In-chat /prism commands
A chat client that cannot add HTTP headers can still steer Prism. Send a
message that consists only of a /prism command. Prism answers it itself, in
the shape of a normal Chat Completions or Responses reply, and does not send it
to any model.
/prism help/prism models/prism session/prism reset-session/prism change-model/prism change-model select<number>/prism cancel/prism next-model <candidate_id>/prism prefer-model <candidate_id> [--session] [--ttl 30m]/prism clear-preference/prism strategy <balanced|quality|cheap|fast|private> [--ttl 30m]/prism clear-strategy/prism mode <manual|auto> [--ttl 2h]/prism clear-mode/prism strict [--ttl 30m]/prism clear-strictSession
Section titled “Session”Commands that change routing apply to one conversation, identified by the
X-Prism-Session header. Without that header Prism rejects them with
Prism command rejected: /prism strategy requires X-Prism-Session. (and the
same for the other session commands). /prism help, /prism models and
/prism session answer without it.
Commands
Section titled “Commands”| Command | Effect |
|---|---|
/prism help |
Lists the commands |
/prism models |
Lists the models Prism can route to, as provider/model with task types and price |
/prism session |
Shows the session’s current affinity, preferred model, strategy override, next model and pending selection |
/prism reset-session |
Clears all of the above for the session |
/prism change-model |
Moves the session to the first eligible model other than the current one. Refused while a preferred model is set |
/prism change-model select, then a number |
Shows a numbered list of candidates; the next message with a number picks one |
/prism cancel |
Cancels a pending numbered selection |
/prism next-model <candidate_id> |
Uses that model for the next request in the session, once |
/prism prefer-model <candidate_id> [--ttl 30m] |
Prefers that model for the session; --ttl sets an expiry (any positive Go duration, such as 30m or 2h) |
/prism clear-preference |
Removes the preferred model |
/prism strategy <name> [--ttl 30m] |
Sets the session’s strategy: balanced, quality, cheap, fast or private |
/prism clear-strategy |
Removes the session’s strategy override |
/prism mode <manual|auto> [--ttl 2h] |
Sets the routing mode of this conversation: manual follows the preferred model and the provider order, auto lets the strategy pick the model. Works without X-Prism-Session in a conversation Prism recognizes |
/prism clear-mode |
Removes the session’s routing mode; the agent’s or the default mode applies |
/prism strict [--ttl 30m] |
Turns the session’s strict-model rule on: an explicit model is served only by the seller it names, no other provider and no other version takes over. Keeps a calibration or judge run on the seller it asked for even when that seller hits a quota |
/prism clear-strict |
Removes the session’s strict-model rule |
<candidate_id> is provider/model as /prism models prints it, for example
openai/gpt-5-mini.
What runs and what does not
Section titled “What runs and what does not”- Only the latest user message runs, and only when it is a command and nothing else. A command inside other text, in a quote or code block, in an earlier turn, or in an assistant, system or tool message is not executed.
/prism-helpand other prefix collisions are not commands.- The reply carries
X-Prism-Control: acceptedorX-Prism-Control: rejected. - A concrete
modelin the request body wins over session controls. - Controls never bypass routing limits: health, capabilities, context size, output limits and policy still apply.
- Defaults for the whole installation cannot be set from chat. Use the web console (Routing) or the local control API.