session-cabinet

Auto-files Claude Code session transcripts into a searchable, date-anchored archive.

What it does

Claude Code stores every session as a UUID-named .jsonl file under ~/.claude/projects/. These files are opaque by default — no titles, no topics, no way to find "that deploy session from three weeks ago."

session_cabinet.py scans those transcripts and files each one into:

~/Desktop/Claude Code Sessions/
  2026/
    06-June/
      2026-06-03_1420__nginx-deploy-fix/
        SESSION.md          ← frontmatter card + summary
        manifest.log        ← one-line event record
        docs/
        artifacts/
        screenshots/
        notes/
    _INDEX.md               ← per-month table (Date · Title · Projects · Folder)
  _INDEX.md                 ← master table (all sessions)

The folder name is derived from the AI-generated title or the first real thing you typed — not the UUID. Every session is findable by date, topic, or keyword grep on the SESSION.md cards.

CLI

# Dry-run (no writes — shows what would be filed)
python3 session_cabinet.py --dry-run

# File everything
python3 session_cabinet.py

# Verbose (shows each session path)
python3 session_cabinet.py --verbose

# Limit to N sessions (useful for testing)
python3 session_cabinet.py --limit 20

# Custom cabinet location
python3 session_cabinet.py --cabinet ~/Documents/SessionArchive

# Or via env var
CC_CABINET_DIR=~/Documents/SessionArchive python3 session_cabinet.py

# Override timezone (default: system local)
python3 session_cabinet.py --tz -5:00

# Recover sessions from an external drive / server backup
python3 session_cabinet.py --source-dir /Volumes/Backup/.claude/projects \
                           --source-label server

Dry-run contract

--dry-run prints a full report — counts of new/migrated/updated/skipped sessions, any day-burst dates, the cabinet tree — without writing a single file. Always run dry first on a new machine.

Folder layout

<CABINET>/
  YYYY/
    MM-MonthName/
      YYYY-MM-DD_HHMM__topic-slug/   ← one folder per session
        SESSION.md                   ← YAML frontmatter + summary
        manifest.log                 ← event log
        docs/                        ← subfolder stubs (empty at creation)
        artifacts/
        screenshots/
        notes/
      _INDEX.md                      ← month table
    [DD/]                            ← day-burst dir when >6 sessions/day
  _INDEX.md                          ← master table

Scheduling (hourly via launchd)

Run bash setup.sh to install a launchd agent that runs the filer hourly. The agent uses the local label io.ariacode.session-cabinet.

To check status after install:

launchctl list | grep session-cabinet
cat ~/Library/Logs/session-cabinet.log

Safety model

• No deletion. Sessions are never deleted — only filed or moved.
• Idempotent. Re-running is safe; already-filed sessions are skipped.
• No LLM calls. Everything is heuristic parsing of the JSONL structure.
• Stdlib only. No pip install required.
• Portable. All paths derived from pathlib.Path.home(). Works on any

user's machine; no hardcoded paths or usernames.