The time envelope (Mode A)

Introduction

The spend policy caps amount and total. It can also bound time — and that is what makes Mode A work: a headless agent runs free inside a budget and a time envelope, and the policy is the consent, with no per-payment approval prompt. It’s a budget that is also a clock.

Two independent leashes, both opt-in. Omit them and behaviour is byte-identical to a time-free policy.

import { PipRailClient } from '@piprail/sdk'

const client = new PipRailClient({
  chain: 'base',
  wallet: { privateKey: process.env.AGENT_KEY! },
  policy: {
    maxAmount: '0.10', maxTotal: '5.00', tokens: ['USDC'],
    ttlSeconds: 3600,                          // the whole session expires in 1h
    windowTotal: '1.00', windowSeconds: 600,   // …and ≤ 1.00 USDC per rolling 10 minutes
  },
})

The session deadline — ttlSeconds / expiresAt

A session deadline is a hard stop. Past it, every payment is refused — even a zero-amount or under-cap one — because expiry is amount-blind and checked first. It’s a headless agent’s time leash: spend dies when the clock runs out.

policy: { ttlSeconds: 3600 }   // dies 1 hour after client construction

The deadline is relative to session start, which is when you construct the client. State is in-memory and resets on restart — the session is the process. For crash-loop-resistant limits, supply a pluggable durable store (the isUsed / markUsed analogue).

expiresAt sets the same deadline absolutely, as epoch milliseconds (matching Date.now()):

policy: { expiresAt: Date.now() + 3_600_000 }   // SDK-only; pass milliseconds

Field	Units	Notes
ttlSeconds	seconds, relative to session start	Positive safe integer; also the MCP’s PIPRAIL_TTL knob.
expiresAt	epoch milliseconds (absolute)	SDK-only — no MCP env knob. Not auto-corrected from seconds.

When both are set, the earlier deadline wins.

Caution

expiresAt is milliseconds, not seconds, and is not auto-corrected. A small value expires immediately by design — expiresAt: 3600 is the year 1970, not “an hour from now”. Use ttlSeconds for a relative leash, or Date.now() + ms for an absolute one.

A payment past the deadline throws PaymentDeclinedError with reasonCode: 'SESSION_EXPIRED', before any funds move. This is terminal — don’t retry: the session is over, and the next payment will be refused too.

import { PipRailClient, PaymentDeclinedError } from '@piprail/sdk'

const client = new PipRailClient({
  chain: 'base',
  wallet: { privateKey: process.env.AGENT_KEY! },
  policy: { ttlSeconds: 3600 },
})
const url = 'https://api.example.com/report'

try {
  await client.fetch(url)
} catch (e) {
  if (e instanceof PaymentDeclinedError && e.reasonCode === 'SESSION_EXPIRED') {
    // terminal — the session's time leash has elapsed; start a fresh client
  } else {
    throw e
  }
}

The rolling window — windowTotal / windowSeconds

The window is a rate limit layered on top of the lifetime maxTotal: at most windowTotal of a given asset may be spent within any trailing windowSeconds. It’s enforced per (network, asset) — the same per-distinct-asset scoping as maxTotal, because summing across tokens is unit-meaningless without a price oracle (which the SDK deliberately omits).

policy: { windowTotal: '1.00', windowSeconds: 600 }   // ≤ 1.00 USDC per rolling 10 min

A payment that would push spend within the last windowSeconds past the cap throws PaymentDeclinedError with reasonCode: 'OUTSIDE_WINDOW'. Unlike a session expiry, this is recoverable — wait for the window to free up (older spend ages out), or raise windowTotal.

Caution

The window needs both windowTotal and windowSeconds, or neither. Setting one without the other is a config error and throws a TypeError at construction — a security leash can’t be silently half-armed. windowSeconds must be a positive integer.

Read the leash before paying — client.budget()

budget() returns a read-only SessionBudget: the time envelope plus the per-asset money leash. A Mode A agent reads this to see its remaining runway rather than discovering the limit by hitting a decline.

const b = client.budget()
// → {
//     session: { start, expiresAt, secondsRemaining },  // ISO strings; null fields when no deadline
//     byAsset: [ /* one SpendRemaining row per (network, asset) the ledger has seen */ ],
//   }

client.remaining() returns just the byAsset half — the per-(network, asset) remaining cap, as a SpendRemaining[]. A pair only appears once it’s been spent on (the SDK learns the token’s decimals from the first payment), so a never-spent asset simply isn’t a row, and a fresh client with a maxTotal returns [] until its first payment.

The time envelope also surfaces on every PaymentPlan as plan.session — present only when a deadline is configured — so the same plan you check for funds also shows the clock:

const plan = await client.planPayment(url)
if (plan?.session) {
  console.log(plan.session.secondsRemaining)   // number | null — seconds left on the session
}

How the guards interleave

Every payment is evaluated against the whole policy in a pinned, first-failure-wins order, so the refusal is specific. Session expiry runs first because it’s session-global, not asset-scoped — an expired session always reports expiry, never some other gate that happens to also fail.

session expiry → chains → hosts → unknown-token → tokens → maxAmount → maxTotal → window

Both time guards collapse to a single decline reasonCode on the client, distinct from the amount-based codes:

Configured by	Decline reasonCode	Retry?
ttlSeconds / expiresAt	SESSION_EXPIRED	No — terminal.
windowTotal + windowSeconds	OUTSIDE_WINDOW	Yes — once the window frees, or raise the cap.

Note

That reasonCode is the coarse code a catch sees on PaymentDeclinedError (a DeclineReasonCode). The read-only quote.policyCode (a finer PolicyDenyCode) names the exact guard: SESSION_EXPIRED passes through unchanged, while the rolling window’s WINDOW_TOTAL maps to OUTSIDE_WINDOW.

See Evaluate policy for the pure decision function behind these checks, and Why payments fail for branching on every reasonCode. The MCP exposes the relative PIPRAIL_TTL knob and a piprail_budget tool — see MCP modes.