/manage

name: manage
description: Manage scheduled Claude Code tasks — add (recurring or one-off), list, pause, resume, remove, view results, and test execution of skills, prompts, and scripts with safety controls and notifications. Use when the user mentions scheduling, cron, automated tasks, recurring tasks, background tasks, running something on a schedule, periodic execution, or wants a skill/prompt/script to run automatically at a set time. Cross-platform (macOS, Linux, Windows).
user-invocable: true

Scheduler

Manage automated Claude Code tasks across macOS, Linux, and Windows. Schedule marketplace skills, freeform prompts, or shell scripts to run on a recurring cron schedule or as one-off tasks, with safety controls and desktop notifications.

Platform Support

| Platform | Scheduler | Notifications | API Key Storage | |----------|-----------|---------------|-----------------| | macOS | launchd (plist + launchctl) | osascript (native) | Keychain (`security` CLI) | | Linux | systemd user timers (.service + .timer) | notify-send (libnotify) | secret-tool (GNOME Keyring) or `~/.config/anthropic/api-key` | | Windows | Task Scheduler (schtasks.exe + XML) | BurntToast / Windows toast | Windows Credential Manager or `%USERPROFILE%\.config\anthropic\api-key` |

The platform is auto-detected at runtime. Tasks record which platform they were created on.

Overview

This orchestrator manages the full lifecycle of scheduled tasks. It delegates all operations to `scheduler.py` and presents results conversationally.

**Core Principle**: This is a thin orchestrator. All scheduling logic, file generation, and platform scheduler interaction are handled by `scheduler.py` with platform-specific backends. This skill manages the conversational flow only.

When to Use

Use this skill when:

• Scheduling a recurring or one-off task (skill, prompt, or script)
• Listing, pausing, resuming, or removing scheduled tasks
• Viewing results or logs from past scheduled runs
• Testing a scheduled task before activating it

Prerequisites

• macOS, Linux, or Windows
• `claude` CLI in PATH (for skill and prompt tasks)
• `uv` installed (for running scheduler.py)
• **Linux only**: `systemctl --user` available (systemd with user session)
• **Windows only**: PowerShell 5.1+ and Task Scheduler access

First-Time Setup

On first invocation, verify `claude` is authenticated and in PATH:

claude -p "hello" --max-turns 1 --output-format text

If this fails, guide the user to run `claude login` before proceeding.

Also check if the scheduler directory exists. The scheduler uses project-level storage by default — it looks for `.git` or `CLAUDE.md` to find the project root, then stores state at `<project>/.claude/scheduler/`:

ls .claude/scheduler/registry.json

If it does not exist, the first `add` command will create it automatically. No manual setup needed. You can override the location with the `SCHEDULER_DIR` environment variable, or fall back to `~/.claude/scheduler/` when outside a project.

Workflow

When the user invokes `/manage`, present the operation menu:

1. **Add** a new scheduled task 2. **List** all scheduled tasks 3. **Pause** a task 4. **Resume** a paused task 5. **Remove** a task 6. **Results** — view output from a task 7. **Logs** — view execution logs 8. **Run now** — test a task immediately 9. **Cleanup** — delete old logs and results 10. **Repair** — regenerate wrappers and schedule artifacts

Then execute the selected operation below.

Operation: Add

Walk the user through each field conversationally:

**Step 1: Task name** Ask: "What should this task be called?" Generate a slug ID from the name (e.g., "Weekly note ideas" → `weekly-note-ideas`).

**Step 2: Task type** Ask: "What type of task?"

• **Skill** — invoke a marketplace skill (e.g., `creator-stack:generate-note-ideas`)
• **Prompt** — run a freeform Claude prompt
• **Script** — execute a shell script

**Step 3: Target** Based on type:

• Skill: Ask which skill. Suggest relevant installed skills if possible.
• Prompt: Ask for the prompt text.
• Script: Ask for the absolute path to the script.

**Step 4: Frequency** Ask: "Recurring or one-off?"

• **Recurring** — runs on a schedule until paused or removed
• **One-off** — runs once at the specified time, then auto-completes

**Step 5: Schedule** Ask: "What schedule? You can use natural language or a cron expression." Examples: "Every Monday at 8am", "Daily at 9:00", "Every weekday at 7:30am", "Tomorrow at 9am", "0 8 * * 1"

Convert natural language to a cron expression. For one-off tasks, convert the target datetime to a specific cron (e.g., "tomorrow at 9am" on Feb 25 → `0 9 26 2 *`).

Confirm with the user: "That's `0 8 * * 1` — Every Monday at 8:00 AM. Correct?"

**Step 6: Safety limits** Ask: "Safety limits? Defaults are max 20 turns, 15 minute timeout." Let user accept defaults or customize.

**Step 6a: Output directory (optional)** Ask: "Want results saved to a specific directory? Default is the scheduler's results folder." If the user provides a path, it will be resolved to an absolute path and used as the flat output directory (no date subdirectories). If skipped, results go to `<scheduler_dir>/results/YYYY-MM-DD/{id}.md`.

**Step 6b: Permissions (skip for script-type tasks)** Scheduled tasks run non-interactively — Claude Code can't ask the user to approve tool use. Without pre-approved permissions, tools that need approval will fail silently.

Ask: "What permissions should this task have?" Offer these presets:

• **Default** — no extra permissions (works if you have `bypassPermissions` set globally)
• **Readonly** — read files, search, browse web (safe for research/reporting tasks)
• **Research** — read files, search, web browse, curl (for data gathering)
• **Full edit** — read, write, edit files, git, npm, uv, python, web (for code maintenance tasks)
• **Bypass** — skip all permission checks (`--dangerously-skip-permissions`)
• **Custom** — spec