Commands

Press / on an empty prompt, or Ctrl+P at any time.

The text field never loses focus. While a dialog is open every key is handled by the palette and marked as consumed, so there is no focus to take away and give back — which is where this kind of thing usually breaks.

Key
⇅	move one option
←→	jump to the previous/next group
PgUp PgDn	a page at a time — /engine has ~150 entries
Home End	first, last
Tab	complete the filter to the highlighted option
⏎	pick
Esc	close, reverting anything a live preview changed

Typing filters. The match is fuzzy but ranked, and the order matters: prefix first, then substring, then skipped letters. Without that hierarchy, typing mo with /model and /move in the list is a coin flip. Descriptions are searched too, at a lower weight — so looking for tema finds /theme even with the interface in English.

The list

It was a board once: every command visible at once, two rows, grouped by family. It looked good and was annoying to use, because on a board ←→ means the one next to it and ⇅ means another family — so moving between two options that read as neighbours could cost a family jump and a jump back.

It is a list now. Four things keep it from being a plain one:

• a row per option with its description; comparing two used to mean moving back and forth, because only the highlighted one showed its description;
• a group header where the group changes, which keeps the grouping the board gave;
• a window with an n/N counter, so you can tell there is more below;
• the typed filter at the top and always visible, rather than tucked in a corner and only drawn once you had typed something.

The commands

Agent

/engine	who answers — ~150 providers, grouped: connected first, then via CLI, then featured, then the rest
/connect	pick a provider and a plan, and connect it
/model	switch model without losing the conversation. Shows each model's context window and price, from models.dev
/effort	low medium high xhigh max
/clear	start over: clears the conversation and the panel

/model and /effort are launch arguments, so they cannot change in place. On the CLI engine the process relaunches on the same session id and the conversation survives — verified: give haiku a fact, relaunch as sonnet, it still remembers. On the HTTP engines the history was ours all along, so there is nothing to preserve.

Appearance

/theme	13 palettes, previewed live as you scroll, reverted if you press Esc
/effects	CRT scanlines and vignette. Off by default: it is character, not legibility
/language	interface and reply language — the language travels in the system prompt too

Packet Tracer

/topology	re-read the topology and refill the panel
/bridge	check the bridge

Both are canned prompts to the agent. Everything else you ask for in plain language, which is more natural than a command per tool — there are 61 of them.

Utility

/usage	how much of the plan is used up
/mcp	whether the Packet Tracer MCP is registered — local, does not spend a turn
/debug	engine, plan, protocol, session id, model, effort, whether a credential exists
/copy	last reply to the clipboard, via OSC 52
/export	conversation and topology to a file
/help /exit

/debug and /connect report whether there is a credential, never which one.

Adding a command

A command is data: a name, a category, and a run that receives a context.

{
  name: "app.thing",
  category: "utilidad",
  run(ctx) { ctx.decir("hello") },
}

One that needs you to choose something has no concept of arguments — its run opens another dialog. That avoids inventing a "command with parameters" type that would then need parsing, validation and completion.

CommandCtx is an interface, so commands are tested without mounting any UI.