticing that something is waiting, and deciding correctly in seconds. B is the only direction that treats state as first-class structure — a semantic rail, a summary strip and aligned monospace metadata — and the only one where approval takes the whole screen with its queue visible, which is the right weight for an irreversible action with a timer on it.
The cost is tone: B reads colder during ordinary use, and its density needs discipline to keep from growing. If the team weights implementation speed over supervisory clarity, A is the fallback — it ships fastest and feels closest to a native system app.