grok-build-upstream-mirror/crates/codegen/xai-grok-pager/docs/tutorial/07-plan-and-permissions.md

39 lines
1.4 KiB
Markdown
Raw Permalink Normal View History

Synced from monorepo Synced from monorepo Changes: - Workspace server: report `/ready` as failed with dwell on hub connect failure - Refresh OIDC token for the Grok agent in the shell - ACP terminal output recorder - Cross-platform provider auth commands in the shell - Default `/resume` to Grok sessions with a hint for hidden external sessions - Resume sessions by title with `--resume` - Limit app-builder archive size - Data-driven tag labels for slash commands - Doctor fixes for tmux - Custom provider gateways and subprocess environment policy in the shell - `/tutorial` — opt-in onboarding tour of Grok Build - Soft and required CLI version checks in the shell - Privacy banner env overrides survive live settings updates - Add remote flag to override the image-edit model - Return profile fields from auth info even when the access token is expired - Add edit control on queued prompt rows - Keep fail-closed policy when clearing orphans with no team - Setting to disable the Ctrl+Space/F8 voice shortcut - Pass `--raw` to pw-record so Linux dictation works on older PipeWire - Validate git URLs when adding marketplace entries - Stop shipping stale tool-doc parameter and tool names - Re-point dashboard attach after `/fork` only when the parent was attached - Surface Grok Computer media-generation results as file-path chunks - Clear web background-task tray on kill and keep the task description - Show privacy upsell banner in agent view until acted on - Add tools-server client callback surface - Protect persistent global hook sources Source-Revision: 95d84f443eddcbed6cbfd6eed22e2eafe6b3939d
2026-07-23 17:12:33 +00:00
# Plan Mode & Permissions
Grok asks before doing anything risky — and can plan before it codes.
## Permissions
When Grok wants to run a risky command or edit a file, it pauses and asks:
allow once, always allow that kind of action, or deny.
Reading is always free: file reads, searches, and safe read-only commands
(`ls`, `git status`, `grep`, …) never prompt. Chained commands are
checked piece by piece — `ls && rm -rf tmp` still prompts for the `rm`.
Trust the session? `/always-approve` (or `Ctrl+O`) skips the prompts.
## Plan mode
For bigger or more ambiguous tasks, use **plan mode**: Grok explores the
codebase read-only, designs an approach, and presents a plan you approve
*before* any code is written.
- **`Shift+Tab`** (prompt focused) cycles the mode: Normal → Plan →
Always-approve.
- **`/plan`** enters plan mode directly; `/plan <task>` plans that task in
one step.
When the plan is ready: `a` approves, `c` comments on a specific line,
`s` requests changes — Grok iterates until you're happy, then implements.
A good habit: plan mode for "how should we even do this?", normal mode for
"just do it".
## Long-running commands
A build or test run hogging the turn? **`Ctrl+B`** sends it to the
background — Grok keeps working and you're notified when it finishes
(`Ctrl+G` shows the tasks pane).
*Go deeper: `/docs Plan Mode` or `/docs Permissions and Safety`*