Expose current usage / limit state to the model (self-pacing for long autonomous runs)

Summary

Expose a read-only signal of current session usage / remaining budget (e.g. % of the 5-hour limit used, time-to-reset, or a remaining-budget estimate) to the model in-context, so it can self-pace long autonomous runs.

Problem

During long autonomous work — orchestrating multiple background Workflow batches near a usage limit — the model has no visibility into its own session usage or how close it is to the 5-hour limit. To decide whether it's safe to launch another batch (and still have headroom to validate + commit the results), it has to ask the human to relay the percentage each time. That puts a person in the loop purely as a usage gauge.

In a real session this looked like: "92% used." → "run one more batch (~3% each)." → launch → "what's the % now?" — repeated. The model was otherwise fully autonomous; the only thing it needed a human for was reading the meter.

Proposed

A read-only in-context signal or tool — e.g. pct_used, remaining_budget, time_to_reset — that lets the model:

• size or defer the next unit of work (batch/workflow) against remaining headroom,
• reserve budget to finish + commit in-flight work before a cutoff,
• write an accurate "resume here" checkpoint before hitting the wall rather than mid-operation.

Note on what already exists

The Workflow tool already exposes a per-turn token budget (budget.total / budget.spent() / budget.remaining()) when a target like +500k is set — so the model can self-pace inside a workflow against that target. What's missing is the session-level 5-hour usage/limit state, which is exactly the value users currently have to relay by hand.

Environment

• Claude Code (CLI), long autonomous multi-workflow runs near the 5-hour usage limit.

_Filed at a user's request after repeatedly relaying usage % to keep an otherwise-autonomous run paced._