v0.21.0

Decision models (opt-in, decision feature — off by default, no new dependencies)

Ask a small decision model typed questions — yes/no (Noul), Choice, Score — about a text or JSON state. Nothing is sent until you construct a DecisionModel; an API key in the environment enables nothing.

• Models: DecisionModel::jev() (TypeSafe), jev_opencode(), local(url) for self-hosted SystemOne servers, and logprobs(url, model) for any OpenAI-compatible endpoint that returns logprobs (vendor-neutral, open-weight models).
• Agent::with_decision_model: advisory skill and tool hints; failures add nothing (fail open).
• Agent::with_tool_gate(ToolGate::new(model)): denies tool calls judged destructive and not requested by the user. Fails closed.
• with_input_guard(InputGuard::new(model)): screens input for prompt injection or harmful content. Fails closed by default.
• Fallback chains (primary.or(backup)) with one overall deadline.
• calibrate: accuracy, Brier score, ECE, thresholds and a fitted temperature over a labelled set.
• Spend: decision spend is recorded in SessionStats::decision and counted in total_cost_usd.

Hooks (no feature needed)

• ToolCallRequest::new builders for unit-testing middleware.
• Structured user_request_parts().
• AsyncInputFilter and TurnHook.
• SubAgentTool::with_turn_hook and with_async_input_filter.

See CHANGELOG.md for the full list.