The Agent Skills Directory

[COMMAND_EXECUTION]: The !exec YAML function allows for the execution of arbitrary shell scripts during stack configuration resolution.
Evidence: Found in SKILL.md and references/yaml-functions.md. It uses a POSIX-compatible interpreter to run multi-line scripts or single commands like !exec date +%s.
[EXTERNAL_DOWNLOADS]: The !include function supports fetching configuration or raw text from remote network locations.
Evidence: Supports multiple protocols including HTTPS, GitHub (github://), S3 (s3::), GCS (gcs::), SCP/SFTP, and OCI registries. Examples show fetching from raw.githubusercontent.com.
[DYNAMIC_EXECUTION]: The skill implements runtime code evaluation via !starlark and !template tags.
Evidence: !starlark executes Atmos Automation Language (Starlark) code directly from the component configuration. !template evaluates Go template expressions and converts JSON output to YAML types.
[DATA_EXFILTRATION]: Numerous functions provide access to sensitive environment data and credentials which could be exfiltrated if combined with network-capable tags.
Evidence: !secret (resolves declared secrets), !env (reads environment variables), !terraform.state (reads remote state backends including S3 and GCS), and !aws.* (account IDs, caller identity ARNs, and organization IDs).
[INDIRECT_PROMPT_INJECTION]: The skill provides a mechanism to ingest untrusted external data that can influence agent behavior.
Ingestion points: The !ci.labels function reads live CI pull request and merge-queue labels from GitHub.
Boundary markers: None documented; labels are processed as a sorted list or boolean flags.
Capability inventory: The skill provides highly privileged capabilities including !exec (shell execution), !include (remote file fetching), and !secret (credential access).
Sanitization: No sanitization or validation of the label content is described before it is injected into the configuration context.

atmos-yaml-functions