Checkpoint

/checkpoint — save full session state for clean restart

Why this exists

Claude Code session state is partly persistent and partly not:

Survives a restart?	Where it lives
✓ Files on disk	regular filesystem
✓ User memory	~/.claude/projects/*/memory/
✓ Conversation context	only via claude --continue / equivalent
✗ TaskList	session-scoped, dies on exit
✗ Cron jobs from CronCreate	session-scoped, dies on exit
✗ Background Bash run_in_background processes	session-scoped, dies on exit
✗ Loaded skill changes since startup	reloaded fresh on next launch
✗ Anything in /tmp not on tmpfs persistence	depends on system reboot vs Claude restart

/checkpoint is a single command that drains the not-survives column to disk so /restore (in the next session) can re-hydrate the parts that can be rehydrated and tell you about the parts that can't.

Behavior summary

When you (the agent) execute /checkpoint:

1. Wait for a safe stopping point. Don't checkpoint mid-tool-call. If a long-running background process is mid-flight (Bash with run_in_background: true), explicitly mention it to the user and ask whether to: a) wait for it to finish before checkpointing b) leave it running (it'll die at exit) and note the loss c) cancel it via TaskStop and then checkpoint Default: ask, don't decide silently.

2. Capture the data (see Steps below).

3. Report back with a clear "safe to restart now" signal AND a one-line summary of what was captured + what couldn't be.

Steps

State directory: $HOME/.claude/.checkpoint/ (mode 700). Create if absent. Delete any prior checkpoint at start — checkpoints are single-shot, not stacked. Warn the user if they're about to overwrite an unrestored prior checkpoint.

1. Tasks

Call TaskList, then for each task call TaskGet to get the full description. Write to $HOME/.claude/.checkpoint/tasks.json:

[
  {
    "subject": "Build foo",
    "description": "Full description here",
    "activeForm": "Building foo",
    "status": "in_progress",
    "owner": "agent-name-or-empty",
    "blockedBy": ["3", "5"]
  },
  ...
]

chmod 600.

2. Active loops

Reuse /loop-pause (invoke via the Skill tool). It writes $HOME/.claude/.paused-loops.json and CronDeletes each. The /checkpoint skill doesn't duplicate that work — just delegates.

3. Background processes

Capture any Bash run_in_background=true tasks that are still running. Use TaskList for the harness's task-runner tasks, OR look at the conversation for recent Bash calls with that flag and check their pid via pgrep. Write to $HOME/.claude/.checkpoint/processes.txt:

# Background processes at checkpoint time
# (cannot be auto-restarted; user must manually re-kick if needed)
2026-05-17T07:50:00Z pid=12345 cmd="nohup cargo build --release ..." cwd=/opt/vaultwarden
2026-05-17T07:50:00Z pid=23456 cmd="nohup aideinit -y -f"             cwd=/

If none: write a single line # no background processes.

4. Git state

For each directory in $HOME/projects/ (and $HOME/code/, $HOME/work/ if they exist), run git status --porcelain and write any non-clean repos to $HOME/.checkpoint/git-status.txt. If there are uncommitted changes, warn the user in the final report. Don't commit automatically — that's a user decision.

5. Handoff note

Write a short prose summary of "what was being worked on" to $HOME/.claude/.checkpoint/handoff.md. Pull from:

• the in-progress tasks (priority 1)
• the last few substantive turns of the conversation (priority 2)
• any loop session log files (priority 3)

Format: ~10 lines max. Bullet form. Read like a sticky note to the next session's agent. Example:

# Checkpoint 2026-05-17T07:50:00Z

Last session was finishing the vaultwarden install on plausiden-prime.

In flight:
- Vaultwarden running on 127.0.0.1:8888 but Caddy vhost not hooked
  in yet (waiting on cert expansion)
- LE cert needs --expand to include vault.plausiden.com
- DNS for vault.plausiden.com is propagated

Tasks list saved: 2 in_progress, 1 pending.
Active loop: 1 paused (security/forge loop, cron */1 * * * *).

Run /restore in the new session to re-hydrate.

chmod 600.

6. Manifest

Write $HOME/.claude/.checkpoint/MANIFEST.json summarizing what was captured (file list, timestamps, counts):

{
  "checkpoint_at": "2026-05-17T07:50:00Z",
  "captured": {
    "tasks": 3,
    "active_loops_paused": 1,
    "background_processes": 0,
    "dirty_git_repos": 1
  },
  "files": [
    "tasks.json",
    "processes.txt",
    "git-status.txt",
    "handoff.md"
  ]
}

7. Tell the user

Format:

Checkpoint complete.

Captured:
- 3 tasks (2 in_progress, 1 pending) → ~/.claude/.checkpoint/tasks.json
- 1 active loop (security/forge, every minute) → paused via /loop-pause
- 0 background processes
- 1 dirty git tree (claude-tools — uncommitted README.md edit)

Safe to /exit and relaunch Claude Code now.

After relaunch, run /restore to re-hydrate.

If anything was lost (a bg process that couldn't be paused, etc.), say so explicitly with a recovery hint.

Don't

• Don't write secret material to the checkpoint files. If a task description contains a secret, the checkpoint will leak it. Trust the user's task descriptions to already be clean.
• Don't auto-commit dirty git trees. Only show them.
• Don't try to capture the conversation history itself — that's Claude Code's --continue flag's job, not ours.
• Don't checkpoint twice without an intervening /restore. Refuse if an old checkpoint is still on disk (or ask to overwrite).

What this is NOT

• It's NOT a transaction log. There's no "rewind" — just save + restore + (eventually) delete-on-restore.
• It's NOT a backup tool. ZFS snapshots / sanoid handle that layer.
• It's NOT an alternative to git commits — uncommitted code stays uncommitted; we just inventory.