Command and read surfaces
Use the authenticated HTTP API for client commands and queries. The API separates commands from reads. Each client operation requires an authorization action and runs in the caller’s tenant. Follow the operation’s idempotency requirements. Retries after timeouts are normal and must reuse the same business identity. A mounted operation is not proof that a specific deployment wires its Dataprev adapter. An unavailable rail integration returns501 Not Implemented for that request instead of disappearing from the API. The capabilities operation reports which operation families your deployment reaches.
The Lender command stream remains available for the supported integration commands. It is not a replacement for the complete HTTP surface. See Consignado events.
Scheduled workers
The gateway runs fourteen scheduled loops. They share delayed first execution, stable per-replica staggering, panic isolation, and graceful shutdown. Every loop except the auction fast lane also runs non-overlapping passes with positive jitter.
Eleven of the fourteen take their cadence from environment configuration and need a restart to change. The auction fast lane, bid payload cleanup, and proposal witness read their cadence through the Systemplane and take effect on the next delay.
Averbação recovery and rail-command recovery exist because a rail command can leave its outcome unestablished: the process died mid-call, or the rail never answered. They converge that state against the rail’s own registry instead of losing it. Bid reconduction resends a bid whose rail answer was lost, under its original proposal number, and each resend uses the tenant’s rail budget. The three loops are off by default. Enable the two recovery loops in any deployment that submits averbações or rail commands.
Once a month, Dataprev closes for its competência turnover and refuses requests. When a background loop meets that refusal, the loops that poll Dataprev for that tenant pause for 5 minutes, then 15, then 30 for each later refusal. Client requests still reach Dataprev and receive its refusal. The scheduled-jobs report shows
paused_until for a paused loop. In multi-tenant mode, the gateway publishes consignado.rail.paused once per turnover and consignado.rail.resumed when Dataprev answers again.
Rail windows
Rail defaults and constraints include:- a 24-hour loan-auction validity window that Dataprev supplies and the gateway does not configure
- a 72-hour portability-request lookback
- 20 calendar days from the origin exclusion to register an inbound portability, when the request names the exclusion date
- 7 business days to reverse refinancing
- a 48-hour auction-discovery window, with a five-minute clock-skew allowance
Throughput and pacing
Dataprev traffic is paced per tenant. The throughput API sets the tenant’s rate, from 0 to 25 requests per second, and no setting raises that ceiling. A tenant with no stored rate runs at 3 requests per second. A stored value of0 pauses rail traffic for that tenant.
Reconciliation safety
Bookkeeping and repasse reads are independent. One failing source must not suppress the other. Expected repasses can emit three states:consignado.repasse.overdue: the expected movement is lateconsignado.repasse.resolved: the movement arrived after an overdue alarmconsignado.repasse.unobserved: the gateway cannot currently prove presence or absence
fresh.

