A /clear hook that archives my Claude conversations

I type /clear several times a day. Each time, the session is effectively gone: the JSONL transcript sits under ~/.claude/projects/, but I never look there.

What I wanted was exactly this: every /clear automatically commits the conversation to a git repo. Nothing to remember, nothing to run.

AI Disclaimer: I used an LLM to correct orthography and general styling on this post.

The hook

There is no PreClear event, but SessionEnd fires on /clear with reason: "clear", and the matcher can filter on it:

1
{
2
  "hooks": {
3
    "SessionEnd": [
4
      {
5
        "matcher": "clear",
6
        "hooks": [
7
          {
8
            "type": "command",
9
            "command": "~/.claude/hooks/export-on-clear.sh",
10
            "timeout": 15
11
          }
12
        ]
13
      }
14
    ]
15
  }
16
}

/export can’t be called from a hook (slash commands are interactive-only), so a small Python script renders the JSONL to text. It skips thinking blocks and subagent sidechains, clips tool output, and ignores records it can’t parse, since the format is internal and changes between releases.

Not making /clear wait

The foreground part only writes the file. Commit and push run detached, so a slow network never blocks /clear:

1
(
2
  cd "$REPO" 2>/dev/null || exit 0
3
  exec 9>"$REPO/.export.lock" 2>/dev/null || exit 0
4
  flock 9 2>/dev/null || :
5

6
  git add -A -- . >>"$LOG" 2>&1 || { log "git add failed"; exit 0; }
7
  # ... commit, then push with BatchMode and a connect timeout
8
) </dev/null >/dev/null 2>&1 &
9
disown 2>/dev/null || :

flock serialises concurrent clears across tmux panes. BatchMode=yes makes a missing SSH key fail fast; an unpushed commit goes out with the next export.

Files are named 2026-09-18-140353-dotfiles-dfdb7f3f.txt: timestamp, project, short session id. The repo is private, since transcripts can contain file contents and secrets.

1
2026-09-18 14:03:53 wrote 2026-09-18-140353-dotfiles-dfdb7f3f.txt (51667 bytes)
2
2026-09-18 14:03:53 pushed 2026-09-18-140353-dotfiles-dfdb7f3f.txt to origin/main

Alternatives

None of these commit conversations on /clear. Most index or render transcripts in place, so they don’t survive rm -rf ~/.claude or sync across machines.

ai-session: stores transcripts as git artifacts linked to commits and PRs.

claude-code-research-transcript-hook: archives sessions into the project as Markdown with research metadata. When I tried it, it couldn’t find my transcript, crashed with SameFileError, and copied notes from unrelated projects into the archive.

claudesessions: auto-archives sessions with full-text search and restore.

claude-transcript: converts JSONL transcripts to Markdown.

ccexport: Ruby gem that exports transcripts to Markdown and HTML.

claude-lens: MCP server with a SQLite full-text index over past sessions.

ClaudeHistoryMCP: MCP server for searching conversation history.

session-indexer: local semantic search over sessions via Ollama embeddings.

What’s left

Search. grep over the exports goes far, but an index with a one-line summary per export would help. If I tire of that, I’ll point claude-lens at the exports directory.

The scripts live in my dotfiles.