Commands and shortcuts

Documentation · CLI flags · Terminal

Type / in the input to see every command. Tab completes a command name without running it, and Enter runs the highlighted command through its normal handler.

/status
/model
/answer Answer from the evidence already gathered. Do not use more tools.

Most commands run right away. Changes to the model or session, compaction, reload and checkout wait until current work finishes, and a change requested while work is active is listed in the queued-input area until it applies. /thinking takes effect without interrupting on explicitly qualified Responses routes, and other routes change at an idle boundary. A queued effort isn't a provider acknowledgement, and unsupported explicit efforts are rejected. Extensions queue the same way through pi.setModel and pi.setThinkingLevel. /answer saves a steering instruction and makes later requests tool-free at the next safe point, or starts a tool-free run when idle. It doesn't undo effects that already happened. Run control.

Slash commands#

Type /log… or /setu… to see both /login and /setup in the suggestions, then pick one with Up and Down and Enter. An ambiguous partial match doesn't Tab-complete on its own, and exact /login and /setup keep their separate actions. /theme previews and selects the built-in Auto, Light and Dark appearances, or valid TOML files in the trusted theme discovery roots. Cancelling leaves the theme and session unchanged (Theme discovery and format). Extensions can add commands, and each package's README lists its arguments.

Local shell escapes#

A draft that starts with ! is a local command, not model input. !command runs through the product's process gates, the ordinary process approval and the bounded capture and cleanup path, then adds the result to the model's context as a user message. !!command takes the same path but durably records the result excluded from model context. Both work while idle and during a run. A mid-run result is appended at the next idle point, after any pending tool call settles, so a call is never separated from its result. --no-process or --no-shell turns the escape off, and a denial is reported, never dropped silently. Capture is bounded by max_output_bytes and bash_timeout_secs, and it isn't a persistent shell session.

Priority and compaction boundaries#

Bare /fast toggles priority, and /fast status shows the current and queued choice. Unsupported routes refuse instead of silently ignoring the switch. The choice survives compatible same-session rebuilds but resets on another session or a process restart, so it isn't a saved preference. Turning it on first records durable priority-cost uncertainty. Turning it off doesn't erase that exposure, so hard cost ceilings stay fail-closed. The default 272K Codex context cap is unchanged, and selecting priority isn't proof that a provider granted it.

Custom /compact instructions are at most 16 KiB and reject unsupported control characters. Native Responses compaction refuses them outright. Local main and split-prefix summaries use host-owned retries, reservations and accounting, and the summary is committed once. That doesn't promise retry-progress UI or exactly-once inference after a transport interruption.

Read-only reports#

Interactive status and fact reports use section headings and bold field labels, including cache warming, safety permissions and telemetry. Cost and cache records stack their named fields rather than clipping wide ASCII tables. Values remain literal text; filenames, model names and diagnostics cannot inject Markdown or terminal controls. Plain/CLI report output keeps its text format.

Up/Down scroll, PageUp/PageDown page, and Escape returns to the unchanged draft. Reports are transient: they do not enter transcript copy, session history or model input, and reading a report never approves an effect.

Settings menu#

Bare /settings opens a filterable menu immediately, including during active work. Choose Theme, Terminal screen (fullscreen, the default, or inline), Steering mode, Follow-up mode, Inline tool-result images (On/Off), Default model for new sessions, Default reasoning for new sessions, or Show effective settings. Up and Down select, typing filters, and Enter opens or applies the highlighted choice. Child menus offer Back; Escape closes the menu without changing a preference.

Theme uses the existing preview picker. Inline images offers explicit On and Off choices for tool-result display only, not upload or attachment consent. Default model choices come from the available credential-scoped catalog; default reasoning choices use available portable levels. Saving either default does not switch the active session: use /model or /thinking for that. The menu does not label the current session or launch selection as a saved default. Selected changes use the same safe display or deferred settings dispatch as explicit commands; opening and browsing the menu does not interrupt active work.

Steering mode and Follow-up mode are Pi's steeringMode and followUpMode (config keys steering_mode and follow_up_mode). Both default to one-at-a-time, as in Pi. Steering messages are sent while the agent works: one-at-a-time delivers one per turn boundary, all delivers every queued message at the next boundary. Follow-ups wait until the agent finishes: one-at-a-time runs each as its own turn, all sends consecutive queued prompts together in one turn (slash commands, ! shell commands and /answer still run on their own). A change applies from the next run. Terminal screen chooses fullscreen (no scrollback replay or flicker) or the inline, differential renderer; --tui-mode overrides it for one launch.

Show effective settings, or /settings show, opens read-only diagnostics. Configured model is a launch fact and Active reasoning is session state; neither is presented as the saved default for new sessions. In narrower terminal panes, settings, extension and setup menus put descriptions below their bold action labels instead of squeezing both into one row. Native Tern sheets keep bold action labels in the list and wrap the full selected description below it. Explicit /settings theme, tui, steering, follow-up, images, default model, default reasoning, transport and padding forms are also available. Transport, padding, permissions and project trust are not editable menu preferences, and the menu adds no cache-policy controls.

Extension menu#

/extensions is the one place to turn extensions on and set them up. It lists installed executable bundles, plus any extension loaded from --extension-dir or the project (whose activation changes where it came from). Up and Down select. Enter opens the selected extension's options, enabling an installed one first if it's disabled. Each menu shows the extension's state and offers only what applies:

• octet-computer-use: Set up computer use, Check status, Jev (optional) and the jev-use recipe.
• octet-web-search: Use Brave Search (recommended), Use SearXNG, Change the SearXNG endpoint and Log out of Brave Search.
• octet-subagents: Enable or disable orchestration. Use /subagents for the live worker roster, inspection, wait, reattach and stop controls; these runtime operations are not extension configuration.
• octet-mcp: Add a server, then per server Show details, Refresh tools, Restart or Start, Stop, Enable or Disable, Edit and Remove.

Every menu also offers Disable. An action shows its steps live while it runs (Esc cancels), and its prompts (confirmations, choices, keys entered as hidden input) appear in place. Actions an extension marks destructive ask first. A third-party extension without its own menu gets one entry per command it declares, which asks for that command's arguments.

Commands published by ready extensions appear in slash completion and execute through their registered handlers. During active work, commands that need an idle boundary are visibly queued. /extensions remains the configuration menu; /subagents remains a runtime slash command while its first-party extension is ready. The menu never writes a trust grant on enablement. Full access implicitly trusts enabled extensions; safe mode starts only enabled extensions with explicit source-bound host authority.

Keys#

The resume picker and the subagent worker list (Ctrl+X stops the selected worker) have their own keys. Terminal scrolling covers the mouse.

User keybindings and transcript search#

The interactive shell loads ~/.octet/keybindings.json. /reload reloads it and /hotkeys shows the resolved platform bindings. Invalid or conflicting bindings are reported, never silently given executable authority. Bindings don't add clipboard image permission or chord handling.

Ctrl+Shift+F (Ctrl+F on Windows and WSL) opens bounded transcript search in the primary-screen viewport. Enter or Ctrl+G selects the next match, Shift+Enter or Ctrl+Shift+G the previous one, and Escape closes search. Search owns its query input, including paste, and never submits it or admits attachments. It doesn't search discarded capture bytes or your terminal's native scrollback. The resume picker's Ctrl+F separately searches bounded session text. Ctrl+Shift+B cycles the hidden, auto and always transcript scrollbar modes for this shell, which isn't a saved preference or an alternate-screen switch.