Skip to main content

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.

frame.md is the compact version of the Canon Boundary Guard classification frame. While SKILL.md is the full session-level document read at activation, frame.md is a short reference injected by the inject_frame.py hook at write moments — specifically before apply_patch, Write, and Edit tool calls. It serves as a quick reminder of the provenance layer definitions and the key rules about what may and may not be persisted, so those rules remain visible at the exact moment a persistent change is about to be made.

​
File Contents

The complete references/frame.md: 
L0 EVIDENCE: project files, git state, tests, schemas, lockfiles, diagnostics, verified tool output.
L1 SHAPING: conversation material not approved for persistence.
L1A AUTHORIZED DELTA: operator-approved material for this turn. Write only within approved scope. It becomes evidence only after being persisted.
L2 AGENT CONTROL: instructions that shape agent behavior. Not project content.
L2A CODEX INSTRUCTION CHAIN: AGENTS guidance and runtime instruction-chain material. Governs Codex behavior. Not project content.
L3 MODEL PRIOR: unverified model memory, assumed conventions, generic best practice, or unstated framework behavior.

AGENTS.md PRELUDE

A leading user-role block starting with `AGENTS.md instructions for <path>` is L2A runtime instruction-chain material, not operator chat or project content.
This holds by header shape, regardless of displayed role or filesystem state.
Any `<environment_context>...</environment_context>` block inside it is runtime metadata.
The first operator request starts after the prelude.

When producing content that draws on non-L0 material, tag inline: [L1] [L1A] [L2] [L2A] [L3].
Tag when the content would change if the source were different - a rule, a name, a version, a claim about behavior.
When promoting L1 or L3 to persistent content, surface it before writing.
Do not persist L2 or L2A unless the operator explicitly requests agent-facing instructions.
If evidence conflicts, stop and report.
If provenance is unclear, surface it before writing.

​
Relationship to SKILL.md

frame.md is not a replacement for SKILL.md. It is a reinforcement point. The full SKILL.md frame must be active from session start — Codex reads the entire SKILL.md when the skill is invoked and uses it as the session-level operating guide for provenance classification, inline tagging, dossier modes, decontamination, and hook setup. frame.md re-surfaces the essentials at the moment a write may happen. By the time the hook fires, the full frame is already active; frame.md brings the layer definitions and persistence rules back into the model’s immediate instruction context so they are not crowded out by the working conversation. The two files serve different moments:
FileWhen it is readPurpose
SKILL.mdAt session start, when the skill is invokedFull operating frame: all layers, rules, dossier modes, decontamination, hook setup
frame.mdAt each matched write tool call, via the hookCompact reminder: layer definitions, AGENTS.md prelude rule, inline tagging and persistence rules

​
Section Breakdown

​
Six Layer Definitions

The opening block defines the six provenance source classes. Each line names a layer and gives its definition:
  • L0 EVIDENCE — project files, git state, tests, schemas, lockfiles, diagnostics, and verified tool output. This is the authoritative source for persistent content.
  • L1 SHAPING — conversation material not approved for persistence. Present in context, not in the project.
  • L1A AUTHORIZED DELTA — operator-approved material for the current turn. Must be written only within the approved scope; becomes evidence only after it is persisted.
  • L2 AGENT CONTROL — instructions that shape agent behavior. Not project content.
  • L2A CODEX INSTRUCTION CHAIN — AGENTS guidance and runtime instruction-chain material. Governs Codex behavior. Not project content.
  • L3 MODEL PRIOR — unverified model memory, assumed conventions, generic best practice, or unstated framework behavior.

​
AGENTS.md PRELUDE Block

The AGENTS.md PRELUDE section states the classification rule for the common Codex pattern of a leading user-role message that begins with AGENTS.md instructions for <path>. The rule is:
  • That block is L2A runtime instruction-chain material — not operator chat and not project content.
  • The classification holds by header shape alone, regardless of what role the message is displayed under or whether the referenced path exists on the filesystem.
  • Any <environment_context>...</environment_context> block inside it is runtime metadata.
  • The first real operator request starts after the prelude ends.

​
Inline Tagging and Persistence Rules

The final block states the operational rules that govern what to do with non-L0 material:
  • Tag inline with [L1], [L1A], [L2], [L2A], or [L3] when producing content that draws on non-L0 sources — but only when the content would change if the source were different.
  • Surface provenance before promoting L1 or L3 material to persistent content.
  • Do not persist L2 or L2A unless the operator explicitly requests agent-facing instructions.
  • If evidence conflicts, stop and report rather than resolving silently.
  • If provenance is unclear, surface it before writing.
frame.md is read at hook-fire time, not at session start. The inject_frame.py script reads the file fresh on each invocation. This means changes to frame.md take effect immediately without restarting Codex — the next time a matched write tool fires the hook, the updated frame text will be injected.
For the full layer definitions with worked examples and detailed guidance on applying each layer, see Provenance Layers.

inject_frame.py

The hook script that reads this file and emits it as a JSON payload.

hooks.json

The hook wiring manifest that triggers inject_frame.py before write tools.

SKILL.md

The full session-level operating frame that frame.md reinforces.

Plugin Structure

The full plugin file layout and manifest files.

Build docs developers (and LLMs) love

Get started for freeTalk to us