Still grounded in real investigation

Because read-only tools continue to run normally while plan mode is active, the plan that comes out the other end is grounded in the same real investigation an ordinary turn would have done, not a guess assembled without actually looking at the code.

A plan for a refactor, for instance, still reflects an actual search for every call site that would be affected by the change, run for real during planning; it is only the edits themselves that get deferred, never the work of figuring out what they should actually be.

This is what keeps a plan trustworthy enough to approve wholesale rather than needing to be re-verified line by line before you say yes to it.

The only category of tool that is genuinely deferred is the permission-gated one, write, edit, and command; read-only tools have nothing to defer in the first place, since there was never anything at stake in running them immediately.

This distinction is what makes plan mode fundamentally different from simply asking tCode to "describe what you would do" without actually looking: a description without investigation is a guess dressed up as a plan, while what plan mode produces has already done the real work of finding out.