agentproto 0.5.0 adds credential brokering, sandboxes, honest cost tracking

agentproto 0.5.0 turns the project's agent-orchestration daemon into a platform you can hand real credentials to, run off your local machine, and actually measure. The release ships 37 packages, six of them new, organized around three themes: authentication, sandboxed execution, and observability.

The new @agentproto/auth package implements the AIP-50 spec end to end, with a CredentialBroker that refreshes tokens automatically and forwards secrets to child MCP servers as headers — never touching env vars, config files, or disk. A companion @agentproto/sandbox package defines a backend-agnostic sandbox lifecycle (AIP-36), with a first e2b-based implementation and GitHub-backed workspace sync, letting agents operate on real repositories without touching the host filesystem.

On observability, cost reporting now carries a four-way source tag — actual, computed, unpriced, or unmeasured — so dashboards can admit uncertainty instead of fabricating numbers. New role-based spawn controls, an eval harness with pluggable LLM-judge scoring, and OTel/Langfuse telemetry with secret redaction round out a release focused on making multi-agent systems auditable rather than just capable.