Feature: fuzzy search for --resume (session picker)

Problem

claude --resume <session-id> requires knowing the exact session ID. In practice, users need to find sessions by what they discussed, not by UUID. There's no built-in way to search, browse, or fuzzy-find past sessions.

With 300+ sessions across multiple projects, finding "that planning session from this morning" means manually scanning .claude/projects/ directories or grepping JSONL files.

Proposed Solution

Add an interactive session picker when --resume is called without an argument (or with a query):

claude --resume              # browse all sessions, sorted by recency
claude --resume "auth bug"   # search sessions by content

Key UX behaviors:

• Full-text search across session content (user messages, assistant responses, summaries)
• Exact phrase support: "week 12" finds that exact phrase, not just sessions with "week" and "12" separately
• Field-weighted ranking: matches in titles/first messages rank higher than matches buried in conversation
• Cross-project: searches all projects, shows project labels for context
• Live preview: see conversation excerpts before resuming
• Recency-aware: among equal relevance, newer sessions surface first

What this looks like

sessions > auth migration

>  today  Fix the login redirect bug — users are getting stuck in a loop...
    1d   Plan: Migrate to React Server Components        · webapp
    6d   Week 12 Sprint Planning                         · webapp
   today  Morning standup — what's on my plate today?    · Personal

  4 matches · ctrl-/ toggle preview

Typing updates results live (not just filtering — re-ranking). Selecting a session resumes it.

Working Implementation

I built this as a standalone tool: claude-sessions

Architecture:

• SQLite FTS5 for full-text search (ships with Python's stdlib sqlite3 — no external dependencies)
• BM25 ranking with field weights (title 10x, preview 5x, body 1x)
• Incremental indexing: ~180ms warm sync for 383 sessions, ~4ms search
• fzf as the interactive UI with --phony + change:reload for live re-searching

The session data is already there (JSONL files in .claude/projects/). The indexer extracts text from user messages and assistant responses, stores it in an FTS5 virtual table, and searches it with BM25 ranking. The whole thing is ~400 lines of Python with no dependencies beyond stdlib.

Why this matters

Session resume is one of Claude Code's best features — but discoverability kills it. If you can't find the session, you can't resume it. Right now the only option is the compact --resume list which shows truncated summaries with no search. A content-searchable picker would make --resume genuinely useful for power users with many sessions.

Alternatives Considered

• Keyword extraction + TF-IDF: Built this first. Required constant patching (token length limits, custom phrase matching, stopword lists). FTS5 handles all of it natively.
• External search engine (Meilisearch, etc.): Overkill. SQLite FTS5 is zero-dependency and fast enough.
• Just use grep: Works but no ranking, no live preview, no cross-project awareness.