Agent

Use the agent when work spans files and you can describe the outcome. You will see which files were read and which patches were proposed. Tool use follows your permission level and the active mode.

Mode table

Agent

Default implementation mode. Can read, search, edit, patch, and — when permitted — use shell, git, and other tools to complete the task. Best for concrete outcomes you will review in diffs.

Plan

Design-first mode. Soft-blocks shell and non-markdown product edits so the agent focuses on approach, trade-offs, and a written plan. Markdown plan artifacts are the intended write surface until you switch to Agent.

Ask

Read-only exploration. Answers questions using reads and search without applying code changes. Prefer Ask when you are still mapping the system.

Debug

Investigation mode centered on evidence and reproduction. Gather logs, failing tests, and a clear repro before proposing fixes. Switch to Agent (or apply a scoped edit) once the failure is understood.

Switching modes

Use SwitchMode when the task shape changes — for example Plan → Agent after you approve the approach, or Agent → Ask when you need a read-only digression. Mode changes keep the conversation context while updating guardrails.

Iterations and review

A single agent run is bounded to roughly 24 tool/model iterations. If work is unfinished, continue with a follow-up message rather than assuming unlimited loops. Always review SCM diffs before you rely on the result.

Language

Persian prompts get Persian replies; English prompts get English replies. Mixed prompts follow the dominant language of the request.