Resellers as upstreams

Fixture onlyChecked 16 November 2026

In short

  • A reseller is an upstream with its own cost and daily capacity.
  • You pay its cost plus the savings fee, never above list.
  • Only fixture resellers on this machine.

Fixture only. A reseller can only be a fixture server on this machine: the reseller transport refuses every host that is not loopback. Pointing one at a real reseller is a live-spend change and is blocked by AGENTS.md.

A reseller is any service that sells Claude tokens in the Anthropic Messages format or the OpenAI Chat Completions format (Surplus-style token markets, OpenRouter-style resellers). Each one becomes an ordinary upstream: it goes through the same routing, capacity holds, failover, stale-hold sweeper and 2% seller re-checks as the direct upstreams.

Config

Set PETROVA_RESELLERS_FILE to a JSON file, and each reseller's key in the variable it names:

{
  "resellers": [
    {
      "name": "surplus-a",
      "format": "anthropic",
      "base_url": "http://127.0.0.1:9101",
      "key_env": "SURPLUS_A_KEY",
      "trust": "standard",
      "capacity_per_day_nusd": "50000000000",
      "models": {
        "claude-sonnet-5-5": {
          "upstream_model": "claude-sonnet-5-5",
          "input_per_mtok_nusd": "1000000000",
          "output_per_mtok_nusd": "5000000000",
          "cache_read_per_mtok_nusd": "50000000",
          "cache_write_5m_per_mtok_nusd": "1250000000",
          "cache_write_1h_per_mtok_nusd": "2000000000"
        }
      }
    }
  ]
}
  • format: anthropic (POST <base_url>/v1/messages, key as x-api-key) or openai (POST <base_url>/chat/completions, key as a bearer token; translated both ways like Venice).
  • base_url: https, or http on loopback.
  • models: keyed by the catalog model id. The five prices are our cost per million tokens, in nano-USD, each above zero. upstream_model is the id the reseller uses (default: the same).
  • capacity_per_day_nusd: how much of the reseller's cost we may use per UTC day. Today's capacity row is created at that limit the first time it is needed; an operator can still change a day's row with setSupplierCapacity.
  • trust: high or standard resellers are tried before direct; low ones only after direct fails. Re-check sampling: 10% for low, 2% for standard and high.

The key is held in memory only and never logged.

Money

  • The customer pays list. The charge is the existing catalog price for the call's usage plus Petrova's fee (usageCostNusd + feeNusd), as on the direct route. The quote and hold are unchanged.
  • Pay per call. There is no prepaid inventory. The reseller's cost for exactly this call is booked as supplier_payable and written in reseller_calls (list, cost, charged, margin). The settlement's supplier_nusd is that cost, so reconciliation against the reseller's invoice works as usual. The margin is booked as platform_revenue.
  • Never below cost, refused at startup. If any configured price is above our list price for that token category (the model's cheapest tier), the server does not start.
  • Not chosen when the margin is too thin. Per call, a reseller is skipped when its cost ceiling for the request plus our fee is above the list ceiling for the same request. A model the reseller does not sell is skipped too.
  • Capacity is held at the reseller's own cost ceiling and settled at its actual cost.

Failover

Order is [high/standard resellers] → [direct upstreams] → [low-trust resellers], with session affinity first as before. A reseller that answers 429/5xx or an overloaded stream event before content is skipped, its capacity freed, and the next upstream tried. A reseller out of capacity for the day is skipped. A suspended reseller (a failed seller re-check) is never routed to, and its charged calls in the last 24 hours are refunded.

Tests

tests/fixtures/reseller-server.ts starts one Anthropic-format and one OpenAI-format fixture reseller on loopback. tests/integration/resellers.test.ts and tests/unit/resellers.test.ts cover pricing, refusal at startup, routing, failover both ways, capacity, re-checks and reconciliation.

Last checked against the code: 16 November 2026. Something wrong? The source of this page is in the repository, so a fix is one edit.