Gemini CLI: Plan Mode Reaches General Availability

Plan Mode Graduates to Stable in Gemini CLI v0.37.0

Gemini CLI's Plan Mode — the read-only agentic workflow that separates codebase analysis from code execution — has reached general availability in v0.37.0, released April 8, 2026. What began as an experimental feature in v0.27.0, became the default in v0.34.0, and is now a fully supported, stable capability with expanded functionality and formal documentation.

What Plan Mode Does

Plan Mode restricts Gemini CLI to a read-only subset of tools: read_file, grep_search, glob, and compatible read-only MCP tools. In this mode, the agent can navigate a codebase, map dependencies, search for patterns, and produce a structured plan — but it cannot write, modify, or delete any project files. The only files the agent may create are its own internal planning documents.

This separation of thinking from doing addresses a core risk in agentic coding: agents that explore and act simultaneously can produce cascading side effects before the developer has reviewed the proposed approach. Plan Mode enforces a checkpoint — the agent researches first, proposes a plan, and waits for explicit approval before any changes are made.

The ask_user Tool

A key capability within Plan Mode is the ask_user tool, which lets the agent pause its research and surface targeted clarifying questions. Rather than silently embedding assumptions deep into a plan, the agent asks about ambiguities early — before writing a single line of code.

What's New in v0.37.0

Untrusted Folder Support

Plan Mode can now operate in untrusted project folders with appropriate safety restrictions applied automatically. Previously, Plan Mode required the project to be in a trusted context. This change makes Plan Mode accessible in more environments — including third-party repositories and freshly cloned projects — without requiring the developer to grant full project trust upfront.

Conditional Enter/Exit Plan Mode Tools

The enter_plan_mode and exit_plan_mode tools now activate conditionally based on current context. If the agent detects it is already in Plan Mode, the enter tool is suppressed; if it is in normal mode, the exit tool is hidden. This prevents mode-toggle confusion during complex multi-step workflows and improves the reliability of automated plan execution pipelines where the agent's mode state must be predictable.

Project-Level Memory Scope

The save_memory tool has gained a project-level memory scope accessible from within Plan Mode. The agent can now persist facts it discovers during a planning session — such as architectural decisions, dependency constraints, or file ownership notes — directly to the project's memory store. These memories are available in subsequent sessions, eliminating the need to re-explore the same context every time a planning task begins.

Policy and Stability Improvements

v0.37.0 relaxes write-file policy patterns to accommodate plan-specific file paths, and the complete_task tool is now explicitly permitted within Plan Mode so agents can signal task completion without triggering a policy violation. A previously reported deadlock during plan file creation due to sandbox restrictions has been resolved, fixing a reliability issue where Plan Mode would hang in sandboxed environments.

Why It Matters

The GA promotion signals that Google considers the Plan Mode workflow production-ready for professional development environments. The combination of read-only safety, interactive clarification via ask_user, MCP tool integration, project-level memory persistence, and now untrusted folder support makes Plan Mode one of the most complete "think before you act" frameworks available in any terminal-based AI coding agent.