Copia: multi-provider compute burn tracking — quota, bandwidth, and local inference as unified resource model

[koad]

The idea

The weekly compute quota is infrastructure, like gas in a tank. We should track it the way we track CAD spend.

What to track

Per-entity burn rate:

• Every Agent tool invocation returns total_tokens in the task notification
• Attribute tokens to the invoking entity (Sibyl, Faber, Veritas, Rufus, etc.)
• Build a per-entity cost model: Sibyl brief = ~X tokens, Faber post = ~Y tokens, Veritas review = ~Z tokens

Known data points from Day 6:

• Full throttle (5-6 parallel agents): burns weekly 5x quota in ~3 days, ~CAD 75/hour
• Free credits (CAD 140 equivalent): burned in ~2 hours at full throttle
• Paced (1-2 agents): unknown — need one week of data

The "gas in the tank" view:

• Weekly quota = full tank (resets Thursday 10pm ET)
• Track % remaining vs. work remaining in the pipeline
• Forecast: can we complete the current arc before the quota runs out?
• Alert at 50%, 75%, 90% consumed

Burst vs. baseline model

The operation runs in two modes:

• Baseline: steady, paced, 1-2 agents at a time — most days
• Burst: full parallel throughput for arc shipping or deadlines — 2-3 sessions/month, ~2-3 hours each

Copia should model both and flag when a planned burst exceeds remaining quota.

Why this matters

At CAD 75/hour full throttle, a 3-hour burst = CAD 225. That's a real cost decision. Knowing burn rate per entity type lets us optimize: cheaper entities (Mercury distribution plans) vs. expensive ones (Veritas deep fact-checks with 80+ tool calls).

Deliverable

A burn ledger format and a weekly quota tracking report — same discipline as the CAD ledger, applied to compute.