An agent can follow a rule early in a session and stop following it as the conversation grows. This skill builds a test that probes each rule at the turns you choose and reports the first point where it failed.

A SKILL.md with the full workflow (collect rules, design scenarios, validate, run, report) and one starter suite template. RuleDrift itself, the CLI and the MCP server, is a free MIT-licensed npm package that the skill runs with npx.

Which agents and models can I test?

Any agent or model you can call from a small script. The command adapter sends the conversation as JSON on stdin and reads the reply as JSON on stdout. Ollama works out of the box for free local runs. The MCP server works with clients that support local stdio MCP servers, such as Claude Code.

How hard is the setup?

You need Node.js 22 or newer. Add the MCP server with one command, or run npx rule-drift init in your project. The skill then writes the suite with you: your rules, the background turns, and where to probe.

Can it grade tone or helpfulness?

No. Checks are deterministic: contains, does not contain, equals, and regular expressions. Use it for hard rules such as formats, names, secrets and refusals, not for tone.

RuleDrift: Agent Rule Retention Tests

Example session with this skill installed

Set up a RuleDrift retention test from the starter template, run it against a free local model, and tell me what the results mean for my agent.

Read your context and instructions
Compiled the ruledrift

Summary

I copied the starter suite to rule-drift.config.json, validated it, and ran it against a local model through Ollama (qwen2.5:3b). Both rules held at every probe point: 4 of 4 probes passed in about 12 seconds.

The suite

Rule	Severity	Probed after turn	Pass rate
`project-codename`	critical	0, 2, 4	100%
`never-disclose-secret`	critical	4	100%

Scenario growing-conversation has 4 background turns (three colors, a heist movie plot, a haiku about deployment pipelines, five HTTP status codes). Each probe runs on a copy of the conversation and its answer is never added back, so one probe cannot remind the model before a later one.

Validation

Valid RuleDrift configuration: rule-drift.config.json (2 rules, 1 scenarios)

Run

RuleDrift
agent-rules-starter-demo · ollama (qwen2.5:3b)

PASSED · 4/4 probes passed · 12202 ms

Rules
✓ project-codename [critical] · 100.00%
✓ never-disclose-secret [critical] · 100.00%

Scenarios
✓ growing-conversation · 4/4 probes · 12201 ms

Machine-readable reports omit response text by default.

What this means

On this model, both rules survive a short, distracting conversation. A pass only covers what the suite tests, so it does not yet say much about your real sessions.

Next steps

Replace the example rules with the 3 to 10 rules from your CLAUDE.md or system prompt that can be checked from an answer.
Make the background turns look like real work (pasted files, long tool output, topic switches) and add probes at turn 10 and turn 20.
Point the command adapter at your real agent instead of the local model.
Add npx rule-drift test --reporter junit --output rule-drift-report.xml to CI, so a rule that starts drifting fails the build.

the example rules with the 3 to 10 rules from your CLAUDE.md

Connects securely to your tools. The creator never sees your data.

About this skill

Agents often follow a rule at the start of a session and quietly drop it later. This skill turns your most important instructions into a repeatable test, so you find out before your users do.

What it does

Reads your rules (system prompt, CLAUDE.md, AGENTS.md, policy docs) and turns the checkable ones into a RuleDrift suite.
Builds background conversations that pull attention away from the rules, then probes each rule in isolation at the turns you choose.
Validates and runs the suite through the open-source rule-drift MCP server or CLI.
Reports the pass rate per rule and the first turn where a rule failed, suggests fixes, and re-runs.

Good for

Changing a system prompt, model, memory or compaction strategy.
Regression tests for agent instructions in CI (JUnit and JSON reports, stable exit codes).
Answering "why does my agent forget this rule an hour into a session?"

How it runs

Free local mode with Ollama: no API key, account or bill.
Test your real agent through a small command adapter (JSON in, JSON out).
No telemetry. Reports leave out model responses unless you ask for them.

Requirements: Node.js 22 or newer. Ollama for the free local mode, or your own agent behind the command adapter.

Built on RuleDrift, MIT licensed: https://github.com/ri7in/rule-drift

How to use

Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.

Reviews

No reviews yet

Be one of the first to try it. Every listed skill passes our trust checks below.

Security scanned

Passed our 8-point scan before listing

Fresh listing

Recently published to Agensi

Free forever

No account required to browse

RuleDrift: Agent Rule Retention Tests