Measuring and reading the chain¶

Measured, not assumed¶

What the relay knows of a chain, it measures, in one batch: its nodes' chain id, the fork, the deployer's code, each token's balanceOf gas, decimals and symbol, each x402 token's transfer, the payee's holdings. Nothing is read on start; the chain is measured when it's first needed, and again once measure_seconds old (an hour) when next needed, so an idle relay reads nothing, and a fork or a token's upgrade goes unnoticed at most that long. On the Worker, each wake measures afresh.

Degrade, never die. A token that can't be measured is suspended and the rest served; one whose x402 transfer fails is taken without x402. A chain whose nodes are another chain, whose fork no schedule knows, or whose deployer isn't its runtime is unavailable: nothing is quoted, taken or sent, and the treasury holds off, until it measures right. Nodes that don't answer keep the last measurements, tried again a minute later; with none yet, the chain is unavailable and its nodes aren't asked again for a minute. A node's errors stay in the log: what clients and the root see names no URL. The root lists what's wrong.

A simulation must deploy the runner: the deployer answers its address, 32 bytes; any other answer (an address without code answers nothing) means no runner ran and no payment was checked, and fails the simulation.

Clock and reads¶

Work only while there's work. A relay's clock runs while a submission is pending or a funding transaction is in flight: a submission taken starts it, and its batch closes at once. With none, it stops; it wakes only to forget finished submissions after retention_seconds, reading nothing, and a submission taken meanwhile starts it again at once, not when that falls due. Quotes, status and the configuration are answered on request. An idle relay reads nothing from its nodes.

A step a slot. Each step reads the latest block and the next base fee, in one request, and prices and schedules all its work by them: the block's work if there's a new block, the batch if it's due. The next step is at the next slot's start, its block looked for a second in, or when the batch falls due, whichever is sooner. Nothing to do reads nothing more: a batch with nothing queued, or nothing that still simulates, and a block with no batch to send. A step that fails, its nodes not answering, is tried again a second later, then twice as long after each failure in a row, up to five minutes.

A slot's work is a few requests, whatever the queue. Where each submission needs the same read (its planning reads, its simulation, its receipts) or each account does (balance and nonce), the reads go as one JSON-RPC batch, of at most 100 calls a request. A batch fails over to the next node whole; a node's error answer to one call fails only that submission. Closing a batch simulates every ranked submission at once, and those that no longer simulate are rejected, rather than simulating down the ranking until the accounts are filled. Measuring a chain is one request too (measured, not assumed), as is a token's price (every pool on its path), and a treasury's reads.