AI coding assistants are fast — sometimes faster than the human who should be reviewing what they're about to do. A database migration that drops a column, a refactor that renames a widely-imported function, an infrastructure change that modifies a security group — these shouldn't run without a human who understands the blast radius saying "yes, do it."
VibeControls Plan Mode inserts a deliberation step between "AI proposes" and "AI executes". Enable it in workspace settings and select a plan provider (Plannotator is the reference implementation). When an AI session proposes a tool call that crosses the blast-radius threshold, it pauses and generates a structured plan: what it intends to do, what it will change, what it expects the result to be.
The plan is surfaced in the Plannotator portal — a separate, calm interface designed for review. The reviewer sees the proposed steps, can annotate them ("this is fine", "change this line"), and approves or rejects. Only after approval does the AI proceed.
An idle watchdog cancels unapproved plans to prevent AI execution of stale proposals. The entire approval is audited: plan content, reviewer identity, approval or rejection, timestamp.
For compliance-heavy teams, this turns AI from a liability into a governed tool.
Do it yourself
Set up Plan Mode so your AI sessions pause on high-risk changes, generate a structured plan, and wait for your annotated approval in the Plannotator portal before executing.
On a healthy, online agent, open the Plugins page and install the plan plugin. You can also run the install command from your terminal.
You should see: The plan provider (Plannotator) becomes available to your agent and appears as a selectable provider.
Ready to make this your story?
