Skip to content

CHANGELOG.md

Latest commit

 

History

History
241 lines (185 loc) · 11.9 KB

CHANGELOG.md

File metadata and controls

241 lines (185 loc) · 11.9 KB

Changelog

0.5.0 – Maintenance Refresh (February 2026)

Added

  • Python 3.13 support: Added to CI test matrix and pyproject classifiers
  • PEP 561 py.typed marker: Enables type-checker support for library consumers
  • Public API exports: hegelion.__init__ now exports key classes (DialecticalPrompt, PromptDrivenDialectic, AutocodingPrompt, PromptDrivenAutocoding, AutocodingState, and workflow helpers)
  • --version flag: Both hegelion-server and hegelion-setup-mcp CLIs now support --version

Changed

  • Updated model references: .env.example now references current models (Claude Sonnet 4.6, GPT-5.2, Gemini 3 Pro)

Fixed

  • CHANGELOG formatting: Corrected malformed duplicate 0.3.0 / 0.2.3 section header
  • README staleness: "Recent Changes" section now includes v0.4.5 and v0.4.6

0.4.6 – CI Fix & Automated Releases (February 4, 2026)

Fixed

  • Fixed CI pipeline and automated release workflow

0.4.5 – Optional CLI Execution for dialectical_single_shot (February 4, 2026)

Added

  • Single-call CLI execution for dialectical_single_shot: execute, timeout_seconds, max_retries, plus env-based configuration (HEGELION_LLM_COMMAND_JSON, HEGELION_LLM_COMMAND, HEGELION_MCP_AUTO_EXECUTE=1)
  • hegelion-setup-mcp flags: --llm-command-json, --llm-command, --auto-execute to configure CLI execution via MCP host config

0.4.4 – Simplified Skill & Command (January 21, 2026)

Changed

  • Condensed /hegelion slash command and Codex skill to minimal routing tables
  • Removed verbose manual workflows in favor of MCP-first approach

0.4.3 – MCP Refactor + Codex Skill (January 12, 2026)

Added

  • Codex skill: Added skills/hegelion/SKILL.md for the /hegelion workflow

Changed

  • Refactored MCP server into tooling, handlers, validation, and response helpers

0.4.2 – Unified /hegelion Command (January 2026)

Added

  • /hegelion command: Single entry point for both dialectical and autocoding workflows
  • Host shortcuts: hegelion-setup-mcp --host claude-desktop|cursor|vscode|windsurf

Changed

  • Consolidated separate autocoding command and skill into unified /hegelion
  • Code formatted with black

0.4.1 – Schema Versioning (December 2025)

Added

  • Schema versioning: All structured outputs include schema_version
  • Phase clarity: player_prompt and coach_prompt include current_phase and next_phase
  • Error handling: Invalid transitions return structured errors with expected, received, hint

0.4.0 – Autocoding System (December 2025)

Added

  • Player-Coach loop: Dialectical autocoding based on Block AI's g3 agent research
  • MCP tools: autocoding_init, player_prompt, coach_prompt, autocoding_advance
  • Session persistence: autocoding_save / autocoding_load
  • Single-shot mode: autocoding_single_shot for simpler tasks
  • Brand entrypoint: hegelion tool with mode parameter

0.3.2 – MCP Server Simplification

Removed

  • MCP run_dialectic tool: Removed the server-side execution tool that required LLM API configuration. Users on Cursor, Claude Desktop, VS Code, etc. should use dialectical_single_shot or dialectical_workflow instead—these generate prompts that your editor's LLM executes directly, with no API keys needed.

Changed

  • Updated mcp_instructions.md: Documentation now guides agents to use prompt-generation tools (dialectical_single_shot, dialectical_workflow) rather than server-side execution.

Why This Change

The run_dialectic MCP tool required server-side API keys (e.g., HEGELION_PROVIDER, ANTHROPIC_API_KEY), which was confusing for MCP users who expected a zero-config experience. The prompt-generation tools provide the same dialectical reasoning without any server-side LLM configuration.

0.3.1 – Response Styles & Gemini CLI Extension Support

Added

  • New Response Styles: Added two new response formats for dialectical workflows:
    • conversational: Natural dialogue format with smooth transitions like "but on the other hand..." and "so perhaps the best way forward is..."
    • bullet_points: Concise bullet-point format with clearly marked Thesis, Antithesis, and Synthesis sections for quick scanning
    • Full compatibility with existing json, sections, and synthesis_only styles
  • New: Gemini CLI Extension: Added official support for the Gemini CLI. You can now install Hegelion directly as an extension to use dialectical tools within your Gemini CLI workflows.
  • New: GEMINI.md Context: Included a persistent context file to guide the Gemini model on how to use Hegelion's dialectical tools effectively.

0.3.0 – Model-Agnostic Support & Feature Expansion

  • New: Model-Agnostic MCP Server: Introduced hegelion-prompt-server which allows using Hegelion with any model configured in your editor/IDE (Cursor, VS Code, Claude Desktop) without needing API keys.
  • New: Persona-Based Critiques: Configure specific personas for the antithesis phase (e.g., "Council of Critics," "Security Engineer," "Ruthless Editor") via the Python API.
  • New: Iterative Refinement: Support for looping the dialectical process (Synthesis Round 1 → Thesis Round 2) for deeper analysis.
  • New: Search Grounding: Instruct models to verify claims with search tools during the antithesis phase.
  • Changed: The hegelion CLI is now primarily a demo/summary tool. Advanced workflows (personas, iterations) are available via the Python API and MCP servers.
  • Documentation: Comprehensive rewrite of README and documentation to emphasize the dual-path usage (Prompt-Based vs. API-Based).

0.2.3 – MCP/assistant integration polish

  • Clarified canonical HegelionResult JSON schema in README.md, including backend and timing fields, and documented when trace and metadata.debug appear.
  • Expanded docs/MCP.md with friendly request/response examples for run_dialectic and run_benchmark, plus explicit assistant-integration guidance for parsing JSON and JSONL outputs.
  • Aligned README/MCP documentation with the actual HegelionResult schema used by the core engine, CLI, and MCP server to ensure AI assistants see a single, stable contract.

0.2.2+phase2 – Search-Grounded Dialectics with Council and Judge

🎉 MAJOR RELEASE: Advanced Dialectical Reasoning System

✨ New Phase 2 Features

🔍 Search-Grounded Antithesis

  • Smart Search Providers: Automatic fallback from Tavily (premium) to DuckDuckGo (free)
  • Real-World Grounding: Antithesis phases can now access current information via web search
  • Context Integration: Search results are seamlessly woven into dialectical reasoning

👥 The Council - Multi-Perspective Critiques

  • Three Expert Perspectives:
    • The Logician: Formal reasoning and logical consistency
    • The Empiricist: Evidence, facts, and empirical grounding
    • The Ethicist: Ethical implications and societal impact
  • Concurrent Processing: Multiple critiques generated simultaneously via async
  • Unified Synthesis: All council perspectives integrated into final reasoning

⚖️ The Iron Judge - Quality Evaluation

  • Structured Scoring: 0-10 quality assessment with detailed reasoning
  • Instructor Integration: Guaranteed structured output using Pydantic models
  • Iterative Improvement: Automatic retry for low-quality results
  • Critique Validity: Assessment of whether antithesis critiques were properly addressed

🛠️ Enhanced APIs

Core API Enhancements

# Phase 2 Enhanced Dialectics
result = await run_dialectic(
    query="Your question",
    use_search=True,        # Web search grounding
    use_council=True,       # Multi-perspective critiques  
    use_judge=True,         # Quality evaluation
    min_judge_score=7,      # Quality threshold
    max_iterations=3        # Automatic improvement
)

MCP Server Phase 2 Support

  • Updated run_dialectic tool with all Phase 2 parameters
  • Enhanced Claude Desktop integration with search and council features
  • Structured quality evaluation for AI assistant workflows

📦 New Dependencies

  • duckduckgo-search>=6.0.0 - Free web search (no API key required)
  • instructor>=1.0.0 - Structured LLM output with Pydantic
  • pydantic>=2.0.0 - Enhanced data validation and parsing
  • Optional: tavily-python>=0.3.0 - Premium search for professional use

🏗️ Architecture Improvements

  • Modular Phase 2: Search, council, and judge features are optional and imported on-demand
  • Graceful Degradation: Falls back to Phase 1 standard dialectics if Phase 2 dependencies unavailable
  • Enhanced Metadata: Judge scores, council perspectives, and search context in debug traces
  • Quality Assurance: Automatic retries and quality thresholds for production use

🔧 Technical Details

  • Backward Compatible: All existing APIs continue to work unchanged
  • Performance: Phase 2 features add ~2-5x processing time but deliver significantly richer analysis
  • Error Handling: Robust fallbacks ensure system never fails due to Phase 2 enhancements
  • Caching: Phase 2 results are fully cacheable with enhanced cache keys

💡 Usage Examples

Search-Enhanced Analysis:

# Ground dialectics in current information
result = await run_dialectic(
    "What are the latest developments in quantum computing?",
    use_search=True
)

Multi-Perspective Council Critique:

# Get logical, empirical, AND ethical perspectives
result = await run_dialectic(
    "Should we implement universal basic income?",
    use_council=True,
    council_members=["The Logician", "The Ethicist"]
)

Quality-Assured Reasoning:

# Ensure high-quality output with automatic improvement
result = await run_dialectic(
    "Explain consciousness and free will",
    use_judge=True,
    min_judge_score=8,
    max_iterations=3
)

Full Phase 2 Stack:

# The complete enhanced dialectical experience
result = await run_dialectic(
    "How should humanity approach artificial general intelligence?",
    use_search=True,      # Current information
    use_council=True,     # Multiple expert perspectives
    use_judge=True,       # Quality assurance
    min_judge_score=7,    # High quality threshold
    max_iterations=2,     # Improvement iterations
    debug=True           # Full diagnostic trace
)

This release transforms Hegelion from a basic dialectical reasoner into a sophisticated multi-agent reasoning system with real-world grounding and quality assurance.

0.2.2 – MCP documentation and GLM backend verification

  • Added comprehensive MCP reference guide (docs/MCP.md) with tool schemas, troubleshooting, and client integration notes
  • Added Claude Desktop configuration example (examples/mcp/claude_desktop_config.json)
  • Verified GLM 4.6 backend via OpenAI-compatible endpoint (Z.AI devpack)
  • Enhanced documentation with verified backends section and log-sharing guidance
  • Improved README with MCP quick start and backend verification details
  • Fixed package discovery in pyproject.toml to exclude logs/ directory from builds

0.2.1 – License format fix

  • Fixed deprecated license format in pyproject.toml (changed from table to SPDX string format)
  • Removed deprecated license classifier

0.2.0 – First public release

  • Refactored into a clean hegelion Python package with clear module boundaries (core, engine, backends, parsing, models, prompts, config, MCP server).
  • Introduced HegelionResult with structured contradictions, research proposals, and provenance metadata.
  • Added CLI tools: hegelion for single queries and hegelion-bench for JSONL benchmarks.
  • Implemented MCP server (hegelion-server) exposing run_dialectic and run_benchmark tools.
  • Wrote comprehensive README and HEGELION_SPEC.md describing the protocol and JSON schema.
  • Added test suite, pre-commit hooks, and GitHub Actions CI for automated testing.