PermitGraph Docs

AI Analysis Guide

How AI agents should inspect PermitGraph output without inventing findings or leaking sensitive data.

What this is

This guide explains how an AI agent should read PermitGraph output. The scanner is the source of truth. Deep Agent review is an explanation layer over scanner artifacts.

When to use it

Use this page when an AI agent needs to summarize a run, explain why a repository is blocked or needs review, compare expected findings with actual findings, or draft reviewer-facing remediation.

Command or example

uv run agent-permit scan . --ci --exclude "tests/fixtures/**"
uv run --extra deep-agent agent-permit investigate .agent-permit/runs/<run_id>

Output to expect

Review artifacts in this order:

  1. summary.md
  2. permit.yaml
  3. raw-findings.json
  4. graph-paths.json
  5. controls.json
  6. agent-investigation.md

How to interpret it

approved means no configured agent-access risk matched in this scanner run.

needs_review means a reviewer must inspect evidence before unattended access continues.

blocked means unattended agent access should stop until remediation or an explicit exception exists.

Common mistakes

  • Creating new findings from model intuition.
  • Claiming a repository is safe beyond the configured scanner scope.
  • Publishing generated artifacts without redaction review.
  • Revealing raw secret values.
  • Executing repository code, MCP servers, CI workflows, package scripts, or external tools to prove a scanner result.

On this page