scheduledtask

Codex-automations-style scheduled tasks for every agent harness and every model — open source and harness-agnostic.

scheduledtask runs recurring prompts headlessly on a schedule, captures the results in an inbox, isolates risky runs in git worktrees, and stays quiet when nothing changed. It works with OpenAI Codex CLI, Claude Code, opencode, or any command you can shell out to.

Packaged as an Agent Plugins v1.0.0 plugin with an agent-facing skill, so your coding agent can manage schedules for you.

Why

OpenAI ships automations only inside the ChatGPT desktop app; Anthropic ships its own inside Claude Code. Everyone else gets nothing. scheduledtask is the missing open layer:

• Any harness — codex / claude-code / opencode / arbitrary shell commands.
• Any model — pass -m-style model ids straight through (gpt-5.6-luna, opencode/x-preview-f-free, whatever your harness supports).
• Two run modes, matching Codex automations:
  • standalone — fresh session every run; results land separately in the inbox.
  • thread — resumes the previous session so each check builds on the last.
• Real schedules — cron expressions and RFC 5545 RRULEs (FREQ=MONTHLY;BYMONTHDAY=1;BYHOUR=9), evaluated in local wall-clock time.
• Worktree isolation — each run can execute in a detached-HEAD worktree so background changes never touch your dirty checkout.
• Dedup + triage inbox — identical findings collapse to silent no-change runs; unread markers, archive, pin.
• Missed-run policies — laptop asleep? Choose latest, all, or skip.
• Notifications — macOS Notification Center out of the box.

Install

Requires Bun — bins are Bun-native TypeScript.

git clone https://github.com/Dhruv2mars/scheduledtask ~/.scheduledtask-cli
echo 'alias scheduledtask="bun ~/.scheduledtask-cli/src/cli.ts"' >> ~/.zshrc
scheduledtask doctor   # verify harnesses are visible

One of codex, claude, or opencode must be on $PATH (or use the shell harness).

Quick start

cd my-project
scheduledtask init                 # creates .scheduledtask/tasks/example.yaml
$EDITOR .scheduledtask/tasks/example.yaml   # set prompt + enabled: true
scheduledtask run example          # test it manually FIRST
scheduledtask inbox                # read the report
scheduledtask daemon               # start scheduling

Create from flags

scheduledtask add --name nightly-brief \
  --cron "0 9 * * *" \
  --harness opencode \
  --model opencode/x-preview-f-free \
  --prompt "Summarize yesterday's commits grouped by workstream." \
  --worktree

Task reference

Field	Values	Default
schedule.cron	cron string	—
schedule.rrule	RFC 5545 RRULE string	—
harness	claude-code | codex | opencode | shell	codex
model	harness-native model id	harness default
effort	low | medium | high	harness default
sandbox	read-only | workspace-write | full	workspace-write
mode	standalone | thread	standalone
worktree	boolean	false
cleanup_worktree	remove worktree even with changes	false
projects	list of dirs — one run per project per slot	task's repo
dedup	collapse identical findings	true
catch_up	latest | all | skip after downtime	latest
notify	[macos]	[macos]
timeout_minutes	kill switch per run	30

Sandbox mapping is honest per-harness, not lowest-common-denominator:

sandbox	codex	claude-code	opencode
read-only	-s read-only	default	default
workspace-write	-s workspace-write	--permission-mode acceptEdits	--auto
full	bypass approvals+sandbox	--dangerously-skip-permissions	--auto (no stronger escalation exists)

Thread mode uses native resume where available (codex exec resume <id>, claude --resume <id>, opencode -s <id>).

Trust model

Tasks found in your repo (.scheduledtask/tasks/) or inside plugins (dev.scheduledtask/tasks/) are checked-in code that will execute unattended. scheduledtask therefore refuses to run them until you mark their tree safe:

scheduledtask trust      # allow repo/plugin tasks under cwd
scheduledtask untrust    # revoke

User-authored tasks (~/.scheduledtask/tasks/) always run. sandbox: full is rejected at load time for anything not user-authored.

The shell harness never interpolates prompt text into the command line (prompts embed model output and repo content). Use {{prompt_file}}, which expands to a path holding the prompt:

harness: shell
command: 'run-analysis.sh "$(cat {{prompt_file}})"'

Inbox lifecycle

Runs land unread; show marks them read; read-all bulk-clears. Archiving an unpinned run releases its worktree; pinned runs keep theirs; scheduledtask gc reclaims archived-unpinned worktrees and prunes history (200 runs/task by default, override with gc --keep <n>).

Agent Plugins integration

This repo is itself a valid Agent Plugin (plugin.json). Tasks can also be distributed inside other plugins:

my-plugin/
├── plugin.json
└── dev.scheduledtask/
    └── tasks/
        └── nightly-bugfix.yaml

Discovery scans <repo>/.agents/plugins/*/dev.scheduledtask/tasks/ and ~/.agents/plugins/*/dev.scheduledtask/tasks/ automatically. The bundled skill (skills/manage-scheduledtask) teaches any Agent Skills-compatible harness how to create and operate tasks through the CLI.

Running as a service

# macOS launchd (keep the daemon alive across logins)
brew install services  # not required; use a plist pointing at:
bun /path/to/scheduledtask/src/cli.ts daemon

The daemon sweeps every 30 s and handles missed slots on wake via each task's catch_up policy. For servers, wire scheduledtask sweep into systemd timers or cron as a watchdog.

Development

bun install
bun test          # unit + security regression tests
bun run typecheck # strict TS, zero errors
bun src/cli.ts doctor

Status & roadmap

Working: standalone + thread modes, cron/RRULE, worktrees, dedup inbox, notifications, multi-project tasks, plugin-distributed tasks, trust gate for checked-in tasks, three first-class adapters plus shell escape hatch.

Next: TUI inbox browser, Slack/ntfy notification channels, cloud catch-up runner for sleeping laptops, tasks/ component-type proposal upstream to the Agent Plugins spec.

MIT licensed. Not affiliated with OpenAI, Anthropic, or SST.