v0.2.5 - TOS Evaluator

Release Notes - v0.2.5

Added

• New skill: compliance/tos_evaluator
  • Local-first website policy evaluation before automated access.
  • Checks robots.txt permissions for the target URL/path and user-agent.
  • Discovers candidate legal pages (Terms, Legal, Acceptable Use, API-related links).
  • Extracts and evaluates policy clauses relevant to automated behaviors (scraping, crawling, indexing, monitoring, downloading, API use).
  • Returns structured verdicts: SAFE, UNSAFE, CAUTION, INSUFFICIENT_EVIDENCE.
  • Includes evidence payloads and recommended next-step guidance.
  • Optional low-cost LLM fallback for ambiguous clauses (Gemini model shown as example; interface remains provider/model configurable).

Skill Package Contents

• Added skills/compliance/tos_evaluator/:
  • manifest.yaml
  • skill.py
  • instructions.md
  • card.json
  • __init__.py
  • test_skill.py

Tests

• Added focused central tests:
  • tests/skills/compliance/test_tos_evaluator.py
• Added local skill test:
  • skills/compliance/tos_evaluator/test_skill.py

Documentation

• Added skill documentation:
  • docs/skills/tos_evaluator.md
• Updated skill catalog:
  • docs/skills/README.md (Compliance section now includes compliance/tos_evaluator)

Examples

• Added integration examples:
  • examples/gemini_tos_evaluator.py
  • examples/claude_tos_evaluator.py
  • examples/ollama_tos_evaluator.py

Dependencies

• Added HTML parsing dependency for deterministic policy extraction:
  • beautifulsoup4 (project dependency)
• Skill manifest requirement uses import-resolvable package name:
  • bs4

Packaging

• Bumped package version:
  • pyproject.toml: 0.2.4 -> 0.2.5

v0.2.4 - MiCA Compliance Module and PII Masking Implementation

This release introduces major compliance capabilities and framework optimizations focused on high-performance agentic workflows.

Key Changes:

1. MiCA Compliance Module: Added the compliance/mica_module skill, featuring in-memory caching for ultra-low latency RAG (~1.7ms) and a weighted surgical router to prevent context window asphyxiation.
2. PII Masking: Integrated the compliance/pii_masker skill for high-precision, local privacy guardrails using micro-f1-masking.
3. Pure Cognitive Framework: Realigned all MiCA examples (Gemini, Claude, Ollama) to follow a prompt-based cognitive pattern that avoids opaque native tool-calling obstacles.
4. Documentation: Comprehensive documentation updates for the new Compliance category and a refined core README.
5. Quality Engineering: Resolved all PEP 8 and Flake8 violations across the registry and verified execution with 100% unit test success.
6. Versioning: Framework version bumped to 0.2.4.

This release ensures that Skillware agents can operate in highly regulated environments with deterministic precision and local privacy.

v.0.2.3

Release Overview: v0.2.3

This release introduces the compliance/pii_masker component into the Skillware framework, providing a strict edge-boundary "Privacy Firewall" to prevent inadvertent data leakage within agentic workflows.

New Features

• Zero-Latency PII Masker Skill: Intercepts, identifies, and scrubs sensitive metadata (Names, Emails, Physical Addresses, Crypto Wallets) locally before executing external LLM dispatch.
• Ollama Edge Interoperability: Formally leverages the 270M parameter arpacorp/micro-f1-mask structure for optimized, offline processing.
• Dynamic Modalities:
  • mask: Preserves contextual entity tags (e.g., [PERSON_1]) for post-processing mapping protocols.
  • redact: Completely overwrites tokens heavily with localized constants (XXXX).
  • remove: Intelligently drops strings from the payload to decrease token size and strictly obscure information.

Chores & Polish

• Rewrote API compliance manifest parameters matching internal JSON Schema architecture.
• Integrated rigorous Pytest mock structures intercepting the edge boundary.

Note on Privacy Cycle Nuance: The pii_masker skill included herein performs stateless forward-pass scrubbing by default. For Full Proxy Middleware architectures incorporating Vault-Mapping Redis reconstruction capabilities, visit the root ARPAHLS/micro-f1-mask repository.

Skillware v0.2.2 - Synthetic Data Generator

What's Changed

• New Skill: Introduced the data_engineering/synthetic_generator skill for bulk-generating high-entropy synthetic training data to combat model collapse (Resolves #22).
• Model Agnosticism: The framework now supports internal routing for the synthetic generator to Ollama, Gemini, and Anthropic.
• Zero-Dependency Entropy Scoring: Added a new zlib compression ratio heuristic to natively validate lexical entropy and block boilerplate outputs without heavy NLP dependencies.
• New Documentation: Launched the Data Engineering category in the central skill registry along with comprehensive integration guides and integration scripts (examples/build_dataset_demo.py).
• Bug Fixes: Addressed all flake8 PEP8 linting issues across the module.

Full Changelog: v0.2.1...v0.2.2

v0.2.1

Release v0.2.1 - The Optimization Release

This release introduces the Optimization category to the Skillware ecosystem, anchored by the first of five planned operational efficiency tools.

New Features

• Prompt Token Rewriter Skill: A deterministic middleware skill (optimization/prompt_rewriter) that heuristically compresses bloated prompts and conversation histories.
  • Reduces token usage by 50-80% while retaining 100% of the semantic meaning and instructions.
  • Includes three levels of aggression (low, medium, high) for granular control over compression.
• Optimization Category: Established the registry infrastructure for skills focused on operational and cost efficiency.

Registry & Documentation Refinements

• Scalable Architecture: Refactored the main README to use generic category/skill patterns, ensuring a clean overview as the registry grows.
• Centralized Documentation: Finalized the docs/skills/prompt_rewriter.md reference card and unified documentation in the docs/ system.
• Skill Chaining Patterns: Updated the Gemini usage guide with examples of using the Rewriter as an automated pre-processor.
• CI/CD Polish: Resolved all white-space and formatting linting issues to ensure a clean build.

Versioning

• Major jump from 0.1.0 -> 0.2.1 to consolidate this first batch of enterprise-grade skill additions.

v0.2.0

Release v0.2.0 - Optimization & Middleware Focus

This release introduces the first set of Optimization capabilities to the Skillware ecosystem, focusing on token efficiency and cost reduction for high-intensity agentic loops.

New Features

• Prompt Token Rewriter Skill: A new middleware skill (optimization/prompt_rewriter) that heuristically compresses bloated prompts into fewer tokens.
  • Saves 50-80% in token costs while retaining 100% of the semantic instructions.
  • Supports three levels of aggression: low (formatting), medium (filler removal), and high (aggressive stop-word stripping).
• Optimization Category: Established a new domain in the skill registry for architectural and operational efficiency tools.

Documentation & Examples

• Skill Reference Card: Added comprehensive documentation for the Rewriter at docs/skills/prompt_rewriter.md.
• Middleware Patterns: Updated the Gemini usage guide with "Skill Chaining" examples showing how to use the rewriter as an automated pre-processor.
• Interactive Demo: Added examples/prompt_compression_demo.py so users can test compression logic offline.

Polish & Soundness

• Standardized Manifests: Aligned all skill metadata with the new parameters and constitution standard.
• CI/CD Alignment: Fixed linting and formatting issues to ensure 100% flake8 compliance in core registry files.

v0.1.0

Features

• Pip Installable: Now installable via pip install git+https://github.com/ARPAHLS/skillware.git or from PyPI as skillware (once this release publishes).
• CI/CD Pipeline: Automated testing with pytest and strict linting with flake8 and black.
• Clean Codebase: Fully linted and formatted.

Packaging

• pyproject.toml configuration added.
• Automated PyPI publishing workflow (publish.yml).

Fixes

• Closes #7 (Make Skillware pip-installable)
• Closes #8 (Add Automated Testing and Strict Linting CI)