Tags: voidstackloop/modelforge
Tags
v1.3.0 — ModelForge Medical Supersedes the original v1.3.0 tag (2026-08-03), which was never actually published as a GitHub Release. This tag now points at the complete, current state of the app — everything below, not just the original highlights. This release moves the project from a general-purpose local AI chat/agent client to a medical-focused clinical workspace. See the notice at the top of README.md and its Product boundary section for what that does and doesn't mean (never an autonomous diagnostician or prescriber, no regulatory certification claimed). Original v1.3.0 highlights (2026-08-03): - Clinical layer: Patient Cases, Evidence Library, per-case Knowledge Graph, and an Audit & Privacy log, plus a structured 8-section response contract and deterministic safety checks (emergency red-flag detection, allergy/medication conflict warnings, transmission preview, redaction) that run independently of the model. - MCP client rework on the official SDK: protocol version negotiation, AJV schema validation, resources/prompts, OAuth 2.1+PKCE, per-tool trust profiles, and a DICOM MCP integration with an enforced tool denylist. - Privacy/security: passphrase-based encryption at rest for patient cases and chat sessions, automatic session locking, configurable audit retention, and a SHA-256 hash-chained tamper-evident audit log. - Approved model registry, clinical-note review sign-off, structured consent records, and a swappable MedicationSafetyProvider. - Rust-backed data store I/O (lib/, napi-rs): atomic JSON read/write and audit-log hashing/appends moved off the O(n) Node path. Added since then: - Ollama removed entirely in favor of llama.cpp (tool-calling, embeddings, model-catalog rebuild). - Enterprise integration: OIDC-connected shared backend for organization membership, patient cases, and chat sessions with staged local-to-shared migration and an encrypted offline cache/outbox; signed central policy enforcement; tenant/cloud backup and restore; a compute-control-plane fleet agent (self-signed mTLS node identity, heartbeat/lease lifecycle, local capacity reservation) with its own Settings UI panel. - Local-inference hardening: a cross-workload resource orchestrator (per-device-id accelerator reservation, atomic multi-GPU groups, pressure-aware admission), a crash-loop breaker for the llama.cpp GPU backend and for Electron's own GPU-acceleration/sandbox subsystems (both stay on by default; only this machine's own crash history ever downgrades them), and a real bound on the manual GPU-layer-count setting. - Clinical safety and platform hardening: medication-safety provider registry, chat-history sanitization, OHIF imaging viewer integration, report-only CSP, IPC sender/frame validation, MasterVault-backed encryption, managed MCP policy enforcement, and GGUF model/download verification. - Experimental opt-in SQLite audit-log backend (retention/cap enforcement, migrate-in/merge-back between backends). - A companion multi-tenant IAM/compute-control-plane server and admin console (server/, admin-console/) for organizations that want shared, authenticated case access instead of this app's default local-only mode — deployed and versioned separately from this desktop release; see docs/SHARED_BACKEND_DESIGN.md. - CI/release pipeline hardening: fixed a missing shared-package build step that silently broke lint/typecheck/test across four CI jobs and the release build itself, real Postgres-backed test deadlocks, a SQL timestamp-formatting bug, and cross-platform (Windows/macOS) test failures — including real, low-CPU-capacity resource-admission failures on GitHub's macOS runners — never caught before this release's first real 3-OS runs. Co-Authored-By: Claude Sonnet 5 <[email protected]>