Scheduled changes (RUDDER)¶
"Apply this at 2 AM" without anyone awake at 2 AM. RUDDER does it on its own: no host cron and no model is woken at 2 AM. The same code works under any MCP host.
How it works¶
you (chat) RUDDER (rudder-mcp) rudder-scheduler (service) you (Telegram, navi bot)
"add Loopback99 on change_propose(at="02:00") ──────▶ ticket: scheduled 02:00 EST ──────▶ 🔐 approval request, with the time
LEAF1A at 2 AM" dry-run diff, device locked /pushok CODE (any time before 01:59:30)
01:59:30 ──────────────────────▶ ⏰ "starts in 30 s"
/pushno CODE2 = stop
/pushok CODE2 = start now
no answer = it runs
02:00:00 apply: drift check, commit (EOS: rollback timer), verify,
confirm or roll back ─▶ ✅ / ❌ result
- Propose with a time. The agent calls
change_propose(..., at="02:00"). RUDDER checks the lines, makes the dry-run diff on the device and locks the device until the window, so nothing else changes it in the meantime. - Approve in advance. The approval request on the navi bot shows the diff and the time.
/pushokapproves exactly this change at exactly this time. (The bot replies "the push will run now"; it runs at the scheduled time.) - The notice. 30 seconds before the start (
NOTICE_SEC), you get a notice with a new code: /pushno CODEstops it. Nothing is changed and the lock is released./pushok CODEstarts it at once.- No answer: it runs. You already approved it, so the default is yes.
- Apply. Same path as a manual
change_apply: the device must still give the same diff (no drift), then commit, verify, confirm, or roll back. You get the result on the navi bot.
Safety rules¶
- The agent cannot apply a scheduled ticket early.
change_applyrefuses it; only the scheduler runs it. - The time is part of the approved hash. Moving it (
change_schedule) always needs a new/pushok, and a ticket file edited by hand is refused. - If it is not approved by the notice time, it is not applied ("missed"), and you are told.
- If the scheduler was down and is more than
LATE_MIN(10 min) late, it is not applied ("missed"). - If the notice cannot be sent, the scheduler keeps retrying. Without a notice nothing runs, and after
LATE_MINthe change counts as missed. - PRODUCTION refusal, the deny list, one kind of device per ticket: all the same as for manual changes.
In chat¶
| Say | Tool |
|---|---|
"on LEAF1A add interface Loopback99 / description test, apply at 2 AM" |
change_propose(at="02:00") |
| "schedule ticket C20261112-ab12cd for 2026-11-14 01:30" | change_schedule (new /pushok needed) |
| "what is scheduled?" | change_list (shows at and the schedule state) |
| "status of C20261112-ab12cd" | change_status |
| "cancel C20261112-ab12cd" | change_cancel (any time before it starts) |
Times: "02:00" (the next 02:00), "2026-11-13 02:00", "2026-11-13 02:00 EST", or ISO with an offset.
A time with no zone uses SCHEDULE.TIMEZONE (America/New_York). EST / ET mean New York wall time,
so EST in July means EDT. At most MAX_DAYS (7) ahead.
Settings (rudder.yml)¶
SCHEDULE:
TIMEZONE: America/New_York
NOTICE_SEC: 30 # per ticket: notice_sec=10..600
LATE_MIN: 10
MAX_DAYS: 7
POLL_SEC: 5
The service¶
systemctl --user status rudder-scheduler # running?
journalctl --user -u rudder-scheduler -f # what it did (notice sent, applied, missed, stopped)
systemctl --user restart rudder-scheduler # after a code update or a rudder.yml change
~/navigator/.venv/bin/rudder-scheduler list # scheduled changes
~/navigator/.venv/bin/rudder-scheduler once # one pass by hand (testing)
Install on a new host:
cp ~/navigator/scripts/rudder-scheduler.service ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now rudder-scheduler
It needs atlas-link (for credentials; the unit starts after it) and the Hermes binary for the
messages. One scheduler serves one consumer (CONSUMER in rudder.yml, or env RUDDER_CONSUMER). For a
second consumer, run a second unit with Environment=RUDDER_CONSUMER=LAB_BOT.
Everything it does is in ~/.rudder/audit.log and in the ticket (change_status).
A restart is safe at any time: the state is in the ticket files. If it stops in the middle of applying a change, the EOS commit timer rolls the switch back by itself. Cisco and PAN-OS keep their rollback point and manual rollback command in the ticket.
Why not the agent host's cron?¶
Those wake the AI model at the set time with a prompt, so what happens at 2 AM depends on the model and on the host being up. The approval would also need a different setup on each host. RUDDER's scheduler is plain code with the same checks as a manual change, and it runs the same on every host.
Results in the originating chat¶
A ticket may carry reply_to (<profile>/<platform>:<chat>[:<thread>]), filled by the MCP host, never by the
model. Every result the scheduler sends to the approval channel (applied, failed, stopped, missed) is then also
posted there with hermes -p <profile> send. Only profiles in APPROVAL.REPLY_PROFILES may be used (default
none = off). Put host-specific values like that in ~/.config/rudder/rudder.local.yml, which is merged over
rudder.yml (or point RUDDER_LOCAL at another file). Approving still happens only on the approval channel.
On Hermes, a small pre_tool_call plugin can fill reply_to from the session's origin.
Change pages¶
With APPROVAL.WEB_BIND set (e.g. 192.0.2.240:8790) the scheduler also serves the read-only change pages on that address, TCP 8790, directly (no proxy). Open that port in the firewall for the approvers' networks only. See APPROVALS.md.