A maxUsd budget refuses every sandbox child because the tangle provider path reports unknown dollar cost

[drewstone]

What happened

A pursuit registered with supervise.budget = { maxIterations: 960, maxTokens: 160000000, maxUsd: 300 } on runtime 0.225.5 lost 3 of its first 4 children at settlement:

budget pool: ticket 1 reported unknown dollar cost under a dollar-capped budget (…:s1)
budget pool: ticket 2 reported unknown dollar cost under a dollar-capped budget (…:s2)
budget pool: ticket 3 reported unknown dollar cost under a dollar-capped budget (…:s3)

Run: capability-per-parameter-cpp-ds-20260915b (records retained). The children were ordinary literature-graph workers on tangle-sandbox; their settled spent reads usd: 0, usdKnown: false, as every sandbox child's did on the earlier runs too (usd=$0.00 billed=unmeasured in disco report).

The pool is right; the reporting is the defect

Refusing to count unknown as zero under a capped budget is correct and should stay. The problem is upstream: the sandbox/provider path never reports dollar cost for a child, although the cost is on the wire. The Tangle router returns usage.cost on every completion (a director's own call log shows non-zero usage.cost per call), and run-level totals are only ever provenance: estimated.

So today a dollar cap is unusable with the fleet, which means the only enforced ceiling is the token pool — and that pool counts re-sent context on every turn, so it does not bound money either.

Ask

Carry the router's usage.cost from the provider result into the child's spent.usd with usdKnown: true (agent-provider-tangle reports usage tokens today; the cost field is beside them in the same response). Where a route genuinely has no cost, keep usdKnown: false and let the cap refuse — that is the pool doing its job.

Until then, maxUsd should be documented as incompatible with sandbox children, because the failure mode is every child dying with a pool error that reads like an accounting bug.

🤖 Generated with Claude Code