Usage and limits

The proposed /usage endpoint reports aggregates; /usage/records specifies the immutable ledger contract but still requires a public quantity-attribution adapter. All API operations remain proposed. Measured supplier cost, rated customer charge, and explicit adjustments are distinct quantities. An estimate is not a finalized invoice.

Mystro reserves worst-case cost before a call, settles measured usage once, and holds the reservation when the outcome is uncertain. If the control plane cannot confirm a budget, it denies new calls. Parallel runs cannot race past the shared limit.

Protective pilot defaults

Limit	Default
Active runs	1 per workspace, 2 per cell
Queue	100 per cell
Start deadline / active execution	5 minutes / 5 minutes
Agent steps	20
Input / output tokens	32,768 / 8,192
Per-run / monthly sandbox budget	$1 / $100
File / environment storage	10 MiB / 1 GiB

These dollar limits are protective budgets, not product prices. The proposed commercial model combines a capacity fee, measured execution, and storage. Prices and unit economics remain unvalidated until capped live cost tests. Billing starts with reviewed manual invoices.

Support uses sanitized request/run IDs. End-user content access requires an explicit time-bounded audited grant; a console role alone is insufficient.

Request and stream protection

The protected host's proposed defaults are 60 requests per second per cell with a burst of 120, plus at most 128 open SSE streams per cell and 16 per actor. Limits are independent of paid model budgets and the run queue. Treat 429 as backpressure and honor Retry-After when supplied; reuse the same mutation idempotency key after a retryable response. Live host verification is still a release gate.

Aggregate measurement boundaries

Cost and token totals use ledger entries recorded within the requested [from, to) interval. reserved_cost_usd is currently held escrow, including reservations for runs outside that interval; reservation_as_of identifies its snapshot time. Runtime is recorded active execution time for owned runs created in the requested interval. Storage is currently retained bytes owned by the caller, rather than historical byte-hours. These counters are not a billing tariff.

Customer charges remain null with rating_status: unrated until a reviewed rating policy exists. Held or unknown operations and explicitly configured fixed-dispatch tool cost policies keep aggregates provisional. Fixed-dispatch tool costs are policy estimates, not measured third-party invoices. Unknown outcomes retain their reservation until authenticated reconciliation.