Headroom — Context Compression Layer

SmartCrusher + CCR (Compress-Cache-Retrieve) for token-saving context compression. Portable Python module — no external dependencies beyond stdlib.

When to Use

• Large tool output (grep results, JSON arrays, file listings) approaching context limit
• Before sending a long context to a model with token cap
• Need to preserve essential items while dropping noise

Quick Start

from skills.headroom import SmartCrusher, CCRStore

# Compress a JSON array (keep most important items)
crusher = SmartCrusher()
result = crusher.crush(large_json, query="relevant keywords")
# result.compressed  → compressed JSON string
# result.items_kept / items_total → retention ratio
# result.compression_ratio → e.g. 0.3 means 70% tokens saved

SmartCrusher — 5-Dimensional Scoring

Keeps items by:

1. First/Last items — pagination context + latest data (30% head + 15% tail)
2. Error items — 100% preserved
3. Statistical outliers — > 2 std from mean
4. Query-relevant — BM25 match against user query
5. Change points — significant transitions in data

Config overrides:

crusher = SmartCrusher(config={
    "max_items_after_crush": 15,
    "first_fraction": 0.3,
    "variance_threshold": 2.0,
})

CCR Store — Compress-Cache-Retrieve

store = CCRStore(max_entries=1000, ttl_seconds=3600)

# Cache original when crushing
store.put(hash_key, original_text)

# Retrieve if LLM needs more detail
full_text = store.get(hash_key)

Token Estimation

from skills.headroom import estimate_tokens
tokens = estimate_tokens("some text — CJK-aware counting")

Integration Notes

• This module is already imported by skills/shared/context_trimming.py (SmartCrusher layer)
• CCR background worker in skills/sakura/app/agent/memory_curator.py writes to Qdrant
• For roleplay context trimming: the context_trimming module wraps SmartCrusher with 24msg/40K char cap