Find a Claude Code Session

You are searching the user's Claude Code session transcripts to find the session(s) they're thinking of. They often have 15+ terminal windows with active sessions and remember only a fragment of what was said.

Query: $ARGUMENTS

Tool

Use the search script that ships with this config (no other transcript spelunking needed):

$HOME/.claude/bin/claude-search PATTERN [--all] [--since 6h|2d|1w] [-m N] [--paths]

• Searches the user + assistant text of session transcripts — what the user could scroll through in a terminal, not tool payloads — across Claude, Codex and Grok.
• PATTERN is a Python regex, always case-insensitive.
• Scope defaults to LIVE sessions only. Pass --all to sweep history; that is the right default for "find the session where we…" questions about anything older than the tabs open right now.
• --since 6h|2d|1w limits by how recently the transcript was touched.
• Each hit prints the session id, project directory, match count, last-active age, the opening prompt (this is what identifies the terminal window), matching lines, and the exact command that reopens it.
• Exit code is nonzero with a "no matches" message when nothing hits.

Optional local extra. If $HOME/projects/claude-beep/claude-grep exists on this machine, it can additionally point at the tab: claude-grep --mark <session-id-prefix> renames that session's terminal tab to 🎯 <project> — FOUND 🎯 and rings its bell; --unmark <prefix> restores the title. It writes to the tab's tty and may need sandbox override / permission approval. It is a separate personal tool, not part of this config — check it exists (test -x) before reaching for it, and just skip the marking step if it does not. If the user says they found the tab, unmark it.

Approach

1. Classify the query.
   • If it reads as a literal phrase or regex (quoted text, regex metacharacters, an exact error message or identifier), run it as-is — one search.
   • If it reads as a natural language description ("the session where we debugged the disk filling up"), derive 3–6 search patterns and run a search per pattern.
2. Derive good patterns from descriptions. Prefer distinctive, rare tokens over common words: filenames, function names, error fragments, domain terms, project-specific vocabulary. Include synonyms the conversation might have used instead of the user's wording (e.g. "disk full" → disk space, df -h, 100%.*use, filled up). Use alternation (foo|bar) to keep the number of runs small.
3. Widen if needed. Start with --all --since 2d for a recent-sounding query; widen to --since 1w, then to a bare --all, when that yields nothing or only weak hits. If the user's wording implies age ("a while back", "last month"), start wide. Drop --all only when the user says the session is open right now.
4. Merge and rank. Combine hits across patterns by session (the resume id). Rank by: number of distinct patterns that hit > recency > match count. A session matched by several independent patterns is almost certainly the one.
5. Point at the tab. A session that still appears in a default (live) claude-search run is open right now. If the best match is clearly the session the user described (one strong match, or several patterns agree) and the optional claude-grep is installed, run claude-grep --mark <id-prefix> so its tab lights up — that is usually what the user actually wants. If multiple candidates are equally plausible, don't mark; list them and let the user choose.

Output

Lead with the best match. For each candidate (best first, at most ~5):

• Project directory, how recently it was active
• Whether it's open right now (it appears in a live-scope search) and, if you marked it, that its tab is now named 🎯 <project> — FOUND 🎯
• Its first prompt (this is what identifies the terminal window)
• The one or two most telling matching lines
• The exact resume command: claude --resume <session-id>

If nothing matched even with --all, say so and show which patterns you tried, so the user can correct your guesses.