What should the first implementation optimize for?

Decide what the first implementation should optimize for before adding stronger channels or presence inputs.

Why now: The next refactor will either harden the unblock loop or drift into nicer artifact browsing without better interrupt-and-resume behavior.
If ignored: The app will keep mixing queue, canvas, and interrupt behavior without a clear priority.

What I’m showing you

Attention router overview

The agent should surface one bounded decision, collect one answer, and resume from durable local state.

dashboardmvpchief-of-staffqueue

Chief-of-staff loop

Route one decision at a time.

The first useful behavior is not a giant dashboard. It is a disciplined loop: package the blocker, show the right context, interrupt through the cheapest sufficient channel, collect one answer, and resume.

Review queue

Everything lands durably, even if the original agent session is already gone.

Focus view

One ask, one artifact bundle, one response. No context dumping.

Resume semantics

The answer writes to disk, lands in the event log, and unblocks the run without replaying chat.

This is the version of humanctl that can actually beat staying in chat: quieter, more durable, and easier to resume.

Other asks can wait

What do you think of this humanctl loop?

1 background watches