HumaneProxy Safety Benchmark

Version updated for https://github.com/Vishisht16/Humane-Proxy to version v0.6.0.

  • This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

What’s Changed

This release is about doing the same work faster and scaling it across workers—three independent performance wins, each behind an opt-in extra with a clean fallback.

Added

  • ONNX Runtime Stage-2 backend (stage2.backend, env HUMANE_PROXY_STAGE2_BACKEND): Stage 2 can now run the embedding model on its pre-exported ONNX graph via a new onnx extra (onnxruntime + tokenizers + huggingface_hub)—no PyTorch, roughly 2 GB lighter to install, and faster on CPU. The default "auto" prefers ONNX when installed and falls back to sentence-transformers; both produce numerically equivalent embeddings, verified by an equivalence test suite (including long-input truncation parity). Model, anchor, and result caches are keyed per backend.

  • Redis-backed trajectory analysis (trajectory.backend: redis, env HUMANE_PROXY_TRAJECTORY_BACKEND): Session risk windows move into Redis sorted sets with an atomic Lua read-append-trim, so every uvicorn worker shares one consistent view of a session’s trajectory (the in-memory default tracks per process). Reuses storage.redis.url unless trajectory.redis.url is set. Sessions auto-expire via TTL (default 2× the decay half-life), fall back to in-memory tracking with a logged warning when Redis is unavailable, and DELETE /admin/sessions/{id} now erases the Redis trajectory keys as well.

Changed

  • orjson JSON fast path (new optional perf extra): When installed, the proxy serializes every /chat response through an orjson-backed response class. Storage trigger serialization, Stage-3 response parsing, and integration tool output also route through an internal shim. Without the extra, everything falls back to the standard library json module with identical behavior.

Install

pip install humane-proxy[onnx]     # Stage 2 without PyTorch
pip install humane-proxy[perf]     # orjson fast path
pip install humane-proxy[redis]    # shared trajectory + storage
pip install humane-proxy[all]      # everything

Full Changelog: https://github.com/Vishisht16/Humane-Proxy/compare/v0.5.6...v0.6.0