// Project
Changelog
Release history and development roadmap. Uneven AI follows semantic versioning.
The first public npm release is v0.4.0 — April 11, 2026. Versions v0.1–v0.3 represent the internal development history shipped as part of that release.
- Rust core scaffold + napi-rs bridge
- TypeScript public API + type declarations
- Basic CLI: uneven-ai init, uneven-ai ask
- Local LLM inference via Candle (LLaMA 3.2 1B Q8)
- External brain providers: OpenAI, Claude, Gemini, Ollama
- Embeddings generation via Candle (1024-dim, L2 normalized)
- Vector store integration (usearch HNSW)
- File and directory indexing (Rust)
- Database connectors: PostgreSQL, MySQL, SQLite, MongoDB
- URL loader and docs scraping (undici + cheerio)
- Terminal watcher via tokio::process
- Stack trace error parser (TS, JS, Python, Rust, Go, Java, PHP, Ruby)
- Retrieval-Augmented Fix (RAF) — KB semantic search before pattern-match fallback
- Auto-fix engine via similar crate (surgical diffs)
- File watcher with auto re-indexing (notify)
- Pentester static mode (OWASP Top 10, secrets, CVEs, injections, headers)
- Pentester active mode with SHA-256 signed authorization scope
- Malware scanner: 8 categories, CI-compatible exit code (uneven scan)
- AI data analyst: natural language → SQL → Excel/dashboard (uneven analyze)
- CI/CD headless pipeline with GitHub Actions integration (uneven ci)
- Session state protocol (.uneven/session.json)
- Security report generator: HTML + Markdown output
- License system: free / Pro tiers with machine fingerprinting and offline grace period
- Process lock protocol — exclusive .uneven/uneven.lock prevents concurrent Uneven instances from corrupting state
- Atomic index saves — crash during indexing no longer corrupts the previous state
- Logger write queue — serializes all log writes to prevent race conditions
- Throttled security scanners — pentest runs at most 3 concurrent filesystem walkers (was 8)
- Timeout kill switch — all LLM inference, DB queries, web fetches and git operations have hard deadlines
- Index preview — cost and time estimate per brain provider before committing to indexing
- User-selectable DB tables for data analysis + LLM-suggested analysis ideas
- Excel folder reading for data analysis (uneven analyze)
- Humanized CLI help screen — plain-language descriptions for every command
- Hardware Acceleration: Automatic GPU/CUDA offloading (32 layers max) for lightning-fast local inference
- Ethical Guardrails: Built-in defensive system that prevents malware, exploit, and cracking generation
- Advanced Document Intelligence: Native support for PDF, Word (.docx), Excel (.xlsx), and CSV Pro (industrial parsing)
- Strategic AI Recommendations: Preferred path for Local AI and Google Gemini Flash for privacy and cost-efficiency
- System Diagnostics: New `uneven info` command for real-time hardware, VRAM, and version auditing
- Memory Hardening: Silent model auto-unload (10min) and anti-loop safety (5-try ceiling)
- Local Data Analyst: Direct natural language queries over local CSV and Excel files without DB requirements
- Stable public release — production-ready across all core features
- Streaming AI responses — output appears token by token instead of waiting for the full answer
- Human-centric fix approval — alertOnly mode suggests fixes for your review before applying anything
- Dynamic autofix toggle — switch between autonomous and assistant mode mid-session with --autofix
- Faster knowledge base — significantly smaller index files and faster search queries
- SBOM generation (uneven sbom) — export a software bill of materials for your project
- Improved provider support — updated model defaults and correct API endpoints for all external providers
- Conversational shell — run uneven-ai with no arguments to open a natural-language interface in any language; available on the free tier
- GPU auto-upgrade — uneven-ai init detects your GPU (NVIDIA CUDA, Apple Metal) and installs the optimised binary automatically; no manual steps or recompilation
- New commands: explain <file> (explain any file in plain language), docs <file> (generate Markdown documentation for a file), review (AI code review — Pro), askf <task> (AI writes or edits files — Pro)
- Free-tier conversational responses — shell responses respect your configured token limit; bring your own API key and chat with no query caps
- Background task concurrency limit — shell safely queues concurrent background jobs to prevent resource exhaustion
- Local model stability — concurrent inference requests are now handled correctly; timed-out requests no longer produce partial output
- Config write safety — configuration changes are now written atomically to prevent data loss on unexpected shutdown
- Race condition fixes — concurrent knowledge base saves and session state operations are now correctly serialised
- Attacker Attribution: Active pentest probes now silently transmit the origin IP for trace purposes
- Threat Detection: Terminal watcher identifies incoming Uneven probes and alerts the administrator
- License Hardening: Active Pentester mode now requires a Team tier license
- Shell Polish: Improved Portuguese intent classification and fixed false success log notifications
- 3-Layer Privacy Protection: Enhanced data safety suite with Schema Filtering, Semantic SQL Auditing, and Live Result Redaction
- Forensic Audit Logs: Detailed local history of AI actions and system fixes stored in structured Markdown format
- Autonomous Security Pentester: Validated engine for detecting hardcoded secrets, injections, and infrastructure header flaws
- Universal Data Context: Seamlessly index and query .csv and .xlsx files for automated data processing tasks