CLI Reference

This is the exhaustive reference for everything you can type into wow-agent. For task-oriented learning, see the Cookbook instead.

Launch-time options

uv run wow [options]

Option	Effect
(none)	interactive session, confirmations on
--auto	auto mode: plan approved automatically (dangerous/egress still confirmed)
--model <id>	start with a specific model, e.g. --model deepseek or --model ollama:llama3
--language en|zh	override UI language for this session
--safe 0	disable sandbox for this session (not recommended)

---

Slash commands

/help

List every available command with a one-line summary. Aliases: /h, ?.

/model

The model control center.

/model              interactive picker: provider → model
/model deepseek     quick switch by provider id
/model ollama:qwen2.5-coder:7b   provider:model syntax
/model --list       show models for current provider
/model --fetch      refresh list from provider API
/model --info       current model, context size, capabilities

Switching models mid-task is safe — conversation context is preserved.

/status

The dashboard. Shows, at a glance:

Session ─────────────────────────────
 Context    38,412 / 128,000 (30%)
 Task       fix-auth (EXECUTING, 3/5 subtasks)
 Undoable   4 snapshots
 Safety     sandbox ON · egress guard ON
 Model      deepseek-coder (128k ctx)

/compact [hint]

Compress conversation history to reclaim context.

/compact                    # automatic summarization
/compact keep the auth refactor details
/compact --aggressive       # maximum compression

Auto-compact also triggers at WOW_AUTO_COMPACT tokens. Hints steer what the summary preserves.

/undo

Time travel for your working tree.

/undo                # revert last change
/undo 3              # revert last 3
/undo --list         # history with snapshot ids
/undo --to=snap_4f2a # restore to a specific point
/undo --all          # revert everything this session
/undo --dry-run      # preview without touching files

Snapshots persist across restarts (see Rollback).

/review [path] [--level=…]

Read-only code review by a review sub-agent. Never modifies files.

/review                          # whole project, medium level
/review src/auth/                # scoped
/review --changed                # only uncommitted changes
/review --level=high             # 🔴 security/correctness only
/review --level=low --format=markdown --output=review.md

Levels: high 🔴 security & correctness · medium 🟡 best practices (default) · low 🟢 polish & performance.

/mem

Long-term memory across sessions.

/mem save "project:auth" "JWT lives in src/auth/, refresh tokens in Redis"
/mem use "project:auth"     # load into current context
/mem list                   # everything, newest first
/mem rm "project:old"       # delete

Keys are namespaced by convention: project:*, team:*, research:*.

/safe

/safe              # toggle master safety switch
/safe status       # per-layer status
/safe logs         # recent events: blocks, confirms, rollbacks
/safe test         # self-test all four layers

/resume

/resume            # pick a previous session interactively
/resume --latest   # most recent, no menu

/snapshot create <label>

Named checkpoint you can /undo --to later — do this before risky work.

/config [get|set|reset]

/config                     # dump effective config
/config get WOW_MAX_ITER
/config set WOW_MAX_ITER 60 # session-only override
/config reset               # back to defaults

/language [en|zh]

Switch UI language immediately. Without argument: interactive menu.

/clear and /exit

/clear wipes conversation (files untouched, asks to confirm). /exit saves session and quits. Aliases: /q, Ctrl+D.

---

Shell escape !

Anything starting with ! runs in your shell without involving the model:

!git status
!pytest -x -q
!docker compose logs -f api

• Output streams into the session (truncated for context safety)
• Exit codes are shown; the agent can see them if you ask it to look
• Shell commands still pass the safety system (sandbox/egress/dangerous guards apply)

Keyboard shortcuts

Keys	Action
Tab	fuzzy-complete slash commands
↑ / ↓	prompt history
Ctrl+C	cancel current agent operation
Ctrl+D	exit
Ctrl+L	redraw screen
Esc	dismiss menu / suggestion

Exit codes (scripting)

Code	Meaning
0	clean exit
1	provider/config error
2	safety guard blocked a required operation
130	interrupted by user (Ctrl+C)