File Slash Commands

A file slash command is a reusable prompt stored as a file. Typing /name arguments in the TUI, ACP client or a chat channel sends the expanded template to the model instead of the literal text. Lemon reads the command formats of Claude Code, Codex, GitHub Copilot, Gemini CLI, omp and generic .agents harnesses in place; nothing is copied or installed.

Built-in commands win: a file named help.md never replaces the TUI's /help, ACP's /plan, /compact or /context, or a channel's own commands. An unknown /name is sent unchanged, exactly as before.

Where commands live

Project roots are relative to the session working directory (not walked up to ancestors). User roots live under your home directory. Project commands win over user commands, Lemon-native roots win over other tools within the same scope, and the first command with a given name wins in this order:

Scope	Root	Files	Name
project	.lemon/command/, .lemon/commands/	*.md	file name
project	.omp/commands/	*.md	file name
project	.agents/commands/, .agent/commands/	*.md	file name
project	.claude/commands/	**/*.md	git/commit.md → git:commit
project	.codex/prompts/, .codex/commands/	*.md	file name
project	.github/prompts/	*.prompt.md	review.prompt.md → review
project	.gemini/commands/	**/*.toml	git/fix.toml → git:fix
user	~/.lemon/agent/command/, ~/.lemon/agent/commands/	*.md	file name
user	~/.omp/agent/commands/	*.md	file name
user	~/.agents/commands/, ~/.agent/commands/	*.md	file name
user	~/.claude/commands/	**/*.md	nested → a:b
user	~/.codex/prompts/, ~/.codex/commands/	*.md	file name
user	~/.gemini/commands/	**/*.toml	nested → a:b

LEMON_AGENT_DIR relocates the ~/.lemon/agent roots. Dot-files, files over 64 KiB, names containing whitespace, and nesting deeper than four directories are ignored. At most 500 files are read per root.

Markdown commands

---
description: Commit staged work
argument-hint: [message]
---

Commit the staged changes. Message: $ARGUMENTS

Frontmatter is optional; without it the whole file is the template. description and argument-hint are shown in the TUI palette, /help and the ACP command list. model and subtask are read into the command definition but do not change how the expanded prompt runs.

Gemini TOML commands

description = "Explain code"
prompt = "Explain {{args}} step by step."

prompt is required. Gemini shell (!{…}) and file (@{…}) injections are left as literal text.

Arguments

Arguments are split on whitespace; single or double quotes group words. Explicitly empty quoted arguments ("" or '') keep their positional slot.

Placeholder	Value
$1 … $9 (any number)	that argument, or empty when missing
$ARGUMENTS, $@, (also , )	the whole argument text as typed

Substitution happens once: argument text containing $1 is not expanded again. When a template has no placeholder and arguments were given, they are appended after a blank line. An expansion larger than 256 KiB is sent as typed.

How expansion works

Expansion happens on the daemon, against the session's working directory, when a local channel or control-plane run starts, or when a local run receives steering or redirect text (CodingAgent.Commands). The model sees, and session history stores, the expanded text. Subagent, delegated agent, scheduled, automation and async completion prompts are never expanded.

Named-node steering/redirect currently carries text without the incoming submission's origin. Such text stays literal; Lemon never reads the controller's command files to expand a prompt destined for another machine.

The TUI discovers the same roots locally so /name autocompletes, appears in the palette and /help, and is sent verbatim; new command files are picked up the next time an unknown name is typed or the palette opens. Namespaced names such as /git:commit are sent as prompt text and expanded by the daemon. When the TUI is attached to a daemon with a different working directory, the daemon's files decide what expands. File commands use the normal submission mode, including the one-shot Alt+Enter override; submitting one consumes that override just like ordinary prompt text.

ACP clients receive the discovered commands in available_commands_update after the built-ins, with argument-hint as the input hint.