Gemini CLI: Plan Mode
Gemini CLIGemini CLI introduced Plan Mode as an experimental feature in v0.27.0, giving the agent a structured "think before you act" workflow for complex tasks. In Plan Mode, the CLI operates under a strict read-only policy — it can research, analyze, and produce a detailed execution plan without modifying any files. Users review the plan and explicitly approve before any changes are made. The v0.28.0 release expanded Plan Mode further with a generic Checklist component and support for custom deny messages in plan mode policies.
Sources & Mentions
4 external resources covering this update
Gemini CLI Weekly Update [v0.28.1]: Custom Themes, Background Shells, and Integrations
GitHub
Conductor: Enforce a "Spec → Plan → Code" Workflow in the Gemini CLI
Hacker News
Introducing a Structured Workflow for Gemini CLI: Safer, More Disciplined, and Surprisingly Capable
GitHub
Philipp Schmid: Here is how you can add Plan Mode to Gemini CLI
X (Twitter)
Plan Mode: Gemini CLI Gains a "Think Before You Act" Workflow
Released as an experimental feature in v0.27.0 (February 3, 2026) and expanded in v0.28.0 (February 10, 2026), Plan Mode introduces a structured, approval-gated workflow to Gemini CLI — one of the most significant behavioral changes to the agent's execution model since launch.
The Core Idea
Plan Mode separates the planning phase from the execution phase of an agentic coding session. When a user initiates Plan Mode, Gemini CLI switches into a strict read-only policy context: the agent can call research tools, inspect the codebase, read documentation, and reason about the best approach — but it cannot write files, run destructive shell commands, or make any changes to the workspace. The result is a detailed, structured plan that the user reviews before anything is actually executed.
This architecture directly addresses one of the most common criticisms of autonomous AI coding agents: the tendency to dive straight into implementation without surfacing a plan for human review, sometimes making changes that are difficult to reverse.
How It Works
Plan Mode integrates the AskUser tool into the planning workflow, allowing the CLI to surface clarifying questions before finalizing a plan. If the agent encounters an ambiguity — which module to refactor, whether a breaking change is acceptable, which test suite to run — it can pause and ask rather than assuming. This produces plans that are both more accurate and more aligned with the user's actual intent.
Once planning is complete, the plan is presented using the new generic Checklist component introduced in v0.28.0. Each step is listed with its associated tool calls and expected outcomes. Users can review the plan, request modifications, and cycle through approval modes using Shift+Tab before the agent transitions to execution.
Policy Enforcement
Plan Mode enforces its read-only constraint at the scheduler level. Any tool call that would result in a file write, deletion, or external side effect is blocked during the planning phase. If a complex skill invocation or nested agent delegation inadvertently attempts such a call, execution is halted with a structured policy violation message rather than silently proceeding.
The v0.28.0 release added support for custom deny messages in plan mode policies, allowing administrators to provide contextual guidance when certain actions are blocked during planning — for example, directing users to an internal approval workflow before a production deployment step.
Integration with Skills and the Broader Agentic Architecture
As Plan Mode matures, Gemini CLI is expanding its integration with Agent Skills. Research-oriented skills can now be activated during the planning phase to provide domain-specific analysis without triggering writes. A codebase_investigator skill, for example, can gather architecture details during planning to inform a more precise refactoring plan. The community has reported that combining skills with Plan Mode produces significantly more accurate and context-aware plans than generic prompting alone.