Checkpoint: Save & Restore Task State

Product-aware runner

Set these once for the active host. checkpoint.ts uses claudePath, so pass the matching state root on every invocation.

# Standalone Codex
CHECKPOINT_RUNNER="${CODEX_HOME:-$HOME/.codex}/darkroom/source/src/scripts/checkpoint.ts"
CC_STATE_ROOT="${CODEX_HOME:-$HOME/.codex}/darkroom/state"

# Claude Code instead uses:
# CHECKPOINT_RUNNER="$HOME/.claude/src/scripts/checkpoint.ts"
# CC_STATE_ROOT="$HOME/.claude"

Subcommands

save [label]

Save current state with an optional label.

CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" save "Completed phase 3 migration"

list

List all checkpoints for the current project.

CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" list

show [checkpoint-id]

Show details of a specific checkpoint.

CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" show chk-20240115-103000

restore [checkpoint-id] [--force]

Real, opt-in rollback — restores tracked files (working tree + index) to exactly what they were at save time, then reapplies whatever uncommitted changes existed at that moment (captured as a patch when the checkpoint was saved). Untracked files are never touched or deleted; if any existed at save time they're just listed as a warning (their content was never captured). Scope note: restore reconstructs file content only — anything that was staged at save time comes back as unstaged working-tree changes (the patch reapplies to the worktree; index state is intentionally not reconstructed).

Safety rails:

• Auto safety-checkpoint first. Before anything is touched, the current state is saved as its own checkpoint and its id is printed — restore is itself reversible by restoring that id (with --force).
• Branch/sha guard. If the current branch or HEAD differs from what the checkpoint recorded, restore refuses and tells you to check out that branch or pass --force.
• Legacy checkpoints (saved before this patch-capture feature existed) have no recorded diff to restore from. Restoring one prints "legacy checkpoint: metadata only, nothing restored" and falls back to the old print-only behavior — review the dumped JSON and continue manually.

# Restore latest
CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" restore

# Restore specific
CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" restore chk-20240115-103000

# Restore across a branch/sha mismatch
CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" restore chk-20240115-103000 --force

clean

Remove old checkpoints, keeping the last 10.

CC_SETTINGS_HOME="$CC_STATE_ROOT" bun "$CHECKPOINT_RUNNER" clean

Examples

User: "save a checkpoint"
  -> /checkpoint save

User: "checkpoint before this refactor"
  -> /checkpoint save "Before auth refactor"

User: "list my checkpoints"
  -> /checkpoint list

User: "restore from last checkpoint"
  -> /checkpoint restore

User: "show checkpoint details"
  -> /checkpoint show <id>

User: "clean up old checkpoints"
  -> /checkpoint clean

Storage

Checkpoints are stored under $CC_STATE_ROOT/checkpoints/<project-name>/ as JSON files with a latest symlink pointing to the most recent. Each checkpoint with uncommitted changes at save time also gets a sibling chk-<id>.patch file (git diff HEAD output) — this is what makes restore a real rollback instead of a metadata dump. Checkpoints saved before this feature existed have no patch file and restore in metadata-only (legacy) mode.