@vibo-dev/context-compactor

🧠 Context Compactor

Condense a long agent session into a handoff memo — decisions, open tasks, risks, facts — without losing what matters.

Every agent developer knows the pain: your session grows to 200k tokens, context gets compacted, and the important stuff vanishes — the decision you made, the task you promised, the URL you need tomorrow.

Context Compactor reads a raw transcript (txt/md, any language mix of Russian/English) and produces a compact handoff memo grouped into:

• ✅ Decisions — what was actually decided
• 📌 Open tasks — what still needs to be done
• ⚠️ Risks — blockers, broken things, dependencies
• 🔗 Facts & links — URLs, handles, versions

Safety first: secrets never leak into the memo

Handoff memos travel to other agents and files. Context Compactor therefore redacts credential values before scoring: password, пароль, token, api key, secret, private key, sk-…/ghp_… tokens and BEGIN PRIVATE KEY blocks become <REDACTED: …>. The decision/context around them is kept — the secret value is not. --self-test asserts this with a planted-secret case.

Why you need it

• Survive context compaction: paste the memo back into the next session.
• Hand work to another agent (or a human) without dumping 500KB of transcript.
• Daily session digest in one command.

Install

# npm (any OS with Python 3.8+)
npx -y @vibo-dev/context-compactor --input session.md

# ClawHub / OpenClaw registry: install "context-compactor-cli"
# GitHub: clone this repo

Quick start

git clone https://github.com/vnbochkarev-netizen/context-compactor
cd context-compactor

# Condense a session transcript
python3 compactor.py --input session.md

# Save to a handoff file for the next session
python3 compactor.py --input session.md --output handoff.md

# Verify the tool itself
python3 compactor.py --self-test

No dependencies. Python 3.8+. Works on Linux/macOS.

How it works

Heuristic line scoring (no LLM, no API costs, runs offline):

• Decision verbs: «решили / договорились / делаем», "decided / agreed / we will"
• Task markers: «нужно / осталось / напомни», "todo / next step / don't forget"
• Risk words: «риск / баг / блокер», "broken / blocker / depends on"
• Links, handles and version-like facts get boosted
• Code dumps, tool output and noise are filtered out
• Bucketing is context-aware: "договорились, ссылка: " is a fact, not a decision; strong verbs («решили», "decided") always win
• Long transcripts: the head AND the tail are analyzed (final decisions usually sit at the end); skipped middle is counted and reported in the memo header, which also carries the tool version, date and source file

Example memo

# Handoff memo (auto-compacted)

## ✅ Decisions / Решения
- решили сделать skill-injection-scanner на python, MIT-лицензия
- договорились: завтра публикуем

## 📌 Open tasks / Открытые задачи
- надо проверить его тестами перед публикацией

## ⚠️ Risks / Риски
- риск: без тестов будет много ложных срабатываний

## 🔗 Facts & links / Факты и ссылки
- репозиторий проекта: https://github.com/vnbochkarev-netizen/skill-injection-scanner

Use cases

• Handoff files after context compaction (Hermes, Claude, OpenClaw, Codex…)
• Agent-to-agent handoff — pass the memo through an agent bridge
• Daily digests — what did we actually do today?

License

MIT © 2026 Viacheslav Bochkarev