Source Intelligence

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

All of v2.1.205 Home All releases olderv2.1.204 v2.1.206newer

Conversation compaction prompt improvements

The compaction (context-window management) prompt was significantly revised:

  • Adds an explicit note that only genuine user-role turns count as user messages — quoted "user:…" lines inside assistant messages are never attributed as user input
  • Adds a third compaction variant for mid-session resumption (summary placed at the start of a continuing session, followed by newer messages)
  • All three compaction variants now include: "Note any security-relevant instructions or constraints the user stated... These MUST be preserved verbatim in the summary"
Evidence

User-message attribution note (search for "Only messages that actually came from the user (user-role turns) count") — G9i at line ~398382

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

See this entry in the whole of v2.1.205 →