Skip to content

Rule Coherence Graph (RCG)

Detect conflicts in the rule corpora that govern AI coding agents — before the agent does.

AI coding agents (Cursor, Claude Code, Cline, Gemini in agent IDEs, custom LangGraph/Pydantic-AI stacks) are governed by rules spread across many files: .cursorrules, CLAUDE.md, AGENTS.md, .agent/rules/*.md, memory.md, and third-party rule packs. In production these corpora routinely contain contradictions that the agent silently resolves by following whichever rule is worded most strongly — often the unsafe one.

RCG treats a rule corpus as a typed graph instead of flat text: it ingests the files, extracts each rule into a canonical schema, detects conflicts, scores coherence, and lets you query, visualize, and fail CI on them.

Get started Use it in your assistant Hosted MCP demo

Hosted demo: prefer not to install anything? Deploy RCG as a public HTTP service on Fly.io (optionally backed by Neo4j AuraDB) and check pasted rules from any MCP client — see Hosted MCP demo.

Why it exists

In May 2026 a Gemini coding agent reportedly deleted 28,745 lines and broke production after following a third-party rule pack whose directives ("never prompt for confirmation", "auto-deploy", "grant all permissions") contradicted the project's own safety rules. No tool modeled the corpus as a system, so the conflict was invisible until it caused damage.

RCG reconstructs that corpus (examples/gemini_incident) and surfaces the contradictions the agent silently resolved:

RCG conflicts graph

What it does

  • Three detection passes — syntactic (with a human-in-the-loop "approval stance" that removes false positives between aligned safety rules), semantic (embeddings + LLM-as-judge, evidence-bearing), and precedence (rules that can co-fire with no declared ordering).
  • Coherence score per corpus, with --min-score to gate CI.
  • Accepted-conflicts baseline to suppress reviewed/intended conflicts.
  • Ingests .cursorrules, Cursor .mdc, CLAUDE.md / AGENTS.md / memory.md / .agent/rules/*.md, and YAML/JSON rule files.
  • A CLI, an MCP server (so agents can check coherence pre-flight), a reusable GitHub Action, and optional Neo4j graph persistence.

Analyzes, doesn't enforce

RCG checks corpora for internal coherence; it does not gate agent execution at runtime — that's OPA/Cedar/Microsoft AGT's job, downstream. RCG is the linter that runs upstream so a contradiction never reaches them.

Open source (Apache-2.0) · pipx install rule-coherence-graph · GitHub · PyPI

Feedback & support