Module sanitize 

Two-layer sanitize pipeline for fetched URL content.

Layer order is fixed:

1. regex_layer::strip — fast, local, Aho-Corasick over a static pattern list. Replaces matches with the literal [STRIPPED].
2. semantic_layer::review — optional LLM review via Anthropic Messages API. Fails open when ANTHROPIC_API_KEY is unset or the API returns non-2xx: regex-only output is passed through unchanged.
3. wrap::envelope — wrap into <untrusted_content> so the caller LLM physically separates instruction from data.

Modules

regex_layer

1st-layer sanitize: static Aho-Corasick pass over the most common instruction-override / role-token markers.

semantic_layer

2nd-layer sanitize: LLM-backed semantic review via Anthropic Messages API.

wrap

Wrap sanitized content into an <untrusted_content> envelope.

Functions

run

Run the 2-layer sanitize pipeline and return (envelope, removed).