What shipped this week.

Durable tool calls, resumable runs and step-level traces for long-running agents.

Compare every case across up to eight models in one grid, with regressions highlighted.

Set hard monthly spend caps per project. The router downgrades gracefully instead of failing.

Traces now appear in under 300 ms at p99, even for agents with 200+ steps.

PII redaction no longer drops characters on multi-byte streamed tokens.

Typed, async-first and fully compatible with the TypeScript SDK. pip install prompt-zero.