Slash commands

Inside a chat you can type /<name> to run a command instead of sending a message. There are two kinds:

Built-ins — REPL/chat controls baked into eVi (/help, /model, /context, …).
User & plugin commands — markdown prompt templates you drop in ~/.evi/commands/ (or that a plugin ships), fired as /<name> with arguments.

Type /help (or /?) at any time to list everything available in the current session; press Tab in the CLI REPL to complete command names.

Built-in commands (CLI REPL)

The full set, from evi chat / evi sessions resume:

Command	What it does
`/help`, `/?`	Show the command list (built-ins + your user/plugin commands).
`/reset`	Clear the conversation history (start fresh, same session).
`/exit`, `/quit`	Leave the REPL.
`/tools`	List the currently active tools.
`/model [id]`	Show the active model, or switch to `id` (persisted to config).
`/goal [text\	clear]`	Set / clear / show the ongoing goal (prepended to each turn).
`/plan [task]`	Run the next turn in plan-only mode (no tools); optional inline task.
`/auto [on\	off]`	Auto-approve every tool call for this session (no `on/off` = show state).
`/compact`	Summarise older history into one note to free context.
`/context`, `/ctx`	Show where the context window is being spent (per-bucket breakdown).
`/recent [n]`	List recent sessions (resume via `evi sessions resume`).
`/image <path>`	Attach an image to the next turn (vision/VLM models).
`/effort [low\	medium\	high\	max]`	Set reasoning effort for subsequent turns.
`/fast [on\	off\	<model-id>]`	Toggle fast mode (swap to a smaller model).
`/json <prompt>`	Force JSON-object output for the next turn.
`/schema <file> [prompt]`	Constrain the next turn to a JSON Schema (file path).
`/notools <prompt>`	Answer the next turn without using any tools.
`/forcetool <name> <prompt>`	Force the model to call a specific tool.
`/reload`	Re-read `config.toml` without restarting.
`/audio <path>`	Transcribe an audio file and send the text as the next turn.
`/audioraw <path> [prompt]`	Attach raw audio (omni models) / auto-transcribe otherwise.
`/speak [on\	off]`	Auto-speak assistant replies sentence-by-sentence (TTS).
`/predict <text\	file <p>\	clear>`	Set a speculative-decoding hint for the next turn.

Aliases: /? = /help, /ctx = /context, /img = /image, /quit = /exit.

Many of these have a GUI equivalent in the web/desktop app — the model picker chip (/model), the usage-chip popover (/context), the auto/plan chips, the 📎 attach button (/image, /audio), the speak toggle (/speak). See cli-parity.md for the full mapping.

Built-ins in the web / desktop chat

The web chat understands a subset of the built-ins (the ones that make sense without a terminal), plus all your user/plugin commands:

/help · /reset · /tools · /model [id] · /goal [text|clear] · /plan [task] · /auto [on|off] · /compact · /reload

The remaining built-ins (/context, /recent, /image, /effort, /fast, /json, /schema, /notools, /forcetool, /audio*, /speak, /predict, /exit) are CLI-only — in the web/desktop app their job is done by GUI controls or per-turn API parameters instead.

User-defined commands

Drop a markdown file at ~/.evi/commands/<name>.md; typing /<name> in the REPL or the web chat sends its expanded content as the next user message. This is modelled on Claude Code's custom commands (evi/commands.py).

evi command list           # your + plugin commands (what /help shows in chat)
evi command new <name>     # scaffold ~/.evi/commands/<name>.md (ns:name → subdir)

(REPL keybindings have the same treatment: evi keybindings list shows the effective bindings from ~/.evi/keybindings.toml.)

File format

---
description: Draft a conventional-commit message
argument-hint: [scope]
model: qwen2.5-coder:14b-instruct-q4_K_M
---

Write a conventional-commit message for the staged changes.
Scope: $1
Full request: $ARGUMENTS
Here is the diff:
@<(git diff --cached)

Frontmatter (optional --- block):
- description — shown in /help. If omitted, the first body line is used.
- argument-hint — documents expected args (e.g. [scope]), shown to you.
- model — a per-command model override surfaced to callers that honour it.
Arguments:
- $ARGUMENTS — everything after the command name.
- $1 … $9 — positional args (the argument string is shlex-split).
- {args} — legacy alias for $ARGUMENTS.
File references — @path/to/file inlines that file's contents (fenced), up to 16 KB, if the file exists; otherwise the token is left as-is. (Emails like a@b.com are not treated as file refs.)
Namespacing — subdirectories become : names: ~/.evi/commands/git/commit.md → /git:commit.

No shell execution. Unlike some tools, eVi does not run !bash blocks when expanding a command — auto-running shell on expansion is too sharp an edge for eVi's permission model. If you need triggering metadata or tool gating, write a skill instead.

Plugin commands

Installed plugins can ship a commands/ directory; those appear as /<plugin>:<command> alongside your own. See plugins.md.

Usage

/help                     # list built-ins + user/plugin commands
/model qwen2.5:7b         # switch model (persisted)
/context                  # see the context-window breakdown
/auto on                  # auto-approve tools for this session
/commit fix(web)          # run ~/.evi/commands/commit.md with $1="fix(web)"
/git:commit               # a namespaced command from commands/git/commit.md

Anything that isn't a built-in and isn't a known user/plugin command prints unknown command: /<name> (try /help).

Examples

Example 1 — install the bundled `/commit` command

The repo ships one example command:

mkdir -p ~/.evi/commands
cp examples/commands/commit.md ~/.evi/commands/

Now /commit in any chat runs git diff and proposes a conventional-commit message. Pass a scope: /commit web.

Example 2 — a namespaced review command

mkdir -p ~/.evi/commands/pr
cat > ~/.evi/commands/pr/describe.md <<'EOF'
---
description: Draft a PR description from the branch diff
argument-hint: [ticket]
---
Write a concise PR description for ticket $1.
Summarise the change, call out risks, and list manual test steps.
Diff:
@<(git diff main...HEAD)
EOF

Fire it with /pr:describe ABC-123.

Notes / limits

Command and skill names must match [A-Za-z0-9_-]+ per path segment (this also blocks path traversal).
The command store rescans on every call, so newly added files show up without restarting evi web.
/json and /schema are per-turn structured-output controls — see structured-and-batch.md. The web/API equivalent is the output_schema field on the chat request.
Built-ins always win over a user command of the same name, so you can't shadow /model with ~/.evi/commands/model.md.