ilyautov/cordon/cordon-install

Claude Code and agent skill by ilyautov/cordon

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

Install Cordon, a deterministic layer between untrusted content and agent actions, and verify that the harness actually calls it. Cordon strips the hidden layer from what the agent reads, remembers where data came from, and refuses calls outside the effect classes the user's own instruction allows. No model call on the hot path. Use when someone asks to protect an agent from prompt injection, to check whether a hook is really firing, or to install Cordon. Triggers: «install cordon», «поставь cordon», «защити агента от prompt injection», «prompt injection protection», «скрытые инструкции в тексте», «agent reads untrusted content», «cordon doctor», «проверь что хук срабатывает».

Install

npx vskill@latest install ilyautov/cordon/cordon-install

SKILL.md

Installing Cordon

Cordon stands between untrusted text and what the agent does with it. It ships as a Claude Code plugin bound to four harness events: the user's message, a tool call, a tool result, the display of the answer.

Node 22 or newer. No keys, no tokens, no network access: there are no network requests and no model calls on the hot path.

Install

/plugin marketplace add ilyautov/cordon
/plugin install cordon@cordon

No build step. plugin/dist/cli.js ships ready to run, because an installed plugin has neither node_modules nor package.json beside it, and a file with external imports would crash node on the first event. The harness reads a crashed hook as «pass», so the defence would switch off silently.

Verify, and do not confuse the two questions

/hooks answers «is it registered». It does not answer «does it work».

cordon doctor answers the second one: it reads the effective policy, runs a built-in attack sample through the whole path, and names the dangerous parts of the configuration. The plugin path carries the marketplace name and the version, so it changes on every update. Ask the harness instead of hard-coding it:

bash
CORDON=$(node -e "
const p = require(process.env.HOME + '/.claude/plugins/installed_plugins.json')
const key = Object.keys(p.plugins).find(k => k.startsWith('cordon@'))
if (!key) { console.error('plugin is not installed'); process.exit(1) }
console.log(p.plugins[key][0].installPath + '/dist/cli.js')
")
node "$CORDON" doctor

self-check: ok means the hidden layer is stripped, a call outside the certificate is refused, and a call inside it goes through. A hook that is never called looks, from outside, exactly like a hook that had no reason to fire.

What to tell the user honestly

Say how far this has been proven, and do not round it up. The wiring was exercised on a live Claude Code session, 2.1.236: all four events fire, the certificate refuses, provenance refuses, the footer is drawn, argument quarantine is applied. The Gemini CLI adapter, the MCP gateway and the LangChain middleware have not been run live. Their install guides are docs/install-gemini.md, docs/install-mcp.md, docs/install-langchain.md.

The price of the data axis is friction: after reading untrusted content, a consequential call waits until the user names its destination in their own message. On the adversarial battery's working profile this took attack success from 79% to 7%. The rule switches off with exposure: false in the policy, and cordon doctor says out loud when it is off.

What Cordon keeps on disk

Policy in ~/.cordon/policy.yaml, session state in ~/.cordon/sessions/, drafts in ~/.cordon/drafts/, the decision journal at the path from notify.file. Session state and drafts hold content that came from untrusted sources, so expired files are deleted automatically.

Nearby

github.com/ilyautov, lab AI Frontier.

Security

  • Security scanPassed, 90 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

7-day trend score
0

Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/ilyautov/cordon/cordon-install/badge)](https://verified-skill.com/skills/ilyautov/cordon/cordon-install)

Something wrong with this skill or its page? Tell us.