Cron Watchdog Authoring

Cron / watchdog job authoring

Class of work: recurring automation that acts when a condition flips (tool version released, PR changed, disk threshold crossed) and stays silent otherwise. Proven shapes for recurring deployments (tool rollouts, PR monitors).

Pitfall: schedule strings — "30m" is ONE-SHOT

cronjob(action=create, schedule="30m") produces "once in 30m", repeat=once — a single run. Recurring requires the literal prefix: schedule="every 30m", "every 2h", or a cron expression ("0 9 * * *"). After creating a job, CHECK the returned schedule/repeat fields and fix with cronjob(action=update) when they say once. (Measured 2026-08-07: created "30m", got one-shot; updated to "every 30m".)

The watchdog pattern (no_agent + script)

For deterministic jobs, skip the LLM: no_agent=true with script= (resolves under ~/.hermes/scripts/; .sh/.bash run via bash, everything else via Python).

• Silent unless action: empty stdout = silent tick (no delivery); non-empty stdout is delivered verbatim; non-zero exit sends an error alert. Design the script to print ONLY when it took real action; write diagnostics to a log file

(~/./rollout.log-style) instead of stdout.

• Exit 0 on transient failures (offline, not-yet-published): log to the file, stay silent — otherwise a laptop offline for a day spams one alert per tick. Reserve non-zero for genuinely broken states.
• Idempotent by construction: every tick must be a no-op when the end state already holds.
• Delivery caveat: CLI/TUI sessions have NO live-delivery channel — output is saved and viewable via cronjob(action=list). If the user wants a ping, deliver must name a gateway-connected platform (telegram, all).

Version-gated rollout watchdog (measured recipe)

Sequence per tick: read installed version → below target: run the update command (dotnet tool update --all --no-http-cache -g), re-read, still below → silent; at/above target → perform the one-time rollout:

1. Marker file for one-time side effects (~/./rollout--done.marker): the kill+restart+settings-arming happens only on the FIRST tick at target version. Later ticks only ensure the service is still up (self-healing).

Without the marker, every tick re-arms settings and fights later manual changes.

1. Surgical replacement of the old server: lsof -ti : → SIGTERM only the port holder — never pkill -f " serve" (matches scratch servers on other ports and can kill the user's other data-root instances). Then start the

new binary detached (subprocess.Popen(..., start_new_session=True) in Python, nohup-style), log to a file, poll the port up to ~20s.

1. Settings arming once: run the config verbs inside the marker branch only.
2. Version parse: tool --version prints SemVer+commit hash — regex the dotted triple; compare tuples. Verify the parse against the real binary before trusting the script.
3. Python compatibility: scripts may run under the macOS system python3 (3.9.x) — add from __future__ import annotations FIRST (before imports) or avoid X | None annotations entirely; tuple[int,int,int] | None raises TypeError at

definition time on 3.9.

Pitfall: verify CLI verb paths BEFORE the rollout branch can execute them

Draft scripts often assume a config parent verb ( config extract enable true). Verb families are frequently ROOT-level ( extract enable true) — there is no config command at all, and the mistake only surfaces when the rollout branch actually runs ("Unrecognized command or argument 'config'"). Caught this way 2026-08-07: the settings-arming verbs were wrong from the first commit and the error stayed invisible through smoke tests that never reached the rollout branch. Fix: during the foreground verification pass, run EVERY command the script will execute (or its --help) against the real binary — a wrong verb path fails silently in a log-only watchdog.

Pitfall: the scheduler may never tick — run the script manually when the trigger condition is met

A job can sit with next_run_at in the past and last_run_at: null after creation/update (observed 2026-08-07 on an 'every 30m' job; the update also left repeat: once). Treat the cron as a self-healing layer, NOT the execution guarantee: when the user reports the trigger condition (e.g. "we updated to 1.1.0"), run the script in the foreground instead of waiting for the tick.

Verification before scheduling

Run the script once in the foreground first. Expect the "nothing to report" path to be silent + logged; the action path can be exercised with a lowered TARGET only if the real side effects are safe (port kill is NOT safe to dry-run against a live server — verify the rollout branch by review instead).

Gotchas

• Bare "30m" is ONE-SHOT, not every-30-minutes — recurring schedules need "every 30m".
• Verify CLI verb paths BEFORE the rollout branch can run them — a typo'd verb fails silently in the watchdog.

Related

• The bank-maintenance design this rollout supports lives in the repo's work docs

(the maintenance-fix shape); keep a pointer to it beside the rollout script.

• Existing instances: -rollout (every 30m) + a PR monitor (30m, local-deliver).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Arasz
• Source: Arasz/ai-badger
• License: MIT
• Homepage: https://github.com/Arasz/ai-badger

Install and usage instructions live in the source repository linked above.