Product Decisions This Supports

Enforcing per-tenant, per-workspace, or per-user AI spend caps before a provider request is made, including multi-worker deployments that need a shared host account of capped spend.
Roadmap decisions around enterprise tiers, prepaid credits, usage-based billing, internal chargeback, and cost-center reporting for AI features.
Build-vs-buy: adopt this package for capped-call lifecycle semantics—admission, reservation, refusal, settlement, and continuation coverage—while building only the host-owned persistence, reconciliation, and expiry logic around it.
Use cases: customer-facing AI features with budget guardrails, agent/workflow runs that mix text, structured, streaming, and approval continuations, and support/finance workflows that need refusal codes and exact charged amounts.
Product observability and incident response: distinguish "budget refused" from provider failure, reconcile harness charges with host ledgers, and detect worker-death reservation leaks via host-side expiry.

When to Consider This Package

You make capped Prism/Harness model calls and need a host-controlled observer that can refuse calls before any provider request, and settle exactly once with the harness-charged amount.
You run multiple workers or processes and need a shared spend account that is authoritative on the host side.
You need coverage across send, sendStructured, stream, and approve/deny/decide continuations on the capped path.
You can implement persistence for reservations and settlements, and expire stale reservations when workers die between reserving() and settled().
You are comfortable injecting via RunContext::root(..., observer: $observer) or configuring prism-harness.cost_observer as a container-resolved class.
Look elsewhere or defer if: you do not use cost caps; you need uncapped concurrent calls with an observer (uncapped runs are refused with budget_observer_uncapped); you expect the harness to meter observer behavior (the observer is trusted host code); or you need out-of-box persistence rather than host-owned ledger storage.

Executives: v0.15.0 adds hard, pre-flight AI spend controls shared across workers, enabling predictable unit economics, tenant-level budget enforcement, and fewer bill-shock incidents. No breaking changes.
Engineering: Implement CallCostObserver with reserving() and settled(); throwing from reserving() refuses the call before any provider request, and settled() provides exact charged amounts for completed, failed, and abandoned-stream calls. Inject through RunContext or config. Key caveats: capped runs only, host owns persistence and stale-reservation expiry, and the observer is trusted host code.