Operations skill

Unattended agent runs

Updated

Recover recurring work without repeating external actions or killing valid long runs.

Works in: Claude Code; Codex: selection smoke passed (October 2026).

Two run tiers

Fresh process

One bounded process per event. Use the autonomy kit as the starter path.

Persistent session

Use a single-flight lease, explicit done step, full-prompt gate, durable checkpoints, real session-loss probes, and a stall alert that keeps the lease.

Recovery contract

  1. Read every prompt page and record the gate.
  2. Checkpoint before and after each external action with exact IDs.
  3. After restart, use a fresh session and the bounded checkpoint.
  4. Inspect side-effect-pending state before any retry.
  5. Write the summary file, then run the explicit done step once.

Delivery and resource safety

An accepted send is not delivery. Verify a received copy and inspect an ambiguous result before retrying.

Cap memory, swap, and CPU at the session scope. Keep worker fan-out small and offload heavy compute.

Pin working and state directories, then prove persistence with a real reboot.

Authorization boundary

External actions require a written policy that states who may send or change what, to whom, and when. The project instruction file controls.

Install

Claude Code, as part of the newsroom-agent-ops plugin:

/plugin marketplace add jamditis/claude-skills-journalism
/plugin install newsroom-agent-ops@claude-skills-journalism

Codex, or any agent that reads Agent Skills from .agents/skills:

npx skills@latest add jamditis/claude-skills-journalism \
  --skill unattended-agent-runs --agent codex --copy -g -y