Book Summarizer AI Agent

A multi-stage sub-agent pipeline that reads whole PDF books and produces summaries you can rely on: a GAME-loop planner (BookExplorer) decides whether to skip, skim or read each page; a NoteTaker writes per-chunk notes; and a SummaryComposer picks a retrieval, single-call or chunked-merge strategy based on corpus size. Each role runs on an independently swappable backend.

Because users upload the documents, the interesting work is defensive: untrusted-content tagging, a per-run canary tripwire that catches and discards leaked instructions, and Pydantic-validated tool inputs that bound the blast radius — with a red-team test suite running on every push. Summary quality is scored by an LLM-as-judge faithfulness pipeline (manual-trigger only, by design), and the whole system is failure-tolerant: pre-emptive rate limiting, retry with backoff, auto-split on quota overflow, resumable state across crashes, and per-run token-cost tracking.