Content Protection

Claudux protects manually curated content from model modification. This page says exactly what is mechanically enforced, what is prompt guidance, and how to tell the difference — so you can decide what to trust it with.

The two modes

With a committed docs-structure.json (enforced)

The manifest flips generation into its guarded mode:

• The model runs read-only (--allowedTools "Read"); it proposes section-scoped patches and cannot write files itself.
• Code validates the entire section-patch batch before touching a target file, then stages and commits the target documentation files transactionally.
• Every patch must pass the single-section boundary, incremental impact allowlist, read-only target, heading-boundary, and safe-path checks.
• Skip-marker blocks, pinned sections, and explicit read-only sections are sha256-hashed in a guard snapshot before generation; if a guarded block changed, generation aborts with Protected documentation structure changed during generation.

This is the mode to use when protection matters. Commit the manifest.

Without a manifest (guidance)

The first generation pass lets the backend edit documentation paths directly. Prompts still steer it toward the intended docs work, but claudux does not pretend prompt wording is the security boundary.

The repository boundary is still mechanically enforced in a Git checkout:

• Bash is disallowed for Claude generation, and Codex runs with approval_policy="never" inside its configured sandbox.
• Before generation, claudux snapshots HEAD and every existing dirty path outside the documentation allowlist.
• After generation, any new source mutation, source commit, rename escape, or change to a pre-existing unrelated dirty path fails the run and is restored.
• Generated documentation changes remain in the working tree for review.

This protects source code from a documentation run. It does not stop the backend from broadly rewriting one documentation page while generating another. Use the manifest when documentation sections themselves need a hard boundary.

Skip markers

Protect specific blocks within files:

<!-- skip -->
This block is hash-guarded in manifest mode.
<!-- /skip -->

Language	Start marker	End marker
Markdown / HTML / XML	<!-- skip -->	<!-- /skip -->
JS / TS / Swift / Go / Rust / Java / C++	// skip	// /skip
Python / shell / Ruby	# skip	# /skip
CSS / SCSS	/* skip */	/* /skip */
SQL	-- skip	-- /skip

Markers must sit on their own line. In manifest mode, every skip block's hash is captured in the guard snapshot, so a silent edit to a protected block fails the run. Without a manifest, markers are guidance the model is prompted to respect.

Best practices

Commit the manifest. docs-structure.json is what turns protection from guidance into enforcement. Everything else on this page is secondary.

Protect only what needs protection. Wrap the sensitive block, not the whole file, so the rest of the page keeps getting maintained:

## Public API Documentation

This section documents our public API endpoints.

<!-- skip -->
### Internal Debugging Endpoints
These are for development use only...
<!-- /skip -->

Commit your markers so protection is consistent for everyone who runs claudux on the repo.

Review the diff. Protection narrows what the model can change; it does not replace reading git diff docs/ before you commit.

Removing protection

Delete the marker pair and run claudux update — the section becomes maintainable again. In manifest mode, update the manifest entry if the section was pinned there.