Skill Refresh

Refresh Skill (Direct Execution)

Direct execution skill for managing Claude Code resources. Performs two operations:

1. Process cleanup: Identify and terminate orphaned Claude Code processes
2. Directory cleanup: Clean up accumulated files in ~/.claude/

This skill executes inline without spawning a subagent.

Execution

Step 1: Parse Arguments

Extract flags from command input:

• --dry-run: Preview mode
• --force: Skip confirmation, use 8-hour default

# Parse from command input
dry_run=false
force=false
if [[ "$*" == *"--dry-run"* ]]; then
  dry_run=true
fi
if [[ "$*" == *"--force"* ]]; then
  force=true
fi

Step 2: Run Process Cleanup

Execute process cleanup script, forwarding --dry-run through when set (reusing the dry_run boolean already parsed in Step 1 -- no new argument parsing):

.claude/scripts/claude-refresh.sh $( [ "$force" = true ] && echo "--force" ) $( [ "$dry_run" = true ] && echo "--dry-run" )

Store process cleanup output for display.

Step 3: Clean Orphaned Postflight Markers

Clean any orphaned postflight coordination files from the specs directory. These files should normally be cleaned up by skills after postflight completes, but may be left behind if a process is interrupted.

echo ""
echo "=== Cleaning Orphaned Postflight Markers ==="
echo ""

# Find orphaned postflight markers (older than 1 hour)
orphaned_pending=$(find specs -maxdepth 3 -name ".postflight-pending" -mmin +60 -type f 2>/dev/null)
orphaned_guard=$(find specs -maxdepth 3 -name ".postflight-loop-guard" -mmin +60 -type f 2>/dev/null)

# Also check for legacy global markers
legacy_pending=""
legacy_guard=""
if [ -f "specs/.postflight-pending" ]; then
    legacy_pending="specs/.postflight-pending"
fi
if [ -f "specs/.postflight-loop-guard" ]; then
    legacy_guard="specs/.postflight-loop-guard"
fi

if [ -n "$orphaned_pending" ] || [ -n "$orphaned_guard" ] || [ -n "$legacy_pending" ] || [ -n "$legacy_guard" ]; then
    if [ "$dry_run" = true ]; then
        echo "Would delete the following orphaned markers:"
        [ -n "$orphaned_pending" ] && echo "$orphaned_pending"
        [ -n "$orphaned_guard" ] && echo "$orphaned_guard"
        [ -n "$legacy_pending" ] && echo "$legacy_pending"
        [ -n "$legacy_guard" ] && echo "$legacy_guard"
    else
        # Delete orphaned task-scoped markers
        find specs -maxdepth 3 -name ".postflight-pending" -mmin +60 -delete 2>/dev/null
        find specs -maxdepth 3 -name ".postflight-loop-guard" -mmin +60 -delete 2>/dev/null

        # Delete legacy global markers
        rm -f specs/.postflight-pending 2>/dev/null
        rm -f specs/.postflight-loop-guard 2>/dev/null

        echo "Cleaned orphaned postflight markers."
    fi
else
    echo "No orphaned postflight markers found."
fi

Step 4: Reap Stale Task Locks

Sweep specs/ for stale task-number .lock directories (see context/patterns/task-lock.md's Reap Contract section for the full threshold reasoning) and report every one found. This is a distinct cleanup target from Step 3's postflight markers, added alongside it rather than merged into it. Reuses the dry_run boolean already parsed in Step 1 -- no new argument parsing.

echo ""
echo "=== Reaping Stale Task Locks ==="
echo ""

if [ "$dry_run" = true ]; then
    .claude/scripts/task-lock.sh reap --dry-run
else
    .claude/scripts/task-lock.sh reap
fi

The per-lock detail (task number, session id, operation, age) is produced by task-lock.sh itself; echo its output verbatim rather than summarizing it away, matching the reap subcommand's own per-item reporting contract.

Step 4.5: Reap Stale Session-Scoped Orchestration Files

Sweep specs/ for stale session-scoped specs/.orchestrator-multi-state-{session_id}.json and specs/.return-meta-multi-{session_id}.json files (see context/standards/orchestrator-runtime-files.md's Class Table) and report every one found. This is a distinct cleanup target from Step 4's task-lock reap — session-scoping the two repo-level batch-orchestration singletons trades collision risk for unbounded litter if an abandoned batch's file is never swept, so this step exists to bound that litter. Uses the X.5 numbering deliberately so Steps 5-7 below keep their existing numbers and no cross-reference to them in refresh.md needs to change. Reuses the dry_run boolean already parsed in Step 1.

echo ""
echo "=== Reaping Stale Session-Scoped Orchestration Files ==="
echo ""

if [ "$dry_run" = true ]; then
    .claude/scripts/reap-session-runtime-files.sh --dry-run
else
    .claude/scripts/reap-session-runtime-files.sh
fi

The per-file detail (filename, embedded session id, age in minutes) is produced by reap-session-runtime-files.sh itself; echo its output verbatim rather than summarizing it away, matching Step 4's own "echo verbatim" instruction. This cleanup runs only on explicit /refresh invocation, not on the hourly systemd cadence.

Step 4.6: Reap Stale Session Registry Entries

Sweep specs/.sessions/ for stale in-flight orchestration session registry entries (see context/patterns/task-lock.md's Session-Registry CLI section) and report every one found. This is a distinct cleanup target from Step 4's task-lock reap and Step 4.5's session-scoped orchestration-file reap — the session registry (specs/.sessions/{session_id}.json) is a separate, additive mechanism produced by task-lock.sh session-register/session-heartbeat, not one of the two files Step 4.5 sweeps. Uses the X.6 numbering deliberately so Steps 5-7 keep their existing numbers and no cross-reference to them in refresh.md needs to change. Reuses the dry_run boolean already parsed in Step 1 — the same --dry-run passthrough branch structure as Step 4.5.

echo ""
echo "=== Reaping Stale Session Registry Entries ==="
echo ""

if [ "$dry_run" = true ]; then
    .claude/scripts/task-lock.sh session-reap --dry-run
else
    .claude/scripts/task-lock.sh session-reap
fi

The per-entry detail (session id, command, task numbers, age, reap reason) is produced by task-lock.sh session-reap itself; echo its output verbatim rather than summarizing it away, matching Step 4 and Step 4.5's own "echo verbatim" instruction. This cleanup runs only on explicit /refresh invocation, never on the hourly claude-refresh.timer cadence, which runs process cleanup only and does not sweep specs/.

Step 5: Clean Stale Backup Files

Scan for and remove any .backup files left over from the deprecated backup mechanism in .claude/:

echo ""
echo "=== Cleaning Stale Backup Files ==="
echo ""

# Find .backup files in .claude/ directory
backup_files=$(find .claude/ -name "*.backup" -type f 2>/dev/null)

if [ -n "$backup_files" ]; then
    backup_count=$(echo "$backup_files" | wc -l)
    if [ "$dry_run" = true ]; then
        echo "Would delete $backup_count .backup file(s):"
        echo "$backup_files"
    else
        echo "$backup_files" | xargs rm -f
        echo "Deleted $backup_count stale .backup file(s)."
    fi
else
    echo "No stale .backup files found."
fi

Step 6: Run Directory Survey

Show current directory status without cleaning yet:

.claude/scripts/claude-cleanup.sh

This displays:

• Current ~/.claude/ directory size
• Breakdown by directory
• Space that can be reclaimed

Step 7: Execute Based on Mode

Dry-Run Mode

If --dry-run is set:

echo ""
echo "=== DRY RUN MODE ==="
echo "Showing 8-hour cleanup preview..."
echo ""
.claude/scripts/claude-cleanup.sh --dry-run --age 8

Exit after showing preview.

Force Mode

If --force is set:

echo ""
echo "=== EXECUTING CLEANUP (8-hour default) ==="
echo ""
.claude/scripts/claude-cleanup.sh --force --age 8

Show results and exit.

Interactive Mode (Default)

If neither flag is set:

1. Check if cleanup candidates exist (claude-cleanup.sh exits with code 1 if candidates found)

2. If no candidates, display message and exit:

No cleanup candidates found within default thresholds.
All files are either protected or recently modified.

3. If candidates exist, prompt user for age selection:

{
  "question": "Select cleanup age threshold:",
  "header": "Age Threshold",
  "multiSelect": false,
  "options": [
    {
      "label": "8 hours (default)",
      "description": "Remove files older than 8 hours - aggressive cleanup"
    },
    {
      "label": "2 days",
      "description": "Remove files older than 2 days - conservative cleanup"
    },
    {
      "label": "Clean slate",
      "description": "Remove everything except safety margin (1 hour)"
    }
  ]
}

4. Map user selection to age parameter:

   • "8 hours (default)" → --age 8
   • "2 days" → --age 48
   • "Clean slate" → --age 0

5. Execute cleanup with selected age:

case "$selection" in
  "8 hours (default)")
    .claude/scripts/claude-cleanup.sh --force --age 8
    ;;
  "2 days")
    .claude/scripts/claude-cleanup.sh --force --age 48
    ;;
  "Clean slate")
    .claude/scripts/claude-cleanup.sh --force --age 0
    ;;
esac

6. Display cleanup results

Example Execution Flows

Interactive Flow

# User runs: /refresh

# Output:
Claude Code Refresh
===================

No orphaned processes found.
All 3 Claude processes are active sessions.

---

Claude Code Directory Cleanup
=============================

Target: ~/.claude/

Current total size: 7.3 GB

Scanning directories...

Directory                   Total    Cleanable    Files
----------                -------   ----------    -----
projects/                  7.0 GB       6.5 GB      980
debug/                   151.0 MB     140.0 MB      650
...

TOTAL                      7.3 GB       6.7 GB     5577

Space that can be reclaimed: 6.7 GB

# Prompt appears:
[Age Threshold]
Select cleanup age threshold:
  1. 8 hours (default) - Remove files older than 8 hours
  2. 2 days - Remove files older than 2 days
  3. Clean slate - Remove everything except safety margin

# User selects option 1

# Cleanup executes:
Cleanup Complete
================
Deleted: 5577 files
Failed:  0 files
Space reclaimed: 6.7 GB

New total size: 600.0 MB

Dry-Run Flow

# User runs: /refresh --dry-run

# Process cleanup half, forwarding --dry-run through to claude-refresh.sh:
Claude Code Refresh
===================

[DRY RUN] Preview only -- no processes will be terminated.

Found 2 orphaned processes using 1.2 MB:
...
Total memory that can be reclaimed: 1.2 MB

# Directory cleanup half -- shows survey, then:
=== DRY RUN MODE ===
Showing 8-hour cleanup preview...

Would delete: 5577 files
Would reclaim: 6.7 GB

Dry Run Summary
===============
No changes made.

Force Flow

# User runs: /refresh --force

# Shows survey, then immediately:
=== EXECUTING CLEANUP (8-hour default) ===

Cleanup Complete
================
Deleted: 5577 files
Space reclaimed: 6.7 GB

Safety Measures

Protected Files (Never Deleted)

• sessions-index.json (in each project directory)
• settings.json
• .credentials.json
• history.jsonl

Safety Margin

Files modified within the last hour are never deleted, regardless of age threshold.

Process Safety

claude-refresh.sh identifies orphans from a single atomic process snapshot, applying four independently-testable predicates rather than an argv-substring match or an ancestor-only process-tree walk (an ancestor walk cannot reach a sibling, such as the invoking session's own sleep inhibitor, and re-querying a transient PID from an earlier snapshot is a race -- see the script's own header comment for the full rationale):

• Executable-identity match: a candidate must match a narrow allow-list on its executable name (comm), never on a substring anywhere in its argv. This is what keeps a system daemon that merely mentions "claude" in one of its own arguments from ever being considered a candidate at all.
• System-slice cgroup exclusion: a candidate under /system.slice/ is never selected, regardless of anything else.
• Invoking-UID ownership: a candidate not owned by the invoking user's UID is excluded.
• Inhibitor-target liveness: a candidate holding a systemd-inhibit ... tail --pid=<N> sleep-inhibitor is excluded when the process it protects (<N>, a DIFFERENT process than the candidate) is still alive -- this is what correctly protects a live session's own inhibitor and any other live session's inhibitor on the machine, not just the invoker's own.
• Zero-query self-exclusion: a candidate whose pid or ppid equals the script's own pid is skipped, known at parse time with no second query and no race window.

TTY == "?" (no controlling terminal) remains a necessary-but-not-sufficient signal -- it is true of every systemd-managed process by construction, so it is combined with the predicates above, never used alone as the sole discriminator.

This design deliberately trades recall for safety: a leaked process this allow-list fails to recognize survives (false negative), which is strictly preferable to ever terminating a live system daemon or another live session's process (false positive).

Error Handling

Scripts Not Found

If scripts don't exist:

Error: Cleanup scripts not found at .claude/scripts/
Please ensure claude-refresh.sh and claude-cleanup.sh are installed.

Permission Denied

If kill/delete fails due to permissions:

Warning: Some operations failed due to insufficient permissions.
Failed files: 5
Successfully deleted: 5572 files

No ~/.claude/ Directory

If directory doesn't exist:

Error: ~/.claude/ directory not found.
Nothing to clean up.