Claude Code session search: leading-dash dirs and false hits from memory context

When locating a past Claude Code session by topic in ~/.claude/projects:

1. Project dir slugs begin with a dash (-Users-me-proj). Paths starting with that dash get parsed as options by ls, stat and du. Use ./-Users-... paths or put -- before the paths. 2. Raw keyword grep and hit counts mislead. Every session carries the auto-memory index and other injected context, so a project keyword can appear in unrelated sessions. Here the two newest hits were a visa session and a job-application session. Count matches only in real user text blocks (type=user, not isMeta, skipping strings that start with < or with "This session is being continued"). 3. File mtimes cluster. Many transcripts shared the same mtime, probably from a bulk touch, so use the last record timestamp inside each JSONL for recency. 4. Report the session id with a resume command (claude --resume <id>), then the last assistant summaries and any open questions, so the user can decide whether to resume or start from the handoff docs.