Codex plugin

Canon Boundary Guard

A cognitive frame for keeping project evidence, chat context, operator instructions, working hypotheses, and model assumptions in separate layers during a Codex session.

View repository Install

The problem is not memory. It is authority.

Codex can hold a long session together and still need a clear way to separate what each source is allowed to mean.

Chat is not project truth

Conversation can shape the task, but it does not automatically become durable project content.

Instructions are not content

Agent-facing rules can govern behavior without belonging in the repository.

Assumptions are not evidence

Model priors and working hypotheses stay separate until verified or approved.

The frame stays on while Codex thinks.

The hook is only a reinforcement point. The skill itself is the posture: classify sources while reading, reasoning, planning, detecting conflicts, and deciding what can persist.

Canon Boundary Guard does not decide project truth. It makes visible what class of source is being used and where that source is allowed to go.

L0Persistent or verified evidence: files, git state, tests, schemas, diagnostics, verified tool output.
L1Conversation material present in context, not approved for persistence.
L1AOperator-approved material for this turn, before it is written.
L2Instructions that shape agent behavior. Not project content.
L2ACodex instruction-chain material, including runtime AGENTS guidance.
L3Unverified model memory, assumed conventions, and unstated framework behavior.
Preserve L0.

Project evidence can be reorganized or used directly when the evidence is clear.

Do not silently persist shaping.

L1 and L3 need verification or approval before becoming project content.

Keep agent-control out of project content.

L2 and L2A persist only when the operator explicitly asks for agent-facing instructions.

AGENTS.md needs authority handling.

In some Codex environments, AGENTS.md instructions may appear as a visible user-role message before the operator's real first request.

Canon Boundary Guard treats a leading block starting with AGENTS.md instructions for <path> as runtime instruction-chain material, not ordinary chat and not repository content.

When this distinction matters, verify the authority layer before using the block as project evidence or writing it into persistent files.

Install

Add the repository to Codex as a plugin source, then install or enable Canon Boundary Guard from the Codex plugin picker.

After it is enabled, select Canon Boundary Guard at the start of a session when you want this frame active. The skill and the recommended hook are already included in the plugin.

Hook reinforcement

The hook is highly recommended. When trusted, it re-surfaces the compact frame before matched write tools so Codex sees the layer separation again at write time. It does not block tool calls, rewrite requested operations, or decide project policy.

License

This project is licensed under CC BY-SA 4.0: Creative Commons Attribution-ShareAlike 4.0 International.

AI-assisted development

This project was developed with AI assistance through human-directed drafting, structuring, review, and refinement.