Skip to content
nimlet

Commands and shortcuts

Slash commands are typed in the composer and start with /. They configure nimlet, manage sessions, and inspect state - they never become a message to the model, with two exceptions described under “Dispatch” below.

/help prints the same list inside the app, built from the same table this page follows. If the two ever disagree, /help is right. The keyboard shortcuts live on the Keyboard shortcuts page.

Getting started

Section titled “Getting started”
Command Description
/plan Investigate and plan with read-only tools
/act Enable implementation tools
/help Show this help
/login [flow] Sign in to Codex with ChatGPT
/logout Sign out of Codex
/auth Show Codex authentication status

/plan and /act switch modes and print the new one. Shift+Tab does the same thing without touching your draft. Both are covered in Plan and act mode.

/login starts the Codex browser login flow and prints the URL to open. /login device prints the device verification URL and code. Codex owns the credential storage and token refresh. /auth reads the current Codex account status, and /logout removes the Codex login. After login, /provider codex uses the local Codex App Server; /models refresh loads its model list and /model <id> selects a model for later turns.

Model

Section titled “Model”
Command Description
/model [name] Show or set the model
/models refresh Refresh cached model metadata
/thinking <level> Show or set reasoning
/provider [name] Show or set the provider
/web [on|off] Show or set hosted web search
  • /model alone prints the current id. /model <id> switches for later turns and remembers that choice per provider, so switching back and forth returns you to the model you last used. Changing it writes to your config file, and records a selection event in the session.
  • Typing /model opens alphabetized suggestions from the catalog for the active provider, including your current and default model. The catalog is filtered from two characters, up to 50 entries, and each suggestion shows the provider and context size.
  • /models refresh fetches models.dev for normal providers. For codex, it calls the local Codex App Server’s model list instead. A stale models.dev cache is also refreshed automatically at startup.
  • /thinking with no argument shows the active level, (provider default), or (unsupported by model) when the model has no reasoning control. Levels are none, minimal, low, medium, high, xhigh, max; which ones do something is up to the model.
  • /provider alone prints the active provider. /provider <name> accepts anthropic, codex, google, hyper, mistral, openai, opencode, opencodezen, or openrouter, selects that provider’s last model, and saves the provider, model, and remembered model choices.
  • /web shows whether hosted search is available. /web on enables it for providers that support it; it is only offered to the model in act mode.

Session

Section titled “Session”
Command Description
/session [rename|delete|restore] ... Show or manage sessions
/stats Show model, context, token usage, and cost
/new Start a new persistent session
/resume [query|ID] List this project’s sessions, or resume one
/fork [message] Fork from a user message and continue in a new session
/copy Copy the latest assistant response
/export [file] Export the session as standalone HTML
/name [title] Show or set the session name
/compact [instructions] Summarize older context
  • /session prints the id, name, event count, file path, workspace, and thinking setting. /session rename ID TITLE, /session delete ID (with confirmation), and /session restore ID manage files by id. Full details in Sessions.
  • /stats reports the provider, model, the latest request’s usage, the context window percentage, and cost, then the same numbers totalled across the session. It prints Usage: no responses yet before your first reply.
  • /new, /resume, and /fork all switch the active session; /copy copies the last assistant response to your clipboard (in the TUI only).
  • /export writes a standalone HTML transcript to nimlet-session-<session-id>.html in the workspace. Pass a relative or absolute path to choose the output file. The export includes messages, tool calls, tool results, thinking, and attachments that are stored in the session. Review it for sensitive data before sharing.
  • /name alone prints the title, or (unnamed).
  • /compact compacts now, and any text you add is passed to the summarizer as an instruction. See Context and compaction.

Trust

Section titled “Trust”
Command Description
/yolo [on|off] Auto-approve tools for this process
/trust [on|off] Show or set project-local resource trust
/permissions [clear] Show or clear remembered tool grants
  • /yolo and /yolo on are the same thing; /yolo off returns to prompting. The state is never written to disk.
  • /trust says No project-local resources require trust. when the project ships nothing, and otherwise reports trusted or not trusted. /trust on and /trust off take effect immediately for tools, skills, prompts, extensions, and system prompt files. Project config values are loaded at startup, so restart to apply a trust change to configuration.
  • /permissions lists grants, project first. /permissions clear removes the project ones from .nimlet/permissions.json.

Details for all three are in Security and Permissions.

UI

Section titled “UI”
Command Description
/theme [name] Show or set the UI theme
/settings Configure message delivery and other settings
  • /theme alone prints the current theme, what auto resolved to, and every available name. /theme <name> applies to the interface straight away; the transcript keeps its old colours until /new, /resume, or a restart, and nimlet tells you so.
  • /settings opens a menu. Today it holds one category, Queue, which chooses how queued messages are delivered: one-at-a-time or all, separately for steering messages and follow-ups. The choice is saved to your config.

Maintenance

Section titled “Maintenance”
Command Description
/doctor [test] Show configuration and key status; optionally test the connection
/reload Rescan tools, hooks, skills, and prompts
/quit Exit
/exit Exit
  • /doctor prints the selected provider and model, the endpoint with credentials and query parameters stripped, every config file it read (and whether each exists), the file it will write to, and the key source plus set/missing state for each provider. It never prints a key.
  • /doctor test sends one small request (before any session history exists) and reports Connection OK: <model>, or a failure pointing at the key, endpoint, model, and account.
  • /reload re-reads skills, prompts, AGENTS.md, external tools, and extensions, and keeps the current session.
  • /quit and /exit leave. Ctrl+C on an empty composer does the same.

Dispatch: which command wins

Section titled “Dispatch: which command wins”

The first word decides. Built-in names are matched exactly, so type them in lowercase as shown - /Model is not a built-in. If the name is not built in, nimlet looks in this order:

  1. extension commands (case-insensitive)
  2. prompt templates, as a bare /name
  3. skills, but only under /skill:<name>

If nothing matches, you get Unknown command '/foo'; try /help. Since skills use their own namespace, prompts/review.md and skills/review/SKILL.md coexist as /review and /skill:review. Extension commands beat prompt templates with the same name, and both lose to built-ins.

A template or skill command becomes a message to the model: /review fix the parser sends the expanded prompt as your turn. An extension command is sent to that extension instead, which may display a message, start a turn, or both.

Usage errors

Section titled “Usage errors”

Argument validation is strict, and the message tells you the shape that works:

/plan extra                  → /plan takes no arguments
/thinking loud               → Invalid thinking level 'loud' (use none|minimal|low|medium|high|xhigh|max)
/web maybe                   → Invalid /web value 'maybe' (use on|off)
/provider gemini             → Unknown provider 'gemini' (use anthropic|codex|google|hyper|mistral|openai|opencode|opencodezen|openrouter)
/model refresh               → Unknown /model option 'refresh'; did you mean /models refresh?
/models                      → Usage: /models refresh
/session delete              → Usage: /session delete ID
/fork zero                   → Usage: /fork [message]

Commands that take no arguments report /copy takes no arguments, and so on. A trailing space is not an error: /model is treated as still being typed, so the suggestion menu opens instead of a usage message appearing under your cursor.

Completion in the composer

Section titled “Completion in the composer”

Type / and the menu lists built-in commands by usage. Tab, Up, and Down move through it; Enter accepts. Commands that take no arguments submit immediately, while argument-taking commands complete with a trailing space so you can keep typing. /skill:<name> suggestions complete the same way.

Several commands have their own suggestions:

Typing Suggests
/model Your recent models, then catalog matches with provider and context size
/provider , /thinking , /web , /theme The valid names
/resume Sessions in this workspace, filtered by id, name, or first message
/session rename, /delete, /restore Matching session ids, or ids in the trash
/fork Your messages in this session, numbered

Extension commands and prompt templates appear in the same menu with their descriptions, and prompt templates whose names collide with built-ins are not offered.

Mentions and shell shortcuts

Section titled “Mentions and shell shortcuts”

Two more things the composer understands, neither of which is a command:

  • @path attaches a file or folder to your message. File contents arrive wrapped as <file path="…">…</file>, folder mentions arrive as a bounded listing (200 entries, generated folders like node_modules, dist, and nimcache skipped). Typing @ completes paths from the workspace, up to 50 results; folder suggestions keep a trailing / so you can keep typing. Images are attached as images, and pasting an image path does the same.
  • !command runs a shell command in your workspace with the shell the tools use, and sends $ command, its output, and (exit 0) to the model as your next message. !!command runs it and only prints the result in the transcript. Neither asks for approval, because you typed it.

While a turn is running, Enter queues a steering message and Alt+Enter queues a follow-up. Slash commands and shell shortcuts cannot be queued; the composer tells you slash commands cannot be queued or shell shortcuts cannot be queued instead of silently dropping them.

CLI flags

Section titled “CLI flags”

These flags apply when you start nimlet from the shell. They override config for one process only and are never persisted. Run nimlet --help for the same list.

Flag Purpose
--help, -h Print usage and exit
--version Print version and exit
--print, -p Print only the final response to stdout
--mode json One turn, versioned JSONL events on stdout
--mode rpc Long-running JSONL command protocol on stdin/stdout
--provider NAME Provider for this process
--model ID Model for this process
--thinking LEVEL Thinking level for this process
--api-key KEY In-memory API key override
--tools LIST Restrict tools (read,grep,... or none)
--no-session Skip session read/write
--session ID Resume a specific session at startup
--resume Resume the latest session for this workspace
--yolo Auto-approve all tools for this process
--fullscreen Use the alternate screen (default)
--no-fullscreen, --regular Keep normal terminal scrollback
--approve Load project customizations without the trust prompt
--no-approve Skip project customizations
--interactive, -i Keep the REPL after a CLI prompt
-- End of flags; remaining words are the prompt

A prompt after the flags runs one turn and exits unless --interactive is set. Piped stdin is merged before the CLI prompt and selects print mode when stdout is not a TTY.

--no-session cannot be combined with --resume or --session. --approve and --no-approve cannot be combined.

Where to go next

Section titled “Where to go next”