Module semantic_layer 

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

Configuration

• ANTHROPIC_API_KEY — required. Absent ⇒ no-op (input returned unchanged with an empty removed vector).
• AGENTSEC_LLM_MODEL — optional override for the model id; defaults to claude-haiku-4-5-20251001.

Fail-open rationale

On any of (a) missing API key, (b) non-2xx HTTP response, the layer returns the input unchanged. The 1st (regex) layer already removed the obvious markers, so this layer is a defense in depth boost — not the floor of protection. Failing closed here would mean a transient API outage breaks all URL fetches, which is worse than missing a few semantic edge cases.

Prompt and response

The prompt wraps the body in a <content> tag and asks for a strict JSON verdict: {"removed": ["span1", "span2", ...]}. The response is parsed tolerantly (prose / code-fence wrapping is stripped) so a conversational model output does not break the layer; malformed JSON degrades silently to no removals.

Functions

review

Run the semantic review layer and return (cleaned, removed).