Skip to main content
Commands are built-in actions you can invoke from the chat input bar. Like skills, commands are triggered with /, but instead of referencing reusable instructions they perform an action on the session directly.

​
Using commands

Type / at the start of the chat input bar to open the slash popup. Built-in commands appear first with a built-in tag, followed by installed skills. Select a command to insert it into the input, or finish typing and submit it as a normal message. Commands must be the first token of your message. Typing /compact after other text — such as “what does /compact do?” — is treated as plain text and the command is not run. If you submit a command while the agent is still streaming, CTOR queues it and runs it at the next turn boundary instead of dropping it.

​
Available commands

​
/compact

Manually compact the current session’s conversation history. Use this when you want to free up context before hitting the automatic compaction threshold — for example, before kicking off a long sub-task. CTOR replaces older messages with a summary while keeping the most recent turns intact. The summary message appears in place of the compacted history; subsequent messages continue against the shorter context. Follow-up prompt. You can chain a prompt onto the same submission:
CTOR runs compaction first, then sends the trailing text as a new prompt against the compacted history. Availability. The slash popup probes whether the current session is long enough to produce a useful cut. When it isn’t, the /compact entry is hidden from the popup — you can still type it manually, but nothing will happen if there’s no cut point. Editing a previous message. Rewriting an earlier user message to /compact and submitting the edit compacts the conversation up to that message. The edit form discards any trailing text after /compact to avoid splitting the action across two operations.

​
/goal

Set a persistent objective for the current session. The agent works toward the goal automatically, evaluating progress after each turn and continuing until the objective is complete.
CTOR sends the objective as the first prompt and starts working. After each agent turn, a lightweight evaluation checks whether the goal is complete:
  • Complete — the goal is marked done and a completion notification appears in the chat.
  • Incomplete — CTOR records the current progress reason, increments the turn count, and automatically sends a continuation prompt so the agent keeps working.
The goal persists outside the message list, so it survives /compact, automatic compaction, and message edits. Normal follow-up messages and steering do not clear it. Status bar. While a goal is active, a status row appears above the input bar showing the objective text, elapsed time, turn count, and the latest progress reason. The row includes Status and Cancel buttons. Checking status. Type /goal with no argument to append a status notification to the chat:
Cancelling. Type any of the following to cancel the active goal and stop automatic continuations:
You can also click the Cancel button on the status bar. Queuing. If you submit /goal <objective> while the agent is streaming, CTOR queues it and starts the goal at the next turn boundary. Status and cancel actions run immediately even during streaming.

​
/plan

Enter plan mode to draft an implementation plan before the agent touches code. The session becomes read-only — file writes are restricted to plan/, shell commands run in a read-only sandbox, and MCP tools are blocked — until you approve the plan. With a task description. Append a task to start planning immediately:
The agent explores the codebase, writes a structured plan to plan/, and presents it for your approval. Without a task. Type /plan alone and the agent asks what you want to plan. Running a saved plan. Implement a plan you saved earlier with /plan execute (or /plan run) followed by its path:
This runs in a normal read-write session rather than entering plan mode: the agent reviews the saved plan, then implements it. The path is relative to the project root. Exiting. Type /plan exit (or off, cancel, stop, quit) to leave plan mode without approving a plan. You can also click Exit plan mode on the indicator bar below the messages.

Plan mode

See plan mode for the full workflow: reviewing, approving, progress tracking, and enforcement details.