Interactive REPL

Running the binary on a terminal starts a persistent conversation. One process serves every turn, so the prompt cache stays warm and a turn boundary costs milliseconds.

oasis-agent

If nothing is configured yet, the first interactive run walks you through setup.

The editor

Action	Keys
Newline without submitting	\ + ↵, or Alt + ↵
Complete the command being typed	Tab (↑/↓ to choose, ↵ accepts)
Cycle the permission mode	Shift + Tab
Clear the input, or cancel a running turn	Ctrl + C
Leave on an empty prompt	Ctrl + D
Release a foreground job mid-turn	Ctrl + B

Word motions and persistent history work as they do in any readline-style editor. Disable history with --no-history or OASIS_AGENT_NO_HISTORY=1.

Slash commands

Command	What it does
/help	Show the command help
/expand [id]	Print a collapsed tool result in full (default: the last one)
/cost	Token and cost totals for this session
/plan	List the model's current plan items
/plan on | off | status	Toggle or report read-only plan mode
/jobs	Background jobs. ↑↓ choose, ↵ output, k kill, d detach
/background	Hand this conversation to a background agent, or let go of one. /bg and /handoff are the same thing
/agents	Background agents and saved conversations; ↵ switches to one in place. /resume is the same
/status	Version, installed versions, launcher, update settings
/config	Open the settings menu. ↑↓ choose, ←→ change, ↵ edit
/config <key> <value>	Set one key without opening the menu
/restart	Restart with the settings on disk. The conversation is not carried over
/update	Update to the newest published version
/exit, /quit	Leave the REPL

A line starting with / is not always a command

An absolute path is the common case, so anything with a / or a . in the first word is treated as ordinary input and sent to the model. /home/user/file.txt reaches the agent; /cost does not.

Commands that work mid-turn

/jobs, /cost, /help, /plan and /background run where you type them, even while the model is working. Anything else you type is queued and becomes the next prompt.

Permission modes

Five modes, cycled with Shift+Tab, live during a turn as well as at the prompt. The active mode shows under the prompt and in the pinned region while a turn runs.

Mode	Behaviour
plan	Read-only. Mutating tools are refused until a plan is approved
manual	Ask before every mutating tool. The default
accept edits	File edits apply without asking, inside the working directory. Bash still asks
auto	Nothing routine is asked, inside the working directory
dangerous	Every waivable request is approved. Shown in red

Hard-stop tools (a PLC write, a motion command) ask in every mode. Full detail in Permission Model.

Output rendering

Feature	Behaviour	Disable with
Unified diffs	File edits render inline as a diff	--no-diffs
Collapsible results	Long tool output collapses to a placeholder; /expand prints it in full	--no-collapse
Colour	Auto-detected from the terminal	--no-color, or NO_COLOR
Cost meter	Running token and cost totals; /cost for the breakdown	—

Collapsed results stay expandable after a resume. A resumed transcript replays through the same renderers a live turn uses.

Background jobs

When the agent starts a long-running task, a row appears under the prompt with a live timer:

›
  2 jobs ⚙ npm test 1m34s ⚙ vite dev 2m07s  ·  /jobs

/jobs opens the list. The cursor follows the job row, not a line number, so a job finishing above the one you are about to kill does not slide a different one under your keypress.

Full detail in Background Agents & Jobs.

Pasting

Paste size	Behaviour
Up to 3 lines and 280 characters	Appears inline
Larger	Collapses to a [Pasted N lines, M chars] placeholder; the full text is sent on ↵

Backspace deletes a collapsed paste whole rather than a character at a time.

Restarting

/restart reloads the settings on disk (endpoint, model, packs) and starts fresh. The conversation is not carried over. If you want to keep it, hand it to a background agent with /background first, or resume it later from Sessions.