Skip to content
Trazum

1.45.0

The agent's budget

Added#

spend_guard: the thing spending the money can finally ask, and be told no. The MCP server told an agent what a prompt costs; it could not tell it whether it was allowed to spend, which is the only answer that changes what a model does next. Fifth of the ten in docs/plan-1.41-1.50.md.

A refusal never arrives bare. An agent told "denied" and nothing else has two moves — send it anyway, or fail the user's request — and both are worse than the call it wanted to make. Every no carries the cheaper ways to make the same call: a smaller model, a batch window, or both. Each is priced for this call rather than for a month, because the caller is deciding one call right now and a monthly figure is the right number at the wrong moment. Each names what it assumes, typed as everywhere since 1.38. Alternatives appear on a yes too: an agent allowed to spend that could spend less should be told.

An alternative the prompt does not fit in is not an alternative. A cheaper model with a smaller context window does not make the call cheaper — it makes it impossible. Those are filtered before they are offered rather than offered and blamed later, and the surviving ones carry fits: true so the rule lives in the type rather than in a filter nobody reads.

Route and batch combine, never add. The batch discount applies to the cheaper model's price, as billLevers has done since 1.23 — the head arithmetic plan exists to kill does not get to reappear at the tool surface.

It never spends to answer, and never says yes to what it cannot judge. No provider call, no model call, no pull — a cost guard that costs money to consult is a joke with a bill attached. And cannot-tell stays cannot-tell: a guard that permits whatever it cannot see permits everything the moment a figure goes missing.

New core module guard.ts (guardSpend), and spend_guard on the MCP server beside the four tools that were already there.

Security#

A tool may not cause spending. An agent that could trigger a provider pull by asking a question is a denial-of-service with good manners, and the bill lands on whoever installed the cost tool. A guard fails the build if the MCP tool surface reaches the connector's fetch path, the node entry point or the filesystem — proven with a planted probe. The exact-set tools test carries the review that admitted spend_guard, as it has done for profile_usage.

Documentation#

The cost answer document is contracted at last. serve's POST /cost response should have been in docs/json-output.md when it shipped in 1.44 and was not — the one machine-readable output consumers build against hardest, undocumented. It is there now, with the spend-guard document beside it and a two-direction parity test over both.

RELEASES.md no longer states a published version nothing checks. Its header claimed 1.28.0 was on npm while the manifests moved through 1.45.0 — seventeen releases of a stranger being told the wrong thing about what npm install gives them. publish.test.js now asserts that number equals the version the manifests publish, which is checkable because the merge that bumps the manifests is the merge that pushes the tag release.yml publishes on. Proven by planting 1.28.0 back and confirming the failure names both numbers.