Release notes

​

Features

• User-facing message field: LLM violations now include a dedicated message field — under 15 words, written directly to the user, with no rule references. This replaces analysis as the string shown in the terminal.
• analysis in JSON output: The analysis field (internal model reasoning) is now included in --output=json violation objects alongside message, giving consumers access to both.

​

Improvements

• Directive guidance updated to explicitly distinguish message (user-facing) from analysis (internal reasoning not shown in UI)

​

Features

• PAT (Pay A Tax) Evaluation Technique: New gate-check evaluation method using structured output for more reliable LLM-based assessments
• Amazon Bedrock Provider: Run VectorLint against AWS-hosted models via the Amazon Bedrock LLM provider
• Vercel AI SDK Integration: Unified LLM provider interface via Vercel AI SDK, enabling consistent behavior across all providers
• PAT Confidence Tuning: Expose confidence thresholds via CLI to tune PAT filter sensitivity
• File-Type Context: Evaluations now receive file-type context to improve rule applicability

​

Improvements

• Aligned CLI surfacing with PAT filters for consistent flag exposure
• Tightened PAT confidence guidance for more predictable gate-check behavior

​

Features

• Zero-Config Mode: Run VectorLint without any config file by adding a VECTORLINT.md to your project root
• User Instructions Support: Global user instructions now run alongside all rules (previously fallback-only)
• Fix Suggestions: New fix field in violation output provides drop-in replacement suggestions for both Check and Judge evaluations
• Init Command Enhancements: New --quick flag creates only VECTORLINT.md, --full flag creates both config and user instruction files
• Improved Evaluation Instructions: Centralized directive with structured Role/Task/Instructions format

​

Improvements

• Renamed internal terminology: “Style Guide” → “User Instructions” for clarity
• Renamed evaluation types: SubjectiveResult → JudgeResult, SemiObjectiveResult → CheckResult
• Removed redundant instruction sections from bundled preset prompts
• Consolidated architectural documentation into CLAUDE.md

​

Features

• Renamed Rule Types: Rule types renamed for clarity — semi-objective → check, subjective → judge. Backward compatibility maintained via automatic mapping.
• Content Chunking: Large documents are now automatically chunked for improved evaluation accuracy
• Bundled Rule Packs (Presets): Built-in rule packs now ship with VectorLint — no custom rules path required to get started
• Global Configuration: New global config file for storing API keys and environment settings
• Init Command: New vectorlint init command for easy project setup
• Token Usage Reporting: CLI now displays input/output token counts after each evaluation
• Hidden Config Support: Configuration can now be stored as a hidden file
• Disable Rules: Support for disabling specific rules within a rule pack
• Improved Line Accuracy: Better issue location with line numbering and fuzzy matching

​

Improvements

• Refactored help output for clearer CLI usage
• Updated documentation to reflect optional rules path
• Improved Anthropic provider configuration and type safety
• Simplified evaluator token usage tracking

​

Bug Fixes

• Fix CLI to require explicit paths, preventing accidental directory scans
• Fix path option description to indicate it’s required

​

Bug Fixes

• Fix Runtime Crash: Resolved Cannot find module error in published package by inlining package.json version
• Replaced relative runtime require with static import to prevent path resolution issues in dist artifacts

​

Features

• Reviewdog Support: Added rdjson (Reviewdog Diagnostic JSON Format) output capability
• Enables automated code review comments and annotations in CI/CD pipelines
• CLI now accepts --output rdjson flag

​

Features

• Cascading Configuration: Configuration now applies from general to specific patterns
• Files can inherit rules from multiple matching patterns instead of just one
• More specific patterns override settings from general patterns

​

Improvements

• Updated documentation with configuration examples

​

Bug Fixes

• Fix configuration resolution to properly handle multiple matching patterns

​

🚀 Highlights

• Multiple LLM Providers: OpenAI, Anthropic Claude, Azure OpenAI, Google Gemini
• Smart Scoring: Density-based scoring (errors per 100 words) for fair evaluation
• Vale Integration: Seamless integration with Vale CLI for combined analysis
• Technical Accuracy: Hallucination detection with search provider support
• Flexible Config: File-centric rule packs with glob-based targeting
• Multiple Outputs: Native JSON and Vale-compatible JSON formats
• MDX Support: Extended support for React-based documentation

​

📦 Installation

npm install -g vectorlint

​

Contributors

• @ayo6706 made their first contribution in https://github.com/TRocket-Labs/vectorlint/pull/3
• @oluwatooki-GA made their first contribution in https://github.com/TRocket-Labs/vectorlint/pull/10
• @oshorefueled made their first contribution in https://github.com/TRocket-Labs/vectorlint/pull/11
• @hurshore made their first contribution in https://github.com/TRocket-Labs/vectorlint/pull/23
• @mtgr18977 made their first contribution in https://github.com/TRocket-Labs/vectorlint/pull/26