Observations
Observations pair each edit with the reasoning recorded near it, plus flags.
Observations is where an edit stops being a diff and becomes something you can judge β because it's paired with the reasoning the agent recorded next to it, and flagged when it contains something worth a second look.
Reasoning beside each edit
Observations pairs each edit with the reasoning the agent recorded near it in the transcript, flags notable content such as debug statements, possible secrets, and large deletions, and shows each file's review history across sessions. Adjacent same-file edits are coalesced into one run, so a heavily edited file reads as a story rather than a wall of rows.

Context
Its Context section names what shaped the session β skills invoked, plans written, memory read, a resumed compaction β and labels every entry with how it is known: transcript when the session demonstrably used it, file-present when the file merely exists where Claude Code loads it. The two are never merged, because "used" and "was available" are different claims.
A possible-secret flag or a large-deletion flag points your attention; it doesn't decide for you. The point is that nothing notable slips through unread in a long session.