Reset burn-down in Power Claude — use residual capacity before the window rolls
Neural-LLM Admin··3 min read
You paid for the seat. The 5-hour window is about to roll — and least-utilized routing still treats every account as if tomorrow is infinite. Reset burn-down is Power Claude’s default-on answer: pull more from seats that are about to reset while residual capacity remains, so usage is not left on the table.
TL;DR
| Item | Detail |
|---|---|
| What | Prefer near-reset seats with residual 5h/7d capacity |
| When | Multi-account rotation path is active |
| Default | **ON** (`powerClaude.proxy.resetBurnDown`) |
| Urgency | Progressive over the last ~90 minutes to reset |
| Math | Anthropic reset headers (exact) + utilization residual |
| Not | Limit bypass, invented capacity, or auth-dead forcing |
The problem least-utilized alone leaves open
Classic **least-utilized** multi-seat routing is capacity-fair: it keeps filling the emptiest window. That is great for long-horizon balance.
It is a poor fit for a **rolling 5-hour (and weekly) cliff**. A seat at 60% utilization with **15 minutes to reset** still has real residual capacity. If the router keeps hammering a far-from-reset “emptier” seat instead, that residual **vanishes at roll** while other seats stay available for later — wasted money in the window that just closed.
What reset burn-down does
When rotation is enabled across seats you own:
1. Read **utilization** and **reset timestamps** already persisted from Anthropic rate-limit headers.
2. Compute **residual** = capacity left in the window (`1 − utilization`), never boosting hard-blocked seats.
3. Compute **urgency** that ramps **progressively** (quadratic) as time-to-reset falls inside a ~90 minute horizon.
4. Rank seats so near-reset residual sorts **ahead** of far-from-reset idle peers for the next request.
Far seats stay fresher for the next stretch of work. Near seats finish what they can before the window is gone.
Defaults (shipped)
| Setting | Default |
|---------|---------|
| `powerClaude.proxy.resetBurnDown` | `true` |
| `powerClaude.proxy.parallelMode` | `smart` (least-utilized + RPM fan-out + burn-down) |
Turn burn-down off only if you want pure least-utilized without near-reset preference.
How it pairs with Runout Forecaster
- **Runout Forecaster** answers: “Will I exhaust *before* reset at this burn rate?”
- **Reset burn-down** acts: “Given residual *and* a near reset, prefer that seat so residual is used.”
Together: see the wall, and stop leaving unpaid capacity on the floor in front of it.
Safe claims (read this)
Power Claude does **not**:
- Increase Anthropic limits or invent tokens.
- Force traffic onto revoked / `invalid_grant` seats.
- Replace human OAuth when a seat is auth-dead.
It only **orders serveable seats** you already own using signals Anthropic already returns.
FAQ
Is this on by default?
Yes. `powerClaude.proxy.resetBurnDown` defaults to **true**. It only matters when multi-account rotation is actually selecting among seats.
Does it replace Power Mode / least-utilized?
No. It is an overlay on least-utilized (and smart). Far from reset, behavior collapses to classic least-utilized.
What if every seat is far from reset?
Urgency is zero → no burn-down discount → normal capacity-fair routing.
