Caprock Files

Added

• A session opening a repository you have worked in before already knows where you stopped. Open claude, and the agent has been handed the last substantial thing an agent said in that repository — no command, no prompt, nothing to press. Verified against a live session, which named the previous session's work while forbidden to read files or git.

The first user to live in Caprock for a week, asked why free tools were unusable, did not say "no stats" — he said they lose context. Measured here before anything was written: 120 of 151 sessions (79%) open in a repository that already had sessions, 36 of them after more than a day. Claude Code keeps only your own prompts across projects and prunes transcripts after 30 days; Caprock holds what the agent said, since May, across 31 repositories. The record was already there — nothing read it back.

Recency beat retrieval, which was a surprise. The first design searched prior prose by the words of the opening prompt: it helped in 4 of 15 resumed sessions and missed the clearest case there is, because an opening question shares no words with its own answer. The last substantial passage answers 12 of 19, so the clever version is not built.

Bounded on purpose: only a genuinely new session (not a resume, not after compaction — that would be noise growing with the session), at most ~1200 characters, nothing shorter than 400 ("Done." says nothing), nothing older than a fortnight. The text says where it came from, because an agent handed an unattributed paragraph treats it as an instruction. Nothing is typed into any session: Claude Code asks its hook a question, and this is the answer.

The status screen says whether it can speak, and for how many repositories — a feature that acts before you type is one nobody can otherwise see working.