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

Independent review of the first candidate

An independent reviewer rejected the first candidate pending correction. Read the current diff and BOOK_SELF_EVOLUTION_REPORT.md, then address these exact findings without expanding scope:

1. The request-local status is not prefix-stable across the tool loop. On call 1 it is appended to the user API copy; on call 2 that historical user copy loses the status. On later calls, the previous tool result loses its old status before the newest tool result gains a new one. This changes prior wire bytes and invalidates the cached suffix, conflicting with Hermes' byte-stability rule. Use a persistent-append design (for example, stable api_content sidecars) or another design that proves previously sent wire messages remain byte-identical. Keep clean transcript content unchanged.
2. The tests exercise only the formatter. Add behavior-contract coverage for the actual attachment/replay mechanism over at least three successive API requests. Assert that each earlier wire message is byte-identical in every later request and that the newest status is closest to generation.
3. Correct the report's contradictory verification wording: the repository wrapper was blocked because its environment lacked pytest, but the focused tests did execute and pass through uv run --with pytest.

Run the focused tests, relevant existing replay/cache tests, compilation, and git diff --check. Update the report with the review round and exact results. Do not claim end-to-end benefit without an ablation campaign, do not push or commit, and do not edit the book repository.