Summary

_spawn_background_review() forks a review agent seeded with a snapshot
of the current conversation (messages_snapshot) plus a review prompt,
then scans _session_messages for tool results that indicate saves. The
bug: the scan iterated the entire _session_messages list, including
the snapshot entries from the main conversation. Tool results in those
old messages (e.g. earlier memory/skill saves from a previous cycle)
were re-scanned and re-triggered the 💾 save notification on every
review cycle, even when the review agent saved nothing new.

Root cause: _spawn_background_review() in run_agent.py passed
the full _session_messages to the action scan without accounting for
the snapshot_len prefix already present from the forked context.

Fix: Slice _session_messages[snapshot_len + 1:] so only messages
produced by the review agent itself are scanned — skipping the
snapshot_len snapshot entries and the +1 review prompt injected before
the agent ran.

Changes

• run_agent.py — _spawn_background_review(): record snapshot_len = len(messages_snapshot) before forking; slice _session_messages by
  snapshot_len + 1 when collecting actions to notify on

Test plan

• Manually verified: save notification fires once per genuine save cycle,
  not on repeat cycles when no new saves occurred
• No automated tests added (existing suite requires dependencies not
  available in this environment)