Chat

gmlx chat is the interactive terminal client for talking to a model. It loads the model itself or sends turns to a running server, saves sessions, and renders replies as markdown in a color theme. The flags are listed under gmlx chat.

Where the model runs

By default, chat loads the GGUF that you name and generates in its own process. With --server, it sends each turn to a model that the gmlx server serves, and it starts the server if the server is down. With --assistant, it works like --server and adds the assistant's tool loop and long-term memory.

Chat also uses the server without --server when your configured server is running and serves the id that you name, or when you name no model. A GGUF path always loads in the chat process, because the file on disk can be newer than what the server holds, and chat prints the served id when the server already serves that file.

Any of --base-url, --host, --port, --api-key and --no-start selects --server on its own. A flag that changes how the model loads keeps the model in the chat process, and --local forces an in-process load in every case.

In server mode, the server owns the model and its chat template. Chat refuses --adapter, --mmproj and the chat template flags. It ignores the other load flags and prints a note that names them. The model's served profile sets the sampling, and chat sends only the settings that you change. /image, /audio, /thinking-budget and Ctrl-T work only on a model loaded in the chat process.

The optional chat extra adds a completion menu, a toolbar and styled markdown. To install it, follow Optional features.

Commands

A line that starts with / runs one of these commands:

Command	Effect
/help	It lists the commands. An unknown command also lists them.
/exit, /quit, Ctrl-D	Each one quits chat. Ctrl-C at an empty prompt also quits.
/reset, /clear	Both start the conversation again. /clear also clears the screen.
/system [text\|off]	It shows or sets the system prompt. Setting it starts the conversation again.
/retry, /undo	/retry generates the last reply again, and /undo removes the last exchange.
/temp, /top-p and the other sampling commands	Each one changes the sampling of the next replies. /sampling shows the current values.
/thinking [on\|off\|adaptive\|default]	It turns a thinking model's reasoning on or off for the next turns.
/thinking-budget [N\|off]	It limits each reply's reasoning tokens.
/reasoning show\|hide\|raw	It sets how reasoning is shown.
/render rich\|lite\|plain	It sets the markdown renderer.
/theme NAME [cb]	It sets the color theme. cb adds colorblind-safe accents.
/model, /stats	/model shows the model card, and /stats shows the session totals. In server mode, /model <id> switches the served id.
/adapter [on\|off\|SCALE]	It turns the --adapter LoRA off or on, or scales it, for the next turns. With no argument, it shows the scale.
/history [on\|off\|clear]	It turns saving of prompt history on or off, or clears the history.
/save [name], /sessions, /load-session <name\|N>	These commands save, list and restore sessions.
/export [file.md]	It writes the conversation as markdown, by default to the session name in the current folder.
/load <file>	It puts a file's text in the prompt for you to edit and send.
/! <command>	It runs a shell command and attaches its output to the next message. /drop discards what is attached.
/image <file>, /audio <file>	Each one attaches a local file to the next message on a multimodal model.
/copy	It copies the last answer, without its reasoning, to the clipboard.
/memory	It shows or edits the assistant's memory in --assistant mode.

Esc or Ctrl-C during a reply cancels it and returns to the prompt. The partial reply stays in the conversation, so /retry generates it again and /undo removes it.

Editing and history

The arrow keys, Ctrl-A and Ctrl-E edit the line. The up arrow recalls earlier prompts, which chat saves across sessions in $XDG_CACHE_HOME/gmlx/chat_history, or in chat_history.ptk beside it with the chat extra. --no-history keeps a session out of the file, and /history off stops saving mid-session.

With the chat extra, a completion menu opens as you type a command, and an earlier prompt that matches your text appears in grey ahead of the cursor, where the right arrow accepts it. A toolbar at the bottom shows the sampling settings, attached items, how full the context is and the last reply's speed.

The extra also keeps the line breaks of pasted text, and Alt-Enter inserts a newline. Shift-Enter does the same when your terminal sends ESC CR for it. Without the extra, readline edits the line and completes with Tab.

Tab completes command names. After a command, it completes the argument:

After	Tab offers
/history, /reasoning, /render, /thinking	Tab offers that command's values.
/thinking-budget	Tab offers off.
/load-session	Tab offers the names of saved sessions.
/theme	Tab offers theme names, and then cb.
/load, /image, /audio, /export, /!	Tab offers file paths.
/model	In server mode, Tab offers the served ids.

Sampling at runtime

The sampling commands are /temp, /top-p, /top-k, /min-p, /max-tokens, /xtc-probability, /xtc-threshold, /repetition-penalty, /repetition-context-size, /presence-penalty and /frequency-penalty. Each one changes the next replies and has a startup flag with the same name, so the sampling that a model card recommends fits on the command line. /max-tokens 0 removes the limit on the length of a reply.

Chat starts from the model's family defaults. An intent suffix on a GGUF path, such as model.gguf@creative, starts from that intent of the family defaults instead. A configured id also takes a profile of the configuration file after the @.

Models with an MTP head use speculative decoding automatically, and --draft-gguf pairs a separate drafter. While a drafter is active, only temperature, top-p, top-k and min-p apply. Chat warns at startup and drops the penalties, logit bias and XTC settings, as well as --max-kv-size and --quantized-kv-start.

Undo, retry and sessions

/retry and /undo return the KV cache to where it was before the last turn, so the earlier conversation is not read again. They restore the system prompt and attached media too, and they work after a cancelled reply. /retry refuses while items are attached for the next message, so send them or /drop them first.

The cache cannot go back in place for a recurrent hybrid model, for a sliding window that has moved past the turn, or for turns restored from a saved session. In these cases, chat prints that it rebuilt the cache, and the earlier conversation is read again with your next message.

Chat saves the session as JSON in $XDG_DATA_HOME/gmlx/chats after each turn, and --no-autosave turns that off. /reset starts a new file, so the old conversation stays saved.

A saved session restores its settings and its conversation together. --resume at startup restores the model's latest session, or a named one, and /load-session does the same from inside a chat. Both refuse a session that was recorded with another model. The restored conversation is read into the KV cache with your next message, not at load time.

In server mode, /model lists the served ids, and /model <id> sends the next turns to another id. The conversation is kept, and the server reads it again under the new id. A base model and its adapters share one loaded model, so switching between them is fast. LoRA adapters uses this to compare adapters in one conversation.

Shell output and media

/! <command> runs a shell command and attaches its output to your next message as a fenced block, with the command above it and the exit status below it. Your question and the output then arrive in one turn. You can attach several blocks, and the prompt shows (+n) >> while any are waiting. Enter on an empty prompt sends the blocks alone.

Output longer than about 16,000 characters is cut in the middle. The command gets no input, so an interactive program cannot leave the chat waiting, and Ctrl-C stops the command instead of the chat.

With --mmproj, /image and /audio attach media the same way, and so does dragging a file from Finder into the terminal. Each attachment stays with the turn that sent it, so a later question can refer to an earlier image. The supported models and file types are listed in Vision and audio.

Reasoning and rendering

For a thinking model, chat shows the reasoning in a frame in the theme's thinking style. The frame closes with a line that says how long the model thought and how many tokens it used. Ctrl-O folds or unfolds the frame during a reply, and Ctrl-T ends the reasoning so that the answer starts.

--reasoning hide shows a spinner while the model thinks, and then the same closing line. --reasoning raw prints the output unchanged, which helps when chat does not recognize a model's reasoning markers. The conversation keeps the full output in every mode, so the display never changes what the model reads in the next turn.

Replies render as markdown while they stream. Finished blocks stay in the terminal's scrollback, and only the block in progress is drawn again. Chat chooses the renderer, and /render or --render overrides it:

Mode	Chosen when	What it shows
rich	The terminal has color and the chat extra is installed.	It shows tables and code blocks with syntax colors.
lite	The terminal has color and the extra is missing.	It shows markdown styles and needs no extra packages.
plain	Output is not a terminal, or NO_COLOR is set.	It shows the raw text.

Themes

--theme and /theme choose from the built-in themes dark, light, dark-hc, nord, dracula, solarized-dark and gruvbox. The default is dark, which uses your terminal's colors. The cb modifier, or --colorblind, changes the accent colors to the colorblind-safe Okabe-Ito palette and works with every theme.

The theme key of the configuration file sets the theme that each chat starts with, and themes defines your own:

theme: my-black

themes:
  my-black:
    extends: dark                  # Slots that you leave out come from this theme.
    thinking: {italic: true, fg16: 94}
    heading:  {bold: true, rgb: "#88c0d0"}
    stat:     {fg16: 90}
    code_theme: nord               # The pygments style of rich code blocks.

A theme with the name of a built-in theme replaces it. A theme is a set of slots, one for each kind of text, named thinking, heading, bold, italic, inline_code, code_block, code_border, bullet, blockquote, link, hr, stat, info and error. Beside its slots, a theme takes extends, code_theme, code_theme_cb for the pygments style under the colorblind modifier, and ptk_toolbar for the toolbar's prompt_toolkit style.

Each slot holds a style, which takes the booleans bold, dim, italic and underline, and the colors fg16 and rgb. rgb takes "#rrggbb" or [r, g, b]. A terminal with 256 colors or more uses it, reduced to the nearest of 256 colors when the terminal has no true color. fg16 is an ANSI color code from 30 to 37 or 90 to 97, which a 16-color terminal uses instead.

Chat prints a warning at startup for a theme that it cannot read, skips that theme and loads the others.