Find Session

Past sessions are on disk as transcripts. Search what the user typed (not tool output), rank by recency, then read the best match and brief the user on it.

Transcripts: ~/.claude/projects/<cwd with every non-alphanumeric char replaced by ->/<session-id>.jsonl, one JSON record per line. User-typed messages have type == "user" and message.content as a string or a list of text parts. Skip isMeta records, tool results, and injected text (<command-...>, <system-reminder>, <task-notification>). Subagent transcripts sit one level deeper and are skipped.

1. Search

Turn the user's words into a case-insensitive regex: 2-4 distinctive keywords, | between them, \b in front so trip doesn't hit strip. Requires jq.

P=~/.claude/projects/$(pwd | sed 's/[^A-Za-z0-9]/-/g')   # all projects: P=~/.claude/projects/
RE='\b(stripe|webhook|retry)'
grep -lisE "$RE" "$P"*/*.jsonl | while read -r f; do
  jq -r --arg re "$RE" --arg f "$f" '
    select(.type == "user" and (.isMeta | not))
    | (.message.content | if type == "string" then . else ([.[]? | select(.type == "text") | .text] | join(" ")) end) as $m
    | select(($m | test("^\\s*(<|Caveat:|Another Claude session)") | not) and ($m | test($re; "i")))
    | ($m | gsub("\\s+"; " ")) as $flat
    | ([($flat | match($re; "i").offset) - 50, 0] | max) as $at
    | [.timestamp[0:16], $f, $flat[$at:$at + 160]] | @tsv' "$f" 2>/dev/null
done | sort -r | awk -F'\t' '
  !($2 in hits) { order[++n] = $2; last[$2] = $1 }
  { if (++hits[$2] <= 3) snips[$2] = snips[$2] "\n    " substr($1, 1, 10) "  " $3 }
  END { for (i = 1; i <= n && i <= 8; i++) { s = order[i]; printf "\n%s  %d hits  %s%s\n", last[s], hits[s], s, snips[s] } }'

Prints up to 8 sessions, newest last-hit first, each with its hit count, path, and three snippets. "$P"* also covers nested dirs and worktrees of this project. The current project takes about a second; all projects can take ~30s.

• The newest hit is often this session (the user's own question). Skip it.
• Nothing found: try synonyms, drop to one keyword, then offer to widen to all projects.
• A link or exact string may only appear in Claude's replies or tool output. Search the raw files: grep -lF 'https://example.com/x' "$P"*/*.jsonl | xargs ls -t.

2. Read the best match

F=<path from step 1>; O=$(mktemp)
jq -r 'select(.type == "custom-title" or .type == "ai-title") | .customTitle // .aiTitle' "$F" | tail -1
jq -r '
  select((.type == "user" and (.isMeta | not)) or .type == "assistant")
  | (.message.content | if type == "string" then . else ([.[]? | select(.type == "text") | .text] | join(" ")) end) as $m
  | select(($m | length) > 0 and ($m | test("^\\s*(<|Caveat:)") | not))
  | "\n## \(if .type == "user" then "USER" else "CLAUDE" end) \(.timestamp[0:16])\n\($m[0:1500])"' "$F" > "$O"
wc -c "$O"; head -c 6000 "$O"; tail -c 25000 "$O"

The head shows what the session set out to do; the tail shows where it ended. Read the middle around the step-1 hits if the decision you need is there. If two sessions both look right, read both before answering.

3. Brief the user

<title or first ask>, <date>, <session id>
Goal:       one line
Decided:    rulings and the reasons given
Open:       what was queued or left unanswered
Last state: where the conversation stopped (as of that date)

It is a transcript, not current state. Check the repo before you rely on any "done" it reports. To keep working from it, claude --resume <session-id> reopens it, or continue here from the brief.

Common mistakes

• Searching tool output by default. Old file dumps match every keyword. Search what the user typed; go raw only for links and exact strings.
• Summarizing from snippets. Snippets choose the session; read the transcript before you brief.
• Treating the transcript as truth. Decisions hold; status has drifted. Say what it showed and when.