Over Reach Detector — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Over Reach Detector (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
<!-- mcp-name: io.github.choreoatlas/over-reach-detector -->
Detect and report when AI coding agents change files outside their declared task scope. Designed to plug into Cursor, Claude Code, and other MCP-compatible AI coding agents via the standard stdio transport.
Compares two things:
If the actual diff exceeds the declared scope, the tool returns status=over_reach and lists the offending files and categories.
over-reach-detector is an audit/disclosure tool, not a sandbox and not a blocker.
It compares a declared task scope with the actual files/categories changed by an AI coding agent. If the actual changes exceed the declared scope, it reports that mismatch.
It does not prevent file writes by itself. It does not decide whether to revert, approve, or block a change. The caller remains responsible for enforcement, rollback, or human review.
From PyPI:
pip install over-reach-detectorFrom source:
git clone https://github.com/choreoatlas/over_reach_detector
cd over_reach_detector
pip install -e .Run all tests: python -m pytest -v
Try the CLI directly: python -m over_reach_detector.detector --input fixtures/example_pr_1.json --format markdown
Start the server (stdio transport): over-reach-detector (or python -m over_reach_detector.server from source)
Register with your AI agent:
~/.cursor/mcp.json, add (after pip install over-reach-detector):{
"mcpServers": {
"over-reach-detector": {
"command": "over-reach-detector"
}
}
}Dev / from source: use "command": "python", "args": ["-m", "over_reach_detector.server"] (run from repo root).
pip install over-reach-detector, run claude mcp add over-reach-detector over-reach-detector (writes to ~/.claude.json). Dev / from source: claude mcp add over-reach-detector /absolute/path/to/python -m over_reach_detector.server.check_scope_tool takes:
declared_files: list of fnmatch globs (e.g. ["docs/*.md", "tests/*.py"])declared_categories: subset of ["tests", "docs", "infra", "config", "code"]actual_files: list of file paths the AI modifiedoutput_format: "json" (default) or "markdown"Returns a report with:
status: in_scope (within declared scope) | over_reach (reported mismatch) | emptyfile_overreach: files not matching any declared globcategory_overreach: inferred categories outside the declared setCurrent scope: CLI + MCP stdio server + 1 tool. Python only. fnmatch-based globs.
Out of scope (forbidden): code quality review, security audit, completeness governance, languages other than Python, multi-tool MCP servers, HTTP/SSE transport, GitHub Actions integration. These are deliberately deferred to later versions or never.
Call check_scope_tool directly from Python (same logic the MCP server exposes):
import json
from over_reach_detector import server
result = server.check_scope_tool(
declared_files=["docs/*.md"],
declared_categories=["docs"],
actual_files=["docs/a.md", "scripts/extra.py"],
output_format="json",
)
report = json.loads(result)
print(report["status"]) # "over_reach"
print(report["file_overreach"]) # ["scripts/extra.py"]MIT — see LICENSE.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.