The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to operate a command-line interface (CLI) tool named sjskills. It instructs the agent to execute various subcommands including status, plan, apply, init, and restore. The instructions guide the agent to use either a system-installed version on the PATH or a local version located at bin/sjskills within the tool's source repository.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection as it ingests untrusted data from project manifests (sjskills.toml), status reports (plan.json), and external registries.
Ingestion points: External configuration files (sjskills.toml), temporary plan artifacts generated by the CLI (plan.json), and content fetched from remote registries.
Boundary markers: The skill explicitly instructs the agent to "Review every operation, warning, current-state fact... before mutation" and to "Summarize the reviewed operations before applying," providing a layer of manual/cognitive boundary.
Capability inventory: The agent has the capability to execute shell commands via the sjskills binary, as seen in SKILL.md and references/global-rollout.md.
Sanitization: The skill employs integrity checks for global rollouts by requiring the agent to compute and verify SHA-256 digests of the plan artifacts (--approved-plan-sha256), which mitigates the risk of processing tampered or malicious plan data.

sjskills