ai-agent-book/chapter9/hermes-self-evolution/review_task_8.md

Eighth independent review: pre-existing sidecar composition

The fifth terminal acceptance review rejected the candidate after all focused checks passed. Continue the same self-update and correct this production-path collision:

• append_persistent_model_status() currently treats any existing non-empty api_content as an already installed status projection and returns it unchanged.
• Hermes already uses api_content for ordinary API-only composition, including memory/plugin prefetch context and sanitization-divergence replay. build_turn_context() can populate the current user sidecar before the model status helper runs. On that normal path, enabling the feature silently adds no <agent_status> block.
• Existing tests cover only messages with no pre-existing sidecar.

Compose safely with pre-existing string sidecars while remaining idempotent. Distinguish a projection owned by this feature from unrelated API-only content using a bounded, deterministic representation that survives persistence and reload. Never parse or delete clean transcript content. If using an ownership marker/suffix, recognize only the exact feature-owned terminal form and preserve the pre-existing sidecar byte-for-byte; malformed/lookalike user or plugin text must be treated as ordinary base content rather than destructively replaced. Retain fail-closed handling for unsupported types and stable-row database backfill.

Add regression coverage for a realistic pre-existing string api_content sidecar (representing memory/plugin prefetch), repeated same-message request builds, database close/reopen, and exact byte preservation of the original sidecar plus exactly one status projection. Include a malformed/lookalike marker case if the chosen ownership scheme can collide with ordinary content.

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