protect-knowledge-boundary

name	protect-knowledge-boundary
description	Use when editing, moving, reviewing, or preparing to commit WiFi Lens documentation and Agent assets that mention Pro, paid editions, private modules, cross-repository references, AGENTS.md, CLAUDE.md, .agents/, or docs/.

Protect Knowledge Boundary

Keep private Pro implementation knowledge inside the Pro/ submodule while allowing the public repository to index approved private entrypoints.

Required context

Read references/boundary-policy.md completely before reviewing or changing knowledge-boundary content. Do not load private Pro documentation into Agent context unless the task is explicitly Pro-scoped. The deterministic scanner may compare private documents locally; it reports only paths and violation categories, never private passages.

Workflow

When this workflow is being used to gate a requested commit, first follow the per-commit consent protocol in .agents/references/collaboration-rules.md. Do not run steps 3 or 4 until the user chooses checks for that commit.

1. Treat the root repository as public and Pro/ as a separate private repository. Inspect their Git status and diffs separately.

2. Review every new public Pro reference against the policy. Existence and approved entrypoint paths are allowed; implementation knowledge is not.

3. Run the deterministic public-content scan:

   python3 .agents/skills/protect-knowledge-boundary/scripts/check_public_knowledge.py
   

4. Run the independent integrity check:

   python3 .agents/skills/protect-knowledge-boundary/scripts/verify_integrity.py
   

5. Report root and Pro results separately. Do not claim completion while either checker fails or a semantic boundary question remains unresolved.

Decision rules

• PASS: deterministic checks pass and manual review finds no private detail.
• REVIEW: a new reference is ambiguous. Stop and ask the user; do not infer permission to expose more context.
• FAIL: private detail, copied private text, invalid private path, missing protection, broken Claude link, or integrity mismatch exists.

Never fix a failure by weakening a rule, excluding a new path, updating hashes, or rewriting the instruction anchor as part of an ordinary check. Changes to protected assets require explicit user approval and a focused review. Generate new hashes only after that review, then run all Skill tests and both checks.

Integrity limits

Local hashes make tampering visible during normal work; they cannot stop a committer who deliberately changes every local anchor. Trusted CI plus CODEOWNERS approval is required for merge-level enforcement.