Run a decision end to end

1. Frame the question

create_decision
  title="Ship v0.4 behind the launch gate, or hold for the audit?"
  question="..."
  alternatives=["Ship now behind the gate", "Hold until the security review lands"]

Write the question, not the answer. A decision titled "Ship v0.4" records that something happened; one titled with the actual choice records why.

Link it to what it touches with link_decision - the product, the epic, the release, the incident. Links are how it turns up later when someone opens any of those.

2. Let agents advise

post_advisory decision=dec_8fa2 confidence=0.72 recommendation="hold"
  evidence="Gate LAUNCH-1 has 4 of 5 conditions met; the security review is unassigned."

Several agents can advise, and they can disagree. An advisory carries its confidence and its evidence, and changes nothing on its own.

3. Decide - once

decide_decision decision=dec_8fa2 status=APPROVED chosenOption="Hold until the review lands"
  resolution="Two customers are mid-migration; a bad week costs more than a slow one."

An agent principal is refused this call regardless of scope. Only a person ends a decision, and the record says who, when, which option and why.

4. Set a review date if it is not final

Some calls are "for now". reviewAt marks one to come back to, and it surfaces in your inbox when the date arrives instead of being remembered by nobody.

Blocking work on a decision

Link with role: "blocks" and the linked task or epic is held until the decision resolves. Agents can add such a link and cannot remove one - otherwise an agent whose work was blocked could unblock itself.

Where they show up

Open decisions appear in the Inbox, in the product's context package, and on the Command Center. Nothing needs to be chased manually.