Agent modes

• Build: default execution mode for practical implementation loops.
• Plan: planning-first behavior that emphasizes exploration and structured plans.
• Agentic (YOLO): high-autonomy mode for rapid end-to-end execution.

• In local Spatial harness sessions, Build and Plan require tool confirmations by default.
• In local Spatial harness sessions, Agentic (YOLO) auto-approves tool confirmations by default.
• Hook policy checks are evaluated first in local harness flows; hook auto-approve/require-confirm rules can override mode defaults.
• In server-streamed session flows, exec/file operation confirmations are auto-approved and question-style confirmations still require user input.

• Plan mode prompts the agent to investigate, ask clarifying questions, and prepare structured execution plans.
• Build mode prompts the agent to execute implementation work with full tooling and quality checks.
• Agentic mode prompts autonomous execution and confirmation-free momentum.

• Start in Plan when scope is unclear or high-risk.
• Move to Build for day-to-day implementation and verification.
• Use Agentic for tightly-scoped tasks where you explicitly want speed over guardrails.

Claude Code and Codex local CLI models expose the same three user-facing modes: Build, Plan, and Agentic.

When switching models, unsupported or invalid mode selections are normalized back to supported defaults to keep sessions runnable.