Canon Boundary Guard gives Codex a structured classification frame for information provenance. When you are running agentic sessions that touch a real codebase, the model works with several fundamentally different kinds of information at once — project files, conversation history, runtime instructions, and its own prior assumptions. Without explicit separation, these sources blur together and safe-looking writes can carry unverified material into persistent project content. Canon Boundary Guard makes the boundaries visible throughout the whole session: while reading, reasoning, planning, and before any write.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/canon-boundary-guard-codex/llms.txt
Use this file to discover all available pages before exploring further.
Quickstart
Install the plugin and activate the frame in under five minutes.
Core Concepts
Understand cognitive-layer separation and why it matters for agentic workflows.
Provenance Layers
Explore the six source classes — L0 through L3 — and how they govern persistence.
Plugin Reference
Inspect the plugin manifest, skill file, hook configuration, and injection script.
How It Works
Canon Boundary Guard packages a Codex skill and an optional lifecycle hook into a single installable plugin.Install the plugin
Add the repository to Codex via the marketplace command and enable it from the plugin list.
Activate the frame
Tell Codex to use Canon Boundary Guard at the start of a session. The skill loads its full operating frame before any action is taken.
Work normally — with layer awareness
As Codex reads, reasons, and plans, it classifies non-L0 material inline. Project evidence (L0) is preserved; chat, instructions, and model assumptions are tagged and kept separate.
Key Capabilities
Six Provenance Layers
L0 Evidence through L3 Model Prior — a clear taxonomy for every information source Codex encounters.
Inline Tagging
Automatic inline markers ([L1], [L2], [L3], etc.) flag non-verified material in generated output.
PreToolUse Hook
Injects the classification frame before apply_patch, Write, and Edit calls without blocking execution.
AGENTS.md Authority
Correctly classifies runtime instruction-chain preludes regardless of their displayed conversation role.
Dossier Protocol
Three escalating dossier modes control how tightly provenance is documented before a write.
Decontamination
A checklist of residue patterns that must not silently enter persistent project content.