Reserves & Pace Mode

Reserves is the section in the saddle that tracks how much of each subscription you’ve used and decides — based on your pace mode — when to stop sending jobs to that subscription.

This is the differentiator. Nobody else has this.

The problem it solves

You have a Claude Max subscription. It has a 5-hour rolling token window, a weekly cap for all models, a separate weekly cap for Sonnet, and an extra-usage budget for overage. Four overlapping limits. None of the existing dispatch tools track any of them. They just hammer the API and hope.

ModelReins tracks all four. Then it lets you dial how close to the cliff you want to ride.

The four pace modes

Each subscription has a pace mode you can change with one click:

Mode	Symbol	Color	Behavior
PACE	🐢	green	Spread remaining quota evenly across days till reset
FAIR	⚖	yellow	Today gets a fair share + 50% headroom for surprise tasks (default)
SPRINT	🔥	orange	Ignore daily pacing, fire until 95% absolute cap
COAST	🚫	gray	Don’t touch this subscription at all — route around

To switch modes, click the colored pill next to the subscription name. It cycles through pace → fair → sprint → coast.

Real example

You open Claude settings and see:

5h session:        59% used, resets in 2h 48m
Weekly all models: 32% used, resets Sat 10pm
Weekly sonnet:     7% used, resets Mon 6am
Extra usage:       95% ($9.51 of $10), resets May 1

Today is Tuesday. You have 4 days until weekly reset.

”I won’t be on much tomorrow, burn it”

Set the dial to 🔥 SPRINT. The router thinks:

• Weekly all: 32% used, sprint cap = 95% absolute, headroom = 63%
• Weekly sonnet: 7% used, headroom = 88%
• 5h session: 59% used, only 0.1 days till reset, just keep firing

It dispatches every job to Claude until the 5h session caps, then auto-fails-over to local models until the session resets, then resumes burning weekly. You get 4 days of quota in 1 day.

”Pace me till the weekend”

Set the dial to 🐢 PACE. The router thinks:

• Weekly all: 32% used, 68% remaining, 4 days till reset
• Today’s fair share: 68% / 4 = 17%
• Soft cap = 32 + 17 = 49% (just for today)

When you hit 49% used today, the router stops sending to Claude Max and routes to local models for the rest of the day. Tomorrow it computes a fresh 17% allowance.

”Default”

Leave it on ⚖ FAIR. Today gets a slightly larger share (1.5× pace) so you have headroom for surprise tasks but you still won’t burn through the week in a day.

Pasting your usage

Subscriptions don’t auto-fetch their usage yet (that’s Phase 2 — workers reporting rate-limit headers). For now you paste it in.

In the saddle, click 📋 paste usage under any subscription. The format:

name:percent:resets_at

Multiple windows separated by ; or newlines:

Weekly all:32:2026-04-12T02:00:00Z
Sonnet only:7:2026-04-13T10:00:00Z
5h session:59:2026-04-07T20:48:00Z
Extra usage:9.51/10:2026-05-01T00:00:00Z

The parser auto-detects three kinds:

• quota (default) — token windows that affect routing
• balance — when the value is current/max and the name contains “credit”, “balance”, “remaining”, “reserves” (e.g. Credits:23120760/25000000 for 1min.ai)
• overage_budget — when the name contains “overage”, “extra usage”, “spend”, “safety”, “budget”, or starts with $ (info-only, never blocks routing)

The parser handles commas in numbers (23,120,760 works).

How the router uses pace

When you dispatch a job:

1. Router gets the list of eligible workers
2. For each worker, looks up its subscription
3. Asks subscription_has_capacity() which checks the pace policy
4. Workers tied to blocked subscriptions are dropped from candidates
5. Worker scoring + selection happens on the remaining set
6. If everything is blocked, the dispatch errors with a clear “all subscriptions at cap” message

The rest of the routing logic (tag overlap, success rate, load) operates on the surviving candidates.

Visual reference

Each window in the saddle shows:

Window name       [████░░░░░░] 47%
                   ↑    ↑    ↑
                   bar  cap  used%

• Bar fill color = current usage zone (green <50%, yellow 50-80%, red 80%+)
• White cap line = where the pace mode says to stop
• Used% = actual usage from the latest paste

When the bar fill catches up to the cap line, the router stops dispatching. Watch the gap between them — that’s your headroom.

What’s coming (Phase 2 / Phase 3)

• Phase 2 — workers report rate-limit headers automatically. The paste form sticks around as the manual override but auto-calibration handles the day-to-day.
• Phase 3 — sparklines on each window showing usage history, “when will I hit my cap” projections, multi-subscription roll-up view.