Conduct LiteLLM Guard 0.2.5 Adds Runtime Governance for AI Agent Calls

Summary

Conduct AI released conduct-litellm-guard 0.2.5, a Python package that connects Conduct Guard to LiteLLM as a runtime guardrail for AI-agent and LLM requests. Every request routed through a configured LiteLLM proxy is checked against the active Conduct policy before it leaves the network. The guard can allow or advise on a request, forward it with a warning, block it, or hold it pending human-in-the-loop approval; blocked requests do not consume an upstream model token. Each check sends Conduct a compact request summary containing fields such as the model, call type, message count, temperature, stream setting, and the latest user message, while Conduct keeps the signed configuration and hash-chained audit log. The package supports an explicit adapter configuration across LiteLLM versions and documents a native configuration that depends on a stated upstream LiteLLM change. It tracks sessions using LiteLLM metadata, an explicit Conduct session header, an adapter-generated session identifier, or a deterministic hash of user content. The default `fail_closed` mode blocks calls when the Guard is unreachable or the policy evaluator fails; `fail_open` instead allows the call and records a warning. Version 0.2.5 renames `fail_mode` to `unreachable_fallback`, retaining the old name with a deprecation warning until its planned removal in 0.3.0. The package requires Python 3.10 or newer and is distributed under Apache 2.0.