🚓 context-police
A Claude Code skill that patrols your context budget — measure the token cost of a runaway skills/agents catalog, trim it per-project, emit an interactive HTML recap, and fix the root cause.
╔═══════════════════════════════════════════════╗
║ 🚓 C O N T E X T P O L I C E 🚨 ║
║ ▛▀▀▀▜ pull over — that catalog's over budget ║
║ ▙▄☆▄▟ -48% 🪙 saved/turn · ×N on fan-out ║
╚═══════════════════════════════════════════════╝
🪙 What it is
Claude Code injects the whole catalog of installed skills + agents into context on every turn — and into every subagent you spin up. That's fine with a handful of skills. But if you run a claudeception-style learning loop that mints ~1 new skill per session, your ~/.claude/skills/ quietly balloons to 800+ entries... and every single one is force-loaded, forever, paid again on every fan-out.
The bill (measured on a real ~925-skill catalog): ~248k tokens of skill descriptions per full injection — ~24× over a 1% context budget — re-read every turn and multiplied across concurrent subagents. Small-context agent types (like claude-code-guide) can even overflow on launch with "Prompt is too long" at 0 tokens.
context-police is the toolkit to handle it, honestly:
- 📏 Measure the real per-turn / per-subagent cost (from the cached-prefix token floor).
- ✂️ Trim it — per-project with the verified
skillOverrides lever, or globally with disable-model-invocation: true — without deleting a single skill (everything stays on disk + /name-invocable).
- 📊 Report it — emit a self-contained, clickable HTML recap of exactly what got hidden and why.
- 🧠 Fix the root cause — recognize that most of the bloat is episodic lessons mis-stored as force-loaded skills, and route them to an on-demand retrieval hook instead — gated on a measured recall@K bar before anything is hidden.
It's allow-by-default the whole way down: a skill is only ever hidden when it's clearly irrelevant to your current work, and every move is reversible.
🕹️ Quick Start
Just describe the symptom and Claude reaches for the skill:
You: my subagents are dying with "Prompt is too long" and I have like 800 skills installed.
Claude: [reads context-police] That's the force-loaded catalog. Let me first probe whether it's a
universal overflow or just one small-context agent type, then I'll measure the cost and draft
a per-project skillOverrides denylist (allow-by-default, fully reversible) you can review.
You: do it for this project.
Claude: [picks domain-irrelevant skills — bio DBs, cloud one-offs, single-incident traps that can
never match this stack — sets them "off" in .claude/settings.json, protects your real stack,
then renders an interactive HTML recap of every decision] Restart CC and the catalog shrinks.
The lever is a plain map in your project's .claude/settings.json:
{
"skillOverrides": {
"alphafold-database": "off", // bio DB — irrelevant to a web app
"scanpy": "off", // single-cell genomics — irrelevant
"some-one-off-flask-trap": "off", // single-incident lesson, not a reusable skill
"react-router-v7-migration": "on" // PROTECT: this is your real stack
}
}
"off" drops the skill from the model-invocable catalog (reclaims its tokens) — the SKILL.md is untouched and still /name-invocable. Set it back to "on" (or delete the entry) to undo. Takes effect on the next Claude Code restart.
📦 Installation
Git clone (always works):
git clone https://github.com/wan-huiyan/context-police.git ~/.claude/skills/context-police
Claude Code plugin (marketplace):
/plugin marketplace add wan-huiyan/context-police
/plugin install context-police@wan-huiyan-context-police
Either way, restart Claude Code so the skill is picked up.
✨ Without vs With
