zedarvates Universal Compressor
Headroom-inspired multi-type compression for botte-secrete. Reduces token usage by 40-90% depending on content type. Works as library, CLI, or MCP server.
Strategies
| Content Type | Strategy | Typical Savings |
|---|---|---|
text |
Dedup lines, collapse blanks | 0-30% |
json |
Compact + truncate large arrays | 20-60% |
log |
Pattern dedup + sampling | 80-98% |
tool_output |
Head+tail, strip ANSI | 50-90% |
code |
Strip comments, collapse imports | 20-40% |
auto |
Auto-detect content type | Best effort |
Usage
from skills.universal_compressor import compress, restore
# Compress with auto-detection
result = compress(content)
print(f"{result.original_size} → {result.compressed_size} ({result.ratio:.0%})")
# Compress with type hint + reversibility
result = compress(big_log, content_type="log", reversible=True, learn=True)
# Restore original
original = restore(result.reversible_key)
learn=True records a verified compressibility_predictor label only when
reversible=True restores the exact original. The ledger stores features and a
fingerprint, never the raw content; repeated samples are deduplicated.
The canonical test runner sets BOTTE_NN_AUTO_LABELS=0, so test traffic cannot
inflate the production ledger.