Unclear It is not clear what turns typedMarks on or where the fingerprints are sent.
What
Claude Code has new code, active only when an internal typedMarks switch is true, for recording which text in a conversation a person actually typed.
- User messages entered from the input box carry a mark (
typed_marks) listing the typed text. Accepting a plan with cleared context carries a mark with the text labelled asplan. No mark is added when the input came from an accepted suggestion. - Feedback given in host dialogs is also tagged.
- A new function builds lists of typed, command, accepted and instruction text for a conversation, each stored as a SHA-256 hash (a one-way fingerprint of the text rather than the text itself). It skips headless, background and nested sessions, and a list that grows past its limit is reported as unknown with the reason
cap. - When accepting a prompt suggestion, submitted text is now normalised and compared with the suggestion, so typing out a copy of the suggestion counts as accepting it. Before, it was a plain exact comparison.
- A key press that delivers more than a few characters at once is now treated as pasted or typed text and sent to a new queue.
Why
This appears to be groundwork for telling text a person typed apart from text that came from elsewhere. Nothing has been read about whether the typedMarks switch is on, so it may have no visible effect yet.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
It is not clear what turns typedMarks on or where the fingerprints are sent.