Planning topologies

A topology defines the order of planning calls. The first ultra argument selects crossreview, duel, or debate. The default is duel. Counts exclude shared research, retries, and later review.

Crossreview: three calls

flowchart LR
  A[Architect plan] --> B[Critic critique]
  B --> C[Architect revision]

The architect revises its own plan against the critique. It records accepted and rejected points with reasons.

Duel: five calls

flowchart TD
  R[Shared cited evidence] --> A[Architect blind plan]
  R --> B[Critic blind plan]
  A --> C[Critic critiques architect plan]
  B --> D[Architect critiques critic plan]
  A --> E[Architect synthesis]
  B --> E
  C --> E
  D --> E

The architect and critic produce plans concurrently without access to the rival plan. Both then critique the rival plan concurrently. The architect combines both plans and both critiques into a final plan.

Debate: seven calls

flowchart TD
  R[Shared cited evidence] --> A[Architect blind plan]
  R --> B[Critic blind plan]
  A --> C[Critic critiques architect plan]
  B --> D[Architect critiques critic plan]
  A --> E[Architect revises own plan]
  B --> E
  C --> E
  A --> F[Critic revises own plan]
  B --> F
  D --> F
  E --> G[Architect synthesis]
  F --> G
  C --> G
  D --> G

After the independent plans and cross-critiques, both seats revise their own plans concurrently. A seat assigns an agent and model. Each revision receives its own plan, the rival plan, and the critique it received.

The architect combines the revised plans and both critiques. This adds exactly one revision round. It does not create an unlimited planning debate.

The critic acts as a planner during blind-plan and own-plan revision calls. Every dependent step receives named prior outputs explicitly. No topology permits missing seats, stale revisions, or undeclared fallback models. A fallback supplies an alternative after failure.

Planning topology does not set review topology. Ultra review always uses two independent judges.