Trigger an agent

Start agent tasks on a schedule, once at a set time, when an RSS or Atom feed changes, or when a webhook arrives. Define triggers in the agent file:

~/.config/wardrik/agents/triage.md

---
tools: [linear/*]
triggers:
  - name: sweep
    every: 30m
    prompt: Check the incident queue, and say what is new.
  - name: digest
    every: "0 17 * * 1-5"
    timezone: Europe/Berlin
    prompt: Write up what happened today.
---
You watch the incident queue.

wardrik reload
wardrik trigger ls triage

TRIGGER  RUNS                  NEXT       ASKS
digest   every 0 17 * * 1-5    in 3h 12m  Write up what happened today.
sweep    every 30m             in 29m     Check the incident queue, and say what is new.

Trigger messages are added to the agent’s conversation history. See all settings in Agent files.

On a schedule

every: "*/15 * * * *"    # cron, quoted; UTC unless timezone is set
every: "@daily"          # also @hourly, @weekly, @monthly, @yearly
every: 2h                # an interval: s, m, h or d, at least a minute

• A run is skipped if the agent is busy, and missed runs run once on startup. overlap, missed and jitter change that. See triggers.
• By default, if the previous turn is still waiting for approval, the next scheduled run cancels it and starts a new turn. Set overlap: skip to preserve the pending turn.

Once, at a set time

triggers:
  - name: launch
    at: "2026-11-17T09:00"
    timezone: America/Los_Angeles
    prompt: Publish the post in drafts/launch.md.

• at is read in timezone, or takes an offset, as 2026-11-17T17:00:00Z, without timezone.
• The trigger runs once, then it's removed from the agent.
• If the agent is busy, the run waits for its turn. A run missed while the daemon was down runs when it starts. Set missed: skip to drop it.
• Setting a time that has passed is refused.
• With self/*, an agent can schedule its own follow-ups with self_schedule. Each message tells the model when it was sent, so it can work out a time like "Tuesday at 9".

When a feed has new entries

triggers:
  - name: news
    every: 6h
    feed:
      - https://github.blog/changelog/feed/
      - https://hnrss.org/newest?q=postgres
    prompt: Summarise each new entry in one line in NOTES.md.

• When a feed has new entries, Wardrik appends them to the prompt and sends a message to the agent.
• A run with no new entries makes no model call.
• The first run records the current entries and sends nothing.

When a service sends a webhook

triggers:
  - name: issues
    webhook:
      secret: ${ISSUES_HOOK_SECRET}
    prompt: An event is below. If it is a new issue, label it and say who should look.

echo "$SECRET" | wardrik vault put ISSUES_HOOK_SECRET --org acme --for signing_secret

The service posts to {base_url}/v1/hooks/{agent}/{trigger}:

https://agents.acme.example/v1/hooks/agt_01JF8X5RQ2M3V9ZK0B7Y4T6WHN/issues

• The request authenticates with the secret, as Authorization: Bearer SECRET or as the last path segment, .../issues/SECRET.
• verify: standard checks a Standard Webhooks signature instead.
• The secret belongs to the agent's owner, stored --for signing_secret.
• A request that fails verification returns 404. The daemon must be reachable at base_url. See Public HTTPS access.
• Each request costs a model call. Subscribe only to the events you need.

Change triggers

wardrik trigger set triage digest --every "0 17 * * 1-5" --prompt "Write up what happened today."
wardrik trigger rm triage digest

Over HTTP:

curl -X PUT -H "authorization: Bearer $TOKEN" -H "content-type: application/json" \
  http://127.0.0.1:4747/v1/agents/$AGENT/triggers/digest \
  -d '{"every": "0 17 * * 1-5", "prompt": "Write up what happened today."}'

• wardrik reload sets a file-defined agent's triggers to those in the file, and removes the rest.
• With self/* in its tools, an agent can set its own triggers. With agents/*, it can set another agent's. self_schedule and agent_schedule set a schedule; *_trigger_set sets any trigger. Each change needs approval unless the policy allows the tool by name.
• prompt_file: HEARTBEAT.md reads the prompt from the agent's files, so the agent can edit it.
• notify: {to: [ops], on: [finished]} sends each run's reply to a target. See Send notifications.

Details

• Feed entries and webhook bodies are untrusted content.
• A webhook with a repeated ID returns 202 and sends one message.
• If the agent can't accept the webhook, it returns 503, so the service retries.
• The agent receives the first 32,768 characters of a body.
• Each trigger's next run is stored, so triggers survive a restart.