cron-worker-guardrails（cron 工作者护栏）

Cron Worker Guardrails (POSIX)

A reliability-first checklist for OpenClaw cron workers and any unattended automation.

Scope (important)

• This skill is POSIX-focused (bash/sh examples).
• The *principles* are portable, but if you're on Windows/PowerShell you'll need equivalent patterns.

The NO_REPLY convention

Many OpenClaw setups treat emitting exactly NO_REPLY as "silent success" (no human notification).

• If your runtime does not support NO_REPLY, interpret it as: print nothing on success.

Quick Start

1) Scripts-first: move logic into a repo script (recommended: tools/<job>.py or tools/<job>.sh). 2) One command in cron: cron should run *one short command* (no multi-line bash -lc '...'). 3) Deterministic cwd/env: cd to the repo (or have the script do it), and document required env vars. 4) Silent on success: print nothing (or exactly NO_REPLY) when OK; only emit a short alert when broken.

Also see:

• references/cron-agent-contract.md
• references/pitfalls.md

Why this skill exists

Cron failures are rarely "logic bugs". In practice they're often:

• brittle shell quoting (bash -lc '...' nested quotes)
• command substitution surprises ($(...))
• one-liners that hide escaping bugs (python -c "...")
• cwd/env drift ("works locally, fails in cron")
• pipelines that fail for the wrong reason (pipefail + head / SIGPIPE)

The fix is boring but effective: scripts-first + deterministic execution + silent-on-success.

Portability rules (still apply)

Even on POSIX, do not hardcode deployment-specific absolute paths tied to one machine.

Prefer:

• repo-relative paths
• environment variables you document
• minimal wrappers that cd into the repo

Common failure patterns -> fixes

1) unexpected EOF while looking for matching ')'

Likely causes:

• unclosed $(...) from command substitution
• broken nested quotes in bash -lc ' ... '

Fix pattern:

• Replace the whole multi-line shell block with a script.
• Cron calls exactly one short command, for example:

- python3 tools/<job>.py

2) False failure from pipefail + head (SIGPIPE)

Symptom:

• command exits non-zero even though the output you wanted is fine

Fix pattern:

• avoid pipefail when piping into head
• or better: do the filtering in a script (read only what you need)

3) "Works locally, fails in cron"

Common causes:

• wrong working directory
• missing env vars
• different PATH

Fix pattern:

• cd into the repo (or have the script do it)
• keep dependencies explicit and documented

Git footgun: git push rejected (non-fast-forward)

Symptom:

• ! [rejected] ... (non-fast-forward) when automation pushes to a long-lived PR/feature branch.

Conservative fix (no force-push):

• On rejection, fetch the remote branch, transplant your new local commits onto it (cherry-pick), then retry push once.

Copy/paste hardening header (portable)

Use this near the top of a cron prompt (2 lines, low-noise):

• Hardening (MUST): follow references/cron-agent-contract.md (scripts-first, deterministic cwd, silent-on-success).
• Also apply the cron-worker-guardrails skill. If parsing/multi-step logic is needed, write/run a small tools/*.py script.