evi-cron-self-revival
Use when scheduled tasks must survive a session restart — a registry plus a SessionStart self-poke that re-registers in-session cron jobs unattended.
SKILL.mdskills/cron-self-revival/SKILL.md
GitHubCron self-revival
In-session cron (CronCreate) dies with the session: a restart wipes the whole
schedule. This skill restores it without a human in the loop, by combining a
registry file (single source of truth), a SessionStart instruction injection,
and a tmux send-keys self-poke that converts the injected instruction into an
actual turn.
What it guarantees
After any session restart, every recurring job in tasks/cron-registry.json is
re-registered automatically. The agent does not wake up having forgotten its
schedule, and no owner message is needed to trigger the restore.
How it works
On startup, session-start-restore.py:
- Reads
tasks/cron-registry.jsonand injects a re-registration instruction block (runCronList, thenCronCreatefor any missing job) intoadditionalContext— the fallback. - Spawns a detached background process that, after
EVI_SELF_POKE_DELAYseconds, sends/recron+ Enter to the tmux pane — the driver.
The poke matters because injected context is not a turn: an agent that only
reads "please re-register your jobs" stays idle until something pokes it. The
/recron command (../../.claude/commands/recron.md)
then reads the registry, runs CronList, registers what is missing, and writes
a HOOK-FIRE → POKE-SCHEDULED → REGISTERED → DONE audit trail.
Field lessons (tmux / launchd send-keys automation)
These were paid for in real operation — keep them in any send-keys driver:
- A poke during an active turn loses the Enter. While the REPL is busy the Enter keystroke is silently swallowed and the command never submits. Mitigation: re-poke until a start marker appears in the log (the driven command writes one as its first action); do not assume a single poke landed.
- tmux send-keys can hang under launchd. A launchd-launched send-keys call
has been observed to block for hours, starving the next interval. Mitigation:
wrap every send-keys in a background ~10-second watchdog that
kill -9s on breach. - Send the text and Enter as separate send-keys calls. A
/-prefixed input opens the slash-command autocomplete popup; sending Enter in the same call races the popup and is dropped. Send the text, sleep ~1.5 s, then send Enter.
Implement in your project
Reference implementations:
../../scripts/hooks/session-start-restore.py(injection + self-poke)../../.claude/commands/recron.md(/recronprocedure)../../tasks/cron-registry.json(registry schema)
Wire the SessionStart hook (shared with the session-restore skill — one entry
does both):
"SessionStart": [
{ "matcher": "", "hooks": [
{ "type": "command", "command": "EVI_PROJECT_DIR=$CLAUDE_PROJECT_DIR EVI_TMUX_BIN=$EVI_TMUX_BIN EVI_TMUX_TARGET=$EVI_TMUX_TARGET EVI_SELF_POKE_DELAY=$EVI_SELF_POKE_DELAY python3 $CLAUDE_PROJECT_DIR/scripts/hooks/session-start-restore.py" }
]}
]
Set EVI_TMUX_TARGET (find it with tmux list-panes -a), EVI_TMUX_BIN, and
EVI_SELF_POKE_DELAY in .env. Populate tasks/cron-registry.json with your
jobs; point each prompt at a tasks/<name>.md procedure.
Adaptation points
- No tmux: keep the
additionalContextfallback and let the owner type/recronafter restart, or trigger it from your own supervisor. You lose the unattended property but keep correctness. - systemd instead of launchd: the hook itself is unaffected (it spawns a child process, not a service). Field lesson 2 still applies to any service-run send-keys.
- Durable cron exists in your runtime: the registry +
/recronremain the recovery path; treat durable persistence as a bonus, not a guarantee.
Supervise the process itself
This skill revives schedules. The REPL process and your MCP servers also
die — give them a supervisor: launchd KeepAlive (see launchd/ for the
plist shape) or systemd Restart=always. Assume every process crashes;
design the resurrection path before you design the personality.
Verify
python3 -m py_compile scripts/hooks/session-start-restore.pyexits 0.python3 -c "import json; print(len(json.load(open('tasks/cron-registry.json'))['jobs']))"prints the job count (registry is valid JSON).- Restart the session, wait
EVI_SELF_POKE_DELAY + 5s, then checkdocs/work-log/cron-reregister.logforREGISTERED ... / DONE N/N. Note: the audit log (docs/work-log/cron-reregister.log) is created on first run by the session-restore hook — it does not need to pre-exist.
Rollback
Remove the SessionStart entry from .claude/settings.json (and delete
.claude/commands/recron.md if you want the command gone).