Modes
One toggle from ask to autonomy.
Click a mode to see what it allows. Every mode shares the same chat, memory and work-item context.
Chat — ask, then approve
The safe default. The assistant can use tools, but every action is presented as an in-chat consent card before it runs — with a live countdown that pauses while you're away. Cards collapse, dismiss, and stay bound to the session that raised them. Read-only commands run instantly when Auto-approve harmless commands is on: no card, no timer.
Plan — look before you leap
Read-only mode. The model inspects the repo, the work item and its acceptance criteria, then proposes a plan. Nothing is edited, no commands run — the plan comes back as a summary you can approve or refine.
Act — the agentic workhorse
Edits are applied directly, terminal commands run against a wildcard allowlist (git *, npm run *…). Independent tool calls execute in parallel, and batched edit_file edits are serialized per file.
YOLO — full autonomy
Every tool and every command runs without prompts — except pushing to the remote repo, which still asks even here (adoCode.yolo.pushApproval, default on; turn it off for full autonomy). Intended for trusted tasks on isolated worktrees where a mistake costs a reset, not a production incident. Checkpoints keep restores one click away.