All the commands you can type to control your agent
Slash commands let you control your agent directly from the chat. Type a
command starting with / and your agent will act on it instead of treating
it as a regular message.
Commands:/think, /verbose, /reasoningThese are directive commands — they modify how your agent processes your
message rather than being separate actions. When you use a directive, the
command is stripped from your text before your agent sees it. Your agent never
sees the /think part — it just receives your message with the thinking level
already adjusted.
Controls how deeply your agent reasons about its response. Higher levels mean
more thorough thinking, which can improve answer quality for complex questions
but uses more tokens.
Level
What it means
off
No extended thinking — fastest, cheapest
minimal
Light reasoning for simple tasks
low
Basic reasoning for straightforward questions
medium
Balanced reasoning for most tasks
high
Deep reasoning for complex problems
xhigh
Maximum reasoning depth for the hardest problems
Without an argument, /think toggles thinking on or off.
Toggles detailed output. When enabled, your agent includes more information
in its responses — useful for debugging or understanding its decision-making.
Commands:/context, /status, /usage, /configRead-only commands that show you what is happening. None of these change
anything — they just report on the current state.
Shows how much of the context window your agent is using. This includes the
sizes of workspace files, tool definitions, and conversation history. Useful
for understanding why your agent might be forgetting earlier parts of a long
conversation.
Shows a breakdown of token usage and costs per provider. If you are using
multiple models or providers, this helps you understand where your budget is
going.
The switch sub-command takes two arguments: the provider name and the model
identifier. The change applies only to the current session — it does not
modify your configuration file.
Starts a completely fresh conversation. Your agent gets a new session with
no history. This is useful when you want to change topics entirely or start
from a clean slate.
Clears the history of the current session but keeps the same session
identifier. Your agent forgets what was said in this conversation but stays
in the same session context.
Manually triggers the context engine’s LLM compaction step. When your
conversation gets long, this command summarizes older messages to free up space
in the context window. Normally, the context engine manages conversation length
automatically through multiple optimization steps (see
Compaction), but you can trigger the summarization step
directly if you want to reclaim space sooner.Optional arguments:
verbose — Include detailed output about what was compacted
Any additional text is used as instructions for the compaction summary
Commands:/stopCancels the current agent execution. If your agent is taking too long, going
in the wrong direction, or stuck in a tool loop, /stop immediately halts
the current run. The session remains intact — you can continue chatting
normally after stopping.
Creates a copy of your conversation at your latest message. This lets you
try a different approach without losing the original conversation. Think of it
as going back to the last thing you said and asking for a different response.
Command:/budgetThe /budget command sets a per-execution token budget for the next agent run. The amount must be a number with a k (thousands) or m (millions) suffix; bare integers are not accepted. Valid range is 10k to 10m tokens.
Usage
Effect
/budget 500k
Set the next execution to use at most 500,000 tokens
/budget 2m
Set the next execution to use at most 2,000,000 tokens
/budget
Show usage help and the configured range
You can also use the inline form +500k or +2m at the start or end of a message (e.g., +500k research the topic).Like /think and /verbose, /budget is a directive command — it modifies how your agent processes the next message rather than being a separate action.
Commands:/skill:nameCustom prompt skills can be invoked directly from the chat using the
/skill:skillname syntax. The skill name comes after the colon, and any
text after the name is passed as arguments.For example, if you have a prompt skill called weather:
/skill:weather London
This invokes the weather skill with “London” as the argument. The skill’s
prompt template is expanded and used to guide your agent’s response.See Prompt Skills for how to create and manage
custom skills.
Directive commands (/think, /verbose, /reasoning) can be combined with
a regular message. For example:/think high Tell me about quantum computingThis sets reasoning to high AND sends the message “Tell me about quantum
computing” to your agent. The directive is stripped before your agent sees
the text.
This works because directives modify the processing of your message, not the
message itself. Other commands (like /status or /new) are standalone — they
do not accept additional message text.
Slash commands are only recognized when your message starts with the /
character. If you write a / in the middle of a sentence, it is treated as
regular text. This means you can safely type things like “I need help with
my /home directory” without triggering a command.Commands are not case-sensitive for the /skill: prefix (/Skill:weather and
/skill:Weather both work), but the built-in command names must be lowercase
(/think, not /Think).If you type an unrecognized command (for example, /hello), it is treated as
a regular message and passed to your agent as-is. Only the 17 built-in
commands and the /skill:name prefix have special behavior.
Slash commands work on every platform — Discord, Telegram, Slack, WhatsApp,
the web dashboard, and all other connected channels. They are processed by
Comis before reaching your agent, so the platform does not need to support
slash commands natively.
