ncode

CLI docs Agents: panel, hints, overlay

ncode CLI Using ncode

Agents: panel, hints, overlay

A big task is split across agents: the assistant or a swarm's lead starts workers, each with its own brief, and every tool call of every agent is its own step. The side panel shows them all at once, and the agent overlay shows one of them in full.

The side panel#

Ctrl+B cycles the panel through four sizes, auto → full → compact → hidden:

SizeWhat you see
autothe default: no panel while a plain chat is the only thing running; the panel as full draws it (a one-line strip under 120 columns) as soon as two or more agents are in the visible runs, or anything needs you (an approval, a question, a worker waiting, a plan)
fullone row per agent with its name, a status line and one figure
compacta narrower panel
hiddenno panel; the transcript takes the width

Under 120 columns the panel becomes a one-line strip at the edge, or is off. /panel auto, /panel full, /panel compact and /panel hidden set the size directly; your choice is kept for the next session. A plain chat's subtitle in the panel reads chat · 2k (its tokens).

Workflow and consensus runs also draw their own sections in the panel: the phases of a workflow, the rounds of a judged plan. When the lead keeps a plan with the update_plan tool, a Plan 3/7 section sits above the agents (✓ done, ▸ in progress, · pending; at most 7 rows, then … N more), and the one-line strip shows Plan 3/7.

The full side panel during a five-agent swarm, with Verify changes waiting for approval first, followed by Ticket flow, Regression audit, Docs polish and the Lead; short model-written status lines, turn counts, and $0.05 spent are visible.
The same five-agent swarm with the compact side panel: the pending verification command above compact rows for Verify changes, Ticket flow, Regression audit, Docs polish and the Lead, with $0.05 spent and the Panel compact confirmation in the status line.

Speed and memory#

The first block of the full and compact panel is not about agents. It shows how fast the models of this conversation write and how much memory ncode takes:

text
speed                                            tok/s
● main       claude-sonnet-5        ▄▅▆▇▇▇▆▇███▇   142
○ worker     gpt-5-mini               ▁▂▂▂▃▂▂▃▃▃    38
○ validator  claude-opus-5                              −
  RAM        ▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁  312 MB
             app 271 MB · tools 41 MB         of 16 GB
  • One row per model. The dot is filled while that model is streaming and hollow while it is idle. A wide panel names the slot (main, worker, validator; a model the conversation used outside those reads other). The name is cut in the middle with … when it is long, so two similar names stay apart.
  • The speed. The number on the right is the latest output tokens per second: an estimate while the call streams, the exact figure when it ends, and − before the first measurement. The small line before it holds the last 12 measurements, newest at the right, one bar for each finished model call (a lone measurement is drawn at half height), all rows on one scale so their heights compare. Only a streaming model is bright; idle rows are dimmed.
  • Which models. A plain chat lists the main model. A consensus or swarm run adds the worker. Ultra lists main, worker and validator. A model that fills two slots is one row (main+worker). At most four rows are shown, then +N more.
  • RAM. The number is what ncode takes on your Mac: the session itself (app) plus the processes it started (tools, such as the terminal renderer and the commands agents run). The gauge is drawn against your Mac's memory (of 16 GB), or against a soft scale when the size is unknown, and it turns to the warning colour at half of the memory and the error colour at 80 %.

It updates at most once a second while a model streams and every few seconds when nothing does, and nothing is measured when no screen is watching, as in ncode -p. A pane too short to keep 12 rows for the run under the table gets one row instead (● 142 tok/s · claude-sonnet-5 RAM 312 MB), and a pane under 10 rows gets none.

When the panel is hidden, or is only a one-line strip (under 120 columns, or auto with nothing to show), the busiest model's speed and the memory move to the strip's right side or to the status line: 142 tok/s · RAM 312 MB. That is the vitals item of the status line; leave it out of Status line items in Settings → Layout & transcript to hide it.

Reading an agent row#

Rows are sorted by attention. Each row has:

  • a name the model gave the agent when it started it (for example "Audit the auth tests");
  • a status line, a short sentence written by a model about what the agent is doing right now;
  • one figure; a $ amount appears only when the agent's cost is known (a price is set for its model).

An agent stopped because it used up its turns is shown as such, not as a finished report. A worker that changed nothing says no file changes; one that changed files shows +N −M in K files, counted from its own isolated copy, never from your uncommitted changes. In the run palette a finished run is drawn with ✓, a failed one with ✕ and a stopped one with ■ instead of a full progress bar.

AI status lines#

The status lines are written by a model. Turn them off with /panel summaries off (and back on with /panel summaries on), or with AI status lines in Settings → Layout & transcript.

Hint mode#

Ctrl+F (or Ctrl+Space) puts a letter badge before every agent in the panel:

  • press an agent's letter to open its overlay;
  • press a digit to show one run, or 0 to open the runs dashboard (every run);
  • press Ctrl+F again to jump to the next approval or question, like Ctrl+N;
  • Esc cancels.

Hint letters never answer an approval. Hint mode also works over an approval card: the letter opens the overlay of the agent that asked. The letters used for badges can be changed under Hint letters in Settings → Keys & input.

Hint mode in the full swarm panel, with orange s, f, g, h and j badges for Verify changes, Ticket flow, Regression audit, Docs polish and the Lead, a 1 run badge, and the footer s-j open, 1 run, Ctrl-F again needs you, Esc.

The agent overlay#

A badge letter, or Enter on an agent row in select mode, opens that agent full screen:

  • its request, if it is waiting for you, with the approval keys;
  • its life on a time axis;
  • its brief and its findings;
  • what it did, grouped; o lists every raw operation;
  • where it sits in the run;
  • a composer that steers only this agent: a message you send here goes to it, not to the whole run.
KeyDoes
[ / ]previous / next agent of the run
Tabmove between the request, the activity and the composer
oevery raw operation
xstop this agent (asks first)
Escback to the chat, at the same scroll and with the same draft

The approval letters and o, [, ], x act only while the overlay's composer is empty. Here a is a second key for "allow once", beside y.

The Ticket flow worker overlay with its life timeline and brief, grouped activity showing three documentation-file writes and git status --short, the run's agent tree, three changed files, token and turn counts, and an empty composer that steers only Ticket flow.

Runs dashboard and run palette#

A conversation can have several runs at once: a chat turn, a swarm, a workflow, a plan. The run tabs above the transcript switch between them (Alt+1 … Alt+4); Ctrl+G opens the runs dashboard and Ctrl+R the run palette, to find and open a run. /stop stops everything running in the conversation.

What an agent can do#

What an agent can do#

Agents act only through tools. Each tool call is its own step you can watch, and file tools are confined to the project folder.

GroupToolsNotes
Readread_file, list_dir, find_files, grep, lspgrep and find_files use ripgrep when rg is installed
Writewrite_file, edit_file, edit_files, move_file, delete_filea snapshot of each file is taken before it changes; edit_files changes several files all or nothing
Runrun_commanda command still running after 10 seconds moves to the background, where it can be checked or stopped
Gitgit_status, git_diff, git_log, git_commit
Webweb_search, web_fetchsearch uses the engines you enabled, in your order
Memoryrememberadds a dated line to the project's memory file
Teamspawn_agent, message_agent, inbox, wait_for_message, agent_result, integrate_agent, start_swarmworkers, messages between agents, merging an isolated worker's changes
Askask_userthe assistant and a lead may ask you a question; workers may not
Planssubmit_plan, write_speconly in judged (consensus) runs
Checklistupdate_planthe assistant and a lead keep a live list of the task's steps; workers do not get it
Workflowsworkflow_list, workflow_smoke_check, workflow_save, workflow_run, workflow_controlonly the top-level assistant
MCPevery tool of every enabled MCP serverper-tool switches let you turn single tools off

run_command keeps up to 160,000 characters of output by default (an agent may ask for up to 512,000). Whether a write or a command needs your approval depends on the project's approval mode.

Limits#

How many agents may run at once, how deep they may start sub-agents, how many turns each may take and how long a worker may run are settings in Agents & limits (by default: 4 at once, depth 2, 60 turns, 30 minutes). Sub-agents work in isolated copies of the project by default, and their changes are merged back when the lead integrates them.