Plan Mode
Put Sapience into Plan mode to first use the best model available to create a complex multi-agent or multi-turn action plan. Amazing.
1. Plan mode (/plan)
What it is. A conversation mode. When you type /plan (alone or as a prefix), the conversation enters plan mode. While plan mode is active:
- A plan artifact (a
Planobject stored in the backend) is created and attached to your conversation.
- Every turn the agent takes, it gets a reminder injected into its prompt that says "you are in plan mode, here is the current plan content, keep it up to date."
- The agent is automatically switched onto the planning model (a higher-reasoning model) for plan-mode turns. It also gives the agent more time, and automatically flips into long polling mode (no 4-minute cut off).
- Five plan-mode tools become available to the agent:
enter_plan_mode,exit_plan_mode,create_plan,read_plan,edit_plan.
How you turn it on.
- Type
/planat the start of any message — for example,/plan let's refactor the upload handler to handle ZIPs. You can also select it from the mode dropdown.
- Plan mode activates and the rest of your message becomes the brief for the plan.
How you turn it off. Either ask the agent to exit plan mode, or the agent calls exit_plan_mode itself when the work is done. The plan stays in the database as an artifact you can revisit.
Important behaviour points:
- Plan mode is per-conversation. One
Planper conversation; sending/planagain on later turns does NOT create a duplicate plan — it's a no-op once plan mode is already active.
- The plan is markdown with a Context / Approach / Files-to-modify / Tasks / Verification skeleton. Tasks use GitHub-flavoured
[ ]/[x]checkboxes so progress is visible at a glance.
- Plan mode is gated by per-agent setting If your agent doesn't have plan mode enabled, the slash-command and the toggle will appear to do nothing — the agent simply won't have the 5 plan tools.
When to use it. Anything that has 3+ distinct steps, spans multiple files or systems, or needs written verification before you'd call it done. Refactors, multi-step features, structured investigations.
When NOT to use it. A one-line fix. A single question. A piece of work the agent can finish in one turn. Plan mode's overhead (the artifact, the per-turn reminder, the slower model) only pays off when the work actually benefits from a written, evolving record.