ncode

Desktop docs Consensus

ncode Desktop The workspace

Consensus

Consensus gives a plan a second opinion before any code changes. A planner writes a plan, a judge (ideally a different model) reviews it against the checks you pick, and they take turns until the judge approves. Then the plan is implemented, or waits for you.

Starting a judged turn#

  • Pick ⚖ Consensus in the mode pill and send your request, or
  • type /consensus <task>.

The consensus settings#

In Consensus mode the composer shows a settings popover for the conversation:

SettingMeaning
Planner modelthe model that writes the plan
Judge modelthe model that reviews it; if unset, the sub-agent model
Implementer modelthe model that carries out the approved plan; by default the planner implements
Rounds1, 2 or 3 review rounds (2 by default)
Judge forthe checks below

The checks#

Seven checks are on in a new conversation; five more are available:

CheckOn by default
Avoid over-engineeringyes
Judge the plan before any changeyes
Keep changes minimalyes
Compare against the codebaseyes
Scope guardyes
Missing requirements & edge casesyes
Ask me before implementingyes
Judge the changes after implementingno
Risk & safety reviewno
Require verificationno
Propose a simpler alternativeno
No open questions leftno

With Ask me before implementing on, the run pauses after the judge approves and asks you: Implement, Plan only or Revise. With it off, the approved plan is implemented right away.

The planner model is the conversation's chat model.

How a round ends#

Each round, the judge either approves the plan or sends findings back, and the planner revises. If the rounds run out, the planner goes on with its best plan and the card says so. The agents pane shows the rounds, the verdicts and each finding.

Card layouts#

The judged turn's card in the agents pane has four layouts: Scales (the default), Rail (the checks on a rail), Spine and Scorecard. Set the default under Settings → Appearance; a conversation can switch its own from the pane.