Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

447 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

OpenClaw Scheduler

CI License Node

A durable continuity and governed-workflow sidecar for OpenClaw agents and shell workflows. Current OpenClaw already provides SQLite cron state, command jobs, retries, and run history, plus background tasks, Task Flow, and Lobster approval checkpoints. Use this scheduler when shell jobs must survive Gateway downtime, the scheduler needs a separate failure domain, workflows need direct conditional job graphs, or @amittell/agentcli needs its durable runtime target.

This is an independent community project and is not affiliated with the OpenClaw project.

It can coexist with native OpenClaw cron. Do not run the same job in both systems.

Release repo: github.com/amittell/openclaw-scheduler Development mirror: writhub.io/alexm/openclaw-scheduler Default location: ~/.openclaw/scheduler/ Service: ai.openclaw.scheduler (macOS launchd: LaunchAgent or LaunchDaemon) Runtime: Node.js 22, 24, 25, or 26 (ESM), SQLite via better-sqlite3, cron parsing via croner Schema: version 28 Tests: run with npm test (legacy suite, isolated focused tests, docs, and agentcli integration when a checkout is present; hosted CI requires a pinned public checkout) Platform: macOS · Linux · Windows (WSL2)

In practice, this gives you:

  • shell jobs that still work when the Gateway is unhealthy
  • a scheduler failure domain separate from the Gateway
  • AI jobs that stay isolated from your personal chats
  • conditional chains, fenced cancellation, atomic approvals, and transactional external delivery

Table of Contents

  1. Why This Exists
  2. Concrete Use Cases
  3. When To Use It
  4. Choosing a Runtime
  5. Quick Start
  6. Five-Minute Setup
  7. Starter Recipes
  8. Common Migrations
  9. Platform Support
  10. Architecture
  11. How Jobs Execute
  12. Delivery Modes
  13. Delivery Aliases
  14. Shell Jobs
  15. HITL Approval Gates
  16. Idempotency
  17. Context Retrieval
  18. Task Tracker
  19. Resource Pools
  20. Workflow Chains
  21. Retry Logic
  22. Chain Safety
  23. Inter-Agent Messaging
  24. Backup & Recovery
  25. Agent Registry
  26. Database Schema
  27. CLI Reference
  28. Configuration
  29. Service Management
  30. Error Handling & Backoff
  31. Migration & History
  32. Upgrading
  33. Removing the Scheduler
  34. Best Practices
  35. File Reference
  36. Testing
  37. Sub-agent Dispatch
  38. Working with agentcli
  39. Trust Architecture
  40. Troubleshooting
  41. Companion Scripts

Why This Exists

OpenClaw's built-in cron and Task Flow are the right default for many automations. This project addresses a narrower operator need: continuity outside the Gateway process and one local runtime for conditional shell and agent graphs.

The pain usually looks like this:

  • A shell script is operationally important, but it should keep running even if the gateway is unhealthy.
  • One step should trigger another, but only on success, only on failure, or only if the output contains a specific signal.
  • A risky action needs a human in the loop instead of firing immediately.
  • An agent needs to hand work to another agent or process, and you want that handoff tracked and auditable.

openclaw-scheduler exists to solve those problems without replacing the native runtime for ordinary cron work. It provides jobs, runs, chains, shell execution, approvals, a delivery outbox, and message routing backed by a separate SQLite database.

Concrete Use Cases

These are the kinds of workflows this scheduler is meant for:

  • metrics capture -> analysis -> approval -> report publish You want each step tracked, retried if needed, and gated before the final action.
  • shell ingest fails -> agent diagnoses failure -> operator approves remediation The ingest should still run without the gateway, but the failure follow-up can use an agent.
  • workspace audit -> diagnosis -> memory compression The audit is a shell step, the diagnosis is an agent step, and the remediation should only run if the diagnosis actually recommends it.
  • bot health check -> alert -> repair action A shell check runs on schedule, an agent summarizes the issue, and a repair step waits for approval.

The differentiator is not just "better cron". It is mixed shell + agent workflows with durable state and control over what happens after success, failure, timeout, or explicit signals in output.

When To Use It

Use it when you want:

  • shell jobs that do not depend on OpenClaw gateway availability
  • an independently operated scheduler and database
  • parent/child workflow chains
  • approval gates before risky steps
  • auditable inter-agent or agent-to-shell handoffs

Do not use it merely to obtain run history, command jobs, or basic retries. Native OpenClaw provides those capabilities. If all you need is one scheduled command or agent turn, this is probably more system than you need.

Choosing a Runtime

If you have never used OpenClaw's built-in cron, skip migration and go directly to Five-Minute Setup.

Requirement Recommended owner
One command or agent turn on a schedule Native OpenClaw cron
Native restart-safe flow or resumable approval checkpoint OpenClaw Task Flow or Lobster
Shell execution while the Gateway is unavailable OpenClaw Scheduler
Direct parent/child triggers on result or output content OpenClaw Scheduler
Declarative governed runbook manifest @amittell/agentcli targeting OpenClaw Scheduler

Version 0.5.0 Runtime Safety

  • A singleton dispatcher lease and fencing token prevent a second dispatcher from taking ownership of live work.
  • Active runs carry dispatcher ownership. Only the owning fence can commit the terminal transition or trigger downstream work.
  • Cancellation is a durable request. Shell jobs terminate their tracked process group; agent jobs record and send the Gateway cancellation request. Check runs get or status for authoritative state.
  • Expired dispatch claims return to the pending queue only when no active run owns them.
  • Run completion, job state, and child dispatch creation commit atomically.
  • External delivery uses a transactional outbox that is separate from agent prompt messages. Completion ownership is scoped by run, multipart deliveries have independently retryable per-part checkpoints, and completion debt closes only after every part is actually delivered. Attachments retain size and SHA-256 metadata.
  • Approval decisions use atomic versioned transitions and cannot dispatch disabled, rejected, expired, or cancelled work.
  • Governance declarations are enforced at execution time. Unsupported sandbox, path, network, credential, trust, proof, or cost requirements fail closed.
  • Database schema and consolidation failures stop startup. openclaw-scheduler doctor --json reports schema, lease, queue, outbox, approval, and cancellation diagnostics.
  • AgentCLI handoff v4 binds the complete canonical execution artifact to every dispatch, approval, run, provider session, credential presentation, runtime event, and signed evidence record. Versions 1 through 3 remain supported.

These controls do not make arbitrary side effects idempotent. Destructive commands must still detect prior completion and be safe to retry.


Quick Start

New to the scheduler? Start with QUICK-START.md -- a focused guide covering installation, converting existing OpenClaw crons, and building your first workflow chain.

For the full reference, use the npm-first path below and then jump straight to Five-Minute Setup.

Option A: npm-first (publish/install flow)

mkdir -p ~/.openclaw/scheduler
npm install --ignore-scripts=false --prefix ~/.openclaw/scheduler openclaw-scheduler@latest
npm exec --prefix ~/.openclaw/scheduler openclaw-scheduler -- setup

This installs the package without cloning the repo. The launcher command maps to:

  • openclaw-scheduler setupsetup.mjs
  • openclaw-scheduler startdispatcher.js
  • openclaw-scheduler webhook-checkscripts/telegram-webhook-check.mjs
  • openclaw-scheduler <anything-else>cli.js

The explicit --ignore-scripts=false is required because better-sqlite3 builds or installs its trusted native binding during the npm lifecycle. If a prior install used ignore-scripts=true, rerun the install command above or run npm rebuild --ignore-scripts=false better-sqlite3 before starting.

For npm installs, scheduler state defaults to ~/.openclaw/scheduler/ rather than node_modules/openclaw-scheduler/, so upgrades do not trample the database path.

If your Node runtime changes later, rebuild the native SQLite binding before restarting the scheduler:

cd ~/.openclaw/scheduler
npm rebuild better-sqlite3 --ignore-scripts=false

This is commonly needed after a Homebrew Node upgrade on macOS or any major Node ABI change.

macOS shell setup for ad hoc commands

If you use zsh on macOS, put your minimal Homebrew PATH bootstrap in ~/.zshenv, not only in ~/.zprofile or ~/.zshrc.

Why:

  • launchd services do not depend on your interactive shell startup files
  • ad hoc commands like ssh host 'node cli.js status' run a non-interactive shell
  • non-interactive zsh reads ~/.zshenv, but does not read ~/.zprofile

Recommended ~/.zshenv:

# ~/.zshenv — sourced by all zsh instances, including non-interactive SSH commands
if [ -x /opt/homebrew/bin/brew ]; then
  eval "$(/opt/homebrew/bin/brew shellenv)"
fi

export PATH="$HOME/.local/bin:$HOME/bin:/opt/homebrew/bin:/opt/homebrew/sbin:/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin"

If you load OpenClaw completions in ~/.zshrc, initialize completions first so compdef is available:

autoload -Uz compinit
compinit

if [ -f "$HOME/.openclaw/completions/openclaw.zsh" ]; then
  source "$HOME/.openclaw/completions/openclaw.zsh"
fi

Avoid pinning a versioned Node path like /opt/homebrew/opt/node@22/bin in shell startup files. Prefer the stable Homebrew symlink /opt/homebrew/bin/node, which survives normal brew upgrade node.

Option B: source clone (dev/contributor flow)

git clone https://github.com/amittell/openclaw-scheduler ~/.openclaw/scheduler
cd ~/.openclaw/scheduler
npm install
npm test                             # should end with: 0 failed
npm run lint                         # static checks
npm run typecheck                    # exported API declarations
npm run coverage                     # coverage summary + lcov report
npm run verify:local                 # full local maintainer gate
npm run verify:smoke                 # lightweight smoke gate used by GitHub Actions

GitHub Actions runs npm run verify:smoke on Linux and macOS across the supported Node.js lines. That gate includes lint, type checking, every repository test, documentation validation, and a package dry run. npm run verify:local adds coverage and is enforced again by prepublishOnly. Focused test files run sequentially with isolated databases. A dedicated required job separately checks the exact public handoff-v3 producer and the previous handoff-v2 producer.

Option C: local npm pack (simulate the published package from source)

Use this when you want to test the exact package layout end users get from npm without publishing to npmjs.org yet.

git clone https://github.com/amittell/openclaw-scheduler /tmp/openclaw-scheduler
cd /tmp/openclaw-scheduler
npm ci
npm run verify:local
npm pack

mkdir -p ~/.openclaw/packages/openclaw-scheduler
npm install --ignore-scripts=false --prefix ~/.openclaw/packages/openclaw-scheduler --omit=dev --no-package-lock ./openclaw-scheduler-*.tgz

Point your service at ~/.openclaw/packages/openclaw-scheduler/node_modules/openclaw-scheduler/dispatcher.js, and keep mutable state in ~/.openclaw/scheduler via SCHEDULER_HOME and SCHEDULER_DB.

The package also exports a small safe programmatic API surface for tooling:

import { db, jobs, runs, shellResults } from 'openclaw-scheduler';

Then run the interactive setup wizard:

npm exec openclaw-scheduler -- setup
# or: node setup.mjs

The wizard will:

  • Run DB migrations
  • Append scheduler queue/inbox-consumer entries to your agent's MEMORY.md and workspace-index.md
  • Create Inbox Consumer + Stuck Run Detector scheduler jobs
  • Configure dispatcher auto-start service:
    • macOS: LaunchAgent (personal auto-login Mac) or LaunchDaemon (headless/pre-login startup)
    • Linux/WSL2: systemd user service (or PM2 fallback)

After setup:

npm exec openclaw-scheduler -- status # verify scheduler is running
node scripts/stuck-run-detector.mjs   # should print: No stale runs older than 15 minute(s).
tail -5 /tmp/openclaw-scheduler.log   # live logs

Dispatcher setup is covered in:


Five-Minute Setup

This is the shortest path from "I installed it" to "I have a real job running."

1. Install and initialize

mkdir -p ~/.openclaw/scheduler
npm install --ignore-scripts=false --prefix ~/.openclaw/scheduler openclaw-scheduler@latest
alias ocs='npm exec --prefix ~/.openclaw/scheduler openclaw-scheduler --'
ocs setup
ocs status

What this does:

  • installs the package into ~/.openclaw/scheduler
  • creates or migrates scheduler.db
  • installs the scheduler service using the launchd mode you choose (agent or daemon)
  • creates the built-in helper jobs like Inbox Consumer and Stuck Run Detector

If you plan to use the scheduler often, add the ocs alias to your shell profile.

2. Add your first real job

This example runs a simple shell health check every 15 minutes.

ocs jobs add '{
  "name": "Disk Space Check",
  "schedule_cron": "*/15 * * * *",
  "session_target": "shell",
  "payload_message": "df -h /",
  "delivery_mode": "none",
  "run_timeout_ms": 120000,
  "origin": "system"
}'

What it means in plain English:

  • schedule_cron: run every 15 minutes
  • session_target: "shell": run a shell command directly, no AI needed
  • payload_message: the command to run
  • delivery_mode: "none": do not send the output anywhere automatically
  • origin: "system": this job was created by the system, not from a user chat

3. Run it now and inspect the result

ocs jobs list
# copy the job ID for "Disk Space Check"

ocs jobs run <job-id>
ocs runs list <job-id> 5

If you want the full run record:

ocs runs get <run-id>

At that point you have a working scheduler install, a real job, and visible run history. Everything after this is layering on more power: AI jobs, delivery, retries, workflow chains, and approvals.


Starter Recipes

These are copy-paste examples for the most common first workflows.

1. Shell health check with failure alerts

Use this when you want a script to run reliably even if the OpenClaw gateway is down.

ocs jobs add '{
  "name": "API Health Check",
  "schedule_cron": "*/15 * * * *",
  "session_target": "shell",
  "payload_message": "curl -fsS http://127.0.0.1:8080/health || exit 1",
  "delivery_mode": "announce",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 120000,
  "origin": "system"
}'

Why this is useful:

  • it runs even if the gateway is unhealthy
  • it only announces on failure
  • every run is stored in history

2. Daily AI summary

Use this when you want a scheduled agent report instead of a shell script.

ocs jobs add '{
  "name": "Daily Ops Summary",
  "schedule_cron": "0 9 * * *",
  "schedule_tz": "America/New_York",
  "session_target": "isolated",
  "payload_message": "Summarize the last 24 hours of important errors, deploys, and follow-ups in 5 bullet points.",
  "delivery_mode": "announce-always",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 300000,
  "origin": "system"
}'

Why this is useful:

  • the agent runs in its own isolated session
  • delivery is retried durably and exhausted failures remain visible to operators
  • the run history stays separate from your personal chat threads

3. Approval-gated follow-up step

Use this when a risky step should wait for a human before it runs.

ocs jobs add '{
  "name": "Delete Old Backups",
  "parent_id": "<parent-job-id>",
  "trigger_on": "success",
  "approval_required": 1,
  "approval_timeout_s": 3600,
  "approval_auto": "reject",
  "session_target": "shell",
  "payload_message": "find /backups -type f -mtime +14 -delete",
  "delivery_mode": "announce-always",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 120000
}'

Why this is useful:

  • the parent job can run automatically
  • the risky cleanup step pauses until someone approves it
  • the scheduler records who approved or rejected it

Approve or reject later with:

ocs approvals list --json
ocs approvals approve <approval-id> --reason "Change window open"
ocs approvals reject <approval-id> --reason "Not today"

Common Migrations

If you already have cron jobs, OpenClaw cron entries, or shell scripts, this is the simplest way to think about the conversion.

1. OpenClaw built-in cron -> inspect, import, test, then disable duplicates

The default importer reads the supported OpenClaw CLI rather than an internal storage file. It calls openclaw cron list --json and openclaw cron get <id> --json:

openclaw-scheduler migrate --dry-run --json > migration-report.json
openclaw-scheduler migrate --json
openclaw-scheduler jobs list --json

The dry run never opens, creates, or migrates the target scheduler database. Existing-ID duplicate checks therefore occur during the real import.

Cron expressions and one-shot timestamps are preserved exactly. Intervals that five-field cron cannot represent exactly are rejected unless the operator explicitly passes --allow-inexact-every. A pre-SQLite export is opt-in:

openclaw-scheduler migrate --legacy-json ~/.openclaw/cron/jobs.json --dry-run --json

After representative imported jobs complete successfully, disable only the corresponding native jobs:

openclaw cron edit <job-id> --disable

Rollback by stopping the scheduler, disabling its imported copies, and re-enabling the native jobs. Retain the scheduler database until rollback verification is complete.

2. Plain shell cron line -> session_target: "shell"

If you have a normal cron line like:

*/5 * * * * /usr/local/bin/check-api.sh

Convert it to:

ocs jobs add '{
  "name": "API Check",
  "schedule_cron": "*/5 * * * *",
  "session_target": "shell",
  "payload_message": "/usr/local/bin/check-api.sh",
  "delivery_mode": "announce",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 120000,
  "origin": "system"
}'

Choose shell when the task is deterministic and you do not need AI reasoning.

3. AI-ish cron job -> session_target: "isolated"

If the old job was really “run a prompt every morning”, use an isolated agent job instead of a shell script:

ocs jobs add '{
  "name": "Daily Status Summary",
  "schedule_cron": "0 8 * * *",
  "schedule_tz": "America/New_York",
  "session_target": "isolated",
  "payload_message": "Summarize the most important errors, deploys, and follow-ups from the last 24 hours in 5 bullet points.",
  "delivery_mode": "announce-always",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 300000,
  "origin": "system"
}'

Choose isolated when the job needs reasoning, writing, summarization, or tools.

4. Two cron jobs with manual ordering -> parent/child chain

If your current workflow is:

  • run a backup
  • wait
  • run verification

Model that as a chain instead of two unrelated cron entries:

ocs jobs add '{
  "name": "Nightly Backup",
  "schedule_cron": "0 2 * * *",
  "session_target": "shell",
  "payload_message": "/usr/local/bin/nightly-backup.sh",
  "delivery_mode": "announce",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 600000,
  "origin": "system"
}'

Then create the follow-up:

ocs jobs add '{
  "name": "Verify Nightly Backup",
  "parent_id": "<backup-job-id>",
  "trigger_on": "success",
  "trigger_delay_s": 60,
  "session_target": "shell",
  "payload_message": "/usr/local/bin/verify-backup.sh",
  "delivery_mode": "announce",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 120000
}'

This is one of the biggest upgrades over plain cron: the second step now runs because the first step succeeded, not because the clock happened to reach another minute.

5. Risky follow-up -> add approval_required

If the current process is “job runs, then a human decides whether to continue,” model that decision directly:

ocs jobs add '{
  "name": "Delete Temp Files",
  "parent_id": "<analysis-job-id>",
  "trigger_on": "success",
  "approval_required": 1,
  "approval_timeout_s": 3600,
  "approval_auto": "reject",
  "session_target": "shell",
  "payload_message": "find /tmp/myapp -type f -mtime +7 -delete",
  "delivery_mode": "announce-always",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 120000
}'

That keeps the job automated, but only up to the point where human judgment is actually needed.

Rule of thumb

When converting existing work:

  • start with shell unless you clearly need AI reasoning
  • add delivery only if someone really needs to see the output
  • use chains when one step depends on another
  • use approvals when the next step would be annoying, expensive, or risky if it ran by mistake

Platform Support

Platform Service Manager Shell Jobs Status
macOS launchd (agent or daemon) /bin/zsh Tested
Linux systemd user service or PM2 fallback /bin/bash Supported
Windows (WSL2) systemd user service or PM2 fallback inside WSL2 /bin/bash Supported
Windows (native or WSL1) Not applicable Not applicable Not supported; use WSL2

Override the shell for shell jobs with the SCHEDULER_SHELL=/path/to/shell environment variable.


Architecture

The scheduler sits alongside the OpenClaw gateway as an independent process. Isolated targets use stable per-job sessions that remain separate from the user's main conversation. Main targets use the existing main session, and shell targets do not create an agent session.

┌─────────────────────────────────────────────┐
│  Host Machine (e.g., scheduler-host.local)  │
│                                              │
│  OpenClaw Gateway (:18789)                   │
│    ├─ Telegram / Discord / etc.              │
│    ├─ Chat completions endpoint (/v1/...)    │
│    ├─ Tool execution (exec, browser, k8s...) │
│    └─ Memory search                          │
│                                              │
│  Scheduler (launchd service)                 │
│    ├─ SQLite DB (scheduler.db)               │
│    ├─ Job dispatch via chat completions      │
│    ├─ Workflow chain engine                  │
│    ├─ Retry logic                            │
│    ├─ Shell job execution                    │
│    ├─ HITL approval gates                    │
│    ├─ Idempotency ledger                     │
│    ├─ Inter-agent message queue              │
│    ├─ Task tracker                           │
│    └─ MinIO backup                           │
└─────────────────────────────────────────────┘

Tick Loop

┌──────────────────────────────────────────────────────────────┐
│                   Dispatcher Loop (10s tick)                  │
│                                                              │
│  1. Gateway health check                                     │
│  2. Find due jobs → dispatch                                 │
│  3. Check running runs (stale/timeout detection)             │
│  4. HITL approval gate check                                 │
│  5. Message delivery + spawn handling                        │
│  6. Task tracker dead-man's-switch                           │
│  7. Expire old messages                                      │
│  8. Prune old runs + WAL checkpoint (hourly)                 │
│  9. Backup to MinIO (every 5 min)                            │
└──────────────────────────────────────────────────────────────┘
         │                              │
         ▼                              ▼
  ┌───────────────────┐        ┌──────────────────────┐
  │     SQLite DB      │        │   OpenClaw Gateway    │
  │                    │        │                       │
  │  • jobs            │        │  • /v1/chat/completions│
  │  • runs            │        │  • /tools/invoke      │
  │  • messages        │        │  • /health            │
  │  • agents          │        │  • system event CLI   │
  │  • approvals       │        └──────────────────────┘
  │  • task_tracker    │
  │  • idempotency_ledger│
  │  • delivery_aliases│
  │  • schema_migrations│
  └───────────────────┘

Session Types

Session Created By Lifetime Used For
User DM Telegram message Persistent per-peer Your conversations
Group chat Group message Persistent per-group Team discussions
Isolated job Dispatcher via API Persistent per job, reused across runs Cron jobs, chain steps
Main session Dispatcher API, or openclaw system event for fire-and-forget Existing main session Jobs needing main context
Shell Dispatcher (direct) Per-job (no session) Cron scripts, backups, maintenance
Sub-agent sessions_spawn Task-scoped Delegated work

Isolated scheduler jobs cannot see user-chat history, but later runs of the same job reuse its warm per-job session and may retain that job's earlier context. Main-session jobs intentionally use the existing main conversation.


How Jobs Execute

Isolated Jobs (default)

Scheduler tick (every 10s)
  │
  ├─ getDueJobs() → "Hourly Workspace Backup is due"
  ├─ hasRunningRun()? → skip if overlap_policy='skip'
  ├─ createRun() → status='running'
  ├─ setAgentStatus('main', 'busy')
  │
  ├─ POST /v1/chat/completions
  │   session: agent:<agent_id>:scheduler:<job_id>  (stable per job, isolated from main)
  │   model: openclaw:main
  │   message: [job prompt + any pending inbox messages]
  │
  │   ← "Committed 3 files, pushed to origin"
  │
  ├─ finishRun('ok', summary)
  ├─ setAgentStatus('main', 'idle')
  ├─ Deliver to Telegram? → delivery_mode + channel + target
  ├─ Queue result message for traceability
  ├─ Advance next_run_at to next cron fire
  └─ Trigger child jobs if any (workflow chain)

Main Session Jobs

For jobs that need the main session context (rare), use payload_kind: "systemEvent". Default, execute, or plan execution waits for the agent response and captures it for normal completion handling:

Dispatcher → POST /v1/chat/completions → wait for main-session response

Set execution_intent: "fire-and-forget" only when the scheduler should inject the event and return immediately without capturing the eventual response:

Dispatcher → exec: openclaw system event --text "..." --mode now

Handoff v4 main-session jobs use only the synchronous path because it carries the artifact, runtime-instance, and fresh capability-nonce binding through the Gateway request. The openclaw system event transport cannot carry that binding, so v4 fire-and-forget jobs fail validation. Earlier handoff versions keep the existing fire-and-forget behavior.

Shell Jobs

Shell Job (session_target='shell')
  │
  ├─ getDueJobs() → "Hourly Backup is due"
  ├─ createRun() → status='running'
  ├─ run "<payload_message>" via shell (platform default or SCHEDULER_SHELL)
  │   (no gateway required)
  │   ← exit 0: "Backup complete, 3 files"
  │
  ├─ finishRun(exit===0 ? 'ok' : 'error')
  ├─ announce: post output if exit ≠ 0
  ├─ announce-always: post output regardless
  └─ Trigger child jobs if any

Prompt Building

Each isolated job prompt includes:

  1. Header: [scheduler:<job_id> <job_name>]
  2. Pending inbox messages for the agent (up to 5)
  3. Context from prior runs (if context_retrieval is set)
  4. The job's payload_message

Delivery Modes

The scheduler delivers job output through the OpenClaw gateway's messaging system. All channels supported by the gateway work with the scheduler: Telegram, Discord, WhatsApp, Signal, iMessage, and Slack. Set delivery_channel to the channel name and delivery_to to the channel-specific target (chat ID, channel ID, phone number, handle, etc.).

Mode When output is delivered
none Never (background jobs)
announce Agent jobs: delivers when run status is not ok. Shell jobs: non-zero exit only. Silently skipped for main session jobs (use announce-always instead)
announce-always Always delivers output (LLM or shell), including main session jobs

Note: non-exempt jobs using announce or announce-always are rejected at validation time when delivery_to is absent. Runtime delivery is written to a transactional outbox, separate from agent prompt messages. Outbox claims expire and recover safely. Multipart output is stored as independently retryable rows with deterministic :part:i/N idempotency keys. Durable enqueue is not a delivery receipt; run-scoped completion debt closes only after every part is delivered. Attachment rows retain size and SHA-256 metadata.

Examples in this document use Telegram for delivery_channel since it is the most common configuration. Replace with your channel of choice.


Delivery Aliases

Delivery aliases let you define named delivery targets (e.g., @my_team) instead of hard-coding channel/target pairs in every job.

# Create a named alias
openclaw-scheduler alias add my_team telegram -100200000000

# Use @alias in job (resolves at dispatch time)
openclaw-scheduler jobs add '{
  "name": "Alert",
  "delivery_mode": "announce",
  "delivery_to": "@my_team",
  ...
}'

# List aliases
openclaw-scheduler alias list

# Remove an alias
openclaw-scheduler alias remove my_team

Aliases are resolved at dispatch time. If an alias is deleted, jobs fall back to suppressed delivery.


Shell Jobs

Shell jobs run a command directly on the host — no gateway or LLM required. Ideal for backups, scripts, maintenance tasks, and anything that doesn't need AI.

openclaw-scheduler jobs add '{
  "name": "Hourly Backup",
  "schedule_cron": "0 * * * *",
  "schedule_tz": "America/New_York",
  "session_target": "shell",
  "payload_message": "/path/to/backup.sh",
  "delivery_mode": "announce",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 600000,
  "origin": "system"
}'

Key properties:

  • No gateway dependency — runs even when gateway is down
  • payload_message is the command to execute (shell string passed to the configured shell)
  • Output captured up to 1MB, with preview/offload budgets to keep large output out of the main run row
  • Shell runs persist structured failure context on runs: shell_exit_code, shell_signal, shell_timed_out, shell_stdout, shell_stderr, plus optional shell_stdout_path / shell_stderr_path when large output is offloaded
  • Failure-triggered agent children receive shell context with separate exit code, stdout, and stderr blocks
  • run_timeout_ms controls max execution time (required, no default)
  • Workflow chains work the same way — shell jobs can trigger children on success/failure
  • Shell jobs now honor max_retries before failure children fire, the same as isolated agent jobs
  • openclaw-scheduler runs output <run-id> stdout|stderr retrieves stored or offloaded shell output on demand

With environment variables:

openclaw-scheduler jobs add '{
  "name": "DB Dump",
  "schedule_cron": "0 3 * * *",
  "session_target": "shell",
  "payload_message": "PGPASSWORD=secret pg_dump mydb > /backups/mydb.sql && echo OK",
  "delivery_mode": "announce-always",
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 600000,
  "origin": "system"
}'

HITL Approval Gates

Jobs with approval_required: 1 pause every dispatch before execution. The gate applies uniformly to root, manual, scheduled, one-shot, retry, and chain-triggered work. Every gated attempt is represented by a durable dispatch row and an awaiting_approval run.

# Scheduled root job that requires a scoped operator approval
openclaw-scheduler jobs add '{
  "name": "Deploy to Prod",
  "schedule_cron": "0 10 * * 1-5",
  "approval_required": 1,
  "approval_risk_level": "high",
  "approval_approver_scope": "user:alex",
  "approval_timeout_s": 3600,
  "approval_auto": "reject",
  "session_target": "shell",
  "payload_kind": "shellCommand",
  "payload_message": "Deploy the application to production",
  "delivery_mode": "none",
  "origin": "system",
  "run_timeout_ms": 300000
}'

When triggered, the job creator receives the pending approval identifier, risk, scope, and CLI commands.

openclaw-scheduler approvals list --json
openclaw-scheduler approvals approve APPROVAL_ID --reason "Change window open"
openclaw-scheduler approvals reject APPROVAL_ID --reason "Postponing until the next change window"

Key notes:

  • approval_risk_level is optional and accepts low, medium, or high.
  • approval_approver_scope accepts an unprefixed exact identity or exact:, user:, uid:, or principal: matching for the local OS account. Domain scopes are not supported.
  • AgentCLI handoff v4 exposes approval-scope enforcement as one coarse capability while its manifests may contain domain scopes. The scheduler therefore advertises approval_scope_enforcement: false so scoped agentcli manifests fail capability negotiation instead of being partially enforced. Direct scheduler jobs may still use the supported local scopes above.
  • The scheduler derives the approver from the invoking operating-system user and UID. CLI flags and environment variables cannot select another identity.
  • Prefer approvals approve/reject APPROVAL_ID; jobs approve/reject JOB_ID remains a legacy convenience for that job's current pending approval and cannot override approver identity.
  • Use approvals list --json to copy the complete approval UUID before making a decision.
  • approval_auto: "approve" is invalid with an approver scope. Scoped timeout resolution fails closed.
  • The approval snapshots risk, scope, and a canonical SHA-256 binding of the execution contract. Changing, disabling, or deleting the job cancels the pending approval instead of applying it to different work.
  • approval_timeout_s controls the auto-resolution deadline and approval_auto selects "approve" or "reject" for unscoped jobs.

Declared output formats

Set output_format to json, ndjson, or text when downstream work requires a validated result shape. Successful JSON output must parse as one JSON value; each nonblank NDJSON line must parse independently; text is stored on the normalized text path. The scheduler persists validity, byte count, SHA-256 digest, and either the parsed value or an artifact reference on the run. Malformed JSON or NDJSON remains a successful execution with structured_output_valid: 0, a structured_output_warning, and a null parsed value; success children may still run because transport success is distinct from output-shape validation.

Post-success verification

Any synchronous job can declare a separate local shell verification command that runs after its primary execution and structured-output parsing, but before terminal evidence, delivery, and child dispatch:

{
  "verify_shell": "test -f /srv/app/healthy",
  "verify_timeout_s": 30,
  "verify_on_failure": "error"
}

verify_on_failure accepts error or warn. An error policy converts a failed verification to a terminal error and blocks success children; a warning policy preserves primary success and records the warning. Verification stores only status, timing, exit metadata, byte counts, and SHA-256 digests, never raw verification output. Fire-and-forget jobs cannot declare verification.


Idempotency

Control what happens when the dispatcher crashes mid-run. Despite the legacy field name delivery_guarantee, this setting governs execution replay, not external message-delivery confirmation.

# Enable at-least-once: crashed runs replay on next startup
openclaw-scheduler jobs update <id> '{"delivery_guarantee":"at-least-once"}'

# Default (at-most-once): no replay
openclaw-scheduler jobs update <id> '{"delivery_guarantee":"at-most-once"}'

How it works:

  • at-most-once (default): if dispatcher crashes mid-run, run is marked crashed and the schedule advances normally. The run is not replayed.
  • at-least-once: on startup, any running run from a crashed dispatcher is replayed with a new run. replay_of field tracks the original run ID for lineage.

Idempotent agents can return IDEMPOTENT_SKIP in their response to acknowledge they've already processed this execution (detected via the idempotency ledger).

The ledger also prevents double-dispatch in concurrent tick scenarios — each run acquires a lock before dispatch.


Context Retrieval

Inject prior run summaries into a job's prompt so the agent has awareness of recent outcomes.

# Inject last 3 run summaries into job prompt
openclaw-scheduler jobs update <id> '{"context_retrieval":"recent","context_retrieval_limit":3}'

# Hybrid: recent runs + TF-IDF search for semantically relevant summaries
openclaw-scheduler jobs update <id> '{"context_retrieval":"hybrid","context_retrieval_limit":5}'

Modes:

Mode Description
none No context injected (default)
recent Last N run summaries, newest first
hybrid Recent runs + TF-IDF similarity search against all prior summaries

Useful for health check jobs that should know about yesterday's failures, or audit jobs that build incrementally on prior work.


Task Tracker

The task tracker provides a dead-man's-switch for coordinating multi-agent sub-agent teams. Create a tracker, assign expected agents, and receive a summary when all agents complete (or time out).

# Create a task group to monitor N sub-agents
openclaw-scheduler tasks create '{
  "name": "v2-release-team",
  "expected_agents": ["schema-agent","frontend-agent","docs-agent"],
  "timeout_s": 1800,
  "delivery_channel": "telegram",
  "delivery_to": "YOUR_CHAT_ID"
}'

# Monitor
openclaw-scheduler tasks list
openclaw-scheduler tasks status <tracker-id>

Each agent in the team must send heartbeat updates. If an agent goes silent past its timeout, it's declared dead. When all agents complete or time out, a summary is delivered to the configured channel.


Resource Pools

Prevent concurrent execution across different jobs that share a resource.

# Two jobs that must not run concurrently
openclaw-scheduler jobs add '{"name":"DB Migration","resource_pool":"database",...}'
openclaw-scheduler jobs add '{"name":"DB Backup","resource_pool":"database",...}'

If one job in a pool is currently running, all other pool members skip their tick (same behavior as overlap_policy: 'skip', but cross-job rather than per-job). Pool membership is set via the resource_pool string field.


Workflow Chains

Jobs can be linked into parent → child chains. When a parent completes, its children fire automatically.

Pattern 1: Chained Jobs

# Parent: runs on cron
openclaw-scheduler jobs add '{
  "name": "Build App",
  "schedule_cron": "0 10 * * *",
  "payload_message": "Build the application",
  "run_timeout_ms": 300000,
  "origin": "system"
}'
# → id: "abc123..."

# Child: fires when parent succeeds
openclaw-scheduler jobs add '{
  "name": "Deploy App",
  "payload_message": "Deploy to production",
  "parent_id": "abc123...",
  "trigger_on": "success",
  "run_timeout_ms": 300000
}'

# Child: fires when parent fails
openclaw-scheduler jobs add '{
  "name": "Build Alert",
  "payload_message": "Build failed -- check logs",
  "parent_id": "abc123...",
  "trigger_on": "failure",
  "delivery_mode": "announce",
  "delivery_to": "YOUR_CHAT_ID",
  "run_timeout_ms": 300000
}'

Trigger types:

  • success — parent run status = ok
  • failure — parent run status = error or timeout
  • complete — any completion (success, failure, or timeout)
  • Child jobs are chain-triggered only. Use trigger_delay_s to delay a child run; one-shot schedule_kind: "at" is for root jobs only.

Pattern 2: Output-Based Trigger Conditions

# Only fire child if parent output contains "ALERT"
openclaw-scheduler jobs add '{
  "name": "Alert Handler",
  "parent_id": "<monitor-job-id>",
  "trigger_on": "success",
  "trigger_condition": "contains:ALERT",
  "payload_message": "Handle the alert",
  "run_timeout_ms": 300000
}'

# Regex condition
openclaw-scheduler jobs add '{
  "name": "Critical Error Handler",
  "parent_id": "<monitor-job-id>",
  "trigger_on": "success",
  "trigger_condition": "regex:ERROR.*critical",
  "payload_message": "Handle critical error",
  "run_timeout_ms": 300000
}'

regex: conditions use RE2 syntax and linear-time matching. Backreferences, lookahead, and lookbehind are rejected. The full condition is limited to 1,024 characters, and regex evaluation fails closed when the parent output exceeds 65,536 UTF-8 bytes. Use contains: when a literal substring is sufficient.

Pattern 3: Multi-Agent Workflows

Chain jobs targeting different agents:

Build (agent: main, cron: 10am)
  └─ Deploy (agent: ops, trigger: success)
      └─ Health Check (agent: main, trigger: success, delay: 60s)
openclaw-scheduler jobs add '{
  "name": "Deploy",
  "payload_message": "deploy",
  "agent_id": "ops",
  "parent_id": "<build-id>",
  "trigger_on": "success",
  "run_timeout_ms": 300000
}'

Pattern 4: Delayed Triggers

openclaw-scheduler jobs add '{
  "name": "Post-Deploy Check",
  "payload_message": "Verify services healthy",
  "parent_id": "<deploy-id>",
  "trigger_on": "success",
  "trigger_delay_s": 60,
  "run_timeout_ms": 300000
}'

Pattern 5: Runtime Spawning

A running agent can create new jobs on the fly by sending a spawn message:

{
  "from_agent": "main",
  "to_agent": "scheduler",
  "kind": "spawn",
  "body": "{\"name\":\"Dynamic Task\",\"payload_message\":\"analyze results\",\"delete_after_run\":true,\"run_now\":true}"
}

Visualizing Chains

openclaw-scheduler jobs tree

# Output (all root jobs and their chains):
# Build App
#   └─ Deploy App [→success] (agent:ops)
#   └─ Build Alert [→failure]
#   └─ Health Check [→complete +60s]

Retry Logic

Jobs can auto-retry before declaring failure and triggering failure children.

openclaw-scheduler jobs add '{
  "name": "Flaky Deploy",
  "schedule_cron": "0 10 * * *",
  "payload_message": "deploy to prod",
  "max_retries": 3,
  "run_timeout_ms": 300000,
  "origin": "system"
}'

How it works:

  1. Job fails → check max_retries
  2. Retries remaining → schedule retry with exponential backoff (30s, 60s, 120s, ...)
  3. Retry run tracks lineage: retry_of → failed run ID, retry_count incremented
  4. All retries exhausted → trigger failure children + apply error backoff
  5. Any retry succeeds → trigger success children, reset consecutive_errors

Key: failure children don't fire until all retries are exhausted. This prevents false alerts on transient failures.

This retry ladder now applies uniformly to shell jobs, isolated agent jobs, and main-session jobs that surface dispatch failures.

Field Default Description
max_retries 0 Max retry attempts (0 = no retry)
runs.retry_of null ID of the failed run being retried
runs.retry_count 0 Which attempt this is (0 = first try)

Chain Safety

Max Chain Depth

MAX_CHAIN_DEPTH = 10 — enforced on:

  • createJob — can't add a child deeper than 10 levels
  • updateJob — can't move a job to create a chain deeper than 10
  • triggerChildren — runtime safeguard stops dispatch at depth 10

Cycle Detection

detectCycle() walks up the parent chain on both create and update. Catches:

  • Self-referential: A → A
  • Deep cycles: A → B → C → A
  • Throws with descriptive error message

Chain Cancellation

openclaw-scheduler jobs cancel <job-id>
# Requests cancellation for active runs on this job and its descendants

Cancellation is fenced against dispatcher ownership. Pending work is cancelled atomically. A running shell job receives process-group termination and cannot finalize or trigger children after cancellation wins. Agent work records and sends a Gateway cancellation request. Finished runs are unchanged. Poll runs get, runs running, or status; chat notifications are asynchronous and can be stale.


Inter-Agent Messaging

Agents exchange messages through the scheduler's queue.

Features

  • Priority: 0 (normal), 1 (high), 2 (urgent) — inbox sorted by priority then time
  • Threading: reply_to links messages into conversations
  • Read receipts: pending → delivered → read (with timestamps)
  • Broadcast: to_agent = 'broadcast' reaches all agents
  • TTL/Expiry: expires_at auto-expires unread messages
  • Metadata: JSON blob for structured data
  • Kinds: text, task, result, status, system, spawn, decision, constraint, fact, preference
  • Owner field: owner tracks message originator for audit
  • Job linking: messages can reference job_id and run_id

Delivery

Messages are delivered inline with job prompts. When the dispatcher builds a prompt, it includes up to 5 pending messages for the target agent, marked as delivered.

Usage

# Send a message
openclaw-scheduler msg send <from-agent> <to-agent> "message body"

# Read inbox
openclaw-scheduler msg inbox <agent-id>

# Mark all read
openclaw-scheduler msg readall <agent-id>

Signal Queue Consumer Example

Use this when you want scripts to enqueue only actionable signals, then a single consumer job pushes those signals to Telegram.

# 1) Enqueue a signal
openclaw-scheduler msg send monitor-agent main "Found 3 critical errors in prod logs"

# 2) Run the inbox consumer daemon (event-driven, watches SQLite WAL)
#    Delivers within ~250ms of a message landing in the DB.
#    No polling cron needed — the daemon watches for WAL changes.
node scripts/inbox-consumer.mjs --watch --to YOUR_CHAT_ID --channel telegram

# Or as a one-shot drain (useful for testing):
node scripts/inbox-consumer.mjs --to YOUR_CHAT_ID --channel telegram

Backup & Recovery

MinIO backups are disabled by default. Set SCHEDULER_BACKUP=1 to enable. Requires mc (MinIO client) installed and configured with a backupstore alias.

The scheduler can back up its SQLite database to MinIO automatically.

# Manual snapshot
node backup.js snapshot

# Manual rollup (hourly aggregate)
node backup.js rollup

# Check backup status
node backup.js status

# Restore from snapshot
node backup.js restore

# Prune old backups
node backup.js prune

Configuration via environment:

Variable Default Description
SCHEDULER_BACKUP_MC_ALIAS backupstore MinIO client alias
SCHEDULER_BACKUP_BUCKET scheduler-backups MinIO bucket name
SCHEDULER_BACKUP_PREFIX scheduler Path prefix within bucket

Requires mc (MinIO client) in PATH and a configured backupstore alias.

Built-in (when running as a background service):

  • Snapshot every 5 minutes (SCHEDULER_BACKUP_MS)
  • Rollup on the first tick of each hour

Agent Registry

Operation Function Description
Register upsertAgent(id, opts) Create or update
Get getAgent(id) Fetch by ID
List listAgents() All agents
Set status setAgentStatus(id, status, sessionKey) idle/busy/offline
Touch touchAgent(id) Update last_seen_at

The dispatcher automatically manages agent status during dispatch (idle → busy → idle).

openclaw-scheduler agents list
openclaw-scheduler agents get <id>
openclaw-scheduler agents register <id> [name]

Database Schema

Schema version: 28 | Mode: WAL | Foreign keys: ON

Tables

Table Description
jobs Job definitions (schedule, payload, chain config, delivery)
runs Execution history (status, timing, summaries, retry lineage)
messages Inter-agent prompt queue (pending, prompt_claimed, delivered/read terminal states)
agents Agent registry (status, capabilities, last seen)
approvals Atomic, versioned HITL decisions and dispatch state
task_tracker Multi-agent task group definitions
task_tracker_agents Per-agent status within a task group
idempotency_ledger Dispatch deduplication and at-least-once tracking
delivery_aliases Named delivery targets (channel + target pairs)
job_dispatch_queue Leased durable dispatch claims and replay state
dispatcher_leases Singleton dispatcher ownership and fencing tokens
delivery_outbox Transactional external delivery queue with multipart group/part coordinates
delivery_attachments Durable attachment content/path and integrity metadata
completion_debts Run-scoped completion delivery ownership and recovery state
evidence_records Immutable, content-addressed SHA-256 evidence, one row per run
message_receipts Delivery receipt tracking for messages
team_tasks Team-scoped task definitions and status
team_mailbox_events Projected events from team mailbox activity
schema_migrations Baseline schema version log

Jobs (key columns)

id, name, enabled, schedule_kind, schedule_cron, schedule_at, schedule_tz,
session_target, agent_id, payload_kind, payload_message,
payload_model, payload_model_fallback, payload_thinking, execution_intent, execution_read_only,
overlap_policy, run_timeout_ms, max_queued_dispatches, max_pending_approvals,
max_trigger_fanout, output_store_limit_bytes, output_excerpt_limit_bytes,
output_summary_limit_bytes, output_offload_threshold_bytes,
max_retries, delivery_mode, delivery_channel,
delivery_to, delivery_guarantee, delete_after_run, ttl_hours,
parent_id, trigger_on, trigger_delay_s, trigger_condition,
resource_pool, auth_profile, auth_profile_fallback,
approval_required, approval_timeout_s, approval_auto,
approval_risk_level, approval_approver_scope, output_format,
verify_shell, verify_timeout_s, verify_on_failure,
context_retrieval, context_retrieval_limit,
preferred_session_key, job_type, watchdog_target_label,
watchdog_check_cmd, watchdog_timeout_min, watchdog_alert_channel,
watchdog_alert_target, watchdog_self_destruct, watchdog_started_at,
next_run_at, last_run_at, last_status, consecutive_errors,
created_at, updated_at

Runs (key columns)

id, job_id, status, started_at, finished_at, duration_ms,
last_heartbeat, session_key, session_id, summary,
error_message, shell_exit_code, shell_signal, shell_timed_out,
shell_stdout, shell_stderr, shell_stdout_path, shell_stderr_path,
shell_stdout_bytes, shell_stderr_bytes, dispatched_at, run_timeout_ms,
triggered_by_run, retry_of, retry_count, replay_of,
delegation_validation, output_format, structured_output, structured_output_valid,
structured_output_warning, structured_output_bytes, structured_output_sha256,
structured_output_path, verification_result, approval_used

Run statuses: pending, running, ok, error, timeout, skipped, cancelled, crashed, recovery_blocked, awaiting_approval, approved

Messages (key columns)

id, from_agent, to_agent, reply_to, kind, subject, body,
metadata, priority, channel, owner, status, delivered_at,
read_at, expires_at, created_at, job_id, run_id

Agents (10 columns)

id, name, status, last_seen_at, session_key, capabilities,
delivery_channel, delivery_to, brand_name, created_at

CLI Reference

# ── Jobs ──────────────────────────────────────────
openclaw-scheduler jobs list                     # List all (supports --type <type>)
openclaw-scheduler jobs list --include-handoff-artifacts --json # Validate and hydrate persisted v4 artifacts
openclaw-scheduler jobs get <id>                 # Full details as JSON
openclaw-scheduler jobs add --file job.json      # Inline JSON and --stdin are also supported
openclaw-scheduler jobs update <id> --stdin      # Partial update from standard input
openclaw-scheduler jobs validate --file job.json # Validate without DB initialization
openclaw-scheduler jobs enable <id>
openclaw-scheduler jobs disable <id>              # NOTE: one-shot at-jobs with delete_after_run: true are auto-pruned after 24h (ordinary disabled cron jobs are kept indefinitely)
openclaw-scheduler jobs delete <id>              # Cascades to runs
openclaw-scheduler jobs tree                     # Visual chain hierarchy
openclaw-scheduler jobs cancel <id>              # Cancel running chain

# ── Runs ──────────────────────────────────────────
openclaw-scheduler runs list <job-id> [limit]    # Run history
openclaw-scheduler runs get <run-id>             # Full run details
openclaw-scheduler runs output <run-id> stdout   # Stored/offloaded stdout or stderr
openclaw-scheduler runs evidence <run-id>        # Verify and return immutable execution evidence
openclaw-scheduler runs running                  # Active runs
openclaw-scheduler runs stale [threshold-s]      # Stale runs (default 90s)

# ── Messages ──────────────────────────────────────
openclaw-scheduler msg send <from> <to> <body>
openclaw-scheduler msg inbox <agent-id> [limit]
openclaw-scheduler msg outbox <agent-id> [limit]
openclaw-scheduler msg thread <message-id>
openclaw-scheduler msg ack <message-id> [actor] [note]
openclaw-scheduler msg receipts <message-id> [limit]
openclaw-scheduler msg team-inbox <team-id> [limit] [member-id] [task-id]
openclaw-scheduler msg read <message-id>
openclaw-scheduler msg readall <agent-id>
openclaw-scheduler msg unread <agent-id>

# ── Agents ────────────────────────────────────────
openclaw-scheduler agents list
openclaw-scheduler agents get <id>
openclaw-scheduler agents register <id> [name]

# ── Approvals ─────────────────────────────────────
openclaw-scheduler approvals list --json                 # Full pending approval IDs
openclaw-scheduler approvals approve <id> [--reason ...] # Approve by approval ID
openclaw-scheduler approvals reject <id> [--reason ...]  # Reject by approval ID
openclaw-scheduler jobs approve <job-id>                  # Legacy current-gate lookup
openclaw-scheduler jobs reject <job-id> [reason]          # Legacy current-gate lookup

# ── Task Tracker ──────────────────────────────────
openclaw-scheduler tasks create '<json>'        # Create task group
openclaw-scheduler tasks list                   # Active task groups
openclaw-scheduler tasks status <id>            # Detailed status
openclaw-scheduler tasks history [limit]        # Recently completed groups
openclaw-scheduler tasks heartbeat <id> <label> running|completed|failed [msg]
openclaw-scheduler tasks register-session <id> <label> <session-key>  # Enable auto-heartbeat

# ── Queue ─────────────────────────────────────────
openclaw-scheduler queue list [agent] [limit]    # Pending + delivered messages
openclaw-scheduler queue clear [agent]           # Mark all messages read
openclaw-scheduler queue prune                   # Prune old messages

# ── Team Adapter ─────────────────────────────────
openclaw-scheduler team map [limit]                          # Project team messages into events
openclaw-scheduler team tasks <team-id> [limit]              # List team tasks
openclaw-scheduler team events <team-id> [limit] [task-id]   # List team events
openclaw-scheduler team gate <team-id> <task-id> <members-json> [timeout-s]
openclaw-scheduler team check-gates [limit]                  # Evaluate task gates
openclaw-scheduler team ack <message-id> [actor] [note]      # Team-aware ACK

# ── Idempotency ──────────────────────────────────
openclaw-scheduler idem status <job-id>          # Recent idempotency keys
openclaw-scheduler idem check <key>              # Check if key is claimed
openclaw-scheduler idem release <key>            # Manually release a key
openclaw-scheduler idem prune                    # Force prune expired entries

# ── Delivery Aliases ──────────────────────────────
openclaw-scheduler alias list                 # List all aliases
openclaw-scheduler alias add <name> <channel> <target> [description]
openclaw-scheduler alias remove <name>

# ── Runtime Diagnostics ──────────────────────────
openclaw-scheduler status --json              # Live DB, lease, queue, outbox, approval state
openclaw-scheduler doctor --json              # Fail-closed schema and operational checks

# ── Schema Introspection ─────────────────────────
openclaw-scheduler schema jobs              # JSON schema for job fields (types, defaults, enums)
openclaw-scheduler schema runs              # Run statuses and key fields
openclaw-scheduler schema messages          # Message kinds and statuses
openclaw-scheduler schema approvals         # Approval statuses
openclaw-scheduler schema dispatches        # Dispatch kinds and statuses
openclaw-scheduler schema all               # Everything

# ── Status ────────────────────────────────────────
openclaw-scheduler status
openclaw-scheduler version                       # Print version (also: --version)

All CLI commands support --json for machine-readable output (useful for piping into jq or agent toolchains).


Configuration

Variable Default Description
OPENCLAW_GATEWAY_URL http://127.0.0.1:18789 Gateway endpoint
OPENCLAW_GATEWAY_TOKEN (required) Gateway auth token
OPENCLAW_GATEWAY_TOKEN_PATH ~/.openclaw/credentials/.gateway-token Path to a gateway token file under ~/.openclaw/credentials, /run/secrets, or /var/run/secrets (used when OPENCLAW_GATEWAY_TOKEN is not set)
SCHEDULER_HOME ~/.openclaw/scheduler Base dir for scheduler data when installed from npm or when the package dir is not a writable source checkout
SCHEDULER_DB auto (existing ~/.openclaw/scheduler/scheduler.db first; otherwise ./scheduler.db in a writable source checkout; otherwise scheduler home) SQLite database path
SCHEDULER_BACKUP_STAGING_DIR ~/.openclaw/scheduler/.backup-staging Temp folder used by backup.js snapshot/restore
SCHEDULER_TICK_MS 10000 Tick interval in ms, from 1000 through 3600000
SCHEDULER_STALE_THRESHOLD_S 90 Stale run threshold in seconds, from 10 through 604800
SCHEDULER_HEARTBEAT_CHECK_MS 30000 Health check interval in ms, from 5000 through 3600000
SCHEDULER_MESSAGE_DELIVERY_MS 15000 Message and spawn interval in ms, from 5000 through 3600000
SCHEDULER_DELIVERY_BATCH_SIZE 10 Delivery batch size, from 1 through 1000
SCHEDULER_PRUNE_MS 3600000 Prune interval in ms, from 60000 through 604800000
SCHEDULER_BACKUP_MS 300000 MinIO backup interval in ms, from 60000 through 604800000
SCHEDULER_LEASE_TTL_MS 30000 Dispatcher lease TTL in ms, from 15000 through 3600000
SCHEDULER_MAX_CONCURRENCY 4 Worker concurrency, from 1 through 64
SCHEDULER_MAX_PENDING_WORK 1000 Pending-work bound, from concurrency through 10000
SCHEDULER_BACKUP (unset) 1 or true enables MinIO backups; 0 or false disables them
SCHEDULER_BACKUP_MC_ALIAS backupstore MinIO alias used by mc for backup snapshots
SCHEDULER_BACKUP_BUCKET scheduler-backups MinIO bucket for snapshots
SCHEDULER_BACKUP_PREFIX scheduler Object prefix inside bucket
SCHEDULER_ARTIFACTS_DIR ~/.openclaw/scheduler/artifacts Directory for offloaded shell stdout/stderr files
SCHEDULER_DEBUG (unset) 1 or true enables debug logging; 0 or false disables it
SCHEDULER_SHELL /bin/zsh (macOS), /bin/bash (Linux/WSL2) Shell used for shell jobs
SCHEDULER_PROVIDER_PATH (unset) Directory of provider plugin *.js files loaded at startup. High trust boundary -- only point at operator-controlled code. See gateway contract
DISPATCH_CONFIG_DIR ~/.openclaw/dispatch Override dispatch config directory for config.json
DISPATCH_STATE_DIR ~/.openclaw/scheduler/dispatch Allowed state root for the dispatch labels ledger
DISPATCH_LABELS_PATH <DISPATCH_STATE_DIR>/labels.json Override the labels ledger path. Relative paths resolve beneath the state root; absolute paths must remain beneath it after traversal normalization, and existing parents must remain beneath it after symbolic-link resolution
DISPATCH_INDEX_PATH (auto) Override path to dispatch/index.mjs (used by watcher)
DISPATCH_HOST hostname() Host identifier sent with dispatch hook events
DISPATCH_WEBHOOK_URL (unset) Webhook URL for dispatch lifecycle events (hooks.mjs)
LOKI_PUSH_URL (unset) Loki push endpoint for dispatch event logging (hooks.mjs)
TELEGRAM_BOT_TOKEN (unset) Bot token for webhook health check utility
TELEGRAM_WEBHOOK_URL (unset) Expected webhook URL for Telegram webhook check
INBOX_AGENT main Target agent for inbox-consumer.mjs
INBOX_DELIVERY_CHANNEL (unset) Delivery channel for inbox-consumer.mjs forwarding
INBOX_DELIVERY_TO (unset) Delivery target for inbox-consumer.mjs forwarding
INBOX_LIMIT 10 Batch size for inbox-consumer.mjs

Provider-backed identity / authorization / proof behavior, including provider-resolved authorization_ref and its fail-closed error semantics, is documented in the gateway contract.


Service Management

Platform note: The commands below are for macOS (launchd). For Linux, see INSTALL-LINUX.md. For Windows, see INSTALL-WINDOWS.md.

Choose the launchd mode that matches your host:

  • LaunchAgent: best for a personal Mac that auto-logs in and should run the scheduler in your user session
  • LaunchDaemon: best for a headless Mac or for starting the scheduler before login

Install either mode with the setup wizard:

openclaw-scheduler setup --service-mode agent
# or
openclaw-scheduler setup --service-mode daemon

macOS LaunchAgent

# Start / bootstrap
launchctl bootstrap gui/$UID ~/Library/LaunchAgents/ai.openclaw.scheduler.plist

# Stop
launchctl bootout gui/$UID/ai.openclaw.scheduler

# Restart
launchctl kickstart -k gui/$UID/ai.openclaw.scheduler

# Status
launchctl print gui/$UID/ai.openclaw.scheduler
ps aux | grep dispatcher | grep -v grep

# Logs
tail -f /tmp/openclaw-scheduler.log

# Quick health
openclaw-scheduler status

macOS LaunchDaemon

# Start / bootstrap
sudo launchctl bootstrap system /Library/LaunchDaemons/ai.openclaw.scheduler.plist

# Stop
sudo launchctl bootout system/ai.openclaw.scheduler

# Restart
sudo launchctl kickstart -k system/ai.openclaw.scheduler

# Status
sudo launchctl print system/ai.openclaw.scheduler
ps aux | grep dispatcher | grep -v grep

# Logs
tail -f /tmp/openclaw-scheduler.log

# Quick health
openclaw-scheduler status

Both modes use RunAtLoad: true and KeepAlive: true. LaunchDaemon also sets UserName: <your-user> so the service runs under your OpenClaw account while still surviving headless reboots.


Error Handling & Backoff

On dispatch failure

  1. Run marked error, consecutive_errors increments
  2. If max_retries > 0 and retries remain → schedule retry (failure children wait)
  3. If retries exhausted → trigger failure children, apply backoff

Backoff schedule

Consecutive errors Delay
1 30s
2 1 min
3 5 min
4 15 min
5+ 1 hour

Backoff is applied on top of the cron schedule (whichever is later). Resets to 0 on success.

Stale run detection

  • Every 30s, dispatcher checks if running runs still have active sessions
  • No activity for 90s → marked timeout
  • Fallback: runs exceeding run_timeout_ms are force-timed-out

Gateway health

GET /health is checked before dispatch. If it is unreachable, isolated jobs are deferred and shell jobs continue. Main-session jobs also require the Gateway: default, execute, or plan jobs use synchronous agent execution, while fire-and-forget jobs use openclaw system event. A failed synchronous health check defers the job for 60 seconds; later request failures and fire-and-forget failures use the configured retry behavior.


Migration & History

Importing from current OpenClaw

openclaw-scheduler migrate --dry-run --json
openclaw-scheduler migrate --json

Use --legacy-json ~/.openclaw/cron/jobs.json only for an old export.

Schema baseline

Version 0.5.0 uses schema version 29.

  • Net-new installs apply schema.sql transactionally.
  • Existing databases run the idempotent consolidation migration, then reapply the current schema. Migration 28 preserves existing state while rebuilding legacy completion debt rows with a derived delivery scope, multipart outbox coordinates, and immutable checksum evidence. Migration 29 adds handoff v4 artifacts, proof replay claims, runtime events, provider sessions, credential presentations, exact source-run bindings, and signed evidence metadata without rewriting earlier jobs or runs.
  • Any schema or consolidation error aborts startup. It is never logged and ignored.

Native-job ownership and release history

Disable only native jobs that have a verified scheduler replacement. Native OpenClaw cron and heartbeat may continue serving unrelated work. Rollback stops or disables the scheduler copies and re-enables those native jobs. See CHANGELOG.md for release and schema history instead of relying on duplicated version tables in this reference.


Upgrading

Already have the scheduler running and need to update? See UPGRADING.md for the full guide. Short version by platform:

macOS (launchd, git-clone install)

cd ~/.openclaw/scheduler
git pull && npm install
npm run verify:local
launchctl kickstart -k gui/$(id -u)/ai.openclaw.scheduler

Linux / Windows WSL2 (systemd, git-clone install)

cd ~/.openclaw/scheduler
git pull && npm install
npm run verify:local
systemctl --user restart openclaw-scheduler

Removing the Scheduler

To stop the scheduler and restore OpenClaw's built-in cron/heartbeat, see UNINSTALL.md.

Quick summary:

  1. Stop the service (launchctl / systemctl / pm2)
  2. Re-enable OC cron globally: openclaw config set cron.enabled true and remove OPENCLAW_SKIP_CRON=1 from gateway service env
  3. Re-enable OC cron jobs: openclaw cron edit <id> --enable for each job
  4. Re-enable heartbeat: openclaw config set agents.defaults.heartbeat.every "5m" and restore any per-agent agents.list[].heartbeat overrides you use
  5. Optionally delete ~/.openclaw/scheduler/

Best Practices

See BEST-PRACTICES.md for:

  • Choosing between shell, isolated, and main session targets
  • Writing effective payload prompts for LLM jobs
  • When to use chains vs standalone jobs
  • Delivery mode selection
  • How to integrate the scheduler with your OpenClaw agent
  • Example MEMORY.md entries for agent awareness

File Reference

~/.openclaw/scheduler/
│
│  Core scheduler
├── dispatcher.js          # Main process — tick loop, dispatch, chains, retry, backups
├── dispatcher-strategies.js  # Dispatch strategy functions (prepare, execute, finalize)
├── dispatcher-maintenance.js # Stale run reaping, TTL pruning, WAL checkpoints
├── dispatcher-approvals.js   # Approval timeout resolution and auto-approve/reject
├── dispatcher-delivery.js    # Post-run delivery pipeline (announce, announce-always)
├── dispatcher-shell.js       # Shell job execution and result normalization
├── dispatcher-runtime.js     # Singleton lease and bounded worker ownership
├── dispatcher-utils.js       # Shared dispatcher helpers and dependency wiring
├── dispatch-queue.js         # Leased durable dispatch queue and stale-claim recovery
├── runtime-lease.js          # Dispatcher lease and fencing token persistence
├── run-state.js              # Fenced cancellation and terminal run transitions
├── run-completion.js         # Atomic completion, job update, and child enqueue
├── delivery-outbox.js        # Transactional external delivery queue
├── attachment-store.js       # Durable attachment staging and integrity metadata
├── governance.js             # Execution-time governance enforcement
├── db.js                  # SQLite connection, fail-closed schema init, WAL checkpoints
├── schema.sql             # Complete schema (v28)
├── migrate-consolidate.js # Single migration for existing DBs through v28
├── jobs.js                # Job CRUD, cron, chains, cycle detection, resource pools, queue
├── runs.js                # Run lifecycle, stale/timeout, cancellation, context summary
├── messages.js            # Inter-agent message queue (priority, TTL, typed messages)
├── agents.js              # Agent registry
├── gateway.js             # OpenClaw API client (chat completions, events, delivery, aliases)
├── approval.js            # HITL approval queries
├── approval-state.js      # Atomic approval decisions and dispatch claims
├── idempotency.js         # Idempotency ledger (execution replay dedup)
├── retrieval.js           # Context retrieval (recent/hybrid run summaries)
├── task-tracker.js        # Dead-man's-switch for multi-agent sub-agent teams
├── team-adapter.js        # Team mailbox/task projection and task completion gates
├── backup.js              # MinIO snapshot/rollup/restore (requires `mc` CLI)
├── cli.js                 # CLI management tool
├── migrate.js             # Import through OpenClaw cron CLI or explicit legacy JSON
├── scripts/
│   ├── dispatch-cli-utils.mjs       # Dispatch CLI path resolution helpers
│   ├── inbox-consumer.mjs           # Drains queue messages and delivers to Telegram
│   ├── stuck-run-detector.mjs       # Detects stale running runs (alert-only via non-zero exit)
│   └── telegram-webhook-check.mjs   # Telegram webhook health check / repair utility
│
│  Service & docs
├── ~/Library/LaunchAgents/ai.openclaw.scheduler.plist  # macOS LaunchAgent location after install
├── /Library/LaunchDaemons/ai.openclaw.scheduler.plist  # macOS LaunchDaemon location after install
├── INSTALL.md             # Full installation guide — macOS (first host)
├── INSTALL-ADDITIONAL-HOST.md  # Installation guide for additional hosts
├── INSTALL-LINUX.md       # Installation guide for Linux (systemd user service)
├── INSTALL-WINDOWS.md     # Installation guide for Windows (WSL2 only)
├── UPGRADING.md           # Upgrade guide (all platforms)
├── UNINSTALL.md           # Removal guide (all platforms)
├── BEST-PRACTICES.md      # Job type selection, prompt writing, agent integration
├── QUICK-START.md         # Focused guide: install, convert crons, first workflow
├── openclaw-scheduler.service  # Linux systemd user service template
├── CHANGELOG.md           # Version history
└── README.md              # This file

Testing

# Run the legacy suite, isolated focused tests, docs, and agentcli integration when present
npm test

# Require scheduler-owned and agentcli-owned integration suites
npm run test:agentcli

# Add lint, types, coverage, and package verification
npm run verify:local

The agentcli checkout defaults to ../agentcli; set AGENTCLI_PATH when it is elsewhere. A missing checkout is reported as an explicit local skip. SKIP_AGENTCLI_INTEGRATION=1 npm test records an intentional scheduler-only iteration. Hosted CI runs the scheduler-owned contract and the compatible upstream-owned tests against exact public handoff-v3 and handoff-v2 commits. At the v3 pin, the upstream test that hard-codes field_version: "2" is reported as an explicit skip while all other upstream integration tests run.

Focused tests/*.test.mjs files run sequentially in separate processes and databases so shared module state cannot contaminate another file.

Test categories

  • Schema creation & integrity
  • Job CRUD, cron parsing, due detection
  • Run lifecycle (create, heartbeat, finish, stale, timeout)
  • Agent registry (upsert, status, capabilities)
  • Message queue (priority, broadcast, TTL, typed messages)
  • Cascade deletes, pruning
  • Workflow chains (parent/child, trigger matching, tree traversal, trigger conditions)
  • Cycle detection (self, deep)
  • Max chain depth enforcement
  • Retry tracking and sequencing
  • Chain cancellation
  • Shell job execution
  • Approval gate lifecycle
  • Idempotency key claiming/releasing
  • Context retrieval (recent/hybrid)
  • Dispatcher integration (full dispatch pipeline with mock gateway)

Sub-agent Dispatch

The dispatch module (dispatch/index.mjs) spawns and steers isolated agent sessions via the OpenClaw Gateway API and tracks them by a human-readable label. Unlike the scheduler's job/run model, dispatch calls the gateway directly -- no scheduler tick delay, no DB write required to start a session. Each session is assigned a unique session key and recorded in the originating host's local labels.json ledger. When the agent calls done from that same local dispatch shell, final delivery is claimed by label, session, and run, then enqueued durably in delivery_outbox. The routed watcher uses the same claim and outbox path, so the two completion paths cannot both enqueue the same run. Do not run the completion marker from inside ssh, Docker, tmux, or another nested shell, because that can update a different label store. The watcher normalizes completions into short human-readable summaries, and it only falls back to terminal assistant output when the transcript contains strict clean completion evidence. The module also supports symlink-based branding: a wrapper directory (such as my-brand) contains a config.json with a custom name and a symlink to dispatch/index.mjs, giving the same CLI a different identity in notifications and logs.

Routed watchers write no delivery body to stdout after a durable enqueue and emit WATCHER_ALREADY_DELIVERED on stderr for the scheduler wrapper. Despite its legacy name, that marker means the completion was already enqueued and the wrapper must not enqueue it again; it is not a channel delivery receipt. The run-scoped completion debt stays open until every outbox part is delivered. A watcher with no configured delivery route retains its historical stdout behavior, but only after it acquires the durable run-scoped completion claim. If completion claim storage is unavailable, delivery fails closed with COMPLETION_CLAIM_UNAVAILABLE. Multipart output is split into independent outbox rows and can resume from its recorded per-part checkpoint.

Quick Example

# Dispatch a sub-agent task and deliver the result to Telegram
openclaw-scheduler enqueue \
  --label       "fix-deploy-script"                                          \
  --message     "Fix the deploy script in ~/app to handle missing .env files" \
  --mode        fresh                                                         \
  --thinking    high                                                          \
  --timeout     3600                                                          \
  --deliver-to  YOUR_CHAT_ID                                                     \
  --delivery-mode announce

# Fallback (if openclaw-scheduler is not in PATH):
node ~/.openclaw/scheduler/dispatch/index.mjs enqueue \
  --label "fix-deploy-script" --message "..." --deliver-to YOUR_CHAT_ID

# Literal-safe prompt input when the text contains shell metacharacters:
cat prompt.md | openclaw-scheduler enqueue \
  --label "fix-deploy-script" \
  --message-stdin \
  --deliver-to YOUR_CHAT_ID

For normal chat-triggered dispatches, always pass --deliver-to from the inbound metadata chat_id. If you omit --origin, dispatch now derives it from that explicit delivery target instead of guessing from whichever session was active most recently. The old active-session lookup is kept only as a manual/local fallback when both values are absent.

Flag Reference

Flag Default Description
--label required Human-readable name for the session. Used for status lookups, reuse, and watchdog tracking.
--message required* Prompt sent to the agent.
--message-file -- Path to a file whose contents are used as the prompt. Use - to read from stdin. Safer than inline shell quoting for prompts with backticks, quotes, or markdown.
--message-env -- Environment variable name whose value is used as the prompt. Useful when the prompt contains shell-significant characters.
--message-stdin -- Read the prompt from stdin explicitly. If stdin is piped and no explicit prompt source is set, dispatch auto-reads stdin.
--mode fresh fresh creates a new session. reuse continues the last session recorded for this label.
--thinking -- Reasoning budget: low, high, or xhigh.
--model configured dispatch default Model override, e.g. anthropic/claude-sonnet-4-6. When omitted, dispatch uses wrapper config.defaultModel, wrapper config.dispatch.model, DISPATCH_DEFAULT_MODEL, agents.defaults.dispatch.model, agents.defaults.model, then the built-in fallback.
--deliver-to -- Delivery target (e.g. Telegram chat ID). Registers the scheduler watcher job for durable final delivery. The gateway spawn itself stays fire-and-forget so raw tool output and internal done payloads cannot leak directly to chat. Chat-triggered callers should pass inbound metadata chat_id here, especially for group chats.
--delivery-mode announce announce delivers only when output is non-empty. announce-always delivers unconditionally. none suppresses delivery.
--timeout 300 Session timeout in seconds.
--monitor on Auto-register a watchdog job that alerts if the session goes silent past the configured threshold.
--no-monitor -- Disable watchdog registration for this dispatch.

*One prompt source is required: --message, --message-file, --message-env, --message-stdin, or piped stdin.

Subcommand Reference

Subcommand Description
enqueue Spawn a new agent session (or resume one with --mode reuse) and optionally register a scheduler watcher for delivery.
status Show current status for a label: session key, spawn time, running/done/error, and liveness data from the sessions store.
stuck Check all running sessions against the stuck threshold. Exits 1 if genuinely stuck sessions remain after auto-resolving completed ones.
result Retrieve the last assistant reply from a session transcript via chat.history.
sync Reconcile labels.json with sessions store state. Auto-marks sessions as done or error based on idle time. Supports --dry-run.
done Agent-side completion signal. The agent calls this as its final action from the originating local dispatch shell to mark itself done immediately (push-based; no idle timeout wait). Do not call it from inside a remote or nested shell, because the label lookup is local to the dispatch host. The stored completion payload is normalized for channel-safe delivery, with checklist/sha metadata used as a fallback when the raw summary is generic or noisy.
send Inject a message into a running session for mid-run steering. The agent sees it as a new user turn.
steer Alias for send. The name makes steering intent explicit.
heartbeat Check whether a session has been active within the last 10 minutes. Accepts --label or --session-key.
list List all tracked labels in labels.json, sorted by most recent. Accepts `--status running

Multi-agent Orchestration

The main agent acts as the orchestrator and delegates parallel units of work to sub-agents via enqueue. Each sub-agent runs in an isolated session, completes its assigned task, and calls done as its last action. Results are enqueued for durable outbox delivery to the requesting chat (Telegram, Discord, WhatsApp, Signal, iMessage, or Slack) without the orchestrator polling.

Spawn depth constraint: The gateway enforces maxSpawnDepth: 2. The main agent (depth 0) spawns sub-agents (depth 1), which can spawn nested sub-agents (depth 2). Depth 3 is blocked. The dispatcher sets spawnDepth: 1 on each fresh session automatically.

Example: 3 parallel workers

# Orchestrator dispatches three workers in parallel.
# All three run concurrently in isolated sessions.

openclaw-scheduler enqueue \
  --label   "worker-schema"   \
  --message "Review the DB schema and write documentation for all tables" \
  --thinking high --timeout 600 --deliver-to YOUR_CHAT_ID

openclaw-scheduler enqueue \
  --label   "worker-frontend" \
  --message "Audit the React components for accessibility issues" \
  --thinking high --timeout 600 --deliver-to YOUR_CHAT_ID

openclaw-scheduler enqueue \
  --label   "worker-docs"     \
  --message "Update the API docs to reflect the new /v2 endpoints" \
  --thinking high --timeout 600 --deliver-to YOUR_CHAT_ID

# Each worker gets an outbox-delivered, channel-safe completion summary when done.
# No polling needed. Watchdog jobs are auto-registered for each.

Check status at any time:

openclaw-scheduler list --status running
openclaw-scheduler dispatch status --label worker-schema

Branding and Configuration

dispatch/index.mjs resolves config.json relative to the directory of the invoking script, not the module itself. This means a symlink at ~/.openclaw/my-brand/index.mjs -> ~/.openclaw/scheduler/dispatch/index.mjs will load ~/.openclaw/my-brand/config.json, giving the same CLI a different brand name and defaults. All config fields are optional.

config.json fields:

Field Default Description
name "dispatch" Brand name shown in Telegram notifications and log output.
startupGraceMs 90000 Grace period (ms) after spawn before stuck detection and auto-resolve activate.
stuckThresholdMs 600000 Silence duration (ms) before a session is considered stuck.
maxWatcherAgeMs 7200000 Max watcher process age (ms) before it is treated as stale.
watchdogIntervalCron "*/15 * * * *" Cron schedule for the auto-registered watchdog job.
watchdogTimeoutMin 60 Sessions running longer than this (minutes) without completing trigger a watchdog alert.
deliver_watcher_ttl_hours 48 TTL (hours) for scheduler-registered deliver-watcher jobs. These jobs are transient; they auto-prune once delivery is confirmed. Lower values prune faster; higher values retain audit history longer.

Environment variables:

Variable Description
DISPATCH_STATE_DIR Allowed state root for dispatch tracking. Default: ~/.openclaw/scheduler/dispatch.
DISPATCH_LABELS_PATH Override path for labels.json. Default: <DISPATCH_STATE_DIR>/labels.json; the resolved path must remain beneath the state root.
OPENCLAW_GATEWAY_TOKEN Gateway auth token. Falls back to ~/.openclaw/openclaw.json if unset.

Minimal config.json:

{
  "name": "my-brand",
  "watchdogIntervalCron": "*/15 * * * *",
  "watchdogTimeoutMin": 60
}

Monitoring and the Watchdog

When --deliver-to is set and --no-monitor is not passed, enqueue automatically registers a watchdog job in the scheduler DB alongside the delivery watcher job. The watchdog runs on the configured cron schedule and calls stuck --threshold-min <watchdogTimeoutMin> for the dispatched label. If the session has been silent past the threshold, the watchdog posts an alert to the configured delivery target and then disables itself.

Check active dispatch sessions:

# List all running dispatch sessions
openclaw-scheduler list --status running

# Check whether any session is stuck (exits 1 if found)
node ~/.openclaw/scheduler/dispatch/index.mjs stuck --threshold-min 15

# Status for a specific label
openclaw-scheduler dispatch status --label fix-deploy-script

The watchdog disarms itself automatically when the agent calls done, when status or sync auto-resolves the session from gateway idle state, or when result is fetched after a successful completion.


Working with agentcli

agentcli is the control-plane companion for the scheduler. It provides manifest authoring, validation, local execution, identity binding, and capability negotiation. The scheduler provides the durable runtime: scheduling, retries, approvals, delivery, and persistent state.

The scheduler works without agentcli -- most jobs are created by the OpenClaw agent itself when a user requests a scheduled task via Telegram or another messaging channel, and operators can also create jobs directly via the CLI. Adding agentcli on top gives you declarative workflow manifests, stable job IDs, v0.2 identity and authorization support, evidence capability negotiation, handoff v4 capability negotiation, and repeatable applies for workflows that outgrow ad-hoc job creation.

Default job listings omit immutable v4 artifact payloads. Control-plane clients that must validate an existing v4 job before updating it use openclaw-scheduler jobs list --include-handoff-artifacts --json. The opt-in read validates each persisted payload against its digest and fails the complete request if any v4 artifact is missing or invalid; legacy jobs are returned unchanged.

Installing agentcli

npm install -g @amittell/agentcli

Starting fresh with both tools

Write a manifest, validate it, then apply it to the scheduler:

# 1. Write a manifest
cat > my-workflow.json <<'JSON'
{
  "version": "0.1",
  "workflows": [{
    "id": "daily-ops",
    "name": "Daily Operations",
    "tasks": [{
      "id": "health-check",
      "name": "Morning Health Check",
      "prompt": "Run the daily health check and report any issues.",
      "target": { "session_target": "isolated", "agent_id": "main" },
      "schedule": { "cron": "0 9 * * *", "tz": "America/New_York" },
      "runtime": { "timeout_ms": 300000 },
      "delivery": { "mode": "announce", "channel": "telegram", "to": "YOUR_CHAT_ID" }
    }]
  }]
}
JSON

# 2. Validate locally (no scheduler needed)
agentcli validate my-workflow.json

# 3. Preview what would be created (dry-run)
agentcli apply my-workflow.json \
  --db ~/.openclaw/scheduler/scheduler.db \
  --scheduler-prefix ~/.openclaw/scheduler \
  --dry-run

# 4. Apply (creates the jobs)
agentcli apply my-workflow.json \
  --db ~/.openclaw/scheduler/scheduler.db \
  --scheduler-prefix ~/.openclaw/scheduler

# 5. Verify
openclaw-scheduler jobs list

Adopting existing scheduler jobs

If you already have jobs created directly via openclaw-scheduler jobs add and want to bring them under agentcli management:

  1. Write a manifest with task names that match your existing job names exactly.

  2. Run a one-time adoption by name:

agentcli apply my-workflow.json \
  --db ~/.openclaw/scheduler/scheduler.db \
  --scheduler-prefix ~/.openclaw/scheduler \
  --adopt-by name \
  --dry-run          # preview first

agentcli apply my-workflow.json \
  --db ~/.openclaw/scheduler/scheduler.db \
  --scheduler-prefix ~/.openclaw/scheduler \
  --adopt-by name    # execute adoption

This replaces each matched job with a new one under agentcli's stable ID scheme (SHA256 of workflow_id:task_id). The old job is deleted after the new one is created.

  1. On subsequent applies, use the default (no --adopt-by flag). Jobs are matched by their stable ID, so the manifest can be renamed or reorganized without losing job mapping.

Workflow chains

agentcli manifests support parent/child task relationships that compile to scheduler trigger chains:

{
  "version": "0.1",
  "workflows": [{
    "id": "deploy-pipeline",
    "name": "Deploy Pipeline",
    "tasks": [
      {
        "id": "build",
        "name": "Build",
        "shell": { "program": "sh", "args": ["-c", "npm run build"] },
        "target": { "session_target": "shell" },
        "schedule": { "cron": "0 2 * * *" },
        "runtime": { "timeout_ms": 600000 }
      },
      {
        "id": "deploy",
        "name": "Deploy",
        "shell": { "program": "sh", "args": ["-c", "fly deploy"] },
        "target": { "session_target": "shell" },
        "trigger": { "parent": "build", "on": "success" },
        "runtime": { "timeout_ms": 300000 }
      },
      {
        "id": "verify",
        "name": "Post-Deploy Verify",
        "prompt": "Verify all services are healthy after deploy.",
        "target": { "session_target": "isolated", "agent_id": "main" },
        "trigger": { "parent": "deploy", "on": "success" },
        "runtime": { "timeout_ms": 300000 },
        "delivery": { "mode": "announce-always", "channel": "telegram", "to": "YOUR_CHAT_ID" }
      }
    ]
  }]
}

This compiles to three scheduler jobs: Build runs on cron, Deploy triggers on Build success, Verify triggers on Deploy success.

v0.2 identity and authorization

agentcli v0.2 manifests add identity profiles, authorization proofs, evidence declarations, and credential handoff. Identity, authorization, and supported handoff fields compile to the scheduler's v0.2 runtime fields and are enforced at dispatch time:

{
  "version": "0.2",
  "identity_profiles": [{
    "id": "stripe-readonly",
    "provider": "stripe",
    "subject": { "kind": "service", "principal": "agent://payments/reader" },
    "auth": { "mode": "service", "scopes": ["read"] },
    "trust": { "level": "supervised" }
  }],
  "workflows": [{
    "id": "payment-ops",
    "tasks": [{
      "id": "check-balance",
      "name": "Check Balance",
      "shell": { "program": "sh", "args": ["-c", "stripe balance retrieve"] },
      "target": { "session_target": "shell" },
      "identity": { "ref": "stripe-readonly" },
      "contract": {
        "required_trust_level": "supervised",
        "trust_enforcement": "strict"
      },
      "schedule": { "cron": "0 9 * * *" },
      "runtime": { "timeout_ms": 60000 }
    }]
  }]
}

Credential materialization is target-specific. Shell jobs receive the exact declared environment, temporary-file, or stdin bindings locally. Provider output must match every artifact binding by name, medium, environment key, file name, cardinality, and required state before any secret is exposed. Isolated agent jobs first require the connected Gateway to advertise chat-completions-env-inject-v1, then send the validated scoped map through x-openclaw-env-inject. If the capability is absent, dispatch fails closed with GATEWAY_ENV_INJECT_UNSUPPORTED before a credential-bearing request is sent. Main-session jobs reject identity.presentation and credential_handoff. Auth-profile-only isolated turns remain compatible with Gateways that do not provide env injection.

Handoff v4 validates delegated identity chains against the exact source run, resolves authorization_ref only through a configured provider, presents credentials through one negotiated medium, and records append-only runtime events. Its evidence envelope binds the canonical artifact, runtime instance, identity, proof, authorization, complete output digests, postcondition, and lineage. Offloaded stdout and stderr are hashed in full, not from their stored excerpts. Historical verification reloads the exact artifact bound to the run. The SSH provider signs and verifies with ssh-keygen -Y; other declared providers must implement equivalent verification. openclaw-scheduler runs evidence RUN_ID --json re-verifies the persisted envelope against the stored execution.

Earlier handoff consumers and direct scheduler job specifications retain the immutable canonical SHA-256 checksum backend. That legacy path remains a separate capability and is never used to downgrade a v4 signed-evidence declaration.

openclaw-scheduler capabilities --json reports handoff_version: "4", the exact top-level handoff_contract canonicalization and binding versions, and the enforcement features root_approval_gate, approval_scope_enforcement: false, structured_output_format, delegation_validation, authorization_ref_resolution, evidence_generation: true, checksum_evidence_generation: true, evidence_integrity: "artifact-bound-signed-or-provider-verified-v4", evidence_contract: "agentcli-handoff-v4", handoff_v4_artifact, artifact_bound_proofs, signed_or_provider_verified_evidence, provider_session_cache, credential_presentation, source_run_bound_delegation, immutable_runtime_events, gateway_capability_discovery, gateway_env_injection_negotiation, multipart_delivery_checkpoints: true, and completion_delivery_scope: "run".

See the agentcli examples directory for fully annotated manifests covering Stripe, Fly.io, Terraform, GitHub CLI, and more. The executable public v4 flow is tests/handoff-v4-e2e.test.mjs; shared positive and negative protocol vectors are under fixtures/handoff-v4/.

Environment variables

Variable Default Description
AGENTCLI_SCHEDULER_DB (none) Path to scheduler SQLite database
AGENTCLI_SCHEDULER_PREFIX (none) npm prefix where scheduler is installed
AGENTCLI_SCHEDULER_BIN (none) Direct path to scheduler CLI binary
AGENTCLI_TARGET standalone Default compilation target (standalone or openclaw-scheduler)
AGENTCLI_OUTPUT (none) Output format (json or ndjson)

Key commands

agentcli validate manifest.json           # Check manifest validity
agentcli compile manifest.json \
  --target openclaw-scheduler --explain   # Preview compiled job specs
agentcli apply manifest.json \
  --db path/to/scheduler.db --dry-run     # Preview changes
agentcli apply manifest.json \
  --db path/to/scheduler.db               # Create/update jobs
agentcli inspect jobs                     # List managed jobs
agentcli exec manifest.json task-id \
  --dry-run --signer none                 # Local execution (no scheduler)

Trust Architecture

The scheduler acts as a control-plane broker for child execution principals. Child tasks are bounded actors that receive only the credentials the scheduler gives them and cannot escalate their own authority. The credential model supports both precreated scoped keys and dynamic per-task key minting via identity providers.

For the full trust model -- including when the scheduler/child boundary is a real security boundary vs. an operational one, the credential flow from operator to child, and what the model does and does not guarantee -- see docs/trust-architecture.md.


Troubleshooting

Dispatcher isn't dispatching

ps aux | grep dispatcher              # Is it running?
tail -20 /tmp/openclaw-scheduler.log   # Any errors?
curl http://127.0.0.1:18789/health     # Gateway reachable?
openclaw-scheduler jobs list                  # Is nextRun in the past?
openclaw-scheduler runs running               # Overlap blocking?

Wrong next_run_at

All dates must be SQLite format (YYYY-MM-DD HH:MM:SS, UTC). nextRunFromCron() handles this. If manually setting dates, don't use ISO format with T/Z.

Force a job to run now

openclaw-scheduler jobs run <job-id>

Check schema version

openclaw-scheduler doctor --json

Service won't start

# LaunchAgent
plutil -lint ~/Library/LaunchAgents/ai.openclaw.scheduler.plist
launchctl bootstrap gui/$UID ~/Library/LaunchAgents/ai.openclaw.scheduler.plist
launchctl print gui/$UID/ai.openclaw.scheduler

# LaunchDaemon
sudo plutil -lint /Library/LaunchDaemons/ai.openclaw.scheduler.plist
sudo launchctl bootstrap system /Library/LaunchDaemons/ai.openclaw.scheduler.plist
sudo launchctl print system/ai.openclaw.scheduler

Logs not updating

Dispatcher logs to stderr (unbuffered). If logs look stale, the process may have crashed. Check the service that matches your launchd mode:

  • LaunchAgent: launchctl print gui/$UID/ai.openclaw.scheduler
  • LaunchDaemon: sudo launchctl print system/ai.openclaw.scheduler

Job shows 'awaiting_approval'

openclaw-scheduler approvals list --json
openclaw-scheduler approvals approve <approval-id> --reason "Reviewed"
# Or: openclaw-scheduler approvals reject <approval-id> --reason "Not ready"

Backup failing

mc alias list   # verify backupstore alias configured
# Check: SCHEDULER_BACKUP_MC_ALIAS, SCHEDULER_BACKUP_BUCKET, SCHEDULER_BACKUP_PREFIX env vars or defaults in backup.js
# Verify MinIO is reachable: mc ls backupstore/

Companion Scripts

The scripts/ directory contains optional operational helpers built on top of core scheduler primitives.

These scripts are not required for scheduling itself, but they are useful for production operations:

  • scripts/inbox-consumer.mjs drains queued messages and delivers them to Telegram.
  • scripts/stuck-run-detector.mjs detects stale running runs and exits non-zero for alerting.

Signal Queue Pattern

The message queue (messages table) plus cli.js msg send implements a signal-only delivery path that complements delivery_mode: announce:

Failure path:  dispatcher → announce → Telegram (immediate, unconditional)
Signal path:   script → cli.js msg send → queue → Inbox Consumer → Telegram

Scripts write to the queue only when they have found something — not unconditionally. A companion scripts/inbox-consumer.mjs shell job (run every 5 min) drains the queue and delivers to Telegram. It exits 0 when the queue is empty, so there is no noise.

Important: The dispatcher does not write to the message queue automatically. Every message in the queue was put there by a script with a specific receiver in mind. Traceability for completed jobs comes from the runs table, delivery_mode: announce, and run history/CLI views — not from queued messages.

About

Durable, governed workflow scheduler and sidecar for OpenClaw agents and shell automation.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages