ai-agent-book/chapter9/hermes-self-evolution/review_task_7.md

Seventh independent review: idempotent request rebuilds

The fourth terminal acceptance review rejected the candidate after all focused tests passed. Continue the same self-update and correct this retry-path defect:

• append_persistent_model_status() is not idempotent when the same newest message is used to build another request (for example, a transient provider failure, empty-response retry, or request rebuild before a new transcript message). It computes updated = base + "\n\n" + status from wire content that already contains the previous projection, so every rebuild adds another <agent_status> block.
• A direct production-helper exercise repeated the call 20 times and produced 20 status blocks, changing previously sent bytes and violating both the prompt-cache stability claim and the total output bound.
• Current tests always append a new transcript message before the next build and therefore miss same-message retries.

Make projection idempotent at the production persistence boundary. Preserve an unmodified base sidecar or safely recognize/replace only a status projection owned by this feature; do not delete user-authored text that merely resembles status markup. Decide and document whether volatile status values remain fixed for a retry of the same message or can be replaced, but previously sent prefix bytes must not drift unexpectedly. Keep clean transcript content unchanged, retain string-only fail-closed behavior, and durably persist the selected sidecar by stable row identity.

Add a regression test that repeatedly builds requests with unchanged source messages, including a persisted/reloaded case. Assert identical wire bytes and exactly one owned status block. Retain realistic tool-call placement, persistence-reload, adversarial-bound, and unsupported-type coverage.

Run the focused tests, existing sidecar/cache/turn-context regressions, compilation, and git diff --check. Update BOOK_SELF_EVOLUTION_REPORT.md with this seventh review and exact results. Do not edit the book, commit, push, or claim downstream task improvement.