Claude Sessions

Read Claude Code's own session transcripts as swamp data.

Claude Code appends one JSONL transcript per session under ~/.claude/projects//.jsonl. That log is write-only in practice — there is no way to ask "which of my sessions actually finished?" without reading raw JSON by hand. This model parses those transcripts into queryable resources:

• index — per-project session counts, bytes and last activity (stats only, no parsing, so it stays cheap on a large history)
• listSessions — one session resource per transcript: ai-title, cwd, git branch, CLI version, models used, start/last-activity, span, staleness, turn and tool-call counts, top tools, the opening prompt and the closing recap
• search — full-text search across transcripts with snippets
• triage — the worklist method: joins sessions to the kanban cards created by the mk-session-card hook (state files under ~/.claude/mk-session-cards/) and ranks them by a completion verdict

The completion verdict reads the session's LAST assistant message, which by the hook's convention is a recap of what was done and what is left. It is a ranking aid, not a judgment: likely-complete, likely-open, unclear, or hook-noise when the session ended on card bookkeeping rather than on a work recap. Every row carries the closing text as evidence so a human or agent makes the actual call. Sidechain (subagent) turns are excluded so a chatty subagent cannot become the session's last word.

All methods fan out over every matching session in a single run, so a whole history is one method call and one model lock.