A skill is a packaged playbook an agent installs and follows. Ours teach the agent the preloop controls — how to run and debug jobs, how to lock a workflow down, and how to read what CI actually did — and the API calls that drive them.
The core playbook: submit runs, stream events, pause at a failed step, apply a fix, retry from that step, and drive runs on the hosted pool. Use when the user asks to run a workflow or fix a failing job.
An audit playbook built on the known CI/CD attack classes: dangerous triggers and pwn requests, template injection, secret bindings, artifact and cache poisoning, action pinning with actions.lock, OIDC trust, and runner hijack. Every finding maps to the preloop control that closes it.
A measurement playbook: read what every job and test actually cost, which steps were reused, what the cache returned, and which tests a change could not affect. Prune the work that cannot change the outcome and show the savings on the next run.
Skills are SKILL.md playbooks your agent runtime installs. They run through the same API documented on the For agents page — no SDK required.
Each skill is a SKILL.md directory in the public preloop-skill repository under skills/. Install all three into every detected agent runtime, or copy a single directory by hand.
# All skills, all detected agents
npx skills add preloopdev/skills
# One skill, one runtime
npx skills use preloopdev/skills@preloop-secure
# Or by hand
cp -r preloop-skill/skills/preloop-secure ~/.config/your-agent/skills/The skill needs a credential to act on: an agent account signed in with federated OIDC or a single-use assertion, or a scoped token delegated by an org owner. ci-agent tokens can run and debug their own jobs but carry no secrets, no OIDC and no approvals — the boundary is enforced by the token, not by the skill.
| Skill | What it does | Where |
|---|---|---|
| preloop | Run workflows, pause at failures, sync edits, retry from a step | skills/preloop |
| preloop-secure | Audit workflows and apply the control that closes each finding | skills/preloop-secure |
| preloop-optimize | Measure job, step, cache and test cost; prune unaffected work | skills/preloop-optimize |
Everything a skill does is an API call. If a runtime cannot use the skills directly, the same operations are documented on the For agents page and in the API reference.
A SKILL.md file: a named playbook with a description the agent matches to a task, the rules it must follow, and the commands it may run. Install it into your runtime and the agent invokes it when a task matches.
Whatever the skill teaches and the agent credential allows. preloop covers running and debugging; preloop-secure audits and hardens; preloop-optimize measures and prunes. ci-agent tokens carry no secrets, no OIDC and no approvals.
No. They drive the API and the CLI. Everything the dashboard does is an API call, so a skill can do the same under its own credential.
Install a skill, hand the agent a scoped credential, and let it drive the loop.