Skip to content

Slash commands

DuckTapeKiller edited this page Sep 13, 2026 · 1 revision

A slash command at the start of a message tells Dive exactly what to do, instead of leaving the choice to the model. There are four kinds: commands that run a native tool, /db for your library, /skill:name for skills, and commands that external tools add. Pi mode has its own set.

Dive only recognises a command at the very start of the message. An unknown command is sent to the model as ordinary text.

Native tool commands

Command Runs Type after it
/wiki, /wikipedia wikipedia A title or question. Start with en:, es: or fr: to choose the language
/britannica britannica A search
/larousse larousse A search
/scholarpedia scholarpedia A search
/book, /isbn, /book_search book_search A title, author or ISBN
/deep_research deep_research A topic
/academic_search academic_search A search
/fetch_paper fetch_paper A URL or DOI
/wiktionary wiktionary A word, with an optional en:, es: or fr: prefix
/etymology, /deep_etymology deep_etymology A word, with an optional en:, es: or fr: prefix
/duckduckgo duckduckgo A search
/scrape, /web_scraper web_scraper A URL
/calc, /calculator calculator An expression
/time, /time_and_date time_and_date Nothing, or an IANA time zone such as Europe/Paris
/factcheck, /fact_check fact_check A claim, with an optional en:, es: or fr: prefix
/notes, /local_notes local_notes Nothing or read to read the open note; append and your text, or just your text, to add to it
/remember, /remember_lesson remember_lesson The lesson
/shell shell_command A shell command. It still asks for permission

Without a language prefix, the language is English. Every command except /time and /notes needs some text after it; without it, Dive replies with an error such as "/wiki requires a query."

Some examples:

/wiki es: Vaslav Nijinsky
/calc (4.5 * 12) / 3
/time Asia/Tokyo
/notes append Call the printer on Monday

What happens when you use one

  1. Dive runs the tool at once, with what you typed as its input.
  2. The result goes to the model with the instruction to answer your request from it, and to say so if the result isn't enough.
  3. In llama.cpp, LM Studio and Cloud mode, the model isn't offered any other tools on that message. In Ollama mode the other tools stay available.

A command doesn't get round the rules: if the tool is switched off in this mode the model is told so, and a tool that asks for permission still asks.

/db

/db answers from your library alone. Type your question after it:

/db What does the introduction say about memory?

For that message Dive searches the library in strict mode and offers the model no tools, so the answer rests on the passages it found. In Pi mode, if the library can't be reached, the request ends with an error rather than an answer that isn't grounded in the library.

/skill:name

/skill:name loads a skill's instructions into your message, followed by User: and your text, and then the turn runs normally with the mode's tools available:

/skill:proofreader Please check this paragraph.

Only skills switched on in the current mode are accepted; any other name ends the turn with an error. Later messages in the conversation keep the skill's instructions, while the saved conversation keeps what you actually typed. Pi mode has its own /skill:name, which works the same way for pi's skills. See Tools and skills.

External tool commands

An external tool can add its own commands through its commands map. They work like the native ones, with a few rules:

  • A command only works in modes where its tool is switched on.
  • It can't replace a native command with the same name.
  • The text you type goes to one of the tool's inputs: its only required text input if it has exactly one, otherwise its first text input, otherwise an input named input.
  • With nothing typed, the tool gets no inputs, unless the chosen input is required, in which case Dive replies with an error such as "/read requires path."

See Create an external tool.

Pi mode commands

In Pi mode, Dive handles these commands itself and sends anything else starting with / to pi as a normal message:

Command What it does
/models Lists pi's models
/model <name> Switches model
/think <level> Sets the thinking level: off, minimal, low, medium or high
/compact Compacts the session
/stats Shows session statistics
/help Lists these commands

Buttons above the message box

Settings > TOOLS > NATIVE TOOLS has an INPUT column. Tick it for a tool, and a button for that tool's command appears above the message box in that mode. Clicking it puts the command at the start of the message, replacing a command already there, and leaves the cursor at the end for you to type.

Seventeen tools can have a button: wikipedia, britannica, larousse, scholarpedia, book_search, deep_research, academic_search, fetch_paper, wiktionary, deep_etymology, duckduckgo, web_scraper, calculator, time_and_date, fact_check, local_notes and remember_lesson. A tool has to be switched on for its button to show, and the buttons never appear in Pi mode. The server keeps the same list, so a button can't be added for any other tool.

For developers

  • slash_commands.js holds the COMMANDS table, parseSlashCommand(), which matches /name at the start of a message, and buildForcedSkillToolCall(), which turns the text into each tool's arguments. pluginSlashCommandArguments() applies the external tool rules. INPUT_SKILL_NAMES is the button allowlist.
  • In routes/chat.js, the forced result is added to the conversation as [FORCED TOOL RESULT: name], and the stream announces the command with a slash_command event.
  • /skill:name is expanded by resolveSkillCommand() and expandSkillCommandsInHistory() in agent-skills.js.
  • Pi mode commands are handled by runPiLocalCommand() in assets/js/06-pi.js.

Read next

Clone this wiki locally