Working with the agent
Slash commands
Everything you can type at the prompt that is not a request, grouped by what you are trying to do.
Type / on an empty prompt to open the picker — arrow keys to navigate, type to
filter, and commands with sub-commands drill down. Every command also works as
plain typed text, so /mode plan is the same as picking it from the menu.
Staying in control
| Command | What it does |
|---|---|
/mode | Toggle plan (read-only) ↔ act (full access) — see Modes |
/rewind | Undo: pop the last auto-checkpoint — see Checkpoints |
/rollback | Reverse the recorded side-effects of this session |
/permissions | Active grants — list and revoke |
/plan | Plan branches (git worktrees): list, new, branch, switch, current, diff, apply, discard, deactivate |
Managing the conversation
| Command | What it does |
|---|---|
/context | Context budget, and when compaction will trigger |
/tokens | Token usage for this conversation |
/compact | Compress the conversation now to reclaim room |
/clear | Reset conversation history |
/save, /load | Write a session to a file, read one back |
/export | Write this conversation to a Markdown file |
/share | Upload this session to the web so you can read it anywhere |
See Context for what these numbers mean and when to act on them.
Looking at the project
| Command | What it does |
|---|---|
/map | Display the repository structure |
/diff | Show uncommitted git changes (--stat) |
/test | Run tests with auto-fix |
/code-review | Activate the code-reviewer skill |
/memory, /init | Project memory — see Memory |
Long-running and background work
| Command | What it does |
|---|---|
/background | Background agent tasks: list, status, cancel |
/monitor | Recurring monitors — prod health, poll an endpoint, watch a file: list, status, pause, resume, cancel |
A monitor runs on a fixed interval and reports each cycle inline in the chat. Two flavours: a raw shell command with a pass condition, which costs nothing per cycle, or a full agent turn, which is smarter and costs accordingly.
Extending the agent
| Command | What it does |
|---|---|
/tools | List the tools available to the agent right now |
/skills | Skills: list, show, activate, deactivate, create, edit, delete, reload, install |
/mcp | MCP servers — connect and list |
Skills are guided workflows that restrict the agent to a role — code reviewer,
test generator, security auditor. /skills install <git-url> pulls one from a
repository. The full treatment, alongside the SPEC registry and create_tool,
lives in Extending the agent.
Housekeeping
| Command | What it does |
|---|---|
/help | All commands and shortcuts |
/doctor | Check toolchain, wasm target, provider, .snaga/ permissions, spec registry |
/rebuild | Rebuild and restart from source |
/refresh | Redraw after a terminal resize glitch |
/metrics | TUI telemetry counters |
/voice | Voice input settings |
/quit, /exit | Leave (Ctrl+C twice or Ctrl+D also work) |
Two commands that are easy to confuse
/mode plan makes the agent read-only. /plan creates a git worktree so the
agent can write into an isolated branch. Opposite intentions, adjacent names —
the built-in help for each points at the other for exactly this reason.