Qompressor

The Qompressor agent is the “Skeletonizer” for QonQrete. It creates lightweight structural representations of the codebase, mirroring qodeyard into bloq.d with implementation details stripped out.

How it works

1. Input and Output:

Takes source_dir (typically qodeyard) and dest_dir (typically bloq.d)
Expects qodeyard to be the source of full code

2. Directory Mirroring and Cleaning:

Aggressively cleans bloq.d by removing its entire contents
Recreates fresh directory structure
Ensures bloq.d is always an up-to-date mirror

3. File Processing Logic:

Files are categorized by filename and extension:

COPY_FILENAMES: Files like Dockerfile, Makefile copied as-is (critical context, low tokens)
COPY_EXTENSIONS: Files like .yaml, .json, .md copied as-is
COMPRESS_EXTENSIONS: Code files (.py, .js, .go) undergo compression

4. Compression Methods:

Python (compress_python):

Uses Python AST to parse code
Keeps: import statements, class definitions, function/method signatures, docstrings
Strips: Implementation bodies → replaced with # ... (body stripped by Qompressor) ...

Generic (compress_generic):

Regex-based approach for other languages
Keeps: comments, structural keywords (function, class, def), block openers
Less precise than AST but still effective

Benefits

Token Cost Reduction: Dramatically reduces context token count
Architectural Focus: LLMs focus on structure and interfaces, not implementation details
Speed: Fast and deterministic processing

What’s configurable

Internal Configuration:

COMPRESS_EXTENSIONS (set): File extensions for compression
COPY_EXTENSIONS (set): File extensions for direct copy
COPY_FILENAMES (set): Specific filenames for direct copy

Implicit via config.yaml:

options.use_qompressor (boolean): Whether Qrane runs Qompressor at all