ai-agent-book/chapter9/hermes-self-evolution/review_task_4.md

Fourth independent review: tool-result replay and output bound

The terminal acceptance reviewer rejected the current candidate after running the requested tests and inspecting the production path. Continue the same self-update and correct these exact defects without expanding scope:

1. Status is currently appended to source_messages[-1] / api_messages[-1] without requiring a user message. After a tool call the newest message is a tool result, but production replay restores api_content sidecars only for user and assistant messages. A sidecar attached to the tool result therefore disappears on the next request, breaking the claimed byte-identical replay. Make attachment and replay behavior consistent with the real persisted message contract. Preserve role ordering and keep clean transcript content unchanged.
2. build_model_status_context() converts the TODO item_id to a string without truncating or otherwise bounding it. Enforce a deterministic total output bound, including adversarially long identifiers and descriptions.
3. Add behavior-contract tests that would have caught both defects: a successive-request sequence whose newest message is a tool result, and an oversized TODO identifier. Exercise the same production helpers/type checks used by conversation_loop.py.

Run the focused tests, the existing replay/cache regression set, compilation, and git diff --check. Update BOOK_SELF_EVOLUTION_REPORT.md with this fourth review round and exact results. Do not edit the book, commit, push, or claim downstream task improvement.