Session modes

Blue's interaction intensity has three session modes, cycled with Shift+Tab under editor focus:

normal → plan → yolo

Outside normal, a mode badge sits in the status bar's first row (plan in the accent tier, with a pending ellipsis while messages are queued; yolo in the warning tier). /yolo [on|off] (alias /yes) is the command form of the yolo toggle — same effect as cycling.

normal

The default. Every tool call pops the four-option approval panel (allow once / allow for this session / reject / reject with feedback, see Approvals & questionnaires); user questions pop as usual.

plan — plan first, act later

In plan mode the agent produces a plan before acting. When the plan is final, the harness's exit_plan_mode request surfaces as the plan review panel (editor-slot replacement, mounted like the approval panel):

• the full plan renders as Markdown inside a bordered plan box;
• beneath it a numbered decision list — number keys pick directly, or ↑↓ + Enter:

Option	Effect
1. Approve	Approve the plan, exit plan mode, start executing
2. Reject	Reject — the model hears "the user chose to keep planning" and reacts in the same turn
3. Revise <text>	Inline revision: keep polishing the plan with your feedback

yolo — auto-approve

yolo auto-approves tool calls — the four-option panel stops popping and the agent runs at full speed. User questions still pop: yolo waives tool authorization, it does not answer questions for you. /yolo, /yes, or another Shift+Tab cycles back at any time.

Relation to /preset

Plan mode is supplied by the harness's plan-mode plugin, composed through agent presets (/preset, see the slash commands reference) — the preset decides which capabilities a session has; the session mode decides how closely this interaction asks.