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.

Version updated for https://github.com/Vishisht16/Humane-Proxy to version v0.5.6. 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 rolls up the 0.5.x bug-fix series (0.5.1-0.5.6) — the result of a full audit of the codebase against its documentation. Every fix ships with regression tests: the suite grew from 298 to 353 tests, with zero skips and zero expected failures.

Version updated for https://github.com/Vishisht16/Humane-Proxy to version v0.5.0. 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 Highlights 🔭 OpenTelemetry distributed tracing (#7) — export pipeline traces to Jaeger, Grafana Tempo, or Datadog via the new [telemetry] extra. Every request produces a pipeline.classify span with privacy-safe attributes (category, score, stage reached, message hash — never raw text). Zero overhead when disabled. 🌍 Region-aware care response — set safety.categories.self_harm.region: "IN" (any ISO country code) to surface that country’s crisis resources first, while always keeping the full international list. ☎️ Crisis helplines for new countries (#26) — Japan, South Korea, Spain, Italy, Mexico, New Zealand and more join the existing US/IN/GB/AU/CA/DE/FR/BR/ZA resources. 🗄️ Storage-backend consistency (#37) — the admin API, CLI, and MCP tools now route through the storage factory instead of raw SQLite, so Redis and PostgreSQL deployments see the same data everywhere. 📊 Query upgrades — escalation queries support date filtering (date_from/date_to) and sorting across all three backends. Fixed Multimodal content arrays: /chat extracts text from OpenAI-style content parts and no longer errors on string or malformed message content (#44, #45, #48). Empty /chat request bodies return a clean 400 (#41). CodeQL SQL-injection alerts resolved with statically generated SQL templates; timezone conversion and CSV triggers serialization fixed in escalation export. Security Admin Bearer-token comparison hardened against timing attacks with hmac.compare_digest (#18, #21). HTTP MCP hardened: binds to 127.0.0.1 by default, warns on public binds without a token, supports Bearer auth via HUMANE_PROXY_ADMIN_KEY, and bounds audit-log queries (#17). Docs & Tests Mermaid.js architecture diagram, README table of contents, corrected Stage-1 transition labels. New test coverage: Unicode/leet-speak heuristic edge cases, pipeline config validation, malformed payloads, decay-weighted-mean edge cases. Full Changelog: https://github.com/Vishisht16/Humane-Proxy/compare/v0.4.0...v0.5.0