dl4rce/flaiwheel

"This is the Way" — Now Actually Works

The bootstrap trigger ("This is the Way" / "42") previously scanned the knowledge repo (always clean) instead of the project repo (where the messy docs live). Fixed.

New: classify_documents(files) MCP Tool

The AI agent scans the project directory locally, reads the first ~2000 chars of each documentation file, and sends them to Flaiwheel via this new tool. Flaiwheel classifies each file using its embedding model (cosine similarity against category templates + keyword classification + three-signal consensus) and returns:

• Per-file classification with category, confidence score, target directory, and suggested write_* tool
• Duplicate detection across the batch
• Rewrite flags for files lacking structure

The agent then presents the migration plan to the user and executes approved actions using the existing write_* tools.

Architecture

| Component | Role |
|-----------|------|
| AI Agent (local) | Scans project files, sends content, rewrites docs, pushes via write_* tools |
| Flaiwheel (Docker) | Classification engine (embeddings + keywords), persistence, quality checks |

No LLM needed for classification — just the existing sentence-transformers embedding model.

Also in this release

• analyze_knowledge_repo() docstring clarified: scans knowledge repo only
• _classify_by_keywords extracted as reusable module-level function
• Install script updated with dual workflow (new project vs existing knowledge repo)
• 61 bootstrap tests (15 new), 216 total tests — all passing

Full Changelog

https://github.com/dl4rce/flaiwheel/compare/v2.4.0...v2.5.0