Surface plan usage limits to the agent for budget-aware planning

Problem

The two pieces of information needed to decide whether to start a piece of
work — how much it will cost, and how much budget remains — are currently
split between the user and the agent, and neither party holds both.

• The user can see plan usage (/usage, claude.ai settings) but has no way

to estimate what a proposed task will consume.

• The agent can estimate task effort reasonably well but has zero visibility

into the user's session/weekly limits, even though both live inside
Anthropic's own product.

The result: the person with the least information about the work is the one
budgeting it. Managing a weekly limit becomes manual oversight by the user,
with no instrumentation beyond watching the usage meter between turns.

Concrete scenario

Mid-project, approaching my weekly limit on a higher-tier model, I asked the
agent whether we should start the next work phase. I was really asking "will
this fit in my remaining budget, or will it bail mid-flight?" The agent had
no way to know my remaining allowance, and I had no way to know the phase
would take N hundred thousand tokens. It answered the technical question and
started the work. It happened to land fine — but only because the work was
structured in small commits, not because either of us could see the runway.

Proposal

Inject the user's current plan usage (session and weekly, per model tier)
into the agent's context, the same way session token budget already is.
With that, the agent could:

1. Estimate before starting: "this refactor is a medium job, roughly X;

you have Y left this week on this model."

1. Warn when a requested task likely won't fit in the remaining allowance,

and offer alternatives (smaller scope, lighter model, defer).

1. Structure work into landable increments sized to the remaining budget,

instead of discovering the ceiling by hitting it.

Alternatives considered

• Status line / statusline scripts can show usage to the user, but the

agent still can't plan against it.

• Asking the user to relay their /usage numbers each session works but is

exactly the manual overhead this should remove.

Read-only visibility is sufficient — no need for the agent to manage or
modify limits. Even a coarse signal (percent of weekly budget remaining per
tier) would enable all three behaviors above.