dl4rce/flaiwheel

What's New

🆕 analyze_codebase(path) — 28th MCP Tool

Zero-token cold-start analysis of legacy source code directories.

The problem it solves: bootstrapping a large undocumented codebase currently requires the LLM agent to read hundreds of files, costing $5–50 in tokens and 30–60 minutes. analyze_codebase() eliminates that.

How it works:

• Runs entirely server-side inside the Docker container
• Python files: parsed with Python's built-in ast module — extracts classes, functions, docstrings, imports
• TypeScript/JavaScript: regex extraction of exports, classes, functions, JSDoc
• Classification: existing MiniLM model + nearest-centroid against Flaiwheel's 9 knowledge categories (zero new models)
• Deduplication: existing cosine similarity pipeline (threshold 0.92)
• Prioritization: documentability score = docstrings(+3) + import density(+2) + public API surface(+0.6/fn) + entry-point name patterns(+2)

Output: single bootstrap_report.md with:

• Language distribution
• Inferred category distribution
• Top 20 files to document first (ranked table)
• Near-duplicate file pairs
• Directories with no parsed code

Cost: ~90% token reduction vs. agent-driven file reading. ~5–30 seconds for a typical repo.

Zero new dependencies — Docker image stays ~1GB.

Other Changes

• 20 new tests — total suite now 259
• Total MCP tools: 27 → 28