Skip to content

YuhaoLin2005/compact-counter

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 

Repository files navigation

compact-counter

Track Claude Code compaction events. Know when to /reset. MIT.

The Problem

Claude Code auto-compacts long conversations. Each compaction compresses old context into an AI-generated summary. Early precise instructions become vague hints. After 16+ rounds, your original rules may be gone — and the model doesn't know it.

Compaction count is invisible by default. This tool makes it visible.

What We Found

Tracking 459 sessions across 6 weeks (June–July 2026):

Compactions Sessions Zone
0 430
1–5 14 Warm-up
6–10 3 Sweet spot
11–20 6 Decay
21–50 4 High-risk
51+ 2 Critical

U-shaped curve: compliance peaks at 3–8 compactions (enough context to learn your workflow; early rules not yet compressed away), then declines. After 16 compactions, instruction fidelity drops sharply.

L8→L12 cliff: Controlled experiment (100 API calls, 5 adversarial rules × 5 degradation levels) confirms the production data — full compliance drops from 80% to 20% when instructions degrade past a critical semantic threshold. Format constraints survive longer than semantic ones.

Quick Start

python compact-counter.py
# Session abc123: 7 compactions ● sweet spot
# Session def456: 23 compactions ● consider /reset

~200 lines, stdlib only. Reads compact-state.json from your Claude Code data directory.

How this fits

This tool was born from the hermes-workspace governance research. Compaction count is a feature in the L4 Drift Predictor — when sessions cross 16 compactions, the risk model flags potential rule decay. The controlled experiment (compact_fidelity_decay.py in hermes-workspace) validated the production finding with deterministic scoring.

Related

  • hermes-workspace — Research hub: 16 experiments, 5-layer governance architecture, compaction decay experiment
  • paper-validator — One-command governance audit. Compaction count feeds L4 Drift Predictor.
  • digital-twin-trainer — QLoRA + DPO training pipeline
  • DEV.to — 31 articles · 掘金 — 中文技术写作

License

MIT

About

LLM 上下文压缩 sweet-spot 检测工具 | Python标准库 | 发现过早压缩反而改善输出 | Part 3

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors