Escalate

The PM is not the system's runtime. Every escalation costs them an interruption, so each one must be decidable in under a minute and must actually need them.

When to escalate

• Deadlock — two full convene rounds without convergence.
• Blocker — work cannot proceed and no team can unblock it.
• Un-inferable acceptance criteria — the goal does not reasonably imply what "done" means. Ask rather than guess; a guessed criterion is the system grading its own homework.
• Any agent error after the single retry. See halt.
• A recorded decision needs reversing and the council cannot agree on the replacement.

When NOT to escalate

Something is merely difficult. A team disagrees but a criterion resolves it. You want reassurance. The answer is recoverable from .agentic/DECISIONS.md, the profile, or the repo. Escalating these trains the PM to skim.

Format — one screen, hard limit

Use templates/escalation.md.

BLOCKED: <what cannot proceed, one line>

QUESTION: <the actual decision, phrased as a choice>

OPTION A — <name>
  case:  <one paragraph, the strongest version of this side>
  costs: <what it gives up>

OPTION B — <name>
  case:  <one paragraph, the strongest version of this side>
  costs: <what it gives up>

RECOMMENDATION: <which, and the criterion>

WHILE YOU DECIDE: <what is still running, what is idle, what is burning cost>

Present each option at its strongest. An escalation where one option is obviously strawmanned is a decision already made, dressed as a question — and the PM will notice.

Never name the models behind the positions. The PM decides on merit for the same reason the council drafts blind.

After the answer

Record the PM's decision as a Decision Record with the PM named as the decider. It is now a recorded decision like any other: reversing it later flags the step.