Erodenn/fetch-guard

What's new

Auto size guard

The pipeline now enforces default size limits to prevent oversized content from reaching LLM context windows:

• 2MB pre-extraction limit — fires before any processing; applies to all content types
• 20KB post-extraction limit — fires after extraction, with a suggested max_words value in the error message

Both limits are bypassed when max_words is set, giving callers explicit control when needed.

Tighter injection pattern

The act_as_if pattern now only fires when followed by an AI role word (model, assistant, agent, bot, gpt, claude, dan, etc.), reducing false positives on phrases like "act as a guide" or "act as a bridge."

Breaking changes

None for callers using max_words. Callers fetching large pages without max_words will now receive a FetchError with a suggested override value instead of receiving oversized content.