ncode

CLI docs Modes and run types

ncode CLI Using ncode

Modes and run types

A conversation has a mode that shapes what your next message does, and each message you send starts a run. Several runs of one conversation can work side by side; each has its own tab above the transcript.

The six modes#

ModeWhat your next message does
Buildthe default: the assistant reads, writes and runs commands to do the task
Planthe assistant uses read-only tools and answers with a step-by-step plan
Goalthe message sets a goal the conversation pursues
Ultrabig tasks run as workflows the assistant writes and runs (missions are in the ncode app)
Workflowthe message authors and launches a workflow
Consensusa second model judges the plan before anything changes

/plan and /ultra turn their mode on and off, and /consensus alone switches to Consensus. Shift+Tab in the composer cycles Ask (Read-only), Auto and Plan; see Approvals and trust. Mode in Settings → Models & effort sets Build, Plan, Consensus, Ultra or Workflow (shown as "Writing a workflow") for this conversation. Goal has no switch in the terminal: /goal <text> sets a goal and starts on it at once.

Plan#

/plan turns plan mode on or off. /plan <task> plans that task right now, as a read-only run beside whatever else is running, without changing the mode. Shift+Tab reaches Plan from Auto and leaves it for Ask.

Goal#

/goal <text> sets the conversation's goal, which every agent keeps in mind. /goal alone shows the current goal and its report; PgUp / PgDn and Home / End scroll a long one.

Ultra and Workflow#

/ultra turns Ultra on or off. In the terminal the assistant handles big tasks by writing and running workflows, and the status chip says Ultra · workflows. Missions (an orchestrator that plans a validation contract and runs isolated workers and validators) are a feature of the ncode app; the terminal does not start them yet. /create-workflow <what it should do> writes one with you, once.

Consensus#

/consensus alone switches the conversation to Consensus mode. /consensus <task> runs that one turn as a judged plan without changing the mode: a planner writes a plan, a judge model checks it against a list of checks over one or more rounds, and only then an implementer carries it out. In Settings, per conversation:

SettingDefault
Consensus checks7 of the 12 checks are on; tick them on or off in Settings
Consensus rounds2
Judge modelthe worker model
Implementer modelthe planner implements
ncode consensus transcript containing a minimal release-check plan and an approved Judge round 1, with passing checks and the judge summary in the side panel.

Swarm#

/swarm <task> starts a team: a lead agent splits the task and starts workers, and you follow them in the side panel. A swarm starts at once, beside any running turn.

Review#

/review reviews the uncommitted changes in the project and reports problems.

Compact#

/compact summarises the conversation so far and continues from the summary. /compact <focus> tells it what to keep. Sent while a turn runs, it waits until the turn ends.

Model and effort#

CommandChanges
/modelthis conversation's chat model; alone, it opens a picker with the model in use first
/model <model>set it directly
/worker_model …the model the conversation's workers use (Worker model); alone, a picker
/effortalone, opens a picker of default and the levels the chat model offers, with the one in force ticked
/effort <level>how hard the chat model thinks: low, medium, high, max; none, minimal and xhigh where the model offers them
/effort defaultclear the conversation's own level, so it follows the default again (the NCODE_EFFORT of this session, else Default effort in Settings)
/worker_effort [level]the same for the workers (Worker effort); alone, a picker with the same keys; /worker_effort default follows the default again

/model also accepts <provider_id>|<model> to name the provider as well. The /effort picker always has a default row first, then only the levels the conversation's model offers, so a level the model does not take is never shown; on default, the picker names the level in force and where it comes from (default · medium (NCODE_EFFORT)); Up and Down move a visible cursor, Enter sets the level under it and Esc leaves it as it was. /swarm_model and /swarm_effort, the names before 0.2.1, still work but are no longer listed. Typing a space after /effort lists default and the levels under the composer instead (see Slash commands), and the status line's model chip ends with the effort in force (claude-sonnet-5 · medium), which is the conversation's own level, else NCODE_EFFORT, else the default in Settings. In --plain mode a bare /effort reports the level and the choices instead of opening a picker. The Validator model and Validator effort settings choose the model that checks a mission's work in the ncode app; the terminal's Ultra does not use them. To try another model for one session without changing the conversation, start ncode --model <model> (or --model provider/model); nothing is written, and an explicit /model ends it. Providers and models covers the defaults per role.